Browse the corpus

Key messages Screen-detected diabetes is usually asymptomatic, but individuals often have multimorbidity. Just under half of individuals with screen-detected diabetes are on drugs not related to cardiovascular disease at diagnosis. Many individuals did not start glucose lowering medication in the 5 years after diagnosis. Introduction Medication burden is high among individuals with established type 2 diabetes. Results from a systematic review indicate that patients with diabetes take in the range of 4 to 10 medications a day.1 In an American study of 875 individuals with diabetes, 50% reported taking seven or more prescription medications a day.2 Estimates from English patients with diabetes suggest an average of six medications a day.3 Individuals with diabetes are prescribed a number of cardioprotective drugs, but there is also evidence to suggest high levels of prescription of other drug classes for example, treatment for neuropathy,4 depression,5 and gastric and rheumatological symptoms.6 In 2012–2013 in England, 9.3% of the total cost of prescriptions in the National Health Service (NHS) was related to diabetes.7 As treatment regimens become more complex, patients are more likely to experience adverse side effects8 and less likely to remain adherent to all prescribed medications.9 10

logical symptoms.6 In 2012–2013 in England, 9.3% of the total cost of prescriptions in the National Health Service (NHS) was related to diabetes.7 As treatment regimens become more complex, patients are more likely to experience adverse side effects8 and less likely to remain adherent to all prescribed medications.9 10 Less is known about treatment burden among individuals with screen-detected or recently diagnosed diabetes. Given that population screening for diabetes has been recommended by several national organizations and the NHS currently includes assessment of risk of diabetes in its Health Checks program,11 more individuals will be found earlier in the disease trajectory. Further, there is growing evidence for the benefit of intensive treatment of risk factors early in the course of the disease,12 13 which suggests that screen-detected patients may be prescribed a larger number of cardioprotective drugs earlier than they might previously have been. Although there is some evidence that improved medication adherence may improve health-related quality of life in symptomatic patients with diabetes,14 15 individuals earlier in the disease trajectory are unlikely to have symptoms and may be less likely to adhere to complex medication regimes.16 17 There is currently little knowledge of medication burden in people with screen-detected diabetes, many of whom will have few or no symptoms. Guidelines promote a multifactorial approach to diabetes care from diagnosis that includes pharmacotherapy for multiple cardiovascular disease (CVD)-related conditions.18 19 Despite the increasing number of individuals with screen-detected diabetes, many of whom have comorbidities, there is an absence of knowledge about what the pharmacotherapy burden is at diagnosis in this population, and how it changes in the first 5 years. It is important that this is described so that patients and practitioners are informed about the likely course and burden of treatment. We aimed to (1) describe medication burden at diagnosis, 1 and 5 years in individuals with screen-detected diabetes and (2) examine if age, sex, intensive treatment or modeled 10-year CVD risk was associated with the number of drugs individuals were prescribed at 5 years after diagnosis.

course and burden of treatment. We aimed to (1) describe medication burden at diagnosis, 1 and 5 years in individuals with screen-detected diabetes and (2) examine if age, sex, intensive treatment or modeled 10-year CVD risk was associated with the number of drugs individuals were prescribed at 5 years after diagnosis. Methods The ADDITION study is a primary care-based screening and intervention study for type 2 diabetes (ClinicalTrials.gov, CNT00237549). It was carried out in Denmark, the Netherlands and two UK centers (Leicester and Cambridge). The study has been described in detail elsewhere.13 20 21 In this paper we describe data from the two UK centers. Briefly, individuals aged 40–69 years in Leicester were invited to attend an Oral Glucose Tolerance Test (OGTT). Individuals in Cambridge aged 40–69 years with a high risk of diabetes in Cambridge (Cambridge Risk Score22 ≥0.17) were invited to a stepwise screening program including a random capillary glucose test and glycated hemoglobin, followed by a fasting capillary glucose test and a confirmatory OGTT. Individuals were diagnosed using the WHO 1999 definition of diabetes.23 Exclusion criteria included pregnancy, lactation, an illness with a likely prognosis of less than 1 year or a psychiatric illness likely to limit study involvement or invalidate informed consent. Individuals found to have diabetes were treated according to the group to which their practice was allocated: routine care using national guidelines19 or the promotion of intensive multifactorial treatment. In the intensive treatment group, general practitioners (GPs) were encouraged through guidelines, educational meetings and audits with feedback to introduce a stepwise target-led drug treatment regime to reduce hyperglycemia, hypertension and hyperlipidaemia20 21 similar to the STENO-2 study.24 Trained staff-assessed patients’ health at baseline, 1 and 5 years and collected biochemical and anthropometric data according to standard operating procedures. Self-report questionnaires were used to collect information on sociodemographic information, lifestyle habits and medication use. The study was approved by the relevant ethics committees13 20 21 and all participants provided written informed consent.

l and anthropometric data according to standard operating procedures. Self-report questionnaires were used to collect information on sociodemographic information, lifestyle habits and medication use. The study was approved by the relevant ethics committees13 20 21 and all participants provided written informed consent. Assessment of medication In Cambridge, participants were encouraged to bring their repeat prescription summaries to each health assessment and self-reported medication was collected via a health economics questionnaire which asks for information on all prescribed medication.25 In Leicester, prescription information could also be sourced directly from the records of a peripatetic clinic. Medication data were coded using the Anatomical Therapeutic Chemical Classification System (ATC).26 ATC codes were used to derive counts for each participant within the following 23 classes of medication: insulin, metformin, sulphonylurea, thiazolidinediones, other glucose lowering medication, ace-inhibitors, β-blockers, calcium channel blockers, diuretics, other blood pressure lowering medications, lipid lowering, antithrombotic, gastrointestinal-related, skin-related, hormone-replacement therapy or urogenital, systemic steroids, thyroid-related, anti-inflammatory, analgesic, antiepileptic, psychiatric, respiratory and eye-related. Medication counts in this analysis refer to the number of the 23 classes prescribed (not overall pill count), while medication agent refers to 1 of the 23 explored classes of medication. For several analyses, these 23 categories were also collapsed into diabetes-related (insulin, metformin, sulphonylurea, thiazolidinediones, other glucose-lowering medication), cardioprotective (ace-inhibitors, β-blockers, calcium channel blockers, diuretics, other hypertension-related medications, lipid-lowering, antithrombotic) and other (remaining 11 classes). Medication types that were not within these categories, for example acute medications like antibiotics, were not included in these analyses.

tective (ace-inhibitors, β-blockers, calcium channel blockers, diuretics, other hypertension-related medications, lipid-lowering, antithrombotic) and other (remaining 11 classes). Medication types that were not within these categories, for example acute medications like antibiotics, were not included in these analyses. Statistical analysis Baseline and 5-year descriptive characteristics of the cohort were summarized using means, medians and proportions. We described the medication profile of the ADDITION-UK cohort at diagnosis, 1 and 5 years following diagnosis. Using complete case linear regression, we explored the mutually adjusted associations between age, baseline 10-year UK Prospective Diabetes Study (UKPDS) CVD risk,27 sex, treatment group and baseline number of medications on (1) change in total number of medications, (2) change in cardioprotective medications and (3) change in other medications between diagnosis and 5 years. Owing to the distribution of change in diabetes-related medication being left-censored at zero an analogous Poisson regression model was used to explore the association between baseline predictors and change in diabetes-related medication over 5 years. SEs were used to adjust for clustering by GP practice in the models. As current guidelines for the treatment of type 2 diabetes are very similar to the protocol used in the intensive treatment arm of ADDITION-UK, and the achieved difference in treatment was small, treatment arms were pooled for the primary analysis.13 28 A sensitivity analyze by randomization arm showed little differences relative to overall changes.

reatment of type 2 diabetes are very similar to the protocol used in the intensive treatment arm of ADDITION-UK, and the achieved difference in treatment was small, treatment arms were pooled for the primary analysis.13 28 A sensitivity analyze by randomization arm showed little differences relative to overall changes. In order to characterize missing data, we used logistic regression models to derive the odds of being included in the complete-case analysis, individually adjusted for age, sex, baseline UKPDS 10-year CVD risk, treatment group and 2004 indices of multiple deprivation (IMD). IMD scores were only available for the 867 individuals (86% of the sample) from the Cambridge area, so the association between missing data and socioeconomic status is described using a smaller data set for this sensitivity analysis.

KPDS 10-year CVD risk, treatment group and 2004 indices of multiple deprivation (IMD). IMD scores were only available for the 867 individuals (86% of the sample) from the Cambridge area, so the association between missing data and socioeconomic status is described using a smaller data set for this sensitivity analysis. The small differences in the outcome and treatment between routine care and intensive treatment in ADDITION-Europe has been linked to the continual improvement of routine care, most likely accelerated through the introduction of the Diabetes National Service Framework in 2001,29 clinical guidelines for targeting blood pressure and lipids in people with diabetes in 2002,19 and the Quality and Outcomes Framework in 2004.13 29 Current guidelines for the treatment of type 2 diabetes are similar to the protocol used in the intensive treatment arm of ADDITION-UK.13 28 As such, while a statistically significant difference in cardioprotective and glucose-lowering drugs is present, absolute differences in the prevalence of medications being reported are small, which is why treatment arms were pooled in this analysis. Statistical analyses were performed using R 3.0.2 (checkpoint 2014-09-18).

The small differences in the outcome and treatment between routine care and intensive treatment in ADDITION-Europe has been linked to the continual improvement of routine care, most likely accelerated through the introduction of the Diabetes National Service Framework in 2001,29 clinical guidelines for targeting blood pressure and lipids in people with diabetes in 2002,19 and the Quality and Outcomes Framework in 2004.13 29 Current guidelines for the treatment of type 2 diabetes are similar to the protocol used in the intensive treatment arm of ADDITION-UK.13 28 As such, while a statistically significant difference in cardioprotective and glucose-lowering drugs is present, absolute differences in the prevalence of medications being reported are small, which is why treatment arms were pooled in this analysis. Statistical analyses were performed using R 3.0.2 (checkpoint 2014-09-18). Results At diagnosis, the ADDITION-UK cohort had a mean age of 61 years (SD 7), a median UKPDS 10-year CVD risk of 19% (IQR 13, 27) and 61% were male (tables 1 and 2). Of the 1026 individuals in the ADDITION-UK cohort, 1024 (99.8%) had medication data at diagnosis, 1008 (99% of living) at 1 year, and 930 (96% of living) at 5 years. Ten people died before 1 year follow-up, and 59 before 5-year follow-up. Table 1 Baseline characteristics of the ADDITION-UK cohort, overall and by previous CVD status and CVD risk quartile

Results At diagnosis, the ADDITION-UK cohort had a mean age of 61 years (SD 7), a median UKPDS 10-year CVD risk of 19% (IQR 13, 27) and 61% were male (tables 1 and 2). Of the 1026 individuals in the ADDITION-UK cohort, 1024 (99.8%) had medication data at diagnosis, 1008 (99% of living) at 1 year, and 930 (96% of living) at 5 years. Ten people died before 1 year follow-up, and 59 before 5-year follow-up. Table 1 Baseline characteristics of the ADDITION-UK cohort, overall and by previous CVD status and CVD risk quartile 10-year UKPDS CVD risk: Lowest quartile 5,17 10-year UKPDS CVD risk: Highest quartile 3692 No CVD Previous CVD* Total N† 244 244 858 106 1026 Mean age in years (SD) 55.6 (7.5) 64.2 (5.3) 60.3 (7.5) 63.1 (5.3) 60.6 (7.4) Male % 40% 83% 60% 74% 61% White % 80% 98% 93% 96% 91% Median 10-year CVD risk (IQR) 14 (11, 15) 47 (40, 56) 24 (17, 33) 45 (35, 56) 25 (17, 36) Mean BMI kg/m2 (SD) 32.8 (5.8) 33.0 (5.8) 33.3 (5.7) 32.9 (6.1) 30.8 (5.4) Mean HbA1C % 6.6 (1.1) 8.3 (2.2) 7.4 (1.7) 7.1 (1.6) 7.3 (1.7) Mean HbA1C mmol/mol 49 (12) 68 (24) 57 (19) 53 (17) 57 (18) Mean systolic BP mm Hg (SD) 133 (16) 153 (23) 143 (19) 139 (22) 146 (17) Mean total cholesterol mmol/L (SD) 5.2 (1.0) 5.5 (1.3) 5.5 (1.1) 4.6 (1.0) 5.6 (1.2) Self-reported CVD* % 1% 30% 0% 100% 11% Self-reported high-blood pressure % 60% 55% 57% 68% 59% Self-reported high cholesterol % 27% 31% 23% 68% 28% *Previous myocardial infarction or stroke. †Number of participants recruited at diagnosis.

10-year UKPDS CVD risk: Lowest quartile 5,17 10-year UKPDS CVD risk: Highest quartile 3692 No CVD Previous CVD* Total N† 244 244 858 106 1026 Mean age in years (SD) 55.6 (7.5) 64.2 (5.3) 60.3 (7.5) 63.1 (5.3) 60.6 (7.4) Male % 40% 83% 60% 74% 61% White % 80% 98% 93% 96% 91% Median 10-year CVD risk (IQR) 14 (11, 15) 47 (40, 56) 24 (17, 33) 45 (35, 56) 25 (17, 36) Mean BMI kg/m2 (SD) 32.8 (5.8) 33.0 (5.8) 33.3 (5.7) 32.9 (6.1) 30.8 (5.4) Mean HbA1C % 6.6 (1.1) 8.3 (2.2) 7.4 (1.7) 7.1 (1.6) 7.3 (1.7) Mean HbA1C mmol/mol 49 (12) 68 (24) 57 (19) 53 (17) 57 (18) Mean systolic BP mm Hg (SD) 133 (16) 153 (23) 143 (19) 139 (22) 146 (17) Mean total cholesterol mmol/L (SD) 5.2 (1.0) 5.5 (1.3) 5.5 (1.1) 4.6 (1.0) 5.6 (1.2) Self-reported CVD* % 1% 30% 0% 100% 11% Self-reported high-blood pressure % 60% 55% 57% 68% 59% Self-reported high cholesterol % 27% 31% 23% 68% 28% *Previous myocardial infarction or stroke. †Number of participants recruited at diagnosis. BMI, body mass index; BP, blood pressure; CVD, cardiovascular disease; HbA1c, glycated hemoglobin; UKPDS, UK Prospective Diabetes Study. Table 2 Association between baseline characteristics at diagnosis and change in medication count between diagnosis and 5 years in the ADDITION-UK cohort

†Number of participants recruited at diagnosis. BMI, body mass index; BP, blood pressure; CVD, cardiovascular disease; HbA1c, glycated hemoglobin; UKPDS, UK Prospective Diabetes Study. Table 2 Association between baseline characteristics at diagnosis and change in medication count between diagnosis and 5 years in the ADDITION-UK cohort Change in total medication count Change in diabetes medication Change in CVD medication Change in other medication β* 95% CI IRR* 95% CI β* 95% CI β* 95% CI Number of medications at diagnosis† −0.49 −0.56 to −0.42 – – −0.50 −0.56 to −0.44 −0.30 −0.37 to −0.22 Male gender −0.25 −0.57 to 0.06 0.86 0.75 to 0.99 −0.11 −0.33 to 0.10 0.12 −0.10 to 0.34 Intensive treatment arm 0.44 0.10 to 0.78 1.14 1.01 to 1.30 0.39 0.09 to 0.69 −0.08 −0.30 to 0.13 Age at diagnosis (years) −0.03 −0.05 to −0.01 0.96 0.95 to 0.97 −0.02 −0.03 to 0.002 0.02 0.01 to 0.04 Modelling 10-year UKPDS CVD risk (%) 0.04 0.02 to 0.05 1.02 1.01 to 1.03 0.02 0.01 to 0.03 0.00 −0.01 to 0.01 *IRR, Incidence Rate Ratio from a Poisson regression model; β, Regression coefficient from a linear regression model. †Number of medications of the medication type that is the dependent variable in that columns regression. CVD, cardiovascular disease; UKPDS, UK Prospective Diabetes Study.

Change in total medication count Change in diabetes medication Change in CVD medication Change in other medication β* 95% CI IRR* 95% CI β* 95% CI β* 95% CI Number of medications at diagnosis† −0.49 −0.56 to −0.42 – – −0.50 −0.56 to −0.44 −0.30 −0.37 to −0.22 Male gender −0.25 −0.57 to 0.06 0.86 0.75 to 0.99 −0.11 −0.33 to 0.10 0.12 −0.10 to 0.34 Intensive treatment arm 0.44 0.10 to 0.78 1.14 1.01 to 1.30 0.39 0.09 to 0.69 −0.08 −0.30 to 0.13 Age at diagnosis (years) −0.03 −0.05 to −0.01 0.96 0.95 to 0.97 −0.02 −0.03 to 0.002 0.02 0.01 to 0.04 Modelling 10-year UKPDS CVD risk (%) 0.04 0.02 to 0.05 1.02 1.01 to 1.03 0.02 0.01 to 0.03 0.00 −0.01 to 0.01 *IRR, Incidence Rate Ratio from a Poisson regression model; β, Regression coefficient from a linear regression model. †Number of medications of the medication type that is the dependent variable in that columns regression. CVD, cardiovascular disease; UKPDS, UK Prospective Diabetes Study. Total medication burden At diagnosis, most individuals reported taking two medications (median 2; IQR 0, 4). This was most commonly a cardioprotective medication (median 1; IQR 0, 3), although some individuals were on more than one non-cardioprotective medication at diagnosis (figure 1). One year after diagnosis a median of 3 medications (IQR 0,6) were recorded. At 5 years, individuals were typically prescribed six medications (median 6; IQR 5, 8), which included one diabetes-related medication (median 1; IQR 0, 1), four cardioprotective medications (median 4; IQR 3, 5) and one other medication (median 1; IQR 0, 2).

diagnosis a median of 3 medications (IQR 0,6) were recorded. At 5 years, individuals were typically prescribed six medications (median 6; IQR 5, 8), which included one diabetes-related medication (median 1; IQR 0, 1), four cardioprotective medications (median 4; IQR 3, 5) and one other medication (median 1; IQR 0, 2). Figure 1 Proportions of self-reported medication use in the ADDITION-UK cohort at diagnosis, 1 and 5 years. COPD, chronic obstructive pulmonary disease; CVD, cardiovascular disease; HRT, hormone replacement therapy. Diabetes-related and cardioprotective medication After diagnosis, both the variety and number of cardioprotective and diabetes medications increased (figure 2). At 1 year, 23% of individuals were prescribed any type of diabetes medication, which increased to 62% at 5 years. Between diagnosis, 1 and 5 years, the prescription of antihypertensive (55% to 51% to 77%), lipid-lowering (24% to 48% to 81%) and anti-thrombotic (20% to 36% to 54%) medication increased. In this screen-detected population, many individuals reported using no glucose lowering medication at 1 and 5 years (78% and 38%, respectively, figures 1 and 2). Figure 2 Count of medication types reported in the ADDITION-UK cohort at diagnosis, 1 and 5 years. CVD, cardiovascular disease.

Diabetes-related and cardioprotective medication After diagnosis, both the variety and number of cardioprotective and diabetes medications increased (figure 2). At 1 year, 23% of individuals were prescribed any type of diabetes medication, which increased to 62% at 5 years. Between diagnosis, 1 and 5 years, the prescription of antihypertensive (55% to 51% to 77%), lipid-lowering (24% to 48% to 81%) and anti-thrombotic (20% to 36% to 54%) medication increased. In this screen-detected population, many individuals reported using no glucose lowering medication at 1 and 5 years (78% and 38%, respectively, figures 1 and 2). Figure 2 Count of medication types reported in the ADDITION-UK cohort at diagnosis, 1 and 5 years. CVD, cardiovascular disease. Other medications At diagnosis, 42% of individuals were prescribed other types of medication, which increased to 62% at 5 years after diabetes diagnosis (figure 2). The most common was for gastrointestinal conditions (13% at diagnosis, and 25% at 5 years). Many individuals also reported anti-inflammatory (12% at diagnosis, and 12% at 5 years), analgesic (12% at diagnosis, and 19% at 5 years) and psychotherapy (11% at diagnosis, and 15% at 5 years)-related prescriptions.

figure 2). The most common was for gastrointestinal conditions (13% at diagnosis, and 25% at 5 years). Many individuals also reported anti-inflammatory (12% at diagnosis, and 12% at 5 years), analgesic (12% at diagnosis, and 19% at 5 years) and psychotherapy (11% at diagnosis, and 15% at 5 years)-related prescriptions. Association between baseline characteristics and number of prescribed drugs at 5 years The baseline characteristics associated with an increase in the total number of prescribed drugs between diagnosis and 5 years were a younger age (β −0.03, 95% CI −0.05 to −0.01), a higher baseline modeled 10-year UKPDS CVD risk score (β 0.04, 95% CI 0.04, 95% CI 0.02 to 0.05), randomisation to the intensive treatment arm of the trial (β 0.44, 95% CI 0.01 to 0.78), and being prescribed less medications at diagnosis (β −0.49, 95% CI −0.56 to −0.42). Sex was not associated with change in total number of medications. Similarly, the baseline characteristics associated with an increase in cardioprotective medication were a higher 10-year CVD risk (β 0.02, 95% CI 0.01 to 0.02), randomization to the intensive treatment arm (β 0.39, 95% CI 0.09 to 0.69) and being prescribed less medication at baseline (β −0.50, 95% CI −0.56 to −0.44). An increase in diabetes-related medication was associated with female sex (Incidence Rate Ratio, IRR 0.86, 95% CI 0.75 to 0.99), younger age (years; IRR 0.96, 95% CI 0.95 to 0.97), having a higher baseline 10-year CVD risk (IRR 1.02, 95% CI 1.01 to 1.02) and randomization to the intensive treatment arm (IRR 1.15, 95% CI 0.01 to 1.30).

iabetes-related medication was associated with female sex (Incidence Rate Ratio, IRR 0.86, 95% CI 0.75 to 0.99), younger age (years; IRR 0.96, 95% CI 0.95 to 0.97), having a higher baseline 10-year CVD risk (IRR 1.02, 95% CI 1.01 to 1.02) and randomization to the intensive treatment arm (IRR 1.15, 95% CI 0.01 to 1.30). Compared to individuals with medication data at 5 years, those without medication data were more likely to be female (OR 0.56; 95% CI 0.35 to 0.89), older (1 year; OR 0.97; 0.94 to 0.999), to have had a previous CVD event (OR 0.49; 95% CI 0.29 to 0.90) and to be in the intensive arm of the trial (OR 2.04; 95% CI 1.32 to 3.20). There was no association between loss to follow-up and ethnicity (White vs other; OR 0.77; 95% CI 0.31 to 1.60) or socioeconomic deprivation (1 point IMD 2004 change; OR 0.99; 95% CI 0.97 to 1.02).

vious CVD event (OR 0.49; 95% CI 0.29 to 0.90) and to be in the intensive arm of the trial (OR 2.04; 95% CI 1.32 to 3.20). There was no association between loss to follow-up and ethnicity (White vs other; OR 0.77; 95% CI 0.31 to 1.60) or socioeconomic deprivation (1 point IMD 2004 change; OR 0.99; 95% CI 0.97 to 1.02). Discussion In a population of individuals with screen-detected type 2 diabetes, we described the prevalence of diabetes-related, cardioprotective and other medications at diagnosis, 1 and 5 years post-diagnosis. Many individuals were on medications not related to cardioprotection before diagnosis (42%), and this increased along with a rise in the number of diabetes-related and cardioprotective drugs. At 5 years, individuals were typically prescribed six medications, including one diabetes-related medication, four cardioprotective medications, and one other medication. This suggests that there is a significant degree of multimorbidity and polypharmacy present in individuals with screen-detected diabetes. Following diagnosis, individuals were more likely to be prescribed diabetes-related medication if they were younger, female, had a high modeled CVD and if they were randomized to the intensive treatment arm of the trial. Younger individuals being prescribed more total and diabetes medication in the 5 years after diagnosis is in line with previous literature that identified those with early diabetes as having worse glycaemic control elevated and CVD risk factors.30 In older individuals, the balance between treatment benefits and harm may also become less clear, which could also lead to the identified association. Higher modeled CVD risk at baseline was associated with a greater increase in cardioprotective medication, but not an increase in other medications. As recommended in national guidelines, our results suggest that the treatment of diabetes was influenced by the underlying risk of CVD.

lead to the identified association. Higher modeled CVD risk at baseline was associated with a greater increase in cardioprotective medication, but not an increase in other medications. As recommended in national guidelines, our results suggest that the treatment of diabetes was influenced by the underlying risk of CVD. This is the first description of total medication burden in a large cohort of individuals with screen-detected diabetes over 5 years of follow-up. In a subset of the Dutch Hoorn Study, among 195 individuals with screen-detected diabetes, 45% were taking blood-pressure lowering medication and 20% were taking lipid-lowering medication at diagnosis.31 In ADDITION-UK at diagnosis, 55% of individuals were taking blood pressure-lowering medication, and 24% lipid-lowering medication, in agreement with the results of the Hoorn screening subsample. In a separate publication from the Hoorn study, 2 weeks after diagnosis 24% of the screen-detected and 78% of the clinically detected individuals were prescribed oral glucose-lowering medication.32 The step-wise screening program carried out in ADDITION-Cambridge used the Cambridge Risk Score to identify those at the highest risk of undiagnosed diabetes.22 This score includes blood pressure medication as a variable, which may have led to an overestimate in the number of individuals taking antihypertensive medication in this sample. In 2005–2006, in an American population with long-standing diabetes, 90% of the population were taking glucose-lowering medications, 78% were taking antihypertensives and 26% were on statins.33 This contrasts with ADDITION-UK, where glucose-lowering medications were less common (62%, at 5 years), and statins were more common (54%, at 5 years). Statin use was the pharmacotherapy that differed by the greatest margin between arms of the ADDITION-UK trial (47% for routine care vs 60% after the promotion of intensive care, at 5 years). Our results suggest that the promotion of statin use is the most readily accepted treatment after diagnosis compared to the introduction of glucose-lowering treatment. In ADDITION-Europe, we have previously demonstrated that individuals with the worst cardiometabolic health at diagnosis were the most likely to be prescribed glucose, blood pressure and lipid lowering medication, and also were likely to achieve the greatest reductions in individual CVD risk factors over the 5 years immediately after diagnosis.34

previously demonstrated that individuals with the worst cardiometabolic health at diagnosis were the most likely to be prescribed glucose, blood pressure and lipid lowering medication, and also were likely to achieve the greatest reductions in individual CVD risk factors over the 5 years immediately after diagnosis.34 Previous literature has noted that the prescription of cardioprotective medication often lags behind glucose-lowering medication, suggesting a disproportionate emphasis on controlling glucose over CVD risk reduction.33 35 In both arms of ADDITION-UK, use of antihypertensive and lipid-lowering medication was reported by around four-fifths of the participants (77% and 81%, respectively), and glucose-lowering and aspirin use was reported for three-fifths of the population (62% and 54%, respectively). Our results suggest that the prescription of cardioprotective medication did not lag behind that of glucose-lowering. Conversely, 20% of individuals were on metformin at 1 year, and 57% at 5 years, despite metformin being recommended as a first line glucose-lowering medication, and immediate initiation being recommended by National Institute for Health and Care Excellence if overweight or non-responsive to lifestyle interventions.19 Variation in treatment could be a positive indicator of patient-centered care or a deficit between patient need and prescribed medication. More detailed knowledge on the circumstances around treatment choices in screen-detected populations would help inform whether the prescription of cardioprotective and glucose-lowering medication should be higher in this population, or that the proportions prescribed medications in this study represent adequate care in relation to GP and patient needs and priorities. An increase in diabetes medication from diagnosis to 5 years was associated with being female, younger, having a GP who was in the trial arm promoted to treat intensively and having a higher baseline risk of a CVD event. In the Hoorn study, 2 weeks after screen-detected diabetes diagnosis, 24% of the population were taking glucose-lowering medication.32 While previous literature suggests there is no association between the prescription of diabetes-related medication and gender.36 37

ely and having a higher baseline risk of a CVD event. In the Hoorn study, 2 weeks after screen-detected diabetes diagnosis, 24% of the population were taking glucose-lowering medication.32 While previous literature suggests there is no association between the prescription of diabetes-related medication and gender.36 37 Strengths and limitations ADDITION-UK is a large cohort (n=1026) with consistency in outcome measurement and little loss to follow-up in individuals prescription histories (4% at 5 years). ADDITION-UK (91% white ethnicity) was less diverse than the UKPDS (81% white ethnicity),38 which may limit generalisability. However, ADDITION-UK remains the only study able to characterize medication changes after screen-detected diabetes diagnosis while receiving contemporary diabetes care. This analysis uses prescribed medications, which is likely to be an over count of the redeemed and consumed prevalence. Some medications may also be available without a prescription. Accuracy of medication data was improved by encouraging participants to bring repeat prescriptions to the health assessment, the use of a health economics questionnaire25 and cross-referencing GP records to collect medication data. For the secondary analysis of change in medications, our analysis assumes that a change from zero to one medication is directly comparable to a change from four to five, or two to one. Medication was coded into 23 classes, but anti-infectives, antiparasitics and antineoplastic medications (as defined by the ATC) were not included as they were acute (eg, infections) or rare (eg, cancer). As this study collected snapshots of medication use at baseline, 1 and 5 years after diagnosis, we are not able to give accurate prevalences for acutely prescribed medications. The number of medical agents was chosen over the raw pill count as some medications can be taken as ‘combination’ pills, or can be split across multiple doses. This could unduly increase the impact of some medications that are taken multiple times a day on the final medication count. There is also likely to be less agreement between the doctor prescribed treatments and daily pill count, compared to reported types of medical agent, as pill count includes both agent and information on frequency and method of dose. Information on non-CVD-related comorbidities that may influence medication was not collected.

is also likely to be less agreement between the doctor prescribed treatments and daily pill count, compared to reported types of medical agent, as pill count includes both agent and information on frequency and method of dose. Information on non-CVD-related comorbidities that may influence medication was not collected. This analysis remains primarily descriptive, and does not directly assess the relationship between changes in cardiometabolic health and pharmacotherapy. This analysis is unable to describe the pharmacotherapy of individuals that died during follow-up, and it is likely that if medication at the time of death was available, it would introduce greater heterogeneity to this analysis. There was no association between loss to follow-up and change in medication, although this analysis was limited to the subsample of Cambridge participants (86% of the sample) due to the IMD scores not being available for all centers.

ime of death was available, it would introduce greater heterogeneity to this analysis. There was no association between loss to follow-up and change in medication, although this analysis was limited to the subsample of Cambridge participants (86% of the sample) due to the IMD scores not being available for all centers. Individuals with screen-detected diabetes are often taking multiple medications before diagnosis, despite being identified early in the diabetes disease trajectory. This includes both cardioprotective medications, and other medications including; gastrointestinal, anti-inflammatories, analgesics and psychiatric/neurological medications. After diagnosis, GPs and patients appear to adopt pharmacological strategies that target both CVD risk reduction and glycemia, providing evidence against concerns of over-prioritizing glycemic targets. The increased prescription of cardioprotective medication was associated with higher baseline CVD risk, indicating an association between need and care. While this result is promising, it remains unclear if the prescription rates of glycemic and cardioprotective medication in this population with elevated cardiovascular risk reflect individualized treatment based on patient led priorities or a deficit in the application of pharmacological intervention.

n need and care. While this result is promising, it remains unclear if the prescription rates of glycemic and cardioprotective medication in this population with elevated cardiovascular risk reflect individualized treatment based on patient led priorities or a deficit in the application of pharmacological intervention. Contributors: RKS, MJD, DW, KK and SJG were involved in designing and implementing the ADDITION trial. All authors viewed and commented on the analysis plan and draft manuscripts of this post hoc analysis of the ADDITION trial. RKS, JAB and SJG reviewed the model results and wrote the first draft and JAB implemented the analysis strategy agreed on by all authors. All authors approve of the content of this manuscript. Funding: This study was supported by the Welcome Trust (grant number G061895), the Medical Research Council (Grant numbers G0001164 and MC_UU_12015/4) and the National Institute for Health Research (Grant number RP-PG-0606-1259).

Contributors: RKS, MJD, DW, KK and SJG were involved in designing and implementing the ADDITION trial. All authors viewed and commented on the analysis plan and draft manuscripts of this post hoc analysis of the ADDITION trial. RKS, JAB and SJG reviewed the model results and wrote the first draft and JAB implemented the analysis strategy agreed on by all authors. All authors approve of the content of this manuscript. Funding: This study was supported by the Welcome Trust (grant number G061895), the Medical Research Council (Grant numbers G0001164 and MC_UU_12015/4) and the National Institute for Health Research (Grant number RP-PG-0606-1259). Competing interests: MJD has acted as consultant, advisory board member and speaker for Novo Nordisk, Sanofi-Aventis, Lilly, Merck Sharp & Dohme, Boehringer Ingelheim, AstraZeneca and Janssen and as a speaker for Mitsubishi Tanabe Pharma Corporation. She has received grants in support of investigator and investigator-initiated trials from Novo Nordisk, Sanofi-Aventis and Lilly. SJG reports non-financial support from Bio-Rad, during the conduct of the study; personal fees from Eli Lilly, outside the submitted work. KK has received funds for research, honoraria for speaking at meetings and served on Advisory Boards for Astra Zeneca, Eli Lily, Novartis, Pfizer, Sanofi Aventis, MSD and Novo Nordisk, Janssen, BI. DW has received funds for research and honoraria for speaking at meetings for Sanofi Aventis, MSD and Novo Nordisk. GL began a job at Roche after the conclusion of this research project and her position as a Career Development Fellow at the MRC.

Lily, Novartis, Pfizer, Sanofi Aventis, MSD and Novo Nordisk, Janssen, BI. DW has received funds for research and honoraria for speaking at meetings for Sanofi Aventis, MSD and Novo Nordisk. GL began a job at Roche after the conclusion of this research project and her position as a Career Development Fellow at the MRC. Ethics approval: Ethical approval was obtained from the Cambridge (ref:01/063), Huntingdonshire (ref:00/609), Peterbor- ough and Fenland (ref:P01/95), West Essex (ref:1511-0103), North and Mid Essex (ref:MH395 MREC02/5/54), West Suffolk (ref:03/002), and Hertfordshire and Bed- fordshire (ref:EC03623) Local Research Ethics Commit- tees, and the Eastern Multi-Centre Research Ethics Committee (ref:02/5/54). Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: The MRC Epidemiology Unit welcomes proposals for projects and aim to make our data as widely available as possible while safeguarding the privacy of our participants, protecting confidential data and maintaining the reputations of our studies and participants. The Data Sharing web portal at http://epi-meta.medschl.cam.ac.uk has a list and detailed description of all variables that were collected in this study, as well as many other studies conducted by this research unit, and via this portal external researchers can submit an application to undertake an analysis using the data from this study.

All subjects (19) p Values Low-fat diet High-fat diet Postabsorptive Glucose 5.23±0.07 5.18±0.08 0.49 Insulin 71.78±4.90 71.22±5.53 0.77 NEFA 0.55±0.03 0.53±0.03 0.47 NPRQ* 0.90±0.01 0.87±0.01 0.12 Steady state Glucose 5.73±0.08 5.67±0.08 0.52 Insulin 1331.78±54.58 1362.86±73.12 0.41 NEFA 0.04±0.00 0.04±0.00 0.79 NPRQ* 1.01±0.01 0.99±0.01 0.15 Ins. clearance 455.73±20.29 453.15±24.69 0.67 GDR† 14.09±0.78 15.51±0.84 0.02 M/I 0.08±0.01 0.09±0.01 0.09 NEFA % supp. 92.90±0.67 92.66±0.73 0.72 ΔNPRQ* 0.11±0.01 0.12±0.01 0.48 Data are mean±SEM. Glucose in mmol/L; insulin in pmol/L; non-esterified fatty acids (NEFA) in mmol/L; non-protein respiratory quotient (NPRQ); insulin clearance (Ins. Clear. in mL m2/BSA/min; glucose disposal rate (GDR) in mg kg/FFM/min; insulin sensitivity (M/I)=GDR/steady-state insulin; % of non-esterified fatty acids suppression (NEFA %); metabolic flexibility (ΔNPRQ). *Data shown for 15 participants (7 African–American and 8 non-Hispanic white women) who completed substrate utilization measures for the LF and HF diets (repeated measures). †p=0.02 for the effect of diet on GDR (higher after the HF diet); other effects of diet were not significant (p range 0.09–0.79). BSA, body surface area; GDR, glucose disposal rate; HF, high-fat diet; LF, low-fat diet; NPRQ, non-protein respiratory quotient; NEFA, non-esterified fatty acids. Figure 1 (A–B) Insulin sensitivity after 7 days of a eucaloric low-fat diet (30%) or high-fat (50%) diet in non-Hispanic white (panel a) and African–American (panel b) women.

BSA, body surface area; GDR, glucose disposal rate; HF, high-fat diet; LF, low-fat diet; NPRQ, non-protein respiratory quotient; NEFA, non-esterified fatty acids. Figure 1 (A–B) Insulin sensitivity after 7 days of a eucaloric low-fat diet (30%) or high-fat (50%) diet in non-Hispanic white (panel a) and African–American (panel b) women. Figure 2 (A–B) Differences in fat oxidation during the hyperinsulinemic euglycemic clamp between African–American and non-Hispanic white women after 7 days of a eucaloric low-fat diet (30%) (A) OR after 7 days of a eucaloric high-fat diet (50%) (B). Open circles, solid lines=African–American women. Open squares, dotted lines=non-Hispanic white women. A general linear model was used to determine whether there were any race differences in response to insulin during the euglycemic hyperinsulinemic clamps, for the LF and HF diets separately (figure 2A and B). Both differences by race and any interactions by race in the effect of insulin during the clamps were computed for the LF and HF diets separately. The difference by race in steady-state insulin levels was also determined after adjusting for the postabsorptive insulin level (as a covariate). No other covariates were included in the analyses. A p value less than 0.05 was considered statistically significant. Statistical analysis was performed using Statistica (V.10.0, Tulsa, Oklahoma, USA).

Key messages There is controversy over whether a eucaloric, moderately high-fat (50%) diet vs a lower fat (30%) diet induces insulin resistance in overweight and obese women; substituting fat for carbohydrates to a moderate degree (50% vs 30%) in a weight-maintaining diet is not deleterious for peripheral insulin action in healthy overweight and obese women, at least in the short term (1 week). Similarly, metabolic flexibility (the ability to suppress fat oxidation by insulin during a hyperinsulinemic clamp) is not affected by a higher (50%) vs a lower fat (30%) eucaloric diet in healthy overweight and obese women, at least in the short term (1 week). African–American women are more insulin resistant and have lower rates of postabsorptive fat oxidation than similar white women, as we have previously reported, but we did not find that a moderately higher fat diet (50%) compared to a lower fat diet (30%) adversely affects their peripheral insulin action or ability to suppress fat oxidation during a high-dose insulin clamp.

ve lower rates of postabsorptive fat oxidation than similar white women, as we have previously reported, but we did not find that a moderately higher fat diet (50%) compared to a lower fat diet (30%) adversely affects their peripheral insulin action or ability to suppress fat oxidation during a high-dose insulin clamp. Introduction The role of the macronutrient composition of the diet with regard to the carbohydrate-to-fat ratio in the treatment of obesity and diabetes prevention has been only partially elucidated. While a low-fat (LF) diet was the mainstay for the diabetes prevention program1 and is the basis for the 2010 Dietary Guidelines for Americans,2 hypocaloric diets of both high-fat (HF) and LF compositions have been effective for weight loss.3 Epidemiologically, higher total fat intake was associated with higher rates of progression to type 2 diabetes in the San Luis Valley Diabetes study4; however, two other large population-based studies in women (Iowa Women's and Nurses’ Health studies) did not replicate these findings.5 6

been effective for weight loss.3 Epidemiologically, higher total fat intake was associated with higher rates of progression to type 2 diabetes in the San Luis Valley Diabetes study4; however, two other large population-based studies in women (Iowa Women's and Nurses’ Health studies) did not replicate these findings.5 6 Whether increasing the fat-to-carbohydrate ratio of a eucaloric, weight-maintaining diet decreases insulin sensitivity is controversial, particularly in women.7–12 One study in women has shown a decrease in insulin sensitivity, measured by a frequently sampled intravenous glucose tolerance test (FSIVGTT), after 3 weeks of an HF diet compared to a LF diet in healthy premenopausal African–American and non-Hispanic (NH) white participants.13 However, other work has demonstrated that peripheral insulin sensitivity, measured by the euglycemic hyperinsulinemic clamp, does not decrease after eucaloric HF diets of various durations (6 days and up to 3 weeks) in lean or obese men7–10 or combined groups of lean men and women.11 12 Metabolic flexibility (the ability to suppress fat oxidation during the euglycemic hyperinsulinemic clamp) has been closely associated with insulin sensitivity14 15 and decreased in response to a HF diet in men,8 9 yet this has never been studied in women.

or obese men7–10 or combined groups of lean men and women.11 12 Metabolic flexibility (the ability to suppress fat oxidation during the euglycemic hyperinsulinemic clamp) has been closely associated with insulin sensitivity14 15 and decreased in response to a HF diet in men,8 9 yet this has never been studied in women. Therefore, we aimed to determine whether insulin sensitivity measured during a euglycemic hyperinsulinemic clamp will be deleteriously affected by a 1 week, eucaloric HF (50% total Kcal from fat) diet in African–American and non-Hispanic white, healthy, premenopausal, overweight and obese women. In addition, we determined the effect of the diets on metabolic flexibility during the clamps. We and others have previously reported lower peripheral insulin sensitivity16–20 differences in muscle adipose tissue distribution19 and lower systemic rates of fat oxidation in African–American vs non-Hispanic white women.15 21 22 Therefore, we also examined any race differences in substrate utilization during the clamps.

nd others have previously reported lower peripheral insulin sensitivity16–20 differences in muscle adipose tissue distribution19 and lower systemic rates of fat oxidation in African–American vs non-Hispanic white women.15 21 22 Therefore, we also examined any race differences in substrate utilization during the clamps. Research design and methods Subjects Twenty-three healthy premenopausal (25–45 years) overweight and obese (body mass index, BMI 25–40 kg/m2) women (11 African–American and 12 non-Hispanic white) participated in the study. Participants were included if they reported all four grandparents to be of African or Caucasian descent, had regular menstrual cycles, and were without diabetes according to an oral glucose tolerance test (75 g glucose load). Self-reported use of any medications (including contraceptive pills), smoking within the past 6 months, and consumption of >2 oz. ethanol/day were exclusionary. All participants signed consent forms approved by the St. Luke's-Roosevelt Institute for Health Sciences Institutional Review Board.

test (75 g glucose load). Self-reported use of any medications (including contraceptive pills), smoking within the past 6 months, and consumption of >2 oz. ethanol/day were exclusionary. All participants signed consent forms approved by the St. Luke's-Roosevelt Institute for Health Sciences Institutional Review Board. Study design In a randomized crossover design, participants consumed a LF (30% fat, 50% carbohydrate and 20% protein) or a HF (50% fat, 30% carbohydrate and 20% protein) weight-maintaining diet for seven consecutive days as per the protocol we had previously published.15 On the morning of day 8 after an overnight admission to the Clinical Research Center at St. Luke's-Roosevelt Hospital Center, insulin sensitivity and substrate utilization were measured before and during a euglycemic hyperinsulinemic clamp. There was a minimum 2-week washout period between diets. All measurements were conducted during the follicular phase of the menstrual cycle.

linical Research Center at St. Luke's-Roosevelt Hospital Center, insulin sensitivity and substrate utilization were measured before and during a euglycemic hyperinsulinemic clamp. There was a minimum 2-week washout period between diets. All measurements were conducted during the follicular phase of the menstrual cycle. Dietary protocol All study participants completed dietary surveys indicating foods they liked and disliked. Eucaloric, weight-maintaining diets were constructed from food items available commercially with known macronutrient and caloric composition. Food item caloric content and macronutrient composition were verified using Nutritionist IV (V.2.0, Nsquared Commuting Co, Salem,Oregon, USA). Total daily calories for weight maintenance were calculated based on resting metabolic rate measured by indirect calorimetry in a fasting state (Horizon metabolic Cart or V-Max29; Sensor Medics, Yorba Linda, California, USA) and multiplied by an activity factor (1.5). Diets were matched in distribution of fat calories with equal parts of saturated fat, monounsaturated fat and polyunsaturated fat. Participants were provided with a 7-day food supply to consume at home. Dietary compliance was assured through weight stability measurements and adjustments were planned for a weight change of more than 1 kg.

in distribution of fat calories with equal parts of saturated fat, monounsaturated fat and polyunsaturated fat. Participants were provided with a 7-day food supply to consume at home. Dietary compliance was assured through weight stability measurements and adjustments were planned for a weight change of more than 1 kg. Insulin sensitivity Following an overnight fast, a three-hour euglycemic hyperinsulinemic clamp (80 mU/m2/min) was performed. We used a high-dose insulin clamp to measure the effect of the diet on peripheral insulin sensitivity in African–American vs non-Hispanic white women as we sought differences between races as well. We, as others, have previously reported lower peripheral insulin sensitivity16–20 in African–Americans vs non-Hispanic whites. Blood samples were collected at 10 min intervals during the postabsorptive and steady state of hyperinsulinemic euglycemic clamp, immediately centrifuged, aliquoted and frozen at −70°C. Insulin was measured by RIA (Linco Research, St. Charles, Missouri, USA), glucose was measured by the Beckman glucose analyzer (Beckman, Fullerton, California, USA) and non-esterified fatty acids (NEFA) were measured by the enzymatic colorimetric method (Wako Chemicals USA, Richmond, Virginia, USA). NEFA suppression was calculated as the difference between the NEFA levels at steady state and the postabsorptive NEFA levels divided by the postabsorptive NEFA levels times 100 (percentage). Insulin clearance was calculated according to DeFronzo23 as the ratio of the difference in insulin concentration between the post-absorptive and steady states and the rate of insulin infusion during the clamp study, which was assumed to be 80 mU/m2/min for all participants. Insulin sensitivity was calculated as M/I using the glucose disposal rate M (mg kg/fat-free mass (FFM)/min) and insulin concentration in the hyperinsulinemic steady state I (µU/mL).

tive and steady states and the rate of insulin infusion during the clamp study, which was assumed to be 80 mU/m2/min for all participants. Insulin sensitivity was calculated as M/I using the glucose disposal rate M (mg kg/fat-free mass (FFM)/min) and insulin concentration in the hyperinsulinemic steady state I (µU/mL). Indirect calorimetry Oxygen consumption (VO2) and carbon dioxide production (VCO2) were measured using a ventilated hood in the postabsorptive and hyperinsulinemic steady states of the euglycemic clamp. In both states, the participants were supine and awake. Substrate oxidation rates were calculated using Frayn's equations,24 and non-protein respiratory quotient (NPRQ) was calculated as a ratio of VCO2 to VO2. Metabolic flexibility was estimated as a change in NPRQ (ΔNPRQ) between the postabsorptive and hyperinsulinemic steady states. Statistics All data are reported as mean±SEM as noted. All variables were checked for normality of distribution; only fasting triglycerides were log transformed for analyses (log10). Statistical comparison of participant characteristics by race was performed using the independent t test (table 1). Selection and confounding biases were controlled for by symmetrical case-crossover methodology with identical length of exposure to the LF and HF diets. While the participants were unaware of the diet composition, there was no allocation concealment from the investigators. Table 1 Participants’ characteristics

Statistics All data are reported as mean±SEM as noted. All variables were checked for normality of distribution; only fasting triglycerides were log transformed for analyses (log10). Statistical comparison of participant characteristics by race was performed using the independent t test (table 1). Selection and confounding biases were controlled for by symmetrical case-crossover methodology with identical length of exposure to the LF and HF diets. While the participants were unaware of the diet composition, there was no allocation concealment from the investigators. Table 1 Participants’ characteristics Characteristics All participants (23) African–American (11) Non-Hispanic white (12) Age (years) 33.61±1.18 32.27±1.80 34.83±1.54 BMI (kg/m2) 29.65±0.90 28.80±1.12 30.42±1.39 Percentage of fat* 42.21±1.72 40.03±2.57 44.22±2.24 FM (kg)* 33.83±2.14 31.32±2.98 36.13±3.02 FFM (kg)* 44.85±0.91 45.39±1.02 44.35±1.51 Fasting triglycerides (mmol/L) 0.85±0.11 0.82±0.16 0.88±0.16 Fasting HDL (mmol/L) 0.90±0.04 0.90±0.07 0.91±0.07 Data are mean±SEM. There were no significant differences in subject characteristics between African–American and non-Hispanic white women (p range 0.23–0.91). *Determined by dual X-ray absorptiometry (DXA). BMI, body mass index; HDL, high-density lipoprotein; FM, fat mass; FFM, fat-free mass.

Characteristics All participants (23) African–American (11) Non-Hispanic white (12) Age (years) 33.61±1.18 32.27±1.80 34.83±1.54 BMI (kg/m2) 29.65±0.90 28.80±1.12 30.42±1.39 Percentage of fat* 42.21±1.72 40.03±2.57 44.22±2.24 FM (kg)* 33.83±2.14 31.32±2.98 36.13±3.02 FFM (kg)* 44.85±0.91 45.39±1.02 44.35±1.51 Fasting triglycerides (mmol/L) 0.85±0.11 0.82±0.16 0.88±0.16 Fasting HDL (mmol/L) 0.90±0.04 0.90±0.07 0.91±0.07 Data are mean±SEM. There were no significant differences in subject characteristics between African–American and non-Hispanic white women (p range 0.23–0.91). *Determined by dual X-ray absorptiometry (DXA). BMI, body mass index; HDL, high-density lipoprotein; FM, fat mass; FFM, fat-free mass. Analysis of variance and multivariate analysis of variance (ANOVA/MANOVA) were used to determine the effects of diet (LF vs HF, repeated measures, within effect) and to compute diet by race interactions (African–Americans vs non-Hispanic whites, between effect) from measures in the postabsorptive state and during the steady state of the clamp (glucose, NEFA and insulin levels, substrate utilization, NEFA suppression, insulin clearance, insulin sensitivity and metabolic flexibility), as shown in table 2 and figures 1A, B and 2A, B. Only data from women who completed either one of the dietary interventions (LF or HF diet) were used for analysis. A power analysis was performed for the effect of diet on peripheral insulin sensitivity, using initial pilot data (first seven participants of the study), which yielded a large effect size, Cohen's d=0.81 (M/I change between diets mean±SD, 0.014262±0.017618) from which the required sample size for 2 tailed α=0.05, power=0.80 was calculated to be n=15.

r the effect of diet on peripheral insulin sensitivity, using initial pilot data (first seven participants of the study), which yielded a large effect size, Cohen's d=0.81 (M/I change between diets mean±SD, 0.014262±0.017618) from which the required sample size for 2 tailed α=0.05, power=0.80 was calculated to be n=15. Table 2 Effect of diet on insulin sensitivity and substrate utilization (repeated measures) All subjects (19) p Values Low-fat diet High-fat diet Postabsorptive Glucose 5.23±0.07 5.18±0.08 0.49 Insulin 71.78±4.90 71.22±5.53 0.77 NEFA 0.55±0.03 0.53±0.03 0.47 NPRQ* 0.90±0.01 0.87±0.01 0.12 Steady state Glucose 5.73±0.08 5.67±0.08 0.52 Insulin 1331.78±54.58 1362.86±73.12 0.41 NEFA 0.04±0.00 0.04±0.00 0.79 NPRQ* 1.01±0.01 0.99±0.01 0.15 Ins. clearance 455.73±20.29 453.15±24.69 0.67 GDR† 14.09±0.78 15.51±0.84 0.02 M/I 0.08±0.01 0.09±0.01 0.09 NEFA % supp. 92.90±0.67 92.66±0.73 0.72 ΔNPRQ* 0.11±0.01 0.12±0.01 0.48 Data are mean±SEM. Glucose in mmol/L; insulin in pmol/L; non-esterified fatty acids (NEFA) in mmol/L; non-protein respiratory quotient (NPRQ); insulin clearance (Ins. Clear. in mL m2/BSA/min; glucose disposal rate (GDR) in mg kg/FFM/min; insulin sensitivity (M/I)=GDR/steady-state insulin; % of non-esterified fatty acids suppression (NEFA %); metabolic flexibility (ΔNPRQ). *Data shown for 15 participants (7 African–American and 8 non-Hispanic white women) who completed substrate utilization measures for the LF and HF diets (repeated measures).

A general linear model was used to determine whether there were any race differences in response to insulin during the euglycemic hyperinsulinemic clamps, for the LF and HF diets separately (figure 2A and B). Both differences by race and any interactions by race in the effect of insulin during the clamps were computed for the LF and HF diets separately. The difference by race in steady-state insulin levels was also determined after adjusting for the postabsorptive insulin level (as a covariate). No other covariates were included in the analyses. A p value less than 0.05 was considered statistically significant. Statistical analysis was performed using Statistica (V.10.0, Tulsa, Oklahoma, USA). Results Participant characteristics are shown in table 1. Twenty-three premenopausal (age 33.61±1.18 years) overweight (BMI 29.65±0.90 kg/m2) women participated in the study. Eight of 11 African–American and 11 out of 12 white women completed insulin sensitivity measurements after both the LF and HF diet periods (repeated measures). Additionally, one African–American woman completed the studies only after the LF diet condition and three women (2 African–Americans and 1 white) completed the studies only after the HF diet condition. For personal reasons, they did not participate in the second dietary period. There were no statistically significant differences in age, BMI, body composition measurements, fasting triglycerides and high-density lipoprotein (HDL)-cholesterol levels between the two races (table 1) or in the subgroups which had repeated measures (not shown).

they did not participate in the second dietary period. There were no statistically significant differences in age, BMI, body composition measurements, fasting triglycerides and high-density lipoprotein (HDL)-cholesterol levels between the two races (table 1) or in the subgroups which had repeated measures (not shown). For the 19 participants who had repeated measures, the effect of diet on insulin sensitivity and metabolic flexibility (ΔNPRQ during the clamp) are shown in table 2. There were no significant diet by race interactions on any of the variables (p range 0.31 to 1.0); thus, the main effects of diet are presented here. Insulin sensitivity computed as the glucose disposal rate per kg of FFM and divided by the steady-state insulin level (M/I) was not significantly decreased by the diet in the African–American (0.06±0.01 vs 0.07±0.01, for LF vs HF diet, respectively, p=0.40) or in the white women (0.09±0.01 vs 0.10±0.01, for LF vs HF diet, respectively, p=0.09). In most women, insulin sensitivity either remained unchanged or was higher after the HF compared to the LF diet (figure 1A and B). Similarly, metabolic flexibility (ΔNPRQ during the clamp) was not significantly altered by the diet type (table 2).

women (0.09±0.01 vs 0.10±0.01, for LF vs HF diet, respectively, p=0.09). In most women, insulin sensitivity either remained unchanged or was higher after the HF compared to the LF diet (figure 1A and B). Similarly, metabolic flexibility (ΔNPRQ during the clamp) was not significantly altered by the diet type (table 2). Using data from all participants, the steady-state insulin levels during the clamp were higher in the African–American vs non-Hispanic white women, after adjustment for the postabsorptive values, after the LF diet (1449.17±40.34 pmol/L vs 1247.48±80.68 pmol/L, p=0.02) or after the HF diet (1490.98±59.59 pmol/L vs 1286.55±103.75 pmol/L, p=0.05). Thus, the calculated insulin clearance was lower in African–American vs white women, after the LF diet (407.55±12.26 mL/m2/min vs 489.60±30.19 mL/m2/min, p=0.03) or after the HF diet ((397.02±14.27 mL/m2/min vs 486.86±34.65 mL/m2/min, p=0.04). There were no other significant differences by race, after the LF diet (p range 0.91–1.0) or after the HF diet (p range 0.14–0.9). Fat oxidation was significantly suppressed by insulin during the euglycemic clamp, for both African–American and white women, after both the LF (p<0.001) and HF (p<0.001) diets (figure 2A), with no significant insulin by race interaction (figure 2B) on either diet (p=0.27 and p=0.28, respectively). ΔNPRQ, that is, metabolic flexibility during the clamp, was not significantly different in African–American vs white women after the LF diet (0.10±0.02 vs 0.12±0.02, respectively, p=0.59) or after the HF diet (0.12±0.02 vs 0.13±0.02, p=0.58).

ace interaction (figure 2B) on either diet (p=0.27 and p=0.28, respectively). ΔNPRQ, that is, metabolic flexibility during the clamp, was not significantly different in African–American vs white women after the LF diet (0.10±0.02 vs 0.12±0.02, respectively, p=0.59) or after the HF diet (0.12±0.02 vs 0.13±0.02, p=0.58). Conclusions Our study did not show a decrease in peripheral insulin sensitivity in response to a short-term (1 week) eucaloric 50% HF diet compared to a 30% LF diet in healthy, overweight, and obese premenopausal African–American and non-Hispanic white women. Metabolic flexibility (ΔNPRQ) was similarly unaffected. The only significant race difference we found was the lower insulin clearance in African–American vs white women, regardless of the diet.

diet compared to a 30% LF diet in healthy, overweight, and obese premenopausal African–American and non-Hispanic white women. Metabolic flexibility (ΔNPRQ) was similarly unaffected. The only significant race difference we found was the lower insulin clearance in African–American vs white women, regardless of the diet. Our results highlight the controversy surrounding the effect of a eucaloric increase in the fat content of a weight-maintaining diet on insulin sensitivity and metabolic flexibility, a precursor of insulin sensitivity. One other study, utilizing FSIVGTT to measure insulin sensitivity in premenopausal obese women showed a deterioration of insulin sensitivity after 3 weeks of a eucaloric HF diet vs a eucaloric LF diet,13 whereas other studies, in agreement with our results, have used a euglycemic hyperinsulinemic clamp to assess insulin sensitivity, which is the ‘gold standard’ for this outcome. A eucaloric HF diet consumed over a period of ∼3 weeks did not alter insulin sensitivity in mixed groups of lean men and women,11 12 and similar results were demonstrated in lean men after just 6 days,7 and in lean and overweight men after 3 weeks,8 10 of a eucaloric HF diet. Thus, diet duration does not seem to account for the discrepancy between our results and other work in a similar population.13 Hepatic insulin sensitivity remained unchanged in two studies with a similar HF diet as used by us,7 10 but was shown to decrease in lean men after 11 days of an 83% HF diet.9 FSIVGTT does not differentiate between hepatic and peripheral insulin×sensitivity. Different effects of a HF diet on hepatic vs peripheral insulin sensitivity may to some extent account for the difference in results noted by us.13 Other factors playing a role may be the account of the menstrual cycle stage when insulin sensitivity was measured,13 25 and the differences in the amounts of saturated fat employed.13

fects of a HF diet on hepatic vs peripheral insulin sensitivity may to some extent account for the difference in results noted by us.13 Other factors playing a role may be the account of the menstrual cycle stage when insulin sensitivity was measured,13 25 and the differences in the amounts of saturated fat employed.13 We also found that the metabolic flexibility, measured as a suppression of fat oxidation during the hyperinsulinemic (80 mU/m2/min) euglycemic clamp (ΔNPRQ),14 was not affected by the 1 week of a eucaloric 50% HF diet in our women. The effect of a eucaloric HF diet on ΔNPRQ has been studied in men, yet the results are inconclusive. In lean men, ΔNPRQ was not decreased in response to an HF (75%) diet compared to a similar LF (35%) diet, after 6 days,7 or 3 weeks,10 but was decreased after 11 days of a HF (83%) diet.9 In overweight men, ΔNPRQ decreased after 3 weeks of an HF (55%) diet.8 No similar studies are available in women. A certain threshold in the fat/carbohydrate ratio of the diet and the effect on hepatic insulin sensitivity26 may modulate the degree of fat oxidation suppression by insulin after a eucaloric HF diet. Hepatic insulin sensitivity and its relationship to metabolic flexibility was not evaluated in our study and needs to be investigated further. Some of the findings in the present study, specifically a lack of differential effect by race, may be due to a lack of power secondary to a small sample size. Furthermore, 1 week of a eucaloric 50% HF diet may have different effects in other populations, with different genetic susceptibility.27 28

investigated further. Some of the findings in the present study, specifically a lack of differential effect by race, may be due to a lack of power secondary to a small sample size. Furthermore, 1 week of a eucaloric 50% HF diet may have different effects in other populations, with different genetic susceptibility.27 28 Finally, we previously reported lower rates of postabsorptive fat oxidation in response to an HF diet and lack of fat oxidation suppression by insulin during a pancreatic clamp in African–American vs white women.15 In this study, we observed similar trends for the postabsorptive fat oxidation values, but the higher dose of insulin during the clamp similarly suppressed fat oxidation in the two races, in agreement with a recent report.29 We also found lower insulin clearance in the African–American women compared to the white women, in contrast to one,30 but in agreement with another study in adult women.31 The lower insulin clearance could be contributing to unmeasured postprandial hyperinsulinemia, which may partly explain the numerous reports of lower fat oxidation rates in African–Americans without diabetes compared to other white populations.21 22 32 33

to one,30 but in agreement with another study in adult women.31 The lower insulin clearance could be contributing to unmeasured postprandial hyperinsulinemia, which may partly explain the numerous reports of lower fat oxidation rates in African–Americans without diabetes compared to other white populations.21 22 32 33 In conclusion, peripheral insulin sensitivity was not deleteriously affected by 1 week of a eucaloric HF diet (50% of total Kcal from fat), compared to a LF (30% of total Kcal from fat) diet, in healthy, premenopausal, overweight and obese African–American and non-Hispanic white women. Our findings need to be verified with regard to the effect on hepatic insulin response and more importantly in other susceptible populations. The authors would like to acknowledge Allan Geliebter, PhD, New York Obesity Research Center, for assistance and review of the statistical analyses. Contributors: NMB analyzed the data, and wrote and prepared the manuscript for publication. ME analyzed the data and reviewed the manuscript. RWW collected the data, and reviewed and critiqued the manuscript. ESB designed the study, collected and analyzed the data, and reviewed and critiqued the manuscript. JBA designed the study, collected and analyzed the data, and wrote and prepared the manuscript for publication. She is also the guarantor of this work. Funding: This work was supported by the following grants: NIH R21DK71171, New York Obesity Research Center Grant P30DK26687, CTSA M01RR00645, DERC P30DK63608 and American Diabetes Association Grant 1-10-CT-01.

Contributors: NMB analyzed the data, and wrote and prepared the manuscript for publication. ME analyzed the data and reviewed the manuscript. RWW collected the data, and reviewed and critiqued the manuscript. ESB designed the study, collected and analyzed the data, and reviewed and critiqued the manuscript. JBA designed the study, collected and analyzed the data, and wrote and prepared the manuscript for publication. She is also the guarantor of this work. Funding: This work was supported by the following grants: NIH R21DK71171, New York Obesity Research Center Grant P30DK26687, CTSA M01RR00645, DERC P30DK63608 and American Diabetes Association Grant 1-10-CT-01. Competing interests: JBA is an Associate Editor for Open BMJ DRC. She is also a reviewer of grants, abstracts and papers for the American Diabetes Association and its journals. JBA reports research grants funding from Weight Watchers, Eli Lilly, Roche, Takeda, Merck and Novo Nordisk, outside the submitted work. ESB is a current employee of GlaxoSmithKline. No other potential duality or conflicts of interest were reported relevant to this article. Ethics approval: St. Luke's-Roosevelt Institute for Health Sciences Institutional Review Board. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: Methodology details information from this study is available through consultation with the corresponding author.

Key messages A novel barrier to physical activity that we have identified previously is that exercise feels more difficult to sedentary premenopausal women with type 2 diabetes mellitus (T2DM) than their similarly obese and sedentary counterparts without diabetes. We found statistically significant differences in plasma lactate during low- to moderate-intensity exercise in postmenopausal women with T2DM, as compared to their counterparts without diabetes; we also found clinically meaningful differences in exercise effort as measured subjectively by the Borg Rating of Perceived Exertion (RPE). An important take-home point for clinicians is to encourage patients to be physically active at a pace that is personally comfortable, as this tends to be associated with both good adherence and a good physiological fitness response. The prevalence of type 2 diabetes mellitus (T2DM) continues to rise worldwide and the highest prevalence rates are found among older adults.1 Exercise is considered a critical cornerstone of treatment for people with T2DM due to its beneficial effects on glycemic control, physical fitness, cardiovascular health and prevention of disability as well as premature mortality.2–4 However, people with T2DM are consistently more sedentary than similarly obese people who do not have diabetes, for reasons that are not clear.5 6 Since lower fitness levels are linked to cardiovascular morbidity and mortality in populations with and without T2DM,7 8 understanding and overcoming barriers to physical activity for people with T2DM is critically important.

ntary than similarly obese people who do not have diabetes, for reasons that are not clear.5 6 Since lower fitness levels are linked to cardiovascular morbidity and mortality in populations with and without T2DM,7 8 understanding and overcoming barriers to physical activity for people with T2DM is critically important. A novel barrier to physical activity that we have identified previously is that exercise feels more difficult to sedentary people with T2DM than their similarly obese and sedentary counterparts without diabetes.9 10 Specifically, we have shown that effort during low intensity exercise is greater in premenopausal women with T2DM versus similarly obese controls, as measured by both the Borg Rating of Perceived Exertion (RPE) and plasma lactate concentrations during exercise. Since RPE is modifiable11–13 and plays a significant role in adherence to prescribed physical activity,14–16 it has great potential relevance as a modifiable barrier to physical activity. In addition, these findings may partly explain some barriers to physical activity identified in prior questionnaires and focus groups of people with T2DM, such as ‘difficulty keeping up with others who don't have T2DM.’17 Finally, RPE is related to the affective response to exercise,18 19 and the 2010 American Diabetes Association physical activity guidelines suggest that “affective responses to exercise may be important predictors of physical activity adoption and maintenance.”3

difficulty keeping up with others who don't have T2DM.’17 Finally, RPE is related to the affective response to exercise,18 19 and the 2010 American Diabetes Association physical activity guidelines suggest that “affective responses to exercise may be important predictors of physical activity adoption and maintenance.”3 Important possible mediators of increased exercise effort include impaired cardiorespiratory fitness levels and submaximal exercise responses. Submaximal exercise impairments include a slowed VO2 uptake kinetics response that represents a delay in achieving steady-state oxygen utilization during constant work rate exercise. Among adolescents and middle-aged adults, participants with T2DM have significantly worse peak cardiorespiratory fitness levels and worse submaximal exercise performance than their counterparts without diabetes.10 20–22 Less is known about these measures of exercise performance in older adults with T2DM23 even though older adults have the highest prevalence of diabetes.1 Therefore, we sought to compare measures of exercise effort and exercise performance during exercise in older women with T2DM versus their counterparts without diabetes. We hypothesized that there would be greater perceived effort during low to moderate intensity exercise in participants with T2DM compared to controls without diabetes. We also hypothesized that both fitness levels and submaximal exercise responses would be more impaired in participants with T2DM than controls, despite recruiting participants of similar weight and similar levels of habitual physical activity. We studied women with T2DM because T2DM confers greater exercise impairment in women than men.24

that both fitness levels and submaximal exercise responses would be more impaired in participants with T2DM than controls, despite recruiting participants of similar weight and similar levels of habitual physical activity. We studied women with T2DM because T2DM confers greater exercise impairment in women than men.24 Research design and methods In this cross-sectional study, we enrolled overweight, sedentary women with or without T2DM from the metropolitan Denver, Colorado, USA community between 2007 and 2011. Participants were 50–75 years old with either T2DM (n=26) or controls without diabetes (n=28). We defined people without diabetes by the prevailing American Diabetes Association guidelines at the time of the research (ie, normal fasting glucose <100 mg/dL and glycated hemoglibin, HbA1c <6%).25 In addition, to minimize finding insulin resistance in control participants, we excluded controls with >1 first-degree relative with T2DM.25 Additional inclusion criteria ensured similar overweight and sedentary status in study groups: overweight/obese body mass index (BMI) (25–39.9 kg/m2) and self-report of leisure physical activity behavior of <60 min/week. Participants were postmenopausal as documented by no menses in >12 months and by measured follicle-stimulating hormone levels.10 Our exclusion criteria included conditions that imposed safety concerns or could impair exercise performance, such as uncontrolled hypertension, a history of atherosclerosis, congestive heart failure, autonomic or peripheral neuropathy, chronic disease of the lung, liver or kidneys, microalbuminuria, or tobacco use within 1 year.26 In addition, we excluded participants with prolonged duration of T2DM or suboptimal disease control: HbA1c >8.5%, duration of diabetes >20 years, microalbuminuria (urine microalbumin/creatinine >30).27 Use of insulin, thiazolidinediones, GLP-1 agonists and DPP-4 inhibitors was excluded since these drugs might either affect exercise capacity or suggest more advanced disease.22 28

of T2DM or suboptimal disease control: HbA1c >8.5%, duration of diabetes >20 years, microalbuminuria (urine microalbumin/creatinine >30).27 Use of insulin, thiazolidinediones, GLP-1 agonists and DPP-4 inhibitors was excluded since these drugs might either affect exercise capacity or suggest more advanced disease.22 28 Exercise effort Measures of exercise effort included RPE (primary outcome) and plasma lactate concentration (secondary outcome) measured during eight bouts of constant work rate submaximal exercise. The Borg RPE scale is the gold standard measure of RPE during exercise in healthy29 30 and diabetic populations.31 The range of scores on this ordinal scale is from 6 to 20 with verbal anchors every two points (eg, RPE=11, ‘light’, RPE=13, ‘somewhat hard’). To minimize bias, participants and the research staff recording RPE were blinded to the work rate. Plasma lactate concentration was measured using the lactate dehydrogenase method on blood drawn in perchloric acid tubes.10 RPE and lactate concentration were measured on two separate study dates at both an absolute work rate (30 W, 4 bouts) and a relative work rate (4 bouts). We designed the relative work rate to be 35% of the VO2peak from each participant's peak exercise test, in order to account for the influence of cardiorespiratory fitness on exercise effort. We chose 35% VO2peak as the relative work rate to ensure participants achieved a steady-state VO2 level that was reliably below yet close to the lactate threshold. In our prior published work in premenopausal women, 35% VO2peak was ∼30 W for participants with T2DM.9

ence of cardiorespiratory fitness on exercise effort. We chose 35% VO2peak as the relative work rate to ensure participants achieved a steady-state VO2 level that was reliably below yet close to the lactate threshold. In our prior published work in premenopausal women, 35% VO2peak was ∼30 W for participants with T2DM.9 Nutritional and body composition assessments prior to exercise testing To control for effects of diet on exercise performance, participants consumed a eucaloric study diet with standardized macronutrient distributions for 3 days prior to exercise testing and fasted for at least 4 h prior to exercise testing.10 Eucaloric diets were developed by registered dieticians based on body composition assessed by Dual-energy X-ray Absorptiometry (DXA scan, Hologic/Discovery W, Hologic Inc, Bedford, Massachusetts, USA). DXA was also used to assess the exercise effort-related covariate of total fat-free mass.

prior to exercise testing.10 Eucaloric diets were developed by registered dieticians based on body composition assessed by Dual-energy X-ray Absorptiometry (DXA scan, Hologic/Discovery W, Hologic Inc, Bedford, Massachusetts, USA). DXA was also used to assess the exercise effort-related covariate of total fat-free mass. Exercise testing The exercise testing procedures have been described in detail previously.10 32 In brief, each exercise test began with the participant seated upright at rest for 3 min on the cycle ergometer (Lode ergometer, MedGraphics, Minneapolis, Minnesota, USA) breathing into a mouthpiece connected to a metabolic cart (Ultima CPX, MedGraphics). Participants first performed a familiarization graded cycle ergometer exercise test (GXT). On a subsequent day, we used a ramping protocol GXT to measure VO2peak. During the ramping protocol GXT, the rate of increase of the work rate to peak exercise capacity was individualized to ensure an optimal test duration of 10–14 min (eg, if the familiarization GXT peak work rate was 55 W, the wattage increased continuously by 0.0833 W/s to reach a work rate of 55 W at 11 min). During exercise, VO2 was measured breath-by-breath and time-averaged over 30 s intervals. We defined VO2peak by standard convention as the peak VO2 associated with an RER ≥1.1 or as a VO2 plateau despite an increase in workload.33

the wattage increased continuously by 0.0833 W/s to reach a work rate of 55 W at 11 min). During exercise, VO2 was measured breath-by-breath and time-averaged over 30 s intervals. We defined VO2peak by standard convention as the peak VO2 associated with an RER ≥1.1 or as a VO2 plateau despite an increase in workload.33 Exercise performance measures and other predictors of exercise effort Our secondary outcomes of exercise performance included cardiorespiratory fitness (VO2peak) and a measure of submaximal exercise response (time constant (τ2) from VO2 uptake kinetics) at both 30 W and 35% VO2peak. Kinetics measurements during constant-load exercise On subsequent study dates, participants performed eight bouts of submaximal constant-load exercise with work rates alternating between 30 W and 35% VO2peak. Each exercise bout was 8 min in duration. The research assistants recording the data and participants were blinded to the work rate. During constant-load exercise, we assessed submaximal exercise response using VO2 kinetics, where a longer time constant (τ2) represents a longer time to achieve steady-state VO2.10 As previously described, for each work rate, gas-exchange data for kinetic analyses were processed using a software program developed in our laboratory.10 32 The pulmonary VO2 kinetic responses data for each of the bouts of oxygen consumption were evaluated using a two-component exponential model.26

steady-state VO2.10 As previously described, for each work rate, gas-exchange data for kinetic analyses were processed using a software program developed in our laboratory.10 32 The pulmonary VO2 kinetic responses data for each of the bouts of oxygen consumption were evaluated using a two-component exponential model.26 Physiological predictors of exercise effort According to our conceptual model (figure 1), we considered the following physiological variables as potential predictors of exercise effort: heart rate during exercise, VO2peak and τ2. In addition, we assessed metabolic and vascular factors that we hypothesized were likely to impair the physiological response to exercise in people with T2DM: Homeostasis Model Assessment of Insulin Resistance (HOMA-IR),34 markers of endothelial dysfunction such as a diagnosis of hypertension and arterial stiffness expressed as pulse-wave velocity35 (SphygmaCor CP system, AtCor Medical). HOMA-IR was calculated per standard convention from blood glucose and insulin measurements collected during a 12 h fast.34 We also measured glucose levels during the 8 min of exercise as an additional potential predictor of effort. A diagnosis of hypertension was obtained by self-report and validated during the medical history and physical examination by the study physician/PI (AGH). Pulse-wave velocity was assessed after a 4 h fast and prior to any exercise testing performed in the same study visit. Figure 1 Conceptual model of perceived effort during exercise.

Physiological predictors of exercise effort According to our conceptual model (figure 1), we considered the following physiological variables as potential predictors of exercise effort: heart rate during exercise, VO2peak and τ2. In addition, we assessed metabolic and vascular factors that we hypothesized were likely to impair the physiological response to exercise in people with T2DM: Homeostasis Model Assessment of Insulin Resistance (HOMA-IR),34 markers of endothelial dysfunction such as a diagnosis of hypertension and arterial stiffness expressed as pulse-wave velocity35 (SphygmaCor CP system, AtCor Medical). HOMA-IR was calculated per standard convention from blood glucose and insulin measurements collected during a 12 h fast.34 We also measured glucose levels during the 8 min of exercise as an additional potential predictor of effort. A diagnosis of hypertension was obtained by self-report and validated during the medical history and physical examination by the study physician/PI (AGH). Pulse-wave velocity was assessed after a 4 h fast and prior to any exercise testing performed in the same study visit. Figure 1 Conceptual model of perceived effort during exercise. Behavioral predictors of exercise effort According to our conceptual model where behavioral factors may influence the interpretation of physiological cues (figure 1), we also considered the following behavioral variables as potential predictors of RPE: self-efficacy to perform physical activity36–38 and depressive symptoms.39 40 We assessed these behavioral variables by paper surveys conducted prior to any exercise testing performed in the same study visit.

ological cues (figure 1), we also considered the following behavioral variables as potential predictors of RPE: self-efficacy to perform physical activity36–38 and depressive symptoms.39 40 We assessed these behavioral variables by paper surveys conducted prior to any exercise testing performed in the same study visit. Power calculation Based on our preliminary data where we observed an RPE difference of 1.3±1.4 (mean±SD) between the T2DM and overweight control groups,9 we estimated that 54 participants would provide 82% power to detect a between-group difference in RPE of 0.8 in the 8 min of exercise at the 0.05 level. We considered a between-group RPE difference of one point to be clinically meaningful.41

fference of 1.3±1.4 (mean±SD) between the T2DM and overweight control groups,9 we estimated that 54 participants would provide 82% power to detect a between-group difference in RPE of 0.8 in the 8 min of exercise at the 0.05 level. We considered a between-group RPE difference of one point to be clinically meaningful.41 Statistical analysis All outcome and predictor variables were examined with descriptive statistics and graphic summaries, overall and by T2DM status. To determine group differences in RPE and lactate, 30 W data collected during four separate bouts of exercise and 35%VO2peak data collected during four separate bouts of exercise were evaluated using separate maximum likelihood repeated measures models (eg, 30 W model assessed 4 repeated measures).42 The primary outcomes were RPE and lactate at minute 8. We also assessed changes in RPE within and between-groups across minutes 2, 4, 6 and 8 at 30 W and 35% VO2peak to determine if RPE levels were at steady-state, defined as equivalent levels at minutes 6 and 8, and to compare the RPE change from minutes 2 to 8 by study group. We also assessed lactate concentrations at rest prior to exercise and the change in lactate concentrations from the resting baseline to the 8 min of exercise (Δlactate), by using separate maximum likelihood repeated measures models. Finally, we compared group differences in heart rate, also using a maximum likelihood repeated measures model at each of 2, 4, 6 and 8 min into the exercise bouts. To ensure that our data at the relative work rate of 35% VO2peak was accounting for fitness levels appropriately, we also conducted a sensitivity analysis to estimate group differences in effort variables (ie, RPE, lactate during minute 8, heart rate during minute 8 of exercise) with adjustment for fitness at the 30 W absolute work rate for comparison with the group differences from the relative work rate of 35% VO2peak that accounted for fitness. We also conducted a sensitivity analysis to estimate group differences in resting lactate concentration with adjustment for fitness.

8 of exercise) with adjustment for fitness at the 30 W absolute work rate for comparison with the group differences from the relative work rate of 35% VO2peak that accounted for fitness. We also conducted a sensitivity analysis to estimate group differences in resting lactate concentration with adjustment for fitness. Continuous variables that were only measured once were compared with a two-sided two-sample t test, and dichotomous variables were compared with either a χ2 test for equal proportions or Fisher's exact test. Secondary analyses were conducted to evaluate potential covariates according to our conceptual model (figure 1). We first estimated Pearson's product-moment correlation coefficients between candidate predictor variables to rule out excessive collinearity (r>0.8). We then estimated Pearson's correlations between the effort variables, RPE and lactate and the potential covariates. Predictor variables with a correlation coefficient ≥0.2 were added to the maximum likelihood models described above; we retained predictor variables in the final model when p<0.05. For groupings of predictor variables with relatively high collinearity (eg, metabolic predictors of a diagnosis of T2DM (dichotomous variable), HOMA and HbA1c), we selected the metabolic predictor that created the best fitting model based on the BIC statistic. All analyses were conducted in SAS V.9.2.43

Secondary analyses were conducted to evaluate potential covariates according to our conceptual model (figure 1). We first estimated Pearson's product-moment correlation coefficients between candidate predictor variables to rule out excessive collinearity (r>0.8). We then estimated Pearson's correlations between the effort variables, RPE and lactate and the potential covariates. Predictor variables with a correlation coefficient ≥0.2 were added to the maximum likelihood models described above; we retained predictor variables in the final model when p<0.05. For groupings of predictor variables with relatively high collinearity (eg, metabolic predictors of a diagnosis of T2DM (dichotomous variable), HOMA and HbA1c), we selected the metabolic predictor that created the best fitting model based on the BIC statistic. All analyses were conducted in SAS V.9.2.43 Results Study population We consented 97 women for participation, 35 women with T2DM and 62 control women without diabetes. After screening laboratory testing and procedures, we excluded 43 women who met exclusion criteria, including 18 women with prediabetes. Women in the T2DM (n=26) and control groups without diabetes (n=28) were similar in age, BMI and baseline physical activity levels, as per study design (table 1). Hypertension prevalence was ∼50% in each study group. As compared with the control group, the T2DM group included more non-white women (p=0.01) and had higher mean HbA1c levels (p<0.001). Table 1 Participant characteristics

Results Study population We consented 97 women for participation, 35 women with T2DM and 62 control women without diabetes. After screening laboratory testing and procedures, we excluded 43 women who met exclusion criteria, including 18 women with prediabetes. Women in the T2DM (n=26) and control groups without diabetes (n=28) were similar in age, BMI and baseline physical activity levels, as per study design (table 1). Hypertension prevalence was ∼50% in each study group. As compared with the control group, the T2DM group included more non-white women (p=0.01) and had higher mean HbA1c levels (p<0.001). Table 1 Participant characteristics Variable Overweight control (n=28) Type 2 diabetes (T2DM) (n=26) p Value Duration of T2DM, years (SD) NA 5.1 (5.0) NA Age, years (SD) 59.8 (5.8) 59.3 (5.7) 0.39 Ethnicity, n (% Hispanic) 1 (4) 3 (12) 0.34 Race, n (% Caucasian) 27 (96) 18 (67) 0.01 Hypothesized physiological RPE predictors BMI, kg/m2 (SD) 30.1 (2.8) 31.3 (3.9) 0.25 HbA1c, % 5.7 (0.3) 6.8 (0.6) <0.001 Homeostasis Model Assessment of Insulin Resistance (HOMA-IR) 4.0 (1.9) 6.7 (4.6) 0.01 HTN prevalence (%) 46 50 0.99 Pulse wave velocity (m/s) 9.2 (0.7) 9.9 (0.8) 0.06 Reported usual physical activity, MET h/week (SD) 222 (35) 231 (25) 0.32 VO2peak (mL/min) 1470 (286) 1313 (244) 0.02 VO2peak (mL/kg/min) 17.8 (3.0) 15.4 (2.5) 0.003 Work rate during 35% VO2peak (SD) 43.5 (7.7) 35.9 (7.1) <0.0001 Heart rate during 8 min of exercise at 30 W (bpm) 91 (2) 102 (2) 0.002 Heart rate during 8 min of exercise at 35% VO2peak 98 (3) 105 (3) 0.04 Resting lactate concentrations (mmol/L) 0.39 (0.03) 0.65 (0.03) <0.0001 Lactate during 8 min of exercise at 30 W (mmol/L) 0.76 (0.10) 1.18 (0.10) 0.004 Lactate at 8 min at 35% VO2peak (mmol/L) 0.99 (0.10) 1.28 (0.10) 0.047 τ2 at 30 W 39.2 (13) 36.2 (9.1) 0.34 τ2 at 35% VO2peak 38.6 (12.2) 38.5 (10.2) 0.98 Hypothesized behavioral RPE predictors Depressive symptoms CES-D 11.4 (3.7) 11.7 (5.6) 0.77 PHQ-9 3.3 (3.3) 3.2 (3.2) 0.96 Self-efficacy scores Endurance self-efficacy 746 (359) 739 (264) 0.94 Lorig self-efficacy 481 (95) 479 (94) 0.92 Sallis self-efficacy 45 (8) 47 (8) 0.23 Data expressed as mean (SE) for heart rate, lactate; data expressed as mean (SD) for all other continuous variables.

.7) 11.7 (5.6) 0.77 PHQ-9 3.3 (3.3) 3.2 (3.2) 0.96 Self-efficacy scores Endurance self-efficacy 746 (359) 739 (264) 0.94 Lorig self-efficacy 481 (95) 479 (94) 0.92 Sallis self-efficacy 45 (8) 47 (8) 0.23 Data expressed as mean (SE) for heart rate, lactate; data expressed as mean (SD) for all other continuous variables. Missing data for <10% of participants with the exception of pulse wave velocity where we are missing 40% of data for each study group. BMI, body mass index; CES-D, Center for Epidemiological Studies Depression; HbA1c, glycated hemoglobin; HTN, hypertension; PHQ-9, Patient Health Questionnaire-9; RER, respiratory exchange ratio; T2DM, type 2 diabetes mellitus. Markers of effort—RPE and lactate At both absolute work rates and relative work rates that accounted for fitness differences among participants, RPE was greater, but not significantly so, in the T2DM versus control group at minutes 2, 4, 6 and 8 of exercise (figure 2A, B). Of note, despite all participants reaching steady-state oxygen uptake levels within the first 2 min of exercise, the RPE levels continued to rise during 8 min of exercise and had not yet reached steady-state at 8 min of exercise (the RPE slope was different from zero (p<0.001) from minutes 6 to 8 in each study group). The change in RPE from minute 2 to 8 was greater, but not significantly so, in the T2DM group compared to the control group (p=0.07).

nued to rise during 8 min of exercise and had not yet reached steady-state at 8 min of exercise (the RPE slope was different from zero (p<0.001) from minutes 6 to 8 in each study group). The change in RPE from minute 2 to 8 was greater, but not significantly so, in the T2DM group compared to the control group (p=0.07). Figure 2 (A) Group differences in rating of perceived exertion over time at 30 W. (B) Group differences in rating of perceived exertion over time at 35% VO2peak. (C) Higher lactate concentrations in T2DM during rest and exercise. Lactate concentrations are shown in figure 2C. To summarize, mean lactate levels during minute 8 of exercise were significantly greater in the T2DM group as compared with control women (p=0.004 at 30 W and p=0.046 at 35% VO2peak, figure 2C). Mean lactate levels were also significantly greater at rest in the T2DM group compared with control women (p<0.0001). The difference between resting and exercise lactate concentrations (Δlactate) was not significantly greater in the T2DM group compared with control women (30 W Δlactate: 0.53±0.08 vs 0.37±0.08 (p=0.17); 35% VO2peak Δlactate: 0.62±0.08 vs 0.60±0.08, p=0.86). Our sensitivity analysis showed that group differences in resting lactate remained significantly different after adjustment for VO2peak (p<0.001), and we observed no statistical association between VO2peak and resting lactate concentration (p=0.98).

±0.08 (p=0.17); 35% VO2peak Δlactate: 0.62±0.08 vs 0.60±0.08, p=0.86). Our sensitivity analysis showed that group differences in resting lactate remained significantly different after adjustment for VO2peak (p<0.001), and we observed no statistical association between VO2peak and resting lactate concentration (p=0.98). Physiological effort-related variables VO2peak was significantly lower in the T2DM group versus study group without diabetes by absolute and weight-adjusted measures (table 1). However, τ2 was not different between study groups at 30 W or 35% VO2peak (table 1). Of the physiological variables that were prespecified as candidate predictors of RPE (figure 1), those that were significantly different between study groups included heart rate during exercise, lactate levels during exercise, VO2peak, HbA1c and HOMA-IR (table 1). In a sensitivity analysis adjusting for fitness levels as VO2peak (mL/min) rather than accounting for fitness differences at the relative work rate of 35% VO2peak, mean lactate levels and heart rate during exercise at 30 W remained significantly greater in the T2DM group as compared with control women (p=0.02 for lactate; p=0.01 for heart rate).

djusting for fitness levels as VO2peak (mL/min) rather than accounting for fitness differences at the relative work rate of 35% VO2peak, mean lactate levels and heart rate during exercise at 30 W remained significantly greater in the T2DM group as compared with control women (p=0.02 for lactate; p=0.01 for heart rate). Association of physiological variables with effort In the regression model assessing predictors of RPE in the 8th min of exercise with adjustment for insulin resistance by HOMA-IR, the physiological variables significantly associated with RPE were heart rate during exercise, diagnosis of hypertension and plasma lactate during exercise (table 2). Although HOMA-IR was not a statistically significant predictor of RPE, it was included in the model as a precision efficiency variable (ie, the model fit was enhanced by including HOMA-IR, and HOMA-IR outperformed the other metabolic candidate variables of T2DM diagnosis, glucose level during exercise and HbA1c). In the regression model assessing predictors of lactate level at minute 8, the only significant predictor at both 30 W and 35% VO2peak was heart rate during exercise. At the 30 W work rate only, τ2 was also significantly associated with lactate. Table 2 Regression models for predictors of RPE and lactate at 30 W and at 35% VO2peak

Association of physiological variables with effort In the regression model assessing predictors of RPE in the 8th min of exercise with adjustment for insulin resistance by HOMA-IR, the physiological variables significantly associated with RPE were heart rate during exercise, diagnosis of hypertension and plasma lactate during exercise (table 2). Although HOMA-IR was not a statistically significant predictor of RPE, it was included in the model as a precision efficiency variable (ie, the model fit was enhanced by including HOMA-IR, and HOMA-IR outperformed the other metabolic candidate variables of T2DM diagnosis, glucose level during exercise and HbA1c). In the regression model assessing predictors of lactate level at minute 8, the only significant predictor at both 30 W and 35% VO2peak was heart rate during exercise. At the 30 W work rate only, τ2 was also significantly associated with lactate. Table 2 Regression models for predictors of RPE and lactate at 30 W and at 35% VO2peak Effort predictor variables 30 W F statistic (p value) 35% VO2peak F statistic (p value) Physiological RPE predictor model Heart rate during 8 min exercise 9.8 (p=0.002) 4.1 (p=0.045) Hypertension diagnosis (yes/no) 6.4 (p=0.02) 10.3 (p=0.003) Lactate during 8 min exercise 6.0 (p=0.02) 4.2 (p=0.04) HOMA-IR 2.3 (p=0.14) 0.8 (p=0.38) Behavioral RPE predictor model CES-D 4.5 (p=0.04) 3.2 (p=0.08) HOMA-IR 2.5 (p=0.11) 1.5 (p=0.23) Physiological lactate predictor model Heart rate during 8 min exercise 5.6 (p=0.02) 8.9 (p=0.0002) T2DM diagnosis (yes/no) 6.2 (p=0.02) 1.6 (p=0.26) Separate linear regression models were conducted for physiological predictors of RPE, behavioral predictors of RPE and physiological predictors of lactate during exercise at 30 W and at 35% VO2peak.

ctor model Heart rate during 8 min exercise 5.6 (p=0.02) 8.9 (p=0.0002) T2DM diagnosis (yes/no) 6.2 (p=0.02) 1.6 (p=0.26) Separate linear regression models were conducted for physiological predictors of RPE, behavioral predictors of RPE and physiological predictors of lactate during exercise at 30 W and at 35% VO2peak. CES-D, Center for Epidemiological Studies Depression; HOMA-IR, Homeostasis Model Assessment of Insulin Resistance; RPE, Rating of Perceived Exertion; T2DM, type 2 diabetes mellitus. Behavioral effort-related variables We found no significant difference in depressive symptoms or self-efficacy to perform physical activity by study group (table 1). In a multivariate regression model assessing predictors of RPE with adjustment for HOMA-IR, we found that depressive symptoms as measured by the Center for Epidemiological Studies—Depression (CESD)39 scale were associated with RPE at 30 W but not at 35% VO2peak. Depressive symptoms as measured by the PHQ-940 scale were not predictive of RPE at either work rate.

g predictors of RPE with adjustment for HOMA-IR, we found that depressive symptoms as measured by the Center for Epidemiological Studies—Depression (CESD)39 scale were associated with RPE at 30 W but not at 35% VO2peak. Depressive symptoms as measured by the PHQ-940 scale were not predictive of RPE at either work rate. Discussion We found that objective measures of exercise effort during submaximal exercise of lactate and heart rate during exercise were significantly greater in older women with T2DM compared to controls without diabetes. In addition, we identified higher levels of perceived effort in women with T2DM than in controls without diabetes, but these group differences in RPE were not statistically significant. Although RPE did not differ significantly between study groups, the RPE group differences observed at both 30 W and 35% VO2peak were ≥1 point—others have reported a 1-point difference to be clinically meaningful on the RPE scale.44 In addition, we found both heart rate and lactate during exercise to be significant predictors of RPE in the multivariate analysis, supporting the existing research that these three measures are all related markers of effort.30 45 Taken together, these measures suggest that women with T2DM experience exercise, even at relatively low levels, to be more strenuous than their counterparts without diabetes. Exercise effort is an important barrier to physical activity because it is modifiable11–13 and the perception of more intense effort during exercise has been associated with lower levels of usual physical activity.14–16

rcise, even at relatively low levels, to be more strenuous than their counterparts without diabetes. Exercise effort is an important barrier to physical activity because it is modifiable11–13 and the perception of more intense effort during exercise has been associated with lower levels of usual physical activity.14–16 As has been previously shown in adolescents and adults with T2DM, we found that mean levels of peak cardiorespiratory fitness were significantly lower in women with T2DM than in controls.10 20 22 23 Similar to the finding of Wilkerson et al23 in older men, but different from our findings in younger adults,10 22 the present study found no differences in mean τ2 levels in the T2DM group as compared to controls (table 1). Of note, the values of τ2 in our control participants were longer than those observed in prior studies of adolescent and premenopausal control participants,22 24 32 suggesting that the kinetic differences observed in younger T2DM may make them physiologically ‘older’ compared to their similarly-aged peers. Conversely, the lack of difference in subjects with T2DM versus controls in older populations (eg, Wilkerson et al23 and the present study) suggests that age-related changes in this response parameter may mask the independent effects of T2DM.

may make them physiologically ‘older’ compared to their similarly-aged peers. Conversely, the lack of difference in subjects with T2DM versus controls in older populations (eg, Wilkerson et al23 and the present study) suggests that age-related changes in this response parameter may mask the independent effects of T2DM. Few studies have compared RPE in people with T2DM to RPE in people without diabetes. The existing literature suggests that RPE differences by T2DM status are more likely to be apparent at absolute work rates than at relative work rates that account for the lower fitness levels exhibited in people with T2DM than controls.9 46 In our prior study of premenopausal women, we assessed RPE at two absolute work rates of 20 W and 30 W during cycle ergometer exercise and found significantly greater RPE in women with T2DM when compared to both overweight controls and normal weight controls with no T2DM.9 Coquart et al46 studied overweight women with T2DM and overweight controls at a relative work rate of 100% ventilatory threshold (mean work rate in watts: T2DM: 41±14; overweight control: 43±11) and found no significant group differences in RPE (T2DM=13.7±2.3, overweight control=13.2±1.6). Consistent with the concept that RPE differences are more attenuated at relative work rates than absolute work rates, in the present study the RPE differences between T2DM women and controls were smaller at the relative work rate of 35% VO2peak than at 30 W, although the group differences were not statistically significant at either 30 W (p=0.08) or at 35% VO2peak (p=0.20).

e attenuated at relative work rates than absolute work rates, in the present study the RPE differences between T2DM women and controls were smaller at the relative work rate of 35% VO2peak than at 30 W, although the group differences were not statistically significant at either 30 W (p=0.08) or at 35% VO2peak (p=0.20). After adjustment for insulin resistance in our population of overweight women with and without T2DM, we found that the objective markers of effort of plasma lactate and heart rate during exercise were very strong predictors of RPE, as has been observed in other healthy populations.47 A somewhat surprising finding was that RPE was more strongly associated with a diagnosis of hypertension than with heart rate and plasma lactate during exercise. This finding warrants confirmation in future studies to ensure that it is not spurious. Others have reported an association between glucose levels and RPE during high-intensity exercise of prolonged duration because the ‘depletion of carbohydrate fuel sources triggers muscular fatigue’.46 47 We did not find an association between glucose levels during exercise and RPE, thus supporting the existing literature that energy substrates such as glucose do not appear to influence RPE at lower exercise intensities.47

nged duration because the ‘depletion of carbohydrate fuel sources triggers muscular fatigue’.46 47 We did not find an association between glucose levels during exercise and RPE, thus supporting the existing literature that energy substrates such as glucose do not appear to influence RPE at lower exercise intensities.47 Our findings add to the literature suggesting that lactate levels during rest and exercise are greater in people with T2DM than in their counterparts without diabetes. Previous studies of differences in lactate levels during exercise in people with T2DM versus controls have been conducted in younger women10 and middle-aged men,48 but we are not aware of other studies in older women with T2DM versus no diabetes. Resting lactate concentrations were greater in people with T2DM as compared to controls without diabetes in prior studies as well as the present study—elevated lactate at rest has even been recognized as a predictor of incident T2DM.49 50 In prior studies by Regensteiner and Mogensen et al, the lactate concentrations during exercise were significantly greater in the T2DM group versus overweight control group, and the differences became larger at higher work rates. With moderate intensity exercise, lactate levels may modestly increase as we observed, and lactate dehydrogenase (LDH) activity increases and can offset/counter-balance hydrogen ion production in order to delay muscle fatigue.51 52 Our observations of higher absolute lactate levels at rest and during exercise may indicate an alteration of the LDH complex and function in T2DM—or perhaps a different position on the LDH complex operating curve. We observed wide interindividual variance in Δlactate concentrations and were not powered to detect significance in these data. The wide interindividual variance in our observed Δlactate concentrations may relate partly to the effect of different LDH isoforms and partly to variance in exercise intensity perturbation.

. We observed wide interindividual variance in Δlactate concentrations and were not powered to detect significance in these data. The wide interindividual variance in our observed Δlactate concentrations may relate partly to the effect of different LDH isoforms and partly to variance in exercise intensity perturbation. In other studies assessing metabolic predictors of exercise effort, absolute lactate concentrations during exercise have been used as RPE predictors rather than Δlactate concentrations.30 45 A scientific rationale for using the absolute concentrations of lactate rather than Δlactate is that lactate is dynamically produced and metabolized at rest and during exercise and levels of metabolism are proportionally higher at greater lactate concentrations.52–54 Thus, the absolute lactate concentration provides a meaningful comparison of the absolute difference between lactate production and metabolism at any given time. The 2010 American Diabetes Association physical activity guidelines suggest that “affective responses to exercise may be important predictors of physical activity adoption and maintenance”.3 Perceptions of disproportionately greater effort during exercise may be one reason that people with T2DM are less active than their counterparts without T2DM, as perceptions of exercise effort over a certain threshold worsen the affective response to exercise and are linked to lower levels of leisure physical activity.14–16 Our effort measures were obtained at work rates that are less intense than many activities of daily living. Since people tend to prefer physical activities with an intensity in the 11–14 range,55 56 our findings provide some additional support to existing concerns that overweight, sedentary individuals with and without T2DM may avoid activities of daily living because they are perceptually too difficult and hence unpleasant. For example, our study population of overweight, sedentary women that reported mean RPE levels in the 10–12 range during cycle ergometry at 30 W (approximately 3 METs of intensity) would be expected to experience a much higher RPE while walking up a flight of stairs that represents approximately 10 METs of intensity.57 Thus, activities of daily living that require greater intensity activity, such as walking up stairs or walking at a faster pace to keep up with others, may be avoided for individuals such as those we studied.

nce a much higher RPE while walking up a flight of stairs that represents approximately 10 METs of intensity.57 Thus, activities of daily living that require greater intensity activity, such as walking up stairs or walking at a faster pace to keep up with others, may be avoided for individuals such as those we studied. In keeping with this theory, a prior barrier reported by people with T2DM is the inability to keep pace with others who do not have diabetes.17 A recent review article suggested that practitioners balance goals for their patients to conduct regular exercise in a range of intensity that will allow for physiological improvement and the need to make physical activity ‘palatable’ in an intensity range that is acceptable to the individual.58 An important take-home point for clinicians is to encourage patients to be physically active at a pace that is personally comfortable as this tends to be associated with both good adherence and a good physiological fitness response.58 59

activity ‘palatable’ in an intensity range that is acceptable to the individual.58 An important take-home point for clinicians is to encourage patients to be physically active at a pace that is personally comfortable as this tends to be associated with both good adherence and a good physiological fitness response.58 59 The strengths of this study include the assessment of effort during exercise by both subjective (RPE) and objective methods (plasma lactate and heart rate), as well as assessments of both physiological and behavioral predictors of exercise effort. The cross-sectional nature of the study prevented us from drawing any causal inferences for the statistical associations that we observed between exercise effort and the predictors of effort. Although the control subjects did not meet a diagnosis of prediabetes by the prevailing criteria of the time, a subset would have met the current criteria for prediabetes by a HbA1c of 5.7–6.4%. The inclusion of control subjects with prediabetes may have diminished the differences in exercise effort observed between the study groups. Another limitation of these findings is that the duration of exercise bouts assessed was only 8 min; evaluating the effort response to longer bouts of duration would also be important to examine in future studies.

with prediabetes may have diminished the differences in exercise effort observed between the study groups. Another limitation of these findings is that the duration of exercise bouts assessed was only 8 min; evaluating the effort response to longer bouts of duration would also be important to examine in future studies. In summary, we found what others have reported to be clinically meaningful differences in RPE44 and statistically significant differences in lactate and heart rate during low-to moderate-intensity exercise in postmenopausal women with T2DM, as compared to their counterparts without diabetes. The group differences in RPE were not as large as we observed in our prior study of younger women,9 suggesting that future studies in older adults should use a more conservative estimate of effect size than is appropriate in younger women. It is possible that the effects of aging may influence T2DM-related exercise impairments. Greater perceived exertion is a modifiable and thereby targetable end point. Therefore, methods to reduce perception of work effort in T2DM should be sought in order to facilitate regular physical activity in people with T2DM. Improving physical activity would also improve the premature disability and mortality experienced by people with T2DM.

n is a modifiable and thereby targetable end point. Therefore, methods to reduce perception of work effort in T2DM should be sought in order to facilitate regular physical activity in people with T2DM. Improving physical activity would also improve the premature disability and mortality experienced by people with T2DM. This work was supported by funding from the University of Colorado Center for Women's Health Research, the University of Colorado Division of General Internal Medicine and the Eugene C and Florence Armstrong Family Foundation. During the period of this project, AGH was supported by NIH/NCATS Colorado CTSA Grant Number KL2 TR001080. Contents are the authors’ sole responsibility and do not necessarily represent official NIH views. The authors have no relevant conflicts of interest to disclose. The authors very much appreciate the participation of our volunteer study participants. Contributors: AGH and LH collected the data. AGH analyzed and interpreted the data and drafted the manuscript. PW supervised the statistical data analysis and revised the manuscript. JEBR, TAB, WMK, SD and JGR interpreted the data and revised the manuscript. All authors read and approved the final version of the manuscript. AGH is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Funding: University of Colorado School of Medicine; Eugene C and Florence Armstrong Family Foundation; National Institutes of Health. Competing interests: None declared.

Contributors: AGH and LH collected the data. AGH analyzed and interpreted the data and drafted the manuscript. PW supervised the statistical data analysis and revised the manuscript. JEBR, TAB, WMK, SD and JGR interpreted the data and revised the manuscript. All authors read and approved the final version of the manuscript. AGH is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Funding: University of Colorado School of Medicine; Eugene C and Florence Armstrong Family Foundation; National Institutes of Health. Competing interests: None declared. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: No additional data are available. Ethics statement: The study consent processes and procedures were approved by the University of Colorado Institutional Review Board.

Key messages Using a database comprising ‘real-world’ observations with a long-term follow-up period, the present study revealed that long-term visit-to-visit variability in glycated hemoglobin (HbA1c) and systolic blood pressure (SBP) represented a combined and additive risk for cardiovascular disease (CVD) incidence in patients with type 2 diabetes. It is suggested that a synergistic effect exists between HbA1c variability and mean SBP levels for the incidence of CVD. In addition, SBP variability can be a risk factor for CVD incidence, even if the mean SBP is maintained within the normal range. Our findings indicate the possibility that stabilization of variability in HbA1c and SBP as well as lowering of their mean levels can be an efficient strategy for preventing the incidence of CVD. Introduction Recent clinical evidence has raised the possibility that visit-to-visit glycated hemoglobin (HbA1c) variability1–4 and blood pressure (BP) variability5–7 independently predict macrovascular complications and/or all-cause mortality in patients with type 2 diabetes. However, to date, no study has examined the combined risk associated with visit-to-visit variability in HbA1c and systolic BP (SBP) simultaneously. In addition, the differences in the effects between HbA1c variability and SBP variability on the incidence of cardiovascular disease (CVD) according to the mean HbA1c and SBP values have been scarcely investigated.

e combined risk associated with visit-to-visit variability in HbA1c and systolic BP (SBP) simultaneously. In addition, the differences in the effects between HbA1c variability and SBP variability on the incidence of cardiovascular disease (CVD) according to the mean HbA1c and SBP values have been scarcely investigated. Basic research data have shown that glucose fluctuations can cause oxidative stress,8–10 chronic inflammation, and endothelial dysfunction, which are involved in the progression of atherosclerosis.11 12 Additionally, increased BP variability may reflect arterial stiffness and baroreceptor dysfunction, which have been associated with arteriosclerosis and can result in cardiovascular events.13–17 However, the precise mechanisms have not been fully elucidated. In this study, we evaluated the combined effect of visit-to-visit variability in HbA1c and SBP on the incidence of CVD using a database of ‘real-world’ observations with long-term follow-up in patients with type 2 diabetes. In addition, we analyzed the differences in the effects between HbA1c variability and SBP variability on the incidence of CVD according to the mean HbA1c and SBP values.

lity in HbA1c and SBP on the incidence of CVD using a database of ‘real-world’ observations with long-term follow-up in patients with type 2 diabetes. In addition, we analyzed the differences in the effects between HbA1c variability and SBP variability on the incidence of CVD according to the mean HbA1c and SBP values. Methods Study participants Of the 1912 patients who first visited the outpatient clinic of our hospital from January 1995 to December 1996, we retrospectively recruited 632 patients with type 2 diabetes who attended at least four clinic visits, with at least one clinic visit per year, and had been followed up for ≥1 year. Patients were excluded if they had impaired glucose tolerance or a history of CVD at the first visit or within 1 year thereafter. They were followed for the incidence of CVD until June 2012. Of the 632 patients, 293 (46.4%) completed the follow-up, and 26 (4.1%) died. In June 2012, a questionnaire was mailed to the remaining 313 patients (49.5%) who had transferred to other hospitals or dropped out. One hundred and thirty-six (21.5%) responses were obtained. Of these, 27 deaths were confirmed. However, there were no responses from 177 patients (28.0%) who were regarded as censored cases at the last visit. Finally, the overall follow-up rate was 72.0% (455/632).

ho had transferred to other hospitals or dropped out. One hundred and thirty-six (21.5%) responses were obtained. Of these, 27 deaths were confirmed. However, there were no responses from 177 patients (28.0%) who were regarded as censored cases at the last visit. Finally, the overall follow-up rate was 72.0% (455/632). The following baseline characteristics of the patients were analyzed: age, sex, diabetes duration, BP, body mass index (BMI), HbA1c level, serum lipid level, serum creatinine (SCr) level, estimated glomerular filtration rate (eGFR), smoking status, alcohol intake, diabetes therapy, use of antihypertensive agents (ACE inhibitor, calcium channel blocker, α-blocker, or β-blocker), and/or use of a lipid-lowering drug. The merely renin-angiotensin system inhibitors available in Japan in 1995–1996 were ACE inhibitors. Initial therapy was defined as treatment started before the first visit, at the first visit, or within 6 months thereafter. Patients who received a combination of insulin and an oral antidiabetic drug were considered as insulin-treated patients. The study design was consistent with the Japanese government's Ethical Guidelines Regarding Epidemiological Studies and was in accordance with the Declaration of Helsinki. The protocol of this study was reviewed and approved by our Institutional Review Board and informed consent was obtained from all enrolled patients.

The study design was consistent with the Japanese government's Ethical Guidelines Regarding Epidemiological Studies and was in accordance with the Declaration of Helsinki. The protocol of this study was reviewed and approved by our Institutional Review Board and informed consent was obtained from all enrolled patients. End point definition The end point was the first CVD event, defined as fatal or non-fatal acute myocardial infarction, coronary artery procedure (bypass surgery or angioplasty), or stroke (ischemic or hemorrhagic), that required hospitalization. These events were determined according to a thorough review of medical records and responses of the questionnaires. Patients who had no CVD event, including those who had died from all other causes except CVD, were considered censored cases at the last clinic visit.

hemic or hemorrhagic), that required hospitalization. These events were determined according to a thorough review of medical records and responses of the questionnaires. Patients who had no CVD event, including those who had died from all other causes except CVD, were considered censored cases at the last clinic visit. Data collection and variables determined Capillary blood was drawn at each visit to determine blood glucose and HbA1c levels, irrespective of fasting or postprandial status. HbA1c levels were determined using an automated glycohemoglobin analyzer (Tosoh Bioscience, Tokyo, Japan); beginning in November 1994, HbA1c levels were determined using high-performance liquid chromatography, as standardized by the Japan Diabetes Society (JDS). HbA1c values obtained before January 2007 were converted to JDS standard values (reference range 4.3–5.8%) using linear regression equations. The equations were derived from duplicate assays using old and/or new devices or standard substances. Beginning in June 2012, we used the National Glycohemoglobin Standardization Program (NGSP)-certified method, and all earlier HbA1c (%) values were converted to NGSP values (%) using the following equation: (HbA1c (NGSP) (%)=1.02×HbA1c (JDS) (%)+0.25 (%)).18 The intrapersonal mean, coefficient of variation (CV), and variation independent of mean (VIM) of all recorded HbA1c measurements were calculated for each patient, and CV and VIM were employed as a measure of visit-to-visit variability in HbA1c.

he following equation: (HbA1c (NGSP) (%)=1.02×HbA1c (JDS) (%)+0.25 (%)).18 The intrapersonal mean, coefficient of variation (CV), and variation independent of mean (VIM) of all recorded HbA1c measurements were calculated for each patient, and CV and VIM were employed as a measure of visit-to-visit variability in HbA1c. BP was typically determined once at each visit in the sitting position by a trained medical technologist using an electronic sphygmomanometer (OMRON, Kyoto, Japan). The intrapersonal mean, CV, and VIM of all recorded SBP measurements were calculated for each patient, and CV and VIM were employed as a measure of visit-to-visit variability in SBP. The recorded BP values were used, in spite of whether the patient initiated or added an antihypertensive agent during the follow-up.

an). The intrapersonal mean, CV, and VIM of all recorded SBP measurements were calculated for each patient, and CV and VIM were employed as a measure of visit-to-visit variability in SBP. The recorded BP values were used, in spite of whether the patient initiated or added an antihypertensive agent during the follow-up. Lipids were measured irrespective of fasting or postprandial status. The total cholesterol (TC) level was determined using an enzymatic method. The high-density lipoprotein cholesterol (HDL-C) level was determined using a dextran sulfate and magnesium precipitation method until April 25, 1996, after which HDL-C was determined using a direct enzymatic method. HDL-C data from the precipitation method were converted to the direct enzymatic method equivalents using a linear regression equation derived from duplicate assays. The baseline TC: HDL-C ratio (TC/HDL-C) was employed as a covariate in the analysis because TC/HDL-C has been shown to be the best predictor of CVD among males with type 2 diabetes.19 20 Moreover, TC/HDL-C was found to be a stronger predictor of CVD compared with non-HDL-C in the UK Prospective Diabetes Study (UKPDS) risk engine.21

TC: HDL-C ratio (TC/HDL-C) was employed as a covariate in the analysis because TC/HDL-C has been shown to be the best predictor of CVD among males with type 2 diabetes.19 20 Moreover, TC/HDL-C was found to be a stronger predictor of CVD compared with non-HDL-C in the UK Prospective Diabetes Study (UKPDS) risk engine.21 The SCr level was determined using the Jaffe-Rate method until June 11, 1995, after which SCr was determined using an enzymatic method. SCr data obtained using the Jaffe-Rate method were converted to enzymatic method equivalents using a linear regression equation derived from duplicate assays. eGFR was determined using the following equation, as advocated by the Japanese Society of Nephrology: eGFR (mL/min/1.73 m2) =194×SCr−1.094×age−0.287 (×0.739 if female).22 Statistical analysis Data are expressed as means±SD for continuous variables or as numbers and percentages for categorical variables. Since the data distributions of the follow-up period and the number of visits were skewed, they were described as median values (IQR). Differences between patients who did and did not develop a CVD event were analyzed using Student's t test for continuous variables and the χ2 test or Fisher's exact test, as needed, for categorical variables. Age, sex, and diabetes duration were adjusted using logistic regression analysis.

ribed as median values (IQR). Differences between patients who did and did not develop a CVD event were analyzed using Student's t test for continuous variables and the χ2 test or Fisher's exact test, as needed, for categorical variables. Age, sex, and diabetes duration were adjusted using logistic regression analysis. The Kaplan-Meier survival curves for a CVD event were given for the four groups classified by median HbA1cCV and SBPCV values, after adjusting for age, mean HbA1c, mean SBP, and the number of visits. Values for the number of visits were ln-transformed for inclusion in the model to adjust for the possibility that the number of visits could influence variability. Multivariate analyses were performed using Cox proportional hazard models to evaluate the respective and combined effects of visit-to-visit variability in HbA1c and SBP as continuous variables on the incidence of CVD. The analysis was performed after adjusting for mean HbA1c, mean SBP, the number of visits (ln-transformed), age, sex, diabetes duration, BMI, TC/HDL-C, eGFR, baseline smoking status, baseline alcohol intake, baseline use of insulin, and baseline use of an antihypertensive agent. HRs are reported in 1SD increments.

CVD. The analysis was performed after adjusting for mean HbA1c, mean SBP, the number of visits (ln-transformed), age, sex, diabetes duration, BMI, TC/HDL-C, eGFR, baseline smoking status, baseline alcohol intake, baseline use of insulin, and baseline use of an antihypertensive agent. HRs are reported in 1SD increments. The combined effects of the visit-to-visit variability in HbA1c and SBP on the incidence of CVD as categorical variables were analyzed. Patients were classified into four groups by median HbA1cCV and SBPCV values. HRs for the incidence of CVD associated with these four groups (with lower HbA1cCV and lower SBPCV serving as the reference group) were calculated using a Cox proportional hazard model after adjusting for the aforementioned covariates. Similarly, the analysis was also performed in respect of the four groups classified by median HbA1cVIM and SBPVIM values (with lower HbA1cVIM and lower SBPVIM serving as the reference group). Furthermore, a stratified analysis was performed according to the mean HbA1c and SBP levels of 7.0% and 130 mm Hg, respectively, and the effects of visit-to-visit variability in HbA1c and SBP on the incidence of CVD were evaluated as continuous variables using multivariate Cox proportional hazard models after adjusting for the aforementioned covariates. The interaction was examined. The SAS V.9.4 software package (SAS Institute, Cary, North Carolina, USA) was used for all statistical analyses. Two-tailed p values <0.05 were considered to indicate significance.

Furthermore, a stratified analysis was performed according to the mean HbA1c and SBP levels of 7.0% and 130 mm Hg, respectively, and the effects of visit-to-visit variability in HbA1c and SBP on the incidence of CVD were evaluated as continuous variables using multivariate Cox proportional hazard models after adjusting for the aforementioned covariates. The interaction was examined. The SAS V.9.4 software package (SAS Institute, Cary, North Carolina, USA) was used for all statistical analyses. Two-tailed p values <0.05 were considered to indicate significance. Results The baseline clinical characteristics of all patients classified according to the incidence of CVD events during follow-up are shown in table 1. During the follow-up period, 81 (12.8%) patients (65 males and 16 females) had suffered a CVD event. After adjusting for age, sex, and diabetes duration, patients who had suffered a CVD event were significantly older, had a significantly longer diabetes duration, a significantly higher TC level, and a significantly lower HDL-C level. In addition, patients with CVD were significantly more likely to be taking an antihypertensive agent (ACE inhibitor, calcium channel blocker, or β-blocker) and a lipid-lowering agent compared with those without CVD. Table 1 Baseline characteristics of all patients following classification according to the incidence of CVD during follow-up

Results The baseline clinical characteristics of all patients classified according to the incidence of CVD events during follow-up are shown in table 1. During the follow-up period, 81 (12.8%) patients (65 males and 16 females) had suffered a CVD event. After adjusting for age, sex, and diabetes duration, patients who had suffered a CVD event were significantly older, had a significantly longer diabetes duration, a significantly higher TC level, and a significantly lower HDL-C level. In addition, patients with CVD were significantly more likely to be taking an antihypertensive agent (ACE inhibitor, calcium channel blocker, or β-blocker) and a lipid-lowering agent compared with those without CVD. Table 1 Baseline characteristics of all patients following classification according to the incidence of CVD during follow-up All CVD event p Value Adjusted p value* No event Event n 632 551 81 Male (%) 519 (82.1) 454 (82.4) 65 (80.3) 0.638 0.997 Age (years) 55.7±9.3 55.2±9.3 58.8±8.7 0.001 0.030 Duration of diabetes (years) 5.7±6.7 5.3±6.4 8.5±8.2 0.001 0.003 BMI (kg/m2) 23.3±3.3 23.3±3.3 23.6±3.1 0.506 0.169 HbA1c (%) 8.0±1.7 8.0±1.7 8.2±1.6 0.235 0.287 (mmol/mol) 64.2±18.7 63.9±18.9 66.5±17.9 0.235 0.287 SBP (mm Hg) 133.4±21.1 132.6±20.6 138.8±23.3 0.014 0.071 DBP (mm Hg) 77.7±12.6 77.5±12.4 79.6±13.8 0.160 0.128 TC (mg/dL) 209.5±37.6 208.0±36.6 219.7±42.6 0.008 0.011 HDL-C (mg/dL) 49.9±12.8 50.3±13.0 46.9±11.5 0.026 0.002 eGFR (mL/min/1.73 m2) 79.8±18.6 80.1±18.5 77.8±19.5 0.306 0.858 Current smoker 267 (42.3) 235 (42.7) 32 (39.5) 0.593 0.889 Alcohol intake 477 (75.5) 422 (76.6) 55 (67.9) 0.090 0.232 Initial therapies Oral antidiabetic drugs† 263 (41.6) 222 (40.3) 41 (50.6) 0.078 0.240 Insulin‡ 83 (13.1) 72 (13.1) 11 (13.6) 0.898 0.598 Antihypertensive agents 138 (21.8) 109 (19.8) 29 (35.8) 0.001 0.023 ACE inhibitors 54 (8.6) 39 (7.1) 15 (18.5) 0.0006 0.004 Calcium channel blockers 102 (16.2) 77 (14.0) 25 (30.9) 0.0001 0.006 β-blockers 19 (3.0) 12 (2.2) 7 (8.6) 0.006 0.003 α-blockers 7 (1.1) 6 (1.1) 1 (1.2) 1.000 0.837 Lipid-lowering agents 68 (10.8) 53 (9.6) 15 (18.5) 0.016 0.045 Values are numbers (percentages) or means±SDs.

tors 54 (8.6) 39 (7.1) 15 (18.5) 0.0006 0.004 Calcium channel blockers 102 (16.2) 77 (14.0) 25 (30.9) 0.0001 0.006 β-blockers 19 (3.0) 12 (2.2) 7 (8.6) 0.006 0.003 α-blockers 7 (1.1) 6 (1.1) 1 (1.2) 1.000 0.837 Lipid-lowering agents 68 (10.8) 53 (9.6) 15 (18.5) 0.016 0.045 Values are numbers (percentages) or means±SDs. *Age, sex, and diabetes duration-adjusted p value. Age, sex, and diabetes duration were adjusted except for itself, respectively. †Excludes patients treated with oral antidiabetic drugs and insulin. ‡Includes patients treated with oral antidiabetic drugs and insulin. BMI, body mass index; CVD, cardiovascular disease; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; HbA1c, glycated hemoglobin; HDL-C, high-density lipoprotein cholesterol; SBP, systolic blood pressure; TC, total cholesterol. The total number of visits was 55 855 (per-patient median, 81; IQR 36–126.5). The median follow-up period was 15.4 years (IQR 6.6–16.4).

BMI, body mass index; CVD, cardiovascular disease; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; HbA1c, glycated hemoglobin; HDL-C, high-density lipoprotein cholesterol; SBP, systolic blood pressure; TC, total cholesterol. The total number of visits was 55 855 (per-patient median, 81; IQR 36–126.5). The median follow-up period was 15.4 years (IQR 6.6–16.4). Correlations between variability in HbA1c and SBP, as well as variability and mean HbA1c or SBP HbA1cCV was correlated with mean HbA1c (r=0.428, p<0.0001). However, no correlation was observed between SBPCV and mean SBP (r=0.038, p=0.341). HbA1cVIM (proportional to SD/mean2.70) was independent of mean HbA1c (r=−0.0006, p=0.989), and SBPVIM (proportional to SD/mean1.11) was independent of mean SBP (r=−0.00008, p=0.998). A weak correlation was observed between HbA1cCV and SBPCV (r=0.107, p=0.007), as well as between HbA1cVIM and SBPVIM (r=0.100, p=0.012), which were statistically significant owing to the large number of patients involved.

.989), and SBPVIM (proportional to SD/mean1.11) was independent of mean SBP (r=−0.00008, p=0.998). A weak correlation was observed between HbA1cCV and SBPCV (r=0.107, p=0.007), as well as between HbA1cVIM and SBPVIM (r=0.100, p=0.012), which were statistically significant owing to the large number of patients involved. Adjusted Kaplan-Meier survival curves for a CVD event In figure 1, the Kaplan-Meier survival curves for a CVD event were given for the four groups classified by median HbA1cCV and SBPCV values, after adjusting for age, mean HbA1c, mean SBP, and the number of visits (ln-transformed). The adjusted survival curves revealed a clear association between variability in HbA1c and SBP and the incidence of CVD. The highest incidence was observed in the high-HbA1cCV and high-SBPCV group, followed by the low-HbA1cCV and high-SBPCV group, and finally the high-HbA1cCV and low-SBPCV group. Conversely, the lowest incidence was observed in the low-HbA1cCV and low-SBPCV group. Figure 1 Kaplan-Meier survival curves for a CVD event classified according to the median HbA1cCV and SBPCV values, after adjusting for age, mean HbA1c, mean SBP, and the number of visits (ln-transformed) (CV, coefficient of variation; CVD, cardiovascular disease; HbA1c, glycated hemoglobin; SBP, systolic blood pressure).

Figure 1 Kaplan-Meier survival curves for a CVD event classified according to the median HbA1cCV and SBPCV values, after adjusting for age, mean HbA1c, mean SBP, and the number of visits (ln-transformed) (CV, coefficient of variation; CVD, cardiovascular disease; HbA1c, glycated hemoglobin; SBP, systolic blood pressure). Respective and combined effects of visit-to-visit variability in HbA1c and SBP on the incidence of CVD The respective and combined effects of HbA1cCV and SBPCV as continuous and categorical variables on the incidence of CVD were evaluated by a multivariate analysis performed using a Cox proportional hazard model (table 2). The analysis was performed after adjusting for the patient characteristics described in the Methods section. The HRs were calculated in accordance with a 1SD increment. In model 1, HbA1cCV but not SBPCV was incorporated as one of the covariates in combination with other clinical variables. In model 2, SBPCV but not HbA1cCV was incorporated. In model 3, both of these were incorporated. In these three models, HbA1cCV and SBPCV were treated as a standardized continuous variable. In model 4, HbA1cCV and SBPCV were also included in the model as model 3, but these were treated as categorical variables, being dichotomized by the respective median value. Table 2 Multivariate Cox proportional hazard models for the incidence of CVD in association with HbA1cCV and SBPCV as continuous and categorical variables divided by their respective median values

Respective and combined effects of visit-to-visit variability in HbA1c and SBP on the incidence of CVD The respective and combined effects of HbA1cCV and SBPCV as continuous and categorical variables on the incidence of CVD were evaluated by a multivariate analysis performed using a Cox proportional hazard model (table 2). The analysis was performed after adjusting for the patient characteristics described in the Methods section. The HRs were calculated in accordance with a 1SD increment. In model 1, HbA1cCV but not SBPCV was incorporated as one of the covariates in combination with other clinical variables. In model 2, SBPCV but not HbA1cCV was incorporated. In model 3, both of these were incorporated. In these three models, HbA1cCV and SBPCV were treated as a standardized continuous variable. In model 4, HbA1cCV and SBPCV were also included in the model as model 3, but these were treated as categorical variables, being dichotomized by the respective median value. Table 2 Multivariate Cox proportional hazard models for the incidence of CVD in association with HbA1cCV and SBPCV as continuous and categorical variables divided by their respective median values Model 1 Model 2 Model 3 Model 4 HR (95% CI) p Value HR (95% CI) p Value HR (95% CI) p Value HR (95% CI) p Value Continuous variables HbA1cCV (1SD increment) 1.39 (1.10 to 1.76) 0.006 NA NA 1.33 (1.04 to 1.70) 0.024 NA NA SBPCV (1SD increment) NA NA 1.33 (1.07 to 1.63) 0.009 1.26 (1.02 to 1.57) 0.032 NA NA Categorical variables Low HbA1cCV and low SBPCV NA NA NA NA NA NA 1 Low HbA1cCV and high SBPCV NA NA NA NA NA NA 2.55 (1.22 to 5.32) 0.013 High HbA1cCV and low SBPCV NA NA NA NA NA NA 2.64 (1.16 to 6.00) 0.020 High HbA1cCV and high SBPCV NA NA NA NA NA NA 3.08 (1.45 to 6.55) 0.003 All models were adjusted for mean HbA1c, mean SBP, number of visits (ln-transformed), age, sex, diabetes duration, BMI, TC/HDL-C, eGFR, baseline smoking status, baseline alcohol intake, baseline use of insulin, and baseline use of antihypertensive agents.

A1cCV and high SBPCV NA NA NA NA NA NA 3.08 (1.45 to 6.55) 0.003 All models were adjusted for mean HbA1c, mean SBP, number of visits (ln-transformed), age, sex, diabetes duration, BMI, TC/HDL-C, eGFR, baseline smoking status, baseline alcohol intake, baseline use of insulin, and baseline use of antihypertensive agents. BMI, body mass index; CV, coefficient of variation; CVD, cardiovascular disease; eGFR, estimated glomerular filtration rate; HbA1c, glycated hemoglobin; HDL-C, high-density lipoprotein cholesterol; NA, not applicable; SBP, systolic blood pressure; TC, total cholesterol. In model 1, HbA1cCV was a significant predictor of the incidence of CVD independent of mean HbA1c. In model 2, SBPCV was a significant predictor of the incidence of CVD independent of mean SBP. In model 3, HbA1cCV and SBPCV as continuous variables were significant predictors of the incidence of CVD, independent of mean HbA1c and mean SBP, simultaneously. In model 4, patients were classified into four groups by median HbA1cCV and SBPCV values. HRs for the incidence of CVD associated with these four groups (with low-HbA1cCV and low-SBPCV serving as the reference group) were calculated. The HRs were highest in the high-HbA1cCV and high-SBPCV group and significantly higher in the low-HbA1cCV and high-SBPCV group and high-HbA1cCV and low-SBPCV group than in the low-HbA1cCV and low-SBPCV group.

ociated with these four groups (with low-HbA1cCV and low-SBPCV serving as the reference group) were calculated. The HRs were highest in the high-HbA1cCV and high-SBPCV group and significantly higher in the low-HbA1cCV and high-SBPCV group and high-HbA1cCV and low-SBPCV group than in the low-HbA1cCV and low-SBPCV group. In table 3, VIM was used instead of CV in all models shown in table 2. The results of model 1, 2, and 3 of table 3 were similar to those of table 2. In model 4, the HRs were significantly higher in the high-HbA1cVIM and high-SBPVIM group, and higher, but not significantly, in the low-HbA1cVIM and high-SBPVIM group and high-HbA1cVIM and low-SBPVIM group than in the low-HbA1cVIM and low-SBPVIM group. Table 3 Multivariate Cox proportional hazard models for the incidence of CVD in association with HbA1cVIM and SBPVIM as continuous and categorical variables divided by their respective median values

In table 3, VIM was used instead of CV in all models shown in table 2. The results of model 1, 2, and 3 of table 3 were similar to those of table 2. In model 4, the HRs were significantly higher in the high-HbA1cVIM and high-SBPVIM group, and higher, but not significantly, in the low-HbA1cVIM and high-SBPVIM group and high-HbA1cVIM and low-SBPVIM group than in the low-HbA1cVIM and low-SBPVIM group. Table 3 Multivariate Cox proportional hazard models for the incidence of CVD in association with HbA1cVIM and SBPVIM as continuous and categorical variables divided by their respective median values Model 1 Model 2 Model 3 Model 4 HR (95% CI) p Value HR (95% CI) p Value HR (95% CI) p Value HR (95% CI) p Value Continuous variables HbA1cVIM (1SD increment) 1.33 (1.07 to 1.65) 0.011 NA NA 1.28 (1.02 to 1.61) 0.035 NA NA SBPVIM (1SD increment) NA NA 1.33 (1.08 to 1.63) 0.008 1.27 (1.03 to 1.57) 0.026 NA NA Categorical variables Low HbA1cVIM and low SBPVIM NA NA NA NA NA NA 1 Low HbA1cVIM and high SBPVIM NA NA NA NA NA NA 1.77 (0.90 to 3.51) 0.10 High HbA1cVIM and low SBPVIM NA NA NA NA NA NA 1.72 (0.82 to 3.63) 0.15 High HbA1cVIM and high SBPVIM NA NA NA NA NA NA 2.19 (1.12 to 4.29) 0.022 All models were adjusted for mean HbA1c, mean SBP, number of visits (ln-transformed), age, sex, diabetes duration, BMI, TC/HDL-C, eGFR, baseline smoking status, baseline alcohol intake, baseline use of insulin, and baseline use of antihypertensive agents.

cVIM and high SBPVIM NA NA NA NA NA NA 2.19 (1.12 to 4.29) 0.022 All models were adjusted for mean HbA1c, mean SBP, number of visits (ln-transformed), age, sex, diabetes duration, BMI, TC/HDL-C, eGFR, baseline smoking status, baseline alcohol intake, baseline use of insulin, and baseline use of antihypertensive agents. BMI, body mass index; CVD, cardiovascular disease; eGFR, estimated glomerular filtration rate; HbA1c, glycated hemoglobin; HDL-C, high-density lipoprotein cholesterol; NA, not applicable; SBP, systolic blood pressure; TC, total cholesterol; VIM, variation independent of mean.

cVIM and high SBPVIM NA NA NA NA NA NA 2.19 (1.12 to 4.29) 0.022 All models were adjusted for mean HbA1c, mean SBP, number of visits (ln-transformed), age, sex, diabetes duration, BMI, TC/HDL-C, eGFR, baseline smoking status, baseline alcohol intake, baseline use of insulin, and baseline use of antihypertensive agents. BMI, body mass index; CVD, cardiovascular disease; eGFR, estimated glomerular filtration rate; HbA1c, glycated hemoglobin; HDL-C, high-density lipoprotein cholesterol; NA, not applicable; SBP, systolic blood pressure; TC, total cholesterol; VIM, variation independent of mean. Stratified analysis by mean HbA1c and SBP levels and the effects of visit-to-visit variability in HbA1c and SBP, as continuous variables, on the incidence of CVD A stratified analysis was performed by mean HbA1c and SBP levels of 7.0% and 130 mm Hg, respectively (table 4). The effects of visit-to-visit variability in HbA1c and SBP on the incidence of CVD were evaluated as continuous variables using multivariate Cox proportional hazard models. The covariates are described in the Methods section. In models 1 and 2, for the stratum of patients with mean HbA1c<7.0%, neither HbA1cCV nor HbA1cVIM was significant, but SBPCV and SBPVIM were borderline significant. For the stratum of patients with mean HbA1c≥7.0%, HbA1cCV and HbA1cVIM were borderline significant, whereas neither SBPCV nor SBPVIM was significant. For the stratum of patients with mean SBP<130 mm Hg, neither HbA1cCV nor HbA1cVIM was significant, whereas SBPCV and SBPVIM were significant. For the stratum of patients with mean SBP≥130 mm Hg, the HRs associated with HbA1cCV and HbA1cVIM were drastically elevated compared with those for the stratum of patients with mean SBP<130 mm Hg (interaction p=0.018 for HbA1cCV; interaction p=0.016 for HbA1cVIM). Thus, HbA1cCV and HbA1cVIM were significant predictors, whereas neither SBPCV nor SBPVIM was significant.

g, the HRs associated with HbA1cCV and HbA1cVIM were drastically elevated compared with those for the stratum of patients with mean SBP<130 mm Hg (interaction p=0.018 for HbA1cCV; interaction p=0.016 for HbA1cVIM). Thus, HbA1cCV and HbA1cVIM were significant predictors, whereas neither SBPCV nor SBPVIM was significant. Table 4 HRs for incidence of CVD associated with variability in HbA1c and SBP stratified according to mean HbA1c and SBP levels Mean HbA1c<7.0% Mean HbA1c≥7.0% Interaction p Mean SBP<130 mm Hg Mean SBP≥130 mm Hg Interaction p HR (95% CI) p Value HR (95% CI) p Value HR (95% CI) p Value HR (95% CI) p Value Events/patients 37/338 44/294 37/306 44/326 Model 1 HbA1cCV (1SD increment) 1.20 (0.76 to 1.88) 0.44 1.37 (1.00 to 1.87) 0.052 0.99 (0.65 to 1.52) 0.97 1.77 (1.28 to 2.46) 0.0007 0.018 SBPCV (1SD increment) 1.33 (0.97 to 1.81) 0.076 1.30 (0.89 to 1.88) 0.18 0.79 1.59 (1.11 to 2.29) 0.013 1.16 (0.86 to 1.56) 0.34 Model 2 HbA1cVIM (1SD increment) 1.17 (0.83 to 1.66) 0.37 1.38 (0.99 to 1.93) 0.054 0.96 (0.64 to 1.44) 0.85 1.73 (1.25 to 2.38) 0.0009 0.016 SBPVIM (1SD increment) 1.32 (0.97 to 1.80) 0.076 1.31 (0.90 to 1.90) 0.16 0.82 1.59 (1.12 to 2.28) 0.011 1.19 (0.89 to 1.60) 0.25 All models were adjusted for mean HbA1c, mean SBP, number of visits (ln-transformed), age, sex, diabetes duration, BMI, TC/HDL-C, eGFR, baseline smoking status, baseline alcohol intake, baseline use of insulin, and baseline use of antihypertensive agents.

90) 0.16 0.82 1.59 (1.12 to 2.28) 0.011 1.19 (0.89 to 1.60) 0.25 All models were adjusted for mean HbA1c, mean SBP, number of visits (ln-transformed), age, sex, diabetes duration, BMI, TC/HDL-C, eGFR, baseline smoking status, baseline alcohol intake, baseline use of insulin, and baseline use of antihypertensive agents. BMI, body mass index; CV, coefficient of variation; CVD, cardiovascular disease; eGFR, estimated glomerular filtration rate; HbA1c, glycated hemoglobin; HDL-C, high-density lipoprotein cholesterol; SBP, systolic blood pressure; TC, total cholesterol; VIM, variation independent of mean. Discussion Long-term visit-to-visit variability in HbA1c and SBP represented a combined and additive risk for the incidence of CVD simultaneously in patients with type 2 diabetes. In addition, the risk of CVD associated with an increase in HbA1c variability was drastically elevated among patients with mean SBP≥130 mm Hg. In contrast, the CVD risk associated with an increase in SBP variability increased significantly among patients with mean SBP<130 mm Hg, whereas an increase in HbA1c variability was likely to have no effect. It is suggested that a synergistic effect exists between HbA1c variability and mean SBP levels for the incidence of CVD (interaction p<0.05).

ociated with an increase in SBP variability increased significantly among patients with mean SBP<130 mm Hg, whereas an increase in HbA1c variability was likely to have no effect. It is suggested that a synergistic effect exists between HbA1c variability and mean SBP levels for the incidence of CVD (interaction p<0.05). The results of the Action in Diabetes and Vascular Disease: Preterax and Diamicron Modified Release Controlled Evaluation (ADVANCE) trial showed that visit-to-visit SBP variability is an independent risk factor for macrovascular complications in patients with type 2 diabetes.6 Recently, the same trial, as the first large-scale study, showed that visit-to-visit HbA1c variability predicts the future development of macrovascular events independent of cardiovascular risk factors, including mean HbA1c.4 However, the study differed from this study in that participants had type 2 diabetes, were ≥55 years old, and had a history of major macrovascular or microvascular disease or at least one other risk factor for vascular disease. Furthermore, no data were reported on the combined effect of visit-to-visit variability in HbA1c and SBP in that trial. Thus, this study is the first to report the combined effect of visit-to-visit variability in HbA1c and SBP on the incidence of CVD simultaneously in patients with type 2 diabetes.

or vascular disease. Furthermore, no data were reported on the combined effect of visit-to-visit variability in HbA1c and SBP in that trial. Thus, this study is the first to report the combined effect of visit-to-visit variability in HbA1c and SBP on the incidence of CVD simultaneously in patients with type 2 diabetes. The relationship between BP variability, calculated for different long-term sequential time frames, and mortality risk was reported.23 It was a large cohort study using real-world clinical BP data of 14 522 treated patients with hypertension who were followed up over 35 years. The results indicated that long-term variability in SBP and diastolic BP (DBP) calculated for the long term (1–4 years) and ultra long term (5–9 years) were significant predictors of mortality, independent of mean BP. This relationship was also evident in subgroups with mean SBP<140 mm Hg, which agrees with our results. In our study, a stratified analysis by mean SBP of 130 mm Hg was also performed. HbA1c variability was a significant predictor among patients with mean SBP≥130 mm Hg, but SBP variability was not. In contrast, SBP variability was a significant predictor, but HbA1c variability was not among patients with mean SBP<130 mm Hg. Thus, even if mean SBP was maintained within the normal range, increased SBP variability was a risk factor for a CVD event.

among patients with mean SBP≥130 mm Hg, but SBP variability was not. In contrast, SBP variability was a significant predictor, but HbA1c variability was not among patients with mean SBP<130 mm Hg. Thus, even if mean SBP was maintained within the normal range, increased SBP variability was a risk factor for a CVD event. We previously reported the relationships between the risk of CVD in patients with type 2 diabetes and both visit-to-visit variability and time-to-effect differences in BP.7 Our earlier study showed that increases in SBP over the preceding 3–5 years resulted in a significant CVD risk. Therefore, increased HbA1c variability over the preceding 3–5 years could emerge as a more harmful risk factor for the incidence of CVD. Thus, stabilizing variability in HbA1c level and lowering BP during these periods seem to be particularly important.

P over the preceding 3–5 years resulted in a significant CVD risk. Therefore, increased HbA1c variability over the preceding 3–5 years could emerge as a more harmful risk factor for the incidence of CVD. Thus, stabilizing variability in HbA1c level and lowering BP during these periods seem to be particularly important. Glycemic variability is associated with a risk of severe hypoglycemia.24 Severe hypoglycemia is also associated with a higher risk of CVD.25 In our study, however, no information about a severe hypoglycemic episode in which a patient required the assistance of another person was available. Therefore, hypoglycemia was defined by a fasting or casual blood glucose level at clinic visits of less than 60 mg/dL at least once during follow-up. Hypoglycemia occurred in 32 patients. There was no association between hypoglycemia at clinic visits and the incidence of CVD (data not shown). Furthermore, the association between visit-to-visit variability in HbA1c and SBP and the incidence of CVD was independent of hypoglycemia at clinic visits (data not shown). Even if hypoglycemia was defined as less than 50 mg/dL, results remained almost unchanged (data not shown).

the incidence of CVD (data not shown). Furthermore, the association between visit-to-visit variability in HbA1c and SBP and the incidence of CVD was independent of hypoglycemia at clinic visits (data not shown). Even if hypoglycemia was defined as less than 50 mg/dL, results remained almost unchanged (data not shown). The possible practical factors, such as age, the number of visits, seasonal changes, lifestyle factors, non-adherence with antidiabetic and antihypertensive medications, and improper titration/dosing of those medications, could contribute to visit-to-visit variability in HbA1c and SBP. In this study, we evaluated what baseline characteristics contributed to subsequent visit-to-visit variability in HbA1c and SBP. Younger age and increased baseline HbA1c contributed to HbA1c variability, while older age, increased baseline SBP, decreased baseline BMI and eGFR, baseline use of insulin, and baseline use of antihypertensive agents contributed to SBP variability (data not shown).

bsequent visit-to-visit variability in HbA1c and SBP. Younger age and increased baseline HbA1c contributed to HbA1c variability, while older age, increased baseline SBP, decreased baseline BMI and eGFR, baseline use of insulin, and baseline use of antihypertensive agents contributed to SBP variability (data not shown). SD as a measure of variability is most familiar to clinicians and easy to calculate, although SD is affected by the mean value. The same analyses were also performed using SD; consequently, similar results were obtained. Concretely, in the participants of our study, the median values (IQR) of the mean HbA1c (%), HbA1cSD (%), mean SBP (mm Hg), and SBPSD (mm Hg) were 6.93 (6.43–7.48), 0.57 (0.38–0.79), 130.6 (121.8–139.5), and 11.6 (9.8–14.1), respectively. The risk of the incidence of CVD significantly increased 2.45-fold for each 1% increase in HbA1cSD and 2.02-fold for every 10 mm Hg increase in SBPSD. It is important that clinicians pay attention to variability in HbA1c and SBP, which may represent a higher cardiovascular risk than their mean values and can lead to combined additive risk for the incidence of CVD.

increased 2.45-fold for each 1% increase in HbA1cSD and 2.02-fold for every 10 mm Hg increase in SBPSD. It is important that clinicians pay attention to variability in HbA1c and SBP, which may represent a higher cardiovascular risk than their mean values and can lead to combined additive risk for the incidence of CVD. One strength of this study was the use of a database comprising ‘real-world’ observations with a long-term follow-up period. In addition, we addressed a topic that may help provide novel and effective strategies for preventing CVD in patients with type 2 diabetes. However, several limitations must also be mentioned. First, this is a retrospective observational cohort study. The results merely indicate the association, but not causation. In addition, potential information biases included changes in sample examination methods with time and differences in the number of visits. However, some data generated by the different measurement methods were converted using linear regression equations derived from duplicate assays. Visit-to-visit BP variability increases with the number of visits.26 To adjust for the possibility that the number of visits could influence variability, the number of visits was included in the model as a covariate after being ln-transformed. Second, BP data were derived typically from a single measurement obtained at each visit. However, visit-to-visit BP variability is a reproducible, not random, phenomenon.27 28 It can be assumed that higher reproducibility is achieved when automated devices are used.27–29 Third, the incidence of CVD was partly self-reported by patients. However, the end point was defined clearly, and 66 (81.5%) of the 81 CVD events were determined according to a thorough review of medical records. Therefore, merely 15 events (18.5%) were based on results of the questionnaire. Fourth, lipids were determined irrespective of fasting or postprandial status. Therefore, we could not conduct an analysis using triglyceride. Nevertheless, TC/HDL-C was used as a covariate for the analysis, because TC/HDL-C is the best lipid predictor of CVD for males with type 2 diabetes.19 20 Fifth, non-adherence with antidiabetic and antihypertensive medications could contribute to visit-to-visit variability in HbA1c and SBP; however, we have no data on adherence with medication.

s used as a covariate for the analysis, because TC/HDL-C is the best lipid predictor of CVD for males with type 2 diabetes.19 20 Fifth, non-adherence with antidiabetic and antihypertensive medications could contribute to visit-to-visit variability in HbA1c and SBP; however, we have no data on adherence with medication. Finally, our study participants were recruited from a single hospital in Japan and included more males than females; however, their clinical characteristics were similar to those of patients in another large-scale study in Japan.30 It is uncertain whether our findings can be generalized to other ethnic groups. Prospective international multicenter trials are needed. In conclusion, long-term visit-to-visit variability in HbA1c and SBP represented a combined and additive risk for the incidence of CVD simultaneously in patients with type 2 diabetes. In addition, it is suggested that a synergistic effect exists between HbA1c variability and mean SBP levels for the incidence of CVD. Even if mean SBP is maintained within the normal range, SBP variability can be a risk factor for a CVD event. Our findings indicate the possibility that stabilization of variability in HbA1c and SBP as well as lowering of their mean levels can be an efficient strategy for preventing the incidence of CVD. The authors thank Kumiko Kimura (the Institute for Adult Diseases, Asahi Life Foundation) for her assistance with the collection of research data, and Akifumi Kushiyama (the Institute for Adult Diseases, Asahi Life Foundation) for his advice on revision of this manuscript.

In conclusion, long-term visit-to-visit variability in HbA1c and SBP represented a combined and additive risk for the incidence of CVD simultaneously in patients with type 2 diabetes. In addition, it is suggested that a synergistic effect exists between HbA1c variability and mean SBP levels for the incidence of CVD. Even if mean SBP is maintained within the normal range, SBP variability can be a risk factor for a CVD event. Our findings indicate the possibility that stabilization of variability in HbA1c and SBP as well as lowering of their mean levels can be an efficient strategy for preventing the incidence of CVD. The authors thank Kumiko Kimura (the Institute for Adult Diseases, Asahi Life Foundation) for her assistance with the collection of research data, and Akifumi Kushiyama (the Institute for Adult Diseases, Asahi Life Foundation) for his advice on revision of this manuscript. Contributors: TT contributed to the study concept and design, collection and recording of data, data analyses and interpretation, and writing of the manuscript. YM and MS contributed to the data interpretation. HY and YI were responsible for intellectual contributions. TT is the guarantor, and had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Competing interests: None declared. Ethics approval: This study was approved by the Institutional Review Board of the Institute for Adult Diseases, Asahi Life Foundation. Provenance and peer review: Not commissioned; externally peer reviewed.

Contributors: TT contributed to the study concept and design, collection and recording of data, data analyses and interpretation, and writing of the manuscript. YM and MS contributed to the data interpretation. HY and YI were responsible for intellectual contributions. TT is the guarantor, and had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Competing interests: None declared. Ethics approval: This study was approved by the Institutional Review Board of the Institute for Adult Diseases, Asahi Life Foundation. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: No additional data are available.

Key messages Financial incentives could reduce inequalities by improving health screening uptake; however, there are ethical concerns about the use of financial incentives in health. We asked people with diabetes about their views on the acceptability of incentives. Vouchers were more acceptable than cash payments, and those in deprived and middle-aged groups found incentives most acceptable. Introduction Deprivation is linked to higher levels of diabetes,1 worse outcomes, and the development of secondary associated conditions. Retinopathy complications are the top cause of blindness in the UK working-aged population.2 In areas with the greatest socioeconomic deprivation, diabetes prevalence is highest and screening attendance is lowest.3 4 Demographic factors such as age and deprivation are predictors of adherence to screening programs.5 If screening programs only reach the less deprived, they have the potential to exacerbate health inequalities. Financial incentives have successfully changed organizational behavior, paying health providers more for improved performance.6 Incentives are also a powerful mechanism to encourage healthy behaviors such as smoking cessation.7–9 Research into whether financial incentives increase participation is largely confined to immunization initiatives.10 Compared with disincentives and increased regulation, incentives could be popular;11 however, ethical concerns around coercion, personal responsibility, and unintended consequences have inhibited research.12

arch into whether financial incentives increase participation is largely confined to immunization initiatives.10 Compared with disincentives and increased regulation, incentives could be popular;11 however, ethical concerns around coercion, personal responsibility, and unintended consequences have inhibited research.12 This study explores these concerns with people who have diabetes and their health professionals. The primary objective was to determine whether these groups find the use of financial incentives in screening ethically acceptable. We analyzed the impact of demographic profile and different incentives on perceived acceptability. Acceptability is a difficult term—this study did not attempt to interrogate moral acceptability, which tends to be inflexible and abstract. Nor was it primarily concerned with personal acceptability, which can be selfish. The ethical acceptability here refers to societal norms, which can be flexible and reflect personal experience and demographic group. Methods We used a questionnaire methodology, as this allowed for the collection of comparable data from a broad sample. Questionnaires also allowed for privacy and anonymity, and minimal participant time commitment.

Acceptability is a difficult term—this study did not attempt to interrogate moral acceptability, which tends to be inflexible and abstract. Nor was it primarily concerned with personal acceptability, which can be selfish. The ethical acceptability here refers to societal norms, which can be flexible and reflect personal experience and demographic group. Methods We used a questionnaire methodology, as this allowed for the collection of comparable data from a broad sample. Questionnaires also allowed for privacy and anonymity, and minimal participant time commitment. Participants The study was conducted across west London, in partnership with local provider 1st Retinal Screen Ltd. All 925 people with diabetes due to be screened in August 2013 were approached. Of these, 828 attended their appointment, during which they were issued with a voluntary questionnaire. There was a response rate of 44%. Questionnaires were posted to those who did not attend, with an invitation to rearrange their appointment. The views of this cohort would be especially valuable, as the success of an incentive scheme would be judged on an increase in participation; however, only six responded. Finally, to ascertain the perspective of diabetes professionals, a questionnaire was issued to the screening team at 1st Retinal Screen, and to all general practitioner (GP) practices in the area for which addresses were available. The screening team were very engaged with the project and all five completed the questionnaire. None of the local GP practices responded.

professionals, a questionnaire was issued to the screening team at 1st Retinal Screen, and to all general practitioner (GP) practices in the area for which addresses were available. The screening team were very engaged with the project and all five completed the questionnaire. None of the local GP practices responded. Questionnaire For attendees and non-attendees, questions 1–3 captured basic demographic information: gender, age group, and postcode. Question 4 comprised statement pairs expressing opposing positions on ethical concerns raised in previous studies.12 13 Owing to the complex nature of ethical debate and the difficulties in coding this into statements, it was not always possible to devise pairs that were logical opposites. However, the intention was to encourage participants to decide in favor of or against incentives. 4.1: The message Offering an incentive sends a positive message to everyone that participation in the screening program is important. Offering an incentive sends out the wrong message—people should not be paid to do the right thing—they should participate in screening for the good of their own health, not for money. The way health services are delivered can communicate underlying messages about health and the National Health Service (NHS). This question captured views on how that message might be interpreted by the target group. 4.2: Fairness/responsibility At the moment, only some people get screened. If this scheme encourages everyone to be screened, this is fairer for society and means everyone has the same chance to be healthy.

The way health services are delivered can communicate underlying messages about health and the National Health Service (NHS). This question captured views on how that message might be interpreted by the target group. 4.2: Fairness/responsibility At the moment, only some people get screened. If this scheme encourages everyone to be screened, this is fairer for society and means everyone has the same chance to be healthy. This ignores where the real responsibility lies. Doctors, the Government, and the NHS should encourage people to look after themselves without paying. Incentive schemes may reduce health inequalities, but this could be perceived as the NHS ignoring its other responsibilities. 4.3: NHS values Incentives complement the values of the NHS, by encouraging those most in need to access the resources that will improve their lives. Incentives undermine the key principles of the NHS, that healthcare should be free when using services. Key NHS principles are equality of access and the absence of financial transaction. Financial incentives may reduce inequality, but introduce cash transfers at the point of care. 4.4: Autonomy/choice It is better than compulsory screening—people can still choose whether to take the incentive. Very poor people may feel that they do not have a choice—offering an incentive might coerce, or force people into doing something they do not want to do. Another key debate around the remit of the NHS is on consent. If incentives are perceived to threaten autonomy, this could render them unacceptable.

4.4: Autonomy/choice It is better than compulsory screening—people can still choose whether to take the incentive. Very poor people may feel that they do not have a choice—offering an incentive might coerce, or force people into doing something they do not want to do. Another key debate around the remit of the NHS is on consent. If incentives are perceived to threaten autonomy, this could render them unacceptable. 4.5: Wider impact If it works, it could be used in other services as a way of encouraging people to look after their health. It could lead to people expecting money to use other NHS services, which could be expensive in the long run. Introducing a new and effective mechanism in one service could lead to proliferation. 4.6: Opportunity cost Offering incentives will cost the NHS less in the long run by preventing bad health before it becomes a problem. Taxpayers’ money should not go to people who are not looking after themselves properly. Should the NHS be a passive provider of care for those who seek it, or take extra steps to combat problems early on? 4.7: Risk Even if screening can never be 100% accurate, if it prevents one person from losing their sight, it is worthwhile. Incentives would only be acceptable if screening was 100% accurate, because people would trust the result. This question attempts to capture the problem of inaccuracy and risk. The NHS makes a concerted effort to communicate the fallibility and stress of screening. Taking stronger measures to encourage participation acts against these warnings, and may persuade those who would otherwise decide against screening.

Incentives would only be acceptable if screening was 100% accurate, because people would trust the result. This question attempts to capture the problem of inaccuracy and risk. The NHS makes a concerted effort to communicate the fallibility and stress of screening. Taking stronger measures to encourage participation acts against these warnings, and may persuade those who would otherwise decide against screening. Question 5 asked participants to consider different types of incentive. A cash reward is often found to be least acceptable but most effective,14 so it would be worth measuring the difference in opinion between cash and healthy food or book vouchers (5.2 and 5.1, respectively) and between covering expenses and additional reward (5.6 and 5.7). These questions test the ethical problem with cash payments against encouraging positive behavior and compensation, respectively. The principles of behavioral economics suggest that we overweight small probabilities, which might make a prize draw effective.15 Alternatively, a financial reward every time might be more acceptable (5.10 and 5.9). For the individual, a large incentive (eg, £60) might be a more agreeable prospect. Equally this might be perceived as wasteful and a smaller reward (eg, £6) might be perceived as more proportionate (5.3 and 5.4).

The principles of behavioral economics suggest that we overweight small probabilities, which might make a prize draw effective.15 Alternatively, a financial reward every time might be more acceptable (5.10 and 5.9). For the individual, a large incentive (eg, £60) might be a more agreeable prospect. Equally this might be perceived as wasteful and a smaller reward (eg, £6) might be perceived as more proportionate (5.3 and 5.4). Targeting certain demographics could help further reduce inequalities and make a scheme more affordable.16 However, in a health system underpinned by solidarity and equality, targeted rewards could cause resentment. We asked whether incentives should only be available for those who had not attended before, or for all (5.7 and 5.8). For those who did not attend questions 6–8 ascertained their reasons, to identify whether they had reservations or were misinformed. A previous study into the reasons for non-attendance informed the question design.5 The questionnaire was tested on patients with diabetes as well as patients without diabetics. Analysis The high response rate from the attendee group allowed full statistical analysis. However, the markedly low response rate from the non-attendees and health professionals only permitted limited analysis. Postcodes gave an Index of Multiple Deprivation score, which allowed the group to be divided into deprivation quintiles, with 1 representing the least deprived and 5 representing the most deprived. Simple summary statistics demonstrated the difference in acceptability for question 4.

Analysis The high response rate from the attendee group allowed full statistical analysis. However, the markedly low response rate from the non-attendees and health professionals only permitted limited analysis. Postcodes gave an Index of Multiple Deprivation score, which allowed the group to be divided into deprivation quintiles, with 1 representing the least deprived and 5 representing the most deprived. Simple summary statistics demonstrated the difference in acceptability for question 4. To ensure question 4 was capturing ethical acceptability, we ran a factor analysis which showed a high degree of correlation between responses (figure 1). The exception being the final statement pair. This outlier question may draw on different constructs, about the risks of screening itself, rather than how this relates to ethical acceptability. Cronbach's α coefficient measure of internal consistency found responses to question 4 to be highly reliable (α=0.91) when 4.7 was excluded. Figure 1 Factor analysis showing correlation between responses to question 4 (Cronbach's α coefficient measure of internal consistency). (NHS, National Health Service). By coding negative views as 0 and positive as 1, giving equal weight to questions 4.1–4.6, a mean score for participants’ perception of ethical acceptability was devised. Zero indicated that incentives were found to be unacceptable for all subquestions, and 1 indicated that incentives were found to be acceptable for all (table 1). Table 1 Percentage of positive responses when investigating participants' perception of ethical acceptability

By coding negative views as 0 and positive as 1, giving equal weight to questions 4.1–4.6, a mean score for participants’ perception of ethical acceptability was devised. Zero indicated that incentives were found to be unacceptable for all subquestions, and 1 indicated that incentives were found to be acceptable for all (table 1). Table 1 Percentage of positive responses when investigating participants' perception of ethical acceptability Question Percentage of positive responses Age Deprivation quintiles Female Male All Least deprived Most deprived 18–39 40–64 65+ 18–39 40–64 65+ 18–39 40–64 65+ 1 2 3 4 5 4.1—The message 33.33 34.38 34.04 44.44 52.17 31.46 41.67 45.81 28.57 30.3 36.76 37.5 45.59 44.45 4.2—Fairness/responsibility 33.33 43.75 34.55 85.71 53.51 48.15 70 50 42.03 39.06 50.88 46.77 53.13 49.21 4.3—NHS values 33.33 43.55 32.69 50 47.37 37.04 45.45 46.02 34.81 34.38 44.46 39.34 50 44.45 4.4—Autonomy and choice 66.67 60.66 47.17 83.33 63.96 56.41 77.78 62.79 51.88 59.38 60 55.17 62.9 57.63 4.5—Wider impact 33.33 47.54 42.59 50 57.02 46.15 45.55 53.11 44.78 37.1 53.45 44.26 57.14 56.45 4.6—Opportunity cost 0 57.89 47.06 83.33 59.82 45 55.56 59.17 45.11 41.67 53.57 58.62 65.08 49.15 4.7—Risk 66.66 74.19 73.77 87.5 77.57 84.71 81.82 76.33 80.41 80.95 79.37 80 73.85 79.37 Overall acceptability (excluding 4.7) 0=completely unacceptable 1=very acceptable Mean 0.33 0.46 0.39 0.53 0.55 0.40 0.48 0.52 0.39 0.37 0.47 0.46 0.55 0.48 SD 0.44 0.41 0.40 0.46 0.42 0.40 0.44 0.42 0.40 0.40 0.41 0.42 0.41 0.43 NHS, National Health Service.

7.5 77.57 84.71 81.82 76.33 80.41 80.95 79.37 80 73.85 79.37 Overall acceptability (excluding 4.7) 0=completely unacceptable 1=very acceptable Mean 0.33 0.46 0.39 0.53 0.55 0.40 0.48 0.52 0.39 0.37 0.47 0.46 0.55 0.48 SD 0.44 0.41 0.40 0.46 0.42 0.40 0.44 0.42 0.40 0.40 0.41 0.42 0.41 0.43 NHS, National Health Service. To understand the effect of sex, age, and deprivation on responses to question 4, a binary logistic regression was used, and a linear regression analysis on the overall mean acceptability score (table 2). A linear regression analysis on question 5 helped to understand the relative impact the demographic variables had on the perceived ethical acceptability for each type of incentive (table 3). Table 2 Binary logistic and linear regression analyses to investigate the effect of sex, age, and deprivation on responses to question 4 using the overall mean acceptability score

To understand the effect of sex, age, and deprivation on responses to question 4, a binary logistic regression was used, and a linear regression analysis on the overall mean acceptability score (table 2). A linear regression analysis on question 5 helped to understand the relative impact the demographic variables had on the perceived ethical acceptability for each type of incentive (table 3). Table 2 Binary logistic and linear regression analyses to investigate the effect of sex, age, and deprivation on responses to question 4 using the overall mean acceptability score Predictors Binary logistic regression Linear regression Overall acceptability (excluding 4.7) 4.1 4.2 4.3 4.4 4.5 4.6 4.7 The message Fairness/ responsibility NHS values Autonomy/ choice Wider impact Opportunity cost Risk Sex β 0.511 0.481 0.133 0.197 0.294 0.021 0.356 0.052 p 0.036 0.045 0.587 0.423 0.223 0.932 0.205 0.266 OR 1.667 1.617 1.142 1.218 1.341 1.021 1.428 Age β −0.540 −0.349 −0.403 −0.460 −0.191 −0.412 0.258 −0.092 p 0.011 0.103 0.060 0.040 0.369 0.059 0.298 0.025 OR 0.583 0.705 0.668 0.631 0.827 0.662 1.295 Deprivation β 0.138 0.083 0.088 −0.029 0.167 0.085 −0.028 0.026 p 0.097 0.313 0.284 0.736 0.044 0.311 0.778 0.105 OR 1.148 1.086 1.093 0.972 1.182 1.089 0.972 R2 0.063 0.035 0.024 0.023 0.029 0.023 0.013 0.029 NHS, National Health Service. Table 3 Linear regression analysis to investigate the effect of sex, age, and deprivation on responses to question 5 using the overall mean acceptability score

Predictors Binary logistic regression Linear regression Overall acceptability (excluding 4.7) 4.1 4.2 4.3 4.4 4.5 4.6 4.7 The message Fairness/ responsibility NHS values Autonomy/ choice Wider impact Opportunity cost Risk Sex β 0.511 0.481 0.133 0.197 0.294 0.021 0.356 0.052 p 0.036 0.045 0.587 0.423 0.223 0.932 0.205 0.266 OR 1.667 1.617 1.142 1.218 1.341 1.021 1.428 Age β −0.540 −0.349 −0.403 −0.460 −0.191 −0.412 0.258 −0.092 p 0.011 0.103 0.060 0.040 0.369 0.059 0.298 0.025 OR 0.583 0.705 0.668 0.631 0.827 0.662 1.295 Deprivation β 0.138 0.083 0.088 −0.029 0.167 0.085 −0.028 0.026 p 0.097 0.313 0.284 0.736 0.044 0.311 0.778 0.105 OR 1.148 1.086 1.093 0.972 1.182 1.089 0.972 R2 0.063 0.035 0.024 0.023 0.029 0.023 0.013 0.029 NHS, National Health Service. Table 3 Linear regression analysis to investigate the effect of sex, age, and deprivation on responses to question 5 using the overall mean acceptability score Linear regression 5.1 5.2 5.3 5.4 5.5 5.6 5.7 5.8 5.9 5.10 Predictors Vouchers Cash Small payment Large payment Expenses Extra reward Targeted For all Pay every time Prize draw Sex β 0.083 −0.204 −0.044 −0.075 −0.022 −0.295 −0.180 −0.091 −0.023 −0.098 p 0.663 0.284 0.822 0.662 0.910 0.088 0.344 0.660 0.904 0.599 Age β 0.405 0.298 0.257 0.420 −0.125 0.516 0.178 0.482 0.207 0.302 p 0.016 0.073 0.134 0.005 0.452 0.001 0.281 0.008 0.220 0.065 Deprivation β −0.015 −0.106 −0.181 −0.100 0.030 −0.052 −0.037 −0.107 −0.098 −0.143 p 0.818 0.102 0.007 0.089 0.651 0.368 0.564 0.129 0.135 0.025 R2 0.020 0.027 0.035 0.042 0.003 0.056 0.009 0.035 0.014 0.033 β, regression coefficient; p, significance; NHS, National Health Service.

05 0.452 0.001 0.281 0.008 0.220 0.065 Deprivation β −0.015 −0.106 −0.181 −0.100 0.030 −0.052 −0.037 −0.107 −0.098 −0.143 p 0.818 0.102 0.007 0.089 0.651 0.368 0.564 0.129 0.135 0.025 R2 0.020 0.027 0.035 0.042 0.003 0.056 0.009 0.035 0.014 0.033 β, regression coefficient; p, significance; NHS, National Health Service. Results Participants attending the clinic Sixty-two percent of participants felt that offering an incentive sends out the wrong message and that people should participate in screening for their own good. Participants also showed concern about the way incentives align with the principles of the NHS, with 58.7% of those who answered feeling that it undermines the lack of financial transaction at the point of healthcare delivery. Responses were balanced on whether it could be a useful tool elsewhere, or simply lead to people expecting payment in other services, with half (50.6%) fearing this possibility. There was less concern about the potential for coercion and impact on the ability to choose, with 58.6% feeling that individual autonomy was retained. There was also slightly less concern about using health resources in this way, with more participants (53.1%) thinking it might cost the NHS less in the long run by preventing ill health. In total, 52.8% of participants felt that other routes should be tried, and more responsibility placed on the NHS to encourage people to take better care of themselves.

about using health resources in this way, with more participants (53.1%) thinking it might cost the NHS less in the long run by preventing ill health. In total, 52.8% of participants felt that other routes should be tried, and more responsibility placed on the NHS to encourage people to take better care of themselves. Overall, participants tended to answer consistently negatively or positively, with 31.3% answering negatively on all questions, and a further 11.2% erring on the side of unacceptability for all but one of the questions. In total, 27.7% responded positively to all questions with a further 5.6% answering positively to all but one. Those in the middle age group were more accepting of incentives generally (mean=0.52, SD=0.42), which is especially evident for questions on the impact of autonomy (4.4) and on opportunity cost (4.6), which were answered positively by 62.79% and 59.17% of participants, respectively. In general, the older age group gave more negative responses, particularly on the message incentives send out, with only 28.57% giving a favorable view. Overall, participants in the two most deprived quintiles felt more positively about incentives than those in the least deprived groups (table 1). Overall acceptability of incentives decreases with age in a statistically significant way (β=−0.092, p=0.025). Increasing age was also found to be a significant predictor on some of the ethical dimensions. Older participants were more likely to be concerned about the message (β=−0.54, p=0.011) and the impact on autonomy and choice (β=−0.46, p=0.04; table 2).

eases with age in a statistically significant way (β=−0.092, p=0.025). Increasing age was also found to be a significant predictor on some of the ethical dimensions. Older participants were more likely to be concerned about the message (β=−0.54, p=0.011) and the impact on autonomy and choice (β=−0.46, p=0.04; table 2). Deprivation and sex were not found to be significant predictors of overall acceptability; however, male participants were more likely to feel positively about the message (β=0.511, p=0.036) and issues of fairness and responsibility (β=0.481, p=0.045). Those in the most deprived groups were statistically more likely to feel positive about the wider impact of incentives (β=0.167, p=0.044; table 2). On incentive types, vouchers were preferred to cash, with 36.5% of participants who answered the question finding vouchers to be very acceptable, compared with 18.6% who found cash very acceptable and 64.4% of participants finding it completely unacceptable. When asked whether small or large cash incentives would be preferable, most responded that neither would be acceptable. However, among those who responded positively, almost twice as many (65) found small incentives very acceptable than those finding large very acceptable (38).

pants finding it completely unacceptable. When asked whether small or large cash incentives would be preferable, most responded that neither would be acceptable. However, among those who responded positively, almost twice as many (65) found small incentives very acceptable than those finding large very acceptable (38). Compensating for expenses incurred was found to be either slightly or completely acceptable by over half (51.1%) of all those who responded. This compares to 72% of participants who felt paying over and above reasonable expenses to be slightly or completely unacceptable. In total, 24.4% of participants found targeted incentives to be wholly or slightly acceptable, whereas 34.5% found incentives for all to be acceptable. More deprived groups were statistically more likely to find small payments (β=−0.181, p=0.007) and prize draws acceptable (β=−0.143, p=0.025). With increasing age, vouchers were found to be more acceptable in a statistically significant way (β=0.405, p=0.016). Other statistically significant preferences for incentive types which increase with age were large payments (β=0.42, p=0.005), paying more than just expenses (β=0.516, p=0.001), and incentives for all (β=0.516, p=0.008; table 3).

rs were found to be more acceptable in a statistically significant way (β=0.405, p=0.016). Other statistically significant preferences for incentive types which increase with age were large payments (β=0.42, p=0.005), paying more than just expenses (β=0.516, p=0.001), and incentives for all (β=0.516, p=0.008; table 3). The results demonstrate a clear preference overall for offering vouchers and expenses over cash and paying more than just expenses. The mean responses to question 5 also demonstrate a clear preference overall for vouchers compared with any other type of incentive. Small cash payments were, in general, preferred to large—although overall the mean score demonstrates less approval of cash payments in general, regardless of size. Prize draws were also found to be more unacceptable than small payments every time; however, this question was framed in the context of cash payments, and the mean responses reflect a general aversion to cash incentives (table 4). Table 4 Mean scores for domains investigated in question 5 by age and deprivation quintile

The results demonstrate a clear preference overall for offering vouchers and expenses over cash and paying more than just expenses. The mean responses to question 5 also demonstrate a clear preference overall for vouchers compared with any other type of incentive. Small cash payments were, in general, preferred to large—although overall the mean score demonstrates less approval of cash payments in general, regardless of size. Prize draws were also found to be more unacceptable than small payments every time; however, this question was framed in the context of cash payments, and the mean responses reflect a general aversion to cash incentives (table 4). Table 4 Mean scores for domains investigated in question 5 by age and deprivation quintile Mean and SD (scale: 1 (very acceptable)—5 (completely unacceptable)) Deprivation quintiles Age groups Least deprived Most deprived Question 18–39 40–64 65+ 1 2 3 4 5 Vouchers 5.1  Mean 2.45 2.59 3.01 2.98 2.83 2.62 2.58 2.92 SD 1.63 1.55 1.68 1.69 1.69 1.62 1.59 1.58 Cash 5.2  Mean 4.33 3.69 4.21 4.16 4.14 3.96 3.43 4.00 SD 1.23 1.70 1.45 1.42 1.49 1.66 1.82 1.56 Small payment (eg, £6) 5.3  Mean 3.80 3.55 3.94 4.20 3.73 3.93 3.33 3.52 SD 1.62 1.63 1.62 1.37 1.69 1.54 1.69 1.77 Large payment (eg, £60) 5.4  Mean 4.33 3.96 4.58 4.51 4.40 4.31 3.73 4.28 SD 1.37 1.59 1.12 1.15 1.29 1.35 1.71 1.48 Expenses 5.5  Mean 3.09 2.66 2.57 2.68 2.38 2.80 2.37 2.91 SD 1.64 1.56 1.65 1.61 1.53 1.67 1.37 1.79 Extra reward 5.6  Mean 4.36 3.79 4.53 4.34 4.18 4.07 3.84 4.14 SD 1.29 1.56 1.09 1.12 1.44 1.49 1.63 1.38 Targeted 5.7  Mean 3.25 3.65 3.85 3.82 3.79 3.62 3.94 3.52 SD 1.76 1.58 1.56 1.50 1.59 1.58 1.48 1.71 For all 5.8  Mean 3.75 3.11 3.85 3.95 3.53 3.27 2.89 3.63 SD 1.60 1.81 1.67 1.51 1.80 1.88 1.81 1.71 Payment every time (eg, £6) 5.9  Mean 4.27 3.58 3.99 4.03 3.78 3.95 3.67 3.53 SD 1.27 1.62 1.58 1.44 1.68 1.57 1.61 1.70 Prize draw (eg, £6000) 5.10  Mean 4.55 3.65 4.24 4.16 4.28 4.11 3.40 3.87 SD .69 1.71 1.36 1.28 1.37 1.52 1.79 1.68 Using a paired samples t test to compare incentive types, vouchers were found to be significantly more acceptable than cash (t=−10.380, p=0.000). Small payments were more acceptable than large (t=−5.048, p=0.000). Expenses were significantly more acceptable than paying extra (t=−12.886, p=0.000).

28 1.37 1.52 1.79 1.68 Using a paired samples t test to compare incentive types, vouchers were found to be significantly more acceptable than cash (t=−10.380, p=0.000). Small payments were more acceptable than large (t=−5.048, p=0.000). Expenses were significantly more acceptable than paying extra (t=−12.886, p=0.000). Non-attendees—postal responses There were a range of views on ethical acceptability in general and for different incentive types, but mostly positive. While it is not possible to draw conclusions or properly analyze the non-attendee group, the mean acceptability for different incentive types broadly aligns to the responses provided by the attendee group, with participants preferring vouchers, small payments, expenses, and incentives for all. Staff On overall ethical acceptability, there was a marked difference from the non-attendee group. There were far more negative responses than positive, with only 41% of all responses to question 4 falling in favor of incentives. On incentive types, staff gave responses broadly in line with those of both patient groups, demonstrating a preference for vouchers and expenses, on average finding these to be acceptable. While neither small nor large cash payments were found to be particularly acceptable, there was a clear preference for smaller incentives. The staff were the only group to prefer a prize draw over payment for all.

t groups, demonstrating a preference for vouchers and expenses, on average finding these to be acceptable. While neither small nor large cash payments were found to be particularly acceptable, there was a clear preference for smaller incentives. The staff were the only group to prefer a prize draw over payment for all. Discussion This study has shown that, across those surveyed, opinion is heavily polarized. Participants were more likely to find incentives ethically acceptable or unacceptable across all counts than to give a mixture of answers. The ethical concerns expressed most strongly across the sample group were that incentives undermine the responsibility of Government and the NHS to promote healthy behaviors, but also of individuals to look after themselves. It was felt that financial transactions sat uncomfortably with the principles of the NHS. This suggests that the most pressing concern is about how financial incentives may derail existing cultural norms around responsibility. One interpretation could be that there is an ingrained societal understanding of the roles and responsibilities between patients, Government, and healthcare professionals, and uneasiness about how introducing financial transactions may affect this delicate balance. This should be seen in the context of the broad program of change from which the NHS has recently emerged, and may reflect a wider sense of anxiety about the state of the NHS, and uncertainty about how services will be delivered in the future.

out how introducing financial transactions may affect this delicate balance. This should be seen in the context of the broad program of change from which the NHS has recently emerged, and may reflect a wider sense of anxiety about the state of the NHS, and uncertainty about how services will be delivered in the future. Age was an important factor, with those in the middle age group most likely to find incentives acceptable. That this group has the most positive view should offer hope to proponents of incentives, as people currently in this age group are at the greatest risk of developing type 2 diabetes in the near future. Incentives may be a useful way to encourage early positive interaction with the necessary screening regime. Those above the age of 65 are much more likely to have a negative view of this approach. It is therefore vital for health organizations to have a full understanding of the age profile of their population in order to communicate and design a successful incentive scheme. Across all groups, there was a strong preference for healthy food or book vouchers over cash, and only covering incurred expenses. This suggests that incentives can be framed in a more positive light if they are shown to fulfill an additional purpose, over and above encouraging greater participation.

Age was an important factor, with those in the middle age group most likely to find incentives acceptable. That this group has the most positive view should offer hope to proponents of incentives, as people currently in this age group are at the greatest risk of developing type 2 diabetes in the near future. Incentives may be a useful way to encourage early positive interaction with the necessary screening regime. Those above the age of 65 are much more likely to have a negative view of this approach. It is therefore vital for health organizations to have a full understanding of the age profile of their population in order to communicate and design a successful incentive scheme. Across all groups, there was a strong preference for healthy food or book vouchers over cash, and only covering incurred expenses. This suggests that incentives can be framed in a more positive light if they are shown to fulfill an additional purpose, over and above encouraging greater participation. For the patient groups, incentives which are targeted at those most in need of screening were found to be less acceptable than incentives for all. This supports the interpretation that there is less concern about escalating costs to the public purse, and suggests an aversion to perceived injustice. Equality, where all are given the same offer regardless of circumstance, may be a more important pillar of the NHS than equity, where resources are allocated according to need.

orts the interpretation that there is less concern about escalating costs to the public purse, and suggests an aversion to perceived injustice. Equality, where all are given the same offer regardless of circumstance, may be a more important pillar of the NHS than equity, where resources are allocated according to need. In considering how to implement any kind of financial incentive scheme, policymakers could look to the results of this study for insight. These results support previous studies which show that vouchers which can be used for healthy foods are often found to be more acceptable than cash.13 This may be because healthy food is seen as a treatment for wider health problems, helping make diabetes more manageable, rather than a reward for attendance. There is the opportunity here to use incentives to achieve other policy goals, such as healthy eating or physical activity through gym membership. This study also points to the possibility that ethically acceptable ways of implementing incentive schemes could align well with another universal policy goal: achieving more with fewer resources. Participants in this study generally preferred small to large payments and favored an approach that solely reimbursed reasonable expenses—this is especially true for the most deprived. This group also found prize draws to be more acceptable, which would be more cost-effective than paying all attendees.

fewer resources. Participants in this study generally preferred small to large payments and favored an approach that solely reimbursed reasonable expenses—this is especially true for the most deprived. This group also found prize draws to be more acceptable, which would be more cost-effective than paying all attendees. In revealing that those in the most deprived groups find incentives more acceptable than those in the least, this study suggests that financial incentives could be used as a lever to reduce inequalities. If universal incentive schemes can have targeted appeal, they could raise the health status of the most deprived in our communities. If incentive schemes can be found to be acceptable and practicable for diabetic retinopathy screening, other national or local screening programs might find it a useful tool, for example, for breast, cervical and bowel cancer, and abdominal aortic aneurysm. Policymakers would be advised to proceed more carefully here, as some of these schemes are screening for arguably more serious conditions, whereas others may involve more invasive methods or subsequent treatment, such as a risky operation. This leads to a shift in the ethical dimensions for different types of screening.

olicymakers would be advised to proceed more carefully here, as some of these schemes are screening for arguably more serious conditions, whereas others may involve more invasive methods or subsequent treatment, such as a risky operation. This leads to a shift in the ethical dimensions for different types of screening. Limitations As an initial investigation, this study has revealed some useful and important findings. An expanded study, with a more refined questionnaire aimed at reaching a wider and more diverse sample group would allow inferences to be made about the whole population. A sample which assessed the views of those from different areas of the country would have been useful; as urban Londoners are not necessarily representative of the country as a whole. The sample was also largely comprised of those who might benefit from an incentive scheme. There is a concern that responses may be distorted in self-interest. We experienced difficulty in gathering the views of those who did not attend their screening appointment: while this was anticipated, it is nonetheless frustrating, as this is the group that an incentive scheme would ultimately target. It might have been more productive to call each non-attendee directly and guide them through the questionnaire over the telephone. In asking participants to choose from predetermined ethical concerns, the argument was framed, potentially influencing participants’ views. A more open approach which asked participants for their views on incentives with fewer prompts may have been less subject to potential bias in questionnaire design.

We experienced difficulty in gathering the views of those who did not attend their screening appointment: while this was anticipated, it is nonetheless frustrating, as this is the group that an incentive scheme would ultimately target. It might have been more productive to call each non-attendee directly and guide them through the questionnaire over the telephone. In asking participants to choose from predetermined ethical concerns, the argument was framed, potentially influencing participants’ views. A more open approach which asked participants for their views on incentives with fewer prompts may have been less subject to potential bias in questionnaire design. The number of people who chose to make additional comments suggests that a free-text field in the questionnaire would have been welcomed and allowed the collection of qualitative data. Some participants expressed difficulty as the ethical statement pairs were not mutually exclusive, and that this made it difficult to choose. Others felt they appreciated both sides of the argument. Further research into this subject might want to allow participants more time for consideration, and revisit the questionnaire design to provide more detailed or granular questions. There is certainly space for further qualitative, quantitative, and philosophical research into the ethical acceptability of incentives, and their implementation and efficacy.

The number of people who chose to make additional comments suggests that a free-text field in the questionnaire would have been welcomed and allowed the collection of qualitative data. Some participants expressed difficulty as the ethical statement pairs were not mutually exclusive, and that this made it difficult to choose. Others felt they appreciated both sides of the argument. Further research into this subject might want to allow participants more time for consideration, and revisit the questionnaire design to provide more detailed or granular questions. There is certainly space for further qualitative, quantitative, and philosophical research into the ethical acceptability of incentives, and their implementation and efficacy. Conclusion This preliminary study suggests that some groups may find financial incentives ethical acceptable, although there will be many who take the opposite view. Replication and extension of this study may offer further insight, and research into the efficacy of incentive schemes in increasing screening program participation will be essential to the development of our understanding in this area. Contributors: HW wrote the manuscript and researched data. IV contributed to discussion and edited the manuscript. CB contributed to discussion and reviewed the manuscript.

Conclusion This preliminary study suggests that some groups may find financial incentives ethical acceptable, although there will be many who take the opposite view. Replication and extension of this study may offer further insight, and research into the efficacy of incentive schemes in increasing screening program participation will be essential to the development of our understanding in this area. Contributors: HW wrote the manuscript and researched data. IV contributed to discussion and edited the manuscript. CB contributed to discussion and reviewed the manuscript. Funding: This research was part funded by the NIHR Health Service and Delivery Research programme. This paper presents independent research funded by the National Institute for Health Research (NIHR grant number - 12/209 HS&DR Researcher Led programme). The research was supported by the NIHR Biomedical Research Centre based at Imperial College Healthcare NHS Trust and Imperial College London. Competing interests: None declared. Patient consent: Obtained. Ethics approval: Social Care Research Ethics Committee http://www.screc.org.uk. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: IV is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. The data are available by contacting the corresponding authors.

Key messages Mice lacking of an adaptive immune system (SCID mice) have disruptions in the concentrations of various circulating cytokines. SCID mice have significant glucose intolerance that is independent of diet and obesity. The absence of functional B and T cells in SCID mice is not protective against diet-induced glucose intolerance progression. Introduction Increasing interest in the convergent biology of insulin resistance and diabetes mellitus has been partly fueled by the need for effective interventions, and by recent data implicating the immune system in the pathogenesis of these diseases and perhaps shared mechanisms. It has long been established that obesity, insulin resistance, and chronic inflammation are often associated, but it has been challenging to determine if a causal relationship exists among these conditions. It is known that the adipose tissue of obese mice and humans is characterized by an accumulation of macrophages and, it is thought, that this is a necessary component of disease progression.1–3 The secretion of cytokines tumor necrosis factor-α and interleukin (IL)-6 by infiltrating macrophages are thought to play a direct role in reducing the insulin sensitivity of regional adipocytes.4–6 Even more compelling, the role of the immune system in insulin resistance is no longer thought to be restricted to just the innate arm, but also includes an adaptive immune response. B and T cells have both been implicated in insulin resistance, with both detrimental and protective effects being reported.7–21

cytes.4–6 Even more compelling, the role of the immune system in insulin resistance is no longer thought to be restricted to just the innate arm, but also includes an adaptive immune response. B and T cells have both been implicated in insulin resistance, with both detrimental and protective effects being reported.7–21 The results to date have been compelling; however, they fail to address if the adaptive immune system is a necessary component for metabolic homeostasis; that is, by manipulating the immune phenotype are we disrupting a physiological homeostatic mechanism? This study aimed to examine this by assessing the effects of immune incompetence on metabolic homeostasis in the presence and absence of a high-fat diet (HFD). Research design and methods Ethics statement All animal studies were completed in accordance with the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health and approved by the Institutional Animal Care and Use Committee of Stanford University.

The results to date have been compelling; however, they fail to address if the adaptive immune system is a necessary component for metabolic homeostasis; that is, by manipulating the immune phenotype are we disrupting a physiological homeostatic mechanism? This study aimed to examine this by assessing the effects of immune incompetence on metabolic homeostasis in the presence and absence of a high-fat diet (HFD). Research design and methods Ethics statement All animal studies were completed in accordance with the Guide for the Care and Use of Laboratory Animals of the National Institutes of Health and approved by the Institutional Animal Care and Use Committee of Stanford University. Mouse manipulations Male, 6-week-old BALB/c control and BALB/c scid mice were purchased from Jackson Laboratories. Blood glucose measurements were taken using the One Touch Ultra blood glucose meter. The HFD, 60% kilocalories from fat, (#D12492) and control, 10% kilocalories from fat, (#D12450B) diets were purchased from Research Diets, Inc. Fasting insulin levels were detected using the Crystal Chem Inc Ultra Sensitive Mouse Insulin ELISA (#90080). The Stanford Human Immune Monitoring Core completed the mouse cytokine Luminex array. The Stanford Department of Comparative Medicine's Histology Lab prepared all histological samples. Glucose tolerance test The glucose tolerance test (GTT) was completed as previously described.22 Briefly, mice were fasted for 8 h followed by intraperitoneal administration of 2 g/kg of D-glucose.

Mouse manipulations Male, 6-week-old BALB/c control and BALB/c scid mice were purchased from Jackson Laboratories. Blood glucose measurements were taken using the One Touch Ultra blood glucose meter. The HFD, 60% kilocalories from fat, (#D12492) and control, 10% kilocalories from fat, (#D12450B) diets were purchased from Research Diets, Inc. Fasting insulin levels were detected using the Crystal Chem Inc Ultra Sensitive Mouse Insulin ELISA (#90080). The Stanford Human Immune Monitoring Core completed the mouse cytokine Luminex array. The Stanford Department of Comparative Medicine's Histology Lab prepared all histological samples. Glucose tolerance test The glucose tolerance test (GTT) was completed as previously described.22 Briefly, mice were fasted for 8 h followed by intraperitoneal administration of 2 g/kg of D-glucose. Insulin tolerance test The insulin tolerance test (ITT) was completed as previously described.22 Briefly, mice were fasted for 6 h followed by intraperitoneal administration of 2 units/kg of Humulin R insulin. Statistical analysis All data were presented as the mean±SEM. For data analysis between two groups, significance was determined by the unpaired Student t test and defined as p≤0.05. For repeated measures data (GTT and ITT), significance was determined by two-way analysis of variance for repeated measures with Bonferroni correction and defined as p≤0.05.

presented as the mean±SEM. For data analysis between two groups, significance was determined by the unpaired Student t test and defined as p≤0.05. For repeated measures data (GTT and ITT), significance was determined by two-way analysis of variance for repeated measures with Bonferroni correction and defined as p≤0.05. Results The phenotypes of BALB/c scid (SCID) and the BALB/c strain controls (control) were compared for variations in glucose and insulin sensitivity. SCID mice had significantly reduced glucose tolerance compared with control mice (p=0.0049; figure 1A). There was no difference observed between the fasting blood glucose levels or in response to insulin administration (figure 1B). Figure 1 The insulin sensitivity of BALB/c scid mice is lower than that in BALB/c mice. (A) Glucose tolerance test of control (white circles) and SCID (black circles) mice following a 2 g/kg glucose challenge (n=4/5). (B) Insulin tolerance test of control (white circles) and SCID (black circles) mice following a 2 units/kg insulin challenge (n=4/5). (C) The fasting, blood-insulin levels of control and SCID mice (n=4/5). (D) The per cent β-cell function as determined by the homeostatic model assessment (HOMA) of control and SCID mice (n=4/5). (E) The per cent insulin sensitivity as determined by the HOMA of control and SCID mice (n=4/5). (Error bars represent the mean±SEM, and *signifies a significant difference between the designated groups of p<0.05.)

-cell function as determined by the homeostatic model assessment (HOMA) of control and SCID mice (n=4/5). (E) The per cent insulin sensitivity as determined by the HOMA of control and SCID mice (n=4/5). (Error bars represent the mean±SEM, and *signifies a significant difference between the designated groups of p<0.05.) On evaluation of fasting insulin levels, SCID mice had a 174% greater fasting blood insulin concentration than that of controls (p=0.0345; figure 1C). Use of the homeostatic model assessment method indicated that SCID mice had significantly greater β-cell function, 205±31%, when compared with control mice, 71±7% (p=0.0183; figure 1D). Correspondingly, the method indicated that SCID mice had a nearly significant reduction in insulin sensitivity (p=0.0977; figure 1E). Despite the apparent insulin resistance of SCID mice, SCID mice weighed significantly less at 23.3±0.3 g (p=0.0332; figure 2A). The total body weight of control mice was 25.0±0.4 g. This difference was predominately due to differences in fat mass. SCID mice had 50% the adipose volume when compared with control mice. SCID mice had 110±26 mg of inguinal subcutaneous adipose tissue and 211±37 mg of visceral epididymal adipose tissue (mean±SEM). In contrast, control mice possessed significantly more adipose tissue with 204±25 and 423±70 mg, respectively (p=0.0383 and 0.0252; figure 2B,C). Histopathology did not reveal any clear differences between the tissues collected from mice (figure 2D,E).

sue and 211±37 mg of visceral epididymal adipose tissue (mean±SEM). In contrast, control mice possessed significantly more adipose tissue with 204±25 and 423±70 mg, respectively (p=0.0383 and 0.0252; figure 2B,C). Histopathology did not reveal any clear differences between the tissues collected from mice (figure 2D,E). Figure 2 The adipose tissue of BALB/c scid mice is quantitatively, but not qualitatively, different than that of BALB/c control mice. (A) The total body weight of control BALB/c and BALB/c scid (SCID) mice (n=4/5). (B) The total weight of the subcutaneous adipose tissue bilaterally isolated from the inguinal region of control and SCID mice (n=4/5). (C) The total weight of the epididymal adipose tissue bilaterally isolated from control and SCID mice (n=4/5). (Error bars represent the mean±SEM, and *signifies a significant difference between the control and SCID groups of p<0.05.) (D) 10× light microscopy of H&E stained sections of liver, pancreas, subcutaneous, and visceral adipose tissue collected from control and SCID mice. Scale bar is 200 μm. (E) 100× light microscopy of H&E stained sections of liver, pancreas, subcutaneous, and visceral adipose tissue collected from control and SCID mice. (Scale bar is 50 μm.)

opy of H&E stained sections of liver, pancreas, subcutaneous, and visceral adipose tissue collected from control and SCID mice. Scale bar is 200 μm. (E) 100× light microscopy of H&E stained sections of liver, pancreas, subcutaneous, and visceral adipose tissue collected from control and SCID mice. (Scale bar is 50 μm.) Luminex was used to evaluate the circulating cytokine profiles of control and SCID mice and revealed that SCID mice had alterations in the levels of circulating cytokines with significantly higher levels of granulocyte colony-stimulating factor (GCSF), IL-4, monocyte chemoattractant protein 3 (MCP3), MCP1, IL-17A, macrophage inflammatory protein 2 (MIP2), IL-1A, IL-28, IL-18, and IL-31 and significantly lower levels of C-C motif chemokine ligand 5 (CCL5) (RANTES) relative to controls (figure 3). The cytokines with the greatest difference in SCID mice were MCP1 and MCP3 and GCSF which were both significantly higher in the immunodeficient SCID mice (p=0.0116, 0.0011, and 0.0207, respectively).

nd IL-31 and significantly lower levels of C-C motif chemokine ligand 5 (CCL5) (RANTES) relative to controls (figure 3). The cytokines with the greatest difference in SCID mice were MCP1 and MCP3 and GCSF which were both significantly higher in the immunodeficient SCID mice (p=0.0116, 0.0011, and 0.0207, respectively). Figure 3 BALB/c scid mice have higher levels of macrophage-stimulating cytokines in circulation compared with BALB/c control mice. The average fold difference in the (MFI) of SCID mice relative to the average MFI of control mice (n=5). (Error bars represent the mean±SEM, and *signifies a significant difference between the control and SCID groups of p<0.05.) GCSF, granulocyte colony-stimulating factor; GMCSF, granulocyte macrophage colony-stimulating factor; GROA, growth-regulated protein alpha; IFNA, interferon alpha; IFNG interferon gamma; IL, interleukin; IP, interferon-inducible protein-10; LIX, lipopolysaccharide-inducible CXC chemokine; MCSF, macrophage colony-stimulating factor; MFI, median fluorescence intensity; MCP, monocyte chemoattractant protein; MIP, macrophage inflammatory protein; TGFB, transforming growth factor beta; TNFA, tumor necrosis factor alpha; VEGF, vascular endothelial growth factor.

IX, lipopolysaccharide-inducible CXC chemokine; MCSF, macrophage colony-stimulating factor; MFI, median fluorescence intensity; MCP, monocyte chemoattractant protein; MIP, macrophage inflammatory protein; TGFB, transforming growth factor beta; TNFA, tumor necrosis factor alpha; VEGF, vascular endothelial growth factor. To assess the role of the adaptive immune response in diet-induced glucose intolerance, control (BALB/c) or BALB/c scid (SCID) mice were fed either a HFD, with 60% of the kilocalories derived from fat, or a control diet, with 10% of the kilocalories derived from fat, for 14 weeks. At completion of the dietary intervention, there was no significant difference between the total body weights for all groups of mice (figure 4A). BALB/c mice fed on a HFD had significantly (p=0.0473) reduced glucose tolerance relative to BALB/c mice maintained on a control diet with significantly higher fasting blood glucose levels. The SCID mice maintained on the control diet also had significantly (p=0.0013) greater fasting blood glucose levels relative to the BALB/c mice on the control diet. Similarly, the SCID mice on the control diet had a significantly (p=0.0005) reduced glucose tolerance compared with BALB/c mice on the control diet. SCID mice maintained on a HFD had significantly (p<0.0104) higher fasting blood glucose levels compared with all other mouse groups. Additionally, SCID mice on a HFD had significantly higher blood glucose levels following a glucose challenge when compared with the SCID mice on a control diet (p=0.0027) and the BALB/c mice on a HFD (p=0.0001; figure 4B).

ignificantly (p<0.0104) higher fasting blood glucose levels compared with all other mouse groups. Additionally, SCID mice on a HFD had significantly higher blood glucose levels following a glucose challenge when compared with the SCID mice on a control diet (p=0.0027) and the BALB/c mice on a HFD (p=0.0001; figure 4B). Figure 4 High-fat diet induces progression of glucose intolerance in BALB/c scid mice. (A) The total body weight of control (BALB/c) and BALB/c scid (SCID) mice following 14 weeks of a control diet (control) or high-fat diet (HFD; n=5). (B) Glucose tolerance test of BALB/c (circle) and SCID (square) mice on the control (white) or HFD (black) diet following a 2 g/kg glucose challenge (n=5). (Error bars represent the mean±SEM, *signifies a significant difference between the designated groups of p<0.05.) Discussion The goal of this work was to assess the role of the adaptive immune system in glucose homeostasis versus diet-induced glucose intolerance. We approached this by comparing the phenotypes of the immunodeficient SCID mouse and its immunocompetent BALB/c strain control. Collectively, our studies indicate that an adaptive immune response is a necessary, physiological component of the metabolic organ required for glucose homeostasis, but its absence does not protect against diet-induced glucose intolerance.

of the immunodeficient SCID mouse and its immunocompetent BALB/c strain control. Collectively, our studies indicate that an adaptive immune response is a necessary, physiological component of the metabolic organ required for glucose homeostasis, but its absence does not protect against diet-induced glucose intolerance. SCID mice were insulin resistant in the absence of obesity and were, in fact, leaner than their counterparts. This result supports the role of the adaptive immune system in metabolic homeostasis by suggesting that obesity is not a required component for immune-mediated disruption in insulin tolerance. The finding that the adaptive immune system is important for glucose homeostasis is consistent with work that demonstrated an early populating of the visceral adipose tissue of lean mice by T regulatory cells in an antigen-dependent fashion. The T regulatory cells performed both inflammatory and metabolic functions.20 A recent human trial suggests that a HFD can induce insulin resistance without causing concurrent detectable changes in immune function. While short-term overfeeding did result in insulin resistance, it did not cause significant alternations in the populations of immune cells or cytokine gene expression profiles within the subcutaneous adipose tissue.23 These findings support that glucose sensitivity is an extraordinarily complex process, melding both physiological and pathological mechanisms of a multitude of organ systems.

cause significant alternations in the populations of immune cells or cytokine gene expression profiles within the subcutaneous adipose tissue.23 These findings support that glucose sensitivity is an extraordinarily complex process, melding both physiological and pathological mechanisms of a multitude of organ systems. Interestingly, cytokine profile analyses in immuocompetent controls and immunodeficient SCID mice demonstrated several significant differences; however, those cytokines that were elevated to a degree as to be functionally significant were all associated with neutrophil proliferation, GCSF, or macrophage recruitment, MCP1 and MCP3. This is of special note as macrophage accumulation within the adipose tissue, especially visceral adipose tissue, is a consistent finding among insulin-resistant mice and humans and is thought to be a necessary component in diet-induced insulin resistance.1–3 Additionally, visceral adipose tissue-associated T cells are thought to be the major source of macrophage-recruiting chemokines.13 Our results indicated that the MCP1 and MCP3 levels were elevated in the absence of T cells, suggesting an alternative source for these chemokines. The likely source for these chemokines is adipocytes. Others have demonstrated that adipocytes secrete macrophage-recruiting chemokines and that their expression increases in adipocytes on introduction of a HFD.24 25 Collectively, these results support the hypothesis that one role of the adaptive immune system in glucose homeostasis is maintaining levels of macrophage-stimulating cytokines and preventing disruption in glucose homeostasis. Additionally, our work indicates that the absence of functional B and T cells does not provide protection against HFD-induced insulin resistance as SCID mice fed on a HFD, although already insulin resistant, developed even greater glucose intolerance.

mulating cytokines and preventing disruption in glucose homeostasis. Additionally, our work indicates that the absence of functional B and T cells does not provide protection against HFD-induced insulin resistance as SCID mice fed on a HFD, although already insulin resistant, developed even greater glucose intolerance. Surprisingly, mice fed on a HFD did not gain significant amounts of weight. This result may be due to our low sample size, and/or our use of the BALB/c mouse strain. The BALB/c strain has previously been shown to demonstrate less pathology in association with a HFD when compared with other mouse strains.26 An additional finding was that 6-week-old SCID mice on a control diet did not have significantly greater fasting blood glucose levels when compared with same-aged BALB/c mice on a control diet (figure 1A and online supplementary material 1); however, 20-week-old SCID mice on a control diet did have significantly higher fasting blood glucose levels when compared with BALB/c mice on a control diet (figure 4B). This result suggests that the glucose intolerant phenotype of SCID mice worsens with age. It is possible that B and T cells have a developmental and/or continuous regulatory role in adipose biology, influencing the gut microbiome, adipose tissue biology, and the adipose tissue inflammatory tone in a way that is exacerbated by age.7–21 27

s that the glucose intolerant phenotype of SCID mice worsens with age. It is possible that B and T cells have a developmental and/or continuous regulatory role in adipose biology, influencing the gut microbiome, adipose tissue biology, and the adipose tissue inflammatory tone in a way that is exacerbated by age.7–21 27 Collectively, this work supports that the adaptive immune system is part of the metabolic organ system and that disruptions in its function can result in insulin resistance; however, an absence of B and T cells does not protect from diet-induced insulin resistance. These findings emphasise that results collected in immune-manipulated and diet-manipulated mice must be interpreted carefully as they can be due to a disruption in physiological glucose homeostasis, pathological insulin resistance or a combination of these two events. Contributors: LLB performed all experiments, data and statistical analysis and wrote the manuscript. CHC contributed to the experimental design, interpretation and revised the manuscript. Funding: This work was funded in part by a generous gift from the Chamber's Family Foundation, and supported, in part, by a grant to Stanford by the March of Dimes, the Child Health Research Institute at Stanford, and by the National Institutes of Health (R24DK096465-01). Competing interests: None declared. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: No additional data are available.

tients receiving hemodialysis, a ratio-of-ratios analysis of influenza VE found no evidence of any protection against influenza-like-illness, influenza/pneumonia hospitalization, or mortality.20 Influenza VE at earlier stages of CKD is unclear, and still less is known about pneumococcal VE among patients with CKD.19 21 We aimed to describe the extent to which the burden of community-acquired LRTI and pneumonia among older people with diabetes may be preventable with pneumococcal and influenza vaccination, and whether this varied according to CKD status. We conducted a retrospective cohort study using linked primary and secondary care electronic health record data to calculate the VE of pneumococcal vaccine against all community-acquired pneumonia. Since influenza vaccine may potentially reduce the incidence of influenza infection and secondary pneumonia, we calculated the influenza VE to prevent all community-acquired LRTI (considered as a broad category of all ‘chest infections’, including influenza infections, and possible secondary infections such as bronchitis and pneumonia), using a ratio-of-ratios analysis to address confounding by indication.

iabetes (from 41% to 55%, p=0.04). Among ED patients with a secondary diagnosis of diabetes, the proportion of Medicare patients also increased (from 56% to 66%, p<0.01). In addition, there was an increase in the proportion of elderly patients presenting with a secondary diagnosis of diabetes (from 50% to 60%, p<0.01). We did not find evidence of any statistically significant changes in the proportion of female, non-Hispanic black, Hispanic, privately insured, or uninsured patients presenting with a primary or secondary diagnosis of diabetes. However, among ED patients with a secondary diagnosis of diabetes, there was a statistically significant decrease in the proportion of Medicaid patients (from 26% to 18%, p=0.02) (figure 1). Figure 1 Daily changes in ED use compared to baseline utilization in October and November 2012 for patients with diabetes in New York City evacuation zone level 1. ED, emergency department. Changes in daily ED visits for a primary versus secondary diagnosis of diabetes in New York City evacuation zone level 1. The highest numbers of ED visits occur on days 4 and 5 postlandfall for a primary diagnosis of diabetes and on day 2 postlandfall for a secondary diagnosis of diabetes. ED, emergency department.

nt. Changes in daily ED visits for a primary versus secondary diagnosis of diabetes in New York City evacuation zone level 1. The highest numbers of ED visits occur on days 4 and 5 postlandfall for a primary diagnosis of diabetes and on day 2 postlandfall for a secondary diagnosis of diabetes. ED, emergency department. Geographic distribution In our evaluation of ED patients from all regions of New York City, we identified ZIP codes with a statistically significant change in the number of ED visits by diabetic adults comparing the week after the landfall of Hurricane Sandy to the average weekly baseline prior to disaster in 2012 (figures 2 and 3). There were significant increases in ED use for a primary diagnosis of diabetes, especially, in flood-prone areas. These regions included the Rockaways (11693, 11694), Howard Beach (11414), Coney Island (11224), eastern Staten Island (10306), and parts of Brooklyn (11201) and the Bronx (10454, 10461). For a secondary diagnosis of diabetes, there was similarly a statistically significant increase in ED use in the Rockaways (11691, 11694) and Coney Island (11224), but also some increases in flood-prone areas of Manhattan (10002, 10011). Figure 2 Geographic areas with increased ED visits after Hurricane Sandy by diabetic adults. Significant changes in ED visits among patients with a primary diagnosis of diabetes. Compares the week after Hurricane Sandy's landfall to baseline weekly data by New York City ZIP Codes in 2012. Flooded areas based on the FEMA Modeling Task Force Hurricane Sandy Impact Analysis. ED, emergency department.

Key messages Elderly diabetic adults are at high risk for requiring postdisaster emergency care. There is need to support diabetic patients in the first week after a disaster. Specific medical issues such as myocardial infarction increase among diabetic adults in the days after a disaster. Introduction On 29 October 2012, Hurricane Sandy made landfall on the East Coast of the USA, causing severe storm surges and immense destruction.1 In New York City alone, 305 000 homes were destroyed, and the city sustained 19 billion dollars in damage including to critical health facilities and infrastructure.2 There were 159 deaths; of which, 72 were directly attributed to the storm and 87 were indirectly due to the storm's impact, ie, extended power outages which led to hypothermia, falls in the dark by the elderly, or carbon monoxide poisoning from improperly used generators and cooking devices.3 Although the acute impact on mortality and cost of destruction was well estimated for Hurricane Sandy, less is known about the acute impact this disaster had on patients with chronic diseases such as diabetes.3 4 Patients with diabetes are a particularly vulnerable population during a disaster. Disrupted access to care, inability to monitor glucose, poor nutrition, limited physical activity, damaged or lost medications, inability to safely store insulin due to power loss, and lack of prescription refills all place diabetic adults at increased risk for developing acute postdisaster medical needs, and potential morbidity or mortality in the days following a disaster.5 6

nutrition, limited physical activity, damaged or lost medications, inability to safely store insulin due to power loss, and lack of prescription refills all place diabetic adults at increased risk for developing acute postdisaster medical needs, and potential morbidity or mortality in the days following a disaster.5 6 There is a paucity of information on the acute impact of disasters on diabetic patients, specifically following Hurricane Sandy. Longer term studies in the literature from other disasters have demonstrated poor glycemic control as measured by HbA1c months later and poorer quality of life both thought to be attributable to the increased physical and emotional stress experienced by diabetic patients.7 8 However, in our recent study, screening for conditions that demonstrated increased postdisaster use of emergency care, we found that there was also a statistically significant increase in the number of New York City emergency department (ED) patients presenting with a primary or secondary diagnosis of diabetes even in the first week after Hurricane Sandy.9 These findings require additional investigation as to which characteristics make certain diabetic patients at higher risk for needing acute medical care after a disaster.

k City emergency department (ED) patients presenting with a primary or secondary diagnosis of diabetes even in the first week after Hurricane Sandy.9 These findings require additional investigation as to which characteristics make certain diabetic patients at higher risk for needing acute medical care after a disaster. The goal of this study was to identify salient patient characteristics that increase the likelihood that a diabetic patient will develop acute medical needs after a disaster such as Hurricane Sandy. We analyzed data on patients over the age of 18 years focusing on those who lived in the highest risk evacuation zone in New York City. We investigated patient demographics, insurance status, and medical comorbidities that were associated with increased ED use among these diabetic adults. This study provides insight into how to support diabetic populations in future disasters and to develop a framework for targeted interventions and deployment of disaster response resources.

demographics, insurance status, and medical comorbidities that were associated with increased ED use among these diabetic adults. This study provides insight into how to support diabetic populations in future disasters and to develop a framework for targeted interventions and deployment of disaster response resources. Research design and methods Study design Using an all-payer claims database of ED visits in New York City for 2012, we compared the demographic characteristics, insurance status, geographic distribution, and health conditions of ED patients with a primary versus secondary diagnosis of diabetes before and after landfall of Hurricane Sandy. To evaluate post-disaster ED utilization by diabetic adults that occurred in geographically vulnerable regions, we compared ED use for the first week after the disaster to baseline ED use in 2012 prior to 29 October the day on which Hurricane Sandy struck the East Coast of the USA. Data source The New York State Department of Health collects claims data from hospitals on ED visits and inpatient hospitalizations. The Statewide Planning and Research Cooperative System (SPARCS) is the most comprehensive resource for ED utilization in New York State.10 It includes privately insured, Medicare, Medicaid, and uninsured patients. In addition, to patient demographic data and insurance status, SPARCS includes diagnosis codes and patient addresses that can be used to identify health conditions and the exact location of a patient's residence.

utilization in New York State.10 It includes privately insured, Medicare, Medicaid, and uninsured patients. In addition, to patient demographic data and insurance status, SPARCS includes diagnosis codes and patient addresses that can be used to identify health conditions and the exact location of a patient's residence. Study population Our study included all adult patients aged 18 years and older who visited a New York City ED in 2012, had a home address located in New York City, and had either a primary or secondary diagnosis of diabetes. We excluded patients from correctional facilities and nursing homes, and patients who visited an ED associated with a specialty hospital (ie, surgical subspecialty, oncological, or Veterans Administration facilities). These exclusions were performed so that we studied non-institutionalized New York City adults with diabetes who visited a 911-receiving ED based at a general acute care hospital in New York City.

an ED associated with a specialty hospital (ie, surgical subspecialty, oncological, or Veterans Administration facilities). These exclusions were performed so that we studied non-institutionalized New York City adults with diabetes who visited a 911-receiving ED based at a general acute care hospital in New York City. ICD-9 diagnosis codes To identify ED patients with a primary or secondary diagnosis of diabetes, we scanned International Classification of Diseases, 9th Edition (ICD-9) codes for a prefix of 250 (ie, 250.00–250.93), which includes diabetes diagnoses with and without complications. In a study of patients at the University of Pittsburgh Medical Center, the presence of a diabetes diagnosis code in ED records was 95% sensitive and 99% specific for identifying diabetic adults among ED patients.11 In fact, ED data had higher accuracy than outpatient records (79% sensitive, 98% specific), inpatient records (84% sensitive, 97% specific), and identification methods based on scanning pharmacy records for diabetes medications and laboratory tests for elevated HgbA1c or blood glucose concentrations.11

g ED patients.11 In fact, ED data had higher accuracy than outpatient records (79% sensitive, 98% specific), inpatient records (84% sensitive, 97% specific), and identification methods based on scanning pharmacy records for diabetes medications and laboratory tests for elevated HgbA1c or blood glucose concentrations.11 In addition, to identify health conditions associated with ED visits for patients with a primary or secondary diagnosis of diabetes, we also evaluated the frequency of other ICD-9 codes based on the first three-digit and/or letter prefix. For patients who had an ED visit with a primary diagnosis of diabetes, we evaluated the other secondary diagnosis codes to evaluate other conditions and comorbidities that may have contributed to increases in ED use after the storm. For patients who had a secondary diagnosis of diabetes, we evaluated the primary diagnosis code associated with the ED visit to evaluate the presenting cause of ED visits among patients who have a history of diabetes, but came to the ED for other conditions.

may have contributed to increases in ED use after the storm. For patients who had a secondary diagnosis of diabetes, we evaluated the primary diagnosis code associated with the ED visit to evaluate the presenting cause of ED visits among patients who have a history of diabetes, but came to the ED for other conditions. Evacuation zones In our prior study of ED utilization patterns, we had found that adults with a history of diabetes presented in significantly increased numbers in the level 1 evacuation zone. In this study, we focused on this subset of diabetic adults by studying diabetic patients located in the level 1 evacuation zone (at the highest risk).9 These revised evacuation zones were developed by the New York City Office of Emergency Management in response to Hurricane Sandy's impact and were generally areas of the city most impacted by the storm. To identify these patients, we geocoded the addresses of diabetic adults in New York City and analyzed the sample of patients whose address was located in the level 1 evacuation zone based on a publicly available geographic shapefile called ‘Atomic Polygons’ (Release 15B, April 2015), which is available from the New York City Department of City Planning.12

d the addresses of diabetic adults in New York City and analyzed the sample of patients whose address was located in the level 1 evacuation zone based on a publicly available geographic shapefile called ‘Atomic Polygons’ (Release 15B, April 2015), which is available from the New York City Department of City Planning.12 Statistical analyses We first evaluated the geographic distribution of ED patients from any part of New York City who presented with a primary or secondary diagnosis of diabetes by comparing the number of ED visits for the week after the storm compared to weekly baseline ED use in 2012 prior to landfall of Hurricane Sandy.13 We analyzed this geographic distribution by postal ZIP code to identify areas of significant changes in ED use. We calculated a Z-score for the number of ED visits for the week after the storm's landfall using an average and SD from the pre-disaster weeks in 2012. Then, we identified areas with significant changes in ED use by mapping ZIP codes with a Z-score of 1.645, 1.96, and 2.545, which we noted as a 90%, 95%, and 99% confidence of an increase or decrease in postdisaster ED use.

the week after the storm's landfall using an average and SD from the pre-disaster weeks in 2012. Then, we identified areas with significant changes in ED use by mapping ZIP codes with a Z-score of 1.645, 1.96, and 2.545, which we noted as a 90%, 95%, and 99% confidence of an increase or decrease in postdisaster ED use. We then examined the demographic characteristics and insurance status of ED patients with a primary or secondary diagnosis of diabetes focusing on patients living in the level 1 evacuation zone. To compare these characteristics for the week after the storm's landfall to baseline utilization patterns, we analyzed the average weekly proportion of elderly (aged 65 years and older), female, non-Hispanic black, and Hispanic patients, in addition to the proportion of privately insured, Medicare, Medicaid, and uninsured patients in 2012 prior to Hurricane Sandy's landfall. For these weeks in 2012 prior to the disaster, we computed an average and SD and then calculated Z-scores for patient characteristics the week after the storm to determine statistically significant changes using a p value of <0.05.

Medicare, Medicaid, and uninsured patients in 2012 prior to Hurricane Sandy's landfall. For these weeks in 2012 prior to the disaster, we computed an average and SD and then calculated Z-scores for patient characteristics the week after the storm to determine statistically significant changes using a p value of <0.05. Finally, to analyze the other primary and secondary diagnoses, we determined the most common primary diagnoses among ED patients with a secondary diagnosis of diabetes and also the most common secondary diagnoses among ED patients with a primary diagnosis of diabetes. This analysis also focused on patients from the level 1 evacuation zone. We compared the top 10 of these primary and secondary diagnoses before and after the disaster, and we also evaluated which of these diagnoses had the highest absolute increases in the number of ED patients in the week following Hurricane Sandy's landfall compared to the pre-disaster weekly average. To be considered, the increase in ED visits for a given primary or secondary diagnosis had to be statistically significant, meaning that the Z-score for the number of ED visits for the week after storm's landfall had to be above 2.81, correlating to a p value of at least 0.005 based on a Bonferroni adjustment for comparisons among the top 10 diagnoses. Statistical analyses were performed using Stata V.12.1 (StataCorp: College Station, Texas, USA, 2011). Geographic analysis was performed using ArcGIS Desktop 10.2 (ESRI: Redlands, California, USA, 2013).

Finally, to analyze the other primary and secondary diagnoses, we determined the most common primary diagnoses among ED patients with a secondary diagnosis of diabetes and also the most common secondary diagnoses among ED patients with a primary diagnosis of diabetes. This analysis also focused on patients from the level 1 evacuation zone. We compared the top 10 of these primary and secondary diagnoses before and after the disaster, and we also evaluated which of these diagnoses had the highest absolute increases in the number of ED patients in the week following Hurricane Sandy's landfall compared to the pre-disaster weekly average. To be considered, the increase in ED visits for a given primary or secondary diagnosis had to be statistically significant, meaning that the Z-score for the number of ED visits for the week after storm's landfall had to be above 2.81, correlating to a p value of at least 0.005 based on a Bonferroni adjustment for comparisons among the top 10 diagnoses. Statistical analyses were performed using Stata V.12.1 (StataCorp: College Station, Texas, USA, 2011). Geographic analysis was performed using ArcGIS Desktop 10.2 (ESRI: Redlands, California, USA, 2013). Results Study population In evacuation zone level 1 in New York City, there were an average of 30.9 ED visits with a primary diagnosis of diabetes and 258.8 ED visits with a secondary diagnosis of diabetes during the weeks in 2012 before Hurricane Sandy's landfall (table 1). For the week after the landfall, there were statistically significant increases with 60 ED visits for a primary diagnosis of diabetes and 420 ED visits for a secondary diagnosis of diabetes in the level 1 evacuation zone (p<0.01).

agnosis of diabetes during the weeks in 2012 before Hurricane Sandy's landfall (table 1). For the week after the landfall, there were statistically significant increases with 60 ED visits for a primary diagnosis of diabetes and 420 ED visits for a secondary diagnosis of diabetes in the level 1 evacuation zone (p<0.01). Table 1 Characteristics of ED users with diabetes in New York City evacuation zone level 1 before and 1 week after Hurricane Sandy's landfall in 2012 Patient characteristics Before Hurricane Sandy landfall After Hurricane Sandy landfall p Value for difference Weekly average 2012 Proportion of patients (%) One week after the landfall Proportion of patients (%) Primary diagnosis of diabetes Total 30.9 100 60 100 <0.01* Demographics Elderly (aged 65 years and older) 9.8 32 26 43 0.16 Female 12.8 41 36 60 0.05 Black 12.7 41 4 27 0.09 Hispanic 6.1 20 13 22 0.76 Insurance Private 3.6 12 4 7 0.42 Medicare 12.7 41 33 55 0.04* Medicaid 10.7 35 17 28 0.48 Self-Pay 3.9 12 6 10 0.69 Secondary diagnosis of diabetes Total 258.8 100 420 100 <0.01* Demographics Elderly (aged 65 years and older) 128.4 50 250 60 <0.01* Female 143.5 55 238 57 0.69 Black 73.9 29 101 24 0.19 Hispanic 50.9 20 81 19 0.92 Insurance Private 32.0 12 37 9 0.11 Medicare 145.9 56 279 66 <0.01* Medicaid 67.8 26 75 18 0.02* Self-Pay 13.1 5 29 7 0.16 ED, emergency department. * = p-value < 0.05

Patient characteristics Before Hurricane Sandy landfall After Hurricane Sandy landfall p Value for difference Weekly average 2012 Proportion of patients (%) One week after the landfall Proportion of patients (%) Primary diagnosis of diabetes Total 30.9 100 60 100 <0.01* Demographics Elderly (aged 65 years and older) 9.8 32 26 43 0.16 Female 12.8 41 36 60 0.05 Black 12.7 41 4 27 0.09 Hispanic 6.1 20 13 22 0.76 Insurance Private 3.6 12 4 7 0.42 Medicare 12.7 41 33 55 0.04* Medicaid 10.7 35 17 28 0.48 Self-Pay 3.9 12 6 10 0.69 Secondary diagnosis of diabetes Total 258.8 100 420 100 <0.01* Demographics Elderly (aged 65 years and older) 128.4 50 250 60 <0.01* Female 143.5 55 238 57 0.69 Black 73.9 29 101 24 0.19 Hispanic 50.9 20 81 19 0.92 Insurance Private 32.0 12 37 9 0.11 Medicare 145.9 56 279 66 <0.01* Medicaid 67.8 26 75 18 0.02* Self-Pay 13.1 5 29 7 0.16 ED, emergency department. * = p-value < 0.05 Comparing the proportion of these patients to the average weekly baseline in New York City in 2012, there was a statistically significant increase in the proportion of Medicare patients presenting to an ED with a primary diagnosis of diabetes (from 41% to 55%, p=0.04). Among ED patients with a secondary diagnosis of diabetes, the proportion of Medicare patients also increased (from 56% to 66%, p<0.01). In addition, there was an increase in the proportion of elderly patients presenting with a secondary diagnosis of diabetes (from 50% to 60%, p<0.01).

etic adults. Significant changes in ED visits among patients with a primary diagnosis of diabetes. Compares the week after Hurricane Sandy's landfall to baseline weekly data by New York City ZIP Codes in 2012. Flooded areas based on the FEMA Modeling Task Force Hurricane Sandy Impact Analysis. ED, emergency department. Figure 3 Significant changes in ED visits among patients with a secondary diagnosis of diabetes. ED, emergency department. Primary diagnoses Among patients in the level 1 evacuation zone, the top 10 primary diagnoses among ED patients with a secondary diagnosis of diabetes for the week after Hurricane Sandy's landfall and the pre-Hurricane Sandy average weekly baseline in 2012 are shown in table 2. Comparatively, the lists are relatively similar with over half of the diagnoses appearing before and after Hurricane Sandy's landfall. Notable exceptions included a statistically significant increase in the number of myocardial infarctions (an increase from 4 to 12 cases per week in evacuation zone level 1), hypertension (an increase from 4 to 15 cases per week), chronic bronchitis (an increase from 4 to 12 cases per week), and hypertensive kidney disease (an increase from 2 to 10 cases per week). In addition, we also found statistically significant increases in the number of ED visits with a primary diagnosis for chronic kidney disease, prescription refills, drug dependence, dialysis dependence, and electrolyte disorders (p<0.005).

eek), and hypertensive kidney disease (an increase from 2 to 10 cases per week). In addition, we also found statistically significant increases in the number of ED visits with a primary diagnosis for chronic kidney disease, prescription refills, drug dependence, dialysis dependence, and electrolyte disorders (p<0.005). Table 2 Most common primary diagnoses of patients with a secondary diagnosis of diabetes who presented for emergency care before and 1 week after Hurricane Sandy's landfall in 2012 Weekly baseline in 2012 before Hurricane Sandy's landfall One week after Hurricane Sandy's landfall Highest increases before and after Respiratory symptoms (16) General symptoms (25) General symptoms (+13) General symptoms (12) Respiratory symptoms (19) Hypertension (+11) Heart failure (10) Hypertension (15) Myocardial infarction (+8) Sepsis (7) Heart failure (13) Hypertensive kidney disease (+8) Cellulitis or abscess (7) Myocardial infarction (12) Chronic bronchitis (+8) Urinary tract infection (7) Chronic bronchitis (12) Chronic kidney disease (+7) Asthma (7) Asthma (12) Prescription refills (+7) Abdominal/pelvic symptoms (6) Hypertensive kidney disease (10) Drug dependence (+7) Cardiac dysrhythmias (5) Sepsis (9) Dialysis dependence (+6) Ischemic heart disease (5) Ischemic heart disease (9) Electrolyte disorder (+6) Diagnoses in bold highlight the differences in categories before and after Hurricane Sandy's landfall.

al/pelvic symptoms (6) Hypertensive kidney disease (10) Drug dependence (+7) Cardiac dysrhythmias (5) Sepsis (9) Dialysis dependence (+6) Ischemic heart disease (5) Ischemic heart disease (9) Electrolyte disorder (+6) Diagnoses in bold highlight the differences in categories before and after Hurricane Sandy's landfall. Secondary diagnoses We also analyzed the top 10 secondary diagnoses among ED patients in the highest level evacuation zone with a primary diagnosis of diabetes to assess associated health conditions and comorbidities (see table 3). Comparing the week after Hurricane Sandy's landfall to the predisaster average weekly baseline for 2012, a majority of the secondary diagnoses overlap before and after the storm. Notable exceptions included a statistically significant increase in the number of ED patients presenting primarily with diabetes who also had a secondary diagnosis code for a postprocedural state (an increase from 3 to 10 cases per week in evacuation zone level 1) and a secondary diagnosis code for overweight or obesity (an increase from 2 to 7 cases per week). In addition, we also found statistically significant increases in the number of ED visits with a secondary diagnosis for cardiac dysrhythmia, chronic airway obstruction, and osteomyelitis. Table 3 Most common secondary diagnoses of patients presenting for EDs for a primary diagnosis of diabetes before and 1 week after Hurricane Sandy's landfall in 2012

Secondary diagnoses We also analyzed the top 10 secondary diagnoses among ED patients in the highest level evacuation zone with a primary diagnosis of diabetes to assess associated health conditions and comorbidities (see table 3). Comparing the week after Hurricane Sandy's landfall to the predisaster average weekly baseline for 2012, a majority of the secondary diagnoses overlap before and after the storm. Notable exceptions included a statistically significant increase in the number of ED patients presenting primarily with diabetes who also had a secondary diagnosis code for a postprocedural state (an increase from 3 to 10 cases per week in evacuation zone level 1) and a secondary diagnosis code for overweight or obesity (an increase from 2 to 7 cases per week). In addition, we also found statistically significant increases in the number of ED visits with a secondary diagnosis for cardiac dysrhythmia, chronic airway obstruction, and osteomyelitis. Table 3 Most common secondary diagnoses of patients presenting for EDs for a primary diagnosis of diabetes before and 1 week after Hurricane Sandy's landfall in 2012 Weekly baseline in 2012 before Hurricane Sandy's landfall One week after Hurricane Sandy's landfall Highest increases before and after Hypertension (13) Hypertension (30) Hypertension (+17) Hyperlipidemia (7) Post-procedural aftercare (17) Post-procedural aftercare (+11) Post-procedural aftercare (6) Chronic skin ulcer (14) Chronic skin ulcer (+9) Electrolyte disorder (6) Electrolyte disorder (13) Electrolyte disorder (+7) Ischemic heart disease (5) Hyperlipidemia (11) Post-procedural state (+7) Chronic skin ulcer (5) Post-procedural state (10) Overweight or obese (+5) Chronic kidney disease (4) Ischemic heart disease (10) Ischemic heart disease (+5) Hypertensive kidney disease (4) Drug abuse (8) Cardiac dysrhythmia (+4) Drug abuse (4) Chronic kidney disease (7) Chronic airway obstruction (+4) Other health hazards (7) Overweight or obese (7) Osteomyelitis (+4) Diagnoses in bold highlight the differences in categories before and after Hurricane Sandy's landfall.

+5) Hypertensive kidney disease (4) Drug abuse (8) Cardiac dysrhythmia (+4) Drug abuse (4) Chronic kidney disease (7) Chronic airway obstruction (+4) Other health hazards (7) Overweight or obese (7) Osteomyelitis (+4) Diagnoses in bold highlight the differences in categories before and after Hurricane Sandy's landfall. ED, emergency department. Conclusions Prior literature has documented the longer term impact of disasters on diabetic adults. Hurricane Sandy was the second most destructive hurricane in US history, only surpassed by Hurricane Katrina in 2005.14 After Hurricane Katrina, there was evidence of long-term effects of the disaster on diabetic patients when glycemic control was measured by HbA1c and quality of life was assessed.7 8 15 16 In studies of international disasters, there has similar been evidence of disruptions in glycemic control and worse quality of life among individuals affected by events like earthquakes in Japan, Italy, and Turkey.17–22 The presumed mechanism of impact is thought to be mediated by the stress caused by the disaster, which has also been seen in other non-natural disasters.23–25 Our study expands this body of literature by specifically analyzing the acute impact of disasters on diabetic patients by evaluating ED use by diabetic adults after Hurricane Sandy.

sm of impact is thought to be mediated by the stress caused by the disaster, which has also been seen in other non-natural disasters.23–25 Our study expands this body of literature by specifically analyzing the acute impact of disasters on diabetic patients by evaluating ED use by diabetic adults after Hurricane Sandy. In our geographic analysis, we found that diabetic adults in evacuation zone level 1 were more likely to have increased the need for post-disaster emergency care. We also found that elderly diabetic adults aged 65 years and older and Medicare patients are particularly at risk for requiring postdisaster emergency care compared to other vulnerable diabetic populations. Finally, through our analysis of diagnosis codes, we found that certain coexisting medical conditions and comorbidities place diabetic patients at higher risk for requiring acute medical care in the week immediately following a disaster such as Hurricane Sandy.

cy care compared to other vulnerable diabetic populations. Finally, through our analysis of diagnosis codes, we found that certain coexisting medical conditions and comorbidities place diabetic patients at higher risk for requiring acute medical care in the week immediately following a disaster such as Hurricane Sandy. There are a unique constellation of factors that combine to increase vulnerability among elderly patients to the negative effects of catastrophic events.26 27 Many elderly patients depend on complex medication regimens or medical equipment.28 Some are dependent on caregivers to help manage their lives, and with disaster-related disruptions, these caregivers may be unable to care for this population.29 Complicating these risk factors, these patients may not be adequately prepared for disasters as at least one study shows that only 31% of elderly had a disaster plan in place.30 Finally, the intersection between multiple chronic diseases and advanced age can push those at the margins of compensated chronic disease into decompensated states.31

factors, these patients may not be adequately prepared for disasters as at least one study shows that only 31% of elderly had a disaster plan in place.30 Finally, the intersection between multiple chronic diseases and advanced age can push those at the margins of compensated chronic disease into decompensated states.31 We analyzed the coexisting medical conditions and comorbidities to identify those that were associated with increased ED use the week after Hurricane Sandy's landfall compared to the pre-disaster average weekly baseline in 2012 among diabetic patients from the highest risk evacuation zone.9 For patients presenting to the ED for a primary diagnosis of diabetes, a majority of the most common secondary diagnoses were no different from those conditions noted before the storm. However, there were a substantial number of patients with a secondary diagnosis of postprocedural status. This finding suggests that diabetic patients who recently had a significant surgical or medical procedure experienced difficulty in accessing appropriate follow-up aftercare, and better planning is required for these patients in the face of a pending disaster.32 33 In a post hoc analysis, we did not find that there was a statistically significant increase in the number of emergency visits for a primary diagnosis of diabetic ketoacidosis or hyperglycemic hyperosmolar syndrome. However, this subanalysis was limited by a small number of events in the highest risk evacuation zone.

33 In a post hoc analysis, we did not find that there was a statistically significant increase in the number of emergency visits for a primary diagnosis of diabetic ketoacidosis or hyperglycemic hyperosmolar syndrome. However, this subanalysis was limited by a small number of events in the highest risk evacuation zone. In addition, when analyzing patients with a secondary diagnosis of diabetes who presented in increased numbers for care in the week following Hurricane Sandy's landfall, we found that there was a substantial increase in the number of patients who had a primary diagnosis of myocardial infarction, hypertension, respiratory conditions, kidney disease, dialysis dependence, and prescription refills. In prior studies, patients with respiratory conditions and kidney conditions were known to be at high risk during disasters as some are ventilator dependent or dialysis dependent.34–36 Following Hurricane Sandy's landfall, electrical outages and damage to infrastructure led to the loss of the ability to use durable equipment at home and the closure of dialysis centers.37–39

nd kidney conditions were known to be at high risk during disasters as some are ventilator dependent or dialysis dependent.34–36 Following Hurricane Sandy's landfall, electrical outages and damage to infrastructure led to the loss of the ability to use durable equipment at home and the closure of dialysis centers.37–39 In addition, the closure of outpatient clinics and hospitals with their associated medical care centers might further limit the access to regular medical care.40 An intervention to address the need for unfilled prescriptions may come in the form of mobile pharmacies that can deliver diabetic medications to patients in need.41 An important window of opportunity may exist as patients with a primary diagnosis of diabetes presented somewhat delayed on days 4 and 5 following Hurricane Sandy's landfall compared to other patients, for example, ventilator-dependent or dialysis-dependent individuals, who came to the ED for care immediately after the storm.9 42

t window of opportunity may exist as patients with a primary diagnosis of diabetes presented somewhat delayed on days 4 and 5 following Hurricane Sandy's landfall compared to other patients, for example, ventilator-dependent or dialysis-dependent individuals, who came to the ED for care immediately after the storm.9 42 Prior disaster literature has documented the increased incidence of acute myocardial infarction in the general population after disasters such as Hurricane Katrina, and the Hanshin-Awaji and Niigata-Chusetsu earthquakes.43–45 The proposed mechanism of this increase in acute myocardial infarction is thought to be related to the mental stress and physical activity induced by disaster-related damage to property, delays in recovery and reconstruction, loss of livelihood, all coupled with disrupted access to healthcare and medications leading to an increased emotional distress, blood pressure, and catecholamine release.44 46 Our data, although based on small numbers of events in the level 1 evacuation zone, demonstrate three times the number of cases of myocardial infarction among diabetic adults, which suggests the need to optimize cardiovascular health and to provide social support for diabetic patients highly affected by a disaster like Hurricane Sandy.47

based on small numbers of events in the level 1 evacuation zone, demonstrate three times the number of cases of myocardial infarction among diabetic adults, which suggests the need to optimize cardiovascular health and to provide social support for diabetic patients highly affected by a disaster like Hurricane Sandy.47 Finally, in our geographic analysis of patients from all of New York City, we found that statistically significant increases in ED use primarily were concentrated in flood-prone areas. These hot spots may be areas in which interventions can be focused and geographically targeted.48 Furthermore, finding statistically significant clusters of diabetic patients in disaster-prone regions using geospatial analysis may help in preplanning to identify areas in which interventions may be targeted to provide home care services for patients with diabetes who had recent surgical or medical interventions, or mobile pharmacies to deliver diabetic medications, or other interventions to optimize cardiovascular health of diabetic patients at higher risk for myocardial infarction.49 In fact, our prior studies demonstrate that geographic analysis of patients with chronic disease and ED use can help enhance disaster planning and response to optimize care for particularly vulnerable populations of patients.

ptimize cardiovascular health of diabetic patients at higher risk for myocardial infarction.49 In fact, our prior studies demonstrate that geographic analysis of patients with chronic disease and ED use can help enhance disaster planning and response to optimize care for particularly vulnerable populations of patients. Limitations Our study is a geographic analysis of ED administrative claims data, which means that it is subject to coding errors that can occur in data collection. In addition, we used municipal administrative evacuation zones developed post hoc based on the damage incurred by Hurricane Sandy. These areas may not precisely represent the true underlying geographic impact of Hurricane Sandy, particularly for electrical failures. Given the number of ICD-9 diagnosis code categories that we examined, we made adjustments to the level of statistical significance required to identify a category with increased ED utilization. Important categories of ED utilization may have been missed by using this stringent requirement for statistical significance. Finally, our study is limited to Hurricane Sandy and New York City, a unique and densely populated urban environment. Findings of our study may not be generalizable to other regions of the country or other types of disasters in which different changes in ED utilization may occur.

quirement for statistical significance. Finally, our study is limited to Hurricane Sandy and New York City, a unique and densely populated urban environment. Findings of our study may not be generalizable to other regions of the country or other types of disasters in which different changes in ED utilization may occur. Conclusions Our findings suggest that there is need to support diabetic adults during disasters by ensuring access to medications, aftercare for patients following a recent procedure, and optimize their cardiovascular health to reduce the risk of heart attacks. This study fills the void on a number of questions: specifically what conditions cause diabetic patients to present for emergency care after a disaster, and which patient characteristics among diabetic adults contribute to a higher risk of requiring acute medical needs after a disaster. In addition, this study demonstrates that geographic analysis can help identify diabetic patients who are most vulnerable after a disaster, which may help targeting interventions to improve disaster preparedness or response. The implications of these results will help shape future disaster management and policies, especially for diabetic patients, who we find are particularly at risk during events such as Hurricane Sandy.

most vulnerable after a disaster, which may help targeting interventions to improve disaster preparedness or response. The implications of these results will help shape future disaster management and policies, especially for diabetic patients, who we find are particularly at risk during events such as Hurricane Sandy. Contributors: DCL, SWS, and LRG are responsible for study conception and design. DCL involved in acquisition of the data. DCL, VKG, BGC, SM, BF, SPW, SWS, and LRG participated in analysis and interpretation of the data. DCL and VKG involved in drafting of the manuscript. BGC, SM, BF, SPW, SWS, and LRG carried out critical revision of the manuscript for intellectual content. DCL, SWS, and LRG involved in obtaining funding. SWS provided administrative, technical, and material support. LRG involved in supervision. DCL takes full responsibility for the work as a whole, including the study design, access to data, and the decision to submit and publish the manuscript. Funding: This work was funded by the US Department of Health and Human Services, Office of the Assistant Secretary for Preparedness and Response (ASPR), award number HITEP-150030-01-00 to the NYU School of Medicine.

Contributors: DCL, SWS, and LRG are responsible for study conception and design. DCL involved in acquisition of the data. DCL, VKG, BGC, SM, BF, SPW, SWS, and LRG participated in analysis and interpretation of the data. DCL and VKG involved in drafting of the manuscript. BGC, SM, BF, SPW, SWS, and LRG carried out critical revision of the manuscript for intellectual content. DCL, SWS, and LRG involved in obtaining funding. SWS provided administrative, technical, and material support. LRG involved in supervision. DCL takes full responsibility for the work as a whole, including the study design, access to data, and the decision to submit and publish the manuscript. Funding: This work was funded by the US Department of Health and Human Services, Office of the Assistant Secretary for Preparedness and Response (ASPR), award number HITEP-150030-01-00 to the NYU School of Medicine. Disclaimer: The funding agency (ASPR) played no role in the design or conduct of the study; collection, management, analysis, or interpretations of the data; preparation of the manuscript; or decision to publish. BGC spends a portion of his time as the Director of the Emergency Care Coordination Center in the US Department of Health and Human Services. The content of this article is the responsibility of the authors and does not necessarily represent the views of the US government, US Department of Health and Human Services, ASPR, NYU School of Medicine, or Sidney Kimmel Medical College. Competing interests: None declared.

Disclaimer: The funding agency (ASPR) played no role in the design or conduct of the study; collection, management, analysis, or interpretations of the data; preparation of the manuscript; or decision to publish. BGC spends a portion of his time as the Director of the Emergency Care Coordination Center in the US Department of Health and Human Services. The content of this article is the responsibility of the authors and does not necessarily represent the views of the US government, US Department of Health and Human Services, ASPR, NYU School of Medicine, or Sidney Kimmel Medical College. Competing interests: None declared. Ethics approval: This study was approved by the Institutional Review Board at the NYU School of Medicine and the SPARCS Data Protection Review Board at the New York State Department of Health. Provenance and peer review: Not commissioned; externally peer reviewed.

Key messages It has previously been shown in the non pregnant state urinary C peptide creatinine ratio (UCPCR) obtained from a spot urine sample correlates with serum C peptide concentration and is a validated method to assess residual β-cell function. The current work shows in pregnant glucose tolerant women at 28 weeks gestation UCPCR correlates with serum C peptide at 0 and 120 minutes during a 75 g OGTT UCPCR is detectable in pregnant women with over 9 years of type 1 diabetes The UCPCR measurement in pregnancy provides a practical method for assessing insulin secretion in pregnancy in women with and without diabetes Introduction Autopsy studies have suggested β-cell proliferation and neogenesis in human pregnancies, possibly due to placental factors.1–3 Three studies involving a total of 55 women with type 1 diabetes measured serum or plasma C peptide in pregnancy, showing 49 women to have detectable C peptide values.4–6 The ratio of urinary C peptide to the urinary creatinine obtained from a spot urine sample and expressed as UCPCR is correlated to serum C peptide outside pregnancy and has been used to assess residual β-cell function in women with type 1 diabetes. The current study investigated the use of UCPCR to assess β-cell function in pregnant women with normal glucose tolerance and with type 1 diabetes. Research design and methods This prospective study carried out at Queen Charlotte's Hospital was ethically approved by the Imperial College Healthcare Tissue Bank and the Research Ethics Committee Wales: 12/WA/0196. All women gave informed written consent.

Introduction Autopsy studies have suggested β-cell proliferation and neogenesis in human pregnancies, possibly due to placental factors.1–3 Three studies involving a total of 55 women with type 1 diabetes measured serum or plasma C peptide in pregnancy, showing 49 women to have detectable C peptide values.4–6 The ratio of urinary C peptide to the urinary creatinine obtained from a spot urine sample and expressed as UCPCR is correlated to serum C peptide outside pregnancy and has been used to assess residual β-cell function in women with type 1 diabetes. The current study investigated the use of UCPCR to assess β-cell function in pregnant women with normal glucose tolerance and with type 1 diabetes. Research design and methods This prospective study carried out at Queen Charlotte's Hospital was ethically approved by the Imperial College Healthcare Tissue Bank and the Research Ethics Committee Wales: 12/WA/0196. All women gave informed written consent. One hundred women were recruited prospectively to provide an extra blood and urine sample during a diagnostic 75 g oral glucose tolerance test (OGTT) at 28 weeks of pregnancy for gestational diabetes mellitus (GDM). All women had one or more risk factors for GDM according to the National Institute for Health and Care Excellence (NICE) guidelines,7 or were 35 years old or above. All women were fasted for 8–10 hours and had passed their first void morning urine.

OGTT) at 28 weeks of pregnancy for gestational diabetes mellitus (GDM). All women had one or more risk factors for GDM according to the National Institute for Health and Care Excellence (NICE) guidelines,7 or were 35 years old or above. All women were fasted for 8–10 hours and had passed their first void morning urine. Blood samples were collected at 0 (fasting) and 120 min (post-75 g OGTT) in 6 mL BD plastic Vacutainer Plus, silicone coated tubes, placed on ice prior to separation of serum by centrifugation. The second void urine samples at 0 and the 120 min urine sample were collected in 30 mL polystyrene universal containers with boric acid preservative. Serum and urine samples were transferred to cryotubes and stored at −80°C before analysis. Seven women with previously diagnosed type 1 diabetes were recruited to give a non-fasting spot urine sample in the antenatal clinic on two separate occasions. All seven women gave urine samples that ranged from 10 to 22 weeks apart (five women between the first and third trimester, one between the first and second and one between the second and third trimester). Urine samples were handled as described above.

pot urine sample in the antenatal clinic on two separate occasions. All seven women gave urine samples that ranged from 10 to 22 weeks apart (five women between the first and third trimester, one between the first and second and one between the second and third trimester). Urine samples were handled as described above. Urine and serum C peptide was measured by a two-step chemiluminescent microparticle immunoassay using an Abbott Diagnostics Architect platform, with a total coefficient of variation (CV) <10% and a detection range of 3.33–10 000 pmol/L for undiluted samples. Initially urinary c peptide measurements, including those of the seven women with type 1 diabetes, were analyzed undiluted. Samples exceeding the upper limit of detection were rerun following an automated 1:10 dilution using the validated Abbott protein containing diluent. Samples still exceeding the upper limit of detection were rerun following a manual 1:20 dilution using the manufactures' multiassay manual diluent.8 Creatinine was measured using the kinetic alkaline picrate method with a total CV of ≤6% (Abbott Architect ci16200 system). The estimated glomerular filtration rate (eGFR) was estimated by the Modification of Diet in Renal Disease (MDRD) formula. Statistical analysis was performed using SPSS V.22. Correlations between UCPCROGTT and serum C peptideOGTT at 0 and 120 min for the glucose-tolerant women were performed by Spearman's rank correlation.

Urine and serum C peptide was measured by a two-step chemiluminescent microparticle immunoassay using an Abbott Diagnostics Architect platform, with a total coefficient of variation (CV) <10% and a detection range of 3.33–10 000 pmol/L for undiluted samples. Initially urinary c peptide measurements, including those of the seven women with type 1 diabetes, were analyzed undiluted. Samples exceeding the upper limit of detection were rerun following an automated 1:10 dilution using the validated Abbott protein containing diluent. Samples still exceeding the upper limit of detection were rerun following a manual 1:20 dilution using the manufactures' multiassay manual diluent.8 Creatinine was measured using the kinetic alkaline picrate method with a total CV of ≤6% (Abbott Architect ci16200 system). The estimated glomerular filtration rate (eGFR) was estimated by the Modification of Diet in Renal Disease (MDRD) formula. Statistical analysis was performed using SPSS V.22. Correlations between UCPCROGTT and serum C peptideOGTT at 0 and 120 min for the glucose-tolerant women were performed by Spearman's rank correlation. Results Of the 100 women who had an OGTT, 90 were included for analysis; excluded were 5 women with GDM by NICE criteria,7 2 with non-singleton pregnancies, 2 with a gestational age above 31 weeks and 1 with a renal transplant.

Statistical analysis was performed using SPSS V.22. Correlations between UCPCROGTT and serum C peptideOGTT at 0 and 120 min for the glucose-tolerant women were performed by Spearman's rank correlation. Results Of the 100 women who had an OGTT, 90 were included for analysis; excluded were 5 women with GDM by NICE criteria,7 2 with non-singleton pregnancies, 2 with a gestational age above 31 weeks and 1 with a renal transplant. The undiluted OGTT urinary samples were above the upper limit of the C peptide assay detection in 65 of the fasting and in all 90 of the 120 min samples. Following an automated 1:10 dilution, 17 of the 120 min samples remained above the upper range of assay detection. A 1:20 manual dilution was performed on these 17 samples; however, due to technical difficulties during the subsequent analysis and multiple freeze–thaw cycles, these samples were discarded.9 Therefore, the analysis of 120 min UCPCR data was performed on the remaining 73 samples. The 90 glucose-tolerant women had a median age (range) of 34 (20–49) years, booking body mass index (BMI) of 23.7 (17.96–39.49) kg/m2 and a gestational age of 28 (24–29) weeks. The 0 and 120 min serum C peptideOGTT median (25th–75th range) was 483 (381–599) and 2254 (1759–2781) pmol/L, respectively. The UCPCROGTT at 0 and 120 min median (25th–75th range) were 2796 (1969–3983) and 12 304 (8621–20 733) pmol/mmol, respectively. The UCPCROGTT and serum C peptideOGTT were significantly correlated at 0 and 120 min, rs 0.675, 0.541 (p<0.0001), respectively (figure 1A, B).

1–599) and 2254 (1759–2781) pmol/L, respectively. The UCPCROGTT at 0 and 120 min median (25th–75th range) were 2796 (1969–3983) and 12 304 (8621–20 733) pmol/mmol, respectively. The UCPCROGTT and serum C peptideOGTT were significantly correlated at 0 and 120 min, rs 0.675, 0.541 (p<0.0001), respectively (figure 1A, B). Figure 1 (A) Scatter plot showing the correlation between fasting second void UCPCR and fasting serum C peptide during a 75 g OGTT (rs 0.675, p<0.0001). (B) Scatter plot showing the correlation between 120 min UCPCR and 120 min serum C peptide at 75 g OGTT (rs0.541, p<0.0001). (C) The changes in UCPCR of seven pregnant women with type 1 diabetes between two samples taken 10 weeks or more apart. OGTT, oral glucose tolerance test; UCPCR, urinary C peptide creatinine ratio. The seven women had previously been clinically diagnosed with type 1 diabetes at a median age of 14.4 (9–25) years. At booking, they were 35 (29–40) years old, with 19 (9–31) years duration of diabetes, BMI of 27.4 (20.3–30.1) kg/m2 and glycated hemoglobin of 6.3% (5.4–9.7%; 47 (36–82) mmol/mol). The eGFR of all women was >90 mL/min/1.73 m2. All seven women with type 1 diabetes had detectable postprandial UCPCR in the first and second undiluted urine samples, median (25th–75th range) 173 (5.4–1014) and 650 (27.5–2250) pmol/mmol. Six women had a rise in UCPCR (median rise (25th–75th range) 477 (23–1221) pmol/mmol) whereas one woman had a fall in UCPCR between the two samples collected at 13 and 31 weeks' gestation (figure 1C).

in the first and second undiluted urine samples, median (25th–75th range) 173 (5.4–1014) and 650 (27.5–2250) pmol/mmol. Six women had a rise in UCPCR (median rise (25th–75th range) 477 (23–1221) pmol/mmol) whereas one woman had a fall in UCPCR between the two samples collected at 13 and 31 weeks' gestation (figure 1C). Discussion Insulin is secreted in equimolar concentrations as C peptide and UCPCR provides an integrated measurement of insulin secretion over the interval of the urine collection.10 As an integrated measure it provides a more informative assessment of insulin secretion over time than a spot serum C peptide concentration that has a 20–30 min circulating half-life.11 In addition, the use of the UCPCR provides a more practical methodology due to its ease in collection and processing than serum C peptide in clinical practice and research.12

ore informative assessment of insulin secretion over time than a spot serum C peptide concentration that has a 20–30 min circulating half-life.11 In addition, the use of the UCPCR provides a more practical methodology due to its ease in collection and processing than serum C peptide in clinical practice and research.12 The correlation between UCPCROGTT and serum C peptideOGTT in 90 glucose-tolerant pregnant women supports the use of the UCPCR to assess insulin secretion during pregnancy. Outside pregnancy, the use of the UCPCR to assess endogenous insulin secretion is well established. The published median (25–75th range) for a mixed meal tolerance test (MMTT) stimulated UCPCR among 27 glucose-tolerant non-pregnant women is 4040 (3000–6990) pmol/mmol; however, this study used a Roche Diagnostic C peptide assay, that using different assay formats and antibodies.8 13 These values are approximately a third of the 120 min UCPCROGTT values of the 73 pregnant glucose-tolerant women (median and 25–75th range of 12 304 (8621–20 733) pmol/mmol) in the current study. An increase in UCPCR and serum C peptide concentration in pregnancy is to be expected due to the physiological decrease in insulin sensitivity that occurs at this time.14 The use of the UCPCR in pregnancy should correct for the physiological increase in glomerular filtration rate that occurs throughout pregnancy.15

dy. An increase in UCPCR and serum C peptide concentration in pregnancy is to be expected due to the physiological decrease in insulin sensitivity that occurs at this time.14 The use of the UCPCR in pregnancy should correct for the physiological increase in glomerular filtration rate that occurs throughout pregnancy.15 All seven pregnant women with long-term type 1 diabetes had detectable non-fasting UCPCR when first measured, albeit with values approximately a 10th of those seen in the 90 pregnant women with normal glucose tolerance tested at 28 (24–29) weeks' gestation. The published median (25th and 75th range) for a MMTT-stimulated UCPCR of 58 non-pregnant women with type 1 diabetes with over 5 years duration is 20 (0–400) pmol/mmol.13 These ranges for UCPCR of non-pregnant women are approximately a fifth lower than those seen for the second urine sample among the seven pregnant women with type 1 diabetes, median (range) 650 (27.5–2250 pmol/mmol), with three women having a UCPCR value >1000 pmol/mmol. However, it has to be recognized that the UCPCR measurements in the current study were performed on undiluted urine that is standard practice in our laboratory for studies in type 1 diabetes. The use of undiluted 24 hours urinary collections has been validated for assaying low C peptide concentrations using the same assay methodology.8

o be recognized that the UCPCR measurements in the current study were performed on undiluted urine that is standard practice in our laboratory for studies in type 1 diabetes. The use of undiluted 24 hours urinary collections has been validated for assaying low C peptide concentrations using the same assay methodology.8 Three separate studies in pregnant women with type 1 diabetes have examined circulating C peptide, reporting it either becomes detectable for the first time in pregnancy or increases during pregnancy in some women. Our findings of detectable UCPCR in pregnancy in a small group of women with longstanding type 1 diabetes suggests that using UCPCR in pregnancy might be a suitable methodology for studying pregnancy-induced β-cell regeneration or neogenesis in humans. There are subtle clinical pointers that pregnancy is related to either β-cell regeneration or neogenesis in women with type 1 diabetes having increased endogenous bioactive insulin secretion. In early pregnancy, there is a decrease of exogenous insulin requirement, and throughout pregnancy a lower than expected incidence of diabetic ketoacidosis despite a fall in serum bicarbonate levels and accelerated maternal lipolysis and ketosis in later pregnancy.16 The possibility that residual β-cells in non-pregnant individuals with type 1 diabetes may emerge from neogenesis of pancreatic ductal cells has been suggested.17 Somatolactogenic hormones and hyperglycemia have been implicated in the enlargement of the pancreatic islets and β-cell induction and proliferation seen in rodents.18

bility that residual β-cells in non-pregnant individuals with type 1 diabetes may emerge from neogenesis of pancreatic ductal cells has been suggested.17 Somatolactogenic hormones and hyperglycemia have been implicated in the enlargement of the pancreatic islets and β-cell induction and proliferation seen in rodents.18 The β-cell adaptation due to neogenesis from other pancreatic cell types, forming new small islets rather than hyperplasia, has been proposed to occur in human pregnancy.1 Pregnancy-related factors capable of neogenesis of the human β-cells could have therapeutic implications for the future treatment of type 1 diabetes. In summary, this study demonstrated that UCPCR provides a robust and practical means for assessing insulin secretion during pregnancy, and provides a practical methodology to assess in future studies the potential for β-cell adaptation in women with type 1 diabetes. The authors would like to thank Dr Shivani Misra for her help with the study protocol and Ludwig Lupak for his expert support in conducting the biochemical analysis of the samples. Contributors: AM designed and conducted the study, analyzed the data, interpreted the results and wrote the manuscript. RV and CT contributed to the conduct of the study and data analysis. AD is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Funding: Imperial College London. Competing interests: None declared.

Contributors: AM designed and conducted the study, analyzed the data, interpreted the results and wrote the manuscript. RV and CT contributed to the conduct of the study and data analysis. AD is the guarantor of this work and, as such, had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Funding: Imperial College London. Competing interests: None declared. Ethics approval: Imperial College Healthcare Tissue Bank and the Research Ethics Committee Wales: 12/WA/0196. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: The raw data for this study are available on request.

Key messages Many estimates for the burden of diabetes come from cohort studies and surveys, which are often not population based. There are very few estimates of incidence of diabetes globally. We used a national administrative pharmacy claims database to calculate prevalence and incidence of type 2 diabetes. This method captures all treated cases of diabetes and thus avoids problems with sampling and generalizing to whole populations that arise from cohort and survey methods. Understanding the true burden of disease in our society is essential for the planning of health services, which in turn help achieve optimal outcomes in terms of diabetes-related morbidity and mortality. Using pharmacy claims data is a straightforward method of achieving up to date estimates. Introduction Diabetes mellitus is a leading cause of death globally, causing almost 4 million deaths in 2010.1 Morbidity arising from the disease is also substantial; the Global Burden of Disease study estimated a 30% increase in disability-adjusted life years for diabetes between 1990 and 2010 due to increasing prevalence of the disease and increased longevity of those living with diabetes.2 To reduce diabetes-related mortality and morbidity, access to appropriate healthcare services for people with diabetes is a necessity. For this to be successful, up-to-date population level estimates of disease burden are required to rationally plan and deliver the required health services.

Introduction Diabetes mellitus is a leading cause of death globally, causing almost 4 million deaths in 2010.1 Morbidity arising from the disease is also substantial; the Global Burden of Disease study estimated a 30% increase in disability-adjusted life years for diabetes between 1990 and 2010 due to increasing prevalence of the disease and increased longevity of those living with diabetes.2 To reduce diabetes-related mortality and morbidity, access to appropriate healthcare services for people with diabetes is a necessity. For this to be successful, up-to-date population level estimates of disease burden are required to rationally plan and deliver the required health services. The Institute of Public Health (IPH) estimates and forecasts for the prevalence of diabetes (combined type 1 and type 2) are often cited and are a valuable resource.3 However, the most recent estimates from the IPH are based on a cross-sectional survey of adults in the Survey of Lifestyle, Attitudes and Nutrition (SLAN) which dates back to 2007.3 4

(IPH) estimates and forecasts for the prevalence of diabetes (combined type 1 and type 2) are often cited and are a valuable resource.3 However, the most recent estimates from the IPH are based on a cross-sectional survey of adults in the Survey of Lifestyle, Attitudes and Nutrition (SLAN) which dates back to 2007.3 4 Other estimates of diabetes prevalence come from The Irish Longitudinal Study of Ageing (TILDA), which is limited to those over 50 years,5 6 the Mitchelstown cohort which was limited to adults aged 50–69 years in one rural area in the South of Ireland7 and the Central Statistics Office (CSO) Quarterly National Household Survey (QNHS) from 2010.8 These estimates are based on a variety of self-reported doctor diagnosis; self-reported diabetes medication usage; or a combination of self-report and HbA1c data; and all are limited to adult populations. Furthermore, there is a scarcity of data available on the incidence of diabetes in Ireland.9 In this study, we used national pharmacy claims data to estimate the prevalence and incidence of type 2 diabetes. These data provide an objective measure of treated diabetes in the total population to complement existing sample-based estimates.

Other estimates of diabetes prevalence come from The Irish Longitudinal Study of Ageing (TILDA), which is limited to those over 50 years,5 6 the Mitchelstown cohort which was limited to adults aged 50–69 years in one rural area in the South of Ireland7 and the Central Statistics Office (CSO) Quarterly National Household Survey (QNHS) from 2010.8 These estimates are based on a variety of self-reported doctor diagnosis; self-reported diabetes medication usage; or a combination of self-report and HbA1c data; and all are limited to adult populations. Furthermore, there is a scarcity of data available on the incidence of diabetes in Ireland.9 In this study, we used national pharmacy claims data to estimate the prevalence and incidence of type 2 diabetes. These data provide an objective measure of treated diabetes in the total population to complement existing sample-based estimates. Methods Health system In Ireland, access to and reimbursement for diabetes medicines occur via two publicly funded community drug schemes. The first is the General Medical Services (GMS) scheme; the main public health insurance program providing primary and secondary healthcare free at the point of access to ∼40% of the Irish population on a means-tested basis.10 Medicines are included under this scheme but are subject to a copayment (€2.50 currently). The second drug scheme is the long-term illness (LTI) scheme. The LTI provides free access to condition-related medicines for individuals diagnosed with any of 16 chronic illnesses including diabetes. LTI coverage is independent of income.

are included under this scheme but are subject to a copayment (€2.50 currently). The second drug scheme is the long-term illness (LTI) scheme. The LTI provides free access to condition-related medicines for individuals diagnosed with any of 16 chronic illnesses including diabetes. LTI coverage is independent of income. Data Pharmacists dispensing medicines to all patients (adults and children) on the GMS and the LTI scheme are reimbursed by the government via the Health Service Executive-Primary Care Reimbursement Service (HSE-PCRS). We used dispensing data from the HSE-PCRS database from July 2011 to December 2012. Data were available for the drug dispensed (classified by WHO Anatomical Therapeutic Chemical (WHO ATC) code), date dispensed, quantity and strength, in addition to patient age and sex. Population denominator data for the year 2012 were population estimates derived by the CSO based on the 2011 census.11 Definitions Type 2 diabetes was classified as using any strength or quantity of an oral hypoglycemic agent (WHO ATC A10B), irrespective of age or insulin use. The different agents included in this study, stratified by age and sex, are given in table 1. Table 1 Types of medicines used in 2012

Population denominator data for the year 2012 were population estimates derived by the CSO based on the 2011 census.11 Definitions Type 2 diabetes was classified as using any strength or quantity of an oral hypoglycemic agent (WHO ATC A10B), irrespective of age or insulin use. The different agents included in this study, stratified by age and sex, are given in table 1. Table 1 Types of medicines used in 2012 Medicine group Biguanides Sulfonylureas Thiazolidinediones DPP-4 Other Total A10BA A10BB A10BG A10BH A10BC, A10BF, A10BX n (%) WHO ATC code n (%) n (%) n (%) n (%) n (%) Total 907 125 (51.9) 527, 886 (30.2) 25 658 (1.5) 126 345 (7.2) 60 069 (3.4) 1 747 755 (100.0) <15 years 1547 (50.0) 914 (29.5) 50 (1.6) 142 (4.6) 139 (4.5) 3096 (0.18) 15–24 years 2959 (69.7) 675 (15.9) 57 (1.3) 178 (4.2) 212(5.0) 4247 (0.24) 25–34 years 12 537 (67.6) 3020 (16.3) 247 (1.3) 734(4.0) 1088 (5.9) 18 554 (1.06) 35–44 48 558 (59.3) 17 542 (21.4) 1243 (1.5) 4414 (5.4) 4787 (5.8) 81 960 (4.7) 45–54 years 125 732 (54.0) 57 833 (24.8) 3553 (1.5) 14 760 (6.3) 14 168 (6.1) 232 955 (13.3) 55–64 years 236 515 (52.6) 122 554 (27.3) 7493 (1.7) 31 253 (6.9) 20 675 (4.6) 449 432 (25.7) 65–69 years 140 569 (52.0) 79 897 (29.6) 4372 (1.6) 20 725 (7.7) 7936 (2.9) 270 106 (15.5) 70+ years 333 853 (49.3) 242 256 (35.8) 8435 (1.3) 53 435 (7.9) 10 600 (1.6) 677 401 (38.8) Women 368 707 (53.2) 204 465 (29.5) 9450 (1.4) 49 794 (7.2) 25 412 (3.7) 692 565 (39.6) Men 536 766 (51.0) 322 465 (30.7) 16 073 (1.5) 76 275 (7.3) 34 545 (3.3) 1 051 910 (60.2) Numbers are numbers of prescriptions in 2012.

5.5) 70+ years 333 853 (49.3) 242 256 (35.8) 8435 (1.3) 53 435 (7.9) 10 600 (1.6) 677 401 (38.8) Women 368 707 (53.2) 204 465 (29.5) 9450 (1.4) 49 794 (7.2) 25 412 (3.7) 692 565 (39.6) Men 536 766 (51.0) 322 465 (30.7) 16 073 (1.5) 76 275 (7.3) 34 545 (3.3) 1 051 910 (60.2) Numbers are numbers of prescriptions in 2012. Other includes sulfonamides, α glucosidase inhibitors and ‘other’ agents as defined by WHO ATC dictionary. DPP-4, dipeptidyl peptidase-4 inhibitors. Calculation of incidence and prevalence To estimate the prevalence of type 2 diabetes, we used dispensing data for 2012. We counted the number of people in the database who met our definition of type 2 diabetes and used this as the numerator. The total population count published by the CSO was the denominator.11 To establish the annual incidence for type 2 diabetes in 2012, we used data from July 2011 to December 2012. An individual's first occurrence in 2012 meeting the definition of type 2 diabetes was referred to as the index date. A 6 month look-back period was used to rule out prior use before the index date. If no prior use of an oral hypoglycemic agent occurred in the look-back period, then the individual was an incident user of oral hypoglycemic medicines and thus an incident case. This count was used as the numerator, while the denominator was the total population count published by the CSO minus the number of prevalent of cases.11 We carried out subgroup analyses by age group (<15 years, ≥15 years, 15–24 years, 25–34 years, 35–44 years, 45–54 years, 55–64 years, 65–69 years and >70+ years) and sex.

To establish the annual incidence for type 2 diabetes in 2012, we used data from July 2011 to December 2012. An individual's first occurrence in 2012 meeting the definition of type 2 diabetes was referred to as the index date. A 6 month look-back period was used to rule out prior use before the index date. If no prior use of an oral hypoglycemic agent occurred in the look-back period, then the individual was an incident user of oral hypoglycemic medicines and thus an incident case. This count was used as the numerator, while the denominator was the total population count published by the CSO minus the number of prevalent of cases.11 We carried out subgroup analyses by age group (<15 years, ≥15 years, 15–24 years, 25–34 years, 35–44 years, 45–54 years, 55–64 years, 65–69 years and >70+ years) and sex. Results In 2012, 1 655 013 people accessed a prescription on the GMS scheme and were available in our data set. The mean age was 42.9 years (SD 25.9), and the population was 54.4% women. On the LTI scheme, 68 996 people accessed at least one prescription in 2012. The mean age was 48.4 years (SD 25.2), and the population was 38% women.

1 655 013 people accessed a prescription on the GMS scheme and were available in our data set. The mean age was 42.9 years (SD 25.9), and the population was 54.4% women. On the LTI scheme, 68 996 people accessed at least one prescription in 2012. The mean age was 48.4 years (SD 25.2), and the population was 38% women. Type 2 diabetes mellitus In 2012, 114 957 people were classified as prevalent type 2 diabetes cases, leading to a prevalence of 2.51% (95% CI 2.49% to 2.52%) in the total population. After excluding those aged <15 years, an adult population prevalence of 3.16% (95% CI 3.15% to 3.18%) was obtained (table 2). Figure 1 demonstrates how the prevalence increased with age; 55–64 years (6.50%), 65–69 years (10.75%) and 70+ years (12.10%). Men had a higher prevalence of type 2 diabetes than women at 2.96% (95% CI 2.94 to 2.98) vs 2.04% (95% CI 2.02% to 2.06%) (χ2 test for homogeneity p<0.0001). Table 2 Prevalence and incidence of type 2 diabetes

econdary pneumonia, we calculated the influenza VE to prevent all community-acquired LRTI (considered as a broad category of all ‘chest infections’, including influenza infections, and possible secondary infections such as bronchitis and pneumonia), using a ratio-of-ratios analysis to address confounding by indication. Research design and methods Data sources We analyzed data from the Clinical Practice Research Datalink (CPRD), a database of anonymized primary care medical records. Data were extracted in May 2011, and contained records for 12.8 million patients at 627 practices across the UK.22 Records include patient demographics, health behaviors, test results, diagnoses, and prescriptions. Diagnoses are recorded using Read codes, and have generally been found to have good positive predictive value in validations.23 The CPRD population is similar to the general UK population in terms of age and sex.24 25 Linked data are available for patients in England, subject to practice-level consent. This study used linked data on all hospital inpatient admissions to NHS hospitals in England from Hospital Episodes Statistics (HES), and socioeconomic status from the Office for National Statistics (ONS).

Type 2 diabetes mellitus In 2012, 114 957 people were classified as prevalent type 2 diabetes cases, leading to a prevalence of 2.51% (95% CI 2.49% to 2.52%) in the total population. After excluding those aged <15 years, an adult population prevalence of 3.16% (95% CI 3.15% to 3.18%) was obtained (table 2). Figure 1 demonstrates how the prevalence increased with age; 55–64 years (6.50%), 65–69 years (10.75%) and 70+ years (12.10%). Men had a higher prevalence of type 2 diabetes than women at 2.96% (95% CI 2.94 to 2.98) vs 2.04% (95% CI 2.02% to 2.06%) (χ2 test for homogeneity p<0.0001). Table 2 Prevalence and incidence of type 2 diabetes GMS LTI Total Population (CSO) Estimate (%) 95% CI Prevalence estimates Total population 81 177 33 780 114 957 4 585 000 2.51 2.49 to 2.52 Total population ≥15 years 80 618 32 987 113 605 3 590 600 3.16 3.15 to 3.18 <15 years 721 1198 1919 994 800 0.19 0.18 to 0.20 15–24 years 594 128 722 553 500 0.13 0.12 to 0.14 25–34 years 1702 765 2467 733 500 0.34 0.32 to 0.35 35–44 years 4309 3081 7390 700 000 1.06 1.03 to 1.08 45–54 years 8964 8088 17 052 586 300 2.91 2.87 to 2.95 55–64 years 16 466 13 938 30 404 468 000 6.50 6.43 to 6.57 65–69 years 12 396 7120 19 516 181 500 10.75 10.61 to 10.90 70+ years 41 098 3397 44 495 367 800 12.10 11.99 to 12.20 Women 36 968 10 231 47 199 2 315 800 2.04 2.02 to 2.06 Men 43 707 23 510 67 217 2 269 600 2.96 2.94 to 2.98 Incidence estimates Total population 15 788 5786 21 574 4 470 043 0.48 0.48 to 0.49 Total population ≥15 years 15 353 5679 21 032 3 476 995 0.60 0.60 to 0.61 <15 years 213 234 447 992 881 0.05 0.04 to 0.05 15–24 years 405 55 460 552 678 0.08 0.08 to 0.09 25–34 years 936 293 1229 731 033 0.17 0.16 to 0.18 35–44 years 1584 819 2403 692 610 0.35 0.33 to 0.36 45–54 years 2459 1523 3982 569 248 0.70 0.68 to 0.72 55–64 years 3485 2026 5511 437 596 1.26 1.23 to 1.29 65–69 years 2406 971 3377 161 984 2.08 2.02 to 2.15 70+ years 5585 413 5998 323 305 1.86 1.81 to 1.9 Women 7242 1959 9201 2 268 601 0.41 0.40 to 0.41 Men 8165 3815 11 980 2 202 383 0.54 0.53 to 0.55 Figure 1 Prevalence and incidence of type 2 diabetes.

0 0.68 to 0.72 55–64 years 3485 2026 5511 437 596 1.26 1.23 to 1.29 65–69 years 2406 971 3377 161 984 2.08 2.02 to 2.15 70+ years 5585 413 5998 323 305 1.86 1.81 to 1.9 Women 7242 1959 9201 2 268 601 0.41 0.40 to 0.41 Men 8165 3815 11 980 2 202 383 0.54 0.53 to 0.55 Figure 1 Prevalence and incidence of type 2 diabetes. In the same year, 21 574 people developed type 2 diabetes giving an incidence of 0.48% (95% CI 0.48% to 0.49%). This was estimated at 0.60% (95% CI 0.60% to 0.61%) in the population aged ≥15 years. The incidence of type 2 diabetes increased with age, reaching its highest level of 2.08% (95% CI 2.02 to 2.15) in people aged 65–69 years. Men had a higher incidence of type 2 diabetes (0.54%) than women (0.41%) (table 2 and figure 1). Discussion This cross-sectional study estimated the prevalence and incidence of type 2 diabetes using population level data from a national pharmacy claims database. The overall prevalence in the adult population was 3.16%. The incidence of type 2 diabetes was 6 cases per 1000 adult people in 2012.

In the same year, 21 574 people developed type 2 diabetes giving an incidence of 0.48% (95% CI 0.48% to 0.49%). This was estimated at 0.60% (95% CI 0.60% to 0.61%) in the population aged ≥15 years. The incidence of type 2 diabetes increased with age, reaching its highest level of 2.08% (95% CI 2.02 to 2.15) in people aged 65–69 years. Men had a higher incidence of type 2 diabetes (0.54%) than women (0.41%) (table 2 and figure 1). Discussion This cross-sectional study estimated the prevalence and incidence of type 2 diabetes using population level data from a national pharmacy claims database. The overall prevalence in the adult population was 3.16%. The incidence of type 2 diabetes was 6 cases per 1000 adult people in 2012. Existing prevalence estimates pertaining to the general adult population range from 3% in the Quarterly National Household survey to 3.5% in those aged ≥18 years using SLAN survey data.4 8 Our estimate of 3.16% is thus comparable to previous figures. In addition, our age stratified estimates for people aged ≥50 years are similar to those based on TILDA using self-report of doctor diagnosis and HbA1c measures.5 6 However, our estimates may underestimate the true burden of diabetes given that we have excluded type 1 diabetes and we also could not account for lifestyle-treated diabetes or undiagnosed diabetes. Despite this, the true prevalence rate of diabetes in Ireland is likely lower than that in the USA which was recently estimated at 8.3% in the adult population.12 In England, diagnosed diabetes in the population aged ≥16 years is estimated at 5.6% from Health Survey for England data.13

diabetes or undiagnosed diabetes. Despite this, the true prevalence rate of diabetes in Ireland is likely lower than that in the USA which was recently estimated at 8.3% in the adult population.12 In England, diagnosed diabetes in the population aged ≥16 years is estimated at 5.6% from Health Survey for England data.13 The only other estimate of incidence of type 2 diabetes in Ireland is 2 cases per 1000, in contrast to the 6 cases per 1000 we found in this study.9 An American study using the National Health Interview Survey found an incidence rate of 7.1/1,000 people aged ≥20 years in 2012, indicating that Irish incidence rates are below those in North America.12 A recent Danish study calculated incidence rates for every year of age.14 While it is difficult to compare single age estimates with estimates for age categories, our estimates appear comparable to those in the Danish study, albeit somewhat higher in the older age groups.14 Making international comparisons is helpful to aid in understanding the plausibility of our estimates, however differences do exist between populations for demographic and methodological reasons.15

ategories, our estimates appear comparable to those in the Danish study, albeit somewhat higher in the older age groups.14 Making international comparisons is helpful to aid in understanding the plausibility of our estimates, however differences do exist between populations for demographic and methodological reasons.15 The study is limited by lack of information on undiagnosed diabetes and lifestyle-treated diabetes. Other data sources, for example the Mitchelstown Cohort and SLAN survey, provide information on undiagnosed diabetes, which can be used in tandem with our results.4 7 Unpublished data from the Mitchelstown Cohort study of over 2000 adults aged 50–69 years reveal that ∼7% of those with self-reported diabetes are treated with diet only.16 Although these data are not nationally representative, they provide some context on the magnitude of underestimation. Furthermore, we relied on diagnosed individuals adhering to their treatment regimens, their dispensed medications thus appearing in the pharmacy claims database. We did not anticipate non-adherence to be a major problem given that medicines are free on the LTI scheme and subject to a small copayment on the GMS scheme (€0.50 per item in 2012).17

diagnosed individuals adhering to their treatment regimens, their dispensed medications thus appearing in the pharmacy claims database. We did not anticipate non-adherence to be a major problem given that medicines are free on the LTI scheme and subject to a small copayment on the GMS scheme (€0.50 per item in 2012).17 The study is strengthened by the objective and reliable nature of the data.18 Additionally, because diabetes medicines are generally provided only through the GMS and LTI drug schemes, data on those with diagnosed and treated diabetes should be nationally complete in this database offering population level data for all ages, including children, in contrast to previous surveys and cohort studies which are limited to adults. Unfortunately, the database does not have access to diagnosis codes, thus we made the assumption that all oral hypoglycemic agents were being used to treat diabetes. A notable exception is the use of metformin for polycystic ovarian syndrome (PCOS). However, as PCOS effects only a small proportion of women of reproductive age, ∼8%, and only some of these will be treated, any effect on our estimates is likely to be small. Further, many women with PCOS will have diabetes, and we will have intended to include them in our estimates.19 Metformin is also used in pre-diabetes, which due to lack of diagnosis codes, we have not been able to separate from our estimates. We know from our prior research that the prevalence of pre-diabetes in those aged ≥45 years is 20%.20 From the most recent audit of diabetes management in General Practice, we know that ∼90% of those with pre-diabetes are treated with dietary intervention (unpublished).21 Thus, any bias contributed to our results from including those with metformin-treated pre-diabetes is likely to be inconsequential.

5 years is 20%.20 From the most recent audit of diabetes management in General Practice, we know that ∼90% of those with pre-diabetes are treated with dietary intervention (unpublished).21 Thus, any bias contributed to our results from including those with metformin-treated pre-diabetes is likely to be inconsequential. Our study has demonstrated the utility of routinely collected administrative claims data in calculating measures of disease burden, including incidence. The method is straightforward, and while we used data from 2012 for this study, more current data would allow estimating disease burden for the most recently completed calendar year along with establishing longitudinal trends. This approach affords advantages in real-time monitoring of disease burden and thus presents a key resource for evaluating public health interventions to reduce the prevalence and incidence of type 2 diabetes, thus informing health policy and health service planning. To address acknowledged weaknesses in using these data, estimates for prevalence and incidence should be considered in combination with cross-sectional and cohort study results to account for undiagnosed and lifestyle-treated cases. Contributors: S-JS, PMK and SMcH conceived of and planned the study. S-JS carried out the analyses and wrote the paper. S-JS, PMK, SMcH, SB, RL and HW all reviewed/edited the manuscript.

Our study has demonstrated the utility of routinely collected administrative claims data in calculating measures of disease burden, including incidence. The method is straightforward, and while we used data from 2012 for this study, more current data would allow estimating disease burden for the most recently completed calendar year along with establishing longitudinal trends. This approach affords advantages in real-time monitoring of disease burden and thus presents a key resource for evaluating public health interventions to reduce the prevalence and incidence of type 2 diabetes, thus informing health policy and health service planning. To address acknowledged weaknesses in using these data, estimates for prevalence and incidence should be considered in combination with cross-sectional and cohort study results to account for undiagnosed and lifestyle-treated cases. Contributors: S-JS, PMK and SMcH conceived of and planned the study. S-JS carried out the analyses and wrote the paper. S-JS, PMK, SMcH, SB, RL and HW all reviewed/edited the manuscript. Funding: At the time the work was carried out, S-JS was funded under Health Research Board in Ireland under grant no. PHD/2007/16. PMK is funded under the Health Research Board Leadership Award in Diabetes (RL/2013/7). SMcH is funded by the Centre for Ageing and Development Research in Ireland (CARDI) Leadership Fellowship. Competing interests: None declared. Patient consent: Obtained. Provenance and peer review: Not commissioned; externally peer reviewed.

Funding: At the time the work was carried out, S-JS was funded under Health Research Board in Ireland under grant no. PHD/2007/16. PMK is funded under the Health Research Board Leadership Award in Diabetes (RL/2013/7). SMcH is funded by the Centre for Ageing and Development Research in Ireland (CARDI) Leadership Fellowship. Competing interests: None declared. Patient consent: Obtained. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: We have provided most of the raw numbers in the included table. We are happy to provide other aggregate level results for anyone who wishes to obtain them, within the limits of our own data agreements.

Key messages Renal transplant recipients are at high risk for development of new-onset diabetes after transplantation and premature mortality. Dietary habits consistent with a Mediterranean style diet are associated with lower risk of development of new-onset diabetes after transplantation and premature mortality in renal transplant recipients. These lower risks were independent of potential confounders, including age, sex, physical activity and smoking behavior. Introduction The prevalence of patients with end-stage renal disease for which patients require chronic dialysis or renal transplantation, also called renal replacement therapy, is increasing at a rate of 7% per year.1 Renal transplantation improves quality of life and increases survival, compared to dialysis treatment.2 3 However, the results are not impressive, since a lot of factors impair survival after renal transplantation. One of these factors is new-onset diabetes after transplantation (NODAT), which is caused by the combination of insulin resistance and deficient insulin production.4 The incidence of NODAT is high and affects ∼20% of renal transplant recipients (RTR).5 However, a systemic review of Montori et al6 concluded that the incidence of NODAT varied widely from 2% to 50% in the first post-transplant year, due to inconsistencies regarding the definition of NODAT. NODAT is an important risk factor for premature mortality in RTR.7 8 This can partly be explained by the fact that NODAT contributes to a high cardiovascular risk, infectious complications and impaired graft survival in RTR.9–11

n the first post-transplant year, due to inconsistencies regarding the definition of NODAT. NODAT is an important risk factor for premature mortality in RTR.7 8 This can partly be explained by the fact that NODAT contributes to a high cardiovascular risk, infectious complications and impaired graft survival in RTR.9–11 Patients with end-stage renal disease are often advised to consume a low protein, low potassium, low sodium diet, which is also often restricted in fluid intake, causing a diet high in energy-rich drinks, carbohydrates and fats to get enough energy.12 This diet prevents excessive generation of urea and the occurrence of hyperphosphatemia, hyperkalemia and hypertension. However, a low-protein diet gives little satiety,13 and adherence to this dietary pattern may induce problems once combined with the stimulation of appetite. This may be a consequence of improved renal function after renal transplantation and as a side effect of the corticosteroid treatment that is part of the immunosuppressive regimen after renal transplantation.14 Furthermore, RTR have low levels of physical activity and gain fat mass after transplantation, resulting in obesity in 26% of the participants,12 an independent risk factor for the development of NODAT.9 To reduce the incidence of NODAT and improve overall transplant success, more attention should be paid to lifestyle modification.

RTR have low levels of physical activity and gain fat mass after transplantation, resulting in obesity in 26% of the participants,12 an independent risk factor for the development of NODAT.9 To reduce the incidence of NODAT and improve overall transplant success, more attention should be paid to lifestyle modification. Meta-analyses showed that adherence to the Mediterranean diet was associated with a lower risk of diabetes.15 16 Furthermore, Estruch et al17 showed in the PREDIMED trial that two components of the Mediterranean diet (additional amount of extra virgin olive oil and nuts) reduced the incidence of cardiovascular events among high-risk patients. In the present study, we aimed to investigate the association of Mediterranean Style diet with the incidence of NODAT and all-cause mortality in RTR that has not yet been examined. We hypothesized that a Mediterranean Style diet is associated with a lower risk of NODAT and all-cause mortality in RTR.

igh-risk patients. In the present study, we aimed to investigate the association of Mediterranean Style diet with the incidence of NODAT and all-cause mortality in RTR that has not yet been examined. We hypothesized that a Mediterranean Style diet is associated with a lower risk of NODAT and all-cause mortality in RTR. Research design and methods Study design and population The study design of this research project is a large single-center prospective cohort of RTR. All adult RTR (≥18 years) with a functioning graft for at least 1 year who visited the outpatient clinic in the University Medical Center Groningen between November 2008 and May 2011 were invited to participate. Patients were only included if this visit was at least 1 year after transplantation. All patients had sufficient knowledge of the Dutch language and according to their patient files and no history of drug or alcohol addiction. In total 707 (86.5%) of the initially 817 invited RTR signed a written informed consent. RTR with missing dietary data (n=81), diabetes mellitus (DM) at baseline (n=152) or who underwent combined pancreas-kidney transplantation (n=6) were excluded, leaving 468 RTR eligible for analyses. This research project was approved by the institutional review board (METc 2008/186), which adhered to the Declaration of Helsinki.

issing dietary data (n=81), diabetes mellitus (DM) at baseline (n=152) or who underwent combined pancreas-kidney transplantation (n=6) were excluded, leaving 468 RTR eligible for analyses. This research project was approved by the institutional review board (METc 2008/186), which adhered to the Declaration of Helsinki. Data collection and clinical end points Baseline data were collected during a morning visit to the outpatient clinic as described previously.18 19 Body weight and height were measured while participants wore indoor clothing without shoes. Waist and hip circumference were measured. Both systolic blood pressure and diastolic blood pressure and also heart rate were measured using a semiautomatic device (Dinamap1846; Critikon, Tampa, Florida, USA). They were measured every minute for 15 min in a half-sitting position to prevent white coat hypertension. Information about daily physical activity was derived using the Short QUestionnaire to ASsess Health-enhancing physical activity (SQUASH) score in time multiplied by intensity.20 Information about smoking behavior was obtained by using a questionnaire. Medication use was acquired from patient records. Furthermore, fasting blood samples were collected and patients were also asked to complete 24 hour urine collection. All RTR were informed to discard their morning urine specimen and then collect their urine for the next 24 hours, including the next morning's first specimen of the day of their visit. Creatinine clearance was based on 24 hour urinary creatinine and serum creatinine. Estimated glomerular filtration rate was calculated using the serum creatinine-based Chronic Kidney Disease Epidemiology Collaboration equation.21

the next 24 hours, including the next morning's first specimen of the day of their visit. Creatinine clearance was based on 24 hour urinary creatinine and serum creatinine. Estimated glomerular filtration rate was calculated using the serum creatinine-based Chronic Kidney Disease Epidemiology Collaboration equation.21 A semiquantitative food frequency questionnaire (FFQ) was used to collect information on dietary intake at baseline during the past month. The FFQ was developed at Wageningen University22 and consisted of 177 items. Patients filled out the self-administered FFQ at home. Frequency was recorded in times per day, week or month for each item. Expression of number of servings was in either natural units such as a slice of bread or an apple, or in household measures, for example, a cup or a teaspoon. Subsequently, all dietary data were converted into total energy and nutrient intake per day using the Dutch Food Composition Table (NEVO 2006). The FFQ was validated by comparing the protein intake of the FFQ with the protein intake calculated by the Maroni Equation, using urinary urea excretion values.19 Protein intake was similar to the estimates derived from the FFQ.23

al energy and nutrient intake per day using the Dutch Food Composition Table (NEVO 2006). The FFQ was validated by comparing the protein intake of the FFQ with the protein intake calculated by the Maroni Equation, using urinary urea excretion values.19 Protein intake was similar to the estimates derived from the FFQ.23 The degree to which the consumed diet resembled the traditional Mediterranean diet was calculated according to a nine-point Mediterranean Diet Score (MDS) of Trichopoulou et al.24 The MDS includes nine food groups: ratio of monounsaturated:saturated fatty acids, intake of legumes, cereals, vegetables, fruit, fish, dairy products, meat products and alcohol.9 14 Food items of the FFQ were divided over these nine food groups (see online supplementary table S1). For each food group, the sex-specific median in grams per day was used as cut-off point for making this division, except for fish and alcohol. Patients received a score of 1 for each of the putative protective components (ratio of monounsaturated:saturated fatty acids, legumes, cereals, vegetables, fruit) if their intake was above the median. The traditional Mediterranean diet was low in dairy and meat products. Therefore, an intake below the median for these food groups was scored 1 and for an intake above the median 1. For the fish component participants received a score of 1 if they consumed more than 5 g/day and a score of 0 if they consumed <5 g/day (<once a month). Alcohol users received a score of 1 and non-users received a score of 0. Moderate alcohol intake is associated with low prevalence of NODAT and lower risk for mortality in RTR, when compared to abstainers.25 The MDS varies between 0 (lowest adherence) and 9 (highest adherence). Subsequently, all patients were divided into two groups based on the frequency distribution of the MDS: group 1 (MDS 0–4) and group 2 (MDS 5–9). We dichotomized data because of the small number of events.

ality in RTR, when compared to abstainers.25 The MDS varies between 0 (lowest adherence) and 9 (highest adherence). Subsequently, all patients were divided into two groups based on the frequency distribution of the MDS: group 1 (MDS 0–4) and group 2 (MDS 5–9). We dichotomized data because of the small number of events. 10.1136/bmjdrc-2016-000283.supp1supplementary table Overview of the food items. In this study primary outcome measurements are the incidence of NODAT and all-cause mortality. Care-based data about the incidence of NODAT and mortality after baseline were retrieved from patient files of all RTR until 1 April 2014. NODAT was defined as a fasting plasma glucose level ≥7.0 mmol/L and/or use of oral hypoglycemic agents or insulin therapy for ≥30 consecutive days.26 Patients developed NODAT when they had one or more of the following conditions; the patient was diagnosed with DM, used antidiabetics (oral hypoglycemic agents or insulin therapy) and/or had a fasting plasma glucose level ≥7.0 mmol/L or non-fasting glucose level ≥11.1 mmol/L. The diagnosis of NODAT was made in outpatient and clinical routine. Patients were tested repeatedly before the diagnosis of NODAT was made and treatment was started. No participants were lost to follow-up.

insulin therapy) and/or had a fasting plasma glucose level ≥7.0 mmol/L or non-fasting glucose level ≥11.1 mmol/L. The diagnosis of NODAT was made in outpatient and clinical routine. Patients were tested repeatedly before the diagnosis of NODAT was made and treatment was started. No participants were lost to follow-up. Data on use of diuretics and/or β blockers, use of ACE-inhibitors or angiotensin II receptor blocker, use of statins, prevalence of polycystic kidney disease and nephrosclerosis, previous viral infections (hepatitis C and cytomegalovirus), cumulative dose of steroids before inclusion, but during follow-up, incidence of acute rejection episodes during follow-up and used of mTOR inhibitors at baseline and during follow-up were retrieved from individual patient files. Cumulative dose of prednisolone was calculated as the sum of maintenance dose of prednisolone until inclusion and the dose of prednisolone or methylprednisolone required for treatment of acute rejection (a conversion factor of 1.25 was used to convert methylprednisolone dose to dose of prednisolone).

ual patient files. Cumulative dose of prednisolone was calculated as the sum of maintenance dose of prednisolone until inclusion and the dose of prednisolone or methylprednisolone required for treatment of acute rejection (a conversion factor of 1.25 was used to convert methylprednisolone dose to dose of prednisolone). Statistical analyses Variable distribution was tested with histograms and probability (Q-Q) plots. For descriptive statistics, data are presented as mean and SD when normally distributed, median and IQR when skewed distributed and number and percentage in case of categorical data. Differences between the two MDS groups to test for potential confounders were tested by an unpaired t-test for continuous variables with a normal distribution, Mann-Whitney U test for continuous variables with a skewed distribution and by means of a χ2 test for categorical variables. All statistical analyses were performed using IBM Statistics SPSS V.22.0. For all statistical tests a statistical significance level of p≤0.05 (two-tailed) was used. GraphPad Prism 5 was used to generate the figures.

or continuous variables with a skewed distribution and by means of a χ2 test for categorical variables. All statistical analyses were performed using IBM Statistics SPSS V.22.0. For all statistical tests a statistical significance level of p≤0.05 (two-tailed) was used. GraphPad Prism 5 was used to generate the figures. Primary analyses concerned Kaplan-Meier curves of the incidence of NODAT and all-cause mortality. For NODAT multivariable logistic regression models were used because the exact dates when patients developed an event were not exactly known. For multivariable Cox regression models patients were censored at the date of last follow-up or death. Owing to the small number of NODAT and all-cause mortality the models were first adjusted for age and sex and additionally for more than two potential confounders.27 Models were checked for fulfillment of the assumptions for logistic regression and Cox regression, including the proportional hazards assumption. The assumptions were met.

l number of NODAT and all-cause mortality the models were first adjusted for age and sex and additionally for more than two potential confounders.27 Models were checked for fulfillment of the assumptions for logistic regression and Cox regression, including the proportional hazards assumption. The assumptions were met. Results In total 468 RTR (56.6% men) were included with a mean±SD age of 51.3±13.2 years. The frequency distribution of the MDS in these 468 RTR is presented in figure 1. The MDS varied between 0 (lowest adherence) and 9 (highest adherence), with a mean score of 4.8±1.7 and 54% of the patients had a high score (>4). Baseline characteristics of the overall RTR population and according to high versus low MDS are shown in table 1. Age and physical activity differed significantly between the groups. Patients with a high MDS were older, had a higher physical activity score, lower fasting triglycerides and higher high-density lipoprotein (HDL)-cholesterol concentrations compared to patients with a low MDS. The percentage of smokers and total energy intake did not differ. A borderline statistical significance was found for pre-emptive transplantation, cold ischemia time and use of tacrolimus. Table 1 Baseline characteristics of the overall RTR population and according to groups based on the MDS

Results In total 468 RTR (56.6% men) were included with a mean±SD age of 51.3±13.2 years. The frequency distribution of the MDS in these 468 RTR is presented in figure 1. The MDS varied between 0 (lowest adherence) and 9 (highest adherence), with a mean score of 4.8±1.7 and 54% of the patients had a high score (>4). Baseline characteristics of the overall RTR population and according to high versus low MDS are shown in table 1. Age and physical activity differed significantly between the groups. Patients with a high MDS were older, had a higher physical activity score, lower fasting triglycerides and higher high-density lipoprotein (HDL)-cholesterol concentrations compared to patients with a low MDS. The percentage of smokers and total energy intake did not differ. A borderline statistical significance was found for pre-emptive transplantation, cold ischemia time and use of tacrolimus. Table 1 Baseline characteristics of the overall RTR population and according to groups based on the MDS Overall RTR (n=468) Group 1 MDS 0–4 (n=217) Group 2 MDS 5–9 (n=251) p Value Demographics Age, years 51.3±13.2 49.9±13.9 52.5±12.4 0.03 Male gender, n (%) 265 (56.6) 124 (57.1) 141 (56.2) 0.83 Smoking behavior (current smoker), n (%) 60 (12.8) 29 (13.4) 31 (12.4) 0.64 Total energy intake, kcal/day 2199±656 2168±696 2225±619 0.35 Physical activity score (time×intensity) 5605 (2885–8647) 5060 (2070–8385) 6000 (3480–8700) 0.03 Weight, kg 78.9±15.8 78.1±16.2 79.6±15.5 0.31 Body composition Height, cm 173.9±9.7 173.4±10.6 174.3±9.0 0.30 BMI, kg/m2 26.0±4.5 25.9±4.6 26.1±4.5 0.60 Waist circumference, cm Men 99.1±12.1 98.7±12.2 99.5±12.1 0.98 Women 93.0±15.8 92.6±15.3 93.4±16.3 0.68 Circulation Heart rate, bpm 67.6±12.0 67.8±12.3 67.4±11.8 0.72 SBP, mm Hg 135.3±17.0 135.7±16.3 135.0±17.7 0.68 DBP, mm Hg 83.0±11.0 83.3±11.1 82.8±11.0 0.61 Renal function eGFR, mL/min per 1.73 m2 53.1±20.2 52.3±21.6 53.7±18.8 0.45 Laboratory parameters Triglycerides, mmol/L 1.6 (1.2–2.1) 1.7 (1.2–2.3) 1.5 (1.1–2.0) 0.04 HDL cholesterol, mmol/L 1.4±0.5 1.4±0.4 1.5±0.5 0.001 Fasting glucose, mmol/L 5.1 (4.7–5.5) 5.1 (4.7–5.5) 5.0 (4.7–5.5) 0.51 Hepatitis C virus, n (%) 6 (1.3) 3 (1.4) 3 (1.2) 0.86 Cytomegalovirus, n (%) Primary infection 99 (21.2) 48 (22.1) 51 (20.3) 0.63 Reactivation 80 (17.1) 38 (17.5) 42 (16.7) 0.69 Primary renal disease Polycystic kidney disease, n (%) 103 (22.0) 47 (21.7) 56 (22.3) 0.87 Nephrosclerosis, n (%) 139 (29.7) 67 (30.9) 72 (28.7) 0.61 Transplant characteristics Transplant vintage, years 5.6 (2.1–12.3) 5.2 (2.2–12.3) 5.8 (1.8–12.3) 0.80 Living donor, n (%) 168 (35.9) 70 (32.3) 98 (39.0) 0.20 Pre-emptive transplant, n (%) 84 (17.9) 32 (14.7) 52 (20.7) 0.09 Dialysis duration, months 37.0 (16.0–60.0) 46.0 (15.0–63.0) 32.0 (17.0–56.0) 0.30 Age donor, years 43.0±15.5 42.6±15.0 43.4±15.9 0.61 Cold ischemia time, hours 14.0 (3.0–21.0) 16.0 (3.0–21.0) 12.0 (3.0–21.0) 0.06 Warm ischemia time, minutes 40 (33–50) 42 (33–51) 39 (34–48) 0.25 Acute rejection, n (%) 114 (24.4) 49 (22.6) 65 (25.9) 0.41 Medication Cyclosporine, n (%) 178 (38.0) 82 (37.8) 96 (38.2) 0.92 Tacrolimus, n (%) 79 (16.9) 44 (20.3) 35 (13.9) 0.07 mTOR inhibitor, n (%) 5 (1.1) 2 (0.9) 3 (1.2) 0.76 Predni

0 (3.0–21.0) 0.06 Warm ischemia time, minutes 40 (33–50) 42 (33–51) 39 (34–48) 0.25 Acute rejection, n (%) 114 (24.4) 49 (22.6) 65 (25.9) 0.41 Medication Cyclosporine, n (%) 178 (38.0) 82 (37.8) 96 (38.2) 0.92 Tacrolimus, n (%) 79 (16.9) 44 (20.3) 35 (13.9) 0.07 mTOR inhibitor, n (%) 5 (1.1) 2 (0.9) 3 (1.2) 0.76 Predni solone dose, mg 10.0 (7.5–10.0) 10.0 (7.5–10.0) 10.0 (7.5–10.0) 0.70 Cumulative prednisolone dose, g 18.3 (7.4–38.1) 18.3 (7.8–36.6) 18.2 (7.4–40.4) 0.86 Diuretics, n (%) 158 (33.8) 76 (35.0) 82 (32.7) 0.59 β blocker, n (%) 284 (60.7) 126 (58.1) 158 (62.9) 0.28 ACE inhibitor, n (%) 158 (33.8) 73 (33.6) 85 (33.9) 0.96 Angiotensin II receptor blocker, n (%) 71 (15.2) 39 (18.0) 32 (12.7) 0.12 Statins, n (%) 232 (49.6) 100 (46.1) 132 (52.6) 0.15 Data are represented as mean±SD, median (IQR) or n (%). Differences were tested by t-test or Mann-Whitney U test for continuous variables and with χ2 test for categorical variables. BMI, body mass index; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; HDL, high-density lipoprotein; MDS, Mediterranean Diet Score; RTR, renal transplant recipients; SBP, systolic blood pressure. Figure 1 Frequency distribution of the Mediterranean Diet Score (MDS) in the overall RTR population (468 participants). RTR, renal transplant recipients.

BMI, body mass index; DBP, diastolic blood pressure; eGFR, estimated glomerular filtration rate; HDL, high-density lipoprotein; MDS, Mediterranean Diet Score; RTR, renal transplant recipients; SBP, systolic blood pressure. Figure 1 Frequency distribution of the Mediterranean Diet Score (MDS) in the overall RTR population (468 participants). RTR, renal transplant recipients. Median time between baseline and transplantation was 5.6 (IQR, 2.1–12.3) years. During a median follow-up of 4.0 (IQR, 0.4–5.4) years from baseline, 22 (5%) patients developed NODAT (17 RTR in low MDS group, 5 RTR in high MDS group) and 50 (11%) patients died (29 RTR in low MDS group, 21 RTR in high MDS group). In the low MDS group, 15 (88%) of the RTR that developed NODAT required treatment with hypoglycemic agents or insulin and in the high MDS group, 5 (100%) of the RTR that developed NODAT required treatment with hypoglycemic agents or insulin. Median intake for the nine food groups of the MDS for both men and women are shown in table 2. In the low MDS group men had a higher intake of cereals and alcohol, whereas women had a higher intake of fruit. In the high MDS group men had a higher intake of cereals, whereas women had a high intake of fruit and vegetables. Intake of other food groups is approximately equal between male and female RTR. The high MDS group had a higher intake of legumes, cereals, vegetables, fruit, fish and alcohol compared to the low MDS group. Table 2 Median intake of the components of the Mediterranean Diet Score

Median time between baseline and transplantation was 5.6 (IQR, 2.1–12.3) years. During a median follow-up of 4.0 (IQR, 0.4–5.4) years from baseline, 22 (5%) patients developed NODAT (17 RTR in low MDS group, 5 RTR in high MDS group) and 50 (11%) patients died (29 RTR in low MDS group, 21 RTR in high MDS group). In the low MDS group, 15 (88%) of the RTR that developed NODAT required treatment with hypoglycemic agents or insulin and in the high MDS group, 5 (100%) of the RTR that developed NODAT required treatment with hypoglycemic agents or insulin. Median intake for the nine food groups of the MDS for both men and women are shown in table 2. In the low MDS group men had a higher intake of cereals and alcohol, whereas women had a higher intake of fruit. In the high MDS group men had a higher intake of cereals, whereas women had a high intake of fruit and vegetables. Intake of other food groups is approximately equal between male and female RTR. The high MDS group had a higher intake of legumes, cereals, vegetables, fruit, fish and alcohol compared to the low MDS group. Table 2 Median intake of the components of the Mediterranean Diet Score Group 1 (0–4) Median (IQR) Group 2 (5–9) Median (IQR) Men Women Men Women Ratio monounsaturated: saturated fatty acids 0.9 (0.8–1.0) 0.9 (0.8–1.0) 1.0 (0.9–1.1) 1.0 (0.9–1.1) Legumes, nuts and soy products (g/day) 29 (16–39) 28 (18–40) 52 (38–72) 44 (32–71) Cereals (g/day) 176 (128–240) 134 (107–178) 210 (170–257) 175 (141–206) Fruit (g/day) 77 (34–137) 105 (57–211) 135 (81–234) 211 (97–249) Vegetables (g/day) 57 (32–77) 64 (48–87) 100 (74–140) 124 (92–153) Meat products (g/day) 109 (82–128) 94 (77–116) 90 (70–115) 79 (58–99) Dairy products (g/day) 357 (234–511) 399 (253–492) 330 (211–481) 369 (217–507) Fish (%) 34 34 57 67 Alcohol (%) 67 84 48 80 The Kaplan-Meier survival curves for the associations of the MDS with NODAT (p=0.003, log rank test) and all-cause mortality (p=0.09, log rank test) are shown in figure 2. RTR with MDS scores ≥5 points were significantly associated with a lower risk of NODAT (OR: 0.23; 95% CI 0.09 to 0.64; p=0.004) and all-cause mortality (hazards ratio (HR): 0.51; 95% CI 0.29 to 0.89, p=0.02), adjusted for age and sex (tables 3 and 4). The results of multivariable analyses, in which we additionally adjusted for use of medication, pre-emptive transplantation and cold ischemia time, total energy intake, smoking behavior and physical activity, fasting triglycerides and HDL-cholesterol concentrations and time between transplantation and baseline, did not materially change the results of the analyses adjusted for age and sex (tables 3 and 4).

ation, pre-emptive transplantation and cold ischemia time, total energy intake, smoking behavior and physical activity, fasting triglycerides and HDL-cholesterol concentrations and time between transplantation and baseline, did not materially change the results of the analyses adjusted for age and sex (tables 3 and 4). Table 3 Multiple logistic regression analysis Group 1 (0–4) Group 2 (5–9) 17 (7.8%) 5 (2.0%) Number of events OR (95% CI) p Value Model 1 1.00 (ref) 0.24 (0.09 to 0.66) 0.006 Model 2 1.00 (ref) 0.23 (0.08 to 0.63) 0.004 Model 3 1.00 (ref) 0.22 (0.08 to 0.62) 0.004 Model 4 1.00 (ref) 0.24 (0.08 to 0.69) 0.008 Model 5 1.00 (ref) 0.23 (0.08 to 0.63) 0.004 Model 6 1.00 (ref) 0.23 (0.08 to 0.65) 0.005 Model 7 1.00 (ref) 0.18 (0.06 to 0.54) 0.002 Model 8 1.00 (ref) 0.23 (0.08 to 0.63) 0.004 The Mediterranean diet is associated with a lower risk to develop NODAT. Model 1, crude. Model 2, adjustment for age and sex. Model 3, model 2+adjustment for cyclosporine, tacrolimus and prednisolone dose. Model 4, model 2+adjustment for pre-emptive transplantation and cold ischemia time. Model 5, model 2+adjustment for total energy intake. Model 6, model 2+adjustment for smoking and physical activity. Model 7, model 2+adjustment for triglycerides and HDL-cholesterol concentrations. Model 8, model 2+adjustment for time between transplantation and baseline. HDL, high-density lipoprotein; NODAT, new-onset diabetes after transplantation. Table 4 Cox Regression analysis

Model 5, model 2+adjustment for total energy intake. Model 6, model 2+adjustment for smoking and physical activity. Model 7, model 2+adjustment for triglycerides and HDL-cholesterol concentrations. Model 8, model 2+adjustment for time between transplantation and baseline. HDL, high-density lipoprotein; NODAT, new-onset diabetes after transplantation. Table 4 Cox Regression analysis Group 1 (0–4) Group 2 (5–9) 29 (13.4%) 21 (8.4%) Number of events HR (95% CI) p Value Model 1 1.00 (ref) 0.62 (0.35 to 1.09) 0.09 Model 2 1.00 (ref) 0.51 (0.29 to 0.89) 0.02 Model 3 1.00 (ref) 0.52 (0.29 to 0.92) 0.03 Model 4 1.00 (ref) 0.52 (0.27 to 0.99) 0.05 Model 5 1.00 (ref) 0.51 (0.29 to 0.89) 0.02 Model 6 1.00 (ref) 0.57 (0.22 to 1.03) 0.06 Model 7 1.00 (ref) 0.57 (0.32 to 1.02) 0.06 Model 8 1.00 (ref) 0.50 (0.29 to 0.89) 0.02 The Mediterranean diet is associated with a lower risk of mortality during follow-up. Model 1, crude. Model 2, adjustment for age and sex. Model 3, model 2+adjustment for cyclosporine, tacrolimus and prednisolone dose. Model 4, model 2+adjustment for pre-emptive transplantation and cold ischemia time. Model 5, model 2+adjustment for total energy intake. Model 6, model 2+adjustment for smoking and physical activity. Model 7, model 2+adjustment for triglycerides and HDL-cholesterol concentrations. Model 8, model 2+adjustment for time between transplantation and baseline. HDL, high-density lipoprotein. Figure 2 Kaplan-Meier survival curves. Probability of survival for NODAT (A) and all-cause mortality (B) for both group 1 and group 2. NODAT, new-onset diabetes after transplantation.

Model 7, model 2+adjustment for triglycerides and HDL-cholesterol concentrations. Model 8, model 2+adjustment for time between transplantation and baseline. HDL, high-density lipoprotein. Figure 2 Kaplan-Meier survival curves. Probability of survival for NODAT (A) and all-cause mortality (B) for both group 1 and group 2. NODAT, new-onset diabetes after transplantation. Conclusions About 50% of the RTR had either a low or a high MDS. Patients with a high MDS had a four times lower risk of NODAT and a two times lower risk of all-cause mortality. These results suggest that a healthy diet is of paramount importance for patients who receive a new kidney. Previous studies on diet and chronic diseases often focused on single nutrients. However, food-based dietary patterns take into account complex interactions between food items and are easier to interpret. Another advantage of foods and dietary patterns is that they can directly be transplanted into dietary recommendations to be used in clinical practice.28 29 A meta-analysis of population-based prospective cohort studies showed that greater adherence to the Mediterranean Style diet was associated with a 20% lower all-cause mortality30 and 20% lower cardiovascular risk.31 A meta-analysis of the adherence to the Mediterranean Style diet showed a 25% lower incidence of diabetes mellitus among prospective cohort studies.15

e cohort studies showed that greater adherence to the Mediterranean Style diet was associated with a 20% lower all-cause mortality30 and 20% lower cardiovascular risk.31 A meta-analysis of the adherence to the Mediterranean Style diet showed a 25% lower incidence of diabetes mellitus among prospective cohort studies.15 To the best of our knowledge, this is the first study to investigate the association between a Mediterranean Style diet and risk of NODAT and all-cause mortality in RTR. In line with our results on NODAT another study showed an association between adherence to a Mediterranean Style diet and a lower incidence of metabolic syndrome after transplantation.32 Prospective cohort studies among patients with cardiovascular disease showed that a Mediterranean Style diet was associated with a lower risk of all-cause mortality.33–35 These results suggest that a Mediterranean Style diet is associated with a lower mortality risk in patients with renal transplant and also in patients with cardiovascular disease. This shows the great potential of a healthy diet in secondary prevention.

diet was associated with a lower risk of all-cause mortality.33–35 These results suggest that a Mediterranean Style diet is associated with a lower mortality risk in patients with renal transplant and also in patients with cardiovascular disease. This shows the great potential of a healthy diet in secondary prevention. There are multiple mechanisms that might explain the protective effect of a Mediterranean Style diet on the development of NODAT and mortality in RTR. It is well known that insulin resistance and pancreatic β cell dysfunction are two fundamental features that play an important role in the development of type 2 DM.36 High adherence to the Mediterranean Style diet is associated with a higher intake of antioxidants, dietary fiber, magnesium and unsaturated fatty acids.36 The Mediterranean diet may prevent cardiometabolic diseases through several pathways. First of all, prolonged oxidative stress contributes to the development of insulin resistance and pancreatic β-cell dysfunction.37 A Mediterranean Style diet might have a protective effect on oxidative stress and antioxidant defense, since this dietary pattern is characterized by high intake of fruit and vegetables.38 Second, the high intake of dietary fiber might reduce plasma insulin levels and have an advantageous effect on glucose metabolism.36 Third, magnesium might play an important role in preventing type 2 DM.36 Previous studies showed that high intake of magnesium is associated with a lower risk of developing type 2 diabetes.39 40

the high intake of dietary fiber might reduce plasma insulin levels and have an advantageous effect on glucose metabolism.36 Third, magnesium might play an important role in preventing type 2 DM.36 Previous studies showed that high intake of magnesium is associated with a lower risk of developing type 2 diabetes.39 40 Also fatty acids could play a role in the prevention of cardiometabolic diseases. A high ratio of monounsaturated: saturated fatty acids improves insulin sensitivity.41 A high intake of monounsaturated fatty acids benefits glycemic control, since it stimulates the secretion of glucagon-like peptide-1 (GLP-1), an antidiabetic hormone.42 GLP-1 activates the GLP-1 receptor in the pancreatic islets, which leads to an increase in secretion of insulin and inhibition of glucagon.43 Furthermore, GLP-1 plays a role in satiety.44 Finally, pathological processes as inflammation and endothelial dysfunction play a role in the etiology of cardiovascular events.45 46 A previous study showed that higher adherence to a Mediterranean Style diet is associated with a lower concentration of biomarkers for inflammation and endothelial dysfunction; C reactive protein, interleukin-6, E-selectin and soluble intercellular cell adhesion molecule-1.47 Furthermore, the literature shows that olive oil, vegetables, cereals and nuts have antithrombotic and/or anticoagulant effects.46 Adherence to a Mediterranean Style diet is associated with lower levels of prothrombotic biomarkers, for example, fibrogen,48 which contributes to a lower cardiovascular risk as well.

-1.47 Furthermore, the literature shows that olive oil, vegetables, cereals and nuts have antithrombotic and/or anticoagulant effects.46 Adherence to a Mediterranean Style diet is associated with lower levels of prothrombotic biomarkers, for example, fibrogen,48 which contributes to a lower cardiovascular risk as well. Our study has several limitations. Although this is a prospective cohort study, causality of the associations cannot be assumed, since this study is of observational nature. Also the number of NODAT cases and the number of deaths is small. Furthermore, the FFQ was originally developed to examine protein intake in RTR. It was only validated by comparing the protein intake of the FFQ with the protein intake calculated by the Maroni Equation, using urinary urea excretion values.19 The Mediterranean Style diet of our RTR is not optimal and may further lower the risk of NODAT and all-cause mortality through a better adherence to the traditional Mediterranean diet. It is the first time the association between the Mediterranean Style diet and NODAT and all-cause mortality is investigated in RTR. Furthermore, strengths include a complete follow-up of the clinically relevant end points: NODAT and all-cause mortality. In conclusion, our prospective cohort study suggests that higher adherence to a Mediterranean Style diet may prevent the development of NODAT and all-cause mortality in RTR. More attention is needed for the nutritional habits of RTR.

Our study has several limitations. Although this is a prospective cohort study, causality of the associations cannot be assumed, since this study is of observational nature. Also the number of NODAT cases and the number of deaths is small. Furthermore, the FFQ was originally developed to examine protein intake in RTR. It was only validated by comparing the protein intake of the FFQ with the protein intake calculated by the Maroni Equation, using urinary urea excretion values.19 The Mediterranean Style diet of our RTR is not optimal and may further lower the risk of NODAT and all-cause mortality through a better adherence to the traditional Mediterranean diet. It is the first time the association between the Mediterranean Style diet and NODAT and all-cause mortality is investigated in RTR. Furthermore, strengths include a complete follow-up of the clinically relevant end points: NODAT and all-cause mortality. In conclusion, our prospective cohort study suggests that higher adherence to a Mediterranean Style diet may prevent the development of NODAT and all-cause mortality in RTR. More attention is needed for the nutritional habits of RTR. The cohort on which the study was based is registered at clinicaltrials.gov as “TransplantLines Food and Nutrition Biobank and Cohort Study (TxL-FN)” with number NCT02811835.

In conclusion, our prospective cohort study suggests that higher adherence to a Mediterranean Style diet may prevent the development of NODAT and all-cause mortality in RTR. More attention is needed for the nutritional habits of RTR. The cohort on which the study was based is registered at clinicaltrials.gov as “TransplantLines Food and Nutrition Biobank and Cohort Study (TxL-FN)” with number NCT02811835. Contributors: MCJO analyzed the data and wrote the first draft of the paper. EC, GJN, CAK, DK and SJLB contributed to the interpretation of the results and important intellectual content. MCJO, EB, CAK and MHdB collaborated in the data collection. All authors had access to the data, contributed to the critical revision of the manuscript and approved the final version of the manuscript. Funding: This work was supported by a grant from the Dutch Top Institute Food and Nutrition (A-1003). Competing interests: None declared. Ethics approval: METc 2008/186. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: No additional data are available.

Significance of this study What is already known about this subject? There is a large and growing burden of community-acquired lower respiratory tract infection and pneumonia among older people with diabetes, much of which can be directly or indirectly attributed to two vaccine-preventable pathogens: pneumococcus and influenza. What are the new findings? We observed only a modest effectiveness of influenza vaccine against community-acquired lower respiratory tract infection (after adjustment for confounding by indication), while the effectiveness of pneumococcal vaccine against pneumonia waned over time. Our results suggested that pneumococcal vaccine effectiveness may be lower among patients with proteinuria but did not otherwise vary according to markers of chronic kidney disease. How might these results change the focus of research or clinical practice? More effective immunization strategies and vaccination schedules may be needed for older people with diabetes. The low influenza vaccine effectiveness (VE) we observed against community-acquired lower respiratory tract infection, when contrasted with the large burden of infection directly and indirectly attributed to influenza, suggests scope for improved influenza immunization among this population, for example, with adjuvants. The suggestion of reduced pneumococcal VE among patients with proteinuria needs confirmation in a repeat study.

The low influenza vaccine effectiveness (VE) we observed against community-acquired lower respiratory tract infection, when contrasted with the large burden of infection directly and indirectly attributed to influenza, suggests scope for improved influenza immunization among this population, for example, with adjuvants. The suggestion of reduced pneumococcal VE among patients with proteinuria needs confirmation in a repeat study. Introduction Hospital admissions for pneumonia are rising rapidly in the UK, most steeply among older people.1 Older people with diabetes have a particularly high burden of lower respiratory tract infection (LRTI), including pneumonia.2 Directly or indirectly, Streptococcus pneumoniae (‘pneumococcus’) and seasonal influenza viruses are responsible for a large burden of community-acquired pneumonia. The most common cause of community-acquired pneumonia is S. pneumoniae.3 Up to a third of community-acquired pneumonia may be influenza-related, due to bacterial coinfection or secondary bacterial pneumonia.4 Vaccination is available against both these pathogens, and recommended in the UK for everyone aged ≥65 years.5 However, the extent to which these vaccines protect against pneumonia among older people remains unclear for both vaccines.

be influenza-related, due to bacterial coinfection or secondary bacterial pneumonia.4 Vaccination is available against both these pathogens, and recommended in the UK for everyone aged ≥65 years.5 However, the extent to which these vaccines protect against pneumonia among older people remains unclear for both vaccines. The effectiveness of 23-valent pneumococcal polysaccharide vaccination against all-cause pneumonia among older people has been questioned, although meta-analyses have been hampered by between-study heterogeneity.6 7 Waning immunity among vaccinated participants has been suggested as a possible cause, but few estimates are available of pneumococcal vaccine effectiveness (VE) according to time since vaccination.8 Traditional observational studies of influenza VE among older people may have overestimated influenza VE due to uncontrolled confounding by indication, in which the patient's functional status affects vaccine uptake.8–12 Observational studies which used strategies to control confounding by indication (such as a ‘ratio-of-ratios’ analysis in which the excess influenza VE during winter compared with summer is calculated) have suggested a null or modest influenza VE against community-acquired pneumonia among older people.9 13–15

take.8–12 Observational studies which used strategies to control confounding by indication (such as a ‘ratio-of-ratios’ analysis in which the excess influenza VE during winter compared with summer is calculated) have suggested a null or modest influenza VE against community-acquired pneumonia among older people.9 13–15 Older people with diabetes have a high prevalence of chronic kidney disease (CKD).16 Even at early stages, patients with CKD have increased incidence of LRTI and pneumonia.16–18 Patients with CKD have a generally reduced response to vaccines, and a faster decline in antibody levels following vaccination.19 Among patients receiving hemodialysis, a ratio-of-ratios analysis of influenza VE found no evidence of any protection against influenza-like-illness, influenza/pneumonia hospitalization, or mortality.20 Influenza VE at earlier stages of CKD is unclear, and still less is known about pneumococcal VE among patients with CKD.19 21

Research design and methods Data sources We analyzed data from the Clinical Practice Research Datalink (CPRD), a database of anonymized primary care medical records. Data were extracted in May 2011, and contained records for 12.8 million patients at 627 practices across the UK.22 Records include patient demographics, health behaviors, test results, diagnoses, and prescriptions. Diagnoses are recorded using Read codes, and have generally been found to have good positive predictive value in validations.23 The CPRD population is similar to the general UK population in terms of age and sex.24 25 Linked data are available for patients in England, subject to practice-level consent. This study used linked data on all hospital inpatient admissions to NHS hospitals in England from Hospital Episodes Statistics (HES), and socioeconomic status from the Office for National Statistics (ONS). Study population The study population comprised all patients in CPRD with diabetes mellitus, aged ≥65 years, with no history of renal replacement therapy, who had at least one valid serum creatinine result recorded in primary care. Diabetes was identified by diagnostic Read codes. For less definitive Read codes, we required confirmation with an antidiabetic medication prescription, as described in detail previously.2

5 years, with no history of renal replacement therapy, who had at least one valid serum creatinine result recorded in primary care. Diabetes was identified by diagnostic Read codes. For less definitive Read codes, we required confirmation with an antidiabetic medication prescription, as described in detail previously.2 Patients met eligibility criteria at the latest time-point of: diabetes diagnosis, 65th birthday, 1 year after practice registration, their general practice fulfilling CPRD quality control standards, or 1 April 1997. Their study entry date was their first valid serum creatinine result after the eligibility criteria were met. Patients left the study at the first time-point of: death, leaving the practice, last data collection from the practice, renal replacement therapy (dialysis or renal transplant), or 31 March 2011. Patients with a diagnosis of HIV or hyposplenia (including celiac disease or sickle cell disease) at any point in their medical record were excluded from the study. Definition of infections LRTI was defined as a broad category of all infections of the lower respiratory tract, including influenza infections, bronchitis, and pneumonia.

Patients met eligibility criteria at the latest time-point of: diabetes diagnosis, 65th birthday, 1 year after practice registration, their general practice fulfilling CPRD quality control standards, or 1 April 1997. Their study entry date was their first valid serum creatinine result after the eligibility criteria were met. Patients left the study at the first time-point of: death, leaving the practice, last data collection from the practice, renal replacement therapy (dialysis or renal transplant), or 31 March 2011. Patients with a diagnosis of HIV or hyposplenia (including celiac disease or sickle cell disease) at any point in their medical record were excluded from the study. Definition of infections LRTI was defined as a broad category of all infections of the lower respiratory tract, including influenza infections, bronchitis, and pneumonia. A clinical diagnosis of infection was identified by a diagnostic Read code in primary care records, or a diagnostic International Classification of Disease 10 (ICD-10) code as the primary cause of hospital admission in secondary care records. To avoid overestimation from repeat attendances for the same infection, diagnostic codes recorded within 28 days of one another were attributed to a single episode of infection. The first consultation for infection was treated as the date of infection onset, and the infection had duration until 28 days after the latest of the last diagnostic code or hospital discharge. All infections with onset date during a HES hospitalization spell, or within 14 days following hospital discharge, or which included a code for postoperative infection, were designated hospital-acquired, and excluded. These methods have been described in detail previously.26

last diagnostic code or hospital discharge. All infections with onset date during a HES hospitalization spell, or within 14 days following hospital discharge, or which included a code for postoperative infection, were designated hospital-acquired, and excluded. These methods have been described in detail previously.26 Time at risk Patients were not at risk of incident community-acquired infection during ongoing infection (community-acquired or hospital-acquired), during any hospitalization, or within 14 days following hospital discharge. These time periods were removed from time at risk. As pneumonia was a subset of LRTI, a patient could be at risk of pneumonia during an ongoing LRTI. Assignment of vaccination status Vaccination status was identified from primary care records using Read codes, prescription data, and immunization record forms. For pneumococcal vaccination, any of these records could define a first vaccination, and any subsequent prescription could identify a booster vaccination. Time-updated pneumococcal vaccination status was classified according to time since the latest pneumococcal vaccination (<1, 1–5, ≥5 years, never vaccinated).

Assignment of vaccination status Vaccination status was identified from primary care records using Read codes, prescription data, and immunization record forms. For pneumococcal vaccination, any of these records could define a first vaccination, and any subsequent prescription could identify a booster vaccination. Time-updated pneumococcal vaccination status was classified according to time since the latest pneumococcal vaccination (<1, 1–5, ≥5 years, never vaccinated). Time-updated influenza vaccination status was assigned within vaccination years (1 September to 31 August). Within each vaccination year, influenza vaccination status was current from the first vaccination record to the subsequent 31 August. Patients without a current vaccination who had received an influenza vaccination within any of the previous five vaccination years were classified as having ‘residual’ influenza vaccination status, and other patients were categorized as unvaccinated. Definition of CKD We studied two markers of CKD: estimated glomerular filtration rate (eGFR) and proteinuria. Estimated GFR was calculated from serum creatinine test results in primary care, using the CKD-EPI equation, including adjustment for black ethnicity.27 Estimated GFR status was time-updated using a last-carried-forward method, with eGFR status assigned according to the most recent creatinine result.17 A history of proteinuria was established from a Read code for persistent proteinuria or proteinuric disease, or a positive test result which did not coincide with a urinary tract infection diagnosis.

Definition of CKD We studied two markers of CKD: estimated glomerular filtration rate (eGFR) and proteinuria. Estimated GFR was calculated from serum creatinine test results in primary care, using the CKD-EPI equation, including adjustment for black ethnicity.27 Estimated GFR status was time-updated using a last-carried-forward method, with eGFR status assigned according to the most recent creatinine result.17 A history of proteinuria was established from a Read code for persistent proteinuria or proteinuric disease, or a positive test result which did not coincide with a urinary tract infection diagnosis. Definitions for covariates Age was categorized in 5-year bands up to a final category of ≥85 years. Socioeconomic status was assigned at a practice level, using 2007 ONS estimates of the Index of Multiple Deprivation, a composite area-level marker of deprivation.28 Smoking status was identified as current, ex-smoker, or non-smoker from HES or CPRD records. Comorbidities were identified from diagnostic Read codes in CPRD and were modeled as separate variables which were: ischemic heart disease, congestive cardiac failure, hypertension, cerebrovascular disease, other dementia, chronic lung disease (which included chronic obstructive pulmonary disease but not asthma), and chronic liver disease. Baseline HbA1C was defined by the most recent HbA1C test result in CPRD prior to (or on) the study entry date. Baseline medication history was identified from CPRD prescription records.

other dementia, chronic lung disease (which included chronic obstructive pulmonary disease but not asthma), and chronic liver disease. Baseline HbA1C was defined by the most recent HbA1C test result in CPRD prior to (or on) the study entry date. Baseline medication history was identified from CPRD prescription records. Data analysis Analysis was conducted separately for pneumococcal VE against pneumonia and for influenza VE against LRTI. We excluded patients with missing smoking status. For comorbidities and proteinuria status, the absence of a positive record was treated as the absence of disease. The absence of a recorded HbA1C test result was included as indicating a relevant category of control.

Data analysis Analysis was conducted separately for pneumococcal VE against pneumonia and for influenza VE against LRTI. We excluded patients with missing smoking status. For comorbidities and proteinuria status, the absence of a positive record was treated as the absence of disease. The absence of a recorded HbA1C test result was included as indicating a relevant category of control. Incidence rates and rate ratios were calculated for each infection using the Poisson regression with lexis expansions for age, and a random effects model to adjust for multiple infection episodes. We adjusted models for prespecified a priori confounders of the association between vaccination status and respiratory infection, and/or the relationship between CKD and respiratory infection. These were: age, sex, socioeconomic status at practice level, residential or nursing home care, baseline smoking status, time-updated comorbidities, steroid use in the 3 months prior to study entry, HbA1C and diabetic medication history at baseline, and date prior to or post 1 April 2004 (when Quality Outcomes Framework guidelines introduced financial incentives for recording CKD status among people with diabetes in primary care which may have improved ascertainment of CKD in primary care).29 No direct biological effect of ethnicity on VE was expected and so we did not adjust for this directly: instead, we adjusted for factors which may mediate any indirect confounding effect of ethnicity, such as CKD and other comorbidities and health behaviors.

ch may have improved ascertainment of CKD in primary care).29 No direct biological effect of ethnicity on VE was expected and so we did not adjust for this directly: instead, we adjusted for factors which may mediate any indirect confounding effect of ethnicity, such as CKD and other comorbidities and health behaviors. For pneumococcal vaccine, VE was calculated as (1−effect estimate). To explore waning of immunity, we described pneumococcal VE according to time since vaccination. To control for confounding by indication in influenza vaccination, we estimated the ratio of influenza VE in summer to influenza VE in winter in a ‘ratio-of-ratios’ analysis by including an interaction term between influenza vaccination status and season, and reporting the antilog of the β coefficient for the interaction term.20 Winter was defined as 1 September to 31 March, to capture excess winter influenza-like-illness.30 Final estimates of VE were stratified by time-updated eGFR and history of proteinuria, as markers of CKD. Stata V.13.1 was used for data analyses. All code lists are available on request. Sensitivity analyses Twenty-three-valent pneumococcal polysaccharide vaccination has been recommended for patients with CKD in the UK since 1992, but in 2003, the recommendation was extended to everyone aged ≥65 years.5 As a sensitivity analysis, we estimated pneumococcal VE separately for the periods before and after 31 March 2003 (to avoid separating the 2002–2003 winter season) to check for bias from secular changes in vaccine uptake.

KD in the UK since 1992, but in 2003, the recommendation was extended to everyone aged ≥65 years.5 As a sensitivity analysis, we estimated pneumococcal VE separately for the periods before and after 31 March 2003 (to avoid separating the 2002–2003 winter season) to check for bias from secular changes in vaccine uptake. The match of influenza vaccine strain to circulating influenza varies each year, which affects VE.20 As a sensitivity analysis, we estimated influenza VE separately for each winter. A further sensitivity analysis defined the start of the influenza season as the first week after 1 September in which weekly influenza-like illness incidence in primary care exceeded 30/100 000 people, limited to the years 2004–2011 due to data availability.31 We also conducted a sensitivity analysis of influenza VE excluding patients with chronic lung disease or congestive heart failure, as the relationship of influenza to LRTI etiology for these patients may differ from that among the general population. For both vaccines, we conducted a sensitivity analysis limited to patients with white ethnicity. Ethics The study was approved by the Independent Scientific Advisory Group of the CPRD (ISAC reference 11_033A) and the London School of Hygiene and Tropical Medicine Ethics Committee (LSHTM reference 6116).

We also conducted a sensitivity analysis of influenza VE excluding patients with chronic lung disease or congestive heart failure, as the relationship of influenza to LRTI etiology for these patients may differ from that among the general population. For both vaccines, we conducted a sensitivity analysis limited to patients with white ethnicity. Ethics The study was approved by the Independent Scientific Advisory Group of the CPRD (ISAC reference 11_033A) and the London School of Hygiene and Tropical Medicine Ethics Committee (LSHTM reference 6116). Results Of 193 470 eligible patients, 1049 patients (0.5%) with a diagnosis of HIV or hyposplenia, 1764 (0.9%) patients with no smoking status available, and 165 (<0.1%) patients who had a record of pneumococcal vaccine administration with a missing date were excluded from both analyses (figure 1). For pneumococcal and influenza vaccinations, unvaccinated patients had a lower recorded prevalence of ischemic heart disease and chronic lung disease than vaccinated patients. Unvaccinated patients may have had poorer diabetic control than vaccinated patients: a higher proportion had poor or unrecorded HbA1C status, and a lower proportion had a history of oral antidiabetic medication and insulin prescription than vaccinated patients. The prevalence of CKD was similar for vaccinated and unvaccinated patients, although unvaccinated patients had a slightly lower prevalence of a recorded history of proteinuria (table 1). Table 1 Baseline description of study population

Results Of 193 470 eligible patients, 1049 patients (0.5%) with a diagnosis of HIV or hyposplenia, 1764 (0.9%) patients with no smoking status available, and 165 (<0.1%) patients who had a record of pneumococcal vaccine administration with a missing date were excluded from both analyses (figure 1). For pneumococcal and influenza vaccinations, unvaccinated patients had a lower recorded prevalence of ischemic heart disease and chronic lung disease than vaccinated patients. Unvaccinated patients may have had poorer diabetic control than vaccinated patients: a higher proportion had poor or unrecorded HbA1C status, and a lower proportion had a history of oral antidiabetic medication and insulin prescription than vaccinated patients. The prevalence of CKD was similar for vaccinated and unvaccinated patients, although unvaccinated patients had a slightly lower prevalence of a recorded history of proteinuria (table 1). Table 1 Baseline description of study population Pneumococcal vaccine status at baseline n=190 492 Influenza vaccine status at baseline n=190 459 Never vaccinated n=79 476 Vaccinated n=111 016 Unvaccinated* n=32 552 Currently vaccinated n=124 130 Residual 1–5 years n=33 777 Age (years) 71 (66–77) 72 (66–78) 71 (66–77) 72 (66–78) 71 (66–78) Median (IQR) n (%) n (%) n (%) n (%) n (%) Female 40 308 (50.7) 53 146 (47.9) 16 603 (51.0) 60 019 (48.4) 16 813 (49.8) Socioeconomic status† 1 (least deprived) 13 701 (17.2) 19 912 (17.9) 5618 (17.3) 22 181 (17.9) 5809 (17.2) 2 14 666 (18.5) 19 591 (17.7) 5799 (17.8) 22 394 (18.0) 6058 (17.9) 3 16 156 (20.3) 23 329 (21.0) 6567 (20.2) 25 957 (20.9) 6956 (20.6) 4 17 758 (22.3) 25 481 (23.0) 7312 (22.5) 28 128 (22.7) 7789 (23.1) 5 (most deprived) 17 195 (21.6) 22 703 (20.5) 7256 (22.3) 25 470 (20.5) 7165 (21.2) Ethnicity White 43 357 (54.6) 63 577 (57.3) 17 037 (52.3) 71 268 (57.4) 18 615 (55.1) South Asian 1515 (1.9) 2353 (2.1) 429 (1.3) 2606 (2.1) 833 (2.5) Black 923 (1.2) 1167 (1.1) 343 (1.1) 1290 (1.0) 457 (1.4) Other 636 (0.8) 717 (0.7) 225 (0.7) 860 (0.7) 267 (0.8) Missing 33 045 (41.6) 43 202 (38.9) 14 518 (44.6) 48 106 (38.8) 13 605 (40.3) Residential care 1697 (2.1) 3274 (3.0) 436 (1.3) 3516 (2.8) 1016 (3.0) Smoking status Non-smoker 38 078 (47.9) 44 713 (40.3) 15 449 (47.5) 52 782 (42.5) 14 543 (43.1) Current smoker 13 901 (17.5) 16 439 (14.8) 6252 (19.2) 18 327 (14.8) 5756 (17.0) Ex-smoker 27 497 (34.6) 49 864 (44.9) 10 851 (33.3) 53 021 (42.7) 13 478 (39.9) Comorbidities Ischemic heart disease 18 886 (23.8) 34 415 (31.0) 6825 (21.0) 36 761 (29.6) 9713 (28.8) Congestive cardiac failure 5935 (7.5) 10 018 (9.0) 2175 (6.7) 10 721 (8.6) 3065 (9.1) Hypertension 46 626 (58.7) 71 311 (64.2) 18 644 (57.3) 78 252 (63.0) 21 024 (62.2) Cerebrovascular disease 9714 (12.2) 14 469 (13.0) 3612 (11.1) 16 021 (12.9) 4540 (13.4) Other dementia 1437 (1.8) 1956 (1.8) 322 (1.0) 2326 (1.9) 729 (2.2) Chronic lung disease 4016 (5.1) 10 881 (9.8) 1515 (4.7) 10 530 (8.5) 2851 (8.4) Chronic liver disease 402 (0.5) 734 (0.7) 168 (0.5) 729 (0.6) 240 (0.7) Steroid use in previous 3 months 2870 (3.6) 5560 (5.0) 1039 (3.2) 5841 (4.7) 1544 (4.6) Latest HbA1C status % (mmol/mol) None recorded 11 202

(1.0) 2326 (1.9) 729 (2.2) Chronic lung disease 4016 (5.1) 10 881 (9.8) 1515 (4.7) 10 530 (8.5) 2851 (8.4) Chronic liver disease 402 (0.5) 734 (0.7) 168 (0.5) 729 (0.6) 240 (0.7) Steroid use in previous 3 months 2870 (3.6) 5560 (5.0) 1039 (3.2) 5841 (4.7) 1544 (4.6) Latest HbA1C status % (mmol/mol) None recorded 11 202 (14.1) 10 620 (9.6) 4872 (15.0) 13 317 (10.7) 3627 (10.7) Good <7% (<53) 34 669 (43.6) 53 305 (48.0) 13 621 (41.8) 58 741 (47.3) 15 596 (46.2) Intermediate 7–10% (53–86) 27 935 (35.2) 41 354 (37.3) 11 389 (35.0) 45 383 (36.6) 12 509 (37.0) Poor >10% (>86) 5670 (7.1) 5737 (5.2) 2670 (8.2) 6689 (5.4) 2045 (6.1) Antidiabetes medication history None 38 755 (48.8) 50 517 (45.5) 16 463 (50.6) 58 195 (46.9) 14 598 (43.2) Oral 33 623 (42.3) 46 949 (42.3) 13 613 (41.8) 51 914 (41.8) 15 031 (44.5) Insulin 2889 (3.6) 4136 (3.7) 1024 (3.2) 4687 (3.8) 1314 (3.9) Oral and insulin 4209 (5.3) 9414 (8.5) 1452 (4.5) 9334 (7.5) 2834 (8.4) Latest eGFR mL/min/1.73 m2 <30 2098 (2.6) 2986 (2.7) 767 (2.4) 3337 (2.7) 977 (2.9) 30–44 7558 (9.5) 10 607 (9.6) 2932 (9.0) 11 964 (9.6) 3254 (9.6) 45–59 18 678 (23.5) 25 508 (23.0) 7337 (22.5) 29 039 (23.4) 7806 (23.1) ≥60 51 142 (64.4) 71 915 (64.8) 21 516 (66.1) 79 790 (64.3) 21 740 (64.4) History of proteinuria No 71 095 (89.5) 94 128 (84.8) 29 231 (89.8) 107 212 (86.4) 28 735 (85.1) Yes 8381 (10.6) 16 888 (15.2) 3321 (10.2) 16 918 (13.6) 5042 (14.9) *Not vaccinated within the five previous years. †Index of multiple deprivation quintile for primary care practice. eGFR, estimated glomerular filtration rate.

(14.1) 10 620 (9.6) 4872 (15.0) 13 317 (10.7) 3627 (10.7) Good <7% (<53) 34 669 (43.6) 53 305 (48.0) 13 621 (41.8) 58 741 (47.3) 15 596 (46.2) Intermediate 7–10% (53–86) 27 935 (35.2) 41 354 (37.3) 11 389 (35.0) 45 383 (36.6) 12 509 (37.0) Poor >10% (>86) 5670 (7.1) 5737 (5.2) 2670 (8.2) 6689 (5.4) 2045 (6.1) Antidiabetes medication history None 38 755 (48.8) 50 517 (45.5) 16 463 (50.6) 58 195 (46.9) 14 598 (43.2) Oral 33 623 (42.3) 46 949 (42.3) 13 613 (41.8) 51 914 (41.8) 15 031 (44.5) Insulin 2889 (3.6) 4136 (3.7) 1024 (3.2) 4687 (3.8) 1314 (3.9) Oral and insulin 4209 (5.3) 9414 (8.5) 1452 (4.5) 9334 (7.5) 2834 (8.4) Latest eGFR mL/min/1.73 m2 <30 2098 (2.6) 2986 (2.7) 767 (2.4) 3337 (2.7) 977 (2.9) 30–44 7558 (9.5) 10 607 (9.6) 2932 (9.0) 11 964 (9.6) 3254 (9.6) 45–59 18 678 (23.5) 25 508 (23.0) 7337 (22.5) 29 039 (23.4) 7806 (23.1) ≥60 51 142 (64.4) 71 915 (64.8) 21 516 (66.1) 79 790 (64.3) 21 740 (64.4) History of proteinuria No 71 095 (89.5) 94 128 (84.8) 29 231 (89.8) 107 212 (86.4) 28 735 (85.1) Yes 8381 (10.6) 16 888 (15.2) 3321 (10.2) 16 918 (13.6) 5042 (14.9) *Not vaccinated within the five previous years. †Index of multiple deprivation quintile for primary care practice. eGFR, estimated glomerular filtration rate. Figure 1 Flow chart of study inclusion. LRTI, lower respiratory tract infection.

(14.1) 10 620 (9.6) 4872 (15.0) 13 317 (10.7) 3627 (10.7) Good <7% (<53) 34 669 (43.6) 53 305 (48.0) 13 621 (41.8) 58 741 (47.3) 15 596 (46.2) Intermediate 7–10% (53–86) 27 935 (35.2) 41 354 (37.3) 11 389 (35.0) 45 383 (36.6) 12 509 (37.0) Poor >10% (>86) 5670 (7.1) 5737 (5.2) 2670 (8.2) 6689 (5.4) 2045 (6.1) Antidiabetes medication history None 38 755 (48.8) 50 517 (45.5) 16 463 (50.6) 58 195 (46.9) 14 598 (43.2) Oral 33 623 (42.3) 46 949 (42.3) 13 613 (41.8) 51 914 (41.8) 15 031 (44.5) Insulin 2889 (3.6) 4136 (3.7) 1024 (3.2) 4687 (3.8) 1314 (3.9) Oral and insulin 4209 (5.3) 9414 (8.5) 1452 (4.5) 9334 (7.5) 2834 (8.4) Latest eGFR mL/min/1.73 m2 <30 2098 (2.6) 2986 (2.7) 767 (2.4) 3337 (2.7) 977 (2.9) 30–44 7558 (9.5) 10 607 (9.6) 2932 (9.0) 11 964 (9.6) 3254 (9.6) 45–59 18 678 (23.5) 25 508 (23.0) 7337 (22.5) 29 039 (23.4) 7806 (23.1) ≥60 51 142 (64.4) 71 915 (64.8) 21 516 (66.1) 79 790 (64.3) 21 740 (64.4) History of proteinuria No 71 095 (89.5) 94 128 (84.8) 29 231 (89.8) 107 212 (86.4) 28 735 (85.1) Yes 8381 (10.6) 16 888 (15.2) 3321 (10.2) 16 918 (13.6) 5042 (14.9) *Not vaccinated within the five previous years. †Index of multiple deprivation quintile for primary care practice. eGFR, estimated glomerular filtration rate. Figure 1 Flow chart of study inclusion. LRTI, lower respiratory tract infection. Pneumococcal vaccine A total of 190 492 patients contributed 811 498 person-years to the pneumococcal vaccine analysis, during which there were 7805 community-acquired pneumonia episodes among 7036 people. At study entry, 58.3% of patients (111 016/190 492) were vaccinated against pneumococcal disease (table 1). Baseline pneumococcal vaccination increased among patients who entered the study after 2003–2004, and did not differ according to eGFR at baseline (see online supplementary figure S1A).

pisodes among 7036 people. At study entry, 58.3% of patients (111 016/190 492) were vaccinated against pneumococcal disease (table 1). Baseline pneumococcal vaccination increased among patients who entered the study after 2003–2004, and did not differ according to eGFR at baseline (see online supplementary figure S1A). 10.1136/bmjdrc-2016-000332.supp1supplementary data Crude rates of pneumonia were lowest among patients within a year of pneumococcal vaccine. The adjusted effectiveness of pneumococcal vaccine for preventing pneumonia was 22% (95% CI 11% to 31%) within the first year after vaccination, and fell with increasing time since vaccination. Pneumonia incidence among patients vaccinated more than 5 years previously was similar to that among patients with no record of vaccination (incidence rate ratio, IRR 1.03: 95% CI 0.95 to 1.11). There was the suggestion of a trend of decreased pneumococcal VE among patients with reduced eGFR, but this was not statistically significant. There was a greater protective effect of pneumococcal vaccine among patients without a history of proteinuria than with a history of proteinuria (table 2). Table 2 Pneumococcal vaccine effectiveness against pneumonia (n=190 492)

Crude rates of pneumonia were lowest among patients within a year of pneumococcal vaccine. The adjusted effectiveness of pneumococcal vaccine for preventing pneumonia was 22% (95% CI 11% to 31%) within the first year after vaccination, and fell with increasing time since vaccination. Pneumonia incidence among patients vaccinated more than 5 years previously was similar to that among patients with no record of vaccination (incidence rate ratio, IRR 1.03: 95% CI 0.95 to 1.11). There was the suggestion of a trend of decreased pneumococcal VE among patients with reduced eGFR, but this was not statistically significant. There was a greater protective effect of pneumococcal vaccine among patients without a history of proteinuria than with a history of proteinuria (table 2). Table 2 Pneumococcal vaccine effectiveness against pneumonia (n=190 492) Pneumococcal vaccination status Never <1 year 1–4 years ≥5 years Person-time (years) 189 776 51 397 275 841 294 484 Infections (n) 1661 326 2255 3563 Crude pneumonia rate/1000 person-years (95% CI) 9.0 (8.6 to 9.5) 6.6 (5.9 to 7.3) 8.7 (8.3 to 9.1) 13.6 (13.1 to 14.1) Adjusted* pneumonia rate ratio (95% CI) 1 (reference) 0.78 (0.69 to 0.89) 0.92 (0.85 to 0.99) 1.03 (0.95 to 1.11) Vaccine effectiveness* % (95% CI) 0 (reference) 22 (11 to 31) 8 (1 to 15) −3 (−11 to 5) Vaccine effectiveness* % (95% CI) stratified by eGFR status (mL/min/1.73 m2) eGFR <30 0 (reference) 6 (−40 to 37) 4 (−22 to 25) 6 (−19 to 26) eGFR 30–44 0 (reference) 16 (−12 to 37) 1 (−18 to 17) −7 (−27 to 11) eGFR 45–59 0 (reference) 21 (−1 to 38) 9 (−6 to 21) −1 (−17 to 14) eGFR ≥60 0 (reference) 26 (11 to 39) 12 (2 to 22) −3 (−15 to 8) p Value (test for trend)† – 0.25 0.49 0.07 Vaccine effectiveness* % (95% CI) stratified by proteinuria status No proteinuria 0 (reference) 28 (16 to 38) 13 (5 to 20) 1 (−8 to 10) Proteinuria 0 (reference) 2 (−25 to 23) −6 (−23 to 9) −19 (−38 to −3) p Value (interaction)‡ – 0.04 0.03 0.04 *Adjusted for: age, sex, socioeconomic status at practice level, residential care, date post 1 April 2004, smoking status, time-updated comorbidities (ischemic heart disease, congestive cardiac failure, hypertension, cerebrovascular disease, other dementia, chronic lung disease, chronic liver disease), time-updated CKD status (eGFR, proteinuria), steroid use in the 3 months prior to study entry, influenza vaccination status, and HbA1C and diabetic medication history at baseline.

rt disease, congestive cardiac failure, hypertension, cerebrovascular disease, other dementia, chronic lung disease, chronic liver disease), time-updated CKD status (eGFR, proteinuria), steroid use in the 3 months prior to study entry, influenza vaccination status, and HbA1C and diabetic medication history at baseline. †Wald test for interaction term of pneumococcal vaccine with eGFR. ‡Wald test for interaction term of pneumococcal vaccine with proteinuria. A sensitivity analysis of pneumococcal VE stratified by date before or after 1 April 2003 suggested that the estimate was not affected by the change in vaccine recommendation in 2003 (see online supplementary table S1). Influenza vaccine For the influenza VE analysis, 190 459 patients contributed 803 230 person-years to time at risk, during which there were 114 313 cases of LRTI among 55 685 patients. At study entry, 65.2% of patients (124 130/190 459) had received a current vaccination against influenza (table 1). Baseline influenza vaccination status increased slightly over time, and did not differ by eGFR status (see online supplementary figure S1B).

g which there were 114 313 cases of LRTI among 55 685 patients. At study entry, 65.2% of patients (124 130/190 459) had received a current vaccination against influenza (table 1). Baseline influenza vaccination status increased slightly over time, and did not differ by eGFR status (see online supplementary figure S1B). Vaccinated patients had a higher crude incidence of LRTI than unvaccinated patients, in winter and summer. After adjustment for age, sex, comorbidities, pneumococcal vaccination, and characteristics of diabetes, the winter incidence rate of LRTI was higher among patients with a current influenza vaccine than unvaccinated patients (IRR 1.19: 95% CI 1.15 to 1.23) and among patients with residual influenza vaccination than unvaccinated patients (IRR 1.23: 95% CI 1.18 to 1.28). Similar or higher, adjusted IRRs were observed in summer. Using the ratio-of-ratios analysis, a 7% effectiveness of current influenza vaccine (95% CI 3 to 12) and a 12% effectiveness of residual influenza vaccination (95% CI 7 to 17) to prevent community-acquired LRTI were observed. There was no evidence to suggest a relationship between VE and eGFR nor proteinuria (table 3). Table 3 Lower respiratory tract infection (LRTI) rates and influenza vaccine effectiveness to prevent LRTI by season (n=190 459)

Vaccinated patients had a higher crude incidence of LRTI than unvaccinated patients, in winter and summer. After adjustment for age, sex, comorbidities, pneumococcal vaccination, and characteristics of diabetes, the winter incidence rate of LRTI was higher among patients with a current influenza vaccine than unvaccinated patients (IRR 1.19: 95% CI 1.15 to 1.23) and among patients with residual influenza vaccination than unvaccinated patients (IRR 1.23: 95% CI 1.18 to 1.28). Similar or higher, adjusted IRRs were observed in summer. Using the ratio-of-ratios analysis, a 7% effectiveness of current influenza vaccine (95% CI 3 to 12) and a 12% effectiveness of residual influenza vaccination (95% CI 7 to 17) to prevent community-acquired LRTI were observed. There was no evidence to suggest a relationship between VE and eGFR nor proteinuria (table 3). Table 3 Lower respiratory tract infection (LRTI) rates and influenza vaccine effectiveness to prevent LRTI by season (n=190 459) Summer Winter Influenza vaccination status Influenza vaccination status >5 years/never Current Residual 1–5 years >5 years/never Current Residual 1–5 years Person-time (years) 35 233 219 456 74 704 47 352 355 766 70 718 Infections (n) 2363 22 726 8496 5751 62 077 12 900 Crude LRTI rate /1000 py (95% CI) 73.0 (69.5 to 76.5) 111.0 (109.2 to 112.9) 121.3 (118.4 to 124.1) 134.8 (130.3 to 139.2) 187.2 (185.4 to 189.4) 195.5 (191.6 to 199.4) Crude LRTI rate ratio (95% CI) 1 (ref) 1.52 (1.45 to 1.60) 1.66 (1.58 to 1.75) 1 (ref) 1.38 (1.34 to 1.44) 1.45 (1.40 to 1.51) Adjusted* LRTI rate ratio (95% CI) Overall 1 (ref) 1.28 (1.21 to 1.35) 1.39 (1.32 to 1.47) 1 (ref) 1.19 (1.15 to 1.23) 1.23 (1.18 to 1.28) Stratified by eGFR (mL/min/1.73 m2) eGFR <30 1 (ref) 1.20 (0.94 to 1.52) 1.29 (1.01 to 1.65) 1 (ref) 1.17 (0.99 to 1.38) 1.20 (1.01 to 1.43) eGFR 30–44 1 (ref) 1.27 (1.10 to 1.45) 1.31 (1.13 to 1.51) 1 (ref) 1.20 (1.10 to 1.32) 1.20 (1.08 to 1.33) eGFR 45–59 1 (ref) 1.32 (1.19 to 1.46) 1.40 (1.26 to 1.56) 1 (ref) 1.16 (1.09 to 1.25) 1.20 (1.12 to 1.30) eGFR ≥ 60 1 (ref) 1.27 (1.19 to 1.37) 1.42 (1.32 to 1.53) 1 (ref) 1.21 (1.15 to 1.27) 1.26 (1.20 to 1.33) Stratified by proteinuria No proteinuria 1 (ref) 1.27 (1.20 to 1.35) 1.36 (1.30 to 1.45) 1 (ref) 1.18 (1.13 to 1.23) 1.22 (1.16 to 1.27) Proteinuria 1 (ref) 1.33 (1.19 to 1.49) 1.50 (1.33 to 1.68) 1 (ref) 1.23 (1.14 to 1.33) 1.27 (1.17 to 1.38) Ratio of incidence rate ratios* winter/summer (95% CI) Overall   1 (ref) 0.93 (0.88 to 0.97) 0.88 (0.83 to 0.93) Stratified by eGFR (mL/min/1.73 m2) eGFR <30   1 (ref) 0.93 (0.73 to 1.17) 0.90 (0.70 to 1.16) eGFR 30–44   1 (ref) 0.88 (0.77 to 1.01) 0.86 (0.75 to 1.00) eGFR 45–59   1 (ref) 0.90 (0.81 to 0.99) 0.87 (0.78 to 0.98) eGFR ≥ 60   1 (ref) 0.95 (0.89 to 1.02) 0.89 (0.83 to 0.96) Stratified by proteinuria No proteinuria   1 (ref) 0.93 (0.88 to 0.99) 0.90 (0.84 to 0.95) Proteinuria   1 (ref) 0.90 (0.81 to 1.00) 0.83 (0.74 to 0.93) Vaccine effectiveness (VE)* based on ratio of incidence rate ratios % (95% CI) Overall   0 (ref) 7 (3 to 12) 12 (7 to 17) Stratified by eGFR (mL/min/1.73 m2) eGFR <30   0 (ref) 7 (−17 to 27) 10 (−16 to 30) eGFR 30–44   0 (ref) 12 (−1 to 23) 14 (

.84 to 0.95) Proteinuria   1 (ref) 0.90 (0.81 to 1.00) 0.83 (0.74 to 0.93) Vaccine effectiveness (VE)* based on ratio of incidence rate ratios % (95% CI) Overall   0 (ref) 7 (3 to 12) 12 (7 to 17) Stratified by eGFR (mL/min/1.73 m2) eGFR <30   0 (ref) 7 (−17 to 27) 10 (−16 to 30) eGFR 30–44   0 (ref) 12 (−1 to 23) 14 ( 0 to 25) eGFR 45–59   0 (ref) 10 (1 to 19) 13 (2 to 22) eGFR ≥ 60   0 (ref) 5 (−2 to 11) 11 (4 to 17) p Value (test for trend)†   – 0.31 0.79 Stratified by proteinuria No proteinuria   0 (ref) 7 (1 to 12) 10 (5 to 16) Proteinuria   0 (ref) 10 (0 to 19) 17 (7 to 26) p Value‡   – 0.56 0.26 *Adjusted for: age, sex, socioeconomic status at practice level, residential care, date post 1 April 2004, smoking status, time-updated comorbidities (ischemic heart disease, congestive cardiac failure, hypertension, cerebrovascular disease, other dementia, chronic lung disease, chronic liver disease), time-updated CKD status (eGFR, proteinuria), steroid use in the 3 months prior to study entry, pneumococcal vaccination, and HbA1C and diabetic medication history at baseline. †Wald test for interaction of eGFR with influenza vaccination status and season, with eGFR modeled as a linear variable. ‡Wald test for interaction of proteinuria with influenza vaccine and season. LRTI, lower respiratory tract infection; py, person-years.

0 to 25) eGFR 45–59   0 (ref) 10 (1 to 19) 13 (2 to 22) eGFR ≥ 60   0 (ref) 5 (−2 to 11) 11 (4 to 17) p Value (test for trend)†   – 0.31 0.79 Stratified by proteinuria No proteinuria   0 (ref) 7 (1 to 12) 10 (5 to 16) Proteinuria   0 (ref) 10 (0 to 19) 17 (7 to 26) p Value‡   – 0.56 0.26 *Adjusted for: age, sex, socioeconomic status at practice level, residential care, date post 1 April 2004, smoking status, time-updated comorbidities (ischemic heart disease, congestive cardiac failure, hypertension, cerebrovascular disease, other dementia, chronic lung disease, chronic liver disease), time-updated CKD status (eGFR, proteinuria), steroid use in the 3 months prior to study entry, pneumococcal vaccination, and HbA1C and diabetic medication history at baseline. †Wald test for interaction of eGFR with influenza vaccination status and season, with eGFR modeled as a linear variable. ‡Wald test for interaction of proteinuria with influenza vaccine and season. LRTI, lower respiratory tract infection; py, person-years. Similar results were obtained in sensitivity analyses of influenza VE stratified by year (see online supplementary table S2), and excluding patients with chronic lung disease and congestive heart failure (see online supplementary table S3). Analyses using the influenza season dates did not change the results materially and are not shown. Sensitivity analysis limited to patients with white ethnicity did not change the results for either vaccine (results not shown).

Similar results were obtained in sensitivity analyses of influenza VE stratified by year (see online supplementary table S2), and excluding patients with chronic lung disease and congestive heart failure (see online supplementary table S3). Analyses using the influenza season dates did not change the results materially and are not shown. Sensitivity analysis limited to patients with white ethnicity did not change the results for either vaccine (results not shown). Conclusions Influenza and pneumococcal vaccine uptake was high among this study population of older people with diabetes mellitus, and did not vary according to markers of CKD. Pneumococcal vaccine had 22% (95% CI 11% to 31%) effectiveness against community-acquired pneumonia within the first year after vaccination. Pneumonia incidence among patients vaccinated more than 5 years previously was similar to that among patients with no record of vaccination (IRR 1.03: 95% CI 0.95 to 1.11). Community-acquired LRTI rates were higher among patients who received an influenza vaccination than among patients who did not, and this relationship remained after adjustment for age, sex, comorbidities, and characteristics of diabetes, and was observed in summer and winter. Traditional analyses would have concluded that influenza vaccination is associated with community-acquired LRTI. However, using the ratio-of-ratios analysis, a 7% effectiveness (95% CI 3% to 12%) of current influenza vaccine against community-acquired LRTI was observed. There was no evidence of a trend in influenza VE according to CKD status. However, there was evidence for a greater protective effect of pneumococcal vaccine among patients without a history of proteinuria than patients with a history of proteinuria.

rrent influenza vaccine against community-acquired LRTI was observed. There was no evidence of a trend in influenza VE according to CKD status. However, there was evidence for a greater protective effect of pneumococcal vaccine among patients without a history of proteinuria than patients with a history of proteinuria. Previous meta-analyses have found insufficient evidence for a protective effect of pneumococcal vaccine against all-cause pneumonia among the adult population due to heterogeneity.6 7 A subgroup analysis of a large Spanish cohort study found that only recent pneumococcal vaccination (<5 years) protected against hospitalization for all-cause community-acquired pneumonia (HR 0.75; 95% CI 0.58 to 0.98) among the general population aged ≥60 years.8 The authors suggested that the heterogeneity observed in meta-analyses might be explained by waning immunity among the vaccinated population. Our results support this view and suggest that pneumococcal vaccination appears to be effective against all-cause community-acquired pneumonia for a year following vaccination among people aged ≥65 years with diabetes, after which time we observed a decrease in pneumococcal VE to a null effect after 5 years.

opulation. Our results support this view and suggest that pneumococcal vaccination appears to be effective against all-cause community-acquired pneumonia for a year following vaccination among people aged ≥65 years with diabetes, after which time we observed a decrease in pneumococcal VE to a null effect after 5 years. Previous cohort studies among older people have provided evidence of a ‘healthy vaccinee effect’, in which higher vaccine uptake among healthier patients resulted in likely overestimation of influenza VE.9 10 12 32 33 Evidence suggesting a healthy vaccinee effect has also previously been found among older people with diabetes.34–38 In contrast, we observed higher rates of LRTI among patients who had received an influenza vaccination than among unvaccinated patients: our vaccinated patients appear, on this outcome measure, to be less healthy than unvaccinated patients. This finding is intriguing. The major difference between our study and most previous studies of this question is that we have included community-acquired LRTIs diagnosed and managed in primary and secondary care. One possible explanation of the difference is that vaccination may reflect health-seeking behavior in primary care. When patients develop symptoms of LRTI, patients who attend primary care for diagnosis and treatment may also be patients who were more likely to take up the influenza vaccine. This ascertainment bias may be less relevant to studies with hospitalization as an outcome—or could even be reversed, as vaccinated patients who attended primary care promptly with LRTI may be less likely to require hospital admission. An alternative explanation is that the healthy vaccinee effect observed in studies of hospitalization for LRTI/pneumonia may reflect residual confounding by ‘frailty’ in which frailer patients are less likely to take up vaccination and more likely to be admitted to hospital when they develop infection. This would be less relevant to diagnosis of LRTI in primary care, and so our outcome may be less vulnerable to residual confounding by indication.

ct residual confounding by ‘frailty’ in which frailer patients are less likely to take up vaccination and more likely to be admitted to hospital when they develop infection. This would be less relevant to diagnosis of LRTI in primary care, and so our outcome may be less vulnerable to residual confounding by indication. Our ‘ratio-of-ratios’ estimate suggested 7% VE of current influenza vaccination against LRTI among older people with diabetes (95% CI 3 to 12). Previous studies using similar strategies among the general population of older people have found no evidence of influenza VE against community-acquired pneumonia (VE 8%: 95% CI −10% to 23%), and evidence of a modest protection against influenza-related excess hospitalization with pneumonia/influenza (VE 19%: 95% CI 4% to 31%).14 15 Our estimate is consistent with both these estimates, and the difference may be due to the higher precision available for the present study due to the large cohort size.

o 23%), and evidence of a modest protection against influenza-related excess hospitalization with pneumonia/influenza (VE 19%: 95% CI 4% to 31%).14 15 Our estimate is consistent with both these estimates, and the difference may be due to the higher precision available for the present study due to the large cohort size. Our results suggested that pneumococcal VE may be reduced among patients with a history of proteinuria. We did not find any evidence of altered influenza VE among patients with CKD, but this may be due to limited power for the stratified ratio-of-ratios analysis. To the best of our knowledge, neither pneumococcal VE against pneumonia nor influenza VE against LRTI using methods to control for confounding by indication has been studied among patients with CKD who are not receiving dialysis. Studies of patients receiving dialysis may give some indication as to whether alteration of VE with CKD status is likely. A large observational study of pneumococcal vaccine found no evidence of effectiveness against hospitalization for pneumonia or respiratory infections among patients receiving dialysis.39 A study of influenza vaccine which calculated a ratio-of-ratios VE comparing influenza effectiveness in years with good match between the vaccine and circulating strain to effectiveness in a poorly matched ‘placebo year’ found no evidence of protection against influenza/pneumonia hospitalization among patients receiving hemodialysis (VE 2%: 95% CI −2% to5%).20 These studies suggest that the suggestion of reduced pneumococcal VE associated with CKD is plausible, but this question requires further investigation before conclusions can be drawn.

no evidence of protection against influenza/pneumonia hospitalization among patients receiving hemodialysis (VE 2%: 95% CI −2% to5%).20 These studies suggest that the suggestion of reduced pneumococcal VE associated with CKD is plausible, but this question requires further investigation before conclusions can be drawn. This study has several strengths. We used large, linked data sets with a careful definition of infection episodes to identify community-acquired infections managed in primary or secondary care, and excluded hospital-acquired infections and hospitalization from time at risk. This avoids differential hospital attendance patterns biasing estimates of VE according to markers of CKD. We adjusted for a wide range of comorbidities, and conducted a ratio-of-ratios analysis for influenza VE to address confounding by indication. We described the effect of pneumococcal vaccine according to time since vaccination, including booster doses, to identify waning immunity following vaccination. Our study population of older people with diabetes is well monitored for CKD,40 and this permitted us to explore the relationship of influenza and pneumococcal VE with CKD among patients not receiving dialysis, which we believe is novel.

ation, including booster doses, to identify waning immunity following vaccination. Our study population of older people with diabetes is well monitored for CKD,40 and this permitted us to explore the relationship of influenza and pneumococcal VE with CKD among patients not receiving dialysis, which we believe is novel. As an observational study of VE using routinely collected health record data, the study has limitations. LRTI/pneumonia is typically diagnosed clinically in general practice, without microbiological testing for the causative pathogen. Thus, we chose broader LRTI/pneumonia outcomes, in common with previous observational studies of influenza VE. We may have underascertained proteinuria and comorbidities; however, the selection of a highly monitored study population should minimize this risk, and the high prevalence of each we observed suggests that this was not a major source of misclassification. Despite adjustment for multiple comorbidities, residual confounding by indication may remain in the pneumococcal VE analysis. Despite our use of large, linked data sets, we had limited power to estimate the relationship of VE according to CKD status, especially in a ratio-of-ratios influenza VE analysis.

lassification. Despite adjustment for multiple comorbidities, residual confounding by indication may remain in the pneumococcal VE analysis. Despite our use of large, linked data sets, we had limited power to estimate the relationship of VE according to CKD status, especially in a ratio-of-ratios influenza VE analysis. Our findings have implications for clinical practice, public health, and future research. Our results should not be interpreted as demonstrating that influenza vaccine is ineffective among this population. We did not study the effectiveness of either vaccine against infection with their specific pathogens. As such, the results should neither discourage patients nor health professionals from influenza and pneumococcal vaccination.

be interpreted as demonstrating that influenza vaccine is ineffective among this population. We did not study the effectiveness of either vaccine against infection with their specific pathogens. As such, the results should neither discourage patients nor health professionals from influenza and pneumococcal vaccination. Our study question was the extent to which the burden of community-acquired LRTI may be preventable with vaccination and our results suggest that the growing burden of community-acquired LRTI and pneumonia among this population cannot be easily tackled by increasing uptake of existing routine vaccination programs. This is relevant for public health—in planning health service provision and designing effective strategies to prevent illness. It should also prompt a call for research into more effective immunization strategies and vaccination schedules. The low influenza VE we observed against community-acquired LRTI, when contrasted with the large burden of infection directly and indirectly attributed to influenza, suggests scope for strategies to improve vaccination effectiveness and better immunization among this population, for example, the use of adjuvants in vaccines. The suggestion of reduced pneumococcal VE among patients with proteinuria is interesting and needs confirmation in a repeat study. The authors are grateful to the RCGP Research and Surveillance Centre for sharing the RCGP Communicable and Respiratory Disease Reports for 2004–2011.

Our study question was the extent to which the burden of community-acquired LRTI may be preventable with vaccination and our results suggest that the growing burden of community-acquired LRTI and pneumonia among this population cannot be easily tackled by increasing uptake of existing routine vaccination programs. This is relevant for public health—in planning health service provision and designing effective strategies to prevent illness. It should also prompt a call for research into more effective immunization strategies and vaccination schedules. The low influenza VE we observed against community-acquired LRTI, when contrasted with the large burden of infection directly and indirectly attributed to influenza, suggests scope for strategies to improve vaccination effectiveness and better immunization among this population, for example, the use of adjuvants in vaccines. The suggestion of reduced pneumococcal VE among patients with proteinuria is interesting and needs confirmation in a repeat study. The authors are grateful to the RCGP Research and Surveillance Centre for sharing the RCGP Communicable and Respiratory Disease Reports for 2004–2011. Contributors: HIMD, SLT, and DN had the idea for the study and its design. HIMD carried out all analyses, using some analytic code provided by ERCM, and wrote the first draft of the manuscript, with input from SLT and DN. JQ contributed to interpretation of results and discussion. All authors contributed to this final manuscript and agreed to its submission. HIMD is the main guarantor of the content of this paper.

nalyses, using some analytic code provided by ERCM, and wrote the first draft of the manuscript, with input from SLT and DN. JQ contributed to interpretation of results and discussion. All authors contributed to this final manuscript and agreed to its submission. HIMD is the main guarantor of the content of this paper. Funding: This work was supported by National Institute for Health Research (grant number CDF 2010-03-32 to SLT) and Kidney Research UK (grant number ST2/2011 to HIMD). JQ was funded on an MRC Population Health Scientist Fellowship (grant number G0902135). Disclaimer: The study funders had no role in the design or conduct of the study, nor the collection, management, analysis nor interpretation of data, nor the preparation, review nor approval of the manuscript, nor the decision to submit it for publication. The views expressed in this publication are those of the authors and not necessarily those of the UK National Health Service, the National Institute for Health Research, the Department of Health, nor Kidney Research UK. Competing interests: None declared. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: The data that support the findings of this study were used under license and belong to the Clinical Practice Research Datalink, which can be contacted at https://www.cprd.com/intro.asp

Significance of the study What is already known about this subject? Previous clinical reports already exhibit that dipeptidyl peptidase-4 (DPP-4) inhibitors ameliorate diabetic nephropathy, such as albuminuria, in a glucose-lowering effect. What are the new findings? Our current study demonstrates that administration of anagliptin, which is switched from other DPP-4 inhibitors, in patients with type 2 diabetes with nephropathy showed that the levels of urinary albumin to creatinine ratio were significantly reduced after 24 weeks. However, the levels of hemoglobin A1c in patients showed no significant change during the treatment. We also found that treatment with anagliptin significantly decreased urinary liver-type  fatty acid-binding protein excretion after 24 weeks. How might these results change the focus of research or clinical practice? Anagliptin may exert beneficial effects for renoprotection in patients with type 2 diabetes with nephropathy in a glucose-lowering-independent manner.

We also found that treatment with anagliptin significantly decreased urinary liver-type  fatty acid-binding protein excretion after 24 weeks. How might these results change the focus of research or clinical practice? Anagliptin may exert beneficial effects for renoprotection in patients with type 2 diabetes with nephropathy in a glucose-lowering-independent manner. Introduction The prevalence of diabetes mellitus has been increasing worldwide in recent years. Long-term diabetes results in vascular changes and dysfunction, and its complications are the major causes of morbidity and mortality in patients. Among diabetic vascular complications, nephropathy is recognized as a leading cause of end-stage renal disease (ESRD) and an independent risk factor for cardiovascular diseases (CVD).1 The early clinical sign of diabetic nephropathy is elevated urinary albumin excretion, referred to as microalbuminuria, which progresses to overt proteinuria. Microalbuminuria in patients with diabetes has been recognized as a useful biomarker for diagnosing diabetic nephropathy and as a predictive factor for progression to ESRD.2 Additionally, microalbuminuria has been shown to be closely associated with an increased risk of cardiovascular morbidity and mortality.3–5 Therefore, microalbuminuria is a biomarker for the diagnosis of diabetic nephropathy and an important therapeutic target for improving the prognosis of renal and cardiovascular risk in patients with diabetes.6 Previous clinical data also showed that urinary liver-type free fatty acid-binding protein (L-FABP), which is associated with renal tubulointerstitial damage and oxidative stress, may be a predictive marker for renal and cardiovascular prognosis in patients with type 2 diabetes.7 8

ular risk in patients with diabetes.6 Previous clinical data also showed that urinary liver-type free fatty acid-binding protein (L-FABP), which is associated with renal tubulointerstitial damage and oxidative stress, may be a predictive marker for renal and cardiovascular prognosis in patients with type 2 diabetes.7 8 Multifactorial management, including diet therapy and glycemic, blood pressure (BP) and lipid control, is recommended for diabetic nephropathy.2 9 10 Among the multifactorial treatments, intensive glycemic control in type 2 diabetes significantly reduced diabetes-induced microvascular events, mainly as a consequence of a reduction in nephropathy.2 However, intensive glycemic control, accompanied by hypoglycemia, is closely related to increased mortality, which is associated with increased incidence of CVD.11 12 Therefore, avoiding hypoglycemia is important in the treatment of patients with diabetes, in particular those who have diabetic nephropathy, because they are a high-risk group for CVD. Treatment with dipeptidyl peptidase-4 (DPP-4) inhibitors, which are oral antidiabetic agents, results in improvements in the blood glucose levels in patients with diabetes following stimulation of endogenous insulin secretion, inhibition of glucagon release and reduction of gastric emptying via the enhanced production of incretin hormones. DPP-4 inhibitors enhance active levels of  glucagon-like peptide-1 (GLP-1) and gastric inhibitory polypeptide (GIP) via inhibition of cleaving and inactivating these incretins by DPP-4 enzyme. DPP-4 inhibitors have become widely accepted in clinical practice because of their low risk of hypoglycemia. In addition to the glucose-lowering effect, previous data from animal and clinical studies demonstrate that DPP-4 inhibitors, including sitagliptin,13–18 linagliptin,19–23 alogliptin,24 vildagliptin25 26 or saxagliptin,27 have pleiotropic beneficial effects such as renoprotection or antiatherogenesis, which are independent of the glucose-lowering effect. Additionally, anagliptin shows serum lipid-lowering effects, which have not yet been observed with the other DPP-4 inhibitors. However, there are not sufficient clinical data regarding the renoprotective effect of anagliptin in patients with diabetes. Therefore, the aim of this study is to investigate the possible effects of anagliptin on glycemic/lipid control and renal function, including albuminuria, in patients with type 2 diabetes with nephropathy.

here are not sufficient clinical data regarding the renoprotective effect of anagliptin in patients with diabetes. Therefore, the aim of this study is to investigate the possible effects of anagliptin on glycemic/lipid control and renal function, including albuminuria, in patients with type 2 diabetes with nephropathy. Research design and method Subjects A total of 48 participants with type 2 diabetes (30 men and 18 women) were selected for the present study from outpatients who visited the Department of Endocrinology and Metabolism at Kanazawa Medical University Hospital. The entry criteria included (1) age ≥20 years old, (2) type 2 diabetes with hemoglobin A1c (HbA1c) ≥6.0%, (3) urinary albumin to creatinine (Cr) ratio (UACR) ≥30 mg/g Cr in spot urine for screening of diabetic nephropathy, and (4) treatment with diet, exercise therapy and oral antidiabetic agents (glimepiride ≤2 mg/day or gliclazide ≤40 mg/day or glibenclamide ≤1.25 mg/day). The exclusion criteria were (1) type 1 diabetes, (2) treatment with insulin therapy, (3) severe diabetic metabolic complications such as ketoacidosis, (4) severe liver dysfunction, (5) hemodialysis, (6) severe chronic heart failure, (7) pregnant or nursing women and those who might be pregnant, and (8) any patient whom the investigator judged to be inappropriate for this study. Patients were given detailed explanations of the study protocol. Informed consent was obtained from each patient. The study protocol was approved by the Ethical Committee of Kanazawa Medical University. The trial was registered with the University Hospital Medical Information Network (UMIN No 000012802).

or this study. Patients were given detailed explanations of the study protocol. Informed consent was obtained from each patient. The study protocol was approved by the Ethical Committee of Kanazawa Medical University. The trial was registered with the University Hospital Medical Information Network (UMIN No 000012802). Study protocol The present study was an open-label, prospective study. At the start of the study, anagliptin 200 mg/day was added to other oral antidiabetic agents such as sulfonylurea (SU), metformin, an α-glucosidase inhibitor (α-GI), pioglitazone and a sodium-glucose cotransporter 2 (SGLT2) inhibitor or, when participants received other DPP-4 inhibitors, the DPP-4 inhibitor was switched to anagliptin 200 mg/day. In addition, in some cases, the anagliptin dose was increased to up to 400 mg/day after 12 weeks, if the physician judged glucose control in the patients to be insufficient. Participants were assessed for the following parameters before the start of the study, 12 and 24 weeks after the addition of anagliptin or when switching to anagliptin: No changes were made to the type and dose of glucose-lowering agents, renin–angiotensin system (RAS) inhibitors such as ACE inhibitors (ACEIs), angiotensin receptor II blockers (ARBs) or spironolactone during the study period. These agents had been prescribed for at least 3 months before the study.

ching to anagliptin: No changes were made to the type and dose of glucose-lowering agents, renin–angiotensin system (RAS) inhibitors such as ACE inhibitors (ACEIs), angiotensin receptor II blockers (ARBs) or spironolactone during the study period. These agents had been prescribed for at least 3 months before the study. After performing a screening of UACR ≥30 mg/g Cr in spot urine, diabetic nephropathy was finally diagnosed by an UACR ≥30 mg/g Cr in the first urine in the early morning. The primary endpoint of the study was change in HbA1c during the treatment with anagliptin. Additionally, we evaluated the changes in lipid data (low-density lipoprotein-cholesterol (LDL-C), high-density lipoprotein-cholesterol (HDL-C) and triglyceride (TG) levels), UACR and urinary L-FABP excretion as secondary endpoints.

primary endpoint of the study was change in HbA1c during the treatment with anagliptin. Additionally, we evaluated the changes in lipid data (low-density lipoprotein-cholesterol (LDL-C), high-density lipoprotein-cholesterol (HDL-C) and triglyceride (TG) levels), UACR and urinary L-FABP excretion as secondary endpoints. Measurements Blood samples were collected in the morning after an overnight fast. First urine in the early morning sample was collected at the home of participants, and urine was carried in a cooling box to the hospital. HbA1c was measured using an automated analyzer, HLC-723 G11 (TOSHO, Tokyo, Japan). Serum LDL-C and HDL-C levels were measured using enzymatic methods (Qualigent HDL-C and Qualigent LDL-C, Sekisui Medical, Tokyo, Japan). Serum TG levels were measured using enzymatic assays (Kyowa Medex, Tokyo, Japan). Urinary albumin was measured by immunonephelometry using a kit from NITTOBO MEDICAL (Tokyo, Japan). Because of their skewed distribution, the UACR data were log10-transformed before analysis. The results of the analysis were back-transformed to obtain geometric means of the UACR of the 24-week value to the baseline value; the values then were expressed as percentage change in the adjusted geometric mean of the UACR ratios of the 24-week value to the baseline value. Serum and urinary Cr were measured using enzymatic assays (Sekisui Medical), and the estimated glomerular filtration rate (eGFR) was calculated as 194×serum creatinine−1.094×age−0.287 in men and as 194×serum creatinine−1.094×age−0.287×0.739 in women.28 Urinary L-FABP was measured by a chemiluminescent enzyme immunoassay, using a Lumipulse L-FABP assay (Fujirebio, Tokyo, Japan), and urinary L-FABP excretion was expressed as the urinary L-FABP to Cr ratio. Serum cystatin C was measured by Latex immunoturbidimetric methods (LSI Medience, Tokyo, Japan).

0.739 in women.28 Urinary L-FABP was measured by a chemiluminescent enzyme immunoassay, using a Lumipulse L-FABP assay (Fujirebio, Tokyo, Japan), and urinary L-FABP excretion was expressed as the urinary L-FABP to Cr ratio. Serum cystatin C was measured by Latex immunoturbidimetric methods (LSI Medience, Tokyo, Japan). Statistical analysis Statistical analyses were performed with a StatView V.5 system (Abacus Concepts, Berkeley, California, USA) for Windows. All values are summarized as the mean and SD unless otherwise indicated. Differences in the percentage change in the UACR and urinary L-FABP before and after administration of anagliptin were assessed by a paired t-test. A Wilcoxon signed-rank test was performed as appropriate for comparison of the two groups. The unpaired t-test was performed as comparison of the two groups on HbA1c, lipid data, BP, body mass index (BMI) and renal function. The correlation of two variables was analyzed by single linear regression analysis. Statistical significance was defined as p<0.05.

ormed as appropriate for comparison of the two groups. The unpaired t-test was performed as comparison of the two groups on HbA1c, lipid data, BP, body mass index (BMI) and renal function. The correlation of two variables was analyzed by single linear regression analysis. Statistical significance was defined as p<0.05. Results Patient disposition is shown in figure 1. Initially, 48 subjects who exhibited albuminuria of more than 30 mg/g Cr in spot urine were enrolled in this study. However, seven subjects were excluded for several reasons, which were diarrhea (n=2), elevation of BP (n=1), colon diverticulitis (n=1), worsening of depression (n=1), increased drowsiness (n=1) and withdrawal of consent before the beginning of treatment with anagliptin (n=1). Furthermore, 16 subjects were also excluded because their albuminuria was less than 30 mg/g Cr in the early-morning first urine. Therefore, we evaluated 25 subjects for the analysis of diabetic nephropathy in this study. Of 25 subjects, 20 subjects were switched to anagliptin from other DPP-4 inhibitors, and in 5 subjects anagliptin was additionally administered. Figure 1 Patient disposition and study protocol. The 25 participants showed 30 mg/g creatinine (Cr) in the urinary albumin to Cr ratio (UACR) in the first urine in the early morning and were diagnosed with diabetic nephropathy. They received anagliptin as an additional treatment (n=5) or were switched from other dipeptidyl peptidase-4 (DPP-4) inhibitors (n=20), and were evaluated at the start of the study and after 12 and 24 weeks. HbA1c, hemoglobin A1c.

CR) in the first urine in the early morning and were diagnosed with diabetic nephropathy. They received anagliptin as an additional treatment (n=5) or were switched from other dipeptidyl peptidase-4 (DPP-4) inhibitors (n=20), and were evaluated at the start of the study and after 12 and 24 weeks. HbA1c, hemoglobin A1c. Baseline clinical and biochemical characteristics in two groups (shown as a group of switch (n=20) and a group of addition (n=5)), as well as concomitant background therapies, are shown in tables 1 and 2. The mean age was 67.6±9.0 and 71.2±4.6 years old, men:women=15:5 and 3:2, baseline BMI of the study population was 25.1±4.0 kg/m2 and 26.6±3.3 kg/m2, and the duration of diabetes was 15.1±7.6 and 12.4±4.8 years, respectively, in the two groups. Baseline HbA1c levels were 7.3%±0.9% and 7.8%±0.9%, and fasting glucose levels were 153.7±38.5 mg/dL and 182.8±80.5 mg/dL. Lipid data were 91.5±25.4 mg/dL and 104.2±17.6 mg/dL for LDL-C, 50.2±13.9 mg/dL and 46.8±11.0 mg/dL for HDL-C, and 165.6±98.7 mg/dL and 123.2±46.6 mg/dL for TG, respectively. Alanine aminotransferase levels were 20.3±9.8 IU/L and 21.0±6.4 IU/L, and uric acid levels were 5.5±1.3 mg/dL and 4.6±1.3 mg/dL, respectively. The median eGFR and serum cystatin C at baseline were 74.0±18.4 mL/min/1.73 m2 and 66.3±23.9 mL/min/1.73 m2, and 0.95±0.23 mg/dL and 0.98±0.35 mg/dL, and the UACR values at baseline were 206.4±343.9 mg/g Cr and 172.3±139.4 mg/g Cr, respectively. We assessed urinary L-FABP excretion in the participants showing more than 5 µg/g Cr at baseline, and the median urinary L-FABP excretion was 8.5±2.8 and 6.4±0.95 µg/g Cr, respectively. At entry of study, in each of the two groups, 80% of the participants had microalbuminuria, and macroalbuminuria was noted in 20% of the individuals. All participants received oral antidiabetic agents including SU (55% and 40%), metformin (75% and 100%), an α-GI (20% and 0%), pioglitazone (10% and 0%) and an SGLT2 inhibitor (5% and 0%) at baseline. Twenty participants received a DPP-4 inhibitor, which includes sitagliptin (n=7, 35%), teneligliptin (n=5, 25%), alogliptin (n=3, 15%), vildagliptin (n=3, 15%) and linagliptin (n=2, 10%). Fifteen participants (75%) in the group of switch (n=20) received antihypertensive therapy at baseline.

tor (5% and 0%) at baseline. Twenty participants received a DPP-4 inhibitor, which includes sitagliptin (n=7, 35%), teneligliptin (n=5, 25%), alogliptin (n=3, 15%), vildagliptin (n=3, 15%) and linagliptin (n=2, 10%). Fifteen participants (75%) in the group of switch (n=20) received antihypertensive therapy at baseline. Of the 15 participants, 13 participants (87%) received RAS inhibitors such as ACEIs (n=4), ARBs (n=9) or spironolactone (n=2) at baseline, with two participants receiving dual RAS blockade, using ACEI and ARBs, or ARBs and spironolactone. Fourteen participants (70%) in the group of switch were treated with lipid-lowering therapy. Of the 14 participants, 12 participants received statins (85%). Table 1 Baseline clinical and biochemical characteristics

Of the 15 participants, 13 participants (87%) received RAS inhibitors such as ACEIs (n=4), ARBs (n=9) or spironolactone (n=2) at baseline, with two participants receiving dual RAS blockade, using ACEI and ARBs, or ARBs and spironolactone. Fourteen participants (70%) in the group of switch were treated with lipid-lowering therapy. Of the 14 participants, 12 participants received statins (85%). Table 1 Baseline clinical and biochemical characteristics n 20 (Switch) 5 (Addition) Male:female 15:5 3:2 Age (years) 67.6±9.0 71.2±4.6 BMI (kg/m2) 25.1±4.0 26.6±3.3 BP (mm Hg) 130.9±12.3/71.8±10.8 145.4±18.5/73.0±6.8 HbA1c (%) 7.3±0.9 7.8±0.9 FPG (mg/dL) 153.7±38.5 182.8±80.5 LDL-C (mg/dL) 91.5±25.4 104.2±17.6 HDL-C (mg/dL) 50.2±13.9 46.8±11.0 TG (mg/dL) 165.6±98.7 123.2±46.6 ALT (IU/L) 20.3±9.8 21.0±6.4 UA (mg/dL) 5.5±1.3 4.6±1.3 eGFR (mL/min/1.73 m2) 74.0±18.4 66.3±23.9 eGFR>90(mL/min/1.73 m2), n (%) 5 (25) 1 (20) eGFR 60–90 (mL/min/1.73 m2), n (%) 10 (50) 2 (40) eGFR 30–60 (mL/min/1.73 m2), n (%) 5 (25) 2 (40) Cystatin C (mg/dL) 0.95±0.23 0.98±0.35 UACR (mg/g Cr) 206.4±343.9 172.3±139.4 UACR 30–300 (mg/g Cr), n (%) 16 (80) 4 (80) UACR>300 (mg/g Cr), n (%) 4 (20) 1 (20) UACR (log) (mg/g Cr) 1.95±0.51 1.99±0.38 ULFABP>5.0 (μg/g Cr), n (%) 8 (40) 3 (60) ULFABP (μg/g Cr) 8.5±2.8 6.4±0.95 Duration of diabetes (years) 15.1±7.6 12.4±4.8 Data are the mean±SD, or n (%).

5 UACR (mg/g Cr) 206.4±343.9 172.3±139.4 UACR 30–300 (mg/g Cr), n (%) 16 (80) 4 (80) UACR>300 (mg/g Cr), n (%) 4 (20) 1 (20) UACR (log) (mg/g Cr) 1.95±0.51 1.99±0.38 ULFABP>5.0 (μg/g Cr), n (%) 8 (40) 3 (60) ULFABP (μg/g Cr) 8.5±2.8 6.4±0.95 Duration of diabetes (years) 15.1±7.6 12.4±4.8 Data are the mean±SD, or n (%). ALT, alanine transaminase; BMI, body mass index; BP, blood pressure; Cr, creatinine; eGFR, estimated glomerular filtration rate; FPG, fasting plasma glucose; Hb1Ac, hemoglobin A1c; HDL-C, high-density lipoprotein-cholesterol; LDL-C, low-density lipoprotein-cholesterol; TG, triglyceride; UA, uric acid; UACR, urinary albumin to creatinine ratio; ULFABP, urinary liver-type fatty acid-binding protein to creatinine ratio. Table 2 Baseline background therapies for diabetes, hypertension and dyslipidemia

ALT, alanine transaminase; BMI, body mass index; BP, blood pressure; Cr, creatinine; eGFR, estimated glomerular filtration rate; FPG, fasting plasma glucose; Hb1Ac, hemoglobin A1c; HDL-C, high-density lipoprotein-cholesterol; LDL-C, low-density lipoprotein-cholesterol; TG, triglyceride; UA, uric acid; UACR, urinary albumin to creatinine ratio; ULFABP, urinary liver-type fatty acid-binding protein to creatinine ratio. Table 2 Baseline background therapies for diabetes, hypertension and dyslipidemia Antidiabetic background therapy at baseline, n (%) 20 (100) (Switch) 5 (100) (Addition) Metformin, n (%) 15 (75) 5 (100) SUs, n (%) 11 (55) 2 (40) DPP-4 inhibitors, n (%) 20 (100) – Sitagliptin, n (%) 7 (35) – Teneligliptin, n (%) 5 (25) – Alogliptin, n (%) 3 (15) – Vildagliptin, n (%) 3 (15) – Linagliptin, n (%) 2 (10) – α-GIs, n (%) 4 (20) 0 Pioglitazone, n (%) 2 (10) 0 SGLT2 inhibitors, n (%) 1 (5) 0 Antihypertensive background therapy at baseline, n (%) 15 (75) 3 (60) RAS inhibitors, n (%) 13 (87) 2 (67) ACEIs, n 4 0 ARBs, n 9 2 ACEI+ARB, n 1 0 Spironolactone, n 2 0 Ca antagonists, n (%) 11 (73) 3 (60) Diuretics, n (%) 0 (0) 1 (20) β blockers, n (%) 1 (7) 1 (20) α-Methyldopa, n (%) 2 (13) 0 Lipid-lowering background therapy, n (%) 14 (70) 2 (40) Statins, n (%) 12 (85) 2 (40) Fibrate, n (%) 2 (14) 0 Ezetimibe, n (%) 1 (0.7) 0 α-GI, α-glucosidase inhibitor; ACEI, ACE inhibitor; ARB, angiotensin II receptor blocker; DPP-4, dipeptidyl peptidase-4; RAS, renin–angiotensin system; SGLT2, sodium-glucose cotransporter 2; SU, sulfonylurea.

therapy, n (%) 14 (70) 2 (40) Statins, n (%) 12 (85) 2 (40) Fibrate, n (%) 2 (14) 0 Ezetimibe, n (%) 1 (0.7) 0 α-GI, α-glucosidase inhibitor; ACEI, ACE inhibitor; ARB, angiotensin II receptor blocker; DPP-4, dipeptidyl peptidase-4; RAS, renin–angiotensin system; SGLT2, sodium-glucose cotransporter 2; SU, sulfonylurea. In a group with additional treatment with anagliptin (n=5, including one participant with anagliptin dose of up to 400 mg/day after 12 weeks), HbA1c was significantly decreased at 12 and 24 weeks (6.9%±0.5%, p<0.01 and 6.7%±0.3%, p<0.05) from baseline (7.8%±0.9%), respectively (see online supplementary figure 1A). The UACR from baseline to 12 or 24 weeks was not significantly decreased in this group, but there was a tendency toward reduction (see online supplementary figure 1B,C). However, a single linear regression analysis between Δ%UACR and ΔHbA1c at 24 weeks showed significant correlation (r=0.904, p=0.035) (see online supplementary figure 1D). However, since the number of participants who received additional treatment with anagliptin was small (n=5) in this study, limited conclusions can be drawn. Therefore, we excluded five patients who received additional treatment with anagliptin, and analyzed 20 subjects who were switched to anagliptin from other DPP-4 inhibitors for evaluating the effect of anagliptin on glycemic control, lipid data and diabetic nephropathy in this study. 10.1136/bmjdrc-2017-000391.supp1Supplementary data 1

In a group with additional treatment with anagliptin (n=5, including one participant with anagliptin dose of up to 400 mg/day after 12 weeks), HbA1c was significantly decreased at 12 and 24 weeks (6.9%±0.5%, p<0.01 and 6.7%±0.3%, p<0.05) from baseline (7.8%±0.9%), respectively (see online supplementary figure 1A). The UACR from baseline to 12 or 24 weeks was not significantly decreased in this group, but there was a tendency toward reduction (see online supplementary figure 1B,C). However, a single linear regression analysis between Δ%UACR and ΔHbA1c at 24 weeks showed significant correlation (r=0.904, p=0.035) (see online supplementary figure 1D). However, since the number of participants who received additional treatment with anagliptin was small (n=5) in this study, limited conclusions can be drawn. Therefore, we excluded five patients who received additional treatment with anagliptin, and analyzed 20 subjects who were switched to anagliptin from other DPP-4 inhibitors for evaluating the effect of anagliptin on glycemic control, lipid data and diabetic nephropathy in this study. 10.1136/bmjdrc-2017-000391.supp1Supplementary data 1 After treatment with anagliptin switching from other DPP-4 inhibitors (n=20), HbA1c was not significantly changed, 7.4%±1.1% at 12 weeks and 7.5%±1.2% at 24 weeks, compared with baseline HbA1c (7.3%±0.9%) (figure 2A). The UACR (log) was significantly reduced after 24 weeks of treatment with anagliptin (1.76±0.53 mg/g Cr, p<0.01), compared with that at baseline (1.95±0.51 mg/g Cr) (figure 2B). The percentage change in UACR (Δ%UACR) from baseline to 12 or 24 weeks was also significantly lower by −8.8% at 12 weeks (p<0.001) and by −10.6% at 24 weeks (p<0.001) (figure 2C). The significant reduction in the UACR induced by anagliptin might be due to switching from sitagliptin, alogliptin and teneligliptin. By contrast, switching from vildagliptin and linagliptin to anagliptin seemed to show no effect on urinary albumin excretion (figure 2D). Lipid data including LDL-C, HDL-C and TG were not changed during treatment with anagliptin (figure 3A–C). We also found no significant change in systolic BP during treatment with anagliptin (figure 3D), and the reduction of Δ%UACR from baseline to 12 or 24 weeks was independent of receiving RAS inhibitors (figure 3E,F). BMI showed no significant change after 12 and 24 weeks of treatment with anagliptin compared with that at baseline (figure 3G). Renal function, which was evaluated by measurement of eGFR and serum cystatin C levels, showed no significant change during anagliptin treatment (figure 3H,I). Furthermore, we assessed urinary L-FABP excretion. We analyzed eight participants who had more than 5 µg/g Cr of urinary L-FABP excretion at baseline in switching to anagliptin treatment group. Urinary L-FABP excretion was significantly decreased from baseline (8.5±2.8 µg/g Cr) to 24 weeks (3.1±1.7 µg/g Cr, p<0.01) after treatment with anagliptin (figure 4A), and the percentage change in urinary L-FABP excretion (Δ%ULFABP (L-FABP to creatinine ratio)) during treatment with anagliptin was −58.1% (p<0.001) (figure 4B).

L-FABP excretion was significantly decreased from baseline (8.5±2.8 µg/g Cr) to 24 weeks (3.1±1.7 µg/g Cr, p<0.01) after treatment with anagliptin (figure 4A), and the percentage change in urinary L-FABP excretion (Δ%ULFABP (L-FABP to creatinine ratio)) during treatment with anagliptin was −58.1% (p<0.001) (figure 4B). Figure 2 (A) Hemoglobin A1c (HbA1c) values at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks.  (B) Urinary albumin to creatinine ratio (UACR) (log) values in 20 participants at baseline and after treatment with anagliptin at 24 weeks. p<0.01 versus baseline. (C) Percentage change in the UACR in 20 participants from baseline to after treatment with anagliptin at 12 and 24 weeks. p<0.001 versus baseline. (D) Change in the UACR after switching from sitagliptin, alogliptin, vildagliptin, teneligliptin or linagliptin to anagliptin. p<0.05 versus baseline. Error bars represent SD. n.s, denotes not significant. Figure 3 (A) Low-density lipoprotein-cholesterol (LDL-C) values at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. (B) High-density lipoprotein-cholesterol (HDL-C) values at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. (C) Triglyceride (TG) values at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. Error bars represent SD. n.s denotes not significant.

weeks. (B) High-density lipoprotein-cholesterol (HDL-C) values at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. (C) Triglyceride (TG) values at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. Error bars represent SD. n.s denotes not significant. Figure 3 (D) Systolic blood pressure (BP) values at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. . (E) Percentage change in the urinary albumin to creatinine ratio (UACR) in 13 participants who received renin–angiotensin system (RAS) inhibitors from baseline to after treatment with anagliptin at 12 and 24 weeks. p<0.05 versus baseline. (F) Percentage change in the UACR in 7 participants who did not receive RAS inhibitors from baseline to after treatment with anagliptin at 12 and 24 weeks. p<0.05 versus baseline. Error bars represent SD. n.s denotes not significant. Figure 3 (G) Body mass index (BMI) values at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. (H) Estimated glomerular filtration rate (eGFR) at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. (I) Serum cystatin C values at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. Error bars represent SD. n.s denotes not significant.

12 and 24 weeks. (H) Estimated glomerular filtration rate (eGFR) at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. (I) Serum cystatin C values at baseline and after treatment with anagliptin in 20 participants at 12 and 24 weeks. Error bars represent SD. n.s denotes not significant. Figure 4 (A) Urinary L-FABP (ULFABP) values in eight participants who had more than 5 µg/g creatinine (Cr) at the start of the study, at baseline and after treatment with anagliptin at 24 weeks. p<0.01 versus baseline. (B) Percentage change in UFABP (Δ%UFABP) in eight participants from baseline to after treatment with anagliptin at 24 weeks. p<0.001 versus baseline. Error bars represent SD. L-FABP, liver-type  fatty acid-binding protein. The single linear regression analysis between Δ%UACR and age, duration of diabetes, ΔHbA1c, ΔLDL-C, Δsystolic BP or ΔBMI for 24 weeks, and HbA1c, UACR, BMI and eGFR at baseline, did not show significant correlation (table 3). In addition, there was no relationship between Δ%ULFABP and ΔHbA1c after 24 weeks, which was evaluated by a single linear regression analysis (r=0.547, 95% CI −0.270 to 1.48, p=0.165). Table 3 The single linear regression analysis between Δ%UACR and clinical parameters

The single linear regression analysis between Δ%UACR and age, duration of diabetes, ΔHbA1c, ΔLDL-C, Δsystolic BP or ΔBMI for 24 weeks, and HbA1c, UACR, BMI and eGFR at baseline, did not show significant correlation (table 3). In addition, there was no relationship between Δ%ULFABP and ΔHbA1c after 24 weeks, which was evaluated by a single linear regression analysis (r=0.547, 95% CI −0.270 to 1.48, p=0.165). Table 3 The single linear regression analysis between Δ%UACR and clinical parameters r 95% CI p Value Δ% UACR Age (year) −0.180 −0.293 to 0.658 0.447 Duration of diabetes (year) 0.317 −0.147 to 0.804 0.173 ΔHbA1c (%) 0.039 −0.437 to 0.514 0.871 ΔSystolic BP (mm Hg) −0.292 −0.190 to 0.790 0.226 ΔBMI (kg/m2) −0.002 −0.473 to 0.477 0.993 ΔLDL-C (mg/dL) −0.149 −0.325 to 0.625 0.531 ΔTG (mg/dL) −0.250 −0.220 to 0.730 0.288 ΔHDL-C (mg/dL) 0.033 −0.442 to 0.509 0.890 ΔeGFR (mL/min/1.73 m2) 0.180 −0.294 to 0.657 0.448 HbA1c at baseline (%) −0.219 −0.253 to 0.698 0.354 UACR (log) at baseline (mg/g Cr) −0.010 −0.465 to 0.486 0.965 BMI at baseline (kg/m2) −0.323 −0.140 to 0.811 0.164 eGFR at baseline (mL/min/1.73 m2) −0.305 −0.161 to 0.790 0.191 Data are the results of a single linear regression analysis for variables at 24 weeks. BP, blood pressure; BMI, body mass index; eGFR, estimated glomerular filtration rate; HbA1c, hemoglobin A1c; HDL-C, high-density lipoprotein-cholesterol; LDL-C, low-density lipoprotein-cholesterol; TG, triglyceride; UACR, urinary albumin to creatinine ratio.

r 95% CI p Value Δ% UACR Age (year) −0.180 −0.293 to 0.658 0.447 Duration of diabetes (year) 0.317 −0.147 to 0.804 0.173 ΔHbA1c (%) 0.039 −0.437 to 0.514 0.871 ΔSystolic BP (mm Hg) −0.292 −0.190 to 0.790 0.226 ΔBMI (kg/m2) −0.002 −0.473 to 0.477 0.993 ΔLDL-C (mg/dL) −0.149 −0.325 to 0.625 0.531 ΔTG (mg/dL) −0.250 −0.220 to 0.730 0.288 ΔHDL-C (mg/dL) 0.033 −0.442 to 0.509 0.890 ΔeGFR (mL/min/1.73 m2) 0.180 −0.294 to 0.657 0.448 HbA1c at baseline (%) −0.219 −0.253 to 0.698 0.354 UACR (log) at baseline (mg/g Cr) −0.010 −0.465 to 0.486 0.965 BMI at baseline (kg/m2) −0.323 −0.140 to 0.811 0.164 eGFR at baseline (mL/min/1.73 m2) −0.305 −0.161 to 0.790 0.191 Data are the results of a single linear regression analysis for variables at 24 weeks. BP, blood pressure; BMI, body mass index; eGFR, estimated glomerular filtration rate; HbA1c, hemoglobin A1c; HDL-C, high-density lipoprotein-cholesterol; LDL-C, low-density lipoprotein-cholesterol; TG, triglyceride; UACR, urinary albumin to creatinine ratio. Discussion The present study showed that the administration of anagliptin for 24 weeks significantly decreased the UACR from the baseline in a glucose-independent, lipid-independent and BP-independent manner. In addition, treatment with anagliptin significantly reduced the levels of urinary L-FABP excretion in the participants, who had more than 5 µg/g Cr at baseline, independent of the change in HbA1c. This is the first report showing a renoprotective effect of anagliptin on patients with type 2 diabetes with nephropathy.

In addition, treatment with anagliptin significantly reduced the levels of urinary L-FABP excretion in the participants, who had more than 5 µg/g Cr at baseline, independent of the change in HbA1c. This is the first report showing a renoprotective effect of anagliptin on patients with type 2 diabetes with nephropathy. DPP-4 is an enzyme that cleaves and inactivates incretin hormones capable of stimulating insulin secretion from pancreatic β cells. DPP-4 inhibitors are now widely used for the treatment of type 2 diabetes. Previous reports have shown that currently available DPP-4 inhibitors, including anagliptin, exert a glucose-lowering effect in patients with diabetes,29 and there is no significant difference in their glucose-lowering efficacy. Our data also demonstrated that HbA1c was significantly decreased at 12 weeks and 24 weeks from baseline levels by additional treatment with anagliptin. However, switching to anagliptin from other DPP-4 inhibitors such as sitagliptin, alogliptin, vildagliptin, linagliptin or teneligliptin exhibited no significant change in HbA1c levels after 12 and 24 weeks of treatment with anagliptin in 20 participants. In addition to the glucose-lowering effect, previous reports have shown that anagliptin has a lipid-lowering effect, decreasing the plasma total cholesterol, LDL-C and TG levels, which was indicated by pooled analysis of phase III clinical trials.30 However, in this study, administration of anagliptin showed no change in lipid data after both 12 and 24 weeks of treatment. Although 70% of participants received lipid-lowering drugs, including statins, fibrate or ezetimibe, anagliptin did not show lipid-lowering effects, independent of receiving lipid-lowering drugs or not receiving them. It is unclear why anagliptin exhibited no lipid-lowering effects in this study.

24 weeks of treatment. Although 70% of participants received lipid-lowering drugs, including statins, fibrate or ezetimibe, anagliptin did not show lipid-lowering effects, independent of receiving lipid-lowering drugs or not receiving them. It is unclear why anagliptin exhibited no lipid-lowering effects in this study. Previous clinical studies have shown a beneficial effect of DPP-4 inhibitors in diabetic nephropathy. Sitagliptin reduced albuminuria in several uncontrolled trials and a small randomized controlled trial, and the reduction of albuminuria was independent of the decrease in HbA1c,14 17 and correlated with decreases in both systolic BP and eGFR.31 Groop et al,23 in a pooled analysis of four studies, demonstrated that treatment with linagliptin in addition to RAS inhibitors reduced UACR by 32% and by 6% compared with placebo, and the efficacy of linagliptin was unaffected by baseline HbA1c levels or systolic BP, changing of HbA1c or systolic BP during the treatment with linagliptin. Therefore, the effect of linagliptin on the reduction of albuminuria was exerted in a glucose-independent and BP-independent manner. Tani et al 26 also evaluated the effects of vildagliptin on atherogenic LDL-C heterogeneity and albuminuria in subjects with diabetes. The UACR decreased significantly by ∼45% after 8 weeks of treatment with vildagliptin, and reduction of UACR by vildagliptin was correlated with change in HbA1c, lipid data, systolic BP and eGFR. Thus, clinically, DPP-4 inhibitors may improve albuminuria and may have a renoprotective effect; however, further study is necessary to identify whether long-term treatment with several DPP-4 inhibitors in patients with diabetes may maintain renal function as well as reduction of albuminuria.

ata, systolic BP and eGFR. Thus, clinically, DPP-4 inhibitors may improve albuminuria and may have a renoprotective effect; however, further study is necessary to identify whether long-term treatment with several DPP-4 inhibitors in patients with diabetes may maintain renal function as well as reduction of albuminuria. There are few reports regarding comparative data among DPP-4 inhibitors on the renoprotective effect, including reduction of albuminuria. Fujita et al 24 reported that in a crossover study with two DPP-4 inhibitors, sitagliptin and alogliptin, in patients with type 2 diabetes who have microalbuminuria and take ARBs, switching from sitagliptin to alogliptin reduced UACR in a glucose-lowering-independent manner. Switching to alogliptin from sitagliptin significantly reduced urinary 8-hydroxy-2’-deoxyguanosine (8-OHdG) excretion, and increased the plasma levels of stromal cell-derived factor-1α (SDF-1α), which is one of the substrates of DPP-4. Therefore, alogliptin might be more effective in the reduction of albuminuria compared with sitagliptin. However, the detailed mechanism is still unclear. In our study, the Δ%UACR from baseline to 12 or 24 weeks after anagliptin treatment was significantly lower, in a glucose-independent, lipid-independent, BP-independent or use of RAS inhibitors-independent manner. The reduction in the UACR induced by anagliptin might be observed by switching from sitagliptin, alogliptin and teneligliptin. By contrast, switching from vildagliptin and linagliptin to anagliptin seemed to show no effect on urinary albumin excretion, but the sample number was very small. What is the difference in the effect of reduction of the UACR among each of the DPP-4 inhibitors? Previously, we reported that linagliptin, but not sitagliptin, inhibits the homodimer formation of DPP-4, which is related to DPP-4 activation, in cultured endothelial cells,32 and this difference may be one of the reasons why the two drugs display different properties. DPP-4 is widely expressed in many cell types, including renal endothelial cells and epithelial tubular cells; therefore, different DPP-4 inhibitors may exhibit diverse biological influences depending on the cell type. However, it is unclear whether anagliptin can inhibit the homodimer formation of DPP-4, similar to linagliptin. In addition, differences in the binding modes in the active site of DPP-4 and the form of binding may contribute to the different effects exhibited by DPP-4 inhibitors.

influences depending on the cell type. However, it is unclear whether anagliptin can inhibit the homodimer formation of DPP-4, similar to linagliptin. In addition, differences in the binding modes in the active site of DPP-4 and the form of binding may contribute to the different effects exhibited by DPP-4 inhibitors. A previous report analyzing the single-crystal structure and enzyme interactions showed that the interacting subsites of anagliptin with DPP-4 are the S1, S2 and S2 extensive subsites, and anagliptin is included in class 3 according to the categorization by Nabeno et al.33 This binding mode of anagliptin leads to high and selective DPP-4 inhibition. Furthermore, anagliptin binds to Ser630 of DPP-4, which is a catalytic residue and a center of its activation in the S1 subsite, through a dipole interaction of the cyanopyrrolidine structure, and anagliptin may possibly lead to the formation of the imidate intermediates through covalent binding to DPP-4.34 Vildagliptin also binds to DPP-4 through covalent binding, which is thought to be a strong binding form.33 Therefore, inhibition of DPP-4 activity due to strength of binding to DPP-4 may be related to the renoprotective effect of anagliptin. In addition, anagliptin is taken twice a day, and therefore the peak of inhibition of DPP-4 activity occurs twice a day, which can lead to strong suppression of DPP-4 activity in the kidney, compared with other DPP-4 inhibitors such as sitagliptin, alogliptin and teneligliptin, which are taken once a day. Uchino and Kaku35 reported that administration of anagliptin twice a day exhibited the significantly increased plasma levels of active GLP-1, particularly after dinner, compared with those in the treatment with sitagliptin once a day, in an open-label, two-period crossover study. Thus, differences in the chemical structure, binding mode of DPP-4 inhibitors and the number of the peak of inhibition of DPP-4 activity may cause different effects on renoprotection; however, further studies are necessary to elucidate differences between anagliptin and other DPP-4 inhibitors, or whether anagliptin has a better effect than other DPP-4 inhibitors, particularly sitagliptin, alogliptin and teneligliptin, on renoprotection.

f DPP-4 activity may cause different effects on renoprotection; however, further studies are necessary to elucidate differences between anagliptin and other DPP-4 inhibitors, or whether anagliptin has a better effect than other DPP-4 inhibitors, particularly sitagliptin, alogliptin and teneligliptin, on renoprotection. Urinary L-FABP is one of the markers for tubulointerstitial damage and an oxidative stress marker. Araki et al reported that urinary L-FABP of more than 5 µg/g Cr may be a predictive marker for renal and cardiovascular prognosis in patients with type 2 diabetes without advanced nephropathy.7 8 Therefore, we evaluated the effect of anagliptin on urinary excretion in patients who had a urinary L-FABP level of more than 5 µg/g Cr. Interestingly, anagliptin clearly decreased the excretion of urinary L-FABP, which indicates a reduction of tubulointerstitial damage, tubular hypoxia and oxidative stress. There are no reports showing a beneficial effect of DPP-4 inhibitors on urinary L-FABP excretion. However, since we could not measure the oxidative stress marker such as urinary 8-OHdG excretion, it is unclear whether anagliptin may provide renal protective effect via stronger antioxidative action than other DPP-4 inhibitors. Thus, our data indicate that anagliptin may suppress both albuminuria and urinary L-FABP, which are predictive markers for renal and cardiovascular prognosis, indicating improvement of glomerular/tubulointerstitial damage, possibly inhibiting the progression of diabetic nephropathy and CVD.

than other DPP-4 inhibitors. Thus, our data indicate that anagliptin may suppress both albuminuria and urinary L-FABP, which are predictive markers for renal and cardiovascular prognosis, indicating improvement of glomerular/tubulointerstitial damage, possibly inhibiting the progression of diabetic nephropathy and CVD. Experimental studies have suggested a renoprotective role of DPP-4 inhibitors in various models of chronic kidney disease (CKD), including diabetic nephropathy, which may be independent of lowering glucose levels. The renoprotective effect of DPP-4 inhibitors in diabetic nephropathy may be exerted through an increase in active GLP-1 or through the inhibition of DPP-4 itself. Previous reports show that GLP-1 receptor agonists may prevent disease progression in diabetic nephropathy through direct effects on the GLP-1 receptor in renal cells including glomerular endothelial cells and monocytes/macrophages.36 37 Higashijima et al 38 also demonstrated that DPP-4 inhibitors, including anagliptin, reduced macrophage infiltration directly via GLP-1-dependent signaling in a rat Thy-1 nephritis model. Therefore, increased GLP-1 induced by DPP-4 inhibition may also lead to renal protection through the GLP-1 receptor and its signaling.39 By contrast, several reports showed that the inhibition of DPP-4 ameliorates kidney injury animal models, including diabetic nephropathy. Tanaka et al 40 also demonstrated that linagliptin significantly inhibited tubulointerstitial injury induced by peritoneal injection of free fatty acid-bound albumin, such as inflammation, fibrosis and apoptosis, in mice without altering blood glucose levels. The anti-inflammatory effect of DPP-4 inhibition in monocytes/macrophages is also associated with renoprotection. In an apolipoprotein E-deficient atherosclerotic mice model, not a kidney disease model, Ervinna et al 41 demonstrated that anagliptin exerted an antiatherosclerotic effect through inhibition of the inflammatory reaction of monocytes and inhibition of smooth muscle cell proliferation. Shinjo et al 42 also demonstrated that anagliptin attenuated inflammatory cytokine expression in lipopolysaccharide-stimulated macrophage, adipocytes and hepatocytes. The in vitro suppressive effects on cytokine production in cultured macrophages by anagliptin suggest the anti-inflammatory effects of these DPP-4 inhibitors to be direct actions rather than via increased concentrations of incretins such as GLP-1.

in lipopolysaccharide-stimulated macrophage, adipocytes and hepatocytes. The in vitro suppressive effects on cytokine production in cultured macrophages by anagliptin suggest the anti-inflammatory effects of these DPP-4 inhibitors to be direct actions rather than via increased concentrations of incretins such as GLP-1. Furthermore, they showed that sitagliptin also exerted anti-inflammation, as well as that of anagliptin; however, the effect of sitagliptin is weaker than that of anagliptin. The treatment with anagliptin and sitagliptin resulted in similar inhibitory effects on DPP-4 activity in the supernatants of both cultured macrophages and adipocytes, whereas anagliptin more strongly inhibited DPP-4 activity in both cell lysates than sitagliptin. The difference in the degrees of anti-inflammatory effects between anagliptin and sitagliptin may be explained by different inhibitory efficiencies against DPP-4 in cell lysates (cell surface DPP-4) and supernatants (soluble form of DPP-4). Oxidative stress also plays a crucial role for the pathogenesis of diabetic nephropathy. Mega et al 43 showed that sitagliptin ameliorated diabetic nephropathy in Zucker diabetic fatty rat, accompanied by reduced lipid peroxidation. Furthermore, teneligliptin works as a direct scavenger of hydroxyl radicals, resulting in reduction of oxidative stress.44 There are few reports regarding the renoprotective effect of anagliptin in both experimental animal models and in human data. Therefore, further study is necessary to evaluate these points.

dation. Furthermore, teneligliptin works as a direct scavenger of hydroxyl radicals, resulting in reduction of oxidative stress.44 There are few reports regarding the renoprotective effect of anagliptin in both experimental animal models and in human data. Therefore, further study is necessary to evaluate these points. There were several limitations in our study design. It was a non-controlled observational study that occurred over a short time period, and the number of participants was small. DPP-4 cleaves a lot of substrates (peptides) including GLP-1, GIP, SDF-1α, brain natriuretic peptide (BNP) and so on. Although we could not show the changes in the levels of these peptides following anagliptin treatment, these changes may be involved in renal protection observed in this study. In addition, we could not evaluate oxidative stress or inflammation to assess the mechanism of the beneficial effect on the diabetic kidney. Therefore, the mechanism by which anagliptin reduced the UACR and urinary L-FABP excretion in a glucose-lowering independent manner is unclear. Further study is necessary to elucidate these points. In conclusion, in the present study of just 24 weeks’ duration, anagliptin caused a decrease in the UACR and urinary L-FABP, which are prognostic markers for CKD and CVD, and the decrease was independent of any change in HbA1c. Therefore, anagliptin may have potential for halting the progression of diabetic nephropathy and the development of CVD through a renoprotective effect. 10.1136/bmjdrc-2017-000391.supp2

In conclusion, in the present study of just 24 weeks’ duration, anagliptin caused a decrease in the UACR and urinary L-FABP, which are prognostic markers for CKD and CVD, and the decrease was independent of any change in HbA1c. Therefore, anagliptin may have potential for halting the progression of diabetic nephropathy and the development of CVD through a renoprotective effect. 10.1136/bmjdrc-2017-000391.supp2 We thank Yuka Kuroshima, Erii Hayashi and Yuka Udagawa, who are clinical research coordinators, and all the staff in the Department of Endocrinology and Metabolism of Kanazawa Medical University Hospital for their great assistance in this study. Contributors: MK and DK designed the study, researched and analyzed the data and wrote and edited the manuscript. KK, ST, MF, MN and AN contributed to the research and the collection of data. KK contributed to the discussion. DK is the guarantor of this work. Funding: This work was financially supported by a grant from the Kidney Foundation, Japan to DK. Competing interests: Boehringer Ingelheim, Mitsubishi Tanabe Pharma, Kyowa Hakko Kirin, Taisho Toyama Pharmaceutical and Ono Pharmaceutical contributed to establishing the Division of Anticipatory Molecular Food Science and Technology. The authors declare that there are no conflicts of interest associated with this manuscript. Patient consent: Obtained. Ethics approval: Kanazawa Medical University. Provenance and peer review: Not commissioned; externally peer reviewed.

Significance of this study What is already known about this subject? Dietary and lifestyle changes delay the onset of type 2 diabetes. The Dietary Approaches to Stop Hypertension (DASH) dietary pattern has also been shown to prevent type 2 diabetes because of its potential to improve insulin resistance and reduce hyperglycemia. What are the new findings? The study reports a stronger association between DASH and controlled self-reported diabetes status and with participants who were not taking antihyperglycemic medications. A higher DASH score was associated with less insulin resistance among Hispanics/Latinos. Differences in DASH scores by Hispanic/Latino heritage did not explain the differences in prevalence of diabetes and insulin resistance reported in the diverse Hispanic/Latino population. How might these results change the focus of research or clinical practice? A higher DASH score is indeed associated with lower insulin resistance. However, further research is needed in order to determine the role of different diet components and diabetes in the diverse Hispanic/Latino population of the USA.

Differences in DASH scores by Hispanic/Latino heritage did not explain the differences in prevalence of diabetes and insulin resistance reported in the diverse Hispanic/Latino population. How might these results change the focus of research or clinical practice? A higher DASH score is indeed associated with lower insulin resistance. However, further research is needed in order to determine the role of different diet components and diabetes in the diverse Hispanic/Latino population of the USA. Introduction Hispanics/Latinos represent 17.6% of the US population,1 and this figure continues to grow. The prevalence of type 2 diabetes mellitus among Hispanics/Latinos has been consistently higher than among non-Hispanic whites.2 In addition, the prevalence of diabetes mellitus has been shown to vary substantially among Hispanic/Latino heritage groups, from 10.2% in those of South American origin to 18% in those of Dominican, Puerto Rican, and Mexican origins.3 Similarly, there are significant differences in the prevalence of metabolic syndrome by Hispanic/Latino heritage.4 Although the specific reasons underlying these differences are unknown, they likely stem from a combination of genetic, biological, and cultural differences.5 Dietary behaviors, in particular, have been proposed as a major contributor in these disparities, since Hispanics/Latinos of diverse origins and heritages have different dietary patterns. In fact, our group recently reported significant variation in the consumption of food and macronutrients among Hispanic/Latino heritage groups.6

viors, in particular, have been proposed as a major contributor in these disparities, since Hispanics/Latinos of diverse origins and heritages have different dietary patterns. In fact, our group recently reported significant variation in the consumption of food and macronutrients among Hispanic/Latino heritage groups.6 Previous research has demonstrated that dietary and lifestyle changes delay the onset of type 2 diabetes, and that certain eating habits may lead to changes in inflammatory markers and insulin resistance.7–10 The Dietary Approaches to Stop Hypertension (DASH) dietary pattern—which includes high intake of fruits, vegetables, whole grains, and low-fat dairy products—was originally developed to treat hypertension.11 12 An index using the DASH dietary pattern that scores various food and nutrient components is thus a measure of diet quality, with higher scores indicating a healthier diet. Following a DASH dietary pattern has also been shown to prevent type 2 diabetes because of its potential to improve insulin resistance and reduce hyperglycemia.13 14 A recent analysis based on the Insulin Resistance Atherosclerosis Study (IRAS, including 548 Hispanics/Latinos) showed an inverse association between adherence to the DASH dietary pattern and the incidence of type 2 diabetes.14 There are otherwise no studies that have examined the role of the DASH dietary pattern on diabetes and insulin resistance outcomes in a large and diverse Hispanic/Latino population.

548 Hispanics/Latinos) showed an inverse association between adherence to the DASH dietary pattern and the incidence of type 2 diabetes.14 There are otherwise no studies that have examined the role of the DASH dietary pattern on diabetes and insulin resistance outcomes in a large and diverse Hispanic/Latino population. Using baseline data from the Hispanic Community Health Study/Study of Latinos (HCHS/SOL), we examined the association between the DASH dietary pattern and type 2 diabetes and insulin resistance in Hispanic/Latino adults, and the extent to which differences in the DASH dietary pattern contribute to differences in diabetes status and insulin resistance across Hispanic/Latino heritage. Enhanced understanding of these associations in the diverse US Hispanic/Latino population may help prioritize the development of interventions targeting modifiable risk factors contributing to the risk for type 2 diabetes in this population.

ces in diabetes status and insulin resistance across Hispanic/Latino heritage. Enhanced understanding of these associations in the diverse US Hispanic/Latino population may help prioritize the development of interventions targeting modifiable risk factors contributing to the risk for type 2 diabetes in this population. Research design and methods The HCHS/SOL is a multicenter, prospective, population-based cohort study, and it is the largest epidemiologic study of Hispanics/Latinos in the USA. A total of 16 415 participants aged 18–74 years at screening from randomly selected households were recruited from four US locations: Bronx, NY; Chicago, IL; Miami, FL; and San Diego, CA. A detailed description of the HCHS/SOL sampling design and methods has been described elsewhere.15 16 In brief, the study was designed to include participants from Cuban, Dominican, Mexican, Puerto Rican, Central American, and South American heritages living in the selected communities, and adults aged 45–74 years were oversampled. All study participants provided informed consent, and the study had institutional review board approval from each institution participating in the study. For this analysis, we used baseline examination data (2008–2011).

American heritages living in the selected communities, and adults aged 45–74 years were oversampled. All study participants provided informed consent, and the study had institutional review board approval from each institution participating in the study. For this analysis, we used baseline examination data (2008–2011). Study measurements and procedures Enrolled participants completed a baseline examination in their preferred language (English or Spanish). All procedures and interviewer-administered questionnaires were conducted by centrally trained and certified bilingual study personnel following a standardized protocol, which included ongoing quality-assurance procedures. During the baseline visit, the following data were collected relevant to our research question and analysis: health behaviors such as dietary behaviors; medical history; and demographics, including age, sex, self-reported Hispanic/Latino heritage, years of education, and household income. Further, anthropometric measurements (including weight in kilogram, and height and waist circumference in centimeter) were performed by trained and certified staff following a standard protocol (www.cscc.unc.edu/hchs). Body mass index (BMI) was calculated as weight in kilogram divided by height in square meter. Blood samples were collected by a non-traumatic venous puncture after a fasting period of at least 8 hours prior to the visit. Participants with a fasting plasma glucose (FPG) <150 mg/dL and no previous diagnosis of diabetes completed a standard 75 g 2-hour oral glucose tolerance test (2hPG). Hemoglobin A1C and fasting insulin levels were collected. The assays’ methodologies and their procedures are described on the HCHS/SOL website (www.cscc.unc.edu/hchs).

ith a fasting plasma glucose (FPG) <150 mg/dL and no previous diagnosis of diabetes completed a standard 75 g 2-hour oral glucose tolerance test (2hPG). Hemoglobin A1C and fasting insulin levels were collected. The assays’ methodologies and their procedures are described on the HCHS/SOL website (www.cscc.unc.edu/hchs). Outcomes Diabetes status Diabetes status was defined from three main sources of information: self-reported physician-diagnosed diabetes, use of antihyperglycemic medications (scanned diabetes medications), and baseline laboratory collection (FPG, 2hPG, and A1C percentage). Specifically, participants were classified into one of the following seven mutually exclusive groups:Normal glucose tolerance: FPG <100 mg/dL, 2hPG <140 mg/dL, A1C <5.7%, no history of diabetes, and not taking antihyperglycemic medications Pre-diabetes FPG 100–125 mg/dL, or 2hPG 140–199 mg/dL, or A1C 5.7%–6.4%, and no history of diabetes and not taking antihyperglycemic medications Participants with self-reported diabetes with optimal glycemic control (A1C <7%)Taking antihyperglycemic medications Not taking antihyperglycemic medications Participants with self-reported diabetes without optimal glycemic control (A1C ≥7%)Taking antihyperglycemic medications Not taking antihyperglycemic medications Unrecognized diabetes based on baseline laboratory collection:FPG ≥126 mg/dL, 2hPG ≥200 mg/dL, A1C ≥6.5%, and no self-reported history of diabetes, and not taking antihyperglycemic medications

Participants with self-reported diabetes without optimal glycemic control (A1C ≥7%)Taking antihyperglycemic medications Not taking antihyperglycemic medications Unrecognized diabetes based on baseline laboratory collection:FPG ≥126 mg/dL, 2hPG ≥200 mg/dL, A1C ≥6.5%, and no self-reported history of diabetes, and not taking antihyperglycemic medications We combined the seven-level diabetes status to create a two-level version of diabetes status (1—normal glucose or pre-diabetes, 2—diabetes) and a three-level version (1—normal glucose, 2—pre-diabetes, and 3—diabetes). Homeostatic model assessment of β-cell function and insulin resistance (HOMA-IR) was calculated for all participants as the product of fasting insulin (µU/mL) and fasting glucose (mmol/L) divided by 22.17

glucose or pre-diabetes, 2—diabetes) and a three-level version (1—normal glucose, 2—pre-diabetes, and 3—diabetes). Homeostatic model assessment of β-cell function and insulin resistance (HOMA-IR) was calculated for all participants as the product of fasting insulin (µU/mL) and fasting glucose (mmol/L) divided by 22.17 DASH score Dietary intake was assessed in all participants using two 24-hour recalls, one in person at the baseline visit and one via unannounced telephone call (30 days after the baseline visit, on average), using the multiple-pass methods of the Nutrition Data System for Research software, V.11, from the Nutrition Coordinating Center at the University of Minnesota. Recalls were excluded if energy intake was below the sequence (first or second)-sex-specific first percentile or above the 99th percentile, or if the recall was unreliable according to the interviewer. The DASH dietary pattern was scored based on the average of the two recalls using the components and standards for minimum and maximum scores from Günther et al.18 Briefly, the DASH score is the sum of eight component scores (grains, vegetables, fruits, dairy, red and processed meat, nuts/seeds/legumes, fats/oils, and sweets), each ranging from 0 (worst) to 10 (best). The grains component is the sum of the scores for the total grains and whole grains subcomponents, and the dairy component is the sum of the scores for total dairy and low-fat dairy subcomponents. Each of the four subcomponents ranges from 0 (worst) to 5 (best). DASH scores can range from 0 to 80. Higher DASH scores (healthier diet) indicate higher consumption of the grains, vegetables, fruits, dairy, and nuts/seeds/legumes components and lower consumption of the red and processed meat, fats/oils, and sweets components.

the four subcomponents ranges from 0 (worst) to 5 (best). DASH scores can range from 0 to 80. Higher DASH scores (healthier diet) indicate higher consumption of the grains, vegetables, fruits, dairy, and nuts/seeds/legumes components and lower consumption of the red and processed meat, fats/oils, and sweets components. Covariates Participants reported their age, sex, Hispanic/Latino heritage, years of education (less than high school, high school, more than high school), annual household income (<$10 000, $10 000–$20 000, $20 000–$40 000, $40 000–$75 000, >$75 000, not reported), dietary acculturation, energy intake, current smoking status, and family history of diabetes. The first item of the dietary behavior questionnaire asked whether the participant’s foods are usually of Hispanic/Latino or American origin (dietary acculturation) using a five-level Likert scale (mainly Hispanic/Latino foods; mostly Hispanic/Latino foods and some American food; equal amounts of both Hispanic/Latino and American foods; mostly American foods and some Hispanic/Latino foods; and mainly American foods). For this analysis, we combined the ‘mainly’ and the ‘mostly’ categories, creating a three-level categorical variable. Energy intake (kcal) was calculated as the average energy (kcal) from both 24-hour dietary recalls.

and American foods; mostly American foods and some Hispanic/Latino foods; and mainly American foods). For this analysis, we combined the ‘mainly’ and the ‘mostly’ categories, creating a three-level categorical variable. Energy intake (kcal) was calculated as the average energy (kcal) from both 24-hour dietary recalls. Statistical analyses We excluded 473 participants due to either age >74 years at baseline (n=9) or missing DASH score (n=234), diabetes status (n=8), or HOMA-IR (n=222), yielding an analytical sample of 15 942 participants. There were no significant differences in baseline characteristics between participants included in the analysis versus those excluded due to missing data. HOMA-IR was log transformed before analyses. Distribution of demographic, health characteristics, and DASH dietary pattern is presented by diabetes status (seven mutually exclusive groups). In model 1, the association between DASH dietary pattern (score or tertiles) and diabetes status (two-level, three-level, and seven-level) was assessed using survey multinomial logistic regression adjusting by age, sex, Hispanic/Latino heritage, education, family income, family history of diabetes, smoking status, dietary acculturation, field center, and energy intake. Model 2 further adjusted by BMI and waist circumference, and model 3 added HOMA-IR. To test whether the association of DASH score and diabetes status differed by Hispanic/Latino heritage, we included the interaction between DASH and heritage. The association of DASH dietary pattern and HOMA-IR was assessed using linear regression adjusted by covariates specified in models 1 and 2 previously, and tested separately the interactions of DASH with seven-level diabetes status and with Hispanic/Latino background. When interactions were significant, at a 0.1 significance level, analyses were stratified; otherwise, models were reduced to exclude the interaction. All analyses accounted for the complex sample design and sampling weights using survey procedures in SAS V.9.3 and SAS-callable SUDAAN V.11.

h Hispanic/Latino background. When interactions were significant, at a 0.1 significance level, analyses were stratified; otherwise, models were reduced to exclude the interaction. All analyses accounted for the complex sample design and sampling weights using survey procedures in SAS V.9.3 and SAS-callable SUDAAN V.11. Results In the target population, the average age was 41.1 years. Overall, 52.3% were female, the average weight was 78.9 kg, and 39.7% were obese. Table 1 provides demographic, diet, health characteristics, glucose, insulin, HOMA-IR, and mean DASH scores overall and by diabetes status. On average, those with diabetes (self-reported and unrecognized) were older and had a higher body weight, BMI, and HOMA-IR than those with normal glucose tolerance and pre-diabetes. HOMA-IR was highest in those with uncontrolled self-reported diabetes either taking or not taking antihyperglycemic medications (5.5 and 6.3, respectively). The mean DASH score was highest in those with self-reported diabetes controlled and no medications (44.8) followed by controlled and on medications (43.9), and the DASH score was lowest in those with normal glucose tolerance (41.3) and pre-diabetes (41.5). Those with uncontrolled self-reported diabetes and taking no medications had the highest consumption of total grains (score of 4.3). Those with uncontrolled self-reported diabetes taking medications had the highest consumption of vegetables (score of 4.6), and those with normal glucose tolerance had the lowest consumption (score of 4.1). Those with controlled self-reported diabetes on medications had the highest consumption of fruits (score of 4.3). Those with self-reported diabetes (both controlled and uncontrolled) and not taking medications also had the highest consumption of low-fat dairy (score of 3.1). Adults with self-reported diabetes controlled and not taking medications had the highest consumption of nuts, seeds, and dried beans (5.7), and they had the lowest (healthiest) consumption of meat, poultry, eggs, and fish (9.7); fats and oils (7.5); and sweets (1.7). Energy intake (kcal/day) was 300 kcal higher in those with normal glucose tolerance compared with those with self-reported diabetes taking medications (1625.6 kcal/day among those with uncontrolled diabetes and 1643.1 kcal/day among those with controlled diabetes) (table 1).

9.7); fats and oils (7.5); and sweets (1.7). Energy intake (kcal/day) was 300 kcal higher in those with normal glucose tolerance compared with those with self-reported diabetes taking medications (1625.6 kcal/day among those with uncontrolled diabetes and 1643.1 kcal/day among those with controlled diabetes) (table 1). Table 1 Demographic, health, and DASH score by diabetes status, HCHS/SOL (2008–2011)

9.7); fats and oils (7.5); and sweets (1.7). Energy intake (kcal/day) was 300 kcal higher in those with normal glucose tolerance compared with those with self-reported diabetes taking medications (1625.6 kcal/day among those with uncontrolled diabetes and 1643.1 kcal/day among those with controlled diabetes) (table 1). Table 1 Demographic, health, and DASH score by diabetes status, HCHS/SOL (2008–2011) Characteristic Overall (n=15 942) Normal glucose tolerance (n=6590) Pre-diabetes (n=6079) SR diabetes controlled (medications) (n=706) SR diabetes controlled (no medications) (n=255) SR diabetes uncontrolled (medications) (n=1040) SR diabetes uncontrolled (no medications) (n=171) Unrecognized diabetes (n=1101) Age (years) 41.1 (40.6, 41.6) 34.0 (33.5, 34.5) 45.2 (44.5, 45.8) 57.9 (56.6, 59.2) 51.5 (49.3, 53.7) 54.6 (53.3, 55.9) 49.0 (46.1, 52.0) 52.3 (51.0, 53.5) Female 52.3 (51.2, 53.4) 54.5 (52.9, 56.1) 48.0 (46.0, 50.0) 56.5 (50.5, 62.3) 51.1 (41.7, 60.4) 55.7 (50.8, 60.5) 45.6 (35.2, 56.4) 56.5 (52.4, 60.5) Diet acculturation Mainly/mostly Hispanic 73.6 (72.2, 75.0) 71.0 (69.2, 72.8) 75.2 (73.0, 77.3) 75.6 (70.5, 80.1) 80.5 (73.0, 86.3) 77.3 (73.2, 80.9) 83.6 (74.9, 89.8) 78.9 (74.7, 82.5) Equally Hispanic/American 22.4 (21.1, 23.6) 24.4 (22.9, 26.0) 21.1 (19.2, 23.2) 21.0 (16.8, 25.9) 14.5 (9.7, 21.0) 19.7 (16.3, 23.6) 14.5 (8.9, 23.0) 18.3 (14.7, 22.4) Mainly/mostly American 4.0 (3.6, 4.6) 4.6 (3.9, 5.4) 3.7 (2.9, 4.7) 3.4 (2.0, 5.6) 5.0 (2.2, 10.9) 3.0 (2.1, 4.4) 1.8 (0.5, 6.2) 2.9 (1.8, 4.5) Current smoker 21.2 (20.0, 22.3) 21.6 (20.0, 23.2) 21.8 (20.2, 23.6) 14.9 (11.3, 19.4) 24.7 (17.1, 34.3) 19.0 (15.6, 23.0) 23.5 (15.6, 33.7) 17.3 (14.3, 20.8) Weight (kg) 78.9 (78.4, 79.4) 75.3 (74.6, 76.0) 81.4 (80.6, 82.3) 85.2 (83.0, 87.3) 82.5 (79.5, 85.5) 85.0 (83.1, 86.9) 84.9 (81.0, 88.8) 83.5 (81.9, 85.1) BMI (kg/m2) Underweight/normal 23.0 (21.9, 24.0) 33.0 (31.4, 34.7) 15.0 (13.6, 16.5) 9.7 (6.5, 14.3) 11.5 (7.3, 17.6) 9.4 (7.2, 12.2) 10.8 (5.5, 20.1) 8.7 (6.7, 11.3) Overweight 37.3 (36.2, 38.5) 37.9 (36.2, 39.7) 38.3 (36.5, 40.2) 28.6 (23.9, 33.8) 35.3 (26.0, 45.9) 33.8 (29.7, 38.1) 37.3 (27.1, 48.7) 34.3 (30.5, 38.5) Obese 39.7 (38.4, 41.0) 29.0 (27.2, 30.9) 46.7 (44.7, 48.6) 61.7 (56.0, 67.1) 53.2 (43.5, 62.6) 56.9 (52.2, 61.4) 51.9 (41.1, 62.5) 56.9 (52.9, 60.9) Waist circumference (cm) 97.4 (97.0, 97.8) 93.1 (92.6, 93.7) 99.8 (99.2, 100.4) 106.3 (104.7, 107.9) 101.8 (99.4, 104.2) 107.0 (105.7, 108.4) 104.6 (101.6, 107.6) 104.2 (103.0, 105.3) Fasting glucose (mmol/L)* 98.7 (98.2, 99.2) 89.5 (89.3, 89.7) 98.1 (97.8, 98.5) 111.8 (109.6, 114.0) 104.0 (100.

62.5) 56.9 (52.9, 60.9) Waist circumference (cm) 97.4 (97.0, 97.8) 93.1 (92.6, 93.7) 99.8 (99.2, 100.4) 106.3 (104.7, 107.9) 101.8 (99.4, 104.2) 107.0 (105.7, 108.4) 104.6 (101.6, 107.6) 104.2 (103.0, 105.3) Fasting glucose (mmol/L)* 98.7 (98.2, 99.2) 89.5 (89.3, 89.7) 98.1 (97.8, 98.5) 111.8 (109.6, 114.0) 104.0 (100. 6, 107.6) 170.5 (165.3, 175.9) 209.0 (191.6, 227.9) 124.2 (120.5, 128.0) Fasting insulin (mU/L)* 10.2 (10.0, 10.5) 8.4 (8.1, 8.6) 11.9 (11.6, 12.2) 13.3 (12.3, 14.3) 12.1 (10.6, 13.8) 13.1 (12.2, 14.1) 12.2 (10.5, 14.2) 15.9 (15.1, 16.8) HOMA-IR* 2.5 (2.4, 2.6) 1.8 (1.8, 1.9) 2.9 (2.8, 3.0) 3.7 (3.4, 4.0) 3.1 (2.7, 3.6) 5.5 (5.1, 6.0) 6.3 (5.5, 7.2) 4.9 (4.6, 5.2) DASH score, 0–80 41.6 (41.3, 42.0) 41.3 (40.9, 41.7) 41.5 (41.0, 41.9) 43.9 (42.8, 44.9) 44.8 (43.1, 46.6) 42.6 (41.7, 43.6) 43.2 (40.8, 45.5) 42.5 (41.6, 43.4) DASH tertiles Less healthy 35.9 (34.4, 37.3) 37.7 (35.8, 39.6) 35.4 (33.3, 37.6) 26.8 (22.6, 31.5) 25.5 (18.4, 34.1) 31.9 (27.6, 36.6) 31.2 (22.0, 42.2) 34.4 (30.5, 38.5) Normal 33.7 (32.5, 34.8) 33.0 (31.5, 34.6) 35.5 (33.4, 37.6) 31.8 (27.2, 36.7) 33.8 (25.6, 43.0) 33.1 (28.1, 38.6) 31.1 (22.5, 41.2) 29.2 (25.8, 32.9) Healthiest 30.5 (29.1, 31.9) 29.3 (27.5, 31.1) 29.1 (27.2, 31.0) 41.4 (35.6, 47.4) 40.8 (32.3, 49.9) 34.9 (30.5, 39.7) 37.7 (27.6, 49.0) 36.4 (32.3, 40.7) Energy (kcal/day) 1914.4 (1893.9, 1934.8) 1958.1 (1930.8, 1985.5) 1943.1 (1908.5, 1977.7) 1643.1 (1558.5, 1727.8) 1822.0 (1672.0, 1971.9) 1625.6 (1566.9, 1684.3) 1895.3 (1740.5, 2050.2) 1771.8 (1710.8, 1832.9) Data are presented as means or percentages (95% CIs).

9) 34.9 (30.5, 39.7) 37.7 (27.6, 49.0) 36.4 (32.3, 40.7) Energy (kcal/day) 1914.4 (1893.9, 1934.8) 1958.1 (1930.8, 1985.5) 1943.1 (1908.5, 1977.7) 1643.1 (1558.5, 1727.8) 1822.0 (1672.0, 1971.9) 1625.6 (1566.9, 1684.3) 1895.3 (1740.5, 2050.2) 1771.8 (1710.8, 1832.9) Data are presented as means or percentages (95% CIs). *Geometric means are reported for variables with skewed distributions. BMI, body mass index; DASH, Dietary Approaches to Stop Hypertension; HCHS/SOL, Hispanic Community Health Study/Study of Latinos; HOMA-IR, homeostatic model assessment of β-cell function and insulin resistance; SR, self-reported.

9) 34.9 (30.5, 39.7) 37.7 (27.6, 49.0) 36.4 (32.3, 40.7) Energy (kcal/day) 1914.4 (1893.9, 1934.8) 1958.1 (1930.8, 1985.5) 1943.1 (1908.5, 1977.7) 1643.1 (1558.5, 1727.8) 1822.0 (1672.0, 1971.9) 1625.6 (1566.9, 1684.3) 1895.3 (1740.5, 2050.2) 1771.8 (1710.8, 1832.9) Data are presented as means or percentages (95% CIs). *Geometric means are reported for variables with skewed distributions. BMI, body mass index; DASH, Dietary Approaches to Stop Hypertension; HCHS/SOL, Hispanic Community Health Study/Study of Latinos; HOMA-IR, homeostatic model assessment of β-cell function and insulin resistance; SR, self-reported. Figure 1 shows the mean DASH component scores by Hispanic/Latino heritage adjusted by age and sex. (In online supplementary figure, the model was further adjusted by diabetes status.) Overall, the mean DASH score was 41.6. Participants of Mexican descent had the highest mean score (45.0 (95% CI 44.7 to 45.4)), and those of Puerto Rican descent had the lowest (37.6 (95% CI 37.0 to 38.1)). Overall, the DASH food group with the healthiest scores was meat, poultry, eggs, and fish, with a mean score of 9.5 (95% CI 9.4 to 9.5); and the food group with the least healthy scores was sweets (mean score of 1.2 (95% CI 1.1 to 1.3) from a maximum score of 10). Those of Mexican descent had the highest score of total grains (mean 4.3 (95% CI 4.3 to 4.4) from a maximum of 5), vegetables (mean 4.8 (95% CI 4.7 to 5.0)), and low-fat dairy (mean 3.1 (95% CI 3.0 to 3.1) from a maximum of 5). Those of Cuban descent had the highest score of total dairy (mean 3.2 (95% CI 3.1 to 3.3)) and the highest score (lowest consumption) of sweets (mean 1.9 (95% CI 1.8 to 2.1) from a maximum of 5). Those of Central American descent had the highest score of nuts, seeds, and dried beans (mean 5.6 (95% CI 5.3 to 5.9)). Those of Dominican decent had the highest score of fruits (mean 4.8 (95% CI 4.5 to 5.1)) and the highest score (lowest consumption) of meats, poultry, eggs, and fish (mean 9.7 (95% CI 9.6 to 9.7)) and fats and oils (mean 7.5 (95% CI 7.2 to 7.8)). Those of South American descent had the lowest score (highest consumption) of sweets (mean 0.8 (95% CI 0.6 to 1.0)).

core of fruits (mean 4.8 (95% CI 4.5 to 5.1)) and the highest score (lowest consumption) of meats, poultry, eggs, and fish (mean 9.7 (95% CI 9.6 to 9.7)) and fats and oils (mean 7.5 (95% CI 7.2 to 7.8)). Those of South American descent had the lowest score (highest consumption) of sweets (mean 0.8 (95% CI 0.6 to 1.0)). 10.1136/bmjdrc-2017-000402.supp1Supplementary data Figure 1 DASH component score by Hispanic/Latino heritage adjusted by age and sex. 1Adjusted by age and sex (mean age: 41.07, % male: 47.71). 2Each component ranges from 0 to 10, except for grains and dairy, for which each subcomponent ranges from 0 to 5. Data are presented as means±SE. DASH, Dietary Approaches to Stop Hypertension. DASH score and diabetes status

Figure 1 DASH component score by Hispanic/Latino heritage adjusted by age and sex. 1Adjusted by age and sex (mean age: 41.07, % male: 47.71). 2Each component ranges from 0 to 10, except for grains and dairy, for which each subcomponent ranges from 0 to 5. Data are presented as means±SE. DASH, Dietary Approaches to Stop Hypertension. DASH score and diabetes status Table 2 presents adjusted ORs and 95% CIs for the association of DASH dietary pattern (score and tertiles) with diabetes status. Interaction terms for DASH dietary pattern and Hispanic/Latino heritage were not significant; hence, results are pooled. After adjusting for heritage, age, sex, family history of diabetes, current smoking, dietary acculturation, education, income, field center, and energy intake, a higher DASH score was associated with diabetes (OR: 1.08 (95% CI 1.01 to 1.15)) self-reported and unrecognized combined. Further, a similar association was observed after adjusting for BMI and waist circumference (OR: 1.13 (95% CI 1.05 to 1.21)) and after adjusting for HOMA-IR (OR: 1.15 (95% CI 1.06 to 1.24)). However, when distinguishing among self-reported diabetes, whether controlled or not, and unrecognized diabetes, a higher DASH score was only associated with self-reported and controlled diabetes (OR: 1.18 (95% CI 1.02 to 1.36)), and it was higher among those not taking medications (OR: 1.40 (95% CI 1.15 to 1.69)). Table 2 Adjusted ORs (95% CIs) for the association of DASH score (10-unit increment or tertiles) and diabetes status

Table 2 presents adjusted ORs and 95% CIs for the association of DASH dietary pattern (score and tertiles) with diabetes status. Interaction terms for DASH dietary pattern and Hispanic/Latino heritage were not significant; hence, results are pooled. After adjusting for heritage, age, sex, family history of diabetes, current smoking, dietary acculturation, education, income, field center, and energy intake, a higher DASH score was associated with diabetes (OR: 1.08 (95% CI 1.01 to 1.15)) self-reported and unrecognized combined. Further, a similar association was observed after adjusting for BMI and waist circumference (OR: 1.13 (95% CI 1.05 to 1.21)) and after adjusting for HOMA-IR (OR: 1.15 (95% CI 1.06 to 1.24)). However, when distinguishing among self-reported diabetes, whether controlled or not, and unrecognized diabetes, a higher DASH score was only associated with self-reported and controlled diabetes (OR: 1.18 (95% CI 1.02 to 1.36)), and it was higher among those not taking medications (OR: 1.40 (95% CI 1.15 to 1.69)). Table 2 Adjusted ORs (95% CIs) for the association of DASH score (10-unit increment or tertiles) and diabetes status Outcome levels compared DASH (10 units) DASH tertile Medium versus low (less healthy) High (healthiest) versus low (less healthy) OR 95% CI OR 95% CI OR 95% CI Diabetes (two levels) Diabetes versus pre-diabetes and normal glucose Model 1 1.08 (1.01 to 1.15) 0.94 (0.79 to 1.11) 1.20 (1.02 to 1.41) Model 2 1.13 (1.05 to 1.21) 0.95 (0.80 to 1.13) 1.31 (1.11 to 1.54) Model 3 1.15 (1.06 to 1.24) 0.95 (0.79 to 1.14) 1.30 (1.08 to 1.57) Diabetes (three levels) Pre-diabetes versus normal glucose Model 1 0.92 (0.87 to 0.97) 1.02 (0.88 to 1.19) 0.84 (0.73 to 0.97) Model 2 0.95 (0.89 to 1.01) 1.03 (0.89 to 1.20) 0.90 (0.77 to 1.03) Model 3 0.95 (0.90 to 1.01) 1.01 (0.87 to 1.18) 0.89 (0.77 to 1.04) Diabetes versus normal glucose Model 1 1.02 (0.95 to 1.11) 0.95 (0.79 to 1.14) 1.08 (0.91 to 1.29) Model 2 1.09 (1.01 to 1.18) 0.97 (0.80 to 1.18) 1.22 (1.02 to 1.47) Model 3 1.11 (1.01 to 1.21) 0.95 (0.77 to 1.17) 1.20 (0.97 to 1.47) Diabetes (seven levels) Pre-diabetes versus normal glucose Model 1 0.92 (0.87 to 0.97) 1.02 (0.88 to 1.19) 0.84 (0.73 to 0.97) Model 2 0.95 (0.89 to 1.01) 1.03 (0.89 to 1.20) 0.90 (0.77 to 1.03) Model 3 0.95 (0.90 to 1.01) 1.01 (0.87 to 1.18) 0.89 (0.77 to 1.04) Self-report controlled diabetes (medications) versus normal glucose Model 1 1.09 (0.95 to 1.25) 1.09 (0.81 to 1.47) 1.34 (0.98 to 1.83) Model 2 1.17 (1.02 to 1.35) 1.12 (0.82 to 1.51) 1.55 (1.12 to 2.14) Model 3 1.18 (1.02 to 1.36) 1.08 (0.79 to 1.48) 1.50 (1.08 to 2.09) Self-report controlled diabetes (no medications) versus normal glucose Model 1 1.31 (1.09 to 1.58) 1.35 (0.82 to 2.22) 1.63 (1.03 to 2.59) Model 2 1.38 (1.14 to 1.66) 1.40 (0.84 to 2.31) 1.81 (1.14 to 2.88) Model 3 1.40 (1.15 to 1.69) 1.36 (0.81 to 2.29) 1.80 (1.11 to 2.90) Self-report uncontrolled diabetes (medications) versus normal glucose Model 1 1.00 (0.89 to 1.13) 1.00 (0.74 to 1.36) 1.05 (0.79 to 1.37) Model 2 1.08 (0.95 to 1.22) 1.04 (0.76 to 1.41) 1.19 (0.90 to 1.58) Model 3 1.10 (0.96 to 1.25) 1.02 (0.73 to 1.42) 1.18 (0.87 to 1.60) elf-report uncontrolled diabetes (no medications) versus normal glucose Model 1 1.02 (0.78 to 1.34) 0.92 (0.52 to 1.62) 1.02 (0.54 to 1.91) Model 2 1.08 (0.82 to 1.43) 0.95 (0.53 to 1.

Model 2 1.08 (0.95 to 1.22) 1.04 (0.76 to 1.41) 1.19 (0.90 to 1.58) Model 3 1.10 (0.96 to 1.25) 1.02 (0.73 to 1.42) 1.18 (0.87 to 1.60) elf-report uncontrolled diabetes (no medications) versus normal glucose Model 1 1.02 (0.78 to 1.34) 0.92 (0.52 to 1.62) 1.02 (0.54 to 1.91) Model 2 1.08 (0.82 to 1.43) 0.95 (0.53 to 1. 69) 1.15 (0.60 to 2.19) Model 3 1.13 (0.85 to 1.50) 0.94 (0.52 to 1.71) 1.17 (0.61 to 2.26) Unrecognized diabetes versus normal glucose Model 1 0.95 (0.85 to 1.06) 0.79 (0.63 to 0.99) 0.93 (0.73 to 1.19) Model 2 1.00 (0.90 to 1.12) 0.80 (0.64 to 1.01) 1.03 (0.81 to 1.32) Model 3 1.00 (0.89 to 1.13) 0.77 (0.60 to 0.98) 0.98 (0.75 to 1.27) Model 1: DASH + heritage + age + sex + diabetes family history + current smoker + dietary acculturation + education + income + field center + energy intake. Model 2: Model 1 + body mass index + waist circumference. Model 3: Model 2 + HOMA-IR. DASH, Dietary Approaches to Stop Hypertension; HOMA-IR, homeostatic model assessment of β-cell function and insulin resistance. DASH dietary pattern and HOMA-IR

69) 1.15 (0.60 to 2.19) Model 3 1.13 (0.85 to 1.50) 0.94 (0.52 to 1.71) 1.17 (0.61 to 2.26) Unrecognized diabetes versus normal glucose Model 1 0.95 (0.85 to 1.06) 0.79 (0.63 to 0.99) 0.93 (0.73 to 1.19) Model 2 1.00 (0.90 to 1.12) 0.80 (0.64 to 1.01) 1.03 (0.81 to 1.32) Model 3 1.00 (0.89 to 1.13) 0.77 (0.60 to 0.98) 0.98 (0.75 to 1.27) Model 1: DASH + heritage + age + sex + diabetes family history + current smoker + dietary acculturation + education + income + field center + energy intake. Model 2: Model 1 + body mass index + waist circumference. Model 3: Model 2 + HOMA-IR. DASH, Dietary Approaches to Stop Hypertension; HOMA-IR, homeostatic model assessment of β-cell function and insulin resistance. DASH dietary pattern and HOMA-IR Table 3 presents adjusted difference in HOMA-IR by DASH score. Interaction terms for DASH dietary pattern (score or tertiles) with seven-level diabetes status (p>0.8) and with Hispanic/Latino background were not significant (p>0.1). Hence, results were not stratified. A 10-unit higher DASH score was associated with a 4% lower HOMA-IR (95% CI -5.97%–-2.06%) after adjusting for diabetes status, family history of diabetes, heritage background, age, sex, education, income, current smoking, dietary acculturation, field center, and energy intake. After adjusting for BMI and waist circumference, a 10-unit higher DASH score was associated with a 1.74% lower HOMA-IR (95% CI -3.34%–-0.14%). The lower HOMA-IR was slightly larger (6%) among adults in the high DASH tertile (healthier diet) than among those in the lowest tertile (less healthy diet).

and energy intake. After adjusting for BMI and waist circumference, a 10-unit higher DASH score was associated with a 1.74% lower HOMA-IR (95% CI -3.34%–-0.14%). The lower HOMA-IR was slightly larger (6%) among adults in the high DASH tertile (healthier diet) than among those in the lowest tertile (less healthy diet). Table 3 Adjusted difference in HOMA-IR by DASH score DASH (10 units) DASH tertile Medium versus low (less healthy) High (healthiest) versus low (less healthy) % Change* 95% CI % Change* 95% CI % Change* 95% CI Model 1 −4.01 (−5.97 to 2.06) 0.96 (−2.72 to 4.65) −5.96 (−10.48 to 1.43) Model 2 −1.74 (−3.34 to 0.14) 1.21 (−1.92 to 4.34) −1.57 (−5.32 to 2.19) Model 1: DASH + Hispanic/Latino heritage + age + sex + diabetes family history + current smoker + dietary acculturation + education + income + field center + energy intake + diabetes status (seven levels). Model 2: Model 1 + body mass index + waist circumference. *Percent change is calculated as 100×β estimate. DASH, Dietary Approaches to Stop Hypertension; HOMA-IR, homeostatic model assessment of β-cell function and insulin resistance.

DASH (10 units) DASH tertile Medium versus low (less healthy) High (healthiest) versus low (less healthy) % Change* 95% CI % Change* 95% CI % Change* 95% CI Model 1 −4.01 (−5.97 to 2.06) 0.96 (−2.72 to 4.65) −5.96 (−10.48 to 1.43) Model 2 −1.74 (−3.34 to 0.14) 1.21 (−1.92 to 4.34) −1.57 (−5.32 to 2.19) Model 1: DASH + Hispanic/Latino heritage + age + sex + diabetes family history + current smoker + dietary acculturation + education + income + field center + energy intake + diabetes status (seven levels). Model 2: Model 1 + body mass index + waist circumference. *Percent change is calculated as 100×β estimate. DASH, Dietary Approaches to Stop Hypertension; HOMA-IR, homeostatic model assessment of β-cell function and insulin resistance. Discussion To our knowledge, the current study is the first to show an association between adherence to the DASH dietary pattern and diabetes status and insulin resistance among the diverse Hispanic/Latino adult population in this country. The DASH dietary pattern has been considered one of the best eating plans in consecutive years and is currently one of the recommended diets for the management of hypertension.19–21 The benefits of the DASH dietary pattern have been documented by several trials in the management of patients with hypertension and for its benefits for inflammatory markers, insulin resistance, and diabetes.8–10

in consecutive years and is currently one of the recommended diets for the management of hypertension.19–21 The benefits of the DASH dietary pattern have been documented by several trials in the management of patients with hypertension and for its benefits for inflammatory markers, insulin resistance, and diabetes.8–10 An analysis based on the IRAS (which included 548 Latinos) demonstrated an inverse association between adherence to the DASH dietary pattern and the incidence of type 2 diabetes after 5 years of follow-up.14 In our cross-sectional analysis, we showed a positive association between DASH score and diabetes, with a stronger association among those who self-reported diabetes and had it controlled without medications. It is to be expected that adults with self-reported diabetes might be more likely to follow dietary recommendations from their providers and more likely to change their eating habits after being diagnosed with diabetes. We previously published data showing that patients with a pre-existing diagnosis of diabetes and/or hypertension were more likely to report receiving lifestyle behavior recommendations from their providers compared with those without diabetes or hypertension.22 Despite the higher prevalence of diabetes among Hispanics/Latinos of Mexican, Puerto Rican, and Dominican heritages in our study, we also observed a higher DASH score among those of Mexican-American heritage and the lowest score in those of Puerto Rican heritage, despite both groups having a high prevalence of diabetes. This discrepancy may be explained in part due to a difference in the consumption of different components of the DASH diet and not only due to the overall score. For example, Mexican-Americans reported a higher consumption of grains, vegetables, and low-fat dairy, whereas Puerto Ricans reported lower consumption of high-fiber grains and vegetables and lower (healthier) consumption of fat and oils. The main reason for this difference is more complex since our data also showed Cuban-Americans—one of the groups with the lowest prevalence of diabetes—reported a healthier mean consumption of sweets; total dairy; and nuts, seeds, and dry beans. Similarly, participants of South American heritage reported a higher consumption of sweets. Thus, the pathophysiology of diabetes and insulin resistance is very complex and has multiple contributing factors.

alence of diabetes—reported a healthier mean consumption of sweets; total dairy; and nuts, seeds, and dry beans. Similarly, participants of South American heritage reported a higher consumption of sweets. Thus, the pathophysiology of diabetes and insulin resistance is very complex and has multiple contributing factors. Diet plays a significant role, but in the case of Hispanics/Latinos, it might not explain the reported difference in prevalence in diabetes and insulin resistance within this population. Our results, however, confirm previously reported findings that a higher DASH score is associated with a lower insulin resistance (HOMA-IR), in the same way in which the DASH dietary pattern has been associated with decreased levels of insulin resistance and inflammatory markers.13 14 Our findings should be considered in light of the following limitations. First, the cross-sectional design of the study precludes causal conclusions. Second, the dietary information is based on self-reported data and is subject to recall and social desirability biases.23 24 However, this study provides valuable new information about associations between the DASH dietary pattern and diabetes and insulin resistance among the large and diverse Hispanic/Latino population in the USA.

ietary information is based on self-reported data and is subject to recall and social desirability biases.23 24 However, this study provides valuable new information about associations between the DASH dietary pattern and diabetes and insulin resistance among the large and diverse Hispanic/Latino population in the USA. Conclusions In the largest epidemiologic study ever conducted in the USA with a diverse Hispanic/Latino population, participants with self-reported diagnosis of diabetes and unrecognized diabetes reported the highest DASH score. Differences in the DASH score do not completely explain the differences in the prevalence of diabetes within Hispanics/Latinos from different heritage backgrounds. Further, we confirm that a higher DASH score is indeed associated with lower insulin resistance, as previously reported in other segments of the US population. Future research is needed to further elucidate the role of different diet components and diabetes in the Hispanic/Latino population. 10.1136/bmjdrc-2017-000402.supp2Supplementary data The authors thank the more than 16 000 participants who generously gave their time and provided study data. The authors also thank the more than 250 staff members of the HCHS/SOL for their dedication and expertise. The study website is www.cscc.unc.edu/hchs. We thank Morgan Deblecourt from Duke University for editorial assistant in the preparation of this manuscript.

ts who generously gave their time and provided study data. The authors also thank the more than 250 staff members of the HCHS/SOL for their dedication and expertise. The study website is www.cscc.unc.edu/hchs. We thank Morgan Deblecourt from Duke University for editorial assistant in the preparation of this manuscript. The opinions shared by the authors of this manuscript do not represent the opinions of the National Heart, Lung, and BloodInstitute, the National Institutes of Health, or the federal government. Contributors: LC developed the research question, researched the literature, interpreted data, and organized and wrote the manuscript. SSA, MJP, CP, RAEG, CMP, AMSR, LVH, and LAS contributed to the interpretation of data, were involved in manuscript preparation, and critically reviewed and edited the manuscript. DSA and NM Butera analyzed the data and were involved in the manuscript preparation. Competing interests: None declared. Patient consent: Obtained. Ethics approval: IRB at each collaborating institution. Provenance and peer review: Not commissioned; internally peer reviewed.

Contributors: LC developed the research question, researched the literature, interpreted data, and organized and wrote the manuscript. SSA, MJP, CP, RAEG, CMP, AMSR, LVH, and LAS contributed to the interpretation of data, were involved in manuscript preparation, and critically reviewed and edited the manuscript. DSA and NM Butera analyzed the data and were involved in the manuscript preparation. Competing interests: None declared. Patient consent: Obtained. Ethics approval: IRB at each collaborating institution. Provenance and peer review: Not commissioned; internally peer reviewed. Data sharing statement: The data and computer code used for this analysis reside at UNC chapel Hill. The HCHS/SOL fully supports data sharing with outside investigators through processes internal to the study, based on a Data and Materials Distribution Agreement (DMDA) to protect the confidentiality and privacy of the HCHS/SOL participants and their families. Alternatively, de-identified HCHS/SOL data are publically available at BioLINCC and dbGaP for the subset of the study cohort that authorized general use of their data at the time of informed consent.

Significance of this study What is already known about this subject? Inpatient hypoglycemia can be predicted with only a modest degree of accuracy on the basis of body weight, renal function, and hospital insulin doses. What are the new findings? Using a broad number of covariates, including age, weight, admitting service, insulin doses, blood glucose data, diabetes type, renal and liver function, and other admission diagnoses, insulin-associated hypoglycemia can be predicted within a 24-hour time window with a good degree of accuracy. How might these results change the focus of research or clinical practice? Development of a real-time informatics alert from an accurate prediction model could be used clinically to prevent insulin-associated hypoglycemia, a potentially serious complication among hospitalized patients.

What are the new findings? Using a broad number of covariates, including age, weight, admitting service, insulin doses, blood glucose data, diabetes type, renal and liver function, and other admission diagnoses, insulin-associated hypoglycemia can be predicted within a 24-hour time window with a good degree of accuracy. How might these results change the focus of research or clinical practice? Development of a real-time informatics alert from an accurate prediction model could be used clinically to prevent insulin-associated hypoglycemia, a potentially serious complication among hospitalized patients. Hypoglycemia is a common occurrence in hospitalized patients and is linked to multiple adverse clinical outcomes and mortality.1 Acute hypoglycemia can provoke cardiac ischemia and arrhythmias, as well as neurologic harm ranging in severity from altered cognition or irritability to focal neurologic deficits, loss of consciousness, stroke, seizures, and coma.2 Besides these potentially life-threatening complications, hypoglycemia can be a source of patient dissatisfaction and worry.1 Approximately half of hypoglycemic events in the hospital are iatrogenic, usually resulting from insulin treatment.3 In observational studies of hospitalized patients, both spontaneous and iatrogenic hypoglycemia are associated with increased mortality.3–6 Considering that 20%–40% of hospitalized patients require glucose-lowering medications,1 prevention of iatrogenic hypoglycemia is a significant patient safety issue and a major challenge to our healthcare system.

italized patients, both spontaneous and iatrogenic hypoglycemia are associated with increased mortality.3–6 Considering that 20%–40% of hospitalized patients require glucose-lowering medications,1 prevention of iatrogenic hypoglycemia is a significant patient safety issue and a major challenge to our healthcare system. Insulin is the recommended therapy for glycemic management in the non-critical care setting, with discontinuation of non-insulin antihyperglycemic agents encouraged for the majority of patients.7 In contrast to the intensive care unit (ICU), where insulin adjustments may be driven by nurse-managed protocols, insulin titration in the non-ICU setting is prescriber-driven and requires evaluation of a complex set of clinical, laboratory, and pharmacologic parameters. Unfortunately, therapeutic inertia—failure to reduce or modify insulin therapy in patients with downward trending blood glucose (BG) readings—is a common cause of insulin-associated hypoglycemia. In over 60% of severe hypoglycemic events, antecedent mild hypoglycemia was observed without any change in diabetes medications.8 Even more concerning, clinicians often fail to modify insulin doses in patients who experience overt hypoglycemia. A retrospective study found that only 44% of patients had the recommended 20% reduction in the insulin total daily dose following a hypoglycemic event.9

bserved without any change in diabetes medications.8 Even more concerning, clinicians often fail to modify insulin doses in patients who experience overt hypoglycemia. A retrospective study found that only 44% of patients had the recommended 20% reduction in the insulin total daily dose following a hypoglycemic event.9 The purpose of this study was to develop and validate a prediction model for insulin-associated hypoglycemia in non-critically ill hospitalized adults. A previous logistic regression model developed using data from 3028 inpatients achieved 54% sensitivity at detecting hypoglycemia using a BG cut-off of 60 mg/dL, missing 46% of acute hypoglycemic events.10 By using a much larger data set and more predictor variables, we hoped to achieve a model with greater predictive accuracy, moving closer to the goal of a real-time alerting or reporting system integrated into the electronic medical record (EMR) to identify inpatients at high risk of incident insulin-associated hypoglycemia in a clinically relevant time window that would permit prophylactic changes to the insulin regimen.

edictive accuracy, moving closer to the goal of a real-time alerting or reporting system integrated into the electronic medical record (EMR) to identify inpatients at high risk of incident insulin-associated hypoglycemia in a clinically relevant time window that would permit prophylactic changes to the insulin regimen. Research design and methods Data source This was a cross-sectional study conducted at Johns Hopkins Hospital, a 1300-bed tertiary care academic medical center in Baltimore, Maryland. Using data from our prior EMR, Sunrise POE, we identified hospitalized adults in the non-critical care, non-obstetrical setting with an admission date on or after 1 January 2013 and a discharge date on or before 31 December 2015. Admissions missing a date of discharge or patient weight measurements were excluded. The primary exposure of interest was treatment with subcutaneous insulin, defined as any administered long-acting, intermediate-acting, rapid-acting, or premixed insulin. At our institution, the formulary long-acting insulin is glargine, the intermediate-acting insulin is Neutral Protamine Hagedorn (NPH), the rapid-acting insulin is aspart, and the premixed insulin is NPH/regular 70/30. Admissions in which subcutaneous insulin was not administered at any time during hospitalization were excluded.

titution, the formulary long-acting insulin is glargine, the intermediate-acting insulin is Neutral Protamine Hagedorn (NPH), the rapid-acting insulin is aspart, and the premixed insulin is NPH/regular 70/30. Admissions in which subcutaneous insulin was not administered at any time during hospitalization were excluded. For eligible admissions, we defined patient-days as 24-hour intervals relative to the admission date and time (online supplementary figure S1). We defined insulin-treated patient-days as patient-days in which at least 1 unit of subcutaneous insulin was administered. We excluded those insulin-treated patient-days in which there was concurrent use of intravenous insulin (as these patients are already under close observation and managed via protocol with hourly BG checks), total parenteral nutrition (TPN; as intravenous insulin is often an additive in the parenteral nutrition), insulin pump use (as we were unable to easily capture patient-administered insulin pump doses from our EMR), or if the insulin-treated patient-day was the date of discharge (as no outcome data would be available). All remaining insulin-treated patient-days were considered index days used for prediction of hypoglycemic outcomes on the next patient-day. Although non-index days were excluded as a unit of observation for prediction, they were included as observations for outcome ascertainment if the previous day was an index day. 10.1136/bmjdrc-2017-000499.supp1Supplementary file 1

For eligible admissions, we defined patient-days as 24-hour intervals relative to the admission date and time (online supplementary figure S1). We defined insulin-treated patient-days as patient-days in which at least 1 unit of subcutaneous insulin was administered. We excluded those insulin-treated patient-days in which there was concurrent use of intravenous insulin (as these patients are already under close observation and managed via protocol with hourly BG checks), total parenteral nutrition (TPN; as intravenous insulin is often an additive in the parenteral nutrition), insulin pump use (as we were unable to easily capture patient-administered insulin pump doses from our EMR), or if the insulin-treated patient-day was the date of discharge (as no outcome data would be available). All remaining insulin-treated patient-days were considered index days used for prediction of hypoglycemic outcomes on the next patient-day. Although non-index days were excluded as a unit of observation for prediction, they were included as observations for outcome ascertainment if the previous day was an index day. 10.1136/bmjdrc-2017-000499.supp1Supplementary file 1 Outcomes The primary outcomes were biochemical or clinically significant hypoglycemia, defined as at least one serum or fingerstick BG of ≤70 mg/dL and <54 mg/dL occurring in a prediction horizon of 24 hours after an index day, respectively.11 Since the degree of hypoglycemia might be influenced by different clinical predictors, we developed separate prediction models for each hypoglycemic outcome.

emia, defined as at least one serum or fingerstick BG of ≤70 mg/dL and <54 mg/dL occurring in a prediction horizon of 24 hours after an index day, respectively.11 Since the degree of hypoglycemia might be influenced by different clinical predictors, we developed separate prediction models for each hypoglycemic outcome. Predictors Candidate predictors of hypoglycemia were selected based on clinical knowledge and previous studies, and with consideration of ease of data extraction from our EMR. Online supplementary table S1 summarizes the definitions, data sources, and timing of collection for each of the predictor variables. Demographic predictors included age, sex, race, and admitting service (medical vs surgical). At our institution, consistent with current practice guidelines, it is recommended that non-insulin antihyperglycemic medications be discontinued and insulin therapy initiated for patients with hyperglycemia persisting for 24–48 hours.7 Therefore, with the exception of subcutaneous insulin, we did not collect information about other hospital-administered antihyperglycemic medications. Administered subcutaneous insulin was categorized as basal, nutritional, and correctional. Although insulin doses were normalized per body weight (unit/kg/index day) in our prediction models, weight was included separately as an independent variable.

tion about other hospital-administered antihyperglycemic medications. Administered subcutaneous insulin was categorized as basal, nutritional, and correctional. Although insulin doses were normalized per body weight (unit/kg/index day) in our prediction models, weight was included separately as an independent variable. For most insulin-treated hospitalized patients on medical/surgical wards, BG measurements are typically obtained four times daily (before each meal and at bedtime) or every 4 hours if nil per os (NPO). We evaluated several glycemic measures as predictors of incident hypoglycemia. Mean BG and coefficient of variation of BG (CVBG) and nadir BG were summarized both on the index day and for all inpatient days up to and including the index day. Both index day and admission-level glycemic measures were evaluated as predictors.

valuated several glycemic measures as predictors of incident hypoglycemia. Mean BG and coefficient of variation of BG (CVBG) and nadir BG were summarized both on the index day and for all inpatient days up to and including the index day. Both index day and admission-level glycemic measures were evaluated as predictors. Using relevant International Classification of Diseases (ICD-9) diagnostic codes, we generated categories of clinical conditions that could affect glucose regulation. We categorized type 1 diabetes mellitus and postsurgical hypoinsulinemia as insulin-deficient states. Without information about home insulin use, we could not identify patients with type 2 diabetes who were insulin-deficient. Acute kidney injury (AKI) and chronic kidney disease (CKD) predispose to hypoglycemia via reduced insulin clearance and reduced gluconeogenesis.12 Liver failure causes hyperglycemia due to increased insulin resistance13 and in decompensated failure may cause hypoglycemia due to reduced gluconeogenesis.14 15 Chronic alcohol use can cause hypoglycemia due to depletion of glycogen stores from prolonged fasting and inhibition of gluconeogenesis.16 We combined alcohol dependency and end-stage liver disease into a category of liver disease. Congestive heart failure is another less common hypoglycemic risk factor.17–19 Acute pancreatitis and pancreatic cancer have both been associated with hyperglycemia (ie, pancreatogenous diabetes).20 Given the low prevalence of each of these conditions, we combined them into one category. Steroids, which are commonly used in hospitalized patients, contribute to hyperglycemia via insulin resistance.21 We evaluated systemic steroid use as a binary predictor, but did not have information about steroid dose to evaluate the effect of steroid tapers on insulin requirements and hypoglycemia risk. Given the very low ICD coding for sepsis among diagnoses (0.07% of index days), sepsis was not included as a predictor in this non-critically ill population. Adrenal insufficiency is another known hypoglycemic risk factor; however, coding for this condition was virtually absent.

ulin requirements and hypoglycemia risk. Given the very low ICD coding for sepsis among diagnoses (0.07% of index days), sepsis was not included as a predictor in this non-critically ill population. Adrenal insufficiency is another known hypoglycemic risk factor; however, coding for this condition was virtually absent. Reduced carbohydrate intake following insulin administration is a common cause of hypoglycemia. In addition to diet orders reflecting varying degrees of carbohydrate intake, we created a category of digestive diseases to include several conditions (nausea, vomiting, abdominal pain, intestinal obstruction) that could influence carbohydrate intake. Although we considered using diagnostic codes associated with malnourishment (eg, failure to thrive, cachexia), coding for these conditions was exceedingly low.

tegory of digestive diseases to include several conditions (nausea, vomiting, abdominal pain, intestinal obstruction) that could influence carbohydrate intake. Although we considered using diagnostic codes associated with malnourishment (eg, failure to thrive, cachexia), coding for these conditions was exceedingly low. Missing data After excluding admissions lacking discharge date and weight information, we had nearly complete information about the predictors and outcomes for our analysis. The variables missing information were admission diagnosis codes and renal laboratory data on patient day 1, which were missing from 10.1% and 30.3% of records, respectively. Patient-days missing admission diagnoses were included in the analysis; however, the diagnoses used to classify clinical conditions as predictor variables were assumed to be absent for these records. Similarly, we assumed that most patients with CKD would have had laboratory assessment of renal function obtained within the first 24 hours of admission. If a glomerular filtration rate measurement was missing, the patient-day was included in the analysis, but the patient was classified as not having CKD.

hese records. Similarly, we assumed that most patients with CKD would have had laboratory assessment of renal function obtained within the first 24 hours of admission. If a glomerular filtration rate measurement was missing, the patient-day was included in the analysis, but the patient was classified as not having CKD. Statistical analysis and methods We adhered to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis guidelines in the design of this study and reporting of results.22 We used split-sample internal validation with 70% and 30% of the index days used for model development and validation, respectively. Records were sorted chronologically by discharge date, and the discharge date corresponding approximately to the 70th percentile of records was used to split the cohort such that it did not partition an individual patient admission. This approach was selected because it allows for non-random variation that would be encountered in the real-world clinical setting if such a model were to be used prospectively for real-time event prediction.

centile of records was used to split the cohort such that it did not partition an individual patient admission. This approach was selected because it allows for non-random variation that would be encountered in the real-world clinical setting if such a model were to be used prospectively for real-time event prediction. A detailed description of our model building strategy is provided in the online supplementary materials. In brief, we used a combination of various selection processes, including trial and error, automated stepwise selection, and evaluation of the information criteria to develop two multivariable logistic regression models: model 1 for prediction of a BG ≤70 mg/dL (biochemical hypoglycemia) and model 2 for a BG <54 mg/dL (clinically significant hypoglycemia). Using the multivariable regression equations for the final models, the probability of hypoglycemia was calculated in the development data set. An empirical probability cut-point was selected using Youden’s Index, which maximizes the sum of sensitivity and specificity.23 The probability of hypoglycemia was then calculated in the validation cohort using the coefficients from the regression equation derived from the development data set. The empirical probability cut-points were then applied to the validation data set, such that a predicted probability at or above the probability cut-point was used to classify the index day as ‘at risk’ for hypoglycemic event.

rt using the coefficients from the regression equation derived from the development data set. The empirical probability cut-points were then applied to the validation data set, such that a predicted probability at or above the probability cut-point was used to classify the index day as ‘at risk’ for hypoglycemic event. Model performance was assessed by comparing true disease status versus classified risk in the validation set. Discrimination was assessed using the c-statistic, which is equal to the area under the receiver operating characteristic curve. c-Statistic values >0.7, >0.8, and >0.9 are considered acceptable, excellent, and outstanding discrimination, respectively.24 Positive and negative predictive values, which are dependent on disease prevalence, indicate the probability of a positive/negative test result among those with/without the disease, respectively.25 Positive and negative likelihood ratios are test characteristics that are independent of disease prevalence and provide information about whether a test result changes the probability of the outcome.25 Positive likelihood ratios of >2, >5, and >10 indicate slight, moderate, and large increases in the likelihood of the outcome with a positive result, respectively; conversely, negative likelihood ratios of <0.5, <0.2, and <0.1 indicate slight, moderate, and large decreases in the likelihood of the outcome with a negative result.26 Statistical analyses were performed using Stata Statistical Software V.14.2. P<0.05 was considered statistically significant.

Model performance was assessed by comparing true disease status versus classified risk in the validation set. Discrimination was assessed using the c-statistic, which is equal to the area under the receiver operating characteristic curve. c-Statistic values >0.7, >0.8, and >0.9 are considered acceptable, excellent, and outstanding discrimination, respectively.24 Positive and negative predictive values, which are dependent on disease prevalence, indicate the probability of a positive/negative test result among those with/without the disease, respectively.25 Positive and negative likelihood ratios are test characteristics that are independent of disease prevalence and provide information about whether a test result changes the probability of the outcome.25 Positive likelihood ratios of >2, >5, and >10 indicate slight, moderate, and large increases in the likelihood of the outcome with a positive result, respectively; conversely, negative likelihood ratios of <0.5, <0.2, and <0.1 indicate slight, moderate, and large decreases in the likelihood of the outcome with a negative result.26 Statistical analyses were performed using Stata Statistical Software V.14.2. P<0.05 was considered statistically significant. Results Characteristics of study population Among 120 224 patient admissions during the study period, 28 899 (24%) had administration of at least 1 unit of subcutaneous insulin (figure 1). Of these eligible admissions, there were a total of 250 747 patient-days, of which 152 821 (61%) were insulin-treated patient-days. After excluding discharge days and days with intravenous insulin, TPN, or insulin pump orders, 128 657 (84%) of the insulin-treated days were identified as index days for prediction of a hypoglycemic outcome within the next 24 hours.

tal of 250 747 patient-days, of which 152 821 (61%) were insulin-treated patient-days. After excluding discharge days and days with intravenous insulin, TPN, or insulin pump orders, 128 657 (84%) of the insulin-treated days were identified as index days for prediction of a hypoglycemic outcome within the next 24 hours. Figure 1 Study flow chart. ICU, intensive care unit; TPN, total parenteral nutrition. The baseline characteristics and hypoglycemic outcomes of the study cohort are shown in table 1. The prevalence of biochemical and clinically significant hypoglycemia was 4.2% and 1.2% of patient-days, respectively. The median length of stay was 6 days, with a slightly higher admission rate to surgical compared with medical services. The median age was 61.2 years, with a slight male predominance. The majority of patients were white (53.8%), followed by black (36.2%), other races (7.5%), and Asian (2.5%). Type 2 diabetes and insulin-deficient diabetes (type 1/pancreatectomy) were present in 56.4% and 4.2% of admissions, respectively. Table 1 Baseline characteristics and hypoglycemic outcomes of study population

The baseline characteristics and hypoglycemic outcomes of the study cohort are shown in table 1. The prevalence of biochemical and clinically significant hypoglycemia was 4.2% and 1.2% of patient-days, respectively. The median length of stay was 6 days, with a slightly higher admission rate to surgical compared with medical services. The median age was 61.2 years, with a slight male predominance. The majority of patients were white (53.8%), followed by black (36.2%), other races (7.5%), and Asian (2.5%). Type 2 diabetes and insulin-deficient diabetes (type 1/pancreatectomy) were present in 56.4% and 4.2% of admissions, respectively. Table 1 Baseline characteristics and hypoglycemic outcomes of study population Variable Results reported per Development data set Validation data set Full cohort Admission Index day Insulin-treated patient-days, n x 90 059 38 598 128 657 Patient admissions, n x 18 867 7399 26 266 Unique patients, n x 13 360 5902 18 196 Hypoglycemic outcome, n (%) Any BG ≤70 mg/dL x 3935 (4.4) 1517 (3.9) 5452 (4.2) Any BG <54 mg/dL x 1081 (1.2) 415 (1.1) 1496 (1.2) Nadir BG, mg/dL x 61 (52–66) 61 (53–66) 61 (52–66) LOS, days x 6 (3–10) 6 (4–11) 6 (3–11) Admitting service, n (%) x Medicine x 9115 (48.3) 3324 (44.9) 12 439 (47.4) Surgery x 9752 (51.7) 4075 (55.1) 13 827 (52.6) Age, years x 61.1 (51.3–70.2) 61.5 (52.2–69.8) 61.2 (51.6–70.1) Sex, male/female (%) x 52.5/47.5 54.1/45.9 52.9/47.1 Race, n (%) x White x 10 125 (53.7) 3997 (54.0) 14 122 (53.8) Black x 6923 (36.7) 2589 (35.0) 9512 (36.2) Asian x 428 (2.3) 224 (3.0) 652 (2.5) Other x 1391 (7.4) 589 (8.0) 1980 (7.5) Admission weight, kg x 81.6 (67.8–98.1) 82.2 (68.2–98.0) 81.7 (68.0–98.1) Insulin doses, unit/kg/day Total daily dose x 0.10 (0.03–0.32) 0.11 (0.03–0.32) 0.10 (0.03–0.32) Basal x 0.00 (0.00–0.15) 0.00 (0.00–0.16) 0.00 (0.00–0.16) Nutritional x 0.00 (0.00–0.00) 0.00 (0.00–0.00) 0.00 (0.00–0.00) Correctional x 0.05 (0.02–0.11) 0.05 (0.02–0.11) 0.05 (0.02–0.11) High-dose SSI, n (%) x 53 488 (59.4) 23 406 (60.6) 76 894 (59.8) Diet orders, n (%) x Nil per os x 20 362 (22.6) 8765 (22.7) 29 127 (22.6) Carbohydrate-controlled x 8832 (9.8) 3450 (8.9) 12 282 (9.6) Regular or other diet x 58 198 (64.6) 25 034 (64.9) 82 232 (64.7) Bolus tube feeds x 2667 (3.0) 1349 (3.5) 4016 (3.1) Steroid use, n (%) x 39 451 (43.7) 17 809 (46.0) 57 260 (44.4) Type 1 diabetes/pancreatectomy, n (%) x 900 (4.8) 217 (2.9) 1117 (4.3) Type 2 diabetes, n (%) x 11 357 (60.2) 3434 (46.4) 14 791 (56.3) Acute kidney injury, n (%) x 191 (1.0) 41 (0.6) 232 (0.9) Chronic kidney disease, n (%) x Stage 3 x 3502 (18.6) 421 (5.7) 3921 (14.9) Stage 4 x 1287 (6.8) 149 (2.0) 1450 (5.5) Stage 5 x 919 (4.9) 89 (1.2) 1016 (3.9) Liver failure, n (%) x 153 (0.8) 58 (0.8) 211 (0.8) Congestive heart failure, n (%) x 431 (2.3) 93 (1.3) 524 (2.0) Digestive disease, n (%) x 935 (5.0) 294 (4.0) 1229 (4.7) Pancreatic disease, n (%) x 468 (2.5) 144 (2.

1 (5.7) 3921 (14.9) Stage 4 x 1287 (6.8) 149 (2.0) 1450 (5.5) Stage 5 x 919 (4.9) 89 (1.2) 1016 (3.9) Liver failure, n (%) x 153 (0.8) 58 (0.8) 211 (0.8) Congestive heart failure, n (%) x 431 (2.3) 93 (1.3) 524 (2.0) Digestive disease, n (%) x 935 (5.0) 294 (4.0) 1229 (4.7) Pancreatic disease, n (%) x 468 (2.5) 144 (2. 0) 612 (2.3) Index day glycemic measures Nadir BG, mg/dL x 121 (100–147) 122 (101–148) 121 (100–147) Mean BG, mg/dL x 158 (134–194) 160 (136–197) 158 (135–195) CV of BG, % x 19.2 (12.8–27.8) 18.9 (12.8–27.1) 19.1 (12.8–27.6) Number of BG measurements x 5 (4–7) 6 (5–7) 5 (5–7) Admission glycemic measures* Nadir BG, mg/dL x 91 (73–115) 91 (73–113) 91 (73–114) Mean BG, mg/dL x 161 (139–194) 163 (141–196) 162 (139–195) CV of BG, % x 24.8 (19.0–32.5) 25.2 (19.3–32.2) 25.0 (19.1–32.4) *Summarized for all patient-days prior to and including index day. Data are median (interquartile range) and n (%). BG, blood glucose; CV, coefficient of variation; LOS, length of stay; SSI, sliding scale insulin.

0) 612 (2.3) Index day glycemic measures Nadir BG, mg/dL x 121 (100–147) 122 (101–148) 121 (100–147) Mean BG, mg/dL x 158 (134–194) 160 (136–197) 158 (135–195) CV of BG, % x 19.2 (12.8–27.8) 18.9 (12.8–27.1) 19.1 (12.8–27.6) Number of BG measurements x 5 (4–7) 6 (5–7) 5 (5–7) Admission glycemic measures* Nadir BG, mg/dL x 91 (73–115) 91 (73–113) 91 (73–114) Mean BG, mg/dL x 161 (139–194) 163 (141–196) 162 (139–195) CV of BG, % x 24.8 (19.0–32.5) 25.2 (19.3–32.2) 25.0 (19.1–32.4) *Summarized for all patient-days prior to and including index day. Data are median (interquartile range) and n (%). BG, blood glucose; CV, coefficient of variation; LOS, length of stay; SSI, sliding scale insulin. Overall, administered insulin doses were relatively low, with median (IQR) total daily dose (TDD) of 0.10 (0.03–0.32) units/kg. The majority of subcutaneous insulin was provided in the form of basal and correctional insulin. Despite hospital protocols and computerized ordersets encouraging use of a basal-bolus insulin regimen, use of nutritional insulin was exceedingly low. For example, the 90th percentile of nutritional insulin dose was only 0.17 units/kg/day. High-dose sliding scale insulin (SSI) was used in 59.8% of index days. There was a high prevalence of steroid use on the index day (44.4%).

aging use of a basal-bolus insulin regimen, use of nutritional insulin was exceedingly low. For example, the 90th percentile of nutritional insulin dose was only 0.17 units/kg/day. High-dose sliding scale insulin (SSI) was used in 59.8% of index days. There was a high prevalence of steroid use on the index day (44.4%). Regarding conditions associated with hypoglycemia, CKD was most prevalent, with 14.9%, 5.5%, and 3.9% having stage 3, stage 4, and stage 5 CKD, respectively. AKI was an admission diagnosis in 0.9%. Digestive diseases affecting nutritional intake were present in 4.7%. Liver diseases and congestive heart failure (CHF) were admission diagnoses in 0.8% and 2.0%, respectively. NPO diet was ordered on 22.6% of index days. With respect to glycemic measures, the mean and nadir BG on the index day were 158 mg/dL and 121 mg/dL, and the mean and nadir BG during admission and up to index day were 162 mg/dL and 91 mg/dL, respectively. As expected, glycemic variability (CVBG) was greater when summarized on all hospital days up to the index day compared with the index day alone (25.0% vs 19.1%). Overall, the development and validation data sets were very similar with respect to the clinical predictors, with the exception of diabetes and CKD, which were more prevalent in the development data set.

greater when summarized on all hospital days up to the index day compared with the index day alone (25.0% vs 19.1%). Overall, the development and validation data sets were very similar with respect to the clinical predictors, with the exception of diabetes and CKD, which were more prevalent in the development data set. Model parameters Table 2 shows the fully adjusted models with ORs, coefficients, and intercepts. The univariate associations for each predictor are shown in online supplementary table S2. For biochemical hypoglycemia, there were a total of 3935 events and 44 variables in the model, with an event per variable (EPV) ratio of 89. For clinically significant hypoglycemia, there were a total of 1081 events and 35 variables, with an EPV ratio of 30.9. Thus, both models exceeded the recommended EPV of 10 or more. Table 2 Logistic regression models with ORs and coefficients from validation data sets

Model parameters Table 2 shows the fully adjusted models with ORs, coefficients, and intercepts. The univariate associations for each predictor are shown in online supplementary table S2. For biochemical hypoglycemia, there were a total of 3935 events and 44 variables in the model, with an event per variable (EPV) ratio of 89. For clinically significant hypoglycemia, there were a total of 1081 events and 35 variables, with an EPV ratio of 30.9. Thus, both models exceeded the recommended EPV of 10 or more. Table 2 Logistic regression models with ORs and coefficients from validation data sets Unit of change Spline knot Model 1: BG≤70 mg/dL Model 2: BG<54 mg/dL OR (95% CI) Coefficient OR (95% CI) Coefficient Intercept 2.140 0.580 Age 10 years Age1 ≤40 1.07 (0.96 to 1.20) 0.066 – – Age2 >40 0.96 (0.92 to 0.99) −0.045 – – Female – – – – 0.76 (0.66 to 0.86) −0.280 Weight 10 kg Weight1 ≤80 0.87 (0.84 to 0.91) −0.136 0.85 (0.80 to 0.91) −0.160 Weight2 >80 0.92 (0.90 to 0.95) −0.079 0.86 (0.81 to 0.91) −0.151 Admission to surgical service 0.91 (0.85 to 0.98) −0.091 – – Basal insulin dose 0.1 units/kg Basal1 ≤0.2 1.85 (1.75 to 1.97) 0.617 2.01 (1.79 to 2.25) 0.697 Basal2 ≤0.8 1.14 (1.10 to 1.17) 0.130 1.15 (1.09 to 1.21) 0.141 Basal3 ≤1.3 0.95 (0.85 to 1.06) −0.050 0.95 (0.80 to 1.14) −0.046 Basal4 ≤1.6 1.29 (0.92 to 1.78) 0.249 1.84 (1.15 to 2.95) 0.609 Basal5 >1.6 0.95 (0.71 to 1.28) 0.755 0.29 (0.07 to 1.28) −1.238 Nutritional insulin dose 0.1 units/kg Nutritional1 ≤0.6 1.05 (1.02 to 1.08) 0.048 1.05 (1.00 to 1.10) 0.049 Nutritional2 ≤0.9 1.05 (0.90 to 1.23) 0.048 1.04 (0.80 to 1.33) 0.035 Nutritional3 ≤1.1 0.59 (0.36 to 0.94) −0.534 0.51 (0.22 to 1.17) −0.674 Nutritional4 >1.1 1.07 (0.86 to 1.35) 0.071 1.19 (0.86 to 1.65) 0.176 Correctional insulin dose 0.1 units/kg Correctional1 ≤0.04 0.69 (0.49 to 0.97) −0.365 – – Correctional2 ≤0.7 1.04 (0.99 to 1.10) 0.041 – – Correctional3 ≤0.9 0.71 (0.39 to 1.30) −0.336 – – Correctional4 >0.9 1.17 (0.79 to 1.72) 0.159 – – High-dose SSI 1.14 (1.03 to 1.27) 0.135 1.18 (0.97 to 1.45) 0.169 Index day mean BG 10 mg/dL Mean1 ≤100 0.73 (0.64 to 0.83) −0.312 0.80 (0.66 to 0.97) −0.220 Mean2 ≤150 0.86 (0.83 to 0.89) −0.155 0.89 (0.84 to 0.94) −0.122 Mean3 >150 0.99 (0.97 to 1.01) −0.011 1.03 (1.01 to 1.05) 0.030 Index day nadir BG 10 mg/dL Nadir1 ≤88 0.99 (0.95 to 1.03) −0.011 0.91 (0.85 to 0.97) −0.099 Nadir2 ≤100 0.64 (0.58 to 0.71) −0.440 0.80 (0.66 to 0.96) −0.228 Nadir3 >100 0.93 (0.91 to 0.95) −0.070 0.93 (0.89 to 0.96) −0.076 Index day CV of BG 10% CV1 ≤10 0.93 (0.61 to 1.42) −0.077 1.04 (0.44 to 2.47) 0.035 CV2 ≤20 1.72 (1.46 to 2.02) 0.541 1.51 (1.10 to 2.09) 0.417 CV3 >20 1.01 (0.96 to 1.05) 0.006 1.02 (0.94 to 1.10) 0.017 Admission nadir BG 10 mg/dL Admission nadir1 ≤100 0.87 (0.86 to 0.89) −0.134 0.86 (0.83 to 0

3 (0.89 to 0.96) −0.076 Index day CV of BG 10% CV1 ≤10 0.93 (0.61 to 1.42) −0.077 1.04 (0.44 to 2.47) 0.035 CV2 ≤20 1.72 (1.46 to 2.02) 0.541 1.51 (1.10 to 2.09) 0.417 CV3 >20 1.01 (0.96 to 1.05) 0.006 1.02 (0.94 to 1.10) 0.017 Admission nadir BG 10 mg/dL Admission nadir1 ≤100 0.87 (0.86 to 0.89) −0.134 0.86 (0.83 to 0 .90) −0.147 Admission nadir2 ≤400 1.03 (1.01 to 1.05) 0.029 1.04 (1.00 to 1.07) 0.035 Admission nadir3 >400 0.84 (0.58 to 1.19) −0.180 0.88 (0.63 to 1.24) −0.122 Admission CV of BG 10% Admission cv1 ≤18 0.78 (0.63 to 0.96) −0.249 0.73 (0.49 to 1.09) −0.311 Admission cv2 >18 1.19 (1.14 to 1.24) 0.171 1.17 (1.10 to 1.25) 0.159 Diet orders NPO 1.00 (ref) – 1.00 (ref) – Carbohydrate-controlled 0.80 (0.71 to 0.92) −0.218 0.67 (0.53 to 0.84) −0.406 Regular or other 0.79 (0.72 to 0.86) −0.235 0.69 (0.59 to 0.81) −0.373 Bolus tube feeds 0.68 (0.54 to 0.86) −0.382 0.76 (0.51 to 1.15) −0.268 Type 1 diabetes/pancreatectomy 1.43 (1.28 to 1.59) 0.357 2.04 (1.73 to 2.39) 0.712 Type 2 diabetes 1.25 (1.13 to 1.38) 0.224 1.29 (1.07 to 1.55) 0.253 AKI 1.27 (0.99 to 1.63) 0.238 – – CKD None 1.00 (ref) – 1.00 (ref) – Stage 3 1.25 (1.15 to 1.37) 0.224 1.37 (1.16 to 1.61) 0.314 Stage 4 1.52 (1.37 to 1.69) 0.418 1.80 (1.49 to 2.17) 0.588 Stage 5 1.76 (1.56 to 1.99) 0.565 2.21 (1.80 to 2.73) 0.795 Liver disease 0.70 (0.48 to 1.01) −0.363 – – Digestive disease 1.20 (1.04 to 1.39) 0.184 – – Steroids on index day – – 0.84 (0.74 to 0.96) −0.170 Bolded values signify P<0.05.

etection of clinically significant compared with biochemical hypoglycemia. Since all of the predictor variables are readily available in the EMR, our models could be used to develop a real-time informatics alert to prevent insulin-associated hypoglycemia in hospitalized patients, a potentially serious clinical outcome. Among candidate predictors, we found that glycemic summary measures were the strongest predictors. Index day mean, nadir, and CVBG were strongly associated with odds of hypoglycemia, as were the admission nadir BG and variability. Basal insulin doses were modestly associated with increased risk, but only for doses ≤0.8 units/kg/day. We suspect that patients requiring basal doses beyond this had some underlying cause of severe insulin resistance, such as morbid obesity or high-dose glucocorticoid use, which may have provided protection against hypoglycemia. Interestingly, nutritional insulin doses were only weakly associated with increased risk in the range <0.6 unit/kg/day, possibly owing to the overall low use of nutritional insulin in this cohort and lack of adequate power to detect effect sizes at larger doses. While use of high-dose SSI was associated with increased risk, administered correctional insulin doses were not.

.37 (1.16 to 1.61) 0.314 Stage 4 1.52 (1.37 to 1.69) 0.418 1.80 (1.49 to 2.17) 0.588 Stage 5 1.76 (1.56 to 1.99) 0.565 2.21 (1.80 to 2.73) 0.795 Liver disease 0.70 (0.48 to 1.01) −0.363 – – Digestive disease 1.20 (1.04 to 1.39) 0.184 – – Steroids on index day – – 0.84 (0.74 to 0.96) −0.170 Bolded values signify P<0.05. AKI, acute kidney; BG, blood glucose; CKD, chronic kidney disease; CV, coefficient of variation; NPO, nil per os; ref, reference; SSI, sliding scale insulin. Predictors associated with reduced risk of hypoglycemia in one or both models included female sex, increasing age over 40 years, admission to surgical service, higher weight, higher index day mean and nadir BG, higher admission nadir BG, non-NPO diet, liver disease, and steroid use. Predictors associated with increased risk of hypoglycemia in one or both models included increasing basal insulin doses (in range of ≤0.8 units/kg), increasing nutritional insulin doses (in range of ≤0.6 units/kg), use of high-dose SSI, higher index day and admission CVBG, NPO diet, type 1 diabetes/pancreatectomy, type 2 diabetes, CKD, and digestive diseases. There was significantly lower total insulin use in surgical compared with medical patients, with median (IQR) doses of 0.07 (0.02–0.24) and 0.14 (0.04–0.38) unit/kg/day, respectively (P<0.001). Lower insulin doses in surgical patients may have been related to higher prevalence of NPO status in this group: 54.8% of index days were NPO in surgery patients compared with 45.2% in medical patients (P<0.001).

ents, with median (IQR) doses of 0.07 (0.02–0.24) and 0.14 (0.04–0.38) unit/kg/day, respectively (P<0.001). Lower insulin doses in surgical patients may have been related to higher prevalence of NPO status in this group: 54.8% of index days were NPO in surgery patients compared with 45.2% in medical patients (P<0.001). Glycemic measures were the strongest predictors of hypoglycemia risk. For example, in model 1, for each 10 mg/dL increase in index day mean BG, the reductions in the adjusted odds of hypoglycemia were 27%, 14%, and 1% in the ranges of BG of ≤100 mg/dL, >100 and ≤150 mg/dL, and >150 mg/dL, respectively. Similarly, each 10% increase in the admission CVBG beyond 18% was associated with a 19% increase in the adjusted odds of hypoglycemia. Online supplementary tables S3 and S4 provide an explanation on how to use each model to calculate prediction for an individual patient-day using mock case examples.

Glycemic measures were the strongest predictors of hypoglycemia risk. For example, in model 1, for each 10 mg/dL increase in index day mean BG, the reductions in the adjusted odds of hypoglycemia were 27%, 14%, and 1% in the ranges of BG of ≤100 mg/dL, >100 and ≤150 mg/dL, and >150 mg/dL, respectively. Similarly, each 10% increase in the admission CVBG beyond 18% was associated with a 19% increase in the adjusted odds of hypoglycemia. Online supplementary tables S3 and S4 provide an explanation on how to use each model to calculate prediction for an individual patient-day using mock case examples. Model performance Table 3 shows the performance characteristics for the prediction models. The selected probability cut-points for prediction of biochemical and clinically significant hypoglycemia were 0.038 and 0.009, respectively. At these cut-points, model 1 achieved a sensitivity of 74.6% and a specificity of 78.5%, with corresponding c-statistic of 0.77, indicating good performance. The positive and negative predictive values were 12.4% and 98.7%, respectively. The positive and negative likelihood ratios were 3.5 and 0.3, respectively, consistent with small to moderate effect in the likelihood of the outcome with a positive or negative test result. Model 2 performed slightly better with a sensitivity of 81.9%, specificity of 78.6%, c-statistic of 0.80, positive likelihood ratio of 3.8, and negative likelihood ratio of 0.2. Table 3 Performance of prediction models

Model performance Table 3 shows the performance characteristics for the prediction models. The selected probability cut-points for prediction of biochemical and clinically significant hypoglycemia were 0.038 and 0.009, respectively. At these cut-points, model 1 achieved a sensitivity of 74.6% and a specificity of 78.5%, with corresponding c-statistic of 0.77, indicating good performance. The positive and negative predictive values were 12.4% and 98.7%, respectively. The positive and negative likelihood ratios were 3.5 and 0.3, respectively, consistent with small to moderate effect in the likelihood of the outcome with a positive or negative test result. Model 2 performed slightly better with a sensitivity of 81.9%, specificity of 78.6%, c-statistic of 0.80, positive likelihood ratio of 3.8, and negative likelihood ratio of 0.2. Table 3 Performance of prediction models Model 1: BG≤70 mg/dL Model 2: BG<54 mg/dL Probability cut-point 0.038 0.009 c-Statistic at probability cut-point 0.77 (0.75–0.78) 0.80 (0.78–0.82) Sensitivity (%) 74.6 (72.3–76.7) 81.9 (77.9–85.5) Specificity (%) 78.5 (78.1–78.9) 78.6 (78.2–79.0) Positive predictive value (%) 12.4 (11.7–13.1) 4.0 (3.6–4.4) Negative predictive value (%) 98.7 (98.6–98.8) 99.8 (99.7–99.8) Positive likelihood ratio 3.5 (3.4–3.6) 3.8 (3.7–4.0) Negative likelihood ratio 0.3 (0.3–0.4) 0.2 (0.2–0.3) BG, blood glucose.

) 81.9 (77.9–85.5) Specificity (%) 78.5 (78.1–78.9) 78.6 (78.2–79.0) Positive predictive value (%) 12.4 (11.7–13.1) 4.0 (3.6–4.4) Negative predictive value (%) 98.7 (98.6–98.8) 99.8 (99.7–99.8) Positive likelihood ratio 3.5 (3.4–3.6) 3.8 (3.7–4.0) Negative likelihood ratio 0.3 (0.3–0.4) 0.2 (0.2–0.3) BG, blood glucose. Conclusions Using EMR data from a large patient population, we developed a model to predict biochemical and clinically significant hypoglycemia in hospitalized patients treated with subcutaneous insulin. Internal validation of our models revealed good performance for detection of hypoglycemic outcomes, with slightly greater accuracy in detection of clinically significant compared with biochemical hypoglycemia. Since all of the predictor variables are readily available in the EMR, our models could be used to develop a real-time informatics alert to prevent insulin-associated hypoglycemia in hospitalized patients, a potentially serious clinical outcome.

ly associated with increased risk in the range <0.6 unit/kg/day, possibly owing to the overall low use of nutritional insulin in this cohort and lack of adequate power to detect effect sizes at larger doses. While use of high-dose SSI was associated with increased risk, administered correctional insulin doses were not. A previous study found that body weight, creatinine clearance, basal insulin dose, basal-only dosing (without mealtime insulin), use of 70/30 insulin, and use of oral antidiabetic agents were predictors of hypoglycemia in hospitalized patients.10 Many of these predictors were also significant in our models; other predictor variables that we identified were type 1 diabetes/pancreatectomy, type 2 diabetes, liver disease, digestive conditions affecting nutritional intake, nutritional status, age, and admitting service. Not surprisingly, one variable that has been shown to be a strong determinant for hypoglycemia is a prior episode of hypoglycemia.27 28 In our models, prior episodes of hypoglycemia were captured in the admission nadir BG variable, which was indeed a strong predictor of hypoglycemia. In the BG range of ≤88 mg/dL, each 10 mg/dL decrease in the admission nadir BG was associated with a 13% and a 14% increase in the odds of biochemical and clinically significant hypoglycemia, respectively (table 2, admission nadir1).

the admission nadir BG variable, which was indeed a strong predictor of hypoglycemia. In the BG range of ≤88 mg/dL, each 10 mg/dL decrease in the admission nadir BG was associated with a 13% and a 14% increase in the odds of biochemical and clinically significant hypoglycemia, respectively (table 2, admission nadir1). In general, prediction in data sets where the prevalence of the target outcome is low (like insulin-associated hypoglycemia) tends to be challenging. A previously published model achieved a sensitivity of 61%, specificity of 65%, positive predictive value of 13% and negative predictive value of 95% for detection of a BG <70 mg/dL in hospitalized patients.10 Accordingly, the positive likelihood ratio was 1.7 and the negative likelihood ratio was 0.6. While the positive predictive value of that model (13%) was comparable with ours (12.4%), our positive likelihood ratio (3.5) was double. Unlike predictive value tests, the likelihood ratio is a test characteristic that does not depend on the prevalence of the disease in the population.26 Our models achieved positive likelihood ratios of 3.5 and 3.8, which would correspond to approximately 23% and 24% increases in the probability of biochemical and clinically significant hypoglycemia, respectively.26

atio is a test characteristic that does not depend on the prevalence of the disease in the population.26 Our models achieved positive likelihood ratios of 3.5 and 3.8, which would correspond to approximately 23% and 24% increases in the probability of biochemical and clinically significant hypoglycemia, respectively.26 Predictive models using hospital information systems have been shown to improve both the quality and cost of care in situations where patient conditions can change rapidly, such as detection of sepsis or septic shock, immediate cardiac arrest, ventilator-induced lung injury, and AKI.29–31 Based on their predictive model, Kilpatrick et al developed a real-time alert process to reduce rates of inpatient hypoglycemia. Their alert, augmented by nurse–physician collaboration, reduced rates of severe hypoglycemia by 68% in high-risk patients.32 The main limitations of alert systems are false alarms and alert fatigue, where providers become desensitized or even confused by alerts. Identifying the appropriate sensitivity and specificity of informatics alerts is an area of ongoing research.33

, reduced rates of severe hypoglycemia by 68% in high-risk patients.32 The main limitations of alert systems are false alarms and alert fatigue, where providers become desensitized or even confused by alerts. Identifying the appropriate sensitivity and specificity of informatics alerts is an area of ongoing research.33 Although our models performed well, there were some limitations in our study, which if addressed in future studies could further enhance their predictive capability and clinical utility as a real-time alerting tool. Importantly, by using a time frame of 24 hours for our prediction window and aggregating information about insulin doses and glucose measures, we were unable to account for the duration of action of insulin on individual BG readings. For patients receiving only rapid-acting insulin (aspart), which has a duration of action of 4–6 hours, it is possible that spontaneous hypoglycemic episodes were misclassified as insulin-associated hypoglycemic events since they may have occurred outside the insulin’s duration of action but within the 24-hour prediction window. We were, unfortunately, unable to narrow down the prediction window to intervals less than 24 hours because the data set only included aggregate insulin information summarized by patient-day, rather than information about individual insulin doses. We are currently working on a new data set derived from our present EMR, EpicCare, to calculate the insulin dose on board relative to each BG reading, which will allow us to more accurately classify insulin-associated hypoglycemic events within more narrow prediction windows based on the pharmacologic actions of the different insulin types. Similarly, information about steroid doses would be useful, since hypoglycemia risk may be increased during steroid tapers. As this was a real-time prediction model, severity of illness and mortality indices (which are calculated after discharge) could not be used but would be expected to be important predictors. Although we did not have information about vital signs in our data set, we are considering using them as surrogate markers of illness severity in a subsequent prediction model.

ity of illness and mortality indices (which are calculated after discharge) could not be used but would be expected to be important predictors. Although we did not have information about vital signs in our data set, we are considering using them as surrogate markers of illness severity in a subsequent prediction model. Our models are dependent to a large extent on the practices at our institution and case-mix of our patient population. Since the validation was performed using the same population as the development data set, the prediction may be overestimated; nonetheless, development of a real-time alerting system from an internally validated model is still useful since the goal is to predict risk within a defined hospitalized population. Validation using an external population would be needed before applying this model for prediction in other hospital populations.

timated; nonetheless, development of a real-time alerting system from an internally validated model is still useful since the goal is to predict risk within a defined hospitalized population. Validation using an external population would be needed before applying this model for prediction in other hospital populations. In this study, it was not possible to extract information from the medical history based on limitations in our previous EMR, and we relied exclusively on the hospital problem list, which may not be a complete reflection of the patient’s clinical conditions due to undercoding. Including all laboratory results regarding renal function, rather than simply using admission-day values, would improve the classification of AKI and CKD, and could account for the impact of hospital-acquired AKI on incident risk of hypoglycemia. We did not have information about administered medications containing dextrose, a potentially important confounder. We chose to exclude patients who received concurrent intravenous and subcutaneous insulin on the index day since patients on intravenous insulin infusions are more likely to be identified as at risk of hypoglycemia according to a protocol requiring hourly BG checks; however, it is possible that such patients might not differ significantly from the population included in our models and could therefore still benefit from a real-time informatics alert. We are currently working on a prediction model using data from our present EMR, which will include information about both subcutaneous and intravenous insulin. Finally, it is possible that there are other unknown predictors or confounders of hypoglycemia that we have not accounted for in our models. Extracting a larger amount of information from the EMR, including all laboratory results and medications, could reveal unexpected associations with hypoglycemia risk. Our current EMR has a more robust data storage system that allows clinicians to aggregate information about clinical conditions from multiple sources. Given the complexity of these models with multiple variables that require mathematical processing, they could only be practically applied in clinical practice with the concurrent use of an automated electronic calculator.

m that allows clinicians to aggregate information about clinical conditions from multiple sources. Given the complexity of these models with multiple variables that require mathematical processing, they could only be practically applied in clinical practice with the concurrent use of an automated electronic calculator. There are several strengths to this study. We adhered to consensus guidelines for development and validation of our models. Given the large sample size, we had sufficiently high event per predictor ratios in both models to minimize model overfitting.34 By restricting to patients on subcutaneous insulin in the non-critical care setting, our findings are generalizable to the majority of hospitalized patients with diabetes/hyperglycemia. We had very few missing data elements, but realize that our assumptions could have resulted in misclassification of some clinical conditions. In conclusion, insulin-associated hypoglycemia in non-critically ill hospitalized adults can be predicted on the basis of EMR data. Further studies using more comprehensive information from these sources will likely improve the predictive accuracy of these models. Integration of such models into the EMR could increase the safety of hospitalized insulin-treated patients by alerting providers in real time about these high-risk patients. We would like to acknowledge partial support for the statistical analysis from the National Center for Research Resources and the National Center for Advancing Translational Sciences (NCATS) of the National Institutes of Health through Grant Number 1UL1TR001079.

In conclusion, insulin-associated hypoglycemia in non-critically ill hospitalized adults can be predicted on the basis of EMR data. Further studies using more comprehensive information from these sources will likely improve the predictive accuracy of these models. Integration of such models into the EMR could increase the safety of hospitalized insulin-treated patients by alerting providers in real time about these high-risk patients. We would like to acknowledge partial support for the statistical analysis from the National Center for Research Resources and the National Center for Advancing Translational Sciences (NCATS) of the National Institutes of Health through Grant Number 1UL1TR001079. Contributors: NNM designed the study, analyzed data, wrote the manuscript, and is the guarantor of this work. EE researched data, contributed to the discussion, and reviewed/edited the manuscript. SR researched data, contributed to the discussion, and reviewed/edited the manuscript. PJP designed the study and reviewed/edited the manuscript. H-CY designed the study and reviewed/edited the manuscript. SHG designed the study and reviewed/edited the manuscript. SS contributed to data analysis, study design, and reviewed/edited the manuscript.

uted to the discussion, and reviewed/edited the manuscript. PJP designed the study and reviewed/edited the manuscript. H-CY designed the study and reviewed/edited the manuscript. SHG designed the study and reviewed/edited the manuscript. SS contributed to data analysis, study design, and reviewed/edited the manuscript. Funding: This study was supported by several grants from the National Institutes of Health. NNM was supported by the National Institute of Diabetes and Digestive and Kidney Diseases (NIDDK) (1K23DK111986-01). EE was supported by the Clinical Research and Epidemiology in Diabetes and Endocrinology Training Grant of the NIDDK through Grant Number T32 DK062707. SR was supported by the NIDDK Summer Medical Student Research Program through the T32 Training Program in Molecular and Cellular Endocrinology (T32 DK007751). H-CY was supported by a Diabetes Research Center’s grant from the NIDDK (P30DK079637). Competing interests: None declared. Patient consent: Detail has been removed from this case description/these case descriptions to ensure anonymity. The editors and reviewers have seen the detailed information available and are satisfied that the information backs up the case the authors are making. Ethics approval: The study was approved by the Johns Hopkins Institutional Review Board. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: No additional data are available.

mpared with the control group at 12 and 36 months (LS-mean difference (95% CI), treatment vs control: 10.4 (4.5 to 16.4) for 12 months, adj-p=0.02; 23.1 (9.2 to 37.1) for 36 months, adj-p=0.04) (figure 1). Similar trends were seen in the exercise and dietary SE component scores (online supplementary figures S1 and S2). 10.1136/bmjdrc-2018-000561.supp1Supplementary data Figure 1 Total scores for psychosocial variables over 3-year follow-up by treatment group: (A) exercise self-efficacy and (B) dietary self-efficacy. Markers represent the least square mean for each group and time, adjusted for sex and baseline age, body mass index, and pre-diabetes type. Error bars represent 95% CIs. Circles: control group; squares: treatment group. Table 1 Demographic, health, and psychosocial characteristics of the analytical sample at baseline by treatment group (N=550)

Significance of this study What is already known about this subject? The nearly 3000 local health departments (LHDs) nationwide are thefrontline of public health and are positioned to implement evidence-based interventions (EBIs) for diabetes control. Little is currently known about use of diabetes-related EBIs among LHDs. What are the new findings? There is wide variation in evidence-based interventions (EBIs) offered at local health departments(LHDs): half of EBIs offered by ≥80% of the sample, and a quarteroffered by fewer than 60%. There are several respondent and LHD characteristics associated with offeringeach of the four diabetes-related EBIs. How might these results change the focus of research or clinical practice? Supporting evidence-based decision making, and increasing the implementation of EBIs by more LHDs can help control diabetes nationwide. Introduction Diabetes causes significant morbidity and mortality.1 Evidence-based interventions (EBIs) are available to help modify lifestyle behaviors related to diabetes (eg, nutrition and physical activity) and enhance its treatment and management.2–4 Tools such as the Community Guide (https://www.thecommunityguide.org/topic/diabetes), What Works for Health, and Cochrane reviews are available to support the use of EBIs to prevent and control diabetes.5–11 There is a strong case for the engagement of organizations such as local health departments (LHDs) in diabetes prevention and control.3 4 12

Guide (https://www.thecommunityguide.org/topic/diabetes), What Works for Health, and Cochrane reviews are available to support the use of EBIs to prevent and control diabetes.5–11 There is a strong case for the engagement of organizations such as local health departments (LHDs) in diabetes prevention and control.3 4 12 The nearly 3000 US LHDs are the ‘frontline’ of public health, and are therefore well positioned to implement EBIs for diabetes control directly and/or in collaboration with other organizations serving the same community.12 13 These departments typically receive funding from state and local governments, and engage in surveillance and prevention activities (eg, tuberculosis screening, child and adult immunization provision), as well as activities related to environmental health (eg, inspecting food service establishments and day care centers).14 As the threats to public health have changed over time, so have the roles of LHDs.12 15 16 Public health departments can and should play an important role in diabetes prevention and management.3 15 17 One study found that for each 10% increase in public health spending, diabetes mortality fell by 1.4%.17 These gains appear to be due, in part, to collaborations and partnerships within communities to provide needed services and achieve common population health goals.18 19 Health departments can extend the reach of healthcare providers and the traditional healthcare system, and are able to provide services to community members who may not otherwise have access to preventive care and health screening due to lack of medical insurance or a feeling of alienation from the medical system.17

19 Health departments can extend the reach of healthcare providers and the traditional healthcare system, and are able to provide services to community members who may not otherwise have access to preventive care and health screening due to lack of medical insurance or a feeling of alienation from the medical system.17 The National Association of County and City Health Officials (NACCHO) conducts an ongoing survey of LHDs, the National Profile of Local Health Departments, to identify the population-based primary prevention activities available in the communities served by LHDs. In 2016, 34% of LHDs reported screening for diabetes, and 74%, 60%, and 57% indicated they offer population-based primary prevention related to nutrition, physical activity, and chronic disease, respectively.14 However, these activities were defined broadly and did not ask about specific EBIs. Despite the critical role LHDs play18 19 and the widespread initiatives LHDs provide, limited information is available about the programs offered and whether these are EBIs. Detail is also lacking with regard to how LHDs are delivering these EBIs (ie, directly by the department and/or in collaboration) at the local level. Further, given the documented gap in translation of research to public health practice,20 a fuller understanding of factors that facilitate and/or hinder EBI implementation is needed.

o lacking with regard to how LHDs are delivering these EBIs (ie, directly by the department and/or in collaboration) at the local level. Further, given the documented gap in translation of research to public health practice,20 a fuller understanding of factors that facilitate and/or hinder EBI implementation is needed. Previous research has suggested that organizational processes can impact uptake of EBIs and that the components of evidence-based decision making (EBDM) can support implementation of EBIs.21–23 EBDM operates at multiple levels within an LHD and includes summarizing the findings from the best available peer-reviewed evidence, using data and information systems, applying program planning frameworks, engaging the community in assessment and decision making, conducting sound evaluation, and synthesizing science and communication skills with common sense and political acumen for dissemination to other stakeholders and decision makers.24 In public health agency settings, management support for EBDM is associated with improved performance.25 This study seeks to assess LHDs’ delivery of EBIs related to diabetes prevention and control in several categories (diabetes-related such as the Diabetes Prevention Program (DPP) or self-management education, obesity, physical activity, nutrition and tobacco), and whether these are delivered directly, in collaboration, and/or both. Further, for diabetes-related EBIs, factors at the level of the LHD, including EBDM, associated with delivering each EBI were explored.

etes Prevention Program (DPP) or self-management education, obesity, physical activity, nutrition and tobacco), and whether these are delivered directly, in collaboration, and/or both. Further, for diabetes-related EBIs, factors at the level of the LHD, including EBDM, associated with delivering each EBI were explored. Research design and methods This cross-sectional survey was part of a larger dissemination study focusing on efforts to improve evidence-based diabetes management and chronic disease control among LHDs.26 Participants and recruitment LHDs were drawn from the 1677 LHDs across the USA which reported in the 2016 NACCHO National Profile that their agency screens for diabetes or body mass index (BMI), or conducts population-based primary prevention activities for nutrition or physical activity (ie, the National Profile survey asks whether the LHDs ‘screen for diabetes or BMI’ and ‘conduct population-based primary prevention activities for nutrition or physical activity’ directly or via contract). A stratified random sample of 600 LHDs were selected according to three jurisdiction population size categories (small <50 000, medium 50 000–199 999, and large ≥200 000). Efforts were made to distribute the sample across LHD jurisdiction population sizes. Within each selected LHD, the lead practitioner working in chronic disease control was invited to participate in the current study (eg, one participant per health department). After excluding non-valid email addresses, the final recruitment sample was 579.

distribute the sample across LHD jurisdiction population sizes. Within each selected LHD, the lead practitioner working in chronic disease control was invited to participate in the current study (eg, one participant per health department). After excluding non-valid email addresses, the final recruitment sample was 579. Data collection Survey invitation emails included study information and a link to complete the survey online via the Qualtrics software. To enhance participation, 1 week prior to the survey invitation, a preinvitation email informing survey contacts about the purpose of the study was sent. If a potential participant did not respond to the invitation, follow-up included three reminder emails and two follow-up calls. As compensation for their time completing the survey, respondents were offered a $20 Amazon.com gift card. Survey development Details of the survey development process have been described elsewhere.26 Briefly, the survey drew on previous research conducted by the project team26 and existing instruments identified through snowball sampling of other researchers’ measures identified by the study team. In addition to three rounds of input, cognitive response testing interviews with 10 practitioners like those in the target audience and an assessment of test–retest reliability were conducted.

am26 and existing instruments identified through snowball sampling of other researchers’ measures identified by the study team. In addition to three rounds of input, cognitive response testing interviews with 10 practitioners like those in the target audience and an assessment of test–retest reliability were conducted. Respondent and LHD characteristics Respondents reported the characteristics of their LHD (eg, current status in Public Health Accreditation Board accreditation efforts) and themselves (eg, age group, years in current position); these characteristics are listed in table 1. The survey also included the Short Grit Scale, which measures respondent characteristics: passion and perseverance for long-term goals.27 Perceived organizational support for EBDM was assessed using six factors derived from the survey using confirmatory factor analysis (full item wording is available in online supplementary table 1; factor development and validation are described elsewhere28). The organizational support for EBDM factors, as shown in Parks et al 26 (figure 1), includes awareness of EBDM (three items), capacity for EBDM (seven items), resource availability (three items), evaluation capacity (three items), EBDM climate cultivation (three items), and partnerships to support EBDM (three items). 10.1136/bmjdrc-2018-000558.supp1Supplementary data Figure 1 Percentage of LHDs that reported delivering EBIs directly and/or collaboratively with a partnering organization. SNAP, Supplemental Nutrition Assistance Program; WIC program, Women, Infants, and Children program.

Respondent and LHD characteristics Respondents reported the characteristics of their LHD (eg, current status in Public Health Accreditation Board accreditation efforts) and themselves (eg, age group, years in current position); these characteristics are listed in table 1. The survey also included the Short Grit Scale, which measures respondent characteristics: passion and perseverance for long-term goals.27 Perceived organizational support for EBDM was assessed using six factors derived from the survey using confirmatory factor analysis (full item wording is available in online supplementary table 1; factor development and validation are described elsewhere28). The organizational support for EBDM factors, as shown in Parks et al 26 (figure 1), includes awareness of EBDM (three items), capacity for EBDM (seven items), resource availability (three items), evaluation capacity (three items), EBDM climate cultivation (three items), and partnerships to support EBDM (three items). 10.1136/bmjdrc-2018-000558.supp1Supplementary data Figure 1 Percentage of LHDs that reported delivering EBIs directly and/or collaboratively with a partnering organization. SNAP, Supplemental Nutrition Assistance Program; WIC program, Women, Infants, and Children program. Table 1 LHD and respondent characteristics of LHDs in the total sample (n=376)

10.1136/bmjdrc-2018-000558.supp1Supplementary data Figure 1 Percentage of LHDs that reported delivering EBIs directly and/or collaboratively with a partnering organization. SNAP, Supplemental Nutrition Assistance Program; WIC program, Women, Infants, and Children program. Table 1 LHD and respondent characteristics of LHDs in the total sample (n=376) n (%*) or mean (SD) Respondent characteristics Age group (years), n (%) 20–29 14 (3.7) 30–39 86 (23) 40–49 111 (30) 50–59 107 (28) 60+ 57 (15) Race/Ethnicity, n (%) White 315 (84.0) Black/African–American 26 (6.9) Other race 27 (7.2) Hispanic or Latino 7 (1.9) Sex, n (%) Male 60 (16) Female 312 (83) Master’s degree or higher in any field, n (%) No 155 (42) Yes 216 (58) Public health master’s or doctorate, n (%) No 253 (68) Yes 118 (32) Position, n (%) Top executive, health director/officer/commissioner 97 (26) Administrator, deputy or assistant director 77 (20) Manager of a division or program 138 (37) Program coordinator 33 (8.8) Technical expert position (evaluator, epidemiologist, health educator)/other 30 (8.0) Years in current position (years), n (%) <5 202 (54) 5–9 87 (23) 10–19 60 (16) 20+ 25 (6.7) Years in public health (years), n (%) <5 41 (11) 5–9 66 (18) 10–19 118 (32) 20+ 149 (40) Short Grit Scale, mean (SD) 4.0 (0.48) LHD characteristics LHD jurisdiction population category, n (%) Small (<50 000) 118 (32) Medium (50 000–199 999) 124 (33) Large (200 000+) 128 (35) PHAB-accredited or preparing to apply, n (%) Currently accredited 113 (30) Recently applied but not yet accredited 42 (11) Yes, but have not yet applied 84 (22) No 107 (28) Unsure 29 (7.7) Currently participate in academic partnerships, n (%) Yes 272 (73) No/Unsure 99 (27) Diabetes prevalence in the state, mean (SD) 9.2 (1.5) *% within respondent and LHD characteristic categories.

ed 113 (30) Recently applied but not yet accredited 42 (11) Yes, but have not yet applied 84 (22) No 107 (28) Unsure 29 (7.7) Currently participate in academic partnerships, n (%) Yes 272 (73) No/Unsure 99 (27) Diabetes prevalence in the state, mean (SD) 9.2 (1.5) *% within respondent and LHD characteristic categories. LHD, local health department; PHAB, Public Health Accreditation Board.

ed 113 (30) Recently applied but not yet accredited 42 (11) Yes, but have not yet applied 84 (22) No 107 (28) Unsure 29 (7.7) Currently participate in academic partnerships, n (%) Yes 272 (73) No/Unsure 99 (27) Diabetes prevalence in the state, mean (SD) 9.2 (1.5) *% within respondent and LHD characteristic categories. LHD, local health department; PHAB, Public Health Accreditation Board. Assessment of EBIs offered For the items assessing EBI delivery, sources such as the Community Guide7 and What Works for Health8 9 were used to identify EBIs, which LHDs might offer either directly or in collaboration. EBIs fell in one of the five categories of diabetes prevention and control activities addressed in the public and community health setting (ie, diabetes-related, obesity, physical activity, nutrition, and tobacco), and were reviewed by the study team to select those with the strongest evidence base. To minimize respondent burden and increase accuracy in reporting, participants were only asked to report on EBIs within a given category (ie, diabetes-related, obesity, physical activity, nutrition, and tobacco), which was determined by the program area in which they reported working (ie, diabetes-related, obesity, physical activity, nutrition, and tobacco). The decision logic was set to increase the sample of participants asked to report on the four diabetes-related EBIs; those who reported diabetes as a program area—whether alone (diabetes only) or in combination with other program areas—were asked to respond to the four diabetes-related EBIs. Thus 240 participants were asked to report on the diabetes-related EBIs, and 24, 31, 38, and 42 participants were asked to report on obesity, physical activity, nutrition, and tobacco EBIs, respectively. Each category included four EBIs and asked participants to report whether their LHD offered the EBI directly, in collaboration with a partner, both (directly/in collaboration), or neither (figure 1 lists the EBIs). The survey defined ‘delivered’ as ‘In the past year, has your agency directly delivered, and has your agency collaborated with organizations to support delivery of the following diabetes interventions’. Collaborated with was defined as ‘served as a community/clinical referral source, or a convener that facilitates the program or referral system’.

d’ as ‘In the past year, has your agency directly delivered, and has your agency collaborated with organizations to support delivery of the following diabetes interventions’. Collaborated with was defined as ‘served as a community/clinical referral source, or a convener that facilitates the program or referral system’. Analysis Participant and LHD characteristics were summarized using descriptive statistics. Descriptive analyses were also used to describe direct EBI implementation and collaborative implementation (ie, if the EBIs were offered, were they delivered directly by the LHD and/or in collaboration). Given the focus of the study, only the diabetes-related EBIs had a large enough sample size to explore in more depth. Bivariate logistic regression models were used to explore whether LHD characteristics and EBDM scores were associated with whether the LHD offered each diabetes-related EBI and whether the LHD offered all four diabetes-related EBIs. Analyses were performed in SPSS V.24; significance levels for the models were set at p<0.05.

iate logistic regression models were used to explore whether LHD characteristics and EBDM scores were associated with whether the LHD offered each diabetes-related EBI and whether the LHD offered all four diabetes-related EBIs. Analyses were performed in SPSS V.24; significance levels for the models were set at p<0.05. Results The 376 responding LHD practitioners (one survey participant per LHD) (65% response rate) were evenly distributed across jurisdiction population size categories; 30% worked for an accredited LHD (table 1). Respondents were primarily female (83%), older than 40 years (73%), and had worked in public health for at least 10 years (72%). In terms of training, 58% of the respondents reported a master’s degree or higher in any field, while 32% reported a master’s or doctorate in public health. Most respondents were a manager of a division or program (37%), the top executive, health director/officer/commissioner (26%), or an administrator, deputy or assistant director (20%) at the LHD. Additional respondent and LHD characteristics are shown in table 1.

le 32% reported a master’s or doctorate in public health. Most respondents were a manager of a division or program (37%), the top executive, health director/officer/commissioner (26%), or an administrator, deputy or assistant director (20%) at the LHD. Additional respondent and LHD characteristics are shown in table 1. There was considerable variation among the diabetes-related EBIs delivered directly and/or in collaboration, with greater than 80% of the respondents reporting they offered the DPP (82%) and diabetes self-management education (81%), compared with 61% offering community health worker programming and 67% offering diabetes screening and treatment referrals (figure 1). Of the 24 LHDs that were asked about obesity EBIs, the only commonly reported EBI was worksite programs, policies or environmental changes to promote nutrition/healthy food and physical activity (83%). Greater than 80% of the 38 LHDs that were asked about the nutrition EBIs reported three of these EBIs were delivered (ie, policies or environmental changes to improve access to healthy foods in worksites, schools, or other local facilities; policies or changes that improve healthier food choices through nutrition assistance programs; and policies, environmental changes or programs promoting breast feeding); school gardens were reported by only 57% of the 38 LHDs. Thirty-one respondents were asked about physical activity promotion EBIs (ie, programs that set up social support for physical activity; programs, policies, or environmental changes that make streets safer for pedestrians and cyclists; programs or policies that create or improve access to places for physical activity; and programs or policies that increase physical activity in schools), and these EBIs were commonly delivered (all ≥70%). Tobacco EBIs were also commonly delivered directly and/or in collaboration, with ≥80% of LHDs delivering each of the three tobacco EBIs and 67% delivering the fourth EBI.

laces for physical activity; and programs or policies that increase physical activity in schools), and these EBIs were commonly delivered (all ≥70%). Tobacco EBIs were also commonly delivered directly and/or in collaboration, with ≥80% of LHDs delivering each of the three tobacco EBIs and 67% delivering the fourth EBI. Five EBIs (including all four physical activity EBIs) were only offered in collaboration or both directly/in collaboration with partners, but were not reported to be offered only directly. Most of the remaining EBIs (n=13) were offered only directly by 3%–11% of LHDs asked. Only two EBIs, both nutrition EBIs (improving food choices in assistance programs and promoting breast feeding) were offered only directly at more than 11% of LHDs.

ion with partners, but were not reported to be offered only directly. Most of the remaining EBIs (n=13) were offered only directly by 3%–11% of LHDs asked. Only two EBIs, both nutrition EBIs (improving food choices in assistance programs and promoting breast feeding) were offered only directly at more than 11% of LHDs. Among the 240 LHDs asked to report on the diabetes-related EBIs, there were several associations between respondent and LHD characteristics, as well as the organizational support for EBDM factors and the EBIs (tables 2 and 3). Most consistently at the respondent level, how long the respondent had been in their current position and their age were both negatively associated with using community health workers to deliver diet and physical activity promotion and/or weight management to groups or individuals with increased risk for type 2 diabetes and with delivering all four diabetes-related EBIs. At the LHD level, diabetes prevalence in the state was associated with offering three of the EBIs: the DPP (OR=1.28 (95% CI 1.02 to 1.62)), diabetes self-management education (OR=1.32 (95% CI 1.04 to 1.67)), and identifying patients and determining treatment (OR=1.27 (95% CI 1.05 to 1.54)). Finally, although all organizational supports for EBDM factors were related in a positive direction with offering the EBIs, the only significant association was between evaluation capacity and identifying patients with diabetes and determining effective treatment (OR=1.54 (95% CI 1.08 to 2.19)).

(95% CI 1.05 to 1.54)). Finally, although all organizational supports for EBDM factors were related in a positive direction with offering the EBIs, the only significant association was between evaluation capacity and identifying patients with diabetes and determining effective treatment (OR=1.54 (95% CI 1.08 to 2.19)). Table 2 LHD and respondent characteristics of LHDs in the sample reporting on diabetes-related EBIs (n=240) LHDs offering* diabetes-related EBIs, n (%†) or mean (SE) Total DPP‡ CHWs§ DSME¶ Identify** All four Respondent characteristics Age group (years), n (%) 20–29 9 7 (78) 7 (78) 8 (89) 7 (78) 6 (67) 30–39 57 47 (82) 35 (61) 42 (78) 37 (69) 22 (39) 40–49 73 62 (85) 52 (71) 64 (89) 51 (72) 38 (52) 50–59 67 54 (81) 37 (55) 51 (80) 39 (63) 22 (33) 60+ 33 26 (79) 12 (36) 25 (76) 17 (52) 7 (21) Pearson’s χ2 p 0.93 0.01 0.36 0.27 0.01

ent characteristics Age group (years), n (%) 20–29 9 7 (78) 7 (78) 8 (89) 7 (78) 6 (67) 30–39 57 47 (82) 35 (61) 42 (78) 37 (69) 22 (39) 40–49 73 62 (85) 52 (71) 64 (89) 51 (72) 38 (52) 50–59 67 54 (81) 37 (55) 51 (80) 39 (63) 22 (33) 60+ 33 26 (79) 12 (36) 25 (76) 17 (52) 7 (21) Pearson’s χ2 p 0.93 0.01 0.36 0.27 0.01 Race/Ethnicity, n (%) White 203 167 (82) 118 (58) 162 (83) 125 (64) 76 (37) Black/African–American 17 14 (82) 11 (65) 14 (82) 13 (76) 9 (53) Other race 14 10 (71) 9 (64) 9 (64) 9 (69) 6 (43) Hispanic or Latino 5 5 (100) 5 (100) 5 (100) 4 (80) 4 (80) Pearson’s χ2 p 0.54 0.27 0.25 0.67 0.16 Sex, n (%) Male 36 31 (86) 22 (61) 31 (91) 25 (78) 16 (44) Female 202 165 (82) 120 (59) 158 (80) 126 (64) 79 (39) Pearson’s χ2 p 0.52 0.85 0.13 0.12 0.55 Master’s degree or higher in any field (n%) No 110 85 (77) 70 (64) 87 (83) 68 (66) 46 (42) Yes 126 108 (86) 71 (56) 100 (81) 80 (65) 47 (37) Pearson’s χ2 p 0.09 0.25 0.67 0.88 0.48 Public health master’s or doctorate, n (%) No 170 139 (82) 105 (62) 138 (84) 109 (67) 70 (41) Yes 66 54 (82) 36 (55) 49 (77) 39 (61) 23 (35) Pearson’s χ2 p 0.99 0.31 0.21 0.37 0.37 Position, n (%) Top executive, health director/officer/commissioner 60 53 (88) 30 (50) 50 (83) 44 (73) 25 (42) Administrator, deputy or assistant director  53 41 (77)  37 (70) 44 (83) 37 (70) 25 (47) Manager of a division or program  83 72 (87)  51 (61) 67 (84) 46 (58) 30 (36) Program coordinator  27 21 (78)  16 (59) 19 (76) 18 (75) 12 (44) Technical expert position (evaluator, epidemiologist, health educator)/other  16 9 (56)  9 (56) 10 (71) 6 (46) 3 (19) Pearson’s χ2 p 0.02 0.31 0.75 0.13 0.29 Years in current position, n (%) <5  135 112 (83)  90 (67) 108 (82) 90 (70) 63 (47) 5–9  57 48 (84)  34 (60) 47 (84) 34 (61) 21 (37) 10–19  31 23 (74)  16 (52) 22 (73) 16 (55) 8 (26) 20+  15 13 (87)  3 (20) 13 (87) 11 (73) 3 (20) Pearson’s χ2 p   0.62  0.00 0.59 0.34 0.05 Years in public health, n (%) <5  28 22 (79)  23 (82) 23 (85) 21 (78) 17 (61) 5–9  38 33 (87)  23 (61) 30 (81) 25 (68) 15 (39) 10–19  81 65 (80)  44 (54) 67 (85) 53 (68) 32 (40) 20+  91 76 (84)  53 (58) 70 (79) 52 (60) 31 (34) Pearson’s χ2 p   0.77  0.07 0.73 0.34 0.10 Short Grit Scale, mean (SE) Not offered   3.96 (0.48)  4.03 (0.49) 4.01 (0.52) 3.95 (0.50) 4.00 (0.48) Offered   4.00 (0.48)  3.97 (0.46) 3.99 (0.47) 4.01 (0.47) 3.98 (0.48) Mean difference   −0.04 (0.08)  0.06 (0.06) 0.02 (0.08) −0.06 (0.07) 0.03 (0.06) t (p)   −0.48 (0.63)  0.99 (0.32) 0.29 (0.77) −0.95 (0.34) 0.41 (0.68) LHD characteristics LHD

t offered   3.96 (0.48)  4.03 (0.49) 4.01 (0.52) 3.95 (0.50) 4.00 (0.48) Offered   4.00 (0.48)  3.97 (0.46) 3.99 (0.47) 4.01 (0.47) 3.98 (0.48) Mean difference   −0.04 (0.08)  0.06 (0.06) 0.02 (0.08) −0.06 (0.07) 0.03 (0.06) t (p)   −0.48 (0.63)  0.99 (0.32) 0.29 (0.77) −0.95 (0.34) 0.41 (0.68) LHD characteristics LHD jurisdiction population category, n (%) Small (<50 000)  79 58 (73)  45 (57) 62 (81) 48 (63) 30 (38) Medium (50 000–199 999)  75 64 (85)  42 (56) 60 (82) 51 (71) 32 (43) Large (200 000+)  84 73 (87)  56 (67) 67 (83) 51 (64) 33 (39) Pearson’s χ2 p   0.05  0.31 0.93 0.55 0.83 PHAB-accredited or preparing to apply, n (%) Currently accredited  69 59 (86)  46 (67) 55 (81) 45 (67) 26 (38) Recently applied but not yet accredited  28 24 (86)  18 (64) 23 (85) 18 (67) 13 (46) Yes, but have not yet applied  43 36 (84)  26 (60) 34 (85) 21 (55) 16 (37) No  78 61 (78)  39 (50) 62 (79) 52 (67) 29 (37) Unsure  21 16 (76)  14 (67) 16 (84) 15 (79) 11 (52) Pearson’s χ2 p   0.71  0.27 0.93 0.49 0.67 Currently participate in academic partnerships, n (%) Yes  173 146 (84)  108 (62) 143 (85) 115 (69) 75 (43) No/Unsure  65 50 (77)  35 (54) 47 (76) 36 (59) 20 (31) Pearson’s χ2 p   0.18  0.23 0.12 0.16 0.08 Diabetes prevalence in the state, mean (SE) Not offered   8.91 (1.47)  9.27 (1.52) 8.85 (1.43) 8.98 (1.47) 9.22 (1.43) Offered*   9.45 (1.51)  9.40 (1.51) 9.45 (1.53) 9.51 (1.51) 9.54 (1.62) Mean difference   −0.54 (0.25)  −0.13 (0.20) −0.60 (0.26) −0.53 (0.21) −0.31 (0.20) t (p)   −2.13 (0.03)  −0.65 (0.52) −2.33 (0.02) −2.55 (0.01) −1.58 (0.12) Organizational support for EBDM (standardized) Factor 1: awareness of EBDM, mean (SE) Not offered   0.01 (0.13)  0.05 (0.08) −0.04 (0.12) −0.06 (0.08) −0.16 (0.28) Offered*   0.05 (0.05)  0.05 (0.05) 0.08 (0.05) 0.11 (0.05) 0.06 (0.04) Mean difference   −0.04 (0.12)  0.00 (0.09) −0.12 (0.11) −0.16 (0.09) −0.22 (0.18) t (p)   −0.34 (0.73)  0.00 (1.00) −1.08 (0.28) −1.77 (0.08) −1.21 (0.23) Factor 2: capacity for EBDM, mean (SE) Not offered   −0.01 (0.14)  0.03 (0.08) −0.03 (0.13) −0.07 (0.09) −0.21 (0.31) Offered*   0.06 (0.05)  0.05 (0.06) 0.08 (0.05) 0.11 (0.06) 0.06 (0.05) Mean difference   −0.06 (0.13)  −0.02 (0.10) −0.11 (0.12) −0.18 (0.10) −0.28 (0.20) t (p)   −0.51 (0.61)  −0.21 (0.83) −0.88 (0.38) −1.80 (0.07) −1.38 (0.17) Factor 3: resource availability, mean (SE) Not offered   −0.07 (0.11)  0.04 (0.07) 0.00 (0.10) −0.04 (0.08) −0.19 (0.24) Offered*   0.06 (0.05)  0.03 (0.05) 0.06 (0.05) 0.09 (0.05) 0.05 (0.04) Mean difference   −0.13

12) −0.18 (0.10) −0.28 (0.20) t (p)   −0.51 (0.61)  −0.21 (0.83) −0.88 (0.38) −1.80 (0.07) −1.38 (0.17) Factor 3: resource availability, mean (SE) Not offered   −0.07 (0.11)  0.04 (0.07) 0.00 (0.10) −0.04 (0.08) −0.19 (0.24) Offered*   0.06 (0.05)  0.03 (0.05) 0.06 (0.05) 0.09 (0.05) 0.05 (0.04) Mean difference   −0.13 (0.11)  0.00 (0.09) −0.07 (0.11) −0.14 (0.09) −0.24 (0.18) t (p)   −1.13 (0.26)  0.03 (0.98) −0.59 (0.56) −1.55 (0.12) −1.37 (0.17) Factor 4: evaluation capacity, mean (SE) Not offered   −0.08 (0.15)  0.09 (0.09) −0.12 (0.13) −0.14 (0.09) −0.27 (0.33) Offered*   0.06 (0.06)  0.00 (0.06) 0.08 (0.06) 0.13 (0.06) 0.06 (0.05) Mean difference   −0.14 (0.14)  0.09 (0.11) −0.20 (0.14) −0.27 (0.11) −0.33 (0.22) t (p)   −1.02 (0.31)  0.82 (0.42) −1.47 (0.14) −2.47 (0.01) −1.51 (0.13) Factor 5: EBDM climate cultivation, mean (SE) Not offered   0.08 (0.10)  0.08 (0.06) 0.02 (0.09) −0.04 (0.07) −0.06 (0.22) Offered*   0.03 (0.04)  0.01 (0.04) 0.05 (0.04) 0.08 (0.04) 0.04 (0.04) Mean difference   0.05 (0.09)  0.07 (0.07) −0.03 (0.09) −0.13 (0.08) −0.10 (0.15) t (p)   0.57 (0.57)  1.01 (0.31) −0.32 (0.75) −1.64 (0.10) −0.69 (0.49) Factor 6: partnerships to support EBDM, mean (SE) Not offered   −0.03 (0.11)  −0.01 (0.07) −0.01 (0.11) −0.10 (0.08) −0.23 (0.25) Offered*   −0.02 (0.04)  −0.03 (0.05) −0.01 (0.04) 0.02 (0.05) −0.01 (0.04) Mean difference   −0.01 (0.11)  0.01 (0.08) 0.00 (0.11) −0.12 (0.09) −0.22 (0.17) t (p)   −0.13 (0.90)  0.15 (0.88) 0.03 (0.98) −1.36 (0.17) −1.31 (0.19) Bold values indicate statistically significant relationships according to a n alpha=0.05 threshold.

−0.02 (0.04)  −0.03 (0.05) −0.01 (0.04) 0.02 (0.05) −0.01 (0.04) Mean difference   −0.01 (0.11)  0.01 (0.08) 0.00 (0.11) −0.12 (0.09) −0.22 (0.17) t (p)   −0.13 (0.90)  0.15 (0.88) 0.03 (0.98) −1.36 (0.17) −1.31 (0.19) Bold values indicate statistically significant relationships according to a n alpha=0.05 threshold. *Each category included four EBIs and asked participants to report whether their LHD offered the EBI directly, in collaboration with a partner, both (directly/in collaboration), or neither. †% within respondent and LHD characteristic categories. ‡Diet and physical activity promotion programs with people at increased risk for type 2 diabetes, such as the Diabetes Prevention Program (DPP). §Community health workers (CHWs) to deliver diet and physical activity promotion and weight management to groups or individuals with increased risk for type 2 diabetes. ¶Diabetes self-management education (DSME) with persons with diabetes delivered in community gathering places. **Diabetes management interventions identifying patients with diabetes and determining effective treatment (identify). EBDM, evidence-based decision making; EBIs, evidence-based interventions; LHDs, local health departments; PHAB, Public Health Accreditation Board. Table 3 Associations between respondent and LHD characteristics and delivering diabetes-related EBIs directly or in collaboration

**Diabetes management interventions identifying patients with diabetes and determining effective treatment (identify). EBDM, evidence-based decision making; EBIs, evidence-based interventions; LHDs, local health departments; PHAB, Public Health Accreditation Board. Table 3 Associations between respondent and LHD characteristics and delivering diabetes-related EBIs directly or in collaboration DPP* CHWs† DSME‡ Identify§ Offering all 4 diabetes EBIs OR (95% CI) OR (95% CI) OR (95% CI) OR (95% CI) OR (95% CI) Respondent characteristics Master’s degree or higher in any field 1.76 (0.90 to 3.45) 0.74 (0.44 to 1.25) 0.86 (0.44 to 1.69) 0.96 (0.55 to 1.66) 0.83 (0.49 to 1.40) Public health master’s or doctorate 1.00 (0.48 to 2.10) 0.74 (0.42 to 1.32) 0.64 (0.31 to 1.30) 0.76 (0.42 to 1.38) 0.76 (0.42 to 1.38) Position (top executive, health director, health officer, commissioner=referent) 0.74 (0.56 to 0.99) 1.07 (0.86 to 1.33) 0.87 (0.65 to 1.15) 0.82 (0.64 to 1.04) 0.86 (0.69 to 1.08) Years in current position 0.93 (0.66 to 1.33) 0.62 (0.46 to 0.82) 0.94 (0.66 to 1.34) 0.88 (0.66 to 1.17) 0.65 (0.47 to 0.88) Years in public health 1.04 (0.75 to 1.45) 0.78 (0.59 to 1.01) 0.88 (0.62 to 1.24) 0.78 (0.59 to 1.04) 0.75 (0.58 to 0.97) Age 0.94 (0.69 to 1.28) 0.71 (0.55 to 0.91) 0.89 (0.66 to 1.22) 0.78 (0.60 to 1.01) 0.72 (0.56 to 0.93)

DPP* CHWs† DSME‡ Identify§ Offering all 4 diabetes EBIs OR (95% CI) OR (95% CI) OR (95% CI) OR (95% CI) OR (95% CI) Respondent characteristics Master’s degree or higher in any field 1.76 (0.90 to 3.45) 0.74 (0.44 to 1.25) 0.86 (0.44 to 1.69) 0.96 (0.55 to 1.66) 0.83 (0.49 to 1.40) Public health master’s or doctorate 1.00 (0.48 to 2.10) 0.74 (0.42 to 1.32) 0.64 (0.31 to 1.30) 0.76 (0.42 to 1.38) 0.76 (0.42 to 1.38) Position (top executive, health director, health officer, commissioner=referent) 0.74 (0.56 to 0.99) 1.07 (0.86 to 1.33) 0.87 (0.65 to 1.15) 0.82 (0.64 to 1.04) 0.86 (0.69 to 1.08) Years in current position 0.93 (0.66 to 1.33) 0.62 (0.46 to 0.82) 0.94 (0.66 to 1.34) 0.88 (0.66 to 1.17) 0.65 (0.47 to 0.88) Years in public health 1.04 (0.75 to 1.45) 0.78 (0.59 to 1.01) 0.88 (0.62 to 1.24) 0.78 (0.59 to 1.04) 0.75 (0.58 to 0.97) Age 0.94 (0.69 to 1.28) 0.71 (0.55 to 0.91) 0.89 (0.66 to 1.22) 0.78 (0.60 to 1.01) 0.72 (0.56 to 0.93) Race/Ethnicity 0.99 (0.60 to 1.63) 1.44 (0.92 to 2.24) 0.88 (0.55 to 1.40) 1.26 (0.80 to 1.98) 1.44 (0.97 to 2.13) Sex 0.72 (0.26 to 1.97) 0.93 (0.45 to 1.93) 0.39 (0.11 to 1.35) 0.50 (0.20 to 1.21) 0.80 (0.39 to 1.64) Short Grit Scale 1.19 (0.59 to 2.38) 0.76 (0.44 to 1.31) 0.90 (0.44 to 1.83) 1.32 (0.75 to 2.33) 0.89 (0.52 to 1.54) LHD characteristics Jurisdiction population categories (<50 000=referent) 1.59 (1.05 to 2.40) 1.23 (0.90 to 1.68) 1.08 (0.72 to 1.61) 1.01 (0.73 to 1.41) 1.03 (0.75 to 1.40) PHAB accreditation status 0.84 (0.66 to 1.08) 0.87 (0.72 to 1.05) 0.99 (0.78 to 1.26) 1.04 (0.85 to 1.26) 1.04 (0.86 to 1.25) Academic partnership 0.62 (0.30 to 1.25) 0.70 (0.39 to 1.25) 0.57 (0.28 to 1.17) 0.65 (0.36 to 1.19) 0.58 (0.32 to 1.07) Diabetes prevalence in the state 1.28 (1.02 to 1.62) 1.06 (0.89 to 1.26) 1.32 (1.04 to 1.67)

status 0.84 (0.66 to 1.08) 0.87 (0.72 to 1.05) 0.99 (0.78 to 1.26) 1.04 (0.85 to 1.26) 1.04 (0.86 to 1.25) Academic partnership 0.62 (0.30 to 1.25) 0.70 (0.39 to 1.25) 0.57 (0.28 to 1.17) 0.65 (0.36 to 1.19) 0.58 (0.32 to 1.07) Diabetes prevalence in the state 1.28 (1.02 to 1.62) 1.06 (0.89 to 1.26) 1.32 (1.04 to 1.67) 1.27 (1.05 to 1.54) 1.15 (0.97 to 1.36) Organizational support for EBDM Factor 1: awareness of EBDM 1.09 (0.67 to 1.75) 1.00 (0.69 to 1.45) 1.31 (0.80 to 2.16) 1.45 (0.96 to 2.20) 1.59 (0.75 to 3.37) Factor 2: capacity for EBDM 1.12 (0.72 to 1.74) 1.04 (0.74 to 1.46) 1.23 (0.78 to 1.95) 1.42 (0.97 to 2.09) 1.62 (0.81 to 3.22) Factor 3: resource availability 1.33 (0.81 to 2.19) 0.99 (0.67 to 1.47) 1.17 (0.70 to 1.96) 1.40 (0.91 to 2.15) 1.72 (0.79 to 3.76) Factor 4: evaluation capacity 1.23 (0.83 to 1.83) 0.88 (0.64 to 1.20) 1.36 (0.90 to 2.05) 1.54 (1.08 to 2.19) 1.60 (0.87 to 2.96) Factor 5: EBDM climate cultivation 0.84 (0.46 to 1.53) 0.79 (0.49 to 1.26) 1.10 (0.60 to 2.02) 1.52 (0.92 to 2.51) 1.38 (0.55 to 3.47) Factor 6: partnerships to support EBDM 1.03 (0.62 to 1.73) 0.97 (0.65 to 1.45) 0.99 (0.59 to 1.68) 1.34 (0.88 to 2.06) 1.65 (0.78 to 3.48) OR from unadjusted bivariate model. Bold values indicate statistically significant relationships according to a n alpha=0.05 threshold. *Diet and physical activity promotion programs with people at increased risk for type 2 diabetes, such as the Diabetes Prevention Program (DPP).

1.27 (1.05 to 1.54) 1.15 (0.97 to 1.36) Organizational support for EBDM Factor 1: awareness of EBDM 1.09 (0.67 to 1.75) 1.00 (0.69 to 1.45) 1.31 (0.80 to 2.16) 1.45 (0.96 to 2.20) 1.59 (0.75 to 3.37) Factor 2: capacity for EBDM 1.12 (0.72 to 1.74) 1.04 (0.74 to 1.46) 1.23 (0.78 to 1.95) 1.42 (0.97 to 2.09) 1.62 (0.81 to 3.22) Factor 3: resource availability 1.33 (0.81 to 2.19) 0.99 (0.67 to 1.47) 1.17 (0.70 to 1.96) 1.40 (0.91 to 2.15) 1.72 (0.79 to 3.76) Factor 4: evaluation capacity 1.23 (0.83 to 1.83) 0.88 (0.64 to 1.20) 1.36 (0.90 to 2.05) 1.54 (1.08 to 2.19) 1.60 (0.87 to 2.96) Factor 5: EBDM climate cultivation 0.84 (0.46 to 1.53) 0.79 (0.49 to 1.26) 1.10 (0.60 to 2.02) 1.52 (0.92 to 2.51) 1.38 (0.55 to 3.47) Factor 6: partnerships to support EBDM 1.03 (0.62 to 1.73) 0.97 (0.65 to 1.45) 0.99 (0.59 to 1.68) 1.34 (0.88 to 2.06) 1.65 (0.78 to 3.48) OR from unadjusted bivariate model. Bold values indicate statistically significant relationships according to a n alpha=0.05 threshold. *Diet and physical activity promotion programs with people at increased risk for type 2 diabetes, such as the Diabetes Prevention Program (DPP). †Community health workers (CHWs) to deliver diet and physical activity promotion and weight management to groups or individuals with increased risk for type 2 diabetes. ‡Diabetes self-management education (DSME) with persons with diabetes delivered in community gathering places. §Diabetes management interventions identifying patients with diabetes and determining effective treatment (identify).

†Community health workers (CHWs) to deliver diet and physical activity promotion and weight management to groups or individuals with increased risk for type 2 diabetes. ‡Diabetes self-management education (DSME) with persons with diabetes delivered in community gathering places. §Diabetes management interventions identifying patients with diabetes and determining effective treatment (identify). EBDM, evidence-based decision making; EBI, evidence-based intervention; LHD, local health department; PHAB, Public Health Accreditation Board. Discussion This study found in a national sample of LHDs a wide variation in EBI offerings by category of EBI (ie, obesity vs physical activity) and by individual EBI, with half of the EBIs offered by at least 80% of the reporting LHDs. Widespread adoption of EBIs in public health practice is an encouraging development for effective prevention and management of diabetes. The results demonstrate that collaboration with other organizations in the community appears to be critical to offering EBIs; very few EBIs were offered only directly by the LHD. Offering healthier food assistance programs and breastfeeding promotion were the EBIs with the greatest percentage only being delivered directly by the LHD (17% and 16%, respectively). These may be thought to be more traditional functions of public health.12 15 16 29 However, when branching out to the other types of EBIs, with more environment and policy focus, LHDs reported collaboration to accomplish implementation.

st percentage only being delivered directly by the LHD (17% and 16%, respectively). These may be thought to be more traditional functions of public health.12 15 16 29 However, when branching out to the other types of EBIs, with more environment and policy focus, LHDs reported collaboration to accomplish implementation. Although half of the EBIs were offered by ≥80% of the sample, a quarter of the EBIs were offered by fewer than 60%. Behavioral interventions to reduce screen time; multicomponent interventions with coaching that uses technology to aid in weight loss or maintenance (eg, pedometers with computer interaction, social media); and school gardens are more newly recommended interventions, which may be, in part, why fewer LHDs reported offering these interventions than more conventional programs such as diabetes self-management education or diet and physical activity promotion programs with people at increased risk for type 2 diabetes, such as the DPP. For example, the oldest reference on the What Works for Health web page for school gardens is from 2005.30 Screening for obesity in adults and referring those with elevated BMI (>30 kg/m2) to behavioral interventions may be offered in a smaller percentage of responding LHDs, as this type of programs may be viewed as more of a clinical service, particularly as the recommendation from the US Preventive Services Task Force is focused on clinicians in primary care settings.31 There may be additional barriers to offering interventions where community health workers deliver diet and physical activity promotion and weight management to those with increased risk for type 2 diabetes, such as licensure, cost/turnover, and fears of deportation.32 33

ocused on clinicians in primary care settings.31 There may be additional barriers to offering interventions where community health workers deliver diet and physical activity promotion and weight management to those with increased risk for type 2 diabetes, such as licensure, cost/turnover, and fears of deportation.32 33 Several factors were found to be related to offering each of the diabetes-related EBIs and all four of the diabetes-related EBIs. At the individual level, older respondents and those who had been in their position longer (likely correlated factors) were less likely to report their LHD offered the EBIs. Previous studies have found that perceptions of public health practice models, such as coordinated chronic disease prevention, vary with duration in a state health department34; however, while one study found barriers to EBDM to be ranked higher by older practitioners,21 another study found older respondents reported higher levels of organizational support for EBDM.35 It is possible that older LHD staff are further removed from training, as has been seen in healthcare,36–38 or prefer to rely more heavily on learned experience than evidence-based resources when selecting interventions to implement. At the organizational level, the size of the jurisdiction served was positively associated with delivering diet and physical activity promotion programs with people at increased risk for type 2 diabetes, such as the DPP. A pilot study of LHDs in Missouri found organizational characteristics such as LHD size and accreditation status were positively associated with delivering EBIs.39 While this cross-sectional study does not allow for assessment of causation, it is notable that, at the LHD level, there was a positive association between diabetes prevalence in the state and offering several of the EBIs. This suggests that higher diabetes prevalence may elevate the issue of diabetes as a priority, and LHDs and their partners may respond with additional EBIs; alternately, higher diabetes prevalence may lead to more funding from the Centers for Disease Control and Prevention. Zhang et al 12 found diabetes prevalence to be associated with LHDs screening for diabetes, but not with delivery of obesity prevention programs.

HDs and their partners may respond with additional EBIs; alternately, higher diabetes prevalence may lead to more funding from the Centers for Disease Control and Prevention. Zhang et al 12 found diabetes prevalence to be associated with LHDs screening for diabetes, but not with delivery of obesity prevention programs. This study provides support for the positive association between organizational support for EBDM and LHDs delivering EBIs. Although all of the organizational supports for EBDM factors were positively associated with offering the EBIs, the only significant association was between evaluation capacity and identifying patients with diabetes and determining effective treatment. This aligns with previous research, which has shown the importance of organizational-level factors related to EBDM and use of research evidence. For example, a pilot study in Missouri LHDs found delivering EBIs to be associated with the perception that the agency gives incentives and rewards to help employees use EBDM principles.39 There is a growing literature that capacity for EBDM can be built with sustained efforts (eg, training, technical assistance).40

e. For example, a pilot study in Missouri LHDs found delivering EBIs to be associated with the perception that the agency gives incentives and rewards to help employees use EBDM principles.39 There is a growing literature that capacity for EBDM can be built with sustained efforts (eg, training, technical assistance).40 There are limitations to this study, including the sample size; respondents were only asked about EBIs in one category, so only the four items in the diabetes-related EBI category had items with adequate sample size. Future work could explore EBIs in the other categories (eg, obesity, nutrition) to identify whether these associations were significant and whether LHDs might be offering other interventions, which may not have had as strong of an evidence base at the time the EBIs were selected. While this was a national study with LHDs from 44 states and a balance of LHDs by jurisdiction population size, only LHDs that offered some diabetes-related services were included; thus, the findings cannot be generalized to other public health settings such as state health departments or community-based organizations or to all LHDs. While there are no directly comparable data at the national level, the NACCHO National Profile of Local Health Departments, an ongoing survey of LHDs, asked whether population-based primary prevention activities (defined broadly, rather than asking about specific EBIs as in the current study) were performed by the LHD directly, contracted out by the LHD, provided by others in the community independent of LHD funding, or not available in the community.41 The 2016 National Profile found a similar percent of LHDs reported programming nutrition (current sample: 97% offer; NACCHO sample: 97% offer), physical activity (current sample: 99% offer; NACCHO sample: 94% offer), and tobacco (current sample: 98% offer; NACCHO sample: 96% offer) were available in their community as was found in the current sample. This suggests the current sample of LHDs is likely representative of those nationwide. Other important limitations include that data were self-reported and there was only one response per LHD. It is possible that LHDs over-reported offering EBIs due to social desirability bias; however, the range of offerings suggests that respondents were willing to report that their LHD did not offer specific EBIs.

de. Other important limitations include that data were self-reported and there was only one response per LHD. It is possible that LHDs over-reported offering EBIs due to social desirability bias; however, the range of offerings suggests that respondents were willing to report that their LHD did not offer specific EBIs. The self-report nature of the data collection also makes it difficult to interpret how respondents conceptualized delivering EBIs in collaboration, where there might be less knowledge of specific EBI delivery. The current study highlights important strengths and gaps in EBI offerings in LHDs and identified correlates at the respondent and LHD levels, as well as correlates related to EBDM that are associated with offering diabetes-related EBIs. While many of the characteristics are non-modifiable (ie, age, jurisdiction population category), it is possible to modify EBDM within an LHD.42 43 Future work could conduct dissemination and implementation studies to better tease out causality, and to determine whether improvements in EBDM support and capacity can lead to increased offering of EBIs by LHDs, which is critical to addressing diabetes in the US and other countries.

to modify EBDM within an LHD.42 43 Future work could conduct dissemination and implementation studies to better tease out causality, and to determine whether improvements in EBDM support and capacity can lead to increased offering of EBIs by LHDs, which is critical to addressing diabetes in the US and other countries. We acknowledge the help of Mackenzie Robinson in data collection and reporting, and the administrative support provided by Linda Dix, Mary Adams, and Cheryl Valko of the Prevention Research Center in St Louis, Brown School, Washington University in St Louis. We also acknowledge the Centers for Disease Control and Prevention and the Robert Wood Johnson Foundation, which provided funding for the 2016 National Profile study, and the National Association of County and City Health Officials (NACCHO). Presented at: An abstract describing similar results has been accepted to the American Diabetes Association 78th Scientific Sessions (June 2018). Contributors: Conceptualization and design: RGT, RCB, RGP, PA, MHC. Survey instrument development and testing: RCB, RGT, PA, KAS, RGP, MHC. Statistical support: RGT, RGP, RRJ, RCB. Writing: RGT, RGP, RCB. Manuscript content revisions: RCB, PA, SM, RGP, MHC, RGT, KAS, MD, DD.

Presented at: An abstract describing similar results has been accepted to the American Diabetes Association 78th Scientific Sessions (June 2018). Contributors: Conceptualization and design: RGT, RCB, RGP, PA, MHC. Survey instrument development and testing: RCB, RGT, PA, KAS, RGP, MHC. Statistical support: RGT, RGP, RRJ, RCB. Writing: RGT, RGP, RCB. Manuscript content revisions: RCB, PA, SM, RGP, MHC, RGT, KAS, MD, DD. Funding: This work was supported by the National Institute of Diabetes and Digestive and Kidney Diseases of the National Institutes of Health under award numbers 5R01DK109913, 2P30DK092949, and P30DK092950. The findings and conclusions in this article are those of the authors and do not necessarily represent the official positions of the National Institutes of Health. Competing interests: None declared. Patient consent: Not required. Ethics approval: The Institutional Review Board at Washington University in St Louis Human Research Protection Office reviewed and approved this study. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: A limited data set without identifiable participant data is available on request in coordination with the Human Research Protection Office.

Significance of this study What is already known about this subject? The influence of psychosocial factors on the effectiveness of translational diabetes prevention program has been studied in certain populations, but thus far no studies have examined this in Asian Indians. What are the new findings? Among Asian Indian adults who participated in a community-based, translational diabetes prevention program, increased exercise self-efficacy at baseline predicted improved health outcomes at intervention completion, and increased change from baseline to intervention completion predicted increased exercise at follow-up. How might these results change the focus of research or clinical practice? Our results support the importance of considering psychosocial health in the development and implementation of translational diabetes prevention programs in order to improve effectiveness among participants.

What are the new findings? Among Asian Indian adults who participated in a community-based, translational diabetes prevention program, increased exercise self-efficacy at baseline predicted improved health outcomes at intervention completion, and increased change from baseline to intervention completion predicted increased exercise at follow-up. How might these results change the focus of research or clinical practice? Our results support the importance of considering psychosocial health in the development and implementation of translational diabetes prevention programs in order to improve effectiveness among participants. Introduction Type 2 diabetes mellitus (T2DM) is a chronic disease of global public health concern that is strongly associated with excess body weight and associated modifiable lifestyle factors, that is, poor diet and physical inactivity.1 In order to slow disease incidence, over the past few decades several randomized controlled trials tested the efficacy of lifestyle interventions for preventing T2DM among individuals at high risk for developing T2DM, which overall yielded strong positive findings.2–6 The largest and most diverse of these studies, the US Diabetes Prevention Program (DPP), found that an intensive lifestyle intervention targeting modest weight loss and increased physical activity reduced the incidence of type 2 diabetes by 58% over an average follow-up of 2.8 years as compared with placebo.3

dings.2–6 The largest and most diverse of these studies, the US Diabetes Prevention Program (DPP), found that an intensive lifestyle intervention targeting modest weight loss and increased physical activity reduced the incidence of type 2 diabetes by 58% over an average follow-up of 2.8 years as compared with placebo.3 The success of these trials has prompted efforts to translate the DPP framework to more ‘real-world’ settings. However, recent systematic reviews and meta-analyses of such translational trials found that there was considerable interstudy variation in program effectiveness for achieving weight loss and/or diabetes risk reduction.7 8 Although this may be explained by a number of factors, such as program adherence, intensity, or delivery, another emerging area of research that may be applicable is on the role of psychosocial factors in promoting or hindering behavior change in the setting of lifestyle interventions.

nd/or diabetes risk reduction.7 8 Although this may be explained by a number of factors, such as program adherence, intensity, or delivery, another emerging area of research that may be applicable is on the role of psychosocial factors in promoting or hindering behavior change in the setting of lifestyle interventions. Self-efficacy (SE), or an individual’s confidence in their ability to perform a task, is one widely studied psychosocial construct in health behavior research.9 Using data from a substudy of participants in the original DPP, Delahanty et al 10 found that self-reported exercise SE at baseline was independently associated with higher levels of leisure physical activity at 1 year and at the end of the study (2–3 years after randomization). Greater exercise SE at baseline was also a significant predictor of achieving the 7% weight loss goal at the end of the study.10 11 In addition to baseline SE scores, 6-month improvements in low-fat dietary SE as a result of the intervention were associated with achieving 7% weight loss at the end of the study.11 This would suggest that individuals with higher SE at baseline or with greater improvements in SE as a result of a diabetes prevention program may be more responsive to lifestyle interventions, although this has not been adequately examined yet in the context of translational diabetes prevention research.

the study.11 This would suggest that individuals with higher SE at baseline or with greater improvements in SE as a result of a diabetes prevention program may be more responsive to lifestyle interventions, although this has not been adequately examined yet in the context of translational diabetes prevention research. Another important gap in the literature is that very few, if any, studies of this nature have been conducted in low-income to middle-income countries to determine if these findings from high-income countries are applicable to other populations. The Diabetes Community Lifestyle Improvement Program (D-CLIP) was a randomized controlled research trial that tested the effectiveness of a translational diabetes prevention program with metformin when needed for preventing diabetes in overweight or obese Asian Indian adults with pre-diabetes, defined by impaired fasting glucose (IFG) and/or impaired glucose tolerance (IGT). Prior analyses have shown that D-CLIP resulted in a 32% reduction in diabetes incidence in the treatment group over a 3-year follow-up compared with control.12 We have also reported on baseline, cross-sectional data from this cohort, and found that SE levels were associated with physical activity levels and fruit and vegetable intake, and inversely associated with body mass index (BMI) and waist circumference (WC).13 In this study we investigated longitudinal changes in self-reported dietary and exercise SE from baseline to intervention completion (4 months), as well as annually until the end of the study (year 3). We also examined whether SE at baseline or improvements after the 4-month intervention were associated with reduced incidence of T2DM (the primary outcome) or greater success in achieving improvements in several secondary outcomes including weight, WC, exercise levels, and total energy intake.

nd of the study (year 3). We also examined whether SE at baseline or improvements after the 4-month intervention were associated with reduced incidence of T2DM (the primary outcome) or greater success in achieving improvements in several secondary outcomes including weight, WC, exercise levels, and total energy intake. Research design and methods Parent study The D-CLIP trial (ClinicalTrials.gov NCT01283308) was a translational, randomized controlled research study in Chennai, India described in detail previously.14 All participants provided written informed consent prior to screening, baseline testing, and study enrollment. The sample included men and women aged 20–65 years old who were overweight or obese, defined according to the WHO’s Asian-specific cut points for BMI (>23 kg/m2) or WC (≥90 cm for men or ≥80 cm for women), and diagnosed with pre-diabetes, defined according to the American Diabetes Association (ADA) criteria for IFG (fasting plasma glucose 5.6–6.9 mmol/L), IGT (2-hour, postload glucose of 7.8–11.0 mmol/L), or both.15 16 After enrollment, participants were randomized to lifestyle treatment or control. For this secondary analysis, we excluded non-compliant individuals who were lost to follow-up after the baseline visit (n=28), resulting in a final sample of 269 treatment and 281 control participants.

e of 7.8–11.0 mmol/L), or both.15 16 After enrollment, participants were randomized to lifestyle treatment or control. For this secondary analysis, we excluded non-compliant individuals who were lost to follow-up after the baseline visit (n=28), resulting in a final sample of 269 treatment and 281 control participants. Study interventions The D-CLIP lifestyle treatment consisted of group-based, culturally appropriate lifestyle classes adapted from the original DPP.14 Briefly, it included 16 weekly core intervention classes on active lifestyle changes (months 0–4), followed by 8 maintenance classes (months 5–6). The two study goals were ≥7% weight loss and ≥150 min weekly of moderate-intensity exercise. After 4 months of the core intervention, lifestyle participants were prescribed metformin (500 mg twice daily) if they were at high risk of conversion to diabetes, defined as having both IFG and IGT or IFG and hemoglobin A1c ≥5.7% (39 mmol/mol). After 6 months when all classes were complete, contact with study staff was minimal, except at follow-up study visits. Control participants received the study site’s standard of care for pre-diabetes, which included a single day of one-on-one visits with a physician, a dietitian, and a fitness trainer, and one group class on diabetes prevention, but no additional contact except at study visits. No control arm participants received metformin because it was not part of standard of care at the study site.

pre-diabetes, which included a single day of one-on-one visits with a physician, a dietitian, and a fitness trainer, and one group class on diabetes prevention, but no additional contact except at study visits. No control arm participants received metformin because it was not part of standard of care at the study site. Measurements Study visits took place at baseline, postcore intervention (4 months), postmaintenance intervention (6 months), 12 months, and every 6 months until study closeout (3–4 years after randomization) or diabetes diagnosis. For this analysis, we only used data collected up to year 3, although a small number of participants (n=81) were followed for another 6–12 months after this. Sociodemographic information was assessed at baseline by self-reported questionnaires. Pre-diabetes category and diabetes incidence were assessed by semiannual fasting blood draws at baseline and every 6 months and/or by oral glucose tolerance tests (OGTTs) at baseline and annually using the ADA criteria.15 Anthropometrics, including weight, height, and WC, were assessed by physical exams at all study visits, and weight and height were used to calculate the BMI. Physical activity levels were estimated in terms of weekly minutes of exercise, which were calculated using survey questions asking individuals how many days per week they exercise and how long each exercise session lasts on average. Total energy intake was assessed at annual visits by a food frequency questionnaire developed for South Indian populations.17

in terms of weekly minutes of exercise, which were calculated using survey questions asking individuals how many days per week they exercise and how long each exercise session lasts on average. Total energy intake was assessed at annual visits by a food frequency questionnaire developed for South Indian populations.17 Psychosocial assessments SE was assessed by self-reported surveys at baseline, 4 months, and at annual visits. Exercise SE was measured using an instrument developed by Sallis et al 18, which asked about an individual’s perception that he/she has the ability to exercise in 12 different situations using a 5-point Likert-type scale. The instrument provided scores for exercise SE on two scales: ‘sticking to it’ (adhering to an exercise regimen regardless of mood and situation) and ‘making time’ (prioritizing exercise over other time demands). These two subscores were summed to yield a total exercise SE score. Dietary SE was measured using the Weight Efficacy Lifestyle (WEL) questionnaire.19 This instrument assesses an individual’s confidence in his/her ability to avoid overeating using a 10-point Likert-type scale. The WEL provides a total score as well as scores on five subscales: negative emotions, availability, social pressure, physical discomfort, and positive activities.

estyle (WEL) questionnaire.19 This instrument assesses an individual’s confidence in his/her ability to avoid overeating using a 10-point Likert-type scale. The WEL provides a total score as well as scores on five subscales: negative emotions, availability, social pressure, physical discomfort, and positive activities. Statistical analysis All statistical analyses were performed in SAS V.9.4 (SAS Institute Inc., Cary, NC, USA). Baseline total and component scores for dietary and exercise SE, as well as other characteristics of the sample, were summarized as mean±SD and counts (percentages). Mixed-effects regression was used to calculate changes in SE total and component scores over time by treatment group. Each model included participants as a random effect, and time, treatment group, and a time × treatment group interaction as fixed effects. Other covariates including sex and baseline age, BMI, and pre-diabetes type were included as fixed effects. Time was treated as a discrete ordinal variable because there was evidence of a non-linear pattern of change over time for the psychosocial variables. Within-group differences were assessed by comparing least square means (LS-means) at each time point (4, 12, 24, and 36 months) to baseline, and between-group differences were assessed by comparing LS-means between groups at each time point. LS-mean differences and 95% CI are reported, as well as p values adjusted for multiple comparisons by the Tukey-Kramer method.

ing least square means (LS-means) at each time point (4, 12, 24, and 36 months) to baseline, and between-group differences were assessed by comparing LS-means between groups at each time point. LS-mean differences and 95% CI are reported, as well as p values adjusted for multiple comparisons by the Tukey-Kramer method. For the subsequent analyses, the analytical sample was limited to treatment participants to test intervention-specific associations. First, Cox proportional hazard models were used to assess whether SE at baseline or SE change from baseline to intervention completion (at 4 months) were associated with time to T2DM, adjusting for sex and baseline age, BMI, and pre-diabetes type. Both exercise and dietary SE variables were entered into the model as pairs consisting of baseline and 4-month change values, which were calculated as SE score at 4 months minus SE score at baseline. Date of last follow-up was marked by either diagnosis of T2DM or censorship due to study completion (3 years) or dropout.

oth exercise and dietary SE variables were entered into the model as pairs consisting of baseline and 4-month change values, which were calculated as SE score at 4 months minus SE score at baseline. Date of last follow-up was marked by either diagnosis of T2DM or censorship due to study completion (3 years) or dropout. Lastly, associations of SE with secondary outcomes, that is, change in weight, WC, weekly exercise, and total energy intake, at 4, 12, 24, and 36 months were evaluated using linear regression. Similar to above, SE variables were entered into the model as baseline values and 4-month change values. Models were adjusted for sex, and baseline age, BMI, and pre-diabetes type. For the models with 12-month, 24-month, and 36-month changes as the outcome, the analytical sample was limited to only participants who had follow-up data at that time point, and a covariate for exercise or dietary SE score at that time point was included in the model. Beta coefficients are reported as standardized coefficients and standard errors, and statistical significance was set at p<0.05.

he analytical sample was limited to only participants who had follow-up data at that time point, and a covariate for exercise or dietary SE score at that time point was included in the model. Beta coefficients are reported as standardized coefficients and standard errors, and statistical significance was set at p<0.05. Results Baseline health and psychosocial characteristics of the sample are summarized in table 1. There were no significant differences between groups for exercise or dietary SE total or component scores at baseline. In mixed-effects models, exercise and dietary SE total scores increased significantly from baseline to 4 months within the treatment group (LS-mean difference (95% CI), 4 months vs baseline: 0.49 (0.20 to 0.79) for exercise SE, adjusted p (adj-p)=0.04; 10.3 (4.0 to 16.6) for dietary SE, adj-p=0.04) but not in the control group (figure 1). There was also a significant between-group difference in exercise and dietary SE at 4 months (LS-mean difference (95% CI), treatment vs control: 0.99 (0.67 to 1.32) for exercise SE, adj-p<0.001; 16.4 (9.4 to 23.5) for dietary SE, adj-p<0.001). However, over long-term follow-up (≥12 months), within-group differences at each time point compared with baseline were no longer significant for either exercise or dietary SE, and between-group differences were no longer significant for exercise SE. For dietary SE, the treatment group sustained a higher total score compared with the control group at 12 and 36 months (LS-mean difference (95% CI), treatment vs control: 10.4 (4.5 to 16.4) for 12 months, adj-p=0.02; 23.1 (9.2 to 37.1) for 36 months, adj-p=0.04) (figure 1). Similar trends were seen in the exercise and dietary SE component scores (online supplementary figures S1 and S2).

Figure 1 Total scores for psychosocial variables over 3-year follow-up by treatment group: (A) exercise self-efficacy and (B) dietary self-efficacy. Markers represent the least square mean for each group and time, adjusted for sex and baseline age, body mass index, and pre-diabetes type. Error bars represent 95% CIs. Circles: control group; squares: treatment group. Table 1 Demographic, health, and psychosocial characteristics of the analytical sample at baseline by treatment group (N=550) Control (n=281) Treatment (n=269) Female, n (%) 110 (39.2) 96 (35.7) Male, n (%) 171 (60.9) 173 (64.3) Age (years), mean±SD 44.3±9.4 44.9±8.8 Anthropometrics, mean±SD BMI (kg/m2) 27.9±3.7 28.0±3.7 Male WC (cm) 90.1±8.3 89.9±9.1 Female WC (cm) 97.7±7.7 97.8±8.2 Pre-diabetes type, n (%) IGT 81 (28.8) 81 (30.1) IFG 81 (28.8) 85 (31.6) IGT+IFG 119 (42.4) 103 (38.3) Dietary self-efficacy, mean±SD Positive activities score 23.6±9.5 23.8±8.8 Availability score 21.0±9.6 22.0±9.4 Physical discomfort score 24.4±8.9 25.0±8.8 Negative emotions score 23.8±9.2 24.7±8.9 Social pressure score 22.1±8.9 23.3±8.5 Total score 114.9±38.9 118.7±37.3 Exercise self-efficacy, mean±SD Sticking to it 3.4±1.0 3.4±1.0 Making time 3.7±1.0 3.8±1.0 Total scores 7.1±1.8 7.2±1.8 Physical activity, mean±SD Exercise, min/week 77.4±103.3 84.2±113.9 Dietary intake, mean±SD Total daily energy intake 3001±838 2950±891 BMI, body mass index; IFG, impaired fasting glucose; IGT, impaired glucose tolerance; WC, waist circumference.

3.4±1.0 3.4±1.0 Making time 3.7±1.0 3.8±1.0 Total scores 7.1±1.8 7.2±1.8 Physical activity, mean±SD Exercise, min/week 77.4±103.3 84.2±113.9 Dietary intake, mean±SD Total daily energy intake 3001±838 2950±891 BMI, body mass index; IFG, impaired fasting glucose; IGT, impaired glucose tolerance; WC, waist circumference. In our analyses of treatment participants only, there was no association of baseline or initial 4-month change in dietary and exercise SE with T2DM incidence over the 3-year follow-up in Cox proportional hazard models (online supplementary table S1). Table 2 reports the standardized beta coefficients, standard errors, and p values for the associations of baseline and initial 4-month change in SE scores with secondary outcomes among treatment participants at 4, 12, 24, and 36 months of follow-up. For short-term (4 months) outcomes, baseline scores for exercise SE were significantly associated with decreased weight and WC, and increased exercise at 4 months among treatment participants in linear regression adjusted for sex, and baseline age, BMI, and pre-diabetes type (all p<0.05). Initial 4-month change in exercise SE was also associated with increased exercise at 4 months (p<0.001). For secondary outcomes at 12-month, 24-month or 36-month follow-up, neither baseline nor initial 4-month change in dietary and exercise SE predicted weight change at any time point. However, the covariates for dietary and exercise SE at 12 months were associated with decreased weight at 12 months; the same was true for dietary SE and weight at 24 months (p<0.05). For the other outcomes, initial 4-month change in exercise SE significantly predicted decreased WC at 24 months, and increased exercise at 12 and 24 months among treatment participants (p<0.05). None of the variables for dietary nor exercise SE were associated with changes in energy intake at any time point (table 2).

For the other outcomes, initial 4-month change in exercise SE significantly predicted decreased WC at 24 months, and increased exercise at 12 and 24 months among treatment participants (p<0.05). None of the variables for dietary nor exercise SE were associated with changes in energy intake at any time point (table 2). Table 2 Estimates from linear regression of dietary and exercise SE total scores with change in secondary health outcomes at 4, 12, 24, and 36 months among treatment participants*† SE variable Weight (kg) WC (cm) Exercise (min/week) Energy intake (kcal) β (standard error) P values β (standard error) P values β (standard error) P values β (standard error) P values 4 months (n=238) 4 months (n=237) 4 months (n=241) 4 months (n=195) Dietary SE Baseline −0.02 (0.01) 0.81 −0.03 (0.01) 0.73 0.02 (0.27) 0.77 −0.02 (2.19) 0.82 4-month Δ −0.02 (0.01) 0.84 0.02 (0.01) 0.80 0.05 (0.27) 0.50 −0.12 (2.21) 0.20 Exercise SE Baseline −0.22 (0.14) 0.02 −0.19 (0.24) 0.04 0.22 (6.0) 0.01 −0.01 (47.8) 0.91 4-month Δ −0.15 (0.12) 0.10 −0.15 (0.20) 0.11 0.39 (5.2) <0.01 0.08 (41.4) 0.43 12 months (n=235) 12 months (n=235) 12 months (n=233) 12 months (n=205) Dietary SE Baseline −0.02 (0.01) 0.76 −0.14 (0.01) 0.11 0.06 (0.35) 0.49 −0.07 (2.33) 0.49 4-month Δ −0.1 (0.01) 0.32 −0.07 (0.01) 0.40 −0.06 (0.32) 0.49 0.03 (2.05) 0.75 12 months −0.13 (0.01) 0.01 −0.07 (0.01) 0.38 −0.02 (0.32) 0.79 0.15 (2.11) 0.08 Exercise SE Baseline −0.03 (0.19) 0.20 −0.19 (0.29) 0.07 0.01 (8.47) 0.95 −0.02 (53.66) 0.88 4-month Δ 0.001 (0.15) 0.19 −0.15 (0.23) 0.14 0.24 (6.66) 0.02 −0.06 (41.84) 0.57 12 months −0.22 (0.15)

Dietary SE Baseline −0.02 (0.01) 0.76 −0.14 (0.01) 0.11 0.06 (0.35) 0.49 −0.07 (2.33) 0.49 4-month Δ −0.1 (0.01) 0.32 −0.07 (0.01) 0.40 −0.06 (0.32) 0.49 0.03 (2.05) 0.75 12 months −0.13 (0.01) 0.01 −0.07 (0.01) 0.38 −0.02 (0.32) 0.79 0.15 (2.11) 0.08 Exercise SE Baseline −0.03 (0.19) 0.20 −0.19 (0.29) 0.07 0.01 (8.47) 0.95 −0.02 (53.66) 0.88 4-month Δ 0.001 (0.15) 0.19 −0.15 (0.23) 0.14 0.24 (6.66) 0.02 −0.06 (41.84) 0.57 12 months −0.22 (0.15) <0.01 −0.08 (0.23) 0.28 0.06 (6.61) 0.48 −0.06 (43.34) 0.49 24 months (n=207) 24 months (n=207) 24 months (n=199) 24 months (n=186) Dietary SE Baseline 0.13 (0.01) 0.17 0 (0.01) 0.99 −0.08 (0.31) 0.42 0.18 (2.41) 0.07 4-month Δ 0.117 (0.01) 0.19 0.06 (0.01) 0.47 −0.06 (0.28) 0.49 0.12 (2.28) 0.22 24 months −0.18 (0.01) 0.03 −0.1 (0.01) 0.24 0.03 (0.29) 0.71 0.06 (2.22) 0.52 Exercise SE Baseline −0.11 (0.24) 0.34 −0.2 (0.33) 0.06 0.028 (6.98) 0.80 0.2 (53.12) 0.09 4-month Δ −0.13 (0.2) 0.25 −0.29 (0.27) 0.01 0.26 (5.67) 0.02 0.14 (44.21) 0.21 24 months −0.01 (0.19) 0.94 0.02 (0.25) 0.79 0.15 (5.26) 0.06 −0.12 (41.15) 0.17 36 months (n=100) 36 months (n=100) 36 months (n=99) 36 months (n=89)

0.03 −0.1 (0.01) 0.24 0.03 (0.29) 0.71 0.06 (2.22) 0.52 Exercise SE Baseline −0.11 (0.24) 0.34 −0.2 (0.33) 0.06 0.028 (6.98) 0.80 0.2 (53.12) 0.09 4-month Δ −0.13 (0.2) 0.25 −0.29 (0.27) 0.01 0.26 (5.67) 0.02 0.14 (44.21) 0.21 24 months −0.01 (0.19) 0.94 0.02 (0.25) 0.79 0.15 (5.26) 0.06 −0.12 (41.15) 0.17 36 months (n=100) 36 months (n=100) 36 months (n=99) 36 months (n=89) Dietary SE Baseline 0.07 (0.02) 0.66 0.06 (0.02) 0.69 −0.14 (0.52) 0.34 0.13 (3.51) 0.37 4-month Δ −0.04 (0.01) 0.74 −0.05 (0.02) 0.72 −0.12 (0.43) 0.38 −0.119 (2.96) 0.37 36 months −0.11 (0.01) 0.40 −0.07 (0.02) 0.60 0.08 (0.46) 0.56 −0.07 (3.1) 0.56 Exercise SE Baseline −0.15 (0.38) 0.44 −0.12 (0.48) 0.52 0.22 (13.44) 0.26 0.31 (90.42) 0.11 4-month Δ −0.1 (0.33) 0.61 −0.04 (0.41) 0.83 0.37 (11.38) 0.07 0.28 (78.02) 0.17 36 months −0.06 (0.28) 0.65 −0.04 (0.35) 0.74 −0.08 (9.54) 0.51 0 (66.83) 0.98 *Models were adjusted for sex, and baseline age, body mass index, and pre-diabetes type. Estimates are standardized beta coefficients. †Change values for each secondary outcome and for 4-month change in SE scores were calculated as follow-up baseline value. ‡Bold values indicate a statistically siganificant association at p<0.05. SE, self-efficacy; WC, waist circumference.

Dietary SE Baseline 0.07 (0.02) 0.66 0.06 (0.02) 0.69 −0.14 (0.52) 0.34 0.13 (3.51) 0.37 4-month Δ −0.04 (0.01) 0.74 −0.05 (0.02) 0.72 −0.12 (0.43) 0.38 −0.119 (2.96) 0.37 36 months −0.11 (0.01) 0.40 −0.07 (0.02) 0.60 0.08 (0.46) 0.56 −0.07 (3.1) 0.56 Exercise SE Baseline −0.15 (0.38) 0.44 −0.12 (0.48) 0.52 0.22 (13.44) 0.26 0.31 (90.42) 0.11 4-month Δ −0.1 (0.33) 0.61 −0.04 (0.41) 0.83 0.37 (11.38) 0.07 0.28 (78.02) 0.17 36 months −0.06 (0.28) 0.65 −0.04 (0.35) 0.74 −0.08 (9.54) 0.51 0 (66.83) 0.98 *Models were adjusted for sex, and baseline age, body mass index, and pre-diabetes type. Estimates are standardized beta coefficients. †Change values for each secondary outcome and for 4-month change in SE scores were calculated as follow-up baseline value. ‡Bold values indicate a statistically siganificant association at p<0.05. SE, self-efficacy; WC, waist circumference. Discussion This study expanded on previous findings from the D-CLIP trial by examining the role of psychosocial variables as correlates of health outcomes among Asian Indian adults with pre-diabetes in a stepwise diabetes prevention program (culturally adapted, group-based, DPP-like program with metformin if needed). Consistent with earlier translational research,20–26 the D-CLIP program resulted in higher scores for exercise and dietary SE for the treatment group compared with the control group at completion of the core intervention. However, these increases in SE within the treatment group were not maintained over long-term follow-up, and returned near baseline levels after 4 months. Nonetheless, there remained a significant difference in dietary SE between the treatment and control groups at 12 and 36 months. However, there were no differences between groups for exercise SE at 12 months or after. These findings indicate that certain SE beliefs may require continual reinforcement to sustain in the long term. During the lifestyle intervention, participants had weekly group-based exercise classes. It is possible that active participation in these classes overcame some of the barriers to exercise and promoted increased exercise SE, which was not sustained once these classes stopped.27 Other studies have also reported inconsistent findings on long-term changes in SE after a translational DPP-style intervention. While some did find differences at 12 months for certain measures of SE,28 other studies found no long-term differences,22 23 and no studies to our knowledge have reported on longer term (≥2 years) follow-up of SE scores. Additional qualitative research may be warranted to elucidate the individual and/or program-related factors that explain why some psychosocial beliefs are sustained over the long term while others are not.

differences,22 23 and no studies to our knowledge have reported on longer term (≥2 years) follow-up of SE scores. Additional qualitative research may be warranted to elucidate the individual and/or program-related factors that explain why some psychosocial beliefs are sustained over the long term while others are not. In our analysis of the relationship of SE and diabetes-related health outcomes among treatment participants, we did not observe significant associations of SE at baseline, nor 4-month improvement in SE as a result of the intervention, with incidence of T2DM over follow-up. This would mean that the relative success among treatment participants in achieving this outcome, that is, prevention of T2DM, was more strongly related to factors other than SE, such as pre-diabetes type, age, and sex, as reported previously.

n SE as a result of the intervention, with incidence of T2DM over follow-up. This would mean that the relative success among treatment participants in achieving this outcome, that is, prevention of T2DM, was more strongly related to factors other than SE, such as pre-diabetes type, age, and sex, as reported previously. For secondary outcomes, exercise SE at baseline was a significant predictor of improved weight and WC at completion of the D-CLIP core intervention at 4 months, supporting that individuals with higher exercise SE going into the intervention experienced greater success with improving obesity-related outcomes. We also found that improved weight at later time points, that is, 12 and 24 months, was significantly related to dietary and/or exercise SE at these time points. These findings are similar to that of other studies showing that SE is a correlate or mediator of weight and WC changes in a DPP-style intervention.22–24 29 Exercise SE at baseline also predicted increased exercise levels at 4 months, while initial 4-month change in exercise SE predicted increased exercise levels at 12 and 24 months. This suggests that greater SE in part explains interindividual differences in improvements in physical activity among participants who received the D-CLIP intervention. This also mirrors findings from other studies, such as the Special Diabetes Program for Indian Diabetes Prevention, which found that participants with higher SE were more likely to be categorized in the ‘Action-Maintenance’ stage (defined by the transtheoretical model of behavioral change), and exhibited higher physical activity and healthier diets compared with the ‘Contemplation’ and ‘Preparation’ stages.30 31 This alignment of improved exercise SE with improved physical activity is logical, and suggests that careful attention should be paid to the type of SE being measured in future studies depending on the health outcomes of interest, as these differences may explain discrepancies when comparing study findings. For example, Gillison et al 28 found significant associations of SE with dietary change, motivation and social support in a group-based translational lifestyle intervention, but not change in physical activity levels, which may be because SE was measured in this study in relation to dietary, but not physical activity behaviors.

le, Gillison et al 28 found significant associations of SE with dietary change, motivation and social support in a group-based translational lifestyle intervention, but not change in physical activity levels, which may be because SE was measured in this study in relation to dietary, but not physical activity behaviors. In contrast to the aforementioned studies, we found no associations of exercise or dietary SE with change in energy intake at any time point. It is possible that, similar to T2DM incidence, success in reducing energy intake over follow-up may have been more strongly related to other factors. Limitations in the instruments used to measure energy intake and SE may also have influenced this finding. First, the food frequency questionnaire used to measure dietary intake relied on self-reported frequencies and portion sizes, which can be subject to recall bias and social desirability bias,32 33 especially in participants with higher BMI,17 thereby potentially limiting our ability to accurately estimate energy intake. However, it is not likely that this significantly impacted the results as the measurement tools used were validated and were able to provide relative measurements of these variables. Alternatively, it is possible that the instruments used to measure SE did not provide sensitive estimates of these variables, as neither the exercise nor dietary SE questionnaires have been validated specifically for the Indian population. Future studies aiming to develop and validate an instrument for measuring health-related SE in this population may be warranted.

ts used to measure SE did not provide sensitive estimates of these variables, as neither the exercise nor dietary SE questionnaires have been validated specifically for the Indian population. Future studies aiming to develop and validate an instrument for measuring health-related SE in this population may be warranted. Other limitations of this analysis included the use of brief questionnaires for assessing psychosocial, dietary, and physical activity variables, which may have been too simple to provide accurate measurements. The self-report surveys were also subject to the biases of self-reported data as previously explained. Another limitation was that participants were no longer followed after being diagnosed with T2DM; thus, we had no data on their clinical course after diagnosis and this limited our sample size for analyses of outcomes over a long-term follow-up. Finally, we only assessed certain psychosocial constructs in this study to reduce respondent burden, and thus did not capture the intervention’s effects on other psychosocial predictors of behavioral change. Future studies may benefit from conducting a full battery of cognitive tests, including other variables such as planning or knowledge variables, which in some studies were a more important predictor of health behavior change than SE.22 34

the intervention’s effects on other psychosocial predictors of behavioral change. Future studies may benefit from conducting a full battery of cognitive tests, including other variables such as planning or knowledge variables, which in some studies were a more important predictor of health behavior change than SE.22 34 The D-CLIP trial has several strengths. It used a randomized controlled trial design, and the sample was balanced after randomization. The trial had a longer duration of follow-up than most translational trials of diabetes prevention program (up to 4 years for some participants), allowing us to understand the longer term impact of a translational diabetes prevention program over time and after the core and maintenance intervention was complete. This study was one of the first to examine the role of psychosocial factors in diabetes prevention and related health outcomes among Asian Indian adults, a population at higher risk for T2DM at younger ages and lower BMIs.35 The sample population also included all forms of pre-diabetes, adding a novel perspective to our understanding of the mechanisms and impact of a community-based diabetes prevention program in this region. The translational nature of the intervention was also a strength in that the D-CLIP classes were designed to be delivered in a way that is lower in cost and less resource-intensive than individualized programs.

r understanding of the mechanisms and impact of a community-based diabetes prevention program in this region. The translational nature of the intervention was also a strength in that the D-CLIP classes were designed to be delivered in a way that is lower in cost and less resource-intensive than individualized programs. Conclusions This study provides additional insights into the potential role of psychosocial factors, in particular SE, in predicting success in achieving primary and secondary outcomes within a community-based, translational diabetes prevention program. Although SE did not impact the risk of developing type 2 diabetes, baseline SE and/or initial improvements in SE as a result of the intervention predicted improved weight loss, reductions in WC, and exercise. Future studies are needed to better understand the specific mechanisms by which psychosocial factors mediate the associations of DPP-style interventions with improved health outcomes as this may also help to improve the effectiveness of interventions.

tion predicted improved weight loss, reductions in WC, and exercise. Future studies are needed to better understand the specific mechanisms by which psychosocial factors mediate the associations of DPP-style interventions with improved health outcomes as this may also help to improve the effectiveness of interventions. Contributors: CEC was the primary author of the manuscript and conducted the statistical analysis. MBW contributed to the conceptualization and design of the study and oversaw the statistical analysis. HR, LRS, RMA and VM provided critical feedback on the statistical analysis, provided edits to the text, and contributed to the discussion and interpretation of the data. All authors reviewed and approved the manuscript. MBW is the guarantor of this work and as such had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis.

edits to the text, and contributed to the discussion and interpretation of the data. All authors reviewed and approved the manuscript. MBW is the guarantor of this work and as such had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Funding: This project is supported by a BRiDGES grant from the International Diabetes Federation (LT07-115). BRiDGES, an International Diabetes Federation project, is supported by an educational grant from Lilly Diabetes. Additional support was provided by the Global Health Institute at Emory University. CEC received funding from a National Institutes of Health T32 grant (T32-DK007734-20). MBW received funding from two National Institute of Diabetes and Digestive and Kidney Diseases T32 grants (5T3-2DK-007298-33 and T32-DK-007734-16) and is currently funded by a National Institute of Diabetes and Digestive and Kidney Diseases Center grant (P30-DK-111024). LRS received funding from the Human Health Molecules to Humankind program funded by the Burroughs Wellcome Fund (grant BWF 1008188). Competing interests: None declared. Ethics approval: The protocol was approved by the Emory University Institutional Review Board and the Madras Diabetes Research Foundation Ethics Committee. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: No additional data are available.

Background Type 2 diabetes mellitus, characterized by abnormal glucose metabolism and an inadequate compensatory insulin secretion response, accounts for over 90% of all cases of diabetes.1 Treatment of type 2 diabetes often targets abnormal glucose levels, yet outcomes measured in clinical trials of glucose-lowering interventions are inconsistent2 and this heterogeneity in outcomes limits the usefulness of trial findings to patients and other decision-makers.3 4 The Selecting Core Outcomes for Randomised Effectiveness trials In Type 2 diabetes (SCORE-IT) study5 aims to address these issues by developing a core outcome set (COS) for use in clinical trials of glucose-lowering interventions in people with type 2 diabetes.

f trial findings to patients and other decision-makers.3 4 The Selecting Core Outcomes for Randomised Effectiveness trials In Type 2 diabetes (SCORE-IT) study5 aims to address these issues by developing a core outcome set (COS) for use in clinical trials of glucose-lowering interventions in people with type 2 diabetes. COS represent agreed standardized sets of outcomes that should be measured and reported, as a minimum, in all clinical trials for a specific health condition.6 The development and implementation of COS can improve the relevance and consistency of trial outcomes and allow the results of clinical research to be pooled and compared, thereby reducing waste in research.7 The first step in the development of a COS typically involves a review of existing knowledge (eg, systematic review of outcomes used in previous studies) to inform the consensus process.8 In the case of the SCORE-IT study, this has involved a systematic review of outcomes used in registered clinical trials of glucose-lowering interventions for type 2 diabetes.2 While this is typical of many COS, clinical trials often overlook the outcomes that are important to patients9 10 and so the outcomes identified in such reviews are likely to predominantly reflect the perspectives of researchers and clinicians.11 This is a concern, as there is evidence that the input of patients leads to the identification of core outcomes beyond those identified by practitioners alone.12–15 Furthermore, the Core Outcome Set-Standards for Development project recently established the inclusion of patient stakeholders within the COS development process to be a minimum standard for COS.16

at the input of patients leads to the identification of core outcomes beyond those identified by practitioners alone.12–15 Furthermore, the Core Outcome Set-Standards for Development project recently established the inclusion of patient stakeholders within the COS development process to be a minimum standard for COS.16 Primary qualitative research is particularly suited to accessing patient perspectives on outcomes, as these studies allow patients to voice their views and experiences in an open-ended way and in their own words.17 The findings from such studies can contribute in several ways to COS development including to the ‘long list’ of outcomes needed in the early stages of the process. However, primary qualitative research can be resource intensive and require study team expertise in qualitative methodology.18 Systematic reviews of qualitative studies potentially provide an alternative to conducting primary qualitative research where suitable published studies are available.17

stages of the process. However, primary qualitative research can be resource intensive and require study team expertise in qualitative methodology.18 Systematic reviews of qualitative studies potentially provide an alternative to conducting primary qualitative research where suitable published studies are available.17 Previous systematic reviews of qualitative studies have identified outcomes of importance to patients and informed the development of COS in critical illness, bariatric and metabolic surgery, neonatal care and tuberculosis.19 20 Such reviews have identified outcomes that were not reported in systematic reviews of clinical trials, indicating that qualitative evidence is needed to ensure that the initial ‘long list’ of outcomes is comprehensive and does not omit outcomes important to patients.19 21 However, these previous systematic reviews of qualitative studies have undertaken exhaustive literature searches and are likely to be time consuming and resource intensive.22 Systematic reviews inform the early phases of COS and comprise one small aspect of the COS development process. As many developers have limited time and resources, there is a need to identify an expedited approach for identifying outcomes that are important to patients, which can subsequently inform the development of COS. Existing qualitative reviews for a specific condition could be considered; however, in the case of type 2 diabetes, these reviews23 24 have not been conducted in the context of COS development and as such group findings into overarching concepts, many of which have little relevance to COS. Here we report an expedited approach for identifying outcomes reported by people with type 2 diabetes when asked about their lived experience.

s, these reviews23 24 have not been conducted in the context of COS development and as such group findings into overarching concepts, many of which have little relevance to COS. Here we report an expedited approach for identifying outcomes reported by people with type 2 diabetes when asked about their lived experience. Aims Our aims in undertaking this review were: to identify an expedited approach for incorporating the patient perspective within the initial stages of COS development; to identify outcomes important to people with type 2 diabetes for inclusion in a ‘long list’ of outcomes for the COS consensus process; and to compare outcomes identified from the qualitative literature with those identified via a previous systematic review of outcomes measured in type 2 diabetes clinical trials. We also aimed to compare outcomes identified from qualitative studies of patients living in low/middle-income countries (LMIC) with outcomes identified from qualitative studies conducted in higher income countries (HIC). The prevalence of diabetes in LMICs is high25 and it is important to examine if outcomes voiced by patients living in LMICs differ from those of patients in HICs.

tudies of patients living in low/middle-income countries (LMIC) with outcomes identified from qualitative studies conducted in higher income countries (HIC). The prevalence of diabetes in LMICs is high25 and it is important to examine if outcomes voiced by patients living in LMICs differ from those of patients in HICs. Methods Search strategy Using rapid review methodology, which involves streamlining traditional systematic review methods to synthesize evidence within a shortened time frame,26 we searched a single health-related database, MEDLINE, with no date restrictions on 22 June 2017. The search terms, which are indicated in table 1, comprised qualitative methodological filters previously shown to identify qualitative research from the MEDLINE electronic database.27 The research field for type 2 diabetes is vast and so search filters designed for maximum specificity27 were selected to minimize irrelevant references. Table 1 MEDLINE search strategy

Methods Search strategy Using rapid review methodology, which involves streamlining traditional systematic review methods to synthesize evidence within a shortened time frame,26 we searched a single health-related database, MEDLINE, with no date restrictions on 22 June 2017. The search terms, which are indicated in table 1, comprised qualitative methodological filters previously shown to identify qualitative research from the MEDLINE electronic database.27 The research field for type 2 diabetes is vast and so search filters designed for maximum specificity27 were selected to minimize irrelevant references. Table 1 MEDLINE search strategy Multifield search (type 2 diabetes OR type II diabetes) Abstract AND patient* Abstract AND (Qualitative OR Themes) Abstract AND (symptom OR treatment OR living with) Abstract NOT (co-morbid* OR foot ulcers OR retinopathy OR nephropathy OR bariatric surgery OR non-alcoholic fatty liver disease OR cardiovascular disease) Abstract Studies reporting qualitative empirical findings of the views and experiences of people with type 2 diabetes on their condition and treatment were eligible for inclusion. Type 1 diabetes, gestational diabetes, type 2 diabetes in children and maturity onset diabetes of the young were outside the scope of this review. Studies where the primary focus was the treatment of diabetes comorbidities or complications (eg, diabetic foot ulcer, diabetic retinopathy, nephropathy, bariatric surgery, non-alcoholic fatty liver disease and cardiovascular disease) were also excluded.

nset diabetes of the young were outside the scope of this review. Studies where the primary focus was the treatment of diabetes comorbidities or complications (eg, diabetic foot ulcer, diabetic retinopathy, nephropathy, bariatric surgery, non-alcoholic fatty liver disease and cardiovascular disease) were also excluded. Study selection SLG and NLH identified and screened titles and abstracts from the MEDLINE search for eligibility, batch checked 10% of these to ensure consistency and discussed uncertainties. There was good agreement between reviewers during the batch check; therefore, the two reviewers each reviewed half the remaining abstracts. Full texts were retrieved and reviewed for articles meeting the following inclusion criteria: participants were people with type 2 diabetes or their partners, the focus was type 2 diabetes and not an associated comorbidity, and qualitative data collection methods (interviews or focus groups) were used. SLG and NLH each reviewed half the full-text papers for eligibility. Data extraction For each included study the following data were extracted: Study aim. Participants (number in study, age, sex, number of years with diabetes). Geographical location of participants. Qualitative data collection methods used. Text excerpts relevant to outcomes.

Study selection SLG and NLH identified and screened titles and abstracts from the MEDLINE search for eligibility, batch checked 10% of these to ensure consistency and discussed uncertainties. There was good agreement between reviewers during the batch check; therefore, the two reviewers each reviewed half the remaining abstracts. Full texts were retrieved and reviewed for articles meeting the following inclusion criteria: participants were people with type 2 diabetes or their partners, the focus was type 2 diabetes and not an associated comorbidity, and qualitative data collection methods (interviews or focus groups) were used. SLG and NLH each reviewed half the full-text papers for eligibility. Data extraction For each included study the following data were extracted: Study aim. Participants (number in study, age, sex, number of years with diabetes). Geographical location of participants. Qualitative data collection methods used. Text excerpts relevant to outcomes. Our approach to identifying text relevant to outcomes was deductive. It is important to note that none of the qualitative studies explicitly aimed to identify outcomes, although they did contain text that we could interpret as relevant to type 2 diabetes outcomes. Such text comprised any reports about how patients felt or functioned in relation to their diabetes and the healthcare or treatment they had received. Others have previously defined such reports as relevant to outcomes if they describe something that could be used to assess the effect of a healthcare intervention on the patient’s life.28 29 We were also guided by this definition. For example, we interpreted the patient quotation, ‘I just think I like to get my blood glucose inside the right range, as we should’ as about the outcome ‘glycaemic control’. All such text, including participants’ quotations about their views and experiences and the authors’ commentary, was extracted verbatim from both the results and discussion sections of included papers. This text was entered as a separate row in a Microsoft excel spreadsheet (available on request).

‘glycaemic control’. All such text, including participants’ quotations about their views and experiences and the authors’ commentary, was extracted verbatim from both the results and discussion sections of included papers. This text was entered as a separate row in a Microsoft excel spreadsheet (available on request). SLG and NLH both reviewed and interpreted outcomes from five included studies and checked these for agreement. They each independently reviewed half of the remaining studies. Quality appraisal The role of quality appraisal of qualitative studies in systematic reviews, and whether quality assessment should be used to exclude studies, is debated.30 One reviewer (SLG) quality appraised included studies using Critical Appraisal Skills Programme checklist31 to facilitate our understanding of them rather than to exclude any studies. Data categorization For the current review, we drew on content analysis to synthesize data from eligible studies. This approach permitted tabulation and frequency counts32 and thereby facilitated comparison with the outcomes reported in the previous systematic review of type 2 diabetes clinical trials. We used the Core Outcome Measures in Effectiveness Trials (COMET) taxonomy to categorize text from the studies.

eligible studies. This approach permitted tabulation and frequency counts32 and thereby facilitated comparison with the outcomes reported in the previous systematic review of type 2 diabetes clinical trials. We used the Core Outcome Measures in Effectiveness Trials (COMET) taxonomy to categorize text from the studies. Two reviewers (SLG and NLH) discussed each text excerpt relevant to outcomes to agree how to categorize them, referring back to the original article when necessary to resolve ambiguities. The COMET taxonomy is an outcome classification system suitable for classifying outcomes across all trials, COS, systematic reviews and trial registries33 regardless of the condition being investigated. It has been designed to provide high-level differentiation between outcome domains to facilitate uniformity of outcome classification in electronic databases.33 Additionally, COS developers have used the taxonomy to assist the classification of outcomes prior to the consensus stage.34–36 The taxonomy comprises 38 core domains structured within five top level core areas: death, physiological/clinical, life impact, resource use, and adverse events.

ion in electronic databases.33 Additionally, COS developers have used the taxonomy to assist the classification of outcomes prior to the consensus stage.34–36 The taxonomy comprises 38 core domains structured within five top level core areas: death, physiological/clinical, life impact, resource use, and adverse events. Reviewers categorized text excerpts, considering all 38 core domains of the taxonomy as they did so. Agreement between reviewers (SLG and NLH) was assessed with three batch checks of 10% of all outcomes until 100% agreement was reached. Where uncertainty about an outcome could not be resolved reviewers sought clinical input (JPHW). Where one outcome included multiple components, for example, ‘fear of death’, which encompasses two discrete outcomes ‘fear’ and ‘death’, the outcome was classified under two domains (eg, ‘emotional functioning/well-being’ and ‘mortality/survival’) as recommended.33 Outcome categorization was verified by the developer of the COMET taxonomy (SD), who was provided with a list of the outcomes extracted from the included studies, which she categorized blind, without seeing the categorization by the two reviewers. Her outcome categorization was compared with that of the two reviewers and any discrepancies were discussed and resolved. Results Study characteristics The search returned 146 articles. Of these, 36 were retained after screening titles and abstracts. Following full-text review a further 10 studies were excluded as they did not meet the inclusion criteria. The flow of studies is shown in figure 1.

Reviewers categorized text excerpts, considering all 38 core domains of the taxonomy as they did so. Agreement between reviewers (SLG and NLH) was assessed with three batch checks of 10% of all outcomes until 100% agreement was reached. Where uncertainty about an outcome could not be resolved reviewers sought clinical input (JPHW). Where one outcome included multiple components, for example, ‘fear of death’, which encompasses two discrete outcomes ‘fear’ and ‘death’, the outcome was classified under two domains (eg, ‘emotional functioning/well-being’ and ‘mortality/survival’) as recommended.33 Outcome categorization was verified by the developer of the COMET taxonomy (SD), who was provided with a list of the outcomes extracted from the included studies, which she categorized blind, without seeing the categorization by the two reviewers. Her outcome categorization was compared with that of the two reviewers and any discrepancies were discussed and resolved. Results Study characteristics The search returned 146 articles. Of these, 36 were retained after screening titles and abstracts. Following full-text review a further 10 studies were excluded as they did not meet the inclusion criteria. The flow of studies is shown in figure 1. Figure 1 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram.

Results Study characteristics The search returned 146 articles. Of these, 36 were retained after screening titles and abstracts. Following full-text review a further 10 studies were excluded as they did not meet the inclusion criteria. The flow of studies is shown in figure 1. Figure 1 Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) flow diagram. These 26 included studies involved either qualitative interviews (69%) or focus groups (31%) with a total of 976 patients (median 23 participants, range 5–246) from five continents. All studies included participants with type 2 diabetes with some studies focusing on particular minority groups.37–42 Time since diagnosis of diabetes ranged from less than 1 year up to 51 years, although this was not reported in all studies. For one study that also included partners of patients,38 content relating to both patients and their partners was included in the synthesis. Another study included both patients and healthcare professionals; however, only data originating from patients were synthesized.43 A summary of included studies is provided in table 2. Table 2 Characteristics of included studies Study ID Aim Year of publication Location of study Participants (n) Participant age Participant sex Time with diabetes Data collection method Outcomes identified (n) 42 Explore behavioral factors affecting what patients do for self-care and why they do it. 1998 USA 51 29–69 years (mean=52.9) Male: 26 (51%), female: 25 (49%) 0.5–1 year: 4 (8%), 1–5 years: 20 (39%), >6 years: 27 (53%) Interviews 24

Study ID Aim Year of publication Location of study Participants (n) Participant age Participant sex Time with diabetes Data collection method Outcomes identified (n) 42 Explore behavioral factors affecting what patients do for self-care and why they do it. 1998 USA 51 29–69 years (mean=52.9) Male: 26 (51%), female: 25 (49%) 0.5–1 year: 4 (8%), 1–5 years: 20 (39%), >6 years: 27 (53%) Interviews 24 44 Investigate the distress associated with type 2 diabetes. 2002 UK 51 Not reported Male: 19 (37%), female: 32 (63%) Not reported Focus groups 16 60 Explore the views and health beliefs of patients who had experienced a new structured diabetes shared care service. 2003 Ireland 25 30–50 years: 6 (24%); 51–70 years: 13 (52%); >70 years: 6 (24%) Male: 15 (60%), female: 10 (40%) Not reported Focus groups 10 58 Describe personal understandings of illness among patients. 2004 Sweden 44 47–80 years (mean=64) Male: 23 (52%), female: 21 (48%) Not reported Interviews 5 39 Explore the relevance of a reframed model of healthcare consultation. 2004 UK 21 Not reported Not reported Not reported Interviews 21 45 Investigate patients’ perceptions about their illness and treatment strategies to facilitate patient-centered, culture-sensitive clinical skills. 2005 Taiwan 22 44–80 years (mean=60.2) Male: 12 (55%), female: 10 (45%) 2–25 years (mean=8.3) Interviews 12

39 Explore the relevance of a reframed model of healthcare consultation. 2004 UK 21 Not reported Not reported Not reported Interviews 21 45 Investigate patients’ perceptions about their illness and treatment strategies to facilitate patient-centered, culture-sensitive clinical skills. 2005 Taiwan 22 44–80 years (mean=60.2) Male: 12 (55%), female: 10 (45%) 2–25 years (mean=8.3) Interviews 12 46 Explore the self-reported healthcare goals, factors influencing these goals, and self-care practices of older patients. 2005 USA 28 65–88 years (mean=74) Male: 12 (43%), female: 16 (57%) 0–5 years: 11 (39%), >5 to 10 years: 6 (21%), >10 years: 11 (39%) Interviews 15 47 Explore medication experiences of patients. 2006 USA 138 >70 years: 30 (22%), >50 years: 102 (74%) Male: 44 (32%), female: 94 (68%) Mean=13 years Focus groups 31 48 Identify the obstacles to adherence for patients. 2007 Europe 246 40–80 years (mean=63.8) Male: 122 (49.6%), female: 124 (49.4%) 1–22 years (mean=10.2) Focus groups 16 61 Describe the experience of benefit and risk assessment for patients when making treatment decisions. 2007 Canada 18 Mean=60 years Male: 8 (44%), female: 10 (55%) Mean=10.7 years Interviews 29 49 Explore the lived experience of patients converting to insulin therapy. 2007 UK 8 49–72 years Male: 4 (50%), female: 4 (50%) Not reported Interviews 25 50 Understand and document the perspectives of patients regarding the processes and strategies used to self-manage their chronic condition. 2008 Taiwan 41 42–81 years Male: 22 (54%), female: 19 (46%) Mean=9.19 years Focus groups 27

49 Explore the lived experience of patients converting to insulin therapy. 2007 UK 8 49–72 years Male: 4 (50%), female: 4 (50%) Not reported Interviews 25 50 Understand and document the perspectives of patients regarding the processes and strategies used to self-manage their chronic condition. 2008 Taiwan 41 42–81 years Male: 22 (54%), female: 19 (46%) Mean=9.19 years Focus groups 27 38 Describe cultural and family challenges to illness management in foreign-born Chinese-American patients and their spouses. 2009 USA 40 (20 patients, 20 partners) Mean=62 years Male: 16 (40%), female: 24 (60%) Mean=8.4 years Interviews 12 51 Explore how women manage their diabetes. 2009 USA 5 50–85 years (mean=70.4) Female: 5 (100%) 1–18 years (mean=8) Interviews 16 62 Explore how living with diabetes in everyday life was experienced following a self-management intervention program based on motivational interviewing. 2011 Denmark 22 30–72 years Male: 10 (45%), female: 12 (55%) 1–11 years Focus groups 10 43 Explore physicians’ and patients’ views of patients’ difficulty achieving treatment goals. 2012 USA 34 patients, 19 physicians* 43–70 years (mean=59.8) Male: 20 (59%), female: 14 (41%) 3–51 years (mean=12) Interviews 10 41 Assess self-management skills of Chinese-Americans. 2012 USA 24 (7 poorly controlled, 17 well controlled) Poorly controlled: mean=56 years, well controlled: mean=60.6 years Poorly controlled—male: 3 (43%), female: 4 (57%); well controlled—male: 13 (76%), female: 2 (12%), NA: 2 (12%) Poorly controlled: mean=6.4 years; well controlled: mean=6 years Focus groups 11

ricans. 2012 USA 24 (7 poorly controlled, 17 well controlled) Poorly controlled: mean=56 years, well controlled: mean=60.6 years Poorly controlled—male: 3 (43%), female: 4 (57%); well controlled—male: 13 (76%), female: 2 (12%), NA: 2 (12%) Poorly controlled: mean=6.4 years; well controlled: mean=6 years Focus groups 11 52 Describe the experiences and ways of coping of older Singaporean Chinese women. 2013 Singapore 10 60–69 years Female: 10 (100%) Not reported Interviews 26 53 Explore the concept of patient values in the context of making decisions about insulin initiation. 2013 Malaysia 21 28–67 years Male: 12 (57%), female: 9 (43%) Not reported Interviews 19 37 Gain a deeper understanding of the difficulties Vietnamese patients experience when accessing services and managing their diabetes. 2013 Australia 15 60 to >70 years: 15 (100%), >70 years: 11 (73%) Male: 4 (27%), female: 11 (73%) >1 year: 6, >5 years: 9 Focus groups 24 54 Better understand barriers to glycemic control from the patient’s perspective. 2013 New Zealand 15 33–90 years (mean=63.3) Male: 5 (33%), female: 10 (67%) 2–30 years (mean=44.3) Interviews 10 55 Explore the barriers to diabetes control of middle-aged women. 2013 Syria 12 40–65 years Female: 12 (100%) 4–23 years Interviews 21 40 Identify issues in self-management, and opportunities for community pharmacies to offer self-management support to these populations. 2013 Australia 24 54–95 years (mean=73) Male: 10 (42%), female: 14 (58%) <5 years: 2, 6–10 years: 8, >10 years: 14 Interviews 14

55 Explore the barriers to diabetes control of middle-aged women. 2013 Syria 12 40–65 years Female: 12 (100%) 4–23 years Interviews 21 40 Identify issues in self-management, and opportunities for community pharmacies to offer self-management support to these populations. 2013 Australia 24 54–95 years (mean=73) Male: 10 (42%), female: 14 (58%) <5 years: 2, 6–10 years: 8, >10 years: 14 Interviews 14 56 Explore patients’ reactions to the diagnosis and their health-related quality of life. 2014 Malaysia 12 50–62 years Male: 5 (42%), female: 7 (58%) 2.5–21 years Interviews 32 57 Explore the illness perceptions of patients attending treatment and better understand how they manage their illness. 2016 Ethiopia 39 >70 years: 30 Male: 20 (51%), female: 19 (49%) 1–25 years Interviews 31 59 Investigate patients’ perceptions and experiences, self-care and engagement with GP-led integrated diabetes care. 2016 Australia 30 <50 to >65 years (mean=60.2) Male: 16 (53%), female: 14 (47%) Mean=12 years Interviews 6 *Data not included in synthesis. GP, general practitioner; NA, not applicable. Quality appraisal The majority of included studies justified the research design (85%), explained details about the recruitment strategy (81%), took ethical issues into consideration (89%), provided an in-depth description of the data collection (89%) and analysis processes (69%), provided a clear statement of findings (96%), and discussed the implications of the research (81%). In contrast, only a minority of studies adequately described the relationship between the researcher and participants (39%).

%), provided an in-depth description of the data collection (89%) and analysis processes (69%), provided a clear statement of findings (96%), and discussed the implications of the research (81%). In contrast, only a minority of studies adequately described the relationship between the researcher and participants (39%). Data categorization A total of 458 individual outcomes were interpreted from the included studies (median 16 outcomes per study, range 5–32) and categorized according to the COMET taxonomy. Thirty-nine outcomes related to multiple domains and were classified under two or more domains. Thus, 501 outcomes were categorized within the 38 taxonomy domains. Table 3 lists the number of outcomes included in each of the taxonomy domains and the number of studies that included outcomes belonging to each domain. Table 3 Outcome categorization according to the COMET taxonomy

Data categorization A total of 458 individual outcomes were interpreted from the included studies (median 16 outcomes per study, range 5–32) and categorized according to the COMET taxonomy. Thirty-nine outcomes related to multiple domains and were classified under two or more domains. Thus, 501 outcomes were categorized within the 38 taxonomy domains. Table 3 lists the number of outcomes included in each of the taxonomy domains and the number of studies that included outcomes belonging to each domain. Table 3 Outcome categorization according to the COMET taxonomy Core area Core domains Studies including one or more outcomes in core domain, n (%) Outcomes included in core domain, n (%) Death Mortality/survival 10 (39) 10 (2) Physiological/clinical Blood and lymphatic system outcomes 0 0 Cardiac outcomes 5 (19) 5 (1) Congenital, familial and genetic outcomes 0 0 Endocrine outcomes 1 (4) 1 (<1) Ear and labyrinth outcomes 0 0 Eye outcomes 8 (31) 9 (2) Gastrointestinal outcomes 1 (4) 2 (<1) General outcomes* 19 (73) 42 (8) Hepatobiliary outcomes 1 (4) 1 (<1) Immune system outcomes 0 0 Infection and infestation outcomes 3 (12) 3 (1) Injury and poisoning outcomes 1 (4) 1 (<1) Metabolism and nutrition outcomes 26 (100) 63 (13) Musculoskeletal and connective tissue outcomes 1 (4) 1 (<1) Outcomes relating to neoplasms: benign, malignant and unspecified (including cysts and polyps) 0 0 Nervous system outcomes 4 (15) 4 (1) Pregnancy, puerperium and perinatal outcomes 0 0 Renal and urinary outcomes 11 (42) 13 (3) Reproductive system and breast outcomes 0 0 Psychiatric outcomes 5 (19) 7 (1) Respiratory, thoracic and mediastinal outcomes 1 (4) 1 (<1) Skin and subcutaneous tissue outcomes 0 0 Vascular outcomes 10 (39) 12 (2) Life impact Social functioning 14 (54) 28 (6) Role functioning 16 (62) 20 (4) Physical functioning 24 (92) 72 (14) Emotional functioning/well-being 24 (92) 106 (21) Cognitive functioning 12 (46) 15 (3) Global quality of life 3 (12) 3 (1) Perceived health status 5 (19) 5 (1) Delivery of care 18 (69) 43 (9) Personal circumstance 7 (27) 12 (2) Resource use Economic 0 0 Hospital 0 0 Need for intervention 9 (35) 10 (2) Societal/carer burden 2 (8) 2 (<1) Adverse events Adverse events/effects 8 (31) 10 (2) *The COMET taxonomy defines ‘general outcomes’ to include those affecting the whole body, which cannot be attributed to a certain body system, for example, fatigue, malaise, pain (unspecified, not associated with a particular body system), fever (not attributable to infection), anthropometric measures (eg, weight), ‘global’ measures, ‘symptoms’ (not associated with a particular body system), ‘physical health’ and fitness.45

ttributed to a certain body system, for example, fatigue, malaise, pain (unspecified, not associated with a particular body system), fever (not attributable to infection), anthropometric measures (eg, weight), ‘global’ measures, ‘symptoms’ (not associated with a particular body system), ‘physical health’ and fitness.45 COMET, Core Outcome Measures in Effectiveness Trials. Of the 501 outcomes, 10 (2%) concerned death, 165 (33%) were physiological/clinical, 304 (61%) were associated with life impact, 12 (2%) related to resource use and 10 (2%) pertained to the adverse events core area. Most outcomes fell within the core domains of ‘emotional functioning/well-being’, ‘physical functioning’ and ‘metabolism and nutrition’. Outcomes relating to each of these three domains were identified in more than 90% of included studies and, when combined, comprised 48% of the total outcomes.

dverse events core area. Most outcomes fell within the core domains of ‘emotional functioning/well-being’, ‘physical functioning’ and ‘metabolism and nutrition’. Outcomes relating to each of these three domains were identified in more than 90% of included studies and, when combined, comprised 48% of the total outcomes. Emotional functioning/well-being Over one-fifth of the derived outcomes related to ‘emotional functioning/well-being’ (n=106). These outcomes were identified in 24 of the 26 included studies (92%). In 16 studies, patients described being fearful,41 42 44–57 with fears relating to medication side effects, treatment escalation, needing insulin injections, dying, uncertainty about the future, and developing complications, such as foot damage, paralysis and loss of eyesight. Relatedly, 10 studies39 42 47–50 52 55–57 described patients feeling worried or anxious about symptoms, complications, health deterioration, and ultimately dying prematurely as a result of their diabetes. Patients also commented on how diabetes was a burden44 47 58 59 and reported experiencing aggression and frustration,38 39 43 47 49 51 56 57 59–61 sadness and depression,37–39 43 47 52 56 59 guilt43 44 54 and hopelessness.43 44 52 54

health deterioration, and ultimately dying prematurely as a result of their diabetes. Patients also commented on how diabetes was a burden44 47 58 59 and reported experiencing aggression and frustration,38 39 43 47 49 51 56 57 59–61 sadness and depression,37–39 43 47 52 56 59 guilt43 44 54 and hopelessness.43 44 52 54 Physical functioning A total of 72 (14%) outcomes concerning physical functioning were derived from 24 included studies (92%). Exercise was identified as an outcome in 16 studies, with patients acknowledging the importance of regular exercise and the benefits it brings; however, some patients labelled it a burdensome activity, which they had difficulty engaging in.37 39 40 42–46 48 50 52 53 56–58 61 In 15 studies, patients referred to the self-monitoring activities they engaged in to manage their diabetes, for example, regularly checking blood glucose levels,37–41 44 47 49 50 52 55–57 61 62 with many emphasizing that the need for a strict self-management regime had become a burden on their lives. Dietary restrictions were a common difficulty relating to self-management, with many patients articulating a desire for dietary freedom, where they could eat what they want, when they want.37 38 44 46 49 57 61

62 with many emphasizing that the need for a strict self-management regime had become a burden on their lives. Dietary restrictions were a common difficulty relating to self-management, with many patients articulating a desire for dietary freedom, where they could eat what they want, when they want.37 38 44 46 49 57 61 Metabolism and nutrition Outcomes relating to metabolism and nutrition were identified in all of the 26 included studies, with 63 (13%) outcomes identified in total. In 20 studies, patients made references to their diet, explaining how healthy eating was necessary for controlling their blood sugar levels.37 39–43 46 48 50–56 58–62 In 17 studies, patients spoke about blood glucose-level fluctuations, the importance of glycemic control and the consequences of blood glucose levels falling outside the appropriate range.38 39 41–43 45–47 49–51 53 55–57 60 62 Relatedly, 12 studies referred to hypoglycemia, including patients’ concerns over what would happen if they did experience a hypoglycemic episode, fears about the physical symptoms and the steps they would take to avoid hypoglycemia.37 39 41 44 47 49 52–54 56 57 61

propriate range.38 39 41–43 45–47 49–51 53 55–57 60 62 Relatedly, 12 studies referred to hypoglycemia, including patients’ concerns over what would happen if they did experience a hypoglycemic episode, fears about the physical symptoms and the steps they would take to avoid hypoglycemia.37 39 41 44 47 49 52–54 56 57 61 Outcomes identified from studies conducted in LMICs Of the 26 included studies, four (15%) were conducted in LMICs.63 These four LMIC studies included one upper middle-income country (Malaysia),53 56 one lower middle-income country (Syria)55 and one least developed country (Ethiopia).57 The most prevalent outcome domains among the LMIC studies were ‘mortality/survival’, ‘general outcomes’, ‘metabolism and nutrition’, ‘role functioning’, ‘physical functioning’, ‘emotional functioning/well-being’ and ‘delivery of care’. Outcomes relating to these seven domains were derived from 100% of the LMIC studies. Five of these seven domains were also among the most reported in the studies conducted in HICs; however, outcomes associated with the ‘mortality/survival’ and ‘role functioning’ domains were less frequently reported in HIC studies, at 27% and 55% of studies, respectively. Additionally, outcomes relating to two of the domains, ‘endocrine outcomes’ (eg, pancreatic function) and ‘hepatobiliary outcomes’ (eg, liver complications), were only derived from the LMIC studies.

al’ and ‘role functioning’ domains were less frequently reported in HIC studies, at 27% and 55% of studies, respectively. Additionally, outcomes relating to two of the domains, ‘endocrine outcomes’ (eg, pancreatic function) and ‘hepatobiliary outcomes’ (eg, liver complications), were only derived from the LMIC studies. Comparison with outcomes identified in systematic review of clinical trials We compared the outcomes derived from the qualitative studies with the outcomes identified from the recent systematic review of type 2 diabetes registered clinical trials.2 Table 4 shows the number of outcomes identified from both reviews according to the five core areas of the COMET taxonomy. In total, 1446 outcomes were identified from the clinical trials review and 458 outcomes from the review of qualitative studies. Both reviews identified a similar proportion of outcomes related to death, resource use and adverse events. However, the systematic review of clinical trials identified a higher proportion of physiological/clinical outcomes (84% vs 33%), whereas the qualitative studies identified a greater proportion of life impact (61% vs 10%) outcomes. Several domains were only identified in one or other of the reviews. Outcomes relating to the ‘blood and lymphatic system’, ‘immune system’, ‘skin and subcutaneous tissue’, ‘economic resource use’ and ‘hospital resource use’ were only extracted from the clinical trials, whereas outcomes associated with ‘injury and poisoning’ (eg, injuries associated with insulin injections), ‘personal circumstances’ (eg, patients’ support networks) and ‘societal/carer burden’ (eg, patients wanting to be independent and not wanting to be a burden to their family) were only identified in the qualitative studies.

utcomes associated with ‘injury and poisoning’ (eg, injuries associated with insulin injections), ‘personal circumstances’ (eg, patients’ support networks) and ‘societal/carer burden’ (eg, patients wanting to be independent and not wanting to be a burden to their family) were only identified in the qualitative studies. Table 4 Number of outcomes identified in the systematic review of clinical trials and synthesis of qualitative literature according to the five core areas within the COMET taxonomy Outcomes identified in systematic review of clinical trials, n (%) Outcomes identified in synthesis of qualitative literature, n (%) Death 3 (<1) 10 (2) Physiological/clinical 1221 (84) 165 (33) Life impact 145 (10) 304 (61) Resource use 31 (2) 12 (2) Adverse events 46 (3) 10 (2) COMET, Core Outcome Measures in Effectiveness Trials. Discussion This review has identified an expedited approach for incorporating patient perspectives within the early stages of COS development. In contrast to previous similar reviews, which have involved exhaustive searches,19 20 the streamlined nature of the current review enabled rapid identification of patient-centered outcomes that can be used to contribute to the development of the ‘long list’ of outcomes to inform the COS consensus process.

development. In contrast to previous similar reviews, which have involved exhaustive searches,19 20 the streamlined nature of the current review enabled rapid identification of patient-centered outcomes that can be used to contribute to the development of the ‘long list’ of outcomes to inform the COS consensus process. To our knowledge, this is the first review of qualitative studies to identify outcomes that are important to people with type 2 diabetes. The findings will be used, alongside a review of outcomes measured in clinical trials,2 in the development of a COS for type 2 diabetes. Importantly, this review of qualitative studies has identified outcomes that have not previously been measured in type 2 diabetes clinical trials. Without this review, these outcomes would not have been identified for inclusion in the ‘long list’ of outcomes to go forward to the Delphi study.

a COS for type 2 diabetes. Importantly, this review of qualitative studies has identified outcomes that have not previously been measured in type 2 diabetes clinical trials. Without this review, these outcomes would not have been identified for inclusion in the ‘long list’ of outcomes to go forward to the Delphi study. Most outcomes identified from the qualitative studies related to life impact, whereas in the review of registered clinical trials a relatively small proportion of outcomes related to life impact.2 If clinical trial reviews are used as the only source for developing ‘long list’ of outcomes for COS consensus processes, life impact outcomes may become sidelined in favor of outcomes more frequently measured in clinical trials. This is reflected in the COMET database, where far fewer COS encompass life impact outcomes, in contrast to many COS that encompass physiological/clinical outcomes.33 Thus, the current review supports the recommendation by Dodd and colleagues33 for COS developers to give greater attention to outcomes of life impact. It also illustrates how conducting reviews of qualitative studies can help, alongside other steps such as including patients as participants in COS studies and involving patients and the public in the design of such studies, to ensure COS reflect outcomes that matter to patients.

r attention to outcomes of life impact. It also illustrates how conducting reviews of qualitative studies can help, alongside other steps such as including patients as participants in COS studies and involving patients and the public in the design of such studies, to ensure COS reflect outcomes that matter to patients. Given uncertainties about how far COS are applicable beyond those countries that the participants in the development process have been drawn from, it is striking that 84% of published COS studies have not included any participants from LMICs.64 While few qualitative studies had been conducted in LMICs, our review has enabled us to compare outcomes identified in LMIC and HIC studies. Outcomes relating to ‘mortality/survival’ were identified in all LMIC studies, yet almost three-quarters of HIC studies made no references to these outcomes. This is most likely due to the higher prevalence of diabetes-related deaths in LMICs.25 Similarly, all LMIC studies reported outcomes relating to ‘role functioning’ (eg, ability to work and managing family responsibilities) while almost half of the HIC studies made no references to these outcomes. Being unable to function in one’s life roles is likely to be more detrimental to patients living in LMICs.65 Furthermore, three domains relating to diabetes-related complications were only reported in the LMIC studies. Complications are expensive to treat in LMICs, which represent less than 20% of the world’s diabetes care-related expenditure.66 It is possible that national economic factors influence which outcomes are important to patients, indicating the importance of ensuring that the perspectives of patients in LMICs are incorporated into COS development processes.

in LMICs, which represent less than 20% of the world’s diabetes care-related expenditure.66 It is possible that national economic factors influence which outcomes are important to patients, indicating the importance of ensuring that the perspectives of patients in LMICs are incorporated into COS development processes. Strengths and limitations This study has identified an expedited approach for incorporating the patient perspective into the initial stages of COS development. Our review has enabled evidence from a large number of patients living in a diversity of countries to contribute to the early stages of COS development. We anticipate that such reviews could be performed by COS developers without specialist expertise in qualitative methods, although relevant training would be helpful. Additionally, the method is less resource intensive than other methods for reviewing qualitative evidence which makes it feasible as part of a wider COS process. However, there are some methodological limitations. In line with our expedited approach and recommendations from previous studies,67 68 we only searched one database (MEDLINE) and did not search gray literature. While this approach may mean that some relevant studies may have been missed, our aim was not to provide a comprehensive overview of all research relating to patients’ views and perceptions of type 2 diabetes. Rather, to provide an expedited approach for identifying outcomes that are important to patients, as part of a wider process to incorporate the patient perspective in COS development. In comparison to other qualitative systematic reviews we identified few articles for screening, which reflects the specificity27 of our search terms. Despite these limitations, the number of studies included in this rapid review is comparable to other reviews of qualitative evidence.19 20

erspective in COS development. In comparison to other qualitative systematic reviews we identified few articles for screening, which reflects the specificity27 of our search terms. Despite these limitations, the number of studies included in this rapid review is comparable to other reviews of qualitative evidence.19 20 A further limitation is that our inclusion of studies relating to a specific experience (eg, patients’ reactions to their diagnosis)56 or intervention (eg, the views of patients who had experienced a new structured diabetes shared care service)60 may have impacted on the outcomes identified. However, when extracting data from such studies we focused on patients’ general views and experiences of diabetes; we did not extract data that focused solely on patients’ perspectives of specific experiences or interventions. An additional limitation is that while many studies in our review were inductive, our approach to reviewing them and deriving outcomes was deductive. Specifically, categorizing text excerpts according to the COMET taxonomy may have transformed their meaning or diluted the patient perspective. However, this categorization enabled us to compare the different literatures in a common ‘currency’, which was key to identifying differences in the types of outcomes given prominence in the clinical trials versus the qualitative literatures, and differences in the qualitative studies from LMICs and HICs. Finally, our review was largely aggregative, collecting findings of previous studies and describing these according to a predefined taxonomy to allow comparison and address specific aims, rather than a configurative review which seeks to interpret the experiences of patients or generate new theory about these.69 Nevertheless, we hope the pragmatic and resource-efficient nature of our method helps the field of COS development by making it easier to incorporate patients’ perspectives.

dress specific aims, rather than a configurative review which seeks to interpret the experiences of patients or generate new theory about these.69 Nevertheless, we hope the pragmatic and resource-efficient nature of our method helps the field of COS development by making it easier to incorporate patients’ perspectives. Conclusion This rapid review of qualitative studies identified outcomes that are important to people with type 2 diabetes and its findings will inform the development of a COS for clinical trials of glucose-lowering interventions in people with type 2 diabetes. These patient-derived outcomes contrasted with those identified from a systematic review of clinical trials, pointing to the importance of incorporating patient perspectives from the outset of COS development. Additionally, this review also emphasized the importance of ensuring that patients in LMICs are able to input into the development of COS. The authors thank Susanna Dodd for verifying the COMET taxonomy outcome classifications during the data categorization stage of the review. Contributors: SLG: investigation, methodology, writing–original draft preparation, writing–review and editing. BY, JPHW: investigation, writing–review and editing. PRW: conceptualization, funding acquisition, supervision, writing–review and editing. NLH: conceptualization, investigation, project administration, writing–original draft preparation, writing–review and editing.

riginal draft preparation, writing–review and editing. BY, JPHW: investigation, writing–review and editing. PRW: conceptualization, funding acquisition, supervision, writing–review and editing. NLH: conceptualization, investigation, project administration, writing–original draft preparation, writing–review and editing. Funding: This work was supported by the Medical Research Council North West Hub for Trials Methodology Research grant number MR/K025635/1 and the European Union's Horizon 2020 research and innovation program (CORBEL, under grant agreement number 654248). Competing interests: None declared. Patient consent for publication: Not required. Provenance and peer review: Not commissioned; externally peer reviewed. Data sharing statement: All the findings from available data have been published in this manuscript. Original extracted data are available from the corresponding author, NLH, on reasonable request.

Significance of this study What is already known about this subject? High refill adherence to lipid-lowering medications associates with lower risk of cardiovascular event and mortality among patients with type 2 diabetes mellitus; the impact of the treating healthcare centers’ adherence to lipid-lowering guidelines is uncertain. What are the new findings? We found patients’ refill adherence to lipid-lowering medications to have greater impact than adherence to guidelines by the treating healthcare center in terms of prevention of cardiovascular event and mortality in both primary and secondary prevention patients. How might these results change the focus of research or clinical practice? Our results emphasize the value of individualized care and the importance of maximizing refill adherence to lipid-lowering medications among patients with type 2 diabetes mellitus.

What are the new findings? We found patients’ refill adherence to lipid-lowering medications to have greater impact than adherence to guidelines by the treating healthcare center in terms of prevention of cardiovascular event and mortality in both primary and secondary prevention patients. How might these results change the focus of research or clinical practice? Our results emphasize the value of individualized care and the importance of maximizing refill adherence to lipid-lowering medications among patients with type 2 diabetes mellitus. Introduction Type 2 diabetes mellitus (T2DM) is a multifactorial disease that requires intensive glycemic control and treatment for comorbidities and complications to reduce the increased risk of cardiovascular disease (CVD) and mortality.1–3 Despite declining mortality rates among Swedish patients with T2DM, the risk is still higher than among patients without diabetes mellitus, and CVD remains the major cause of death.4 Improved control of low-density lipoprotein (LDL) cholesterol with lipid-lowering medications has been shown to reduce the risk of CVD and mortality among patients with T2DM.5–8 Thus, national9 and international1 10 treatment guidelines recommended lipid-lowering medications to patients with T2DM to lower LDL cholesterol below 2.5 mmol/L. For patients with established CVD, LDL cholesterol below 1.8 mmol/L is desirable.

shown to reduce the risk of CVD and mortality among patients with T2DM.5–8 Thus, national9 and international1 10 treatment guidelines recommended lipid-lowering medications to patients with T2DM to lower LDL cholesterol below 2.5 mmol/L. For patients with established CVD, LDL cholesterol below 1.8 mmol/L is desirable. Adherence is defined as the extent to which individuals follow agreed recommendations.11 Previous studies have shown that patients with T2DM with adherence of at least 80% to lipid-lowering medications face a lower risk of CVD and mortality than those with adherence below 80%.5 6 8 12 13 Among multiple levels of adherence to lipid-lowering medications, a gradual increase in CVD risk was observed with declining adherence by patients with T2DM.6 Previous studies have reported adherence by healthcare providers to recommended treatment guidelines at 24%–80% among patients with diabetes mellitus.14–19 Patients with T2DM or established CVD were more likely to receive lipid-lowering medications. However, little is known about the impact of healthcare provider adherence to lipid-lowering medications on the risk of cardiovascular (CV) events or mortality among patients with T2DM. Our primary objective was to analyze the risk of CV events and mortality in relation to T2DM patients' adherence to lipid-lowering medications and providers' adherence to lipid-lowering guidelines. Our secondary objective was to identify differences in the risk of CV events and mortality among various patient characteristics.

mary objective was to analyze the risk of CV events and mortality in relation to T2DM patients' adherence to lipid-lowering medications and providers' adherence to lipid-lowering guidelines. Our secondary objective was to identify differences in the risk of CV events and mortality among various patient characteristics. Methods Data sources The unique Swedish personal identity number was used to link individual-based data from six national registers. Filled prescriptions were collected from the Swedish Prescribed Drug Register (SPDR), which contains information about all prescriptions filled at Swedish pharmacies since July 2005.20 Swedish prescriptions include information about the medication, patient, prescriber and the prescribed daily dosage. Clinical and health-related data about risk factors and complications of diabetes, as well as healthcare center characteristics, were collected from the Swedish National Diabetes Register (NDR), which contains patient information reported by physicians and nurses at hospitals and primary care clinics nationwide.21 Healthcare providers were identified at the center level, not by individual practitioners. Data concerning CV events were collected from the Swedish National Patient Register, which contains information about inpatient and outpatient care in Sweden.22 The date and cause of death were collected from the Cause of Death Register.23 Information about primary tumors was collected from the Swedish Cancer Registry.24 Individual data about socioeconomic status were collected from the Longitudinal Integration Database for Health Insurance and Labour Market Studies (LISA).25

.71 to 1.88) 1.85 (1.69 to 2.03) 1.17 (1.10 to 1.25) 1.72 (1.59 to 1.87) 1.74 (1.55 to 1.96) 1–2 times per week 1.10 (1.05 to 1.15) 1.26 (1.20 to 1.33) 1.30 (1.17 to 1.44) 1.05 (0.98 to 1.12) 1.18 (1.07 to 1.29) 1.19 (1.04 to 1.38) 3–5 times per week 0.99 (0.94 to 1.03) 1.04 (0.98 to 1.10) 0.98 (0.88 to 1.10) 1.00 (0.9 3 to 1.07) 1.05 (0.94 to 1.16) 1.04 (0.89 to 1.21) Smoking¶, ref=no Yes 1.16 (1.11 to 1.22) 1.38 (1.31 to 1.44) 1.38 (1.26 to 1.51) 1.07 (1.00 to 1.14) 1.17 (1.07 to 1.28) 1.23 (1.08 to 1.41) Adjusted HR (with 95% CI) for any CV events, all-cause mortality and CV mortality for patient characteristics among patients with type 2 diabetes mellitus by prevention type. *Includes Belgium, Denmark, Germany, Greece, Spain, France, Ireland, Italy, Luxembourg, The Netherlands, Austria, Portugal, Finland, Sweden and Great Britain. †If age 65 years or older and unemployed. ‡Within 5 years prior to baseline. §30 min walk or equivalent. ¶At least one cigarette or pipe per day or quit smoking within 3 months. BMI, body mass index;CV, cardiovascular;HDL, high-density lipoprotein;HR, hazard ratio;HbA1c, hemoglobin A1c;IQR, interquartile range;LDL, low-density lipoprotein;NA, not applicable;SD, standard deviation;TSEK, thousand Swedish Krona; eGFR, estimated glomerular filtration rate; ref, reference.

The date and cause of death were collected from the Cause of Death Register.23 Information about primary tumors was collected from the Swedish Cancer Registry.24 Individual data about socioeconomic status were collected from the Longitudinal Integration Database for Health Insurance and Labour Market Studies (LISA).25 Study population and period Patients age 18 years or older with a clinical T2DM diagnosis26 who had filled at least one prescription for lipid-lowering medications between July 1, 2006 and December 31, 2012 were eligible for inclusion (figure 1). Prescriptions for bile acid sequestrants were excluded since the indication is typically not hyperlipidemia.27 The date of the first filled prescription was defined as the index date. To establish a new-user design, patients who had filled prescriptions for lipid-lowering medications within 1 year prior to the index date were regarded as prevalent users and excluded. Patients who filled prescriptions for a lipid-lowering combination therapy or lipid-lowering extemporaneous preparation that lacked information about package size were also excluded. Combination therapy has been described elsewhere.6 28 Patients who experienced CVD prior to or on the index date were classified as prescribed lipid-lowering secondary prevention; all other patients were defined as primary prevention. Figure 1 Inclusion and exclusion of the study population. CVD, cardiovascular disease; NDR, National Diabetes Register.

Study population and period Patients age 18 years or older with a clinical T2DM diagnosis26 who had filled at least one prescription for lipid-lowering medications between July 1, 2006 and December 31, 2012 were eligible for inclusion (figure 1). Prescriptions for bile acid sequestrants were excluded since the indication is typically not hyperlipidemia.27 The date of the first filled prescription was defined as the index date. To establish a new-user design, patients who had filled prescriptions for lipid-lowering medications within 1 year prior to the index date were regarded as prevalent users and excluded. Patients who filled prescriptions for a lipid-lowering combination therapy or lipid-lowering extemporaneous preparation that lacked information about package size were also excluded. Combination therapy has been described elsewhere.6 28 Patients who experienced CVD prior to or on the index date were classified as prescribed lipid-lowering secondary prevention; all other patients were defined as primary prevention. Figure 1 Inclusion and exclusion of the study population. CVD, cardiovascular disease; NDR, National Diabetes Register. Patients were followed from cessation of the first filled supply (baseline date) until the first filled prescription for multi-dose dispensed medications, migration, CV event, death or December 31, 2016. The study period was broken down into subsequent intervals of 122 days until December 31, 2014, followed by annual intervals until December 31, 2016.

tion of the first filled supply (baseline date) until the first filled prescription for multi-dose dispensed medications, migration, CV event, death or December 31, 2016. The study period was broken down into subsequent intervals of 122 days until December 31, 2014, followed by annual intervals until December 31, 2016. The healthcare center at baseline was assigned by selecting the entry in the NDR closest to the baseline date and remained the same throughout the study period unless the NDR indicated otherwise. A comparison between the assigned healthcare center from the NDR with the healthcare provider of the first filled lipid-lowering prescription in the SPDR showed agreement of 98% for county council and 87% for type of care. Refill adherence Patients’ (refill) adherence to lipid-lowering medications was assessed with data from the SPDR. We assumed that patients initiated medication use on the index date. We measured refill adherence with the medication possession ratio (MPR), representing the proportion of days with medications available. The supply of each prescription was calculated by dividing the number of tablets filled by the daily dosage stated as a free-text variable by the prescriber. Interpretation of the free-text variable to obtain the daily dosage has been described elsewhere.6 28 Overlapping supplies for the same substance and strength were added to the latter supply. Overlapping supplies for different substances or strengths were deleted. MPR was assessed for each subsequent interval and categorized as high or low based on the common cut-off of 80%.29 30

osage has been described elsewhere.6 28 Overlapping supplies for the same substance and strength were added to the latter supply. Overlapping supplies for different substances or strengths were deleted. MPR was assessed for each subsequent interval and categorized as high or low based on the common cut-off of 80%.29 30 Guideline adherence Based on data from the NDR, we assessed healthcare center adherence to national lipid-lowering prescription guidelines. Guideline adherence was defined as the prescription prevalence of lipid-lowering medications among patients with T2DM and LDL cholesterol above the recommended target levels existing at the time for the study (≥2.5 mmol/L for primary prevention31 and ≥1.8 mmol/L for secondary prevention1 10). Between 2007 and 2014, guideline adherence was assessed for each healthcare center and year, for primary and secondary prevention patients separately. Guideline adherence was linked to patient intervals based on the year in which the interval started. For intervals starting in 2006, guideline adherence for 2007 was used. For healthcare centers that lacked information about adherence, we imputed the preceding year’s figure or the mean annual adherence for the county council and type of care. Guideline adherence was categorized as high or low based on a cut-off that represented the median for primary (48%) and secondary prevention (78%).

For healthcare centers that lacked information about adherence, we imputed the preceding year’s figure or the mean annual adherence for the county council and type of care. Guideline adherence was categorized as high or low based on a cut-off that represented the median for primary (48%) and secondary prevention (78%). Outcomes The outcomes of interest were CV events and mortality. A CV event was defined as a diagnosis of unstable angina pectoris, myocardial infarction (including percutaneous coronary intervention, coronary artery bypass grafting), stroke or ischemic heart disease. All-cause mortality was defined as death from any cause. CV mortality was defined as death from a cause of CVD or a CV event entered in the National Patient Register within 28 days prior to death. Starting from the second interval, the risk of CV events and mortality was analyzed for each interval until migration, CV event, death or December 31, 2016.

s death from any cause. CV mortality was defined as death from a cause of CVD or a CV event entered in the National Patient Register within 28 days prior to death. Starting from the second interval, the risk of CV events and mortality was analyzed for each interval until migration, CV event, death or December 31, 2016. Covariates Covariates regarded as potential confounders included sex, age, socioeconomic status (country of birth, marital status, education level, employment status, profession, income), concurrent medications (filled prescriptions for diabetes medications, anticoagulants and antihypertensives), and clinical and health-related characteristics (diabetes duration, hemoglobin A1c [HbA1c], estimated glomerular filtration rate [eGFR], blood pressure, cholesterol values, microalbuminuria, macroalbuminuria, kidney disease, cancer, body mass index [BMI], physical activity and smoking). These covariates have previously been shown to be important factors when analyzing adherence, as well as the risk of CV events and mortality.6 28

filtration rate [eGFR], blood pressure, cholesterol values, microalbuminuria, macroalbuminuria, kidney disease, cancer, body mass index [BMI], physical activity and smoking). These covariates have previously been shown to be important factors when analyzing adherence, as well as the risk of CV events and mortality.6 28 Sex, age and socioeconomic characteristics were collected from the LISA database. Sex and country of birth were regarded as constant, and age was based on the year of birth. Information about marital status, education level, employment status, profession and income was collected prior to or during the baseline year. Income was regarded as a continuous variable. The remaining socioeconomic variables included the following categories: country of birth: Sweden, other Nordic country, other European Union (EU)-15 country or the Soviet Union, rest of Europe, the Americas, Asia or Oceania, or unknown; marital status: unmarried, married or registered partner, divorced, or widow/widower; education level: compulsory school or lower, upper secondary school, or postsecondary; employment status: unemployed, employed or retired; profession: upper white-collar, lower white-collar, blue-collar, or other. At baseline, cancer and kidney disease were defined as diagnosis within 5 years prior to the baseline date and were collected from the Swedish Cancer Registry and the National Patient Register, respectively. Cancer diagnosis included primary tumors, while kidney disease included acute or chronic kidney failure, as well as glomerular or renal complication due to T2DM.

ed as diagnosis within 5 years prior to the baseline date and were collected from the Swedish Cancer Registry and the National Patient Register, respectively. Cancer diagnosis included primary tumors, while kidney disease included acute or chronic kidney failure, as well as glomerular or renal complication due to T2DM. Filled prescriptions for diabetes medications (anatomical therapeutic chemical (ATC): A10), anticoagulants (ATC: B01), and antihypertensives (ATC: C02, C03, C04, C05, C07, C08, and C09) were collected from the SPDR. Filled prescriptions within 12 months prior to the baseline date were considered concurrent use. The remaining clinical and health-related characteristics were collected from the NDR. At baseline, data were collected between 24 months prior to and 14 days after the baseline date by selecting the value closest to baseline. Diabetes duration was based on the year of birth and diabetes diagnosis. Microalbuminuria and macroalbuminuria were dichotomously categorized. BMI and eGFR were categorized according to recommended references values.32 33 HbA1c, blood pressure and cholesterol levels were categorized as high or low according to recommended target values at the time of the study.31 Physical activity was defined as a 30 min walk or equivalent, categorized as less than once a week, 1–2 times a week, 3–5 times a week or daily. Smoking was dichotomized and defined as at least one cigarette/pipe daily or having quit within the past 3 months.

ing to recommended target values at the time of the study.31 Physical activity was defined as a 30 min walk or equivalent, categorized as less than once a week, 1–2 times a week, 3–5 times a week or daily. Smoking was dichotomized and defined as at least one cigarette/pipe daily or having quit within the past 3 months. A total of 23% of patient characteristics at baseline were missing: 4% of socioeconomic status and 43% of clinical and health-related characteristics. No information was missing about age, sex or concurrent medications. Missing information at baseline was replaced with multiple imputations. Potential confounders (except for cancer) were assessed for each interval during the study period by assuming the imputed baseline value until the information had been updated in the registers. Sensitivity analyses To evaluate the cut-offs used to categorize refill and guideline adherence as high or low, new ones were set. For refill adherence, cut-offs of 60%, 70%, and 90% were applied to evaluate the 80% level that had been used to define patients as low-adherent or high-adherent. For guideline adherence, new cut-offs were set at 30% and 60% (from 48%) for primary prevention and 50% and 90% (from 78%) for secondary prevention. The results of the statistical analyses were compared between the cut-offs. To estimate the impact of multiple imputations, the risk of CVD and mortality was assessed and compared between complete cases and imputed data.

Sensitivity analyses To evaluate the cut-offs used to categorize refill and guideline adherence as high or low, new ones were set. For refill adherence, cut-offs of 60%, 70%, and 90% were applied to evaluate the 80% level that had been used to define patients as low-adherent or high-adherent. For guideline adherence, new cut-offs were set at 30% and 60% (from 48%) for primary prevention and 50% and 90% (from 78%) for secondary prevention. The results of the statistical analyses were compared between the cut-offs. To estimate the impact of multiple imputations, the risk of CVD and mortality was assessed and compared between complete cases and imputed data. Statistical analyses The association between refill (MPR) and guideline adherence was analyzed by means of general linear regression. Multivariate imputations by chained equations (MICE) were used to replace missing information among baseline variables; 10 imputed data sets with 10 iterations each were generated. The risk of CV events and mortality was analyzed for each interval based on the 10 imputed data sets with Cox proportional hazard regression and Kaplan-Meier, adjusted for potential confounders. Covariates and guideline adherence for one interval were regarded as potential confounders for the subsequent interval of MPR measures. Similarly, MPR for one interval was regarded as the exposure for the subsequent interval of outcome measures (online supplementary figure S1). HRs generated for each imputed data set were pooled and one final set was established. The same procedure was performed to assess adjusted and pooled survival estimates and obtain Kaplan-Meier survival curves for CV events and mortality. All hypothesis tests were evaluated at a 5% significance level.

entary figure S1). HRs generated for each imputed data set were pooled and one final set was established. The same procedure was performed to assess adjusted and pooled survival estimates and obtain Kaplan-Meier survival curves for CV events and mortality. All hypothesis tests were evaluated at a 5% significance level. To present baseline characteristics for the study population based on imputed data, the mean values of continuous variables and the most frequent category of categorical variables were obtained from the imputed data sets. These generated baseline values were used to descriptively present the study population but were not used in the statistical analyses. Multiple imputations were performed in R V.3.3.234 using the MICE package.35 All other data management and statistical analyses were performed with SAS V.9.4 software. Results Study population and period A total of 121 914 patients with T2DM were included (figure 1). Of them, a total of 11.8% started on lipid-lowering secondary prevention. The mean age was 63 years, 57% were men and the mean diabetes duration was 5 years (table 1). Approximately 80% were born in Sweden and more than 50% were married. Among primary prevention patients, the mean age was 62 years, 56% were men and the mean diabetes duration was 5 years. Among secondary prevention patients, the mean age was 70 years, 61% were men and the mean diabetes duration was 6.5 years. Table 1 Baseline characteristics (imputed data) of all 121 914 new users of lipid-lowering medications with type 2 diabetes mellitus and by prevention type

Results Study population and period A total of 121 914 patients with T2DM were included (figure 1). Of them, a total of 11.8% started on lipid-lowering secondary prevention. The mean age was 63 years, 57% were men and the mean diabetes duration was 5 years (table 1). Approximately 80% were born in Sweden and more than 50% were married. Among primary prevention patients, the mean age was 62 years, 56% were men and the mean diabetes duration was 5 years. Among secondary prevention patients, the mean age was 70 years, 61% were men and the mean diabetes duration was 6.5 years. Table 1 Baseline characteristics (imputed data) of all 121 914 new users of lipid-lowering medications with type 2 diabetes mellitus and by prevention type Total population, n=121 914 Primary prevention, n=107 587 Secondary prevention, n=14 327 n (%) n (%) n (%) Demographic and socioeconomic status Sex Male 68 828 (56.5) 60 043 (55.8) 8785 (61.3) Age (years) 18–40 3230 (2.7) 3187 (3.0) 43 (0.3) 41–60 42 954 (35.2) 40 466 (37.7) 2388 (16.7) 61–80 68 893 (56.5) 59 406 (55.2) 9487 (66.2) >80 6837 (5.6) 4428 (4.1) 2409 (16.8) Mean (SD) 63.3 (11.2) 62.4 (11.1) 70.3 (10.2) Median (IQR) 64.0 (15.0) 63.0 (15.0) 71.0 (14.0) Country of birth Sweden 96 174 (78.9) 84 645 (78.7) 11 529 (80.5) Other Nordic country 6761 (5.6) 5786 (5.4) 975 (6.8) Other European Union-15* country 2097 (1.7) 1827 (1.7) 270 (1.9) Other European country/Soviet Union 5874 (4.8) 5143 (4.8) 731 (5.1) Africa 1850 (1.5) 1748 (1.6) 102 (0.7) The Americas 1285 (1.1) 1171 (1.1) 114 (0.8) Asia/Oceania 7715 (6.3) 7120 (6.6) 595 (4.2) Unknown 158 (0.1) 147 (0.1) 11 (0.1) Marital status Unmarried 20 024 (16.4) 18 398 (17.1) 1626 (11.4) Married/Registered partner 66 760 (54.8) 59 306 (55.1) 7454 (52.0) Divorced 22 258 (18.3) 19 502 (18.1) 2756 (19.2) Widow/Widower 12 872 (10.6) 10 381 (9.7) 2491 (17.4) Education level Compulsory school or lower 46 718 (38.3) 39 773 (37.0) 6945 (48.5) Upper secondary school 53 717 (44.1) 48 222 (44.8) 5495 (38.4) Postsecondary 21 479 (17.6) 19 592 (18.2) 1887 (13.2) Employment status Unemployed 17 661 (14.5) 15 990 (14.9) 1671 (11.7) Employed 59 290 (48.6) 55 232 (51.3) 4058 (28.3) Retired† 44 963 (36.9) 36 365 (33.8) 8598 (60.0) Profession Upper white-collar 32 535 (26.7) 29 145 (27.1) 3390 (23.7) Lower white-collar 10 232 (8.4) 9207 (8.6) 1025 (7.2) Blue-collar 75 708 (62.1) 66 308 (61.6) 9400 (65.6) Others 3439 (2.8) 2927 (2.7) 512 (3.6) Income (thousand Swedish krona) Per household member, mean (SD) 214.3 (429.2) 216.9 (422.0) 195.3 (478.8) Per household member, median (IQR) 180.0 (117.0) 185.0 (120.0) 156.5 (84.9) Concurrent medications Diabetes medications Any 80 889 (66.4) 72 126 (67.0) 8763 (61.2) Anticoagulants Any 48 138 (39.5) 35 387 (32.9) 12 751 (89.0) Antihypertensives Any 90 160 (74.0) 76 996 (71.6) 12 164 (91.9) Clinical and health-related characteristics Diabetes duration (years) Mean (SD) 5.0 (6.4) 4.8 (6.1) 6.5

9) Concurrent medications Diabetes medications Any 80 889 (66.4) 72 126 (67.0) 8763 (61.2) Anticoagulants Any 48 138 (39.5) 35 387 (32.9) 12 751 (89.0) Antihypertensives Any 90 160 (74.0) 76 996 (71.6) 12 164 (91.9) Clinical and health-related characteristics Diabetes duration (years) Mean (SD) 5.0 (6.4) 4.8 (6.1) 6.5 (7.7) Median (IQR) 3.0 (8.0) 2.2 (7.0) 4.0 (10.0) HbA1c, mmol/mol (%) <42 (<5) 9706 (8.0) 8754 (8.1) 952 (6.6) 42–52 (5–6) 61 162 (50.2) 53 525 (49.8) 7637 (53.3) >52 (>6) 51 046 (41.9) 45 308 (42.1) 5738 (40.1) Mean (SD) 53.0 (12.0) 53.1 (12.1) 52.4 (10.9) Median (IQR) 50.0 (12.6) 50.0 (12.8) 50.0 (12.0) eGFR <60 9606 (7.9) 7521 (7.0) 2085 (14.6) (mL/min/1.73 m†) ≥60 112 308 (92.1) 100 066 (93.0) 12 242 (85.5) Mean (SD) 85.3 (20.3) 86.3 (20.2) 77.5 (18.9) Median (IQR) 84.3 (22.6) 85.3 (22.4) 77.4 (21.1) BMI <18.5 173 (0.1) 155 (0.1) 18 (0.1) (kg/m†) 18.5–24.9 10 015 (8.2) 8795 (8.2) 1220 (8.5) 25.0–29.9 52 074 (42.7) 45 169 (42.0) 6905 (48.2) ≥30.0 59 652 (48.9) 53 468 (49.7) 6184 (43.2) Mean (SD) 30.3 (4.5) 30.4 (4.5) 29.8 (4.2) Median (IQR) 29.9 (4.8) 30.0 (4.9) 29.5 (4.2) Systolic pressure (mm Hg) <130 27 429 (22.5) 25 080 (46.3) 2349 (16.4) ≥130 94 485 (77.5) 57 826 (53.8) 11 978 (83.6) Mean (SD) 137.6 (13.7) 137.4 (13.7) 139.5 (13.7) Median (IQR) 137.3 (14.2) 136.8 (14.0) 139.9 (13.4) Diastolic pressure (mm Hg) <80 58 138 (47.7) 49 761 (46.3) 8377 (58.5) ≥80 63 776 (52.3) 57 826 (53.8) 5950 (41.5) Mean (SD) 79.1 (8.0) 79.3 (8.0) 77.4 (7.8) Median (IQR) 80.0 (8.0) 80.0 (8.5) 78.2 (7.2) LDL cholesterol (mmol/L) <2.5 11 556 (9.5) 9937 (9.2) 1619 (11.3) ≥2.5 110 358 (90.5) 97 650 (90.8) 12 708 (88.7) Mean (SD) 3.4 (0.7) 3.4 (0.7) 3.2 (0.7) Median (IQR) 3.4 (0.7) 3.4 (0.7) 3.3 (0.7) HDL cholesterol (mmol/L) <1.0 (men)/<1.3 (women) 30 489 (25.0) 27 425 (25.5) 3064 (21.4) ≥1.0 (men)/≥1.3 (women) 91 425 (75.0) 80 162 (74.5) 11 263 (78.6) Mean (SD), men/women 1.2 (0.3)/1.4 (0.3) 1.2 (0.3)/1.4 (0.2) 1.2 (0.2)/1.4 (0.3) Median (IQR), men/women 1.2 (0.3)/1.3 (0.3) 1.1 (0.3)/1.4 (0.3) 1.2 (0.3)/1.4 (0.3) Triglycerides (mmol/L) <2.0 73 909 (60.6) 64 400 (59.9) 9509 (66.4) ≥2.0 48 005 (39.4) 43 187 (40.1) 4818 (33.6) Mean (SD) 2.0 (0.7) 2.0 (1.1) 1.9 (0.9) Median (IQR) 1.8 (0.9) 1.8 (0.9) 1.7 (0.7) Microalbuminuria Yes 10 693 (8.8) 9138 (8.5) 1555 (10.9) Macroalbuminuria Yes 3890 (3.2) 3222 (3.0) 668 (4.7) Kidney disease Yes 1859 (1.5) 1412 (1.3) 447 (3.1) Cancer diagnosis Yes 4521 (3.7) 3817 (3.6) 704 (4.9) Physical activity‡ Less than once per week 21 201 (17.4) 18 071

Median (IQR) 1.8 (0.9) 1.8 (0.9) 1.7 (0.7) Microalbuminuria Yes 10 693 (8.8) 9138 (8.5) 1555 (10.9) Macroalbuminuria Yes 3890 (3.2) 3222 (3.0) 668 (4.7) Kidney disease Yes 1859 (1.5) 1412 (1.3) 447 (3.1) Cancer diagnosis Yes 4521 (3.7) 3817 (3.6) 704 (4.9) Physical activity‡ Less than once per week 21 201 (17.4) 18 071 (16.8) 3130 (21.9) 1–2 times/week 23 635 (19.4) 21 017 (19.5) 2618 (18.3) 3–5 times/week 29 670 (24.3) 26 803 (24.9) 2867 (20.0) Daily 47 408 (38.9) 41 696 (38.8) 5712 (39.9) Smoking§ Yes 13 718 (11.3) 12 581 (11.7) 1137 (7.9) *Includes Belgium, Denmark, Germany, Greece, Spain, France, Ireland, Italy, Luxembourg, The Netherlands, Austria, Portugal, Finland, Sweden and Great Britain. †If age 65 years or older and unemployed. ‡30 min walk or equivalent. §At least one cigarette or pipe daily or quit smoking within 3 months. BMI, body mass index;HDL, high-density lipoprotein;HbA1c, hemoglobin A1c;LDL, low-density lipoprotein;eGFR, estimated glomerular filtration rate.