Browse the corpus

Introduction Over the past decade the emerging global threat of antibacterial resistance has alarmingly come to the forefront. Antibiotics, thought one of the greatest medical discoveries of the 20th century, and still pivotal in medicine now, are becoming increasingly compromised. The diminishing effectiveness of many antibiotics is due to the emergence of antibacterial resistance, and although a natural phenomenon, the inappropriate use of antibiotics in both human beings and animals worldwide, has accelerated the emergence and spread of highly resistant bacterial clones.1, 2 As regularly highlighted, between 1929 and the 1970s, 20 new classes of antibiotics were introduced to the market; since then, there has been a discovery void, with only two new classes reaching this stage.3 The speed at which bacteria have evolved to become resistant to antibiotics has surpassed the speed of drug discovery. This exacerbates the issue of resistance and stresses the need to preserve the efficacy of existing antibiotics. In short, without effective treatment, not only would bacterial epidemics become a substantial public health threat once again, but advances in modern medicine, ranging from minor surgery to cancer therapy, would also be jeopardised.3 Research in context Evidence before the study

Introduction Over the past decade the emerging global threat of antibacterial resistance has alarmingly come to the forefront. Antibiotics, thought one of the greatest medical discoveries of the 20th century, and still pivotal in medicine now, are becoming increasingly compromised. The diminishing effectiveness of many antibiotics is due to the emergence of antibacterial resistance, and although a natural phenomenon, the inappropriate use of antibiotics in both human beings and animals worldwide, has accelerated the emergence and spread of highly resistant bacterial clones.1, 2 As regularly highlighted, between 1929 and the 1970s, 20 new classes of antibiotics were introduced to the market; since then, there has been a discovery void, with only two new classes reaching this stage.3 The speed at which bacteria have evolved to become resistant to antibiotics has surpassed the speed of drug discovery. This exacerbates the issue of resistance and stresses the need to preserve the efficacy of existing antibiotics. In short, without effective treatment, not only would bacterial epidemics become a substantial public health threat once again, but advances in modern medicine, ranging from minor surgery to cancer therapy, would also be jeopardised.3 Research in context Evidence before the study On Jan 5, 2015, we searched PubMed and Google using the following search terms: “antibiotic research funding”, “antibiotic resistance research funding”, “antimicrobial research funding”, “antimicrobial resistance research funding”, “antibacterial resistance research funding”, “antibacterial research funding”, “AMR funding”, “infection research funding”, “bacteriology research funding”, for manuscripts published between Jan 1, 1995, and Jan 5, 2015, with no language restrictions. Our search identified two studies with some relevance to this study. Many reports and papers exist detailing the threat of antibacterial resistance; however, what research had been funded to address this pressing public health issue within and between countries was unknown. We concluded that the scale and scope of the two similar studies identified investigating the UK (Head and colleagues, 2014, and Bragginton and Piddock, 2014) were limited, only taking the UK into account, and their inclusion criteria did not span the breadth of our work. The Joint Programme–Neurodegenerative Disease initiative did a similar mapping exercise in 2011, which we took into consideration when designing this study (Joint Programme–Neurodegenerative Disease Mapping Exercise Report, 2011). Since we did not identify any similar exercise in antibacterial resistance, we did a comprehensive analysis to identify all the research funded across 19 Joint Programming Initiative on Antimicrobial Resistance (JPIAMR) countries and European Union (EU)-level organisations (Directorate General for Research and Innovation [DG Research], European Centre for Disease Prevention and Control [ECDC], Directorate General for Health and Consumer Affairs [DG-SANCO], and Innovative Medicines Initiative) related to antibacterial resistance from 2007 to 2013 to have the evidence base to influence future work. The JPIAMR strategic research agenda provided details on the six priority topics into which the research projects were categorised to produce meaningful results.

irs [DG-SANCO], and Innovative Medicines Initiative) related to antibacterial resistance from 2007 to 2013 to have the evidence base to influence future work. The JPIAMR strategic research agenda provided details on the six priority topics into which the research projects were categorised to produce meaningful results. Added value of this study

irs [DG-SANCO], and Innovative Medicines Initiative) related to antibacterial resistance from 2007 to 2013 to have the evidence base to influence future work. The JPIAMR strategic research agenda provided details on the six priority topics into which the research projects were categorised to produce meaningful results. Added value of this study This study takes a widely holistic and encompassing approach to look at a range of research areas that the JPIAMR strategic research agenda has identified as priorities to tackle the global problem of antibacterial resistance. We included research into treatments and preventive measures in human and veterinary medicine, diagnostics, surveillance, transmission, and interventions to prevent resistance emergence, acquisition, transmission, and infection. Additionally, this study analysed data from 19 countries (17 European countries, Canada, and Israel) and the EU, including DG Research (Framework Programme 6 and 7, European Research Council, and Innovative Medicines Initiative), DG-SANCO, and the ECDC. Hence, this study was the first comprehensive systematic analysis of research specifically relevant to antibacterial resistance. By including veterinary, public health, infection control, and diagnostic research in our study, we are better able to define the scale and scope of research to reduce the burden of antibacterial resistance at a national level across 19 JPIAMR countries and at the EU level through the European Commission and related EU agencies. Results for the UK were compared with Head and colleagues, 2014, and Bragginton and Piddock, 2014, who also analysed UK data. However, their inclusion criteria did not span the breadth of this study (since we also included measures to prevent antibacterial resistance, optimisation of the use of existing antibiotics, diagnostics, surveillance, veterinary and environmental research, and interventions). Additionally, data for this study were mainly extracted through internal databases, probably resulting in more hits than manual searching through websites as was done for the two UK studies.

isation of the use of existing antibiotics, diagnostics, surveillance, veterinary and environmental research, and interventions). Additionally, data for this study were mainly extracted through internal databases, probably resulting in more hits than manual searching through websites as was done for the two UK studies. Implications of all the available evidence

isation of the use of existing antibiotics, diagnostics, surveillance, veterinary and environmental research, and interventions). Additionally, data for this study were mainly extracted through internal databases, probably resulting in more hits than manual searching through websites as was done for the two UK studies. Implications of all the available evidence Both UK-based studies and several other reports have claimed investment in antibacterial resistance research is inadequate when compared with the burden of resistance or investment in other areas of health research. We too came to the same conclusions; however, although the provision of new funds is necessary, what these studies did not address is that the burden of antibacterial resistance cannot be tackled by focusing solely on antibiotic development research. To overcome this global threat, a multidisciplinary and transnational approach is needed and all priorities identified by the JPIAMR strategic research agenda require due consideration and investment. Overall, our results provide a baseline from which JPIAMR countries and the EU-level organisations can measure their investment in this area from 2013 onwards and could provide the evidence needed to influence practice and policy at the national level and EU level. This comprehensive analysis also lays the groundwork for future follow-on studies in the area to capture the number of new initiatives in antibacterial resistance in recent times, such as the UK antimicrobial resistance cross-council initiative and the diagnostic prizes announced by the UK and the European Commission, among others. The database of funded research on antibacterial resistance being created and maintained by the JPIAMR will help facilitate studies on antibacterial resistance more readily in the future.

crobial resistance cross-council initiative and the diagnostic prizes announced by the UK and the European Commission, among others. The database of funded research on antibacterial resistance being created and maintained by the JPIAMR will help facilitate studies on antibacterial resistance more readily in the future. In Europe alone, an estimated 25 000 people died from resistant bacterial sepsis in 2007, costing €1·5 billion, placing a substantial strain on already stressed health budgets.4 A review5 in 2014 estimated that an additional 10 million lives a year will be lost by 2050 worldwide as a result of antimicrobial resistance in six key pathogens, four being bacterial, resulting in a cumulative cost of US$100 trillion. Because of the assumptions made and data analysed, these figures are predicted to be an underestimate,5, 6 but the trends are clear and cannot be ignored. We are rapidly returning to a pre-antibiotic era, potentially resulting in the next world health crisis. The pressing urgency of this issue has been pointed out by several organisations worldwide, including WHO, with the poignant slogan used on World Health Day 2011, “No action today—no cure tomorrow”.7

In Europe alone, an estimated 25 000 people died from resistant bacterial sepsis in 2007, costing €1·5 billion, placing a substantial strain on already stressed health budgets.4 A review5 in 2014 estimated that an additional 10 million lives a year will be lost by 2050 worldwide as a result of antimicrobial resistance in six key pathogens, four being bacterial, resulting in a cumulative cost of US$100 trillion. Because of the assumptions made and data analysed, these figures are predicted to be an underestimate,5, 6 but the trends are clear and cannot be ignored. We are rapidly returning to a pre-antibiotic era, potentially resulting in the next world health crisis. The pressing urgency of this issue has been pointed out by several organisations worldwide, including WHO, with the poignant slogan used on World Health Day 2011, “No action today—no cure tomorrow”.7 Antibacterial resistance is a multifaceted problem needing vast and versatile solutions. No individual sector or nation has the capacity to independently handle this major societal challenge. Therefore, to collectively address antibacterial resistance at a national level and to increase the current impact of public research through more effective, efficient, and aligned investment, the Joint Programming Initiative on Antimicrobial Resistance (JPIAMR) was established in 2011.8 This initiative brings together 19 JPIAMR member countries, consisting of 17 European countries (Belgium, Czech Republic, Denmark, Finland, France, Germany, Greece, Italy, the Netherlands, Norway, Poland, Romania, Spain, Sweden, Switzerland, Turkey, the UK), Canada, and Israel, with Estonia and Argentina as JPIAMR observers.

nitiative brings together 19 JPIAMR member countries, consisting of 17 European countries (Belgium, Czech Republic, Denmark, Finland, France, Germany, Greece, Italy, the Netherlands, Norway, Poland, Romania, Spain, Sweden, Switzerland, Turkey, the UK), Canada, and Israel, with Estonia and Argentina as JPIAMR observers. The JPIAMR launched its strategic research agenda in April, 2014, outlining the member states' common vision to tackle antibacterial resistance.8 To ensure comprehensive actions are pursued, the strategic research agenda identifies six holistic and encompassing priority topics. The strategic research agenda acts as a dynamic framework on which the JPIAMR will continue to launch joint activities to guide and align research and investment to reduce the burden of antibacterial resistance across Europe and beyond. The JPIAMR will maintain and extend engagement activities internationally with different stakeholders, including industry, health service organisations, policy makers, the European Commission, and the research community, among others.

investment to reduce the burden of antibacterial resistance across Europe and beyond. The JPIAMR will maintain and extend engagement activities internationally with different stakeholders, including industry, health service organisations, policy makers, the European Commission, and the research community, among others. To guide future research activities and underpin the implementation of the JPIAMR strategic research agenda, an understanding of the present research landscape is necessary, which can be achieved by obtaining an objective insight into the scale and scope of research specifically relevant to antibacterial resistance across JPIAMR countries, the European Commission, and related European Union (EU) agencies. No detailed analyses on research portfolios and associated investment to address antibacterial resistance—including human, veterinary, and environmental research—across multiple countries have been done previously. In this Article we present in-depth analyses examining research specifically relevant to antibacterial resistance from major funding organisations funded within a 7 year period (2007–13) across the 19 JPIAMR countries and at EU level. The EU-level funding includes funding from the Directorate General for Research and Innovation (DG Research) through Framework Programme (FP) 6 and FP7 (including the Innovative Medicines Initiative first programme [IMI-1] and the European Research Council [ERC]), the Directorate General for Health and Consumer Affairs (DG-SANCO), and the European Centre for Disease Prevention and Control (ECDC). We will then provide recommendations for how JPIAMR and member countries could proceed in the future.

icines Initiative first programme [IMI-1] and the European Research Council [ERC]), the Directorate General for Health and Consumer Affairs (DG-SANCO), and the European Centre for Disease Prevention and Control (ECDC). We will then provide recommendations for how JPIAMR and member countries could proceed in the future. Methods Participating countries and data sources We did a systematic observational analysis surveying the 19 JPIAMR countries to establish levels of publicly funded research into antibacterial resistance. A questionnaire was sent to, and completed by, all JPIAMR national representatives who contacted suitable public funding agencies within their respective country. The full list of funding agencies included in this survey and the questionnaire used are in the appendix. All national representatives were briefed on the survey through presentations and discussions at JPIAMR meetings and through one-to-one communications with the data analyser. The variables collected included organisation name, principal investigator, lead institution, title, abstract or summary, start and end dates, and the total investment in euros. Only research funded by public funding bodies (which can invest in both public and private organisations) was collected and no private organisations were contacted for their investments because of difficulties in obtaining data. Data were also provided by DG Research (including FP6, FP7, ERC, and IMI), DG-SANCO, and ECDC, hereby collectively referred to as EU level. Additional desk study was done, mainly using websites and published databases from research organisations, at national and EU level, to identify through keyword searches additional publicly funded research that might have been missed. These projects were then verified for inclusion by the organisation or national representative. The data collected were the most comprehensive data available at the time of the survey from the organisations contacted.

to identify through keyword searches additional publicly funded research that might have been missed. These projects were then verified for inclusion by the organisation or national representative. The data collected were the most comprehensive data available at the time of the survey from the organisations contacted. We used the title and abstract where available (an abstract or summary was available for 984 [87%] of the 1129 JPIAMR country-level and 112 [98%] of the 114 EU-level projects) as a filter to establish research specifically relevant to antibacterial resistance. The inclusion criteria were based on the six encompassing priority topics identified in the JPIAMR strategic research agenda. This Article concentrates on all active and completed research projects with committed funding of €100 000 or more from January, 2007, up to, and including, December, 2013, in basic, applied, and clinical research, including trials, epidemiological, public health, and veterinary research. Some projects might not have begun spending until early 2014 and projects less than €100 000 were not regarded by funders as research projects but as network funding and therefore were not captured for this Article. Projects included were in the areas of therapeutics, such as from basic research to market, ranging from understanding the molecular mechanisms of resistance, to the development of new antibiotics and therapeutic alternatives to antibiotics (such as antivirulence drugs, vaccines, coatings on implants, bacteriophages, drug delivery, etc), and the optimisation of the use of existing antibiotics (eg, stewardship and clinical trials on combination treatment options); development of new diagnostics, such as a point-of-care test to effectively differentiate between bacterial and viral infections, or a rapid test to identify resistant bacteria and its resistance or sensitivity profile; surveillance, such as monitoring resistance rates, or antibiotic use in human or agricultural settings at local, national, or international levels; the transmission dynamics of resistance between different (human and animal) reservoirs; the assessment of the effect of environmental pollution (eg, water, soil, sewage) containing antibiotics, antibiotic residues, and resistant bacteria on the spread of resistance; and interventions to prevent the acquisition, transmission, and infection caused by antibacterial resistant bacteria (eg, infection control procedures, hospital layout, and education programmes).

n (eg, water, soil, sewage) containing antibiotics, antibiotic residues, and resistant bacteria on the spread of resistance; and interventions to prevent the acquisition, transmission, and infection caused by antibacterial resistant bacteria (eg, infection control procedures, hospital layout, and education programmes). Since funding mechanisms across different countries and agencies vary and to maximise consistency, investment allocated to infrastructure (eg, buildings, faculties, and networks) was not included, unless embedded within large grants. All financial information is as reported by the funders; grants awarded in a currency other than euros were converted to euros by the organisations at the time of data collection. No adjustments were made by the data analyst for inflation since information about whether adjustments for inflation had been previously made by individual organisations was not available before submission.

arded in a currency other than euros were converted to euros by the organisations at the time of data collection. No adjustments were made by the data analyst for inflation since information about whether adjustments for inflation had been previously made by individual organisations was not available before submission. Statistical analysis To ensure the data collected met the established inclusion criteria, that no duplication of projects occurred, and that the data provided were complete and accurate, all data were checked and validated twice. First, this was done by the representatives within each participating JPIAMR country and EU-level organisations, and second, in more detail, by the data analyser to ensure consistency across funding organisations. For projects where only a proportion of the project met the inclusion criteria, for example a project looking at fungal and bacterial resistance, only a proportion of funding was allocated to this project; this was done on a case-by-case basis by the data analyser. To ensure consistent classification of projects, the data analyser read the title, abstract, and any further information of each individual project and classified them into one or more of the six priority topics. The JPIAMR Scientific Advisory Board was consulted for any uncertainties. Data were sourced, categorised, and analysed during the period of July 1, 2013, to Feb 1, 2015. Data analyses and generation of figures and graphs were done with Microsoft Excel 2010.

Statistical analysis To ensure the data collected met the established inclusion criteria, that no duplication of projects occurred, and that the data provided were complete and accurate, all data were checked and validated twice. First, this was done by the representatives within each participating JPIAMR country and EU-level organisations, and second, in more detail, by the data analyser to ensure consistency across funding organisations. For projects where only a proportion of the project met the inclusion criteria, for example a project looking at fungal and bacterial resistance, only a proportion of funding was allocated to this project; this was done on a case-by-case basis by the data analyser. To ensure consistent classification of projects, the data analyser read the title, abstract, and any further information of each individual project and classified them into one or more of the six priority topics. The JPIAMR Scientific Advisory Board was consulted for any uncertainties. Data were sourced, categorised, and analysed during the period of July 1, 2013, to Feb 1, 2015. Data analyses and generation of figures and graphs were done with Microsoft Excel 2010. Role of the funding source The funders of the study had no role in the study design, data analysis, data interpretation, or writing of the Article. Members of the JPIAMR and the European Commission provided data included in this work. To the authors' knowledge, all data analysed for this Article are already publicly available and therefore have no restrictions. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Commission provided data included in this work. To the authors' knowledge, all data analysed for this Article are already publicly available and therefore have no restrictions. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results We identified 1234 publicly funded antibacterial resistance projects across 19 JPIAMR countries and at the EU level, with a total investment of €960·7 million. Additionally, DG Research invested in another nine antibacterial resistance research projects via the IMI-1 programme in partnership with the European Federation of Pharmaceutical Industries and Associations, bringing the total public investment in antibacterial resistance research from 2007 to 2013 across 1243 projects to more than €1·3 billion (table). We found the 19 JPIAMR countries collectively contributed €646·6 million (67%) of the €960·7 million total investment in antibacterial resistance research, whereas the remaining €314·1 million (33%) was provided at the EU level. However, when DG Research's contribution to the IMI-1 is considered, JPIAMR countries accounted for only 49·5% (€646·6 million of €1·3 billion) of this investment and 50·5% (€659·2 million of €1·3 billion) was provided at the EU level (table). We investigated patterns over time by analysing the committed budget per year in 19 JPIAMR countries collectively and in EU-level organisations, where positive patterns in funding from 2007 to 2013 were observed (appendix).

this investment and 50·5% (€659·2 million of €1·3 billion) was provided at the EU level (table). We investigated patterns over time by analysing the committed budget per year in 19 JPIAMR countries collectively and in EU-level organisations, where positive patterns in funding from 2007 to 2013 were observed (appendix). At the national level, we identified 1129 projects funded across 19 JPIAMR countries, with a total public sector investment of €646·6 million across the 7 year period (2007–13). Most public sector funding went to universities, some went to hospitals, and only a small proportion went to private organisations. We analysed the total number of projects and associated investment within each of the strategic research agenda priority topics and between countries. Because antibacterial resistance is a complex issue, some multidisciplinary projects spanned more than one priority topic. As such, 69 (6·1%) of the 1129 projects in total funded at national level are classified under more than one priority topic, slightly inflating the total number of projects to 1208. Overlap was most evident between therapeutics and transmission in basic underpinning science projects and between transmission, environment, and surveillance. No duplication of funding occurred since an equal proportion of investment was assigned to each priority topic in projects classified under more than one priority topic.

was most evident between therapeutics and transmission in basic underpinning science projects and between transmission, environment, and surveillance. No duplication of funding occurred since an equal proportion of investment was assigned to each priority topic in projects classified under more than one priority topic. Of 1208 projects investigated, with €646·6 million invested overall, 763 (63%) were within the area of therapeutics, with a total investment of €428·2 million (66%); 185 (15%) were based on the transmission of antibacterial resistance, with a total investment of €55·5 million (9%); 131 (11%) were about the development of new diagnostics, with a total investment of €90·4 million (14%); 53 (4%) were within the area of interventions, with a total investment of €35 million (5%); 39 (3%) were within the area of surveillance, with a total investment of €25·1 million (4%); and 37 (3%) focused on the environment, with a total investment of €12·5 million (2%; figure 1).

tal investment of €90·4 million (14%); 53 (4%) were within the area of interventions, with a total investment of €35 million (5%); 39 (3%) were within the area of surveillance, with a total investment of €25·1 million (4%); and 37 (3%) focused on the environment, with a total investment of €12·5 million (2%; figure 1). We identified substantial variations in funding across countries at the national level, both in terms of number of projects and investment. Figure 2 shows the number of projects per country by priority topic; similar results were evident for investment (appendix). In an attempt to gauge the number of projects in this area per person of population, we also analysed the ratio of the number of projects to the population of that country (figure 3). The mean country population from 2007 to 2013 was used.9, 10 Again, substantial variation existed between countries in the number of projects per person of the population with a broad spread around the mean, but the data suggested that Denmark, Estonia, Finland, the Netherlands, Sweden, and the UK funded more projects than the other countries included (figure 3). Similar results were evident for the ratio of investment per person of the population (appendix).

er person of the population with a broad spread around the mean, but the data suggested that Denmark, Estonia, Finland, the Netherlands, Sweden, and the UK funded more projects than the other countries included (figure 3). Similar results were evident for the ratio of investment per person of the population (appendix). We found that EU-level organisations invested €314·1 million in 105 research projects specifically related to antibacterial resistance from 2007 to 2013 via DG Research (FP6, FP7, ERC), DG-SANCO, and ECDC (DG Research's contribution to IMI is not shown). Since substantial investment was made through large multinational consortia, 19 (18%) of the 105 projects were classified under more than one priority topic. No duplication of funding occurred since a percentage of investment has been assigned to each priority topic in projects classified under more than one priority topic. Of the 133 projects funded at the EU level (105 projects plus the duplicates), receiving an investment of €314·1 million overall, 71 (53%) were in the area of therapeutics, with a total investment of €197·4 million (63%); 18 (14%) were classified as transmission and received an investment of €42·7 million (13%); 16 (12%) were classified as surveillance, but only received an investment of €8·5 million (3%); 13 (10%) were classified as diagnostics and received €38·3 million (12%); 11 (8%) were classified as interventions, with investments of €21·2 million (7%); and four (3%) were classified as the environment, with investments of €5·9 million (2%; figure 4).

rveillance, but only received an investment of €8·5 million (3%); 13 (10%) were classified as diagnostics and received €38·3 million (12%); 11 (8%) were classified as interventions, with investments of €21·2 million (7%); and four (3%) were classified as the environment, with investments of €5·9 million (2%; figure 4). We captured nine projects within IMI-1 meeting the study inclusion criteria, with a total investment of €723·5 million. DG Research has committed €345·1 million to these nine projects via FP7. Most projects address more than one JPIAMR priority topic, although funding is mainly within priority topic A: therapeutics, to strengthen clinical research on antibacterial resistance in Europe. In addition to preclinical and clinical research, the IMI-1 funded projects also focus on developing research infrastructure—for example, all projects, with the exception of one, are required to submit information to a specific database, and three projects aim to develop a drug discovery platform and pan-European clinical trials and laboratory networks.

ical research, the IMI-1 funded projects also focus on developing research infrastructure—for example, all projects, with the exception of one, are required to submit information to a specific database, and three projects aim to develop a drug discovery platform and pan-European clinical trials and laboratory networks. Discussion To our knowledge, this study is the first systematic analysis of research funding specifically relevant to antibacterial resistance across 19 countries, and at EU level, including the IMI-1. This study was also the first, to our knowledge, to take human, veterinary, and environmental research, and all areas identified in the JPIAMR strategic research agenda, including therapeutics, diagnostics, surveillance, transmission, environment, and interventions, into account. IMI-1, the world's largest public–private partnership in life sciences, involving the European Commission and European Federation of Pharmaceutical Industries and Associations, was important to capture as the programme is investing substantially to accelerate the development of better and safer medicines for patients, including antibiotics and alternative treatments for bacterial infections.11

lving the European Commission and European Federation of Pharmaceutical Industries and Associations, was important to capture as the programme is investing substantially to accelerate the development of better and safer medicines for patients, including antibiotics and alternative treatments for bacterial infections.11 When looking at the total spend and number of awards made in antibacterial resistance research from 2007 to 2013, investment across the JPIAMR countries seems to be substantial. However, when compared with total spend on research across the disciplines, the amount spent on antibacterial resistance seems to be very small. For example, in the UK, one of the biggest investors, antibacterial resistance research spend was about 1% of the total research spend for the same period for the four research councils included.12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22 Also evident from the analyses of available data is that, at the EU level, substantially more was invested in antibacterial resistance research in 2007–13 (€314 million and an additional €345 million in IMI-1) than by the JPIAMR countries, which collectively invested €647 million. Therefore, not taking into account the DG Research's contribution to the IMI-1, 33% of the total investment was at the EU level versus 67% from all 19 countries (table). By contrast, when looking at other major societal challenges, such as neurodegenerative diseases, the proportion of annual funding allocated in projects active on Jan 1, 2011, by DG Research (FP7) was 15% (€57 million) versus 85% (€314 million) by the 20 countries participating in the Joint Programme for Neurodegenerative Disease.23 This proportion of funding invested in neurodegenerative disease research by DG Research is more in line with the EU research FP spend. As in 2007–08, EU FP funds represented about 7·5% of all civil research and development expenditure financed by governments of EU member states and European Free Trade Association countries. Although we recognise the 7·5% captures all civil research and development spend, comparing it with what we found (33% funded at the EU level vs 67% by governments in antibacterial resistance research; table) emphasises the need for national budgets to redress the balance.24

d European Free Trade Association countries. Although we recognise the 7·5% captures all civil research and development spend, comparing it with what we found (33% funded at the EU level vs 67% by governments in antibacterial resistance research; table) emphasises the need for national budgets to redress the balance.24 Furthermore, countries varied substantially in terms of total number of antibacterial resistance projects funded (figure 2) and number of projects funded per person of population (figure 3), with Denmark, Estonia, Finland, the Netherland, Sweden, and the UK investing in a greater number of projects than the other countries investigated. Interestingly, those countries that invest the most have lower levels of antibacterial resistance than many of the countries included in this Article that invest the least.25 Also, within countries, funding for other high priority health needs varied substantially—for example, according to the Joint Programme for Neurodegenerative Disease findings, Germany was the biggest funder of neurodegenerative disease research,23 yet, we found Germany was only a very minor funder of antibacterial resistance research. This disparity between funding stresses the need for new funds for antibacterial resistance within countries rather than redistributing funds from other essential areas of research. The exact reasons for the disparate funding within and between countries are unclear, but we believe they argue for improved coordination between countries to share best practice and experience and to embark on joint research projects. We also believe these observations argue for greater sharing of results and data so that the outputs and effects of large investments can be realised across the EU and beyond.

e believe they argue for improved coordination between countries to share best practice and experience and to embark on joint research projects. We also believe these observations argue for greater sharing of results and data so that the outputs and effects of large investments can be realised across the EU and beyond. The burden of disease caused by antibacterial resistance is on the increase. In 2007, 25 000 people died in Europe from resistant bacterial sepsis, costing €1·5 billion.4 The 2014 review from the O'Neill commission5 established by the UK Government estimated that an additional 10 million lives per year will be lost by 2050 globally as a result of antimicrobial resistance in six key pathogens, resulting in a cumulative cost of US$100 trillion. This impending health challenge clearly needs to be addressed through research, but this research also needs to extend beyond the boundaries of individual states. The number of projects funded by individual states is high; 1129 (91%) of the 1243 projects identified in this survey were funded at the national level, but they only account for €646·6 million (49%) of the total investment, suggesting that these are relatively small awards and highly focused. These awards would certainly have been made on the basis of scientific excellence through normal schemes.

43 projects identified in this survey were funded at the national level, but they only account for €646·6 million (49%) of the total investment, suggesting that these are relatively small awards and highly focused. These awards would certainly have been made on the basis of scientific excellence through normal schemes. In view of the relatively small amount of funds invested at the national level in antibacterial resistance compared with the EU level, other health priorities, and other countries, changes might now be needed at the national level in countries that correspond with the strategic importance and growth of antibacterial resistance. The idea behind joint programming is that resources are scarce and the availability of public investments in research has limits.26 Hence, participating countries will not only need to close the gap between the health research needs and the actual research funded but also make strategic and coordinated investments with existing and new funds. These strategies entail that countries work both alone and together in a more efficient way than they are now to increase the effectiveness of research through strengthening national and international coordination and collaborations, harmonising research activities, and collectively pooling resources. From our analysis, some countries already invest substantial funds into antibacterial resistance but a new way of funding multidisciplinary and transnational antibacterial resistance research is needed. The UK Cross Research Council Initiative aims to achieve this goal, not only by linking different research disciplines, both nationally and internationally, but also through coordination of different sectors and funders through the UK antimicrobial resistance funders forum.27, 28

acterial resistance research is needed. The UK Cross Research Council Initiative aims to achieve this goal, not only by linking different research disciplines, both nationally and internationally, but also through coordination of different sectors and funders through the UK antimicrobial resistance funders forum.27, 28 The JPIAMR has also developed the strategic research agenda to coordinate research in close collaboration with the funding instruments of the European Commission, specifically Horizon 2020, IMI, and the ERA-NET scheme to increase effectiveness and avoid duplication. Countries could use the JPIAMR strategic research agenda now as a template for coordination. Beyond Europe, government priorities are beginning to change, and new increased budget commitments are being discussed, with the US Government proposing to invest more than $1·2 billion of funding for 2016 to improve antibiotic stewardship, strengthen antibiotic resistance risk assessment, surveillance, and reporting capabilities, and drive research and innovation in the human health and agricultural sectors.29 Activity within the area of diagnostics has also increased, with the US Department of Health and Human Services announcing a diagnostic prize of up to $20 million,30 the European Commission announcing a €1 million diagnostic prize,31 and the UK announcing a £10 million Longitude prize in diagnostics.32

tural sectors.29 Activity within the area of diagnostics has also increased, with the US Department of Health and Human Services announcing a diagnostic prize of up to $20 million,30 the European Commission announcing a €1 million diagnostic prize,31 and the UK announcing a £10 million Longitude prize in diagnostics.32 As with similar exercises,33, 34 the data presented have limitations. To ensure the information captured was comparable and consistent, only research projects were captured, and institutional funding was unobtainable. Additionally, because the remit of the study was to capture research projects only, surveillance systems that are now core programmes—such as ECDC in-house programmes and national surveillance programmes, including the German KISS and the French RAISIN programmes, among others—were not captured in this study and therefore the picture of surveillance funding captured by this study is not comprehensive. We relied on the accuracy of the data provided by funding bodies, although all data were verified by each national or EU representative and any apparent discrepancies or duplication of projects were dealt with by the data analyst.

herefore the picture of surveillance funding captured by this study is not comprehensive. We relied on the accuracy of the data provided by funding bodies, although all data were verified by each national or EU representative and any apparent discrepancies or duplication of projects were dealt with by the data analyst. All projects were determined for inclusion, first, by each organisations representative, and second, by the data analyst. Some projects might have been missed because of the search terms used or the difficulties in accessing data in some countries. The rules regarding what a research grant will cover can vary between countries (eg, salary of investigators); we made no attempt to remove any indirect and estate costs included in the funding amounts and did not adjust for inflation. However, to deal with these financial confounders, we analysed the number of projects and compared trends across countries using number of projects rather than focusing solely on investment. If this exercise were to be extended globally and include countries such as China and India with fluctuating currency and inflation rates over time, the method described by Young and colleagues35 could be followed to ensure careful comparison between countries with variable exchange and inflation rates. The subjective classification of these cross-disciplinary projects, by the data analyst, to one or more of the priority topics leaves the exercise somewhat open to question. Additionally, for projects classified under more than one priority topic, we could not ascertain the exact proportion of grant funding to be allocated to each of the topics, hence, to ensure consistency an equal proportion of funding was assigned to each. Overlap was often evident between therapeutics and transmission in basic underpinning science projects and also between transmission, environment, and surveillance.

exact proportion of grant funding to be allocated to each of the topics, hence, to ensure consistency an equal proportion of funding was assigned to each. Overlap was often evident between therapeutics and transmission in basic underpinning science projects and also between transmission, environment, and surveillance. Currency conversions to euros will not be precise because of variations in the exchange rates across the data collection process. Although private funding invested in IMI-1 was available, this study did not include private sector funding because of difficulties in openly accessing private funding and project information beyond IMI-1.

sions to euros will not be precise because of variations in the exchange rates across the data collection process. Although private funding invested in IMI-1 was available, this study did not include private sector funding because of difficulties in openly accessing private funding and project information beyond IMI-1. As a result of this study, several recommendations have been made to the JPIAMR members to consider, with some activities already underway. At present, no comprehensive database exists to document research at both national and international levels, and, in view of this study, improvements in data sharing and communication clearly need to be achieved at the national level in several countries. The JPIAMR is actively working to improve data sharing and has turned the research data collected for this study into a useful, freely accessible, and searchable database available on the JPIAMR website. The database will enable researchers and funders to set strategic priorities by revealing what has already been funded across the different areas and what is still needed within the different priority areas. Furthermore, the database is intended to be used by researchers for networking and collaboration and to avoid duplication. If funders from other countries provide similarly detailed information about projects on antibacterial resistance research, global gaps and priorities could be assessed.

e different priority areas. Furthermore, the database is intended to be used by researchers for networking and collaboration and to avoid duplication. If funders from other countries provide similarly detailed information about projects on antibacterial resistance research, global gaps and priorities could be assessed. The substantial allocation of antibacterial resistance projects and investment within the area of therapeutics, both at the national and EU level, is probably because of the existing strength of basic bacteriology research across many of the JPIAMR countries. What will be important from now on is for the JPIAMR and other organisations to ensure that this level of basic research underpins the development of therapies, diagnostics, and intervention strategies for use in clinical and veterinary practice and in the environment to reduce resistance. However, the development of new treatment in a timely manner is challenging since antibiotics take years to develop; therefore actions need to be taken that can have an immediate effect on the rate of acquisition, transmission, and infection by resistant bacteria. Consequently, the JPIAMR have plans to launch a call in January, 2016, to boost funding in this key area, focusing on the One Health Agenda, which uses the ERA-NET scheme. Additionally, increased research effort in affordable, reliable, and rapid point-of-care diagnostics and interventions is necessary. Additionally, the effect of diligent hygiene practices and antimicrobial stewardship in human and agricultural settings, along with improved public and professional education, should not be underestimated.

eased research effort in affordable, reliable, and rapid point-of-care diagnostics and interventions is necessary. Additionally, the effect of diligent hygiene practices and antimicrobial stewardship in human and agricultural settings, along with improved public and professional education, should not be underestimated. To conclude, investment at present might not correspond with the burden of antibacterial resistance and the looming health, social, and economic threat it poses on the treatment of infections and on medicine in general. Antibacterial resistance clearly warrants increased and new investment from a range of sources, but improved coordination and collaboration with more informed resource allocation are needed to make a true impact. Hopefully, this analysis will prompt nations to pay due consideration to the existing research landscape when considering future investments. The analysis will act as a guide for the JPIAMR to ensure research is complementary and that no major overlaps exist, while aiming to identify gaps and opportunities to be exploited. The benefits of working across national boundaries, sharing experiences, and pooling resources are substantial, and these are yet to be fully taken advantage of by the JPIAMR countries. The entire JPIAMR strategic research agenda cannot be tackled at once by each country; thus, prioritisation of research within each country to where it is best placed to deliver useful results will ensure an efficient use of resources and avoid the duplication of efforts. Supplementary Material Supplementary appendix

To conclude, investment at present might not correspond with the burden of antibacterial resistance and the looming health, social, and economic threat it poses on the treatment of infections and on medicine in general. Antibacterial resistance clearly warrants increased and new investment from a range of sources, but improved coordination and collaboration with more informed resource allocation are needed to make a true impact. Hopefully, this analysis will prompt nations to pay due consideration to the existing research landscape when considering future investments. The analysis will act as a guide for the JPIAMR to ensure research is complementary and that no major overlaps exist, while aiming to identify gaps and opportunities to be exploited. The benefits of working across national boundaries, sharing experiences, and pooling resources are substantial, and these are yet to be fully taken advantage of by the JPIAMR countries. The entire JPIAMR strategic research agenda cannot be tackled at once by each country; thus, prioritisation of research within each country to where it is best placed to deliver useful results will ensure an efficient use of resources and avoid the duplication of efforts. Supplementary Material Supplementary appendix Acknowledgments RK, GZ, DW, and RW work for the Medical Research Council UK. The Medical Research Council received a grant from the JPIAMR to conduct this independent study. DW and GZ are JPIAMR Management Board Members. We thank the JPIAMR Scientific Advisory Board for helping to design the questionnaire and providing input throughout the work. We thank all the participating countries and organisations—including the national funding agencies, the JPIAMR Management Board members and the JPIAMR national contact points, the Directorate General for Research and Innovation, the Directorate General for Health and Consumer Affairs, and the European Centre for Disease Prevention and Control—for providing data for this study.

including the national funding agencies, the JPIAMR Management Board members and the JPIAMR national contact points, the Directorate General for Research and Innovation, the Directorate General for Health and Consumer Affairs, and the European Centre for Disease Prevention and Control—for providing data for this study. Contributors RW designed the questionnaire and began the initial data collection. RK collected the final data from key national contact points and collected data from DG Research (including FP6, FP7, ERC, and IMI-1), DG-SANCO, and ECDC, and searched for additional data in publicly available databases. RK collated and validated the dataset, and read all titles and abstracts or summaries to classify each project. RK did the data analysis and created the graphs and figures. RK, HG, DW, and GZ interpreted the data and RK wrote the first draft of the Article. HG, GZ, and DW all provided input to the Article. All authors had final approval of the Article. HG is a guarantor of the Article. Declaration of interests All authors declare no competing of interests. Figure 1 Antibacterial resistance projects funded at national level between 2007 and 2013 by priority topic with total funding

Contributors RW designed the questionnaire and began the initial data collection. RK collected the final data from key national contact points and collected data from DG Research (including FP6, FP7, ERC, and IMI-1), DG-SANCO, and ECDC, and searched for additional data in publicly available databases. RK collated and validated the dataset, and read all titles and abstracts or summaries to classify each project. RK did the data analysis and created the graphs and figures. RK, HG, DW, and GZ interpreted the data and RK wrote the first draft of the Article. HG, GZ, and DW all provided input to the Article. All authors had final approval of the Article. HG is a guarantor of the Article. Declaration of interests All authors declare no competing of interests. Figure 1 Antibacterial resistance projects funded at national level between 2007 and 2013 by priority topic with total funding Proportions of pie chart represent number of projects by priority topic and not total funding. Projects funded at the European Union level are not included. Some projects are classified under more than one priority topic; hence, the numbers of projects are duplicated. Investment is not duplicated because a percentage of investment has been assigned to each priority topic in projects classified under more than one priority topic. Percentages do not add up to 100% because of rounding. Figure 2 Total number of projects per country by priority topic funded at national level

Proportions of pie chart represent number of projects by priority topic and not total funding. Projects funded at the European Union level are not included. Some projects are classified under more than one priority topic; hence, the numbers of projects are duplicated. Investment is not duplicated because a percentage of investment has been assigned to each priority topic in projects classified under more than one priority topic. Percentages do not add up to 100% because of rounding. Figure 2 Total number of projects per country by priority topic funded at national level Totals include national data from participating countries from 2007 to 2013 and do not include projects funded at the European Union level within these countries. Some projects are classified under more than one priority topic (hence, the numbers of projects are duplicated). Figure 3 Ratio of total number of projects per country to the total national population in millions Total number of project includes national data from participating countries from 2007 to 2013 and does not include projects funded at European Union level. The mean country population from 2007 to 2013 was used.9, 10 The mean ratio is shown by the dotted line. Figure 4 Antibacterial resistance projects funded at the European Union level between 2007 and 2013 by priority topic with total funding

Total number of project includes national data from participating countries from 2007 to 2013 and does not include projects funded at European Union level. The mean country population from 2007 to 2013 was used.9, 10 The mean ratio is shown by the dotted line. Figure 4 Antibacterial resistance projects funded at the European Union level between 2007 and 2013 by priority topic with total funding Proportions of pie chart represent number of projects by priority topic and not total funding. European Union level funding collectively refers to funding by the Director General for Research and Innovation (Framework Programme 6 and 7 and the European Research Council), the Directorate General for Health and Consumer Affairs, and the European Centre for Disease Prevention and Control; Director General for Research and Innovation's contribution to Innovative Medicines Initiative first programme is not included in this chart. Some projects are classified under more than one priority topic; as such, the numbers of projects are duplicated. Investment is not duplicated within the priority topics because a percentage of investment has been assigned to each priority topic in projects classified under more than one priority topic. Table Total committed public funding to antibacterial resistance research by JPIAMR countries and the EU

Proportions of pie chart represent number of projects by priority topic and not total funding. European Union level funding collectively refers to funding by the Director General for Research and Innovation (Framework Programme 6 and 7 and the European Research Council), the Directorate General for Health and Consumer Affairs, and the European Centre for Disease Prevention and Control; Director General for Research and Innovation's contribution to Innovative Medicines Initiative first programme is not included in this chart. Some projects are classified under more than one priority topic; as such, the numbers of projects are duplicated. Investment is not duplicated within the priority topics because a percentage of investment has been assigned to each priority topic in projects classified under more than one priority topic. Table Total committed public funding to antibacterial resistance research by JPIAMR countries and the EU Total number of projects, 2007–13 Total funding (€), 2007–13 Proportion of total funding (excluding EC contribution to IMI) Proportion of total funding 19 JPIAMR countries 1129 646 646 541 67·3% 49·5% EU level* 114 659 201 418 NA 50·5% EU level (excluding IMI) 105 314 072 980 32·7% 24·1% IMI (EC contribution only) 9 345 128 438 NA 26·4% Overall† 1243 1 305 847 959 100% 100% The EU-level funding includes funding from the Director General for Research and Innovation (Framework Programme 6 and 7, including the IMI first programme and the European Research Council), the Directorate General for Health and Consumer Affairs, and the European Centre for Disease Prevention and Control from 2007 to 2013. JPIAMR=Joint Programming Initiative on Antimicrobial Resistance. EU=European Union. EC=European Commission. IMI=Innovative Medicines Initiative. NA=not applicable.

an Research Council), the Directorate General for Health and Consumer Affairs, and the European Centre for Disease Prevention and Control from 2007 to 2013. JPIAMR=Joint Programming Initiative on Antimicrobial Resistance. EU=European Union. EC=European Commission. IMI=Innovative Medicines Initiative. NA=not applicable. * Sum total of EU level (excluding IMI) and IMI (EC contribution only) funding. † Sum total of 19 JPIAMR countries and EU level funding.

Introduction In 2014, 91% of 3·2 million HIV-infected children lived in sub-Saharan Africa, but less than 25% of those needing antiretroviral therapy (ART) were receiving it.1 Low-cost, scored, dispersible fixed-dose combination (FDC) paediatric tablets of stavudine plus lamivudine plus nevirapine in child-appropriate drug ratios2 drove initial ART roll-out to African children, replacing separate syrups, which are costly for programmes and difficult for carers to transport and administer.3 However, stavudine was discouraged in 20104 and 20135 WHO guidelines because of high lipodystrophy rates in adults and adolescents. In children, stavudine-associated toxicity has mainly been noted with higher doses than those recommended by WHO and in older children.6, 7, 8

cult for carers to transport and administer.3 However, stavudine was discouraged in 20104 and 20135 WHO guidelines because of high lipodystrophy rates in adults and adolescents. In children, stavudine-associated toxicity has mainly been noted with higher doses than those recommended by WHO and in older children.6, 7, 8 Alternative nucleoside reverse-transcriptase inhibitors (NRTIs) for children younger than 12 years are abacavir or zidovudine. Tenofovir is not licensed for those younger than 2 years and is not recommended by WHO5 in those younger than 10 years, primarily because of concerns regarding long-term effects on bone metabolism and renal function in growing children,9 although more data are needed. Zidovudine is associated with anaemia, which is of particular concern in malnourished children in endemic malaria areas where underlying anaemia is prevalent. Abacavir is associated with hypersensitivity reactions, although these are rare in Africa10 because of a lower risk-allele prevalence.11 However, two South African cohorts recently reported lower virological suppression with abacavir than with stavudine,12, 13 and abacavir is also the most costly NRTI.14 Therefore, whether stavudine, given at the WHO recommended doses, should remain an option for young children was unclear. Research in context Evidence before this study

Alternative nucleoside reverse-transcriptase inhibitors (NRTIs) for children younger than 12 years are abacavir or zidovudine. Tenofovir is not licensed for those younger than 2 years and is not recommended by WHO5 in those younger than 10 years, primarily because of concerns regarding long-term effects on bone metabolism and renal function in growing children,9 although more data are needed. Zidovudine is associated with anaemia, which is of particular concern in malnourished children in endemic malaria areas where underlying anaemia is prevalent. Abacavir is associated with hypersensitivity reactions, although these are rare in Africa10 because of a lower risk-allele prevalence.11 However, two South African cohorts recently reported lower virological suppression with abacavir than with stavudine,12, 13 and abacavir is also the most costly NRTI.14 Therefore, whether stavudine, given at the WHO recommended doses, should remain an option for young children was unclear. Research in context Evidence before this study We searched PubMed up to April 27, 2015, using the keywords “HIV”, “child*”, (“stavudine” or “zidovudine” or “abacavir”), not “prevent*” (to exclude a large number of studies looking at zidovudine to prevent mother-to-child HIV transmission), dated after Jan 1, 1996, (when combination ART was introduced). The most relevant nucleoside reverse-transcriptase inhibitors (NRTIs) for treating HIV-infected children when the study started were abacavir, zidovudine, and stavudine; didanosine and tenofovir were not used because of toxicity (genuine or a potential concern, respectively). The WHO conducts systematic reviews as part of guideline development. No existing systematic reviews of randomised controlled trials comparing these NRTIs head-to-head in HIV-infected children were identified in 2010 or 2013, with only one randomised trial directly comparing abacavir and zidovudine in 128 European children, which identfied that abacavir was virologically superior to zidovudine over 5 years follow-up. Recommendations for preferential ordering of zidovudine, abacavir, then stavudine in 2010, and abacavir, zidovudine, then stavudine in 2013, were therefore based primarily on expert opinion balancing toxicity (estimated from observational studies and randomised trials not containing head-to-head comparisons), cost (greater with abacavir), and practicality (particularly availability as part of fixed-dose-combination tablets and once-daily dosing); and, in 2013, also evidence on accumulation of different resistance mutations with sequential use.

rvational studies and randomised trials not containing head-to-head comparisons), cost (greater with abacavir), and practicality (particularly availability as part of fixed-dose-combination tablets and once-daily dosing); and, in 2013, also evidence on accumulation of different resistance mutations with sequential use. Added value of the study This is the first randomised controlled trial in African children, conducting a head-to-head comparison of the three most relevant NRTIs for paediatric treatment, coformulated in NNRTI/NRTI generic fixed-dose-combination paediatric tablets and dosed with WHO drug ratios and weight bands. We identified no major differences between the NRTIs in adverse events, toxicity, clinical, immunological, or viral load endpoints, but did find higher drug susceptibility to relevant second-line NRTIs if abacavir was used first-line, thus providing evidence to support the WHO 2013 recommendation for its use as the preferred first-line NRTI for children. Use of abacavir also enables a once-daily ART regimen to be constructed for children, in line with adults. Implications of the available evidence

This is the first randomised controlled trial in African children, conducting a head-to-head comparison of the three most relevant NRTIs for paediatric treatment, coformulated in NNRTI/NRTI generic fixed-dose-combination paediatric tablets and dosed with WHO drug ratios and weight bands. We identified no major differences between the NRTIs in adverse events, toxicity, clinical, immunological, or viral load endpoints, but did find higher drug susceptibility to relevant second-line NRTIs if abacavir was used first-line, thus providing evidence to support the WHO 2013 recommendation for its use as the preferred first-line NRTI for children. Use of abacavir also enables a once-daily ART regimen to be constructed for children, in line with adults. Implications of the available evidence Excellent outcomes were obtained on all regimens, showing the importance of widening treatment access for HIV-infected children worldwide. Efforts need to be made to provide abacavir-based combinations where this is possible; but there is no need to move children who are stable on zidovudine-based regimens to abacavir. Further research should investigate the potential for once-daily triple abacavir-based fixed-dose combinations with efavirenz or dolutegravir to further simplify and improve durability of first-line ART for children who will need treatment for much longer than adults.

re stable on zidovudine-based regimens to abacavir. Further research should investigate the potential for once-daily triple abacavir-based fixed-dose combinations with efavirenz or dolutegravir to further simplify and improve durability of first-line ART for children who will need treatment for much longer than adults. Since 2003, changes in NRTIs recommended by WHO for children, followed by changes in national guidelines and clinical practice, have occurred with little evidence and no new randomised trials. Therefore, in 2010, when most African children were receiving stavudine-based ART, we aimed to compare stavudine, zidovudine, or abacavir fixed-dose combinations for first-line ART.

ildren, followed by changes in national guidelines and clinical practice, have occurred with little evidence and no new randomised trials. Therefore, in 2010, when most African children were receiving stavudine-based ART, we aimed to compare stavudine, zidovudine, or abacavir fixed-dose combinations for first-line ART. Methods Study design and participants In this open-label, parallel-group, randomised controlled trial (CHAPAS-3), we enrolled confirmed HIV-infected children from Zambia and Uganda—centres were from Zambia—the University Teaching Hospital (UTH), Lusaka; and from Uganda Baylor-Uganda Centre of Excellence, Kampala, and Joint Clinical Research Centre (JCRC), Kampala and Gulu (satellite site)—aged 1 month to 13 years if they were either previously untreated and met WHO 20104 criteria for ART (ART naive; <5 years in Uganda), or on stavudine-containing first-line (non-nucleoside reverse-transcriptase inhibitors [NNRTI]-containing) ART for 2 years or more with screening viral load less than 50 copies per mL and stable CD4 and/or CD4 cell % (ART-experienced; no signs of lipodystrophy; see appendix p 2 for additional eligibility criteria). All children were already on or initiated co-trimoxazole prophylaxis at enrolment (or dapsone if unable to take co-trimoxazole). Caregivers gave written consent; older children aware of their HIV status also gave assent or consent following national guidelines. The trial was approved by Research Ethics Committees in Zambia, Uganda, and the UK.

on or initiated co-trimoxazole prophylaxis at enrolment (or dapsone if unable to take co-trimoxazole). Caregivers gave written consent; older children aware of their HIV status also gave assent or consent following national guidelines. The trial was approved by Research Ethics Committees in Zambia, Uganda, and the UK. Randomisation and masking Children were randomly assigned (1:1:1) to receive open-label stavudine, zidovudine, or abacavir, together with lamivudine and either nevirapine or efavirenz (at treating paediatrician's discretion: all <3 years received nevirapine). Randomisation was stratified by age (younger than 5 years vs 5 years or older), previous ART (naive vs experienced), NNRTI (nevirapine vs efavirenz), and clinical centre. A computer-generated sequential randomisation list, using the urn probability method15 was prepared by the trial statistician and incorporated securely into the trial database at each centre. The list was concealed until allocation, which occurred after eligibility was confirmed by local centre staff, who then did the randomisation.

erated sequential randomisation list, using the urn probability method15 was prepared by the trial statistician and incorporated securely into the trial database at each centre. The list was concealed until allocation, which occurred after eligibility was confirmed by local centre staff, who then did the randomisation. Procedures Scored dispersible fixed-dose combinations of abacavir plus lamivudine, zidovudine plus lamivudine, zidovudine plus lamivudine plus nevirapine, stavudine plus lamivudine, and stavudine plus lamivudine plus nevirapine as so-called baby and junior tablets (Cipla Pharmaceuticals, Mumbai, India) were prescribed following WHO weight bands5 (stavudine at lower doses than previous studies6, 7, 8 at 2–4 mg/kg [<10 kg] and at 1·4–2·4 mg/kg [>10 kg] daily). Efavirenz (600 mg double-scored, allowing daily doses of 200 mg, 300 mg, 400 mg, 500 mg, and 600 mg) and nevirapine (200 mg scored) were also supplied for children taking dual NRTI fixed-dose combinations.

(stavudine at lower doses than previous studies6, 7, 8 at 2–4 mg/kg [<10 kg] and at 1·4–2·4 mg/kg [>10 kg] daily). Efavirenz (600 mg double-scored, allowing daily doses of 200 mg, 300 mg, 400 mg, 500 mg, and 600 mg) and nevirapine (200 mg scored) were also supplied for children taking dual NRTI fixed-dose combinations. Children exited the trial from Oct 30, 2013, to Jan 23, 2014, after a minimum of 96 weeks follow-up. At nurse (6-weekly) and doctor (12-weekly) visits, children were examined, medical history was recorded, adherence was assessed (self-report), and ART was dispensed. At weeks 6, 12, 24, and then 24-weekly, five skinfold thicknesses (triceps, biceps, sub-scapular, supra-iliac, and mid-thigh) and five body circumferences (waist, hip, mid-thigh, mid-upper-arm [MUAC], and torso) were measured to assess lipodystrophy (mean of three measurements); haematology, biochemistry, and CD4 tests were done (results available to clinicians); and plasma was stored for retrospective viral load and resistance testing (results not available to clinicians in real time). Substitutions for toxicity and switches to second-line for failure were at the treating physician's discretion, following WHO guidelines.5

CD4 tests were done (results available to clinicians); and plasma was stored for retrospective viral load and resistance testing (results not available to clinicians in real time). Substitutions for toxicity and switches to second-line for failure were at the treating physician's discretion, following WHO guidelines.5 Outcomes The primary outcome was grade 2 or greater clinical adverse events, confirmed grade 3 laboratory adverse events, or any grade 4 laboratory adverse events16 (neutrophils17). Clinical primary endpoints were adjudicated against protocol-defined criteria by an endpoint review committee (ERC), masked to allocation, and were also adjudicated for relation to antiretroviral drugs, without knowing the specific ART received. Secondary toxicity outcomes were specific subsets of the primary endpoints (anaemia, neutropenia, lipodystrophy or lipoatrophy, hypersensitivity [also including grade 1 events]), serious adverse events, ART-modifying toxicity (any grade), grade 3/4 adverse events possibly, probably, or definitely related to zidovudine or abacavir or stavudine, and changes in skinfold-thicknesses-for-age and body-circumference-for-age. Secondary efficacy outcomes were viral load suppression, clinical disease progression, change in weight-for-age, height-for-age, and CD4 and ART adherence. Laboratory measures, including viral load, were assayed blind to randomisation. HIV-1 viral load was assayed with the Roche COBAS Ampliprep/Taqman version 2.0 in both Uganda (Joint Clinical Research Centre [JCRC]) and Zambia (Centre for Infectious Disease Research in Zambia [CIDRZ]). Because of small stored sample volumes, most samples were run with a 1/5 dilution with Basematrix 53, giving a lower limit of detection of 100 copies per mL. Drug resistance genotyping was done with either in-house primers (JCRC) or primers from Inqaba Biotec (CIDRZ), with both laboratories using an automated ABI 3730xl sequencer.

mall stored sample volumes, most samples were run with a 1/5 dilution with Basematrix 53, giving a lower limit of detection of 100 copies per mL. Drug resistance genotyping was done with either in-house primers (JCRC) or primers from Inqaba Biotec (CIDRZ), with both laboratories using an automated ABI 3730xl sequencer. Statistical analysis Recruiting 470 children gave 85% power to detect a reduction from 20% to 10% in the cumulative incidence of the primary endpoint across the three randomised groups (two-sided α=0·05; appendix p 2). Interim data were reviewed by an independent data monitoring committee (two meetings, approximately annually) using the Haybittle-Peto criterion (p<0·001). Randomised groups were compared with intention-to-treat analysis with log-rank tests for time-to-event outcomes, exact tests for binary outcomes, and generalised estimating equations with independent working correlation for global tests of repeated measures. Analyses were stratified by age group, naive or experienced, and NNRTI, but not by clinical centre because this was not expected to affect outcome (randomisation was stratified by centre for practical reasons; see appendix p 3 for more detail). Data were analysed with Stata version 13.1. This trial is registered with the ISRCTN Registry number, 69078957.

Statistical analysis Recruiting 470 children gave 85% power to detect a reduction from 20% to 10% in the cumulative incidence of the primary endpoint across the three randomised groups (two-sided α=0·05; appendix p 2). Interim data were reviewed by an independent data monitoring committee (two meetings, approximately annually) using the Haybittle-Peto criterion (p<0·001). Randomised groups were compared with intention-to-treat analysis with log-rank tests for time-to-event outcomes, exact tests for binary outcomes, and generalised estimating equations with independent working correlation for global tests of repeated measures. Analyses were stratified by age group, naive or experienced, and NNRTI, but not by clinical centre because this was not expected to affect outcome (randomisation was stratified by centre for practical reasons; see appendix p 3 for more detail). Data were analysed with Stata version 13.1. This trial is registered with the ISRCTN Registry number, 69078957. Role of the funding source The funder of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

This trial is registered with the ISRCTN Registry number, 69078957. Role of the funding source The funder of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Between Nov 8, 2010, and Dec 28, 2011, 480 children were randomly assigned: 156 to stavudine, 159 to zidovudine, and 165 to abacavir. After two were excluded due to randomisation error, 156 children were analysed in the stavudine group, 158 in the zidovudine group, and 164 in the abacavir group. More children were ART naive (365 [76%]) than ART experienced (113 [24%]); more were younger than 5 years (337 [71%]); and consequently more received nevirapine (353 [74%]) than efavirenz (more similar percentages >3 years received nevirapine (155 [57%]) and efavirenz (116 [43%]; table 1).

nd 164 in the abacavir group. More children were ART naive (365 [76%]) than ART experienced (113 [24%]); more were younger than 5 years (337 [71%]); and consequently more received nevirapine (353 [74%]) than efavirenz (more similar percentages >3 years received nevirapine (155 [57%]) and efavirenz (116 [43%]; table 1). Baseline characteristics were well balanced between randomised groups (table 1). ART-naive children were substantially younger than ART-experienced children (median 2·6 years [IQR 1·6–4·0] vs 6·2 years [5·5–7·2], with lower CD4% (median 20% [IQR 13–25] vs 35% [30–39]). Median retrospectively assayed viral load was 270 670 copies per mL in ART-naive children (79% >100 000 copies per mL), with three (1%) confirmed less than 100 copies per mL at both screening and enrolment (carers reported no previous ART, no previous samples available). ART-experienced children (all <50 copies per mL) had taken stavudine-based ART for median 3·5 years (IQR 2·6–4·2). The mother or child had received nevirapine or NRTIs for prevention of mother-to-child transmission in 56 (15%) ART-naive and nine (8%) ART-experienced children (table 1). Median follow-up was 2·3 years among children completing the study (range 1·8–3·1; total 1057 child-years). 25 (5%) children were lost (last seen before October, 2013), including eight (2%) who withdrew consent. 8967 (98%) of 9143 scheduled nurse visits were completed.

Baseline characteristics were well balanced between randomised groups (table 1). ART-naive children were substantially younger than ART-experienced children (median 2·6 years [IQR 1·6–4·0] vs 6·2 years [5·5–7·2], with lower CD4% (median 20% [IQR 13–25] vs 35% [30–39]). Median retrospectively assayed viral load was 270 670 copies per mL in ART-naive children (79% >100 000 copies per mL), with three (1%) confirmed less than 100 copies per mL at both screening and enrolment (carers reported no previous ART, no previous samples available). ART-experienced children (all <50 copies per mL) had taken stavudine-based ART for median 3·5 years (IQR 2·6–4·2). The mother or child had received nevirapine or NRTIs for prevention of mother-to-child transmission in 56 (15%) ART-naive and nine (8%) ART-experienced children (table 1). Median follow-up was 2·3 years among children completing the study (range 1·8–3·1; total 1057 child-years). 25 (5%) children were lost (last seen before October, 2013), including eight (2%) who withdrew consent. 8967 (98%) of 9143 scheduled nurse visits were completed. Initial ART followed randomisation for 473 (99%) children (figure 1). 445 (93%) remained on their initial treatment throughout follow-up. 33 first-line ART changes occurred among 30 (6%) children: ten (6%) allocated stavudine, 16 (10%) allocated zidovudine, and four (2%) allocated abacavir (p=0·02). Nine changes (three stavudine, four zidovudine, and two abacavir) were nevirapine substitutions for rifampicin-based tuberculosis co-treatment, 14 were nevirapine or NRTI toxicity substitutions, and ten were mostly dispensing errors. Five children ([1%]; all ART-naive, one [1%] stavudine, two [1%] zidovudine, two [1%] abacavir; p=1·0) switched to second-line ART (two clinical, three immunological or virological failure). There was no evidence that self-reported adherence (proportion reporting missing ART doses in the last 4 weeks) across visits through 96 weeks differed between randomised groups (p=0·82).

o [1%] zidovudine, two [1%] abacavir; p=1·0) switched to second-line ART (two clinical, three immunological or virological failure). There was no evidence that self-reported adherence (proportion reporting missing ART doses in the last 4 weeks) across visits through 96 weeks differed between randomised groups (p=0·82). 917 grade 2–4 clinical or grade 3/4 laboratory adverse events (877 clinical; 40 laboratory) occurred in 312 children (104 [67%] children allocated stavudine, 103 [65%] children allocated zidovudine, and 105 [64%] children allocated abacavir; p=0·63; figure 2, table 2; appendix p 7). Events were more common in younger ART-naive children than in ART-experienced children (figure 2A), but there was no evidence of heterogeneity in differences between randomised groups (p=0·41). 634 clinical events were grade 2 (481 non-serious respiratory tract infections); excluding grade 2 events gave similar results (p=0·48; figure 2B). 199 serious adverse events occurred in 132 (28%) children, with no difference between randomised groups (p=0·46; table 2). Six (4%) children allocated stavudine, 12 (8%) allocated zidovudine, and five (3%) allocated abacavir had grade 3/4 adverse events judged by the ERC (masked to randomisation) to have at least a possible relation to one of the randomised NRTIs (p=0·10; table 2). No grade 3/4 adverse events or serious adverse events were judged definitely or probably related to stavudine, zidovudine, or abacavir.

(3%) allocated abacavir had grade 3/4 adverse events judged by the ERC (masked to randomisation) to have at least a possible relation to one of the randomised NRTIs (p=0·10; table 2). No grade 3/4 adverse events or serious adverse events were judged definitely or probably related to stavudine, zidovudine, or abacavir. 14 (3%) children modified ART for toxicity; with significantly more in the zidovudine group (p=0·03; table 2) where eight children substituted zidovudine with stavudine or abacavir for anaemia (n=4), neutropenia (n=3), or leucopenia (n=1). However, there was no evidence of differences between groups in grade 3/4 anaemia (p=0·42 overall; pairwise p>0·25), although more grade 3/4 neutropenia occurred in the zidovudine group (p=0·04 overall; zidovudine vs stavudine p=0·03, zidovudine vs abacavir p=0·06, stavudine vs abacavir p=0·79). Three children substituted ART (all nevirapine to lopinavir plus ritonavir) for hypersensitivity reactions (one grade 2, one grade 3, one grade 4 [Stevens-Johnson; recovered]; none on abacavir). Masked to NRTI received, the ERC adjudicated five stavudine, one zidovudine, and two abacavir primary endpoints as grade 2–4 hypersensitivity reactions (p=0·21; table 2, supplementary tables 1 and 2; appendix p 7); however, neither child on abacavir stopped the drug with no adverse consequences. One additional grade 1 hypersensitivity reaction was reported in the abacavir group; this child also continued abacavir without adverse effects.

2–4 hypersensitivity reactions (p=0·21; table 2, supplementary tables 1 and 2; appendix p 7); however, neither child on abacavir stopped the drug with no adverse consequences. One additional grade 1 hypersensitivity reaction was reported in the abacavir group; this child also continued abacavir without adverse effects. Two ART-experienced children substituted stavudine with abacavir after developing facial lipoatrophy (grade 2 [boy, age 6 years, 2·5 years on stavudine]; grade 3 [boy, age 8 years, 5 years on stavudine]). Body circumference increased with time at all measured sites, as expected, while the five skinfold thicknesses decreased similarly in ART-naive and ART-experienced children (appendix p 12, 19), with few differences between randomised groups (appendix p 5). There was no evidence that randomised groups differed in body circumference or skinfold thickness ratios or the sum of the four skinfolds (p>0·1; table 3), or in changes in total cholesterol, LDL, HDL, or triglycerides (p>0·4; appendix p 14).

dix p 12, 19), with few differences between randomised groups (appendix p 5). There was no evidence that randomised groups differed in body circumference or skinfold thickness ratios or the sum of the four skinfolds (p>0·1; table 3), or in changes in total cholesterol, LDL, HDL, or triglycerides (p>0·4; appendix p 14). Disease progression was rare and similar across randomised groups (p>0·3; table 2). All 19 deaths, and 12 of the 14 WHO stage 3 or 4 events, occurred in ART-naive children (seven pneumonia, three tuberculosis WHO stage 3/4 events). Nine of 19 deaths and five of 14 WHO 3/4 events occurred less than 12 weeks after ART initiation, related to pre-enrolment disease severity. There was very little evidence of drug-related mortality (appendix p 15). Change in weight-for-age, height-for-age, or body-mass index-for-age to 96 weeks did not differ significantly between groups (p>0·2).

f 14 WHO 3/4 events occurred less than 12 weeks after ART initiation, related to pre-enrolment disease severity. There was very little evidence of drug-related mortality (appendix p 15). Change in weight-for-age, height-for-age, or body-mass index-for-age to 96 weeks did not differ significantly between groups (p>0·2). Most ART-naive children achieved viral load less than 400 copies per mL by 48 weeks (figure 3A), with no differences between randomised groups (p=0·58; figure 3; appendix p 16). Viral load less than 400 copies per mL was maintained at 48 weeks by more than 96% ART-experienced children (p=1·0). Results were similar between groups at 96 weeks in ART-naive and ART-experienced children (p>0·4), as was viral load suppression less than 100 copies per mL at 48 weeks and 96 weeks (figure 3B). Among ART-naive children, 48-week suppression was better in those with viral load less than 100 000 copies per mL at enrolment (66 [93%] of 71 vs 202 [80%] of 254), consistently across randomised groups with no evidence that any NRTI had superior performance in these strata (pinteraction=0·85). 48-week suppression was similar in ART-naive children aged 3 years or older at enrolment receiving nevirapine (42 [89%] of 47) and efavirenz (88 [91%] of 97), also with no evidence of variation across randomised groups (pinteraction=0·25). There was also no evidence that 48-week viral load suppression in ART-naive children older than 1 year varied by previous prevention of mother-to-child transmission exposure to nevirapine without NRTI cover (34 [79%] of 43) versus those who had not (219 [85%] of 258; pinteraction=0·09). There was no evidence of differential CD4% recovery across randomised groups (p=0·09; appendix p 16).

-naive children older than 1 year varied by previous prevention of mother-to-child transmission exposure to nevirapine without NRTI cover (34 [79%] of 43) versus those who had not (219 [85%] of 258; pinteraction=0·09). There was no evidence of differential CD4% recovery across randomised groups (p=0·09; appendix p 16). Resistance mutations were assayed in 58 (84%) of 69 children with viral load greater than 500 copies per mL at 96 weeks (19 allocated stavudine, 22 allocated zidovudine, and 17 allocated abacavir; remaining samples failed). Seven children (five allocated stavudine, one allocated zidovudine, and one allocated abacavir) had no NNRTI or NRTI mutations. As expected, M184V and NNRTI mutations were common in all groups, thymidine-analogue mutations (TAMs) were common in stavudine and zidovudine groups (although TAM-1 41L/210W/215Y were only seen in the zidovudine group), and 74V/115F mutations were common in the abacavir group (appendix p 18). However, only one K65R mutation was identified in the abacavir group. In the abacavir group, sensitivity to second-line NRTI options was 100% for zidovudine and 94% for tenofovir. In the zidovudine and stavudine groups, sensitivity to tenofovir remained high (86% and 100%, respectively; p=0·22 across randomised NRTIs; appendix p 21), but, as expected, was lower for their alternative second-line NRTI abacavir (64% and 89%, respectively; p=0·008 comparing susceptibility to the non-tenofovir second-line NRTI option across randomised groups).

itivity to tenofovir remained high (86% and 100%, respectively; p=0·22 across randomised NRTIs; appendix p 21), but, as expected, was lower for their alternative second-line NRTI abacavir (64% and 89%, respectively; p=0·008 comparing susceptibility to the non-tenofovir second-line NRTI option across randomised groups). Discussion In the first African paediatric trial comparing three NRTIs coformulated in NNRTI/NRTI generic fixed-dose-combination paediatric tablets, dosed using WHO drug ratios and weight bands,2, 5 we identified no major differences in any adverse event or toxicity endpoint during nearly 2·5 years follow-up in ART-naive and ART-experienced children. First-line drug substitutions occurred in only 6% of children, with nearly one-third due to starting anti-tuberculosis treatment. ART-naive children had good clinical, immunological, and virological responses, regardless of backbone NRTI; CD4 cell count and virological responses were maintained among almost all ART-experienced children. As expected, most deaths occurred early in children starting ART and only 1% switched to second-line therapy.

. ART-naive children had good clinical, immunological, and virological responses, regardless of backbone NRTI; CD4 cell count and virological responses were maintained among almost all ART-experienced children. As expected, most deaths occurred early in children starting ART and only 1% switched to second-line therapy. Paediatricians have long debated the relative advantages and disadvantages of different so-called backbone NRTIs combined with lamivudine, particularly because harmonising with adult tenofovir-based once-daily ART is not possible because of concerns about bone toxicity in growing children and absence of paediatric fixed-dose combinations or doses in those younger than 2 years. In the past decade, WHO guidelines have promoted paediatric fixed-dose combinations, first used in the CHAPAS-1 trial18 and licensed in 2007. However, preferred NRTI recommendations have changed from stavudine (2006) to zidovudine (20104) to abacavir (20135), based on minimal paediatric data and no randomised trials.

. In the past decade, WHO guidelines have promoted paediatric fixed-dose combinations, first used in the CHAPAS-1 trial18 and licensed in 2007. However, preferred NRTI recommendations have changed from stavudine (2006) to zidovudine (20104) to abacavir (20135), based on minimal paediatric data and no randomised trials. 91% of children needing ART live in Africa, where genetic and environmental factors determine the relative effect of different ART toxicity profiles. We found no major differences across randomised NRTIs in grade 2–4 clinical or grade 3/4 laboratory adverse events, in either ART-naive or ART-experienced children. The only grade 3/4 event with marginally increased frequency was neutropenia in children allocated zidovudine; its significance is uncertain because African children have low neutrophil counts,19 and it rarely led to zidovudine substitution. As previously described,20 haemoglobin increased regardless of backbone NRTI, and severe anaemia occurred no more frequently in children who received zidovudine versus those who received stavudine or abacavir, suggesting HIV-related rather than drug-related cause. However, although infrequent, drug substitution was more common in the zidovudine group than both other groups, as was also reported in the ARROW trial,20 mainly for anaemia. These combined trial results reassure clinicians that zidovudine substitution is rarely needed for anaemia among children on ART. However, an important caveat is that severe anaemia and neutropenia were an exclusion criteria in both trials; if anaemia is HIV related, initiating zidovudine might also lead to good haemoglobin responses in anaemic children, as observed here, but we did not assess this.

arely needed for anaemia among children on ART. However, an important caveat is that severe anaemia and neutropenia were an exclusion criteria in both trials; if anaemia is HIV related, initiating zidovudine might also lead to good haemoglobin responses in anaemic children, as observed here, but we did not assess this. Clinical lipodystrophy was not recorded up to 3 years follow-up of children aged younger than 5 years at ART initiation. Absence of blinding cannot rule out ascertainment bias, but lack of significant differences in body circumferences or skinfold thicknesses between NRTIs supports anecdotally reported rarity of lipodystrophy among young children, and suggests that longer-term consequences of stavudine exposure in young children are likely to be limited. We also found no evidence of a difference between NRTIs in changes in lipids on ART. Nevertheless, lipodystrophy undoubtedly occurs in older children and adolescents; the only lipodystrophy noted during the trial was facial in two older ART-experienced children already taking stavudine for more than 2·5 years. For this reason, and despite little evidence of harm in young children, the WHO 2013 recommendation that stavudine should be used only where other drugs are unavailable seems reasonable because it harmonises with adult and adolescent recommendations where evidence is strong. However, our results suggest that stavudine could be safely used for at least 2 years in young children (eg, with severe anaemia at ART initiation), if alternatives are not available, supporting WHO5 and the European Medicines Agency who recommended that stavudine for children should not be discontinued completely.

ng. However, our results suggest that stavudine could be safely used for at least 2 years in young children (eg, with severe anaemia at ART initiation), if alternatives are not available, supporting WHO5 and the European Medicines Agency who recommended that stavudine for children should not be discontinued completely. Despite no HLA-B5701 testing, no hypersensitivity reactions to abacavir were observed, in agreement with previous data reporting its rarity in African adults21 and children.10 The only three hypersensitivity reactions leading to a change in ART were substitutions from nevirapine to lopinavir plus ritonavir, albeit at a lower rate than in adults,22 consistent with previous paediatric reports.18 Reassuringly, and providing the first randomised data in children, a CHAPAS-3 substudy showed no difference in cardiovascular measurements or biomarkers between randomised NRTI groups.23, 24 One limitation is that our trial recruited more ART-naive and fewer ART-experienced children than was planned, reducing the power to detect differences between these subgroups, although no major interactions were identified.

o difference in cardiovascular measurements or biomarkers between randomised NRTI groups.23, 24 One limitation is that our trial recruited more ART-naive and fewer ART-experienced children than was planned, reducing the power to detect differences between these subgroups, although no major interactions were identified. When this trial was designed, the major questions related to toxicity profiles of the three NRTIs, with concerns over the potency of abacavir12, 13 only arising later. However, 478 children still provided good power to detect 10–15% differences in viral load suppression. CD4 recovery and retrospectively assayed viral load suppression to less than 100 copies per mL, less than 400 copies per mL, or less than 1000 copies per mL (data not shown) did not differ by randomised NRTI (appendix p 16). Overall suppression was better in ART-experienced than in ART-naive children, as expected, because ART-experienced children were suppressed at enrolment. Similarly to ARROW, there were no interactions suggesting differences in viral load suppression by NRTIs by age (<3 years vs >3 years),20 and, also in agreement with other reports, there was no evidence that viral load suppression depended on intrauterine nevirapine exposure beyond infancy25 or NNRTI in children older than 3 years.26 Although many seminal trials have been done in HIV-infected children by the IMPAACT/PACTG group, their randomised comparisons of combination therapy have focused on the third (non-NRTI) drug (eg, PACTG-1060 nevirapine vs lopinavir plus ritonavir), older drugs (eg, PACTG-327 didanosine), or on receiving an additional NRTI (eg, PACTG-300). No IMPAACT/PACTG trial has directly compared abacavir, zidovudine, or stavudine head-to-head within combination therapy. Our results differ from the only previous randomised, smaller trial of zidovudine versus abacavir (PENTA-5), which showed virological superiority of abacavir versus zidovudine over 5 years in children in well-resourced settings.27, 28 However, children received two NRTIs alone or with nelfinavir; with a potent third drug, as in CHAPAS-3, any superiority of abacavir over zidovudine could well be masked.

sus abacavir (PENTA-5), which showed virological superiority of abacavir versus zidovudine over 5 years in children in well-resourced settings.27, 28 However, children received two NRTIs alone or with nelfinavir; with a potent third drug, as in CHAPAS-3, any superiority of abacavir over zidovudine could well be masked. Our results provide reassurance following recent observational analyses reporting poorer virological responses to abacavir versus stavudine in South African children.12, 13 Possible explanations for the difference include unmeasured confounding or drug–drug interactions between abacavir and lopinavir plus ritonavir (the standard third drug in South Africa).29 Of interest, we did not find that abacavir did worse in children with higher viral loads in CHAPAS-3, but only 24 ART-naive children were younger than 1 year, by contrast with the South African studies where many were younger than 1 year with high viral loads. The contribution of the fixed-dose combination rather than separate pills to virological success is difficult to estimate, but cannot affect our within-trial comparisons as all were using fixed-dose combinations.

year, by contrast with the South African studies where many were younger than 1 year with high viral loads. The contribution of the fixed-dose combination rather than separate pills to virological success is difficult to estimate, but cannot affect our within-trial comparisons as all were using fixed-dose combinations. Finally, these first randomised resistance data in African children on different NRTI plus NNRTI first-line ART reassuringly show that most children remained susceptible to second-line NRTIs over the medium term, regardless of initial NRTI. In particular, while those taking first-line zidovudine had significantly reduced susceptibility to abacavir second-line, those taking first-line abacavir retained high susceptibility to zidovudine; both retained high susceptibility to tenofovir, increasingly used in children older than 10 years who weigh more than 35 kg.

particular, while those taking first-line zidovudine had significantly reduced susceptibility to abacavir second-line, those taking first-line abacavir retained high susceptibility to zidovudine; both retained high susceptibility to tenofovir, increasingly used in children older than 10 years who weigh more than 35 kg. At trial closure (before trial results were known), all carers and children were offered continuing follow-up in the research trial centres (but without the transport refund provided by the trial) or moving to an ART programme site closer to where they lived. Children moving to ART programme sites were moved onto the ART regimen provided by the site (predominantly zidovudine at trial closure, abacavir for some Ugandan sites) to ensure that the ART programme site could continue to provide uninterrupted ART, in terms of drug provision and forecasting. Children staying at the research sites could continue their randomised regimen, because there was no reason to change drugs in children doing well and stable on a WHO recommended regimen, and being carefully followed for toxicity. However, although stavudine remains an option for children not able to take other NRTIs in 20104 and 20135 WHO guidelines, at trial closure Uganda national guidelines no longer recommended stavudine for children (previously, stavudine was an option for first-line ART in children <5 years). In Zambia, guidelines were based on duration on stavudine, with age being also used more recently. The recommended substitutions were therefore on a case-by-case basis. As a result of all these factors, as well as reduced demand for stavudine-based products by programmes (and hence scarcity from manufacturers), almost all children moved off stavudine at trial closure.

avudine, with age being also used more recently. The recommended substitutions were therefore on a case-by-case basis. As a result of all these factors, as well as reduced demand for stavudine-based products by programmes (and hence scarcity from manufacturers), almost all children moved off stavudine at trial closure. In conclusion, CHAPAS-3 shows primarily that children respond well to all NRTI/NNRTI recommended fixed-dose combinations in 2013 WHO guidelines with minimal drug toxicity. Most primary endpoints were morbid events, showing the very small contribution of antiretroviral toxicity to managing the HIV-infected child. The population was generally young, with early disease, and hence highly generalisable to increasing numbers entering ART programmes under universal treatment for those younger than 5 years. The fixed-dose combinations have different advantages and disadvantages in terms of number and frequency of tablets, cost, and availability as dual or triple drug fixed-dose combinations. Abacavir has very low toxicity in African children, a superior resistance profile for second-line NRTI sequencing, and is the only once-daily licensed NRTI fixed-dose combination (with lamivudine) for children, supporting its preferred use in first-line ART.5 Its only disadvantage is that it has a higher cost than zidovudine and stavudine (US$0·09 per baby tablet vs $0·05 for zidovudine and $0·03 for stavudine).14 A WHO survey in 2014 showed that paediatric use of abacavir was increasing (34%), whereas stavudine was decreasing (12%); zidovudine was 51% and also decreasing, thus data strongly arguing for further abacavir price reductions. Potential future triple abacavir-based combinations with efavirenz or dolutegravir could further simplify and improve durability of once-daily first-line ART for children who will need ART for much longer than adults.

ovudine was 51% and also decreasing, thus data strongly arguing for further abacavir price reductions. Potential future triple abacavir-based combinations with efavirenz or dolutegravir could further simplify and improve durability of once-daily first-line ART for children who will need ART for much longer than adults. Supplementary Material Supplementary appendix Acknowledgments This study was funded by the European Developing Countries Clinical Trials Partnership (IP.2007.33011.006), Medical Research Council UK, Department for International Development UK, and Ministerio de Sanidady Consumo Spain. Cipla Ltd donated first-line antiretrovirals. We thank all the children, carers, and staff from all the centres participating in the CHAPAS-3 trial. Contributors DMG, ASW, VMul, VMus, AK, DB, HM, CChi, MJT, and CK designed the study. VMul, CChi, and CCha did the study in Zambia, and VMus, AK, GA, GM, AA, and CK did the study in Uganda. DMG, MJT, EO-P, JK, NK, AC, and ASW coordinated the study from the UK. ADC and ASW wrote the trial analysis plan, which all authors then reviewed; and ADC and ASW did the analyses. All authors contributed to interpretation of the data. DMG, ADC, and ASW wrote the first draft of the report. All authors revised the report critically and approved the final version.

ted the study from the UK. ADC and ASW wrote the trial analysis plan, which all authors then reviewed; and ADC and ASW did the analyses. All authors contributed to interpretation of the data. DMG, ADC, and ASW wrote the first draft of the report. All authors revised the report critically and approved the final version. The CHAPAS-3 trial group University Teaching Hospital, Lusaka, Zambia—Chifumbe Chintu, Veronica Mulenga, Desiree Kabamba, Dorothy Kavindele, Chishala Chabala, Musaku Mwenechanya, Monica Kapasa, Caroline C Zulu, Mox Kalumbi, Elias Chambula, Joyce Lungu, Marjory N Liusha, Dorothy Zangata, Dorica Masuka, Elias Chambula, Shadreck Chanshi, Terence Chipoya, Semy Zulu, Daniel Chola, Betty Chanda, Steven Malama, Chama Chama, Sylvia Mulambo, Mpala Mwanza. Baylor Centre of Excellence at Mulago Hospital, Kampala, Uganda—R Alice Asiimwe, J Vicent Tukei, Violet Korutaro, Justine Komunyena, Isaac Sebuliba, Muzamil Kisekka, Carolyn Nansubuga, N Justine Mpanga, Moses Matovu, Charles Okello, Sharon Kesande, Gladys Namutebi, E Glorius Tumuheirirwe, Immaculate Nagawa, Sarah Nakimera, Geoffrey Onen, Fatuma Kabasita, Fred Sunday, Dick Isabirye. Joint Clinical Research Centre, Kampala, Uganda—Cissy Kityo, Victor Musiime, Grace Mirembe, Elizabeth Kaudha, Amos Drasiku, Bernard Bainomuhwezi, Priscilla Wavamunno, Florence Odongo, Constance Lukowe, Winnie Namala, Daniel Sseremba, Alison Balaba, Alice Kwaga, Joshua Kayiwa, Matthew Odera, Paul Oronon, Edith Bagurukira, Phyllis Mwesigwa, Philip Apugulu, Lincoln Mugarura, Eram David Williams, Denis Odoch, Immaculate Nankya, Emmanuel Ndashimyeeva, Eva Nabulime. Joint Clinical Research Centre, Gulu, Uganda—George Abongomera, James Abach, Willy Agings Odong, Beatrice Arach, Irene Claren Aciro, Joseph Omongin, Geoffrey Amone, Peter Okello, Philliam Aleti, Edward Otim, Patrick Kidega, Emmanuel Achol. TASO Gulu—Gladys Aloyo, Robert Alani. Gulu Regional Referral Hospital—Alex Akera, Ciprian Odong. Centre for Infectious Disease Research in Zambia—Mpanji Siwingwa, Innocent Mwape, Joshua Zulu, Gabriel Chipili, Linda Chibesa. MRC Clinical Trials Unit at UCL, London, UK—Diana M Gibb, A Sarah Walker, Margaret J Thomason, Adrian Cook, Ellen Owen-Powell, Alex Ferrier, David Baptiste, Charlotte Male, Brendan Murphy, Moira Spyer. Institute of Child Health, London, UK—Julia Kenny, Nigel Klein. Radboud University Nijmegen Medical Centre, Nijmegen, Netherlands—David Burger, Quirine Fillekes, Angela Colbers.

Sarah Walker, Margaret J Thomason, Adrian Cook, Ellen Owen-Powell, Alex Ferrier, David Baptiste, Charlotte Male, Brendan Murphy, Moira Spyer. Institute of Child Health, London, UK—Julia Kenny, Nigel Klein. Radboud University Nijmegen Medical Centre, Nijmegen, Netherlands—David Burger, Quirine Fillekes, Angela Colbers. University of Cape Town, Cape Town, South Africa—Helen McIlleron. Trial Steering Committee (independent members)—Elwyn Chomba (chair), Jose Ramos, Zainab Akol, Peter Elyanu, Harriet Nakimuli (community). Data Monitoring Committee—Tim E A Peto (chair), Margaret Siwale James Tumwine. Endpoint Review Committee—Hermione Lyall (chair), Julia Kenny, Diana M Gibb. Declaration of interests We declare no competing interests. Figure 1 Trial profile ART=antiretroviral treatment. *Includes one not seen after randomisation. †One participant started stavudine and substituted zidovudine at 12 weeks, two started abacavir and did not change (both prescribing errors). ‡Two started zidovudine and did not change (one prescribing error and one child changed regimen to match twin sibling). Figure 2 Primary endpoint (clinical adverse event grade 2 or higher, confirmed laboratory grade 3 adverse event, or any laboratory grade 4 adverse event; A) and grade 3 or 4 primary endpoint (B) ART=antiretroviral treatment. Figure 3 Viral suppression in patients with less than 400 copies per mL (A) and viral less than 100 copies per mL (B) Data are the absolute (95% CI) between-group differences in overall suppression. ART=antiretroviral treatment. S=stavudine. Z=zidovudine. A=abacavir. Table 1 Baseline characteristics

ART=antiretroviral treatment. Figure 3 Viral suppression in patients with less than 400 copies per mL (A) and viral less than 100 copies per mL (B) Data are the absolute (95% CI) between-group differences in overall suppression. ART=antiretroviral treatment. S=stavudine. Z=zidovudine. A=abacavir. Table 1 Baseline characteristics Naive Experienced Stavudine (n=123) Zidovudine (n=112) Abacavir (n=130) All (n=365) Stavudine (n=33) Zidovudine (n=46) Abacavir (n=34) All (n=113) Centre UTH, Lusaka, Zambia 30 (24%) 25 (22%) 34 (26%) 89 (24%) 15 (45%) 22 (48%) 15 (44%) 52 (46%) Baylor, Kampala, Uganda 42 (34%) 36 (32%) 41 (32%) 119 (33%) 7 (21%) 8 (17%) 7 (21%) 22 (19%) JCRC, Kampala, Uganda 34 (28%) 36 (32%) 32 (25%) 102 (28%) 11 (33%) 16 (35%) 12 (35%) 39 (35%) JCRC, Gulu, Uganda 17 (14%) 15 (13%) 23 (18%) 55 (15%) 0 0 0 0 Age (years) 2·6 (1·6–4·1) 2·6 (1·7–3·9) 2·7 (1·7–4·0) 2·6 (1·6–4·0) 6·5 (5·9–7·3) 6·0 (5·5–7·2) 5·9 (5·4–7·2) 6·2 (5·5–7·2) Sex Male 66 (54%) 54 (48%) 67 (52%) 178 (49%) 23 (70%) 22 (48%) 14 (41%) 59 (52%) Female 57 (46%) 58 (52%) 63 (48%) 187 (51%) 10 (30%) 24 (52%) 20 (59%) 54 (48%) Z score Weight-for-age −2·3 (1·8) −2·2 (1·6) −1·9 (1·6) −2·1 (1·6) −1·1 (0·7) −1·0 (1·2) −1·4 (1·0) −1·1 (1·0) Height-for-age −2·5 (1·7) −2·6 (1·7) −2·3 (1·7) −2·5 (1·7) −1·6 (0·8) −1·4 (1·2) −1·8 (0·9) −1·6 (1·0) Body-mass index-for-age −0·7 (1·6) −0·4 (1·4) −0·5 (1·5) −0·5 (1·5) −0·1 (0·8) −0·1 (1·0) −0·3 (0·9) −0·2 (0·9) WHO stage* 1 17 (14%) 10 (9%) 14 (11%) 41 (11%) 8 (24%) 10 (22%) 7 (21%) 25 (22%) 2 45 (37%) 46 (41%) 48 (37%) 139 (38%) 8 (24%) 9 (20%) 7 (21%) 24 (21%) 3 50 (41%) 41 (37%) 56 (43%) 147 (40%) 8 (24%) 24 (52%) 11 (32%) 43 (38%) 4 11 (9%) 15 (13%) 12 (9%) 38 (10%) 9 (27%) 3 (7%) 9 (26%) 21 (19%) Viral load (copies per mL)† Log10 5·6 (0·7) 5·4 (0·8) 5·3 (0·8) 5·4 (0·8) <50 <50 <50 <50 Absolute 328 320 (191 770–926 170) 252 390 (107 830–808 330) 217 540 (78 760–609 520) 270 670 (116 330–738 360) <50 <50 <50 <50 >100 000 copies per mL 100 (84%) 85 (79%) 95 (74%) 280 (79%) 0 (0%) 0 (0%) 0 (0%) 0 (0%) CD4 cell count CD4% 19% (12–23) 21% (15–26) 19% (11–24) 20% (13–25) 35% (28–39) 35% (30–40) 35% (31–39) 35% (30–39) Absolute CD4 865 (581– 1236) 925 (675–1434) 813 (490–1353) 893 (597–1299) 1143 (987–1414) 1164 (916–1641) 1362 (1072–1656) 1191 (962–1587) Stavudine, years .. .. .. ..

80 (79%) 0 (0%) 0 (0%) 0 (0%) 0 (0%) CD4 cell count CD4% 19% (12–23) 21% (15–26) 19% (11–24) 20% (13–25) 35% (28–39) 35% (30–40) 35% (31–39) 35% (30–39) Absolute CD4 865 (581– 1236) 925 (675–1434) 813 (490–1353) 893 (597–1299) 1143 (987–1414) 1164 (916–1641) 1362 (1072–1656) 1191 (962–1587) Stavudine, years .. .. .. .. 3·0 (2·3–3·6) 3·9 (3·0–4·5) 3·5 (2·5–4·2) 3·5 (2·6–4·2) Any pMTCT received by mother or child 15 (12%) 20 (18%) 21 (16%) 56 (15%) 3 (9%) 3 (7%) 3 (9%) 9 (8%) Nevirapine‡ only 8 (7%) 12 (11%) 18 (14%) 38 (10%) 3 (9%) 3 (7%) 3 (9%) 9 (8%) Nevirapine‡ and NRTI§ 4 (3%) 3 (3%) 2 (2%) 9 (2%) 0 0 0 0 NRTI only§ 3 (2%) 5 (4%) 1 (1%) 9 (2%) 0 0 0 0 Received nevirapine with randomised NRTIs in ART 87 (71%) 75 (67%) 90 (69%) 252 (69%) 29 (88%) 42 (91%) 30 (88%) 101 (89%) Data are n (%), median (IQR), or mean (SD). UTH=University Teaching Hospital. JCRC=Joint Clinical Research Centre. pMTCT=prevention of mother-to-child transmission. NRTI=nucleoside reverse-transcriptase inhibitors. ART=antiretroviral treatment. * Derived from pre-trial WHO event history before and after ART initiation in ART experienced. † ART naive: four missing in stavudine group, four in zidovudine group, and two abacavir group due to samples not being stored. ‡ Single-dose nevirapine to either mother or child or both, or less than 2 days nevirapine to the child (with or without single-dose nevirapine to the mother). § Zidovudine to the child or zidovudine (majority) or zidovudine plus lamivudine to the mother. Table 2 Primary and secondary endpoints (time to event)

† ART naive: four missing in stavudine group, four in zidovudine group, and two abacavir group due to samples not being stored. ‡ Single-dose nevirapine to either mother or child or both, or less than 2 days nevirapine to the child (with or without single-dose nevirapine to the mother). § Zidovudine to the child or zidovudine (majority) or zidovudine plus lamivudine to the mother. Table 2 Primary and secondary endpoints (time to event) Stavudine (n=156); N (%) Zidovudine (n=157) Abacavir (n=164) Abacavir vs zidovudine; HR (95% CI) N (%) HR vs stavudine (95% CI) N (%) HR vs stavudine (95% CI) p value* Primary endpoint adverse event† 104 (67%) 103 (65%) 0·99 (0·75–1·29) 105 (64%) 0·88 (0·67–1·15) 0·63 0·89 (0·68–1·17) Specific subsets of primary endpoint adverse events Anaemia, grade 3/4 5 (3%) 9 (6%) 1·93 (0·64–5·76) 6 (4%) 1·15 (0·35–3·78) 0·42 0·60 (0·21–1·69) Anaemia, grade 4 5 (3%) 7 (4%) 1·45 (0·46–4·57) 3 (2%) 0·57 (0·14–2·38) 0·38 0·39 (0·10–1·52) Neutropenia, grade 3/4 4 (3%) 12 (8%) 3·21 (1·03–9·98) 5 (3%) 1·21 (0·32–4·49) 0·04 0·38 (0·13–1·07) Neutropenia, grade 4 3 (2%) 10 (6%) 3·55 (0·97–12·9) 4 (2%) 1·29 (0·29–5·75) 0·06 0·36 (0·11–1·16) Hypersensitivity reaction‡ 5 (3%) 1 (0·6%) 0·22 (0·03–1·86) 2 (1%) 0·38 (0·07–1·95) 0·21 1·75 (0·16–19·3) Lipodystrophy/lipoatrophy 2 (1%) 0 .. 0 .. 0·08 .. Mitochondrial disease§ 1 (0·6%) 0 .. 1 (0·6%) .. 0·65 .. Grade 3/4 adverse events¶ 46 (29%) 53 (34%) 1·24 (0·83–1·84) 51 (31%) 1·01 (0·68–1·50) 0·48 0·82 (0·56–1·20) Grade 3/4 adverse events adjudicated as NRTI related‖ 6 (4%) 12 (8%) 2·12 (0·79–5·66) 5 (3%) 0·80 (0·25–2·63) 0·10 0·38 (0·13–1·08) Serious adverse events 46 (29%) 44 (28%) 0·98 (0·65–1·48) 42 (26%) 0·78 (0·51–1·19) 0·46 0·80 (0·52–1·22) Serious adverse events adjudicated as NRTI related‖ 8 (5%) 12 (12%) 1·50 (0·61–3·67) 6 (6%) 0·64 (0·22–1·85) 0·22 0·43 (0·16–1·14) Toxicity causing ART modification** 4 (3%) 9 (6%) 2·24 (0·69–7·32) 1 (1%) 0·23 (0·03–2·09) 0·03 0·10 (0·01–0·83) New WHO stage 3 or 4 event or death 9 (6%) 7 (4%) 0·84 (0·31–2·26) 13 (8%) 1·37 (0·58–2·30) 0·55 1·62 (0·65–4·07) Death†† 7 (4%) 3 (2%) 0·48 (0·12–1·85) 9 (5%) 1·23 (0·46–3·29) 0·35 2·56 (0·69–9·45) All HRs were stratified for randomisation stratification factors.

·69–7·32) 1 (1%) 0·23 (0·03–2·09) 0·03 0·10 (0·01–0·83) New WHO stage 3 or 4 event or death 9 (6%) 7 (4%) 0·84 (0·31–2·26) 13 (8%) 1·37 (0·58–2·30) 0·55 1·62 (0·65–4·07) Death†† 7 (4%) 3 (2%) 0·48 (0·12–1·85) 9 (5%) 1·23 (0·46–3·29) 0·35 2·56 (0·69–9·45) All HRs were stratified for randomisation stratification factors. No evidence of interaction between naive versus experienced strata on any outcome in table 2 (p>0·1; 21 tests), except for serious adverse events (p=0·02; naive children with serious adverse events: 46 allocated stavudine, 40 allocated zidovudine, and 39 allocated abacavir; experienced children: none allocated stavudine, four allocated zidovudine, and three allocated abacavir; serious adverse events were most commonly lower respiratory tract infections or other specific infections). HR=hazard ratio. NRTI=nucleoside reverse-transcriptase inhibitors. ART=antiretroviral treatment. * Stratified log-rank test. † Clinical grade 2 or greater, laboratory grade 3 (confirmed), laboratory grade 4 (all). ‡ Includes grade 4 Stevens-Johnson syndrome from appendix p 7; adjudicated blind to actual NRTI received. Both children in the abacavir group continued abacavir without adverse effects. See appendix p 11 for details of adjudicated relation to antiretrovirals. One additional grade 1 hypersensitivity reaction (not a primary endpoint) also occurred in the abacavir group: again the child continued abacavir without adverse effects. § One cardiomyopathy (stavudine group) and one myopathy (abacavir group).

‡ Includes grade 4 Stevens-Johnson syndrome from appendix p 7; adjudicated blind to actual NRTI received. Both children in the abacavir group continued abacavir without adverse effects. See appendix p 11 for details of adjudicated relation to antiretrovirals. One additional grade 1 hypersensitivity reaction (not a primary endpoint) also occurred in the abacavir group: again the child continued abacavir without adverse effects. § One cardiomyopathy (stavudine group) and one myopathy (abacavir group). ¶ Clinical grade 3 or greater, laboratory grade 3 (confirmed), laboratory grade 4 (all). ‖ See appendix p 4 for details of grade 3/4 adverse events and serious adverse events judged by the endpoint review committee as possibly having some relation to any of stavudine or zidovudine or abacavir (all events adjudicated for each drug blind to NRTI actually received, so toxicity from any of the three NRTIs could be attributed to each event). No grade 3/4 adverse events or serious adverse events were judged probably or definitely related. ** Not including changes for tuberculosis treatment (see main text); one toxicity substitution in child randomised to abacavir was actually from zidovudine to nevirapine for anaemia, following previous substitution of nevirapine to zidovudine (triple NRTI regimen) for tuberculosis treatment. Two stavudine and one zidovudine substitutions from nevirapine to lopinavir/ritonavir for rash/hypersensitivity reactions; all other substitutions were from stavudine or zidovudine. †† See appendix p 15 for relation between NRTIs and deaths.

** Not including changes for tuberculosis treatment (see main text); one toxicity substitution in child randomised to abacavir was actually from zidovudine to nevirapine for anaemia, following previous substitution of nevirapine to zidovudine (triple NRTI regimen) for tuberculosis treatment. Two stavudine and one zidovudine substitutions from nevirapine to lopinavir/ritonavir for rash/hypersensitivity reactions; all other substitutions were from stavudine or zidovudine. †† See appendix p 15 for relation between NRTIs and deaths. Table 3 Secondary endpoints (continuous) Stavudine change* Zidovudine change* Difference†(95% CI) Abacavir change* Difference†(95% CI) p value‡ Difference†(95% CI) Growth Weight-for-age 0·91 0·84 0·08 (−0·15 to 0·30) 0·75 −0·06 (−0·32 to 0·10) 0·21 −0·18 (−0·39 to 0·02) Height-for-age 0·62 0·67 0·08 (−0·12 to 0·28) 0·61 0·02 (−0·21 to 0·24) 0·72 −0·06 (−0·27 to 0·15) BMI-for-age 0·63 0·40 0·05 (−0·22 to 0·31) 0·42 −0·13 (−0·41 to 0·15) 0·40 −0·17 (−0·42 to 0·08) Body circumference and skinfolds Waist:hip ratio −0·03 −0·04 −0·01 (−0·02 to 0·00) −0·05 −0·01 (−0·03 to 0·01) 0·33 0·00 (−0·01 to 0·02) Waist:arm ratio −0·03 −0·04 −0·04 (−0·10 to 0·00) −0·05 0·03 (−0·03 to 0·09) 0·13 0·07 (0·00 to 0·13) Torso:arm skinfold ratio −0·02 −0·02 −0·00 (−0·04 to 0·03) −0·03 0·01 (−0·03 to 0·05) 0·87 0·01 (−0·03 to 0·05) Sum of four skinfolds (mm) −1·88 −3·75 −1·13 (−2·86 to 0·60) −2·70 −0·28 (−2·07 to 1·52) 0·42 0·86 (−1·00 to 2·72) * Mean change from baseline at 96 weeks. † Mean difference in change in first 96 weeks from generalised estimating equation model.

Stavudine change* Zidovudine change* Difference†(95% CI) Abacavir change* Difference†(95% CI) p value‡ Difference†(95% CI) Growth Weight-for-age 0·91 0·84 0·08 (−0·15 to 0·30) 0·75 −0·06 (−0·32 to 0·10) 0·21 −0·18 (−0·39 to 0·02) Height-for-age 0·62 0·67 0·08 (−0·12 to 0·28) 0·61 0·02 (−0·21 to 0·24) 0·72 −0·06 (−0·27 to 0·15) BMI-for-age 0·63 0·40 0·05 (−0·22 to 0·31) 0·42 −0·13 (−0·41 to 0·15) 0·40 −0·17 (−0·42 to 0·08) Body circumference and skinfolds Waist:hip ratio −0·03 −0·04 −0·01 (−0·02 to 0·00) −0·05 −0·01 (−0·03 to 0·01) 0·33 0·00 (−0·01 to 0·02) Waist:arm ratio −0·03 −0·04 −0·04 (−0·10 to 0·00) −0·05 0·03 (−0·03 to 0·09) 0·13 0·07 (0·00 to 0·13) Torso:arm skinfold ratio −0·02 −0·02 −0·00 (−0·04 to 0·03) −0·03 0·01 (−0·03 to 0·05) 0·87 0·01 (−0·03 to 0·05) Sum of four skinfolds (mm) −1·88 −3·75 −1·13 (−2·86 to 0·60) −2·70 −0·28 (−2·07 to 1·52) 0·42 0·86 (−1·00 to 2·72) * Mean change from baseline at 96 weeks. † Mean difference in change in first 96 weeks from generalised estimating equation model. ‡ Global test from generalised estimating equations with normally distributed errors and independent covariance. BMI=body-mass index.

Introduction Enteric (typhoid) fever, a systemic infection caused by the Salmonella enterica serovars Typhi and Paratyphi A, B, and C, is a leading cause of febrile disease in many low-income countries. 27 million new infections and more than 200 000 deaths are estimated to be attributable to enteric fever worldwide each year.1, 2 In Kathmandu, the capital of Nepal and the setting of this study, the burden of enteric fever is particularly high, and is the leading cause of febrile bacterial disease in adults and children.3, 4 Research in context Evidence before this study We searched MEDLINE, PubMed, and Scopus without date restrictions for English-language articles with the search terms “randomized controlled trial” (“RCT” or “randomized* control* trial*”) AND “typhoid fever”, “enteric fever”, AND “ceftriaxone”. We also noted relevant articles outlined in a Cochrane review, and a meta-analysis, of fluoroquinolones versus other antimicrobials in the treatment of enteric fever. We identified 11 randomised trials that used intravenous ceftriaxone in one of their treatment groups. The selected trials had small sample sizes (ranging from 15 to 43 patients) and the definitions of outcomes for the primary and secondary outcomes were not standardised. Three trials gave ceftriaxone for 7 days in one of their groups (for 73 patients), the same duration as in our trial, but the drug dose was variable. The mean fever clearance times ranged from 3·9 days to 5·4 days, the number of clinical failures ranged from none to six, and the number of relapses ranged from one to four in these small trials.

xone for 7 days in one of their groups (for 73 patients), the same duration as in our trial, but the drug dose was variable. The mean fever clearance times ranged from 3·9 days to 5·4 days, the number of clinical failures ranged from none to six, and the number of relapses ranged from one to four in these small trials. Added value of this study Our data augment previous findings, predicting that ceftriaxone is safe and effective for the treatment of enteric fever and out-performs gatifloxacin, with only 7% of culture-positive patients failing treatment, and a median fever clearance time of 2·78 days. However, our study, by contrast with the outlined studies, also investigated the clinical outcome in culture-negative patients with suspected enteric fever—in this group, gatifloxacin out-performed ceftriaxone with median fever clearance times of 1·12 days and 3·03, respectively. Furthermore, our work is the first to describe the clinical implications of fluoroquinolone-resistant Salmonella enterica serovar Typhi. Implications of all the available evidence In view of the emergence of fluoroquinolone-resistant S Typhi in this setting and the poor efficacy of ceftriaxone in the culture-negative group, we advocate better diagnostic testing for febrile diseases in low-income countries, and suggest that fluoroquinolones are no longer effective for treatment of enteric fever in Nepal.

of the emergence of fluoroquinolone-resistant S Typhi in this setting and the poor efficacy of ceftriaxone in the culture-negative group, we advocate better diagnostic testing for febrile diseases in low-income countries, and suggest that fluoroquinolones are no longer effective for treatment of enteric fever in Nepal. Resistance and reduced susceptibility to antimicrobials are the major challenges to successful treatment of enteric fever.5, 6 We have previously reported3, 7 a high prevalence of S Typhi and S Paratyphi A strains in Nepal that show resistance against the quinolone nalidixic acid (minimum inhibitory concentration [MIC] ≥256 μg/mL) with a corresponding decreased susceptibility against fluoroquinolones such as ciprofloxacin (MIC ≥0·125 μg/mL).3, 7 Ceftriaxone, a parenteral, third-generation cephalosporin, is a common empirical therapy for febrile disease in endemic enteric fever locations, and is used for the treatment of enteric fever in south Asia and other regions where nalidixic acid-resistant strains predominate. Furthermore, ceftriaxone is also advocated for the treatment of travellers returning with enteric fever from areas of enteric fever endemicity.8, 9 Investigators for three randomised controlled trials have compared fluoroquinolones with ceftriaxone for treatment of enteric fever.10, 11, 12 Their findings generally favoured the fluoroquinolones, but the studies were insufficiently powered (only 15 to 25 patients per treatment group) to reach significance and data for the prevalence of nalidixic acid-resistant strains were not reported.13

uinolones with ceftriaxone for treatment of enteric fever.10, 11, 12 Their findings generally favoured the fluoroquinolones, but the studies were insufficiently powered (only 15 to 25 patients per treatment group) to reach significance and data for the prevalence of nalidixic acid-resistant strains were not reported.13 Intravenous therapy is expensive and difficult to give reliably (particularly to outpatients) in most countries where enteric fever is endemic; therefore, effective oral antimicrobials are more practical for treatment of this disease. Previously, we have shown that even without reported resistance, the oral third-generation cephalosporin, cefixime, did poorly in Nepalese patients with enteric fever—treatment failure was reported in 26 (37%) of 70 patients receiving cefixime versus three (3%) of 88 patients receiving gatifloxacin.14 Conversely, we have also shown in Nepalese and Vietnamese children and adults with uncomplicated enteric fever that the fourth-generation 8-methoxy-fluoroquinolone gatifloxacin is safe and effective despite an increase in prevalence of S Typhi strains with reduced ciprofloxacin susceptibility.14, 15, 16, 17 Therefore, because third-generation cephalosporins are generally associated with slow clinical improvement and high relapse burden,18, 19 and 7 days of oral gatifloxacin is associated with rapid fever clearance (≤4 days) and low relapse burden, we postulated that gatifloxacin is superior to ceftriaxone in treating enteric fever, and did a study to test this hypothesis.

are generally associated with slow clinical improvement and high relapse burden,18, 19 and 7 days of oral gatifloxacin is associated with rapid fever clearance (≤4 days) and low relapse burden, we postulated that gatifloxacin is superior to ceftriaxone in treating enteric fever, and did a study to test this hypothesis. Methods Study design and participants We did an open-label, randomised, controlled, superiority trial at Patan Hospital and the Civil Services Hospital in the Kathmandu valley, Nepal. The study protocol was reviewed and approved by the Ethics Committee of the Nepal Health Research Council and the Oxford Tropical Research Ethics Committee (UK).

participants We did an open-label, randomised, controlled, superiority trial at Patan Hospital and the Civil Services Hospital in the Kathmandu valley, Nepal. The study protocol was reviewed and approved by the Ethics Committee of the Nepal Health Research Council and the Oxford Tropical Research Ethics Committee (UK). We screened children aged 2–13 years and adults aged 14–45 years with suspected enteric fever. The criteria for suspected enteric fever were body temperature at least 38·0°C for 4 days or more without a focus of infection, as assessed by physical examination and laboratory tests, and as previously described.14, 16, 17 Patients were excluded if they were pregnant; had diabetes mellitus, signs of severe infection (eg, obtundation, shock, clinical jaundice, or active gastrointestinal bleeding), or a history of hypersensitivity to either of the trial drugs; or had been given a fluoroquinolone, a third-generation cephalosporin, or a macrolide within the previous week. Patients who had received chloramphenicol, amoxicillin, or co-trimoxazole could be included, provided the treating clinician reported no clinical response. Written, informed consent to participate in the study was required from all patients. For patients younger than 18 years, we obtained written, informed consent from their parent or an adult guardian.

enicol, amoxicillin, or co-trimoxazole could be included, provided the treating clinician reported no clinical response. Written, informed consent to participate in the study was required from all patients. For patients younger than 18 years, we obtained written, informed consent from their parent or an adult guardian. Randomisation and masking We randomly assigned patients (1:1) without stratification to 7 days of treatment with either oral gatifloxacin (10 mg/kg) once per day or intravenous ceftriaxone (60 mg/kg up to a maximum of 2 g for patients aged 2–13 years or 2 g for patients aged ≥14 years) once per day. The randomisation list was computer-generated with blocks of four and six (with equal probability) and maintained by a clinical trials pharmacist. We concealed treatment allocations inside opaque sealed envelopes, which were numbered sequentially to correspond to patient enrolment numbers. Envelopes were kept in a locked drawer and were opened in strictly numerical order by a study clinician (who had previously screened the patients and obtained consent). Treatment allocation was open-label; masking was not possible because of a difference in the administration route of the two drugs.

enrolment numbers. Envelopes were kept in a locked drawer and were opened in strictly numerical order by a study clinician (who had previously screened the patients and obtained consent). Treatment allocation was open-label; masking was not possible because of a difference in the administration route of the two drugs. Procedures Gatifloxacin 400 mg tablets (Square Pharmaceuticals, Bangladesh) were weighed and cut at a dose of 10 mg/kg once per day. Ceftriaxone (Powercef, 1000 mg injection vial, Wockhardt Ltd, India), was injected slowly over 10 min once per day. Patients received the first dose (on day 1) of the study drug in hospital to monitor for anaphylaxis. Patients receiving ceftriaxone were discharged with an intravenous cannula in situ and had a new cannula inserted on day 4 of treatment. Home treatment was monitored by trained community medical auxiliaries (CMAs), as described in previous studies.14, 16, 17 A CMA visited each patient assigned to treatment twice per day for at least 10 days or until the patient was asymptomatic. The CMAs gave the drugs, and recorded drug doses, administration times, oral temperatures, symptoms, and potential adverse effects in a standard case-record form. We measured complete blood count, serum creatinine, liver-function parameters (total bilirubin, aspartate aminotransferase, and alanine aminotransferase), and serum glucose at enrolment and on day 8 of treatment. We did a finger-prick test for glucose each day on days 2–7 after randomisation, and measured random serum glucose on day 8, day 15, and at 1 month.

reatinine, liver-function parameters (total bilirubin, aspartate aminotransferase, and alanine aminotransferase), and serum glucose at enrolment and on day 8 of treatment. We did a finger-prick test for glucose each day on days 2–7 after randomisation, and measured random serum glucose on day 8, day 15, and at 1 month. We took blood from all patients (3 mL from those aged <14 years; 8 mL from those aged ≥14 years) for bacterial culture at enrolment and on day 8 after randomisation if S Typhi or S Paratyphi were isolated at enrolment, or if their symptoms suggested a clinical relapse. We inoculated blood samples from adults into media-containing tryptone soya broth and sodium polyanethol sulphonate, up to a total volume of 50 mL. We used BactecPeds Plus culture bottles (Becton Dickinson, New Jersey, USA) for paediatric blood samples. Culture results were reported for up to 7 days; positive bottles were subcultured onto blood, chocolate, and MacConkey agar, and colonies presumptive of salmonella were identified using standard biochemical tests and serotype-specific antisera (Murex Biotech, Dartford, UK). We measured antimicrobial sensitivities by the modified Kirby-Bauer disc diffusion method with zone size interpretation based on guidelines from the Clinical and Laboratory Standards Institute.20 Antimicrobial discs tested were ceftriaxone (30 μg), ciprofloxacin (5 μg), gatifloxacin (5 μg), and nalidixic acid (30 μg). MICs against these antimicrobials were measured by Etest, according to the manufacturer's recommendations (BioMérieux, France). All patients were requested to attend a clinic at Patan Hospital on day 8, day 15, and 1 month, 3 months, and 6 months after randomisation for clinical assessments and stool culture.

(30 μg). MICs against these antimicrobials were measured by Etest, according to the manufacturer's recommendations (BioMérieux, France). All patients were requested to attend a clinic at Patan Hospital on day 8, day 15, and 1 month, 3 months, and 6 months after randomisation for clinical assessments and stool culture. Outcomes The primary endpoint of this trial was a composite of treatment failure, defined as the occurrence of at least one of the following events: fever clearance time (ie, time from the first dose of a study drug until the temperature fell to ≤37·5°C and remained there for at least 2 days) more than 7 days after treatment initiation; the need for rescue treatment as judged by the treating physician (the recommended rescue treatment was azithromycin; however, any treatment other than the assigned treatment was acceptable); blood-culture positivity for S Typhi or S Paratyphi on day 8 of treatment (microbiological failure); culture-confirmed or syndromic enteric fever relapse within 28 days of treatment initiation; or development of any enteric fever-related complication (eg, clinically significant bleeding, a fall in the patient's Glasgow Coma Score, perforation of the gastrointestinal tract, or admission to hospital) within 28 days after treatment initiation. Time to treatment failure was defined as the time from the first dose of treatment until the date of the earliest failure event; patients without a failure event were censored on day 28 or the date of their last follow-up.

the gastrointestinal tract, or admission to hospital) within 28 days after treatment initiation. Time to treatment failure was defined as the time from the first dose of treatment until the date of the earliest failure event; patients without a failure event were censored on day 28 or the date of their last follow-up. Secondary endpoints were fever clearance time only; time to relapse until day 28 or at any time during follow-up; confirmed and syndromic relapse until day 28; confirmed or syndromic relapse at any time; and faecal carriage of S Typhi or S Paratyphi at 1 month, 3 months, or 6 months after randomisation assessed in culture-positive patients only. We calculated fever clearance times electronically using temperatures recorded twice per day. We treated these times as interval-censored outcomes to show that fever clearance was known to have occurred at some unknown point in the interval from the last febrile temperature assessment until the first afebrile assessment. We treated patients without fever clearance or relapse as right-censored. Safety and adverse events were assessed each day by the CMAs at the patient's home, by giving a symptom questionnaire and simple physical examinaton. Any patient who had unexpected symptoms was assessed by a study clinician in the hospital. Each patient was also seen by the study clinician in the hospital on the scheduled follow-up days and asked about any symptoms, and a physical examination was undertaken to assess for possible adverse events.

physical examinaton. Any patient who had unexpected symptoms was assessed by a study clinician in the hospital. Each patient was also seen by the study clinician in the hospital on the scheduled follow-up days and asked about any symptoms, and a physical examination was undertaken to assess for possible adverse events. Statistical analysis In this study, we aimed to address the hypothesis that gatifloxacin was superior to ceftriaxone. On the basis of our previous data,17 we predicted that about 7% of patients with a positive culture given gatifloxacin would have treatment failure. To detect an increase in the risk of failure by 20% (from 7% to 27%) in the ceftriaxone group with 80% power at the two-sided 5% significance level, and allowing for a 10% loss to follow-up, we calculated that a sample size of 120 culture-positive patients (60 per treatment group) was needed. We assumed a culture-positive rate of at least 40%, and designed the trial to randomly assign 300 patients to treatment.

with 80% power at the two-sided 5% significance level, and allowing for a 10% loss to follow-up, we calculated that a sample size of 120 culture-positive patients (60 per treatment group) was needed. We assumed a culture-positive rate of at least 40%, and designed the trial to randomly assign 300 patients to treatment. For treatment failure, we based the comparison of the absolute risk of treatment failure until day 28 on Kaplan-Meier estimates and corresponding standard errors according to Greenwood's formula.21 We used survival methods for the analysis of the time to treatment failure, fever clearance time, and time to relapse. For the times to treatment failure and relapse, we used the Kaplan-Meier method to calculate the cumulative incidence of events and Cox regression models with treatment group as the only covariate used for comparison between treatment groups. For the interval-censored fever clearance times, we used the non-parametric maximum likelihood estimator (NPMLE) to estimate their distribution, and parametric Weibull accelerated failure time models for the estimation of quantiles of the fever clearance time in each group and for the comparison between groups.21 We based median (IQR) calculations of fever clearance times on models for each treatment group separately, and acceleration factors on models that included treatment as the only covariate.

lure time models for the estimation of quantiles of the fever clearance time in each group and for the comparison between groups.21 We based median (IQR) calculations of fever clearance times on models for each treatment group separately, and acceleration factors on models that included treatment as the only covariate. We undertook all analyses for each of the three main analysis populations: a modified intention-to-treat (ITT) population (consisting of all randomised patients who received at least one dose of study treatment and did not have a confirmed alternative diagnosis), and the subpopulations with either confirmed blood-culture positivity, or blood-culture negativity. Treatment failure and fever clearance time were also assessed in predefined subgroups (age <16 years; age ≥16 years; female; male; recruited before or after April, 2013; MIC against ciprofloxacin <0·12 μg/mL, 0·12–2·00 μg/mL, or >2·00 μg/mL; MIC against gatifloxacin ≤1·00 μg/mL or >1·00 μg/mL; S Typhi infection; and S Paratyphi infection). We tested for heterogeneity of treatment effects between subgroups with a Cox regression model (for analysis of treatment failure) or a Weibull accelerated failure time model (for fever clearance times) that included an interaction term between the treatment and the subgrouping variable. We did all analyses using the statistical software R version 3.0.1,22 based on available data without imputation of missing data. The safety of the trial was overseen by an independent data and safety monitoring board. This trial was registered at ClinicalTrials.gov, number NCT01421693.

nt and the subgrouping variable. We did all analyses using the statistical software R version 3.0.1,22 based on available data without imputation of missing data. The safety of the trial was overseen by an independent data and safety monitoring board. This trial was registered at ClinicalTrials.gov, number NCT01421693. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

nt and the subgrouping variable. We did all analyses using the statistical software R version 3.0.1,22 based on available data without imputation of missing data. The safety of the trial was overseen by an independent data and safety monitoring board. This trial was registered at ClinicalTrials.gov, number NCT01421693. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Between Sept 18, 2011, and July 14, 2014, we screened 725 patients with suspected enteric fever for enrolment (figure 1). The data and safety monitoring board reviewed the outcome data after 100 patients, and then 200 patients, were randomised. At the 200-patient review, the board requested an additional review within 3 months of MICs against ciprofloxacin and gatifloxacin against all bacterial isolates. Data from 109 culture-confirmed patients (and 233 patients in total) were analysed at this additional review. A comparison of treatment failures between treatment groups did not cross the predefined Haybittle-Peto stopping boundary23 of p less than 0·001 (overall or in any subgroup), but the emergence of S Typhi strains with MICs against ciprofloxacin that were greater than 16 μg/mL and against gatifloxacin that were greater than 1 μg/mL, and a significant difference (p=0·0002 for susceptive vs resistant strains in the gatifloxacin group) in treatment response between patients with fluoroquinolone-resistant strains and those with susceptible strains, led the board to recommend stopping the trial on clinical grounds supported by data on the changing in-vitro susceptibility.

fference (p=0·0002 for susceptive vs resistant strains in the gatifloxacin group) in treatment response between patients with fluoroquinolone-resistant strains and those with susceptible strains, led the board to recommend stopping the trial on clinical grounds supported by data on the changing in-vitro susceptibility. The trial was stopped on July 14, 2014. At this point, we had recruited and randomly assigned 246 eligible patients to treatment (including 116 patients with microbiologically confirmed disease; figure 1). Seven patients were excluded from the ITT population—four withdrew consent after randomisation but before receiving the first dose of study drug, and three had an alternative diagnosis. On stopping the trial, 120 patients had received gatifloxacin and 119 patients had received ceftriaxone, totalling 239 analysed in the modified ITT population. S Typhi or S Paratyphi A was isolated from the blood of 62 patients in the gatifloxacin group and 54 patients in the ceftriaxone group (figure 1). Analyses were not adjusted for early stopping of the trial.

ed gatifloxacin and 119 patients had received ceftriaxone, totalling 239 analysed in the modified ITT population. S Typhi or S Paratyphi A was isolated from the blood of 62 patients in the gatifloxacin group and 54 patients in the ceftriaxone group (figure 1). Analyses were not adjusted for early stopping of the trial. The baseline characteristics of patients were balanced between the two treatment groups in the modified ITT population (table 1) except for a larger proportion of men in the gatifloxacin group. Similar numbers of patients in each group had received a non-exclusion antimicrobial in the 2 weeks before randomisation. However, culture-negative patients were more likely to have had enteric fever previously and to report coughing, and had lower serum transaminase concentrations, than patients with blood culture-confirmed S Typhi or S Paratyphi A (appendix). Moreover, patients with S Typhi were more likely to report anorexia, nausea, and diarrhoea, and had a lower haematocrit compared with the other two patient groups. The MICs against ciprofloxacin and the study drugs were also balanced between the treatment groups (table 2). The first patient with a ciprofloxacin-resistant S Typhi culture (MIC >32 μg/mL) was enrolled on April 30, 2013. From that date, 118 additional patients were recruited, 55 of whom had positive blood cultures. Among these, 14 (25%) of 55 patients with S Typhi infections with high MICs against ciprofloxacin (12 >32 μg/mL, two 24 μg/mL) were assigned to a study drug; all 14 strains were also highly resistant to gatifloxacin (MIC ≥1·5 μg/mL).

date, 118 additional patients were recruited, 55 of whom had positive blood cultures. Among these, 14 (25%) of 55 patients with S Typhi infections with high MICs against ciprofloxacin (12 >32 μg/mL, two 24 μg/mL) were assigned to a study drug; all 14 strains were also highly resistant to gatifloxacin (MIC ≥1·5 μg/mL). Treatment failure in the modified ITT population was similar between treatment groups: 18 (15%) of 120 patients who received gatifloxacin had treatment failure, compared with 19 (16%) of 119 who received ceftriaxone (hazard ratio [HR] of time to failure 1·04 [95% CI 0·55–1·98]; p=0·91 [table 3]). Details for each event in the composite endpoint are in the appendix. However, there was significant heterogeneity in the primary outcome between the subpopulations of blood culture-confirmed and culture-negative patients (pinteraction<0·0001; table 3, figure 2). In the culture-confirmed population, 16 (26%) of 62 patients given gatifloxacin had treatment failure, compared with four (7%) of 54 patients given ceftriaxone (HR 0·24 [95% CI 0·08–0·73, p=0·01; table 3, absolute risks of failure in appendix). For the four patients with treatment failure in the ceftriaxone group, MICs against ceftriaxone ranged from 0–0·2 μg/mL, and were similar to MICs in patients without treatment failure. Treatment failure was associated with the emergence of S Typhi exhibiting resistance against fluoroquinolones. None of the subgroup analyses for culture-positive patients showed significant treatment effect heterogeneity of the primary endpoint (table 3).

and were similar to MICs in patients without treatment failure. Treatment failure was associated with the emergence of S Typhi exhibiting resistance against fluoroquinolones. None of the subgroup analyses for culture-positive patients showed significant treatment effect heterogeneity of the primary endpoint (table 3). By contrast, in culture-negative patients, only two (3%) of 58 who received gatifloxacin failed treatment compared with 15 (23%) of 65 who received ceftriaxone (HR 7·50 [95% CI 1·71–32·80]; p=0·01 [table 3, absolute risks of failure in appendix]). The most common cause of treatment failure in culture-negative patients treated with ceftriaxone was fever for more than 7 days (12 [80%] of 15 patients) and nine [60%] of 15 received rescue treatment (appendix). In the modified ITT population, fever clearance times did not differ between the two treatment groups (table 4, figure 2). Furthermore, the incidence of microbiological failure or syndromic relapse at any time until 6 months did not differ between the two treatment groups by day 28 or by 6 months (appendix).

By contrast, in culture-negative patients, only two (3%) of 58 who received gatifloxacin failed treatment compared with 15 (23%) of 65 who received ceftriaxone (HR 7·50 [95% CI 1·71–32·80]; p=0·01 [table 3, absolute risks of failure in appendix]). The most common cause of treatment failure in culture-negative patients treated with ceftriaxone was fever for more than 7 days (12 [80%] of 15 patients) and nine [60%] of 15 received rescue treatment (appendix). In the modified ITT population, fever clearance times did not differ between the two treatment groups (table 4, figure 2). Furthermore, the incidence of microbiological failure or syndromic relapse at any time until 6 months did not differ between the two treatment groups by day 28 or by 6 months (appendix). We noted significant heterogeneity (pinteraction<0·0001) of the treatment effect for fever clearance times in the blood culture-positive and blood culture-negative subgroups (table 4, figure 2). In culture-positive patients, median fever clearance times were longer in those treated with gatifloxacin than ceftriaxone (p=0·001) and outcomes with gatifloxacin were worse for patients with a raised MIC against ciprofloxacin and gatifloxacin (table 4). Occurrence of relapse did not differ between treatment groups in culture-positive patients (appendix). At 1-month follow-up, only two patients had positive stool cultures (one for S Typhi and one for S Paratyphi A), both in the gatifloxacin group. We noted no positive stool cultures at 3 months or 6 months in culture-positive patients (appendix). In culture-negative patients, fever clearance times were shorter in those treated with gatifloxacin than ceftriaxone (p<0·0001; table 4), but occurrences of relapse did not differ between treatment groups (appendix).

p. We noted no positive stool cultures at 3 months or 6 months in culture-positive patients (appendix). In culture-negative patients, fever clearance times were shorter in those treated with gatifloxacin than ceftriaxone (p<0·0001; table 4), but occurrences of relapse did not differ between treatment groups (appendix). Over the course of the trial, 122 adverse events occurred in the gatifloxacin group and 120 in the ceftriaxone group (appendix); no serious adverse events were reported. The most common adverse events reported were vomiting (in 27 [23%] of 120 patients receiving gatifloxacin and 17 [14%] of 119 receiving ceftriaxone; p=0·13), and cough (which did significantly differ between groups: in 15 [12%] patients receiving gatifloxacin and 29 [24%] patients receiving ceftriaxone; p=0·02). No adverse events were attributed to any of the study treatments. The frequency of dysglycaemia and abnormal liver-function tests did not differ between the treatment groups (appendix), and none of the study participants died.

5 [12%] patients receiving gatifloxacin and 29 [24%] patients receiving ceftriaxone; p=0·02). No adverse events were attributed to any of the study treatments. The frequency of dysglycaemia and abnormal liver-function tests did not differ between the treatment groups (appendix), and none of the study participants died. Discussion In patients with clinically suspected enteric fever, we showed that outcomes did not differ between patients receiving gatifloxacin and those receiving ceftriaxone. However, patients with blood culture-confirmed enteric fever fared less well when given gatifloxacin, as suggested by an increased likelihood of treatment failure and protracted fever clearance times. This finding was apparent only in the last year of recruitment into the trial as S Typhi strains with high-level resistance to ciprofloxacin and gatifloxacin (MICs >16 μg/mL and >1 μg/mL, respectively) emerged during the study, leading to the trial being stopped in July, 2014.

lure and protracted fever clearance times. This finding was apparent only in the last year of recruitment into the trial as S Typhi strains with high-level resistance to ciprofloxacin and gatifloxacin (MICs >16 μg/mL and >1 μg/mL, respectively) emerged during the study, leading to the trial being stopped in July, 2014. The data resulting from this trial contradict our hypothesis, because before the emergence of fluoroquinolone-resistant S Typhi in Nepal, we had shown in a series of clinical trials (done between 2004 and 2011) that gatifloxacin was both a safe and effective treatment for uncomplicated enteric fever.14, 16, 17 This antimicrobial has provided good clinical outcomes despite the continuing isolation of S Typhi and S Paratyphi A organisms showing reduced susceptibility against ciprofloxacin (MICs from 0·1 μg/mL–1 μg/mL).7, 24 Although the treatment failure results with gatifloxacin in our study are based on data from a few patients, these findings are highly consistent with the results of larger numbers for the secondary endpoint of fever clearance, which lend support to the credibility of our results.

in (MICs from 0·1 μg/mL–1 μg/mL).7, 24 Although the treatment failure results with gatifloxacin in our study are based on data from a few patients, these findings are highly consistent with the results of larger numbers for the secondary endpoint of fever clearance, which lend support to the credibility of our results. Generally, we still regard gatifloxacin as a safe drug for use in this setting since we have no evidence for an increased risk of dysglycaemia or hepatitis. However, our new data suggest that the efficacy of gatifloxacin (and older-generation fluoroquinolones) for the treatment of enteric fever in Nepal is now compromised by the emergence of high-level fluoroquinolone-resistant S Typhi. As a result, we no longer advocate gatifloxacin as a treatment for enteric fever in Nepal.

ever, our new data suggest that the efficacy of gatifloxacin (and older-generation fluoroquinolones) for the treatment of enteric fever in Nepal is now compromised by the emergence of high-level fluoroquinolone-resistant S Typhi. As a result, we no longer advocate gatifloxacin as a treatment for enteric fever in Nepal. Despite being a two-centre study in Kathmandu, our findings raise substantial questions regarding the use of fluoroquinolones in south Asia and other endemic regions for treating enteric fever. Genomic data has shown that strains with reduced susceptibility to fluoroquinolones are now globally dominant.25 The process of resistance development against the fluoroquinolones is a stepwise process; mutations are sequentially acquired in the target genes, thus determining higher MICs.7 Data obtained from in-vitro experiments suggest that strains with fluoroquinolone-resistance-associated mutations might actually have a selective advantage over wild-type strains.26 Fluoroquinolone resistance is clearly increasing, not only in Nepal, but also in neighbouring areas and other parts of the world where resources are low.24 Therefore in view of this combination of evidence, we predict a short window before the international emergence of S Typhi, and potentially S Paratyphi A, strains with high-level fluoroquinolone resistance, thus rendering this important group of antimicrobials globally ineffective for enteric fever. Conversely, the outcomes for ceftriaxone-treated, culture-positive patients were good, with short fever clearance times and few relapses. The optimum duration of ceftriaxone treatment is not clearly defined; in WHO guidelines, it is recommended for 10–14 days,6, 19 but our data lend support to a 7-day treatment course for patients with uncomplicated enteric fever in an endemic setting.

atients were good, with short fever clearance times and few relapses. The optimum duration of ceftriaxone treatment is not clearly defined; in WHO guidelines, it is recommended for 10–14 days,6, 19 but our data lend support to a 7-day treatment course for patients with uncomplicated enteric fever in an endemic setting. Alongside antimicrobial-resistant S Typhi, our clinical findings pose an additional clinical and public health challenge regarding ceftriaxone treatment. In patients with a negative blood-culture result, the absolute risk of treatment failure was 0·24 in the ceftriaxone group versus 0·04 in the gatifloxacin group. Furthermore, the median fever clearance times in this group were 3·03 days in the ceftriaxone group, versus 1·12 days in the gatifloxacin group. These contrasting outcomes for ceftriaxone in the two predefined patient populations were not predicted and previous similar data from other enteric fever trials are scarce. The reason for this shortage of data is because the results of patients enrolled in enteric fever trials who did not have a positive blood culture were, until recently, not reported. From a small ceftriaxone study11 done in Vietnamese patients with enteric fever, the investigators reported that two of the six culture-negative patients given ceftriaxone failed treatment, but responded to rescue treatment with ofloxacin. Only four randomised trials for enteric fever14, 15, 16, 17 did an intention-to-treat analysis and incorporated an analysis for the blood culture-negative patients. In these trials, culture-negative patients given gatifloxacin, ofloxacin, or azithromycin achieved similar (or better) outcomes than those reported in blood culture-confirmed patients with enteric fever.

15, 16, 17 did an intention-to-treat analysis and incorporated an analysis for the blood culture-negative patients. In these trials, culture-negative patients given gatifloxacin, ofloxacin, or azithromycin achieved similar (or better) outcomes than those reported in blood culture-confirmed patients with enteric fever. Better clinical outcomes in patients with syndromic enteric fever but with a negative blood-culture result might be attributed to the low sensitivity of blood-culture tests (estimated to be 50–60%)5 and the possibility of fewer bacteria in the bloodstream, corresponding with less severe disease. We do not know how many culture-negative patients actually had enteric fever in our study; some might have had alternative bacterial infections. Our previous study27 examined archived blood samples from 765 adults presenting at Patan Hospital, Nepal, with undifferentiated febrile illness in 2001. 50 (7%) patients had Rickettsia typhi (murine typhus) DNA detected by PCR amplification. Furthermore, we investigated the infectious cause of culture-negative patients enrolled into one of our other enteric fever trials in Nepal and noted serological evidence of murine typhus in 21 (22%) of 96 blood culture-negative patients, with 12 (57%) of 21 testing positive for R typhi with PCR amplification.28 Thus, we surmise that a substantial proportion of culture-negative patients with suspected enteric fever in Nepal might actually have other bacterial infections, included those caused by the rickettsiaceae. Ceftriaxone is not regarded as an effective treatment for murine typhus and other rickettsial illnesses, whereas fluoroquinolones do seem to have clinical activity against these pathogens.28 However, no rapid diagnostic tests are available that can accurately differentiate between rickettsial infections, enteric fever, or indeed any bacteraemia, and inform patient management in a timely manner. Without such a test, we suggest that a more pragmatic approach would be to combine ceftriaxone with doxycycline for patients without a positive blood culture who do not respond to ceftriaxone monotherapy.

rickettsial infections, enteric fever, or indeed any bacteraemia, and inform patient management in a timely manner. Without such a test, we suggest that a more pragmatic approach would be to combine ceftriaxone with doxycycline for patients without a positive blood culture who do not respond to ceftriaxone monotherapy. In conclusion, the results of our trial underline two substantial problems for patients with enteric fever. First, the continued development of antimicrobial resistance in the pathogens causing the disease; second, the absence of point-of-care diagnostic tests for febrile diseases in low-income settings.29 We have shown that high-level fluoroquinolone-resistant S Typhi is now likely to be endemic in Nepal and suggest that fluoroquinolones—even the fourth-generation fluoroquinolone gatifloxacin—cannot be recommended as empirical therapy for enteric fever in this setting. Azithromycin and ceftriaxone are alternative options, although sporadic cases of resistance have been reported, and data comparing in-vitro azithromycin susceptibility against clinical outcomes are poor.30, 31 Additionally, under our study conditions, ceftriaxone was suboptimum in a large proportion of culture-negative patients with suspected enteric fever, which further emphasises the need for diagnostic tests for enteric fever and other common febrile diseases. Since antimicrobials, specifically fluoroquinolones, are one of the only routinely used control measures for enteric fever, effective surveillance programmes, the assessment of novel diagnostics, new treatment options, and the use of existing vaccines and development of next-generation vaccines are now a greater priority than ever.

robials, specifically fluoroquinolones, are one of the only routinely used control measures for enteric fever, effective surveillance programmes, the assessment of novel diagnostics, new treatment options, and the use of existing vaccines and development of next-generation vaccines are now a greater priority than ever. Supplementary Material Supplementary appendix Acknowledgments This work was supported by the Wellcome Trust (London, UK) and the Li Ka Shing Foundation Global Health Programme at the University of Oxford (Oxford, UK). SB is funded by a Sir Henry Dale Fellowship from the Wellcome Trust and the Royal Society (100087/Z/12/Z). CD was funded by the Li Ka Shing Foundation Global Health Programme. We wish to acknowledge Mangal Rawal, Sumi Munankarmi, Bibek Karki, Radheshyam KC, Sudeep Dhoj Thapa, Rabin Gautam, Priyanka Tiwari, Manisha Risal, Surendra Shrestha, Balmukunda Neupane, Nabraj Regmi, Krishna Prajapati, Bimal Thapa, the trial monitors Nguyen Thi Phuong Dung and Nguyen Thi Thanh Thuy, and the members of our Data Safety Monitoring Board (Zulfikar A Bhutta, Keith P Klugman, Sue J Lee, Buddhi P Paudyal, and Nicholas J White). Contributors AA, BB, SB, GET, MW, and CD designed and conceived the study. AA, BB, SK, AG, NJ, MS, KRP, SPM, SPP, NA, RT, LM, PS, DG, KL, DL, BKY, GS, SD, AK, and CD ran the study and contributed data. HTN, CNT, MW, and CD did the data analysis. AA, NJ, SD, AK, NTVT, DPT, and SB did the experiments contributing to this study. BB, HTN, CNT, SB, GET, MW, and CD wrote the manuscript. Declaration of interests We declare no competing interests.

Contributors AA, BB, SB, GET, MW, and CD designed and conceived the study. AA, BB, SK, AG, NJ, MS, KRP, SPM, SPP, NA, RT, LM, PS, DG, KL, DL, BKY, GS, SD, AK, and CD ran the study and contributed data. HTN, CNT, MW, and CD did the data analysis. AA, NJ, SD, AK, NTVT, DPT, and SB did the experiments contributing to this study. BB, HTN, CNT, SB, GET, MW, and CD wrote the manuscript. Declaration of interests We declare no competing interests. Figure 1 Trial profile Salmonella enterica Typhi or Salmonella enterica Paratyphi A were isolated from the blood of patients with culture-confirmed enteric fever. Figure 2 Time to treatment failure and fever clearance time Time to treatment failure shown in the (A) modified intention to treat, (B) culture-confirmed, and (C) culture-negative populations. Fever clearance times shown in the (D) modified intention to treat, (E) culture-confirmed, and (F) culture-negative populations. Fever clearance times were interval-censored; because numbers at risk are not well defined in this setting they are not shown for graphs D, E, or F. HR=hazard ratio. AF=acceleration factor. Table 1 Baseline characteristics of the modified intention-to-treat population

Time to treatment failure shown in the (A) modified intention to treat, (B) culture-confirmed, and (C) culture-negative populations. Fever clearance times shown in the (D) modified intention to treat, (E) culture-confirmed, and (F) culture-negative populations. Fever clearance times were interval-censored; because numbers at risk are not well defined in this setting they are not shown for graphs D, E, or F. HR=hazard ratio. AF=acceleration factor. Table 1 Baseline characteristics of the modified intention-to-treat population Gatifloxacin (N=120) Ceftriaxone (N=119) N n (%) or median (IQR) N n (%) or median (IQR) Age (years) 120 19·0 (15·0−23·0) 119 20·0 (14·0−23·5) Sex Male 120 99 (83%) 119 81 (68%) Female 120 21 (18%) 119 38 (32%) Temperature (°C) 116 38·8 (38·3−39·4) 116 38·8 (38·3−39·4) Days of illness before enrolment 120 5·0 (4·0−6·0) 119 5·0 (4·0−7·0) Treatment with antibiotics in past 2 weeks 120 21 (18%) 119 17 (14%) Previous history of typhoid 120 18 (15%) 118 19 (16%) Family history of typhoid 120 18 (15%) 119 17 (14%) Typhoid vaccination 119 6 (5%) 119 5 (4%) Fever 120 120 (100%) 118 118 (100%) Cough 115 38 (33%) 113 42 (37%) Constipation 117 9 (8%) 116 16 (14%) Headache 119 99 (83%) 116 108 (93%) Diarrhoea 117 25 (21%) 116 28 (24%) Vomiting 116 32 (28%) 116 30 (26%) Abdominal pain 114 31 (27%) 115 27 (23%) Anorexia 118 88 (75%) 116 80 (69%) Nausea 116 60 (52%) 114 55 (48%) Splenomegaly 117 0 114 2 (2%) Hepatomegaly 117 0 114 0 Random blood glucose (mmol/L) 117 5·38 (4·77−6·11) 117 5·38 (4·94−5·88) Creatinine (μmol/L) 116 70·72 (61·88−79·56) 114 70·72 (61·88−79·56) Total bilirubin (μmol/L) 117 13·68 (10·26−17·10) 117 11·97 (10·26−15·39) Leucocyte cell count (×109/L) 120 6·9 (4·8−7·2) 119 5·8 (4·7−7·3) Haematocrit (%) 119 39·6 (37·0−43·0) 116 38·7 (35·8−44·0) Platelet cell count (×109/L) 120 170·0 (150·0−210·0) 119 167·0 (145·5−203·0) AST (U/L) 117 46·0 (32·0−66·0) 116 51·5 (38·8−80·0) ALT (U/L) 117 46·0 (30·0−63·0) 117 45·0 (33·0−63·0) Culture positive Salmonella Paratyphi A isolated 120 19 (16%) 119 16 (13%) Salmonella Typhi isolated 120 43 (36%) 119 38 (32%) No growth or culture negative 120 58 (48%) 119 65 (55%) N refers to the number of patients with non-missing data in each group. AST=serum aspartate aminotransferase. ALT=serum alanine aminotransferase.

positive Salmonella Paratyphi A isolated 120 19 (16%) 119 16 (13%) Salmonella Typhi isolated 120 43 (36%) 119 38 (32%) No growth or culture negative 120 58 (48%) 119 65 (55%) N refers to the number of patients with non-missing data in each group. AST=serum aspartate aminotransferase. ALT=serum alanine aminotransferase. Table 2 Minimum inhibitory concentration of organism in the culture-confirmed population at enrolment

positive Salmonella Paratyphi A isolated 120 19 (16%) 119 16 (13%) Salmonella Typhi isolated 120 43 (36%) 119 38 (32%) No growth or culture negative 120 58 (48%) 119 65 (55%) N refers to the number of patients with non-missing data in each group. AST=serum aspartate aminotransferase. ALT=serum alanine aminotransferase. Table 2 Minimum inhibitory concentration of organism in the culture-confirmed population at enrolment All patients Gatifloxacin Ceftriaxone Salmonella Typhi N=81 N=43 N=38 MIC against ciprofloxacin (μg/mL) n=78 n=41 n=37 MIC 50 0·38 0·38 0·38 MIC 90 >32·00 >32·00 13·40 Range* 0·008–>32·00 0·008–>32·00 0·016–>32·00 MIC against gatifloxacin (μg/mL) n=78 n=41 n=37 MIC 50 0·125 0·125 0·125 MIC 90 2·000 2·000 1·250 Range* 0·006−3·000 0·006−3·000 0·006−3·000 MIC against ceftriaxone (μg/mL) n=78 n=41 n=37 MIC 50 0·094 0·094 0·125 MIC 90 0·190 0·190 0·190 Range* 0·032−0·640 0·032−0·250 0·047−0·640 Salmonella Paratyphi A N=35 N=19 N=16 MIC against ciprofloxacin (μg/mL) n=34 n=18 n=16 MIC 50 0·500 0·625 0·500 MIC 90 0·925 1·000 0·750 Range* 0·380−1·000 0·380−1·000 0·380−1·000 MIC against gatifloxacin (μg/mL) n=34 n=18 n=16 MIC 50 0·500 0·500 0·500 MIC 90 0·750 0·575 0·750 Range* 0·380−0·750 0·380−0·750 0·380−0·750 MIC against ceftriaxone (μg/mL) n=34 n=18 n=16 MIC 50 0·125 0·125 0·125 MIC 90 0·190 0·145 0·220 Range* 0·064−0·500 0·094−0·190 0·064−0·500 n refers to the number of patients with non-missing data in each group. MIC=minimum inhibitory concentration. MIC 50=minimum inhibitory concentration at the 50th percentile. MIC 90=minimum inhibitory concentration at the 90th percentile.

25 0·125 0·125 MIC 90 0·190 0·145 0·220 Range* 0·064−0·500 0·094−0·190 0·064−0·500 n refers to the number of patients with non-missing data in each group. MIC=minimum inhibitory concentration. MIC 50=minimum inhibitory concentration at the 50th percentile. MIC 90=minimum inhibitory concentration at the 90th percentile. * Range from the minimum to the maximum noted MIC. Table 3 Treatment failure (primary endpoint) overall and in predefined subgroups

25 0·125 0·125 MIC 90 0·190 0·145 0·220 Range* 0·064−0·500 0·094−0·190 0·064−0·500 n refers to the number of patients with non-missing data in each group. MIC=minimum inhibitory concentration. MIC 50=minimum inhibitory concentration at the 50th percentile. MIC 90=minimum inhibitory concentration at the 90th percentile. * Range from the minimum to the maximum noted MIC. Table 3 Treatment failure (primary endpoint) overall and in predefined subgroups Gatifloxacin (events/n [%]) Ceftriaxone (events/n [%]) Hazard ratio of time to failure (95% CI); p value Heterogeneity test (pinteraction value) All patients (modified intention-to-treat population) 18/120 (15%) 19/119 (16%) 1·04 (0·55−1·98); p=0·91 Culture-negative or culture-positive populations <0·0001 Culture negative 2/58 (3%) 15/65 (23%) 7·50 (1·71−32·80); p=0·01 Culture positive 16/62 (26%) 4/54 (7%) 0·24 (0·08−0·73); p=0·01 Pathogen (culture-confirmed population) 0·25 Salmonella Paratyphi A 1/19 (5%) 1/16 (6%) 1·13 (0·07−18·02); p=0·93 Salmonella Typhi 15/43 (35%) 3/38 (8%) 0·18 (0·05−0·62); p=0·01 Age (modified intention-to-treat population) 0·25 <16 years 6/32 (19%) 4/36 (11%) 0·57 (0·16−2·00); p=0·38 ≥16 years 12/88 (14%) 15/83 (18%) 1·31 (0·61−2·80); p=0·48 Age (culture-confirmed population) 0·76 <16 years 6/21 (29%) 1/16 (6%) 0·19 (0·02−1·62); p=0·13 ≥16 years 10/41 (24%) 3/38 (8%) 0·27 (0·07−0·98); p=0·047 Sex (modified intention-to-treat population) 0·52 Female 3/21 (14%) 4/38 (11%) 0·69 (0·15−3·07); p=0·62 Male 15/99 (15%) 15/81 (19%) 1·21 (0·59−2·47); p=0·61 Sex (culture-confirmed population) 0·08 Female 3/11 (27%) 0/17 0 (0–∞); p=1·00 Male 13/51 (25%) 4/37 (11%) 0·37 (0·12−1·15); p=0·09 Recruitment date (modified intention-to-treat population) 0·15 Before April 1, 2013 7/62 (11%) 11/59 (19%) 1·69 (0·66−4·36); p=0·28 April 1, 2013, or later 11/58 (19%) 8/60 (13%) 0·65 (0·26−1·61); p=0·35 Recruitment date (culture-confirmed population) 0·70 Before April 1, 2013 6/33 (18%) 1/28 (4%) 0·18 (0·02−1·46); p=0·11 April 1, 2013, or later 10/29 (34%) 3/26 (12%) 0·28 (0·08−1·00); p=0·05 MIC against ciprofloxacin (culture-confirmed population) 0·15 <0·12 μg/mL 0/4 1/3 (33%) ∞ (0–∞); p=1·00 0·12−2·00 μg/mL 8/45 (18%) 2/46 (4%) 0·22 (0·05−1·05); p=0·06 >2·00 μg/mL* 8/10 (80%) 1/4 (25%) 0·17 (0·02−1·38); p=0·10 MIC against gatifloxacin (culture-confirmed population) 0·58 ≤1 μg/mL 8/49 (16%) 3/49 (6%) 0·34 (0·09−1·28); p=0·11 >1 μg/mL 8/10 (80%) 1/4 (25%) 0·17 (0·02−1·38); p=0·10 MIC=minimum inhibitory concentration.

μg/mL 8/45 (18%) 2/46 (4%) 0·22 (0·05−1·05); p=0·06 >2·00 μg/mL* 8/10 (80%) 1/4 (25%) 0·17 (0·02−1·38); p=0·10 MIC against gatifloxacin (culture-confirmed population) 0·58 ≤1 μg/mL 8/49 (16%) 3/49 (6%) 0·34 (0·09−1·28); p=0·11 >1 μg/mL 8/10 (80%) 1/4 (25%) 0·17 (0·02−1·38); p=0·10 MIC=minimum inhibitory concentration. * Among the 14 strains, two had a ciprofloxacin MIC of 24 μg/mL and 12 had a ciprofloxacin MIC>32 μg/mL. Table 4 Fever clearance time (secondary endpoint) overall and in predefined subgroups

μg/mL 8/45 (18%) 2/46 (4%) 0·22 (0·05−1·05); p=0·06 >2·00 μg/mL* 8/10 (80%) 1/4 (25%) 0·17 (0·02−1·38); p=0·10 MIC against gatifloxacin (culture-confirmed population) 0·58 ≤1 μg/mL 8/49 (16%) 3/49 (6%) 0·34 (0·09−1·28); p=0·11 >1 μg/mL 8/10 (80%) 1/4 (25%) 0·17 (0·02−1·38); p=0·10 MIC=minimum inhibitory concentration. * Among the 14 strains, two had a ciprofloxacin MIC of 24 μg/mL and 12 had a ciprofloxacin MIC>32 μg/mL. Table 4 Fever clearance time (secondary endpoint) overall and in predefined subgroups Gatifloxacin Ceftriaxone Acceleration factor (95% CI), p value Heterogeneity test (pinteraction value) n Median (IQR) days n Median (IQR) days All patients (modified intention-to-treat population) 120 2·43 (1·09−4·56) 119 2·93 (1·44−5·12) 0·89 (0·72−1·11); p=0·31 Culture-negative or culture-positive population <0·0001 Culture negative 58 1·12 (0·39−2·58) 65 3·03 (1·31−5·85) 0·44 (0·30−0·65); p<0·0001 Culture positive 62 4·21 (2·63−6·10) 54 2·78 (1·62−4·26) 1·42 (1·15−1·76); p=0·001 Pathogen (culture-confirmed population) 0·57 Salmonella Paratyphi A 19 3·68 (2·50−4·98) 16 2·24 (1·18−3·69) 1·31 (0·88−1·94); p=0·19 Salmonella Typhi 43 4·51 (2·79−6·58) 38 3·03 (1·86−4·46) 1·47 (1·15−1·88); p=0·002 Age (modified intention-to-treat population) 0·29 <16 years 32 3·02 (1·70−4·75) 36 2·23 (0·93−4·46) 1·08 (0·72−1·60); p=0·72 ≥16 years 88 2·22 (0·92−4·45) 83 3·25 (1·70−5·41) 0·83 (0·64−1·08); p=0·16 Age (culture-confirmed population) 0·46 <16 years 21 3·82 (2·45−5·43) 16 3·04 (1·94−4·32) 1·26 (0·90−1·76); p=0·18 ≥16 years 41 4·43 (2·77−6·43) 38 2·66 (1·47−4·22) 1·51 (1·15−1·97); p=0·003 Sex (modified intention-to-treat population) 0·99 Female 21 2·41 (1·13−4·39) 38 2·78 (1·34−4·94) 0·89 (0·56−1·41); p=0·61 Male 99 2·44 (1·09−4·59) 81 2·99 (1·48−5·20) 0·89 (0·68−1·14); p=0·35 Sex (culture-confirmed population) 0·68 Female 11 4·27 (3·12−5·47) 17 3·06 (1·93−4·41) 1·25 (0·86−1·83); p=0·25 Male 51 4·18 (2·56−6·16) 37 2·66 (1·50−4·17) 1·46 (1·13−1·89); p=0·004 Recruitment date (modified intention-to-treat population) 0·09 Before April 1, 2013 62 2·30 (1·10−4·09) 59 2·79 (1·16−5·55) 0·74 (0·53−1·03); p=0·08 April 1, 2013, or later 58 2·60 (1·12−5·04) 60 3·05 (1·76−4·69) 1·09 (0·82−1·45); p=0·56 Recruitment date (culture-confirmed population) 0·12 Before April 1, 2013 33 3·88 (2·63−5·26) 28 2·54 (1·31−4·27) 1·21 (0·90−1·63); p=0·22 April 1, 2013, or later 29 4·68 (2·82−6·97) 26 3·00 (1·96−4·20) 1·68 (1·26−2·23); p=0·0004 MIC against ciprofloxacin (culture-confirmed population) 0·02 <0·12 μg/mL 4 2·55 (1·82−3·32) 3 4·98 (4·09−5·82) 0·58 (0·35−0·94); p=0·028 0·12−2·00 μg/mL 45 3·88 (2·67−5·21) 46 2·63 (1·49−4·12) 1·24 (0·99−1·56); p=0·06 >2·00 μg/mL* 10 8·20 (5·99−10·50) 4 3·66 (2·

·82−6·97) 26 3·00 (1·96−4·20) 1·68 (1·26−2·23); p=0·0004 MIC against ciprofloxacin (culture-confirmed population) 0·02 <0·12 μg/mL 4 2·55 (1·82−3·32) 3 4·98 (4·09−5·82) 0·58 (0·35−0·94); p=0·028 0·12−2·00 μg/mL 45 3·88 (2·67−5·21) 46 2·63 (1·49−4·12) 1·24 (0·99−1·56); p=0·06 >2·00 μg/mL* 10 8·20 (5·99−10·50) 4 3·66 (2· 84−4·46) 2·36 (1·58−3·51); p=<0·0001 MIC against gatifloxacin (culture-confirmed population) 0·049 ≤1·00 μg/mL 49 3·76 (2·56−5·08) 49 2·75 (1·58−4·27) 1·17 (0·94−1·45); p=0·15 >1·00 μg/mL 10 8·20 (5·99−10·50) 4 3·66 (2·84−4·46) 2·36 (1·58−3·51); p<0·0001 Percentages not added to this table because the denominators for populations change and are not clearly specified. MIC=minimum inhibitory concentration. * Among the 14 strains, two had a ciprofloxacin MIC of 24 μg/mL and 12 had an MIC >32 μg/mL.

Introduction Several African countries including Malawi, Liberia, and Ethiopia have met their Millennium Development Goals for child mortality reduction; however, neonatal deaths caused by infections, preterm birth, and birth asphyxia account for 44% of mortality for children younger than 5 years.1 Group B streptococcus (GBS) has been identified as a leading cause of neonatal sepsis and meningitis in several countries across sub-Saharan Africa,2 and is therefore a crucial target for public health intervention. In Africa, the reported incidence of early-onset invasive GBS disease varies across studies from 0 to 2·1 per 1000 livebirths, whereas late-onset invasive GBS disease varies from 0 to 0·89 per 1000 livebirths, with case fatality rates ranging from 13% to 46%.2 Intrapartum antibiotic prophylaxis has substantially reduced, although not eliminated, early-onset invasive GBS disease in high-income countries.3 Intrapartum antibiotic prophylaxis is difficult to implement in resource-poor settings and has little effect on the incidence of late-onset invasive GBS disease.4 Research in context Evidence before this study We searched PubMed and Web of Science for studies on group B streptococcus (GBS) vaccines published before Dec 1, 2015, using the search terms “Group B Streptococcus vaccine” and combinations thereof. We did not find any previous GBS vaccines that had been tested on pregnant women infected with HIV, and no published research on CRM197-conjugated GBS vaccines in human beings. Added value of this study

We searched PubMed and Web of Science for studies on group B streptococcus (GBS) vaccines published before Dec 1, 2015, using the search terms “Group B Streptococcus vaccine” and combinations thereof. We did not find any previous GBS vaccines that had been tested on pregnant women infected with HIV, and no published research on CRM197-conjugated GBS vaccines in human beings. Added value of this study This is the first study investigating the safety and immunogenicity of a candidate glycoconjugate GBS vaccine in pregnant women infected with HIV and their infants, and shows that the vaccine is both immunogenic and has a good safety profile, although it was less immunogenic in women infected with HIV. The study also adds to previous evidence of the safety and immunogenicity of GBS vaccines in healthy pregnant and non-pregnant women, in earlier clinical trials. This is the first GBS vaccine immunogenicity study to be done on the African continent that has included participants from outside South Africa. Implications of all the available evidence

This is the first study investigating the safety and immunogenicity of a candidate glycoconjugate GBS vaccine in pregnant women infected with HIV and their infants, and shows that the vaccine is both immunogenic and has a good safety profile, although it was less immunogenic in women infected with HIV. The study also adds to previous evidence of the safety and immunogenicity of GBS vaccines in healthy pregnant and non-pregnant women, in earlier clinical trials. This is the first GBS vaccine immunogenicity study to be done on the African continent that has included participants from outside South Africa. Implications of all the available evidence Given that GBS is a leading cause of neonatal sepsis and meningitis across sub-Saharan Africa, this vaccine offers a potential pathway to reduce infection-related neonatal death and disability in high-burden settings. However, since the vaccine was less immunogenic in women infected with HIV than in those not infected, irrespective of CD4 cell count, validated correlates of protection need to be identified to improve understanding of the potential protective value of this GBS vaccine for pregnant women and their infants, particularly in high HIV seroprevalence countries.

unogenic in women infected with HIV than in those not infected, irrespective of CD4 cell count, validated correlates of protection need to be identified to improve understanding of the potential protective value of this GBS vaccine for pregnant women and their infants, particularly in high HIV seroprevalence countries. Glycoconjugate vaccines against other capsulated bacteria, such as Neisseria meningitidis, Streptococcus pneumoniae, and Haemophilus influenzae type b, have proved highly effective in many settings.5, 6, 7 Because neonatal GBS disease develops rapidly after birth, administration of a glycoconjugate GBS vaccine to infants is unlikely to prevent disease and therefore maternal immunisation during pregnancy has been identified as the best strategy to prevent both early-onset and late-onset invasive GBS disease, and therefore possibly to reduce the incidence of stillbirth and miscarriage.8 Maternal immunisation against other diseases, such as tetanus, has proved highly effective and acceptable in many low-income and middle-income countries.9, 10

the best strategy to prevent both early-onset and late-onset invasive GBS disease, and therefore possibly to reduce the incidence of stillbirth and miscarriage.8 Maternal immunisation against other diseases, such as tetanus, has proved highly effective and acceptable in many low-income and middle-income countries.9, 10 About 5·3% of pregnant women in sub-Saharan Africa are thought to be infected with HIV, and although efforts to prevent mother-to-child transmission across the continent are having an effect, rates of neonatal exposure to HIV remain high.11 Neonates born to women infected with HIV are at increased risk for invasive GBS disease in high-income and middle-income countries.12, 13 The risk for late-onset invasive GBS disease (occurring in neonates aged between 7 and 90 days) is 3·18 to 19 times greater in neonates born to mothers infected with HIV than in those born to mothers not infected with HIV, and the risk for early-onset GBS disease (occurring in neonates aged <7 days) is 1·7 times greater.13, 14, 15 A study11 in Belgium reported that 1·55% of infants born to mothers infected with HIV developed invasive GBS disease, compared with 0·08% of infants born to mothers not infected with HIV over the same time period. Similarly, a study15 in South Africa reported the incidence of invasive GBS disease to be higher in infants exposed to HIV (4·46 per 1000 livebirths) compared with infants not exposed to HIV (1·98 per 1000 livebirths). The increased susceptibility to invasive GBS disease in neonates born to women infected with HIV, despite most of these babies not being infected with HIV, is probably due to lower concentrations of naturally acquired serotype-specific capsular antibodies and reduced transplacental transfer in mother–newborn dyads in which the mother is infected with HIV than in dyads in which the mother is not infected.15 Response to vaccination is known to be impaired in individuals with HIV infection.16 A previous trial17 of three doses (0·5 μg, 2·5 μg, and 5·0 μg) of a glycoconjugate GBS trivalent vaccine in women not infected with HIV in South Africa showed that the vaccine was well tolerated and immunogenic.

ther is not infected.15 Response to vaccination is known to be impaired in individuals with HIV infection.16 A previous trial17 of three doses (0·5 μg, 2·5 μg, and 5·0 μg) of a glycoconjugate GBS trivalent vaccine in women not infected with HIV in South Africa showed that the vaccine was well tolerated and immunogenic. In the present study, we explore the hypothesis that HIV infection affects the safety and immunogenicity of the 5·0 μg formulation of GBS vaccine in pregnant women and their infants in Malawi and South Africa. We aimed to assess placental transfer of GBS serotype-specific antibodies in patients with and without HIV, as well as post-vaccination concentrations in mothers and safety. Methods Study design This non-randomised phase 2, open-label, multicentre study was done in two antenatal clinics, one in Blantyre, Malawi, and one in Soweto, Johannesburg, South Africa, between September, 2011, and December, 2012. Pregnant women were enrolled sequentially by study nurses into three groups in a 1:1:1 ratio on the basis of their HIV infection status and CD4 cell count until each of the groups was filled; the groups were HIV uninfected, HIV infected with a high CD4 cell count (>350 cells per μL), and HIV infected with a low CD4 cell count (>50 to ≤350 cells per μL).

tially by study nurses into three groups in a 1:1:1 ratio on the basis of their HIV infection status and CD4 cell count until each of the groups was filled; the groups were HIV uninfected, HIV infected with a high CD4 cell count (>350 cells per μL), and HIV infected with a low CD4 cell count (>50 to ≤350 cells per μL). The study was done in accordance with the Declaration of Helsinki and the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use Guidelines for Good Clinical Practice. Written informed consent was obtained from women before enrolment. The protocol was approved by the National Health Sciences Research Committee (Malawi) and the University of Witwatersrand, Human Research Ethics Committee (South Africa). Use of an investigational GBS vaccine was approved by the Pharmacy, Medicines & Poisons Board, Malawi, and the Medicine Control Council, South Africa.

he protocol was approved by the National Health Sciences Research Committee (Malawi) and the University of Witwatersrand, Human Research Ethics Committee (South Africa). Use of an investigational GBS vaccine was approved by the Pharmacy, Medicines & Poisons Board, Malawi, and the Medicine Control Council, South Africa. Participants Pregnant women aged 18–40 years between 24 and 35 weeks' gestation were eligible. All women attending the clinics at the study sites were informed of the study. Those who fulfilled the study eligibility criteria and agreed to take part were enrolled after giving informed written consent. Women infected with HIV were eligible if their CD4 cell count was more than 50 cells per μL and they had WHO stage I or II disease. Exclusion criteria included receipt of a vaccine 15 days before enrolment (except tetanus toxoid and non-alum adjuvanted influenza vaccines); severe allergic reaction or hypersensitivity to previous vaccinations or vaccine component; fever at enrolment or acute infection up to 7 days before enrolment; any disorder associated with prolonged bleeding; immunosuppressive treatment within 30 days before enrolment; receipt of blood or blood products in the 12 weeks before enrolment; behavioural or cognitive impairment; and progressive or severe neurological disorder, seizure disorder, epilepsy, or Guillain-Barré syndrome. Twin pregnancies were included in the study (appendix). Carriage of GBS was not assessed or deemed an exclusion criterion for this study. Treatment of mothers and their babies, including HIV management, was provided by government services according to national guidelines, free at the point of care.

lain-Barré syndrome. Twin pregnancies were included in the study (appendix). Carriage of GBS was not assessed or deemed an exclusion criterion for this study. Treatment of mothers and their babies, including HIV management, was provided by government services according to national guidelines, free at the point of care. Gestational age was estimated by (in order of preference) ultrasonography before 24 weeks' gestation, date of last menstrual period, fundal height at 20–35 weeks' gestation, or ultrasonography at 25–27 weeks' gestation. CD4 cell counts were assessed using BD FACSCount (BD Biosciences, San Jose, CA, USA) and viral load was measured in Malawi using the Abbott RealTime HIV-1 assay (Abbott Laboratories, Abbott Park, IL, USA), and in South Africa using the Taqman version 2 HIV-1 assay (Roche, Pleasanton, CA, USA). Procedures Women were immunised with a 0·5 mL dose of non-adjuvanted CRM197-conjugated GBS vaccine (Novartis Vaccines, Siena, Italy) containing 5 μg of each capsular polysaccharide of serotypes Ia, Ib, and III, reconstituted in 0·9% sodium chloride. The vaccine was given intramuscularly into the non-dominant arm at 24–35 weeks' gestation. For immunogenicity analysis, blood was collected from women before vaccination on day 1, at days 15 and 31 post-vaccination, and at delivery. For women infected with HIV, HIV-1 viral load and CD4 cell counts were measured before vaccination and the day of delivery. Cord blood or peripheral blood was collected from infants at birth and day 42.

sis, blood was collected from women before vaccination on day 1, at days 15 and 31 post-vaccination, and at delivery. For women infected with HIV, HIV-1 viral load and CD4 cell counts were measured before vaccination and the day of delivery. Cord blood or peripheral blood was collected from infants at birth and day 42. GBS serotype-specific antibody concentrations in women and infants were estimated using a previously described ELISA protocol at GlaxoSmithKline Clinical Sciences Laboratory, Marburg, Germany.18 In brief, antibody concentrations were assessed using 96-well plates, which were coated in 1 μg/mL human serum albumin-conjugated GBS polysaccharide representing the three vaccine serotypes. Serially diluted serum samples were incubated on the coated plates for 1 h at 37°C and an alkaline-phosphatase-conjugated goat anti-human IgG secondary antibody was added after washing and incubated for a further 90 min. After further washing, SeramunGelb pNPP was added to the plates and incubated for 30 min at room temperature, then the reaction was stopped with SeramunGelb stop. Antibody concentrations were calculated using Mikrowin 2000 software from optical density values measured at 405 nm using a BEP III ELISA processor. Because of lab constraints, serotype Ia ELISA testing was done in three batches, whereas all the testing for serotype Ib and III was done in one batch. Antibody concentrations were expressed as geometric mean concentrations (GMCs) with 95% CIs, calculated with the Clopper-Pearson method. Subgroup analysis was done for women on the basis of baseline serotype-specific antibody titres.

three batches, whereas all the testing for serotype Ib and III was done in one batch. Antibody concentrations were expressed as geometric mean concentrations (GMCs) with 95% CIs, calculated with the Clopper-Pearson method. Subgroup analysis was done for women on the basis of baseline serotype-specific antibody titres. Outcomes The primary objective of this study was to compare the amount of placental transfer of GBS serotype-specific antibodies to the infants of pregnant women infected with HIV and those not infected after administration of investigational GBS vaccine. As a secondary objective, the concentrations of maternal serotype-specific GBS antibodies were assessed post-vaccination. An exploratory objective was the assessment of the kinetics of maternally derived antibodies in infants born to vaccinated women. Safety objectives were the assessment of solicited adverse reactions, unsolicited adverse events, serious adverse events, and obstetric outcomes.

S antibodies were assessed post-vaccination. An exploratory objective was the assessment of the kinetics of maternally derived antibodies in infants born to vaccinated women. Safety objectives were the assessment of solicited adverse reactions, unsolicited adverse events, serious adverse events, and obstetric outcomes. After immunisation, the women were observed for 30 min for any immediate adverse reactions. Adverse events and serious adverse events were graded on severity and possible relation to study vaccine by the investigator. A standardised approach was used to identify adverse events, which were defined as any untoward medical occurrence in a mother given the investigational vaccine or her infant that did not necessarily have a causal relation with this treatment. Solicited adverse reactions were recorded for 7 days post-vaccination with diary cards. Adverse events were collected up to day 31 post-vaccination, and adverse events requiring a physician's visit, serious adverse events, and deaths were recorded for women and infants throughout the study. As part of the safety analysis the outcome of the pregnancy was assessed, including the health status of neonates (including Apgar score). Admission to hospital for a normal delivery was not deemed a serious adverse event in the context of this study.

and deaths were recorded for women and infants throughout the study. As part of the safety analysis the outcome of the pregnancy was assessed, including the health status of neonates (including Apgar score). Admission to hospital for a normal delivery was not deemed a serious adverse event in the context of this study. Statistical analysis The sample sizes were calculated on the conservative assumption that only 60 participants per group would be enrolled, that there would be a 15% dropout rate, and that 80% of women would deliver infants that are at 37 weeks or more of gestation or weighed 2500 g or more at birth. Thus, a sample size of 40 per group would give about 95% power to detect an 8% difference in the percentage of placental transfer between any two study groups. However, if the maximum number of participants per group (90) were to be reached, under the same assumptions for the dropout rate and for the percentage of babies born at 37 weeks or more of gestation or weighing 2500 g or more at birth, the number of assessable participants would increase up to 60 per group, giving about 95% power to detect a 6·5% difference in the percentage of placental transfer between any two study groups. After determination of the sample size for the trial with and without babies born at less than 37 weeks' gestation or less than 2500 g, we decided that both of these early gestation or underweight groups should be retained within the analysis. All serum analyses were done on the full analysis set, which included all mothers who correctly received the vaccine and who provided at least one valid assessable serum sample, and their neonates. Safety data were analysed descriptively for the safety set, which included all mothers who correctly received the study vaccination and who provided safety data, and their neonates.

h included all mothers who correctly received the vaccine and who provided at least one valid assessable serum sample, and their neonates. Safety data were analysed descriptively for the safety set, which included all mothers who correctly received the study vaccination and who provided safety data, and their neonates. The primary objective was analysed using an ANCOVA model on log10-transformed maternal and infant GBS antibody levels, with HIV group and country as qualitative factors, and gestational age at delivery and birthweight as covariates. Multiplicity of testing across the three participant groups was adjusted for by using a significance level of 0·016 (ie, p=0·05/3). The null hypothesis was rejected if a significant difference was seen for all three serotypes. The secondary objective was analysed using an ANCOVA on log10-transformed maternal antibody concentrations at each relevant timepoint, with HIV group and country as qualitative factors, and baseline log10-antibody concentration as a covariate. Infant antibody concentrations at each timepoint were analysed using an ANOVA model, with maternal HIV group and country as qualitative factors and a significance level of 0·05. For analysis, any antibody concentrations that were below the lower limit of quantification (<LLQ) were set as half the LLQ (LLQ serotype Ia: 0·326 μg/mL, serotype Ib: 0·083 μg/mL, serotype III: 0·080 μg/mL). All statistical analyses were done with SAS version 9.1. This study is registered with ClinicalTrials.gov, number NCT01412801.

For analysis, any antibody concentrations that were below the lower limit of quantification (<LLQ) were set as half the LLQ (LLQ serotype Ia: 0·326 μg/mL, serotype Ib: 0·083 μg/mL, serotype III: 0·080 μg/mL). All statistical analyses were done with SAS version 9.1. This study is registered with ClinicalTrials.gov, number NCT01412801. Role of funding source The study sponsor was involved in all stages of the study, including manuscript development. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. All authors agreed to submit the manuscript for publication. No honorarium, grant, or other form of payment was provided to authors, with the exception of funding needed to do the study.

sponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. All authors agreed to submit the manuscript for publication. No honorarium, grant, or other form of payment was provided to authors, with the exception of funding needed to do the study. Results Of 398 women screened across two sites in Malawi and South Africa, 270 women and their infants were enrolled in the study between Sept 26, 2011, and Dec 4, 2012: 90 without HIV (45 from each country), 89 with HIV and high CD4 cell counts (44 from Malawi and 45 from South Africa), and 91 with HIV and low CD4 cell counts (46 from Malawi and 45 from South Africa; figure). 254 (94%) women completed the study across the three groups; seven were lost to follow-up, six withdrew consent, one died, and two relocated away from the study area. 256 (96%) of 266 infants enrolled completed the study across the groups. 270 mothers were enrolled in the study but only 260 remained in the study at delivery, five of whom had twins, resulting in 265 livebirths. One mother died during labour and had a stillbirth that was included in the enrolled infants (n=266). Reasons for withdrawal were death or stillbirth (n=8), and loss to follow-up (n=2).

others were enrolled in the study but only 260 remained in the study at delivery, five of whom had twins, resulting in 265 livebirths. One mother died during labour and had a stillbirth that was included in the enrolled infants (n=266). Reasons for withdrawal were death or stillbirth (n=8), and loss to follow-up (n=2). Baseline characteristics for women and infants were similar across the groups, although, consistent with the natural history of HIV acquisition in these populations, women in the HIV-uninfected group were younger than those in the HIV-infected groups (table 1). No notable differences in demographics existed between sites except that the maternal median body-mass index was higher in all groups in South Africa (26·8–29·3 kg/m2) compared with Malawi (22·8–24·7 kg/m2). As expected, viral loads were higher in mothers infected with HIV and low CD4 cell counts than they were in those with high CD4 cell counts. Because most women who were not already receiving antiretroviral treatment were given treatment during the study, viral load decreased between screening and delivery in both HIV-infected groups. Rates of premature birth for the HIV-uninfected group were similar to those reported for sub-Saharan Africa, and were higher in HIV-infected groups.19

re not already receiving antiretroviral treatment were given treatment during the study, viral load decreased between screening and delivery in both HIV-infected groups. Rates of premature birth for the HIV-uninfected group were similar to those reported for sub-Saharan Africa, and were higher in HIV-infected groups.19 Rates of women reporting at least one solicited adverse reaction were highest in the HIV-uninfected group (60 [67%] of 90 women), compared with the low CD4 cell count (39 [44%] of 88 women) and high CD4 cell count (52 [59%] of 88 women) groups. Most reactions were grade 2 or less, with 4% or less of participants reporting a severe reaction for any of the solicited adverse reactions (table 2). Local reactions were reported by 18–39% of women across the groups, and 40–59% reported systemic reactions. The most frequently reported local adverse reaction was injection site pain, with severe pain being reported by two (2%) of 87 women with HIV and low CD4 cell count group and four (4%) of 90 in the HIV-uninfected group (table 2). The most often reported systemic adverse reactions were fatigue and headache, which were reported most frequently by women without HIV and least by those with HIV and low CD4 cell counts (table 2).

d by two (2%) of 87 women with HIV and low CD4 cell count group and four (4%) of 90 in the HIV-uninfected group (table 2). The most often reported systemic adverse reactions were fatigue and headache, which were reported most frequently by women without HIV and least by those with HIV and low CD4 cell counts (table 2). The percentage of women reporting adverse events was similar across the three groups, with 74–78% reporting at least one adverse event (table 3). Of these, 7–23% were deemed to be at least possibly related to the study vaccine. Serious adverse events were reported by 28–32% of women, but none were deemed to be caused by the study vaccine. Similarly, no differences in reporting of adverse events were seen across the infant groups. 41–49% of infants reported adverse events and 18–19% reported serious adverse events (table 3), none of which was deemed to be caused by vaccination. One woman with HIV in the high CD4 cell count group died from uterine rupture and her infant was stillborn (appendix). Seven other infant deaths occurred in the study, mostly in Malawi; no deaths in the study were deemed to be caused by the study vaccine (table 3, appendix). No differences in obstetric outcomes and pregnancy events were recorded across the three groups (appendix). No association between vaccine administration and change in viral load was seen in the HIV-infected groups.

The percentage of women reporting adverse events was similar across the three groups, with 74–78% reporting at least one adverse event (table 3). Of these, 7–23% were deemed to be at least possibly related to the study vaccine. Serious adverse events were reported by 28–32% of women, but none were deemed to be caused by the study vaccine. Similarly, no differences in reporting of adverse events were seen across the infant groups. 41–49% of infants reported adverse events and 18–19% reported serious adverse events (table 3), none of which was deemed to be caused by vaccination. One woman with HIV in the high CD4 cell count group died from uterine rupture and her infant was stillborn (appendix). Seven other infant deaths occurred in the study, mostly in Malawi; no deaths in the study were deemed to be caused by the study vaccine (table 3, appendix). No differences in obstetric outcomes and pregnancy events were recorded across the three groups (appendix). No association between vaccine administration and change in viral load was seen in the HIV-infected groups. For all groups, GMCs of antibodies were higher post-vaccine than at baseline at all tested timepoints (table 4). The highest responses were seen in the HIV-uninfected group, for which ratios between baseline and delivery against the three serotypes were 15·0 μg/mL (Ia), 9·14 μg/mL (Ib), and 30 μg/mL (III); whereas these were 7·21 μg/mL (Ia), 4·83 μg/mL (Ib), and 9·99 μg/mL (III) for the low CD4 cell count group; and 8·78 μg/mL (Ia), 7·34 μg/mL (Ib), and 8·65 μg/mL (III) for the high CD4 cell count group. The biggest differences between women infected and not infected with HIV were reported against serotype III. No difference in immunogenicity was reported between mothers in the high and low CD4 cell count groups. Despite differences in magnitude, the kinetics of the antibody response in women infected and not infected with HIV (table 4), and the variation in individual responses were similar across groups (appendix). Antibody concentrations at baseline were undetectable (<LLQ) for about 69–80% of women against serotype Ia (72 in the low CD4 cell count, 66 in high CD4 cell count, and 62 in HIV-uninfected groups), 1–6% of women against serotype Ib (one in the low CD4 cell count, one in high CD4 cell count, and five in HIV-uninfected groups), and 34–43% of women against serotype III (22 in the low CD4 cell count, 33 in high CD4 cell count, and 33 in HIV-uninfected groups). Because too few participants had antibody concentrations below the LLQ against serotype Ib, subgroup analysis was only done on serotypes Ia and III. When stratified by baseline antibody concentration (<LLQ or ≥LLQ), differences in antibody GMCs between the HIV-uninfected and HIV-infected groups against serotypes Ia and III were less pronounced for women with undetectable GBS antibody concentrations at baseline. Antibody GMCs post-vaccination were higher in those who had detectable antibody concentrations at baseline, with GMCs at delivery of 18–61 μg/mL across groups for serotype Ia and 1·8–8·8 μg/mL for serotype III, compared with GMCs of 1·0–1·8 μg/mL and 0·4–1·1 μg/mL for the same serotypes for women who had undetectable antibody concentrations at baseline (appendix).

those who had detectable antibody concentrations at baseline, with GMCs at delivery of 18–61 μg/mL across groups for serotype Ia and 1·8–8·8 μg/mL for serotype III, compared with GMCs of 1·0–1·8 μg/mL and 0·4–1·1 μg/mL for the same serotypes for women who had undetectable antibody concentrations at baseline (appendix). Ratios of antibody GMCs to baseline could not be accurately calculated for this analysis because of the high frequency of women below the LLQ at baseline. For serotype Ia, a site difference in response was recorded, with women in South Africa having much higher baseline and vaccine-induced antibody responses than the women in Malawi, across the HIV-infected and HIV-uninfected groups (appendix). Placental GBS serotype-specific antibody transfer ratios were similar between HIV-uninfected and HIV-infected groups for each serotype (0·49–0·72; table 5). Infant antibody GMCs followed the same pattern as maternal antibody GMCs, with infants in the HIV-uninfected group having higher GMCs at birth and day 42 than those in HIV-infected groups (table 5). Rates of antibody decay from day 1 to day 42 were similar across the three groups, with geometric mean ratios over this period of 0·50–0·58 for serotype Ia, 0·28–0·56 for serotype Ib, and 0·31–0·40 for serotype III.

HIV-uninfected group having higher GMCs at birth and day 42 than those in HIV-infected groups (table 5). Rates of antibody decay from day 1 to day 42 were similar across the three groups, with geometric mean ratios over this period of 0·50–0·58 for serotype Ia, 0·28–0·56 for serotype Ib, and 0·31–0·40 for serotype III. Discussion In this small phase 2 trial in high disease burden settings in Malawi and South Africa, we showed the trivalent GBS vaccine to be well tolerated by pregnant women irrespective of HIV status, and have not identified any concerning safety signals in these women or their infants. The vaccine was substantially more immunogenic in women not infected with HIV than it was in those with the infection, and generated antibody GMCs to serotypes Ib and III similar to those recorded previously in pregnant women given this vaccine (GMCs at delivery from a previous study18 in Belgium and Canada: 2·41 μg/mL for serotype Ib and 1·90 μg/mL for serotype III). However, antibody GMCs against serotype Ia in mothers not infected with HIV in the present study were substantially lower than those recorded for the same vaccine dose in a previous study17 of women not infected with HIV in South Africa (4·49 μg/mL for the HIV-uninfected group in the present study compared with five times higher values for the group receiving the equivalent vaccine dose in the previous South African study). Carriage was not measured in the present study, but previous reports suggest that although South Africa and Malawi have similar prevalence in carriage isolates for serotype Ib (6·7% [South Africa] and 6·2% [Malawi]) and III (37% [South Africa] and 39% [Malawi]), they do have different prevalence for serotype Ia (30% [South Africa] and 18% [Malawi]).20, 21 In our study, of women not infected with HIV, ten (22%) of 45 in Malawi and 18 (40%) of 45 in South Africa had antibody GMCs more than or equal to the LLQ against serotype Ia at baseline, whereas in the low and high CD4 cell count groups, the numbers of women with GMCs in these ranges were two (4%) of 45 versus one (2%) of 44 in Malawi, and 16 (36%) of 45 versus 22 (49%) of 45 in South Africa, which might be due to this difference in carriage. However, the difference between these countries for serotype Ia should be interpreted with caution because of the small numbers of women in each group. We saw no similar patterns for the other two serotypes in our study.

6%) of 45 versus 22 (49%) of 45 in South Africa, which might be due to this difference in carriage. However, the difference between these countries for serotype Ia should be interpreted with caution because of the small numbers of women in each group. We saw no similar patterns for the other two serotypes in our study. Although these studies had no major methodological differences between the assays used, these data emphasise the need to have an internationally standardised assay for the assessment of the response to GBS conjugate vaccines. In Malawi and South Africa, 18–30% of pregnant women are infected with HIV and their infants are exposed to HIV.22, 23 Since HIV infection has been shown to reduce vaccine efficacy for a number of vaccines irrespective of CD4 cell counts,16 the lower immune response seen in the HIV positive groups compared with the uninfected group is unsurprising. This impaired response in individuals infected with HIV has also been seen for other conjugate vaccines.24

n has been shown to reduce vaccine efficacy for a number of vaccines irrespective of CD4 cell counts,16 the lower immune response seen in the HIV positive groups compared with the uninfected group is unsurprising. This impaired response in individuals infected with HIV has also been seen for other conjugate vaccines.24 Antibody transfer ratios from mother to infant were similar across the HIV-infected and HIV-uninfected groups, and were similar to those ratios reported in previous trials25, 26 of this vaccine, and for other glycoconjugate vaccines given to pregnant mothers. A previous candidate conjugate GBS vaccine against serotype III led to antibody transfer rates of 77%, which was slightly higher than those seen in our study.27 Similarly, rates of antibody decay in infants were similar in our study to those seen in previous studies with candidate GBS vaccines; however, whether the antibody concentrations seen by day 42 would be sufficient to potentially protect against late-onset invasive GBS disease is not known. Baker and colleagues27 reported infant GBS serotype III-specific antibody concentrations of 30% of those recorded at birth at 2 months of age, and previous studies18 with the study vaccine reported infant concentrations of 22–32% of birth levels at 3 months of age. Although the functionality of maternal antibodies was not measured in our study, previous research has shown that functional GBS antibodies can persist for up to 2 years post-immunisation.28 Future vaccine studies would need to assess opsonophagocytic activity in the infant, in conjunction with an established correlate of protection, to estimate potential infant protection. Additionally, further research should concentrate on the responses of women who had undetectable antibody concentrations at baseline and vaccine strategies that might be needed to potentially provide protection to infants born to these women.

established correlate of protection, to estimate potential infant protection. Additionally, further research should concentrate on the responses of women who had undetectable antibody concentrations at baseline and vaccine strategies that might be needed to potentially provide protection to infants born to these women. The assessment of GBS vaccines as they reach late-phase development would be greatly enhanced if the concentration of maternal and neonatal GBS antibody that correlates with passive protection against neonatal disease could be defined. Assay-specific antibody concentrations have been adopted to predict protection for conjugate vaccines against H influenzae type b and S pneumoniae, but although several thresholds for anti-GBS antibody have been proposed on the basis of naturally acquired and vaccine-acquired immunity,29, 30, 31 these thresholds have not been widely accepted, particularly in an African setting, and the absence of a standardised assay across studies makes the selection of a reliable correlate of protection difficult.32 A limitation of the immunogenicity testing in the present study is that the samples were analysed over a long period of time and multiple antigen lots, which could have contributed to the variation recorded. Several laboratories are working to generate a robust and widely applicable assay for approval by the regulatory authorities. Future larger scale studies should therefore focus on the use of a validated assay so that immunogenicity parameters can be readily compared between studies, countries, and over time. Our study is also limited by the small number of participants and the variation in baseline antibody concentrations, which led to differing magnitudes of individual responses to vaccination. Additional, larger studies are needed that include clinically relevant endpoints such as carriage, invasive disease, and mortality, in the appropriate populations. Future research should also include analysis of the effect of a booster vaccine dose, or a higher dose, on responses in pregnant women infected with HIV, or women with undetectable antibody concentrations at baseline.

relevant endpoints such as carriage, invasive disease, and mortality, in the appropriate populations. Future research should also include analysis of the effect of a booster vaccine dose, or a higher dose, on responses in pregnant women infected with HIV, or women with undetectable antibody concentrations at baseline. In conclusion, the investigational glycoconjugate GBS vaccine was less immunogenic in women infected with HIV than those not infected, irrespective of CD4 cell count. The lower amounts of serotype-specific maternal antibody transferred to the infants of women infected with HIV compared with those not infected could reduce vaccine protection from neonatal GBS. At present, this vaccine is undergoing further phase 2 trials and studies of antibody persistence in pregnant women. Once a validated, approved assay and a correlate of protection has been established, researchers will be better able to understand the potential protective value of this vaccine for pregnant women and their infants in high HIV seroprevalence countries. Supplementary Material Supplementary appendix

In conclusion, the investigational glycoconjugate GBS vaccine was less immunogenic in women infected with HIV than those not infected, irrespective of CD4 cell count. The lower amounts of serotype-specific maternal antibody transferred to the infants of women infected with HIV compared with those not infected could reduce vaccine protection from neonatal GBS. At present, this vaccine is undergoing further phase 2 trials and studies of antibody persistence in pregnant women. Once a validated, approved assay and a correlate of protection has been established, researchers will be better able to understand the potential protective value of this vaccine for pregnant women and their infants in high HIV seroprevalence countries. Supplementary Material Supplementary appendix Acknowledgments The authors would like to thank all the administrative, clinical, and laboratory staff at the participating centres and all women and infants who participated in the study. Additionally, the authors would like to thank Jennifer Howie (Novartis Vaccines, now part of the GlaxoSmithKline group of companies) for editorial assistance in production of the manuscript. This study was funded by Novartis Vaccines and Diagnostics, now part of the GlaxoSmithKline group of companies. The MLW Clinical Research Programme is supported by a Strategic Award from the Wellcome Trust, UK.

now part of the GlaxoSmithKline group of companies) for editorial assistance in production of the manuscript. This study was funded by Novartis Vaccines and Diagnostics, now part of the GlaxoSmithKline group of companies. The MLW Clinical Research Programme is supported by a Strategic Award from the Wellcome Trust, UK. Contributors SAM, KS, and NF designed the study; RSH, SAM, NF, CC, BN, DK, RM, AK, LJ, and MO did the study; and RSH, SAM, NF, CC, MO, BN, FW, KS, and PMD were involved in data acquisition, analysis, and interpretation. RSH drafted the manuscript and all authors reviewed, commented on, and approved the final manuscript for submission. Declaration of interests MO, FW, KS, and PMD were permanent employees of the Novartis group of companies at the time of the study. MO is now a permanent employee of the GSK group of companies, following the acquisition of the Novartis human non-influenza vaccines business in March, 2015. NF is in receipt of investigator-led research grants from GlaxoSmithKline. SAM's institution receives grant funding for work on group B streptococcus disease from Novartis and the Bill & Melinda Gates Foundation. All other authors declare no competing interests, except receiving funding for this study from Novartis Vaccines and Diagnostics, Inc. Figure Study profile

Declaration of interests MO, FW, KS, and PMD were permanent employees of the Novartis group of companies at the time of the study. MO is now a permanent employee of the GSK group of companies, following the acquisition of the Novartis human non-influenza vaccines business in March, 2015. NF is in receipt of investigator-led research grants from GlaxoSmithKline. SAM's institution receives grant funding for work on group B streptococcus disease from Novartis and the Bill & Melinda Gates Foundation. All other authors declare no competing interests, except receiving funding for this study from Novartis Vaccines and Diagnostics, Inc. Figure Study profile Most participants who did not complete the study were from the Malawi site, with the exception of the two women who relocated: one woman in the high CD4 cell count group who was lost to follow-up after delivery and one infant in the high CD4 cell count group who died between birth and day 42. Table 1 Demographics and baseline characteristics of enrolled participants

Most participants who did not complete the study were from the Malawi site, with the exception of the two women who relocated: one woman in the high CD4 cell count group who was lost to follow-up after delivery and one infant in the high CD4 cell count group who died between birth and day 42. Table 1 Demographics and baseline characteristics of enrolled participants HIV-infected, low CD4 cell count HIV-infected, high CD4 cell count HIV-uninfected Women n 91 89 90 Age (years) 28·0 (18–39) 28·0 (18–38) 24·0 (18–39) Black, n (%) 91 (100%) 89 (100%) 90 (100%) Body-mass index (kg/m2) 26·0 (17·9–48·7) 24·8 (16·0–43·5) 25·8 (20·3–43·4) Time from vaccination to delivery (weeks) 11 (1–19) 10 (0–18) 9 (1–17) Gestational age at vaccination (weeks) 27·0 (22–35) 29·0 (23–35) 29·5 (24–34) CD4 cell count at baseline (cells per μL) 233 (55–348); n=89 478 (352–1099) NA CD4 cell count at delivery (cells per μL) 262 (30–988); n=81 502 (30–1326); n=80 NA Viral load at baseline (RNA copies per mL) 2760 (20–814 356); n=90 305 (20–334 981); n=88 NA Viral load at delivery (RNA copies per mL) 75 (20–240 965); n=80 75 (20–729 085); n=81 NA Baseline HAART, n (%) 51 (56%) 47 (53%) NA Infants n 91 88 87 Male sex, n (%) 49 (54%) 48 (55%) 46 (53%) Gestational age at birth (weeks) 39 (27–42) 38 (29–44) 39 (29–42) Premature (<37 weeks), n (%) 19 (21%) 25 (28%) 14 (16%) Early premature* (<32 weeks), n (%) 3 (3%) 2 (2%) 1 (1%) Low birthweight (≤2·5 kg), n (%) 12 (13%)* 7 (8%)* 9 (10%) Data are presented as the median and range unless stated otherwise and are from the full analysis set. NA=not applicable. HAART=highly active antiretroviral treatment.

s), n (%) 19 (21%) 25 (28%) 14 (16%) Early premature* (<32 weeks), n (%) 3 (3%) 2 (2%) 1 (1%) Low birthweight (≤2·5 kg), n (%) 12 (13%)* 7 (8%)* 9 (10%) Data are presented as the median and range unless stated otherwise and are from the full analysis set. NA=not applicable. HAART=highly active antiretroviral treatment. * Two (2%) infants in each group were of very low birthweight (≤1·5 kg). Table 2 Number of women reporting local and systemic adverse reactions during the first 7 days after vaccination HIV-infected, low CD4 cell count (n=87) HIV-infected, high CD4 cell count (n=88) HIV-uninfected (n=90) Local Any local 16 (18%) 26 (30%) 35 (39%) Ecchymosis 0 0 1 (1%) Erythema 0 1 (1%) 1 (1%); n=89 Induration 0 1 (1%) 0 Swelling 0 0 2 (2%) Pain Any 16 (18%) 26 (30%) 35 (39%) Severe 2 (2%) 0 4 (4%) Systemic Any systemic 35 (40%) 48 (55%) 53 (59%) Chills Any 8 (9%) 12 (14%) 20 (22%) Severe 1 (1%) 1 (1%) 0 Nausea Any 11 (13%) 15 (17%) 20 (22%); n=89 Severe 1 (1%) 0 2 (2%); n=89 Malaise Any 9 (10%) 17 (19%) 21 (23%) Severe 1 (1%) 0 1 (1%) Myalgia 7 (8%) 15 (17%) 21 (23%) Arthralgia Any 11 (13%) 20 (23%) 26 (29%) Severe 0 0 1 (1%) Headache Any 21 (24%) 28 (32%) 39 (43%) Severe 1 (1%) 1 (1%) 2 (2%) Fatigue Any 21 (24%) 27 (31%) 42 (47%) Severe 2 (2%) 1 (1%) 3 (3%) Rash 3 (3%) 1 (1%) 1 (1%) Fever (≥38°C) 3 (3%) 0 0 For local adverse reactions, no participants reported severe reactions (>100 mm), with the exception of pain. Data are given for local reactions (≥25 mm). Table 3 Number of women and infants reporting unsolicited adverse events and serious adverse events

HIV-infected, low CD4 cell count (n=87) HIV-infected, high CD4 cell count (n=88) HIV-uninfected (n=90) Local Any local 16 (18%) 26 (30%) 35 (39%) Ecchymosis 0 0 1 (1%) Erythema 0 1 (1%) 1 (1%); n=89 Induration 0 1 (1%) 0 Swelling 0 0 2 (2%) Pain Any 16 (18%) 26 (30%) 35 (39%) Severe 2 (2%) 0 4 (4%) Systemic Any systemic 35 (40%) 48 (55%) 53 (59%) Chills Any 8 (9%) 12 (14%) 20 (22%) Severe 1 (1%) 1 (1%) 0 Nausea Any 11 (13%) 15 (17%) 20 (22%); n=89 Severe 1 (1%) 0 2 (2%); n=89 Malaise Any 9 (10%) 17 (19%) 21 (23%) Severe 1 (1%) 0 1 (1%) Myalgia 7 (8%) 15 (17%) 21 (23%) Arthralgia Any 11 (13%) 20 (23%) 26 (29%) Severe 0 0 1 (1%) Headache Any 21 (24%) 28 (32%) 39 (43%) Severe 1 (1%) 1 (1%) 2 (2%) Fatigue Any 21 (24%) 27 (31%) 42 (47%) Severe 2 (2%) 1 (1%) 3 (3%) Rash 3 (3%) 1 (1%) 1 (1%) Fever (≥38°C) 3 (3%) 0 0 For local adverse reactions, no participants reported severe reactions (>100 mm), with the exception of pain. Data are given for local reactions (≥25 mm). Table 3 Number of women and infants reporting unsolicited adverse events and serious adverse events HIV-infected, low CD4 cell count (90 women and 91 infants) HIV-infected, high CD4 cell count (89 women and 88 infants) HIV-uninfected (90 women and 87 infants) Women Any adverse event 67 (74%) 68 (76%) 70 (78%) Adverse event possibly related to vaccine 6 (7%) 12 (13%) 21 (23%) Serious adverse event* 25 (28%) 28 (31%) 29 (32%) Adverse event leading to participant withdrawing from the trial 0 1 (1%) 0 Medically attended adverse event 42 (47%) 44 (49%) 49 (54%) Death† 0 1 (1%) 0 Infants Any adverse event 37 (41%) 43 (49%) 37 (43%) Adverse event possibly related to vaccine 0 2 (2%) 1 (1%) Serious adverse event* 17 (19%) 16 (18%) 16 (18%) Adverse event leading to participant withdrawing from the trial 4 (4%) 2 (2%) 2 (2%) Medically attended adverse event 21 (23%) 29 (33%) 21 (24%) Death† 4 (4%) 2 (2%) 2 (2%) * No serious adverse events were deemed to be at least possibly related to vaccination.

%) 1 (1%) Serious adverse event* 17 (19%) 16 (18%) 16 (18%) Adverse event leading to participant withdrawing from the trial 4 (4%) 2 (2%) 2 (2%) Medically attended adverse event 21 (23%) 29 (33%) 21 (24%) Death† 4 (4%) 2 (2%) 2 (2%) * No serious adverse events were deemed to be at least possibly related to vaccination. † See appendix for details of deaths. Table 4 Maternal geometric mean antibody concentrations and ratios to baseline

%) 1 (1%) Serious adverse event* 17 (19%) 16 (18%) 16 (18%) Adverse event leading to participant withdrawing from the trial 4 (4%) 2 (2%) 2 (2%) Medically attended adverse event 21 (23%) 29 (33%) 21 (24%) Death† 4 (4%) 2 (2%) 2 (2%) * No serious adverse events were deemed to be at least possibly related to vaccination. † See appendix for details of deaths. Table 4 Maternal geometric mean antibody concentrations and ratios to baseline HIV-infected, low CD4 cell count HIV-infected, high CD4 cell count HIV-uninfected Low CD4vshigh CD4 Low CD4vsHIV-uninfected High CD4vsHIV-uninfected Serotype Ia Data available, n 74 75 77 .. .. .. Geometric mean antibody concentration (μg/mL) Day 1 pre-vaccination 0·24 (0·19–0·30) 0·25 (0·20–0·32) 0·38 (0·30–0·48) p=0·709 p=0·007 p=0·019 Day 15 post-vaccination 2·62 (1·64–4·17) 2·95 (1·86–4·69) 5·61 (3·56–8·84) p=0·713 p=0·023 p=0·053 Day 31 post-vaccination 2·68 (1·74–4·10) 3·26 (2·14–4·98) 6·63 (4·37–10) p=0·513 p=0·003 p=0·020 Delivery 2·22 (1·50–3·29) 2·69 (1·82–3·98) 4·49 (3·06–6·60) p=0·493 p=0·013 p=0·067 Geometric mean ratio to day 1 Day 15 8·36 (5·27–13) 9·54 (6·02–15) 19 (12–30) p=0·690 p=0·011 p=0·032 Day 31 8·55 (5·60–13) 11 (6·91–16) 23 (15–35) p=0·492 p=0·001 p=0·010 Delivery 7·21 (4·88–11) 8·78 (5·95–13) 15 (10–22) p=0·479 p=0·007 p=0·046 Serotype Ib Data available, n 44 53 66 .. .. .. Geometric mean antibody concentration (μg/mL) Day 1 pre-vaccination 0·51 (0·38–0·70) 0·36 (0·27–0·48) 0·40 (0·31–0·51) p=0·104 p=0·205 p=0·652 Day 15 post-vaccination 2·93 (1·73–4·95) 3·50 (2·17–5·65) 6·07 (3·98–9·26) p=0·617 p=0·035 p=0·090 Day 31 post-vaccination 2·62 (1·62–4·24) 3·68 (2·38–5·70) 5·35 (3·63–7·87) p=0·300 p=0·024 p=0·209 Delivery 2·12 (1·36–3·31) 3·04 (2·03–4·56) 3·84 (2·69–5·50) p=0·236 p=0·042 p=0·392 Geometric mean ratio to day 1 Day 15 6·87 (4·08–12) 8·23 (5·12–13) 14 (9·36–22) p=0·611 p=0·033 p=0·089 Day 31 6·11 (3·79–9·85) 8·70 (5·63–13) 13 (8·56–19) p=0·277 p=0·021 p=0·211 Delivery 4·83 (3·10–7·52) 7·34 (4·90–11) 9·14 (6·38–13) p=0·165 p=0·029 p=0·425 Serotype III Data available, n 53 55 70 .. .. ..

042 p=0·392 Geometric mean ratio to day 1 Day 15 6·87 (4·08–12) 8·23 (5·12–13) 14 (9·36–22) p=0·611 p=0·033 p=0·089 Day 31 6·11 (3·79–9·85) 8·70 (5·63–13) 13 (8·56–19) p=0·277 p=0·021 p=0·211 Delivery 4·83 (3·10–7·52) 7·34 (4·90–11) 9·14 (6·38–13) p=0·165 p=0·029 p=0·425 Serotype III Data available, n 53 55 70 .. .. .. Geometric mean antibody concentration (μg/mL) Day 1 pre-vaccination 0·12 (0·09–0·17) 0·10 (0·07–0·14) 0·14 (0·10–0·18) p=0·407 p=0·572 p=0·145 Day 15 post-vaccination 1·24 (0·79–1·95) 1·52 (0·97–2·36) 5·90 (3·99–8·72) p=0·528 p<0·0001 p<0·0001 Day 31 post-vaccination 1·51 (0·97–2·35) 1·31 (0·85–2·02) 5·35 (3·66–7·83) p=0·643 p<0·0001 p<0·0001 Delivery 1·25 (0·81–1·94) 1·07 (0·70–1·65) 3·80 (2·61–5·55) p=0·616 p=0·0002 p<0·0001 Geometric mean ratio to day 1 Day 15 9·85 (6·25–16) 12 (7·64–18) 47 (32–70) p=0·555 p<0·0001 p<0·0001 Day 31 12 (7·70–19) 10 (6·62–16) 43 (29–63) p=0·599 p<0·0001 p<0·0001 Delivery 9·99 (6·44–15) 8·65 (5·65–13) 30 (21–44) p=0·638 p=0·0002 p<0·0001 Data are geometric means (95% CI), unless otherwise stated. p values for pairwise comparisons are given in the last three columns (ANCOVA). Table 5 Infant antibody GMCs, GMRs, and mean placental transfer ratios

Geometric mean antibody concentration (μg/mL) Day 1 pre-vaccination 0·12 (0·09–0·17) 0·10 (0·07–0·14) 0·14 (0·10–0·18) p=0·407 p=0·572 p=0·145 Day 15 post-vaccination 1·24 (0·79–1·95) 1·52 (0·97–2·36) 5·90 (3·99–8·72) p=0·528 p<0·0001 p<0·0001 Day 31 post-vaccination 1·51 (0·97–2·35) 1·31 (0·85–2·02) 5·35 (3·66–7·83) p=0·643 p<0·0001 p<0·0001 Delivery 1·25 (0·81–1·94) 1·07 (0·70–1·65) 3·80 (2·61–5·55) p=0·616 p=0·0002 p<0·0001 Geometric mean ratio to day 1 Day 15 9·85 (6·25–16) 12 (7·64–18) 47 (32–70) p=0·555 p<0·0001 p<0·0001 Day 31 12 (7·70–19) 10 (6·62–16) 43 (29–63) p=0·599 p<0·0001 p<0·0001 Delivery 9·99 (6·44–15) 8·65 (5·65–13) 30 (21–44) p=0·638 p=0·0002 p<0·0001 Data are geometric means (95% CI), unless otherwise stated. p values for pairwise comparisons are given in the last three columns (ANCOVA). Table 5 Infant antibody GMCs, GMRs, and mean placental transfer ratios HIV-infected, low CD4 cell count HIV-infected, high CD4 cell count HIV-uninfected Low CD4vshigh CD4 Low CD4vsHIV-uninfected High CD4vsHIV-uninfected n Data n Data n Data Serotype Ia Transfer ratio* 79 0·58 (0·49–0·69) 79 0·60 (0·51–0·72) 83 0·72 (0·61–0·85) p=0·755 p=0·089 p=0·169 GMC day 1 (μg/mL) 79 1·01 (0·66–1·56) 81 1·22 (0·80–1·87) 83 3·91 (2·56–5·96) p=0·548 p<0·0001 p=0·0002 GMC day 42 (μg/mL) 81 0·61 (0·41–0·89) 83 0·64 (0·44–0·94) 79 1·97 (1·33–2·91) p=0·848 p<0·0001 p<0·0001 GMR day 42:day 1 70 0·58 (0·50–0·67) 78 0·50 (0·44–0·57) 75 0·50 (0·44–0·58) p=0·296 p=0·344 p=0·925 Serotype Ib Transfer ratio* 41 0·51 (0·38–0·69) 54 0·64 (0·50–0·83) 57 0·49 (0·38–0·63) p=0·251 p=0·796 p=0·131 GMC day 1 (μg/mL) 44 1·31 (0·78–2·19) 56 1·62 (1·03–2·56) 57 2·67 (1·70–4·20) p=0·537 p=0·040 p=0·125 GMC day 42 (μg/mL) 75 0·29 (0·19–0·44) 70 0·44 (0·29–0·68) 67 1·16 (0·75–1·80) p=0·174 p<0·0001 p=0·002 GMR day 42:day 1 34 0·28 (0·20–0·38) 46 0·35 (0·27–0·46) 46 0·56 (0·42–0·73) p=0·297 p=0·0006 p=0·008 Serotype III Transfer ratio* 53 0·60 (0·44–0·82) 44 0·51 (0·36–0·72) 66 0·56 (0·43–0·75) p=0·515 p=0·791 p=0·670 GMC day 1 (μg/mL) 54 0·60 (0·36–0·99) 51 0·52 (0·31–0·88) 66 3·88 (2·47–6·10) p=0·713 p<0·0001 p<0·0001 GMC day 42 (μg/mL) 77 0·21 (0·14–0·31) 80 0·15 (0·10–0·22) 77 0·86 (0·58–1·28) p=0·258 p<0·0001 p<0·0001 GMR day 42:day 1 43 0·39 (0·32–0·49) 46 0·40 (0·33–0·49) 59 0·31 (0·26–0·38) p=0·574 p=0·186 p=0·051 95% CIs are given in parenthesis. p values for pairwise comparisons are given in the last three columns (ANOVA for transfer ratio and ANCOVA for geometric mean antibody concentrations and GMRs). GMC=geometric mean antibody concentrations. GMR=ratio of GMCs to baseline.

–0·49) 59 0·31 (0·26–0·38) p=0·574 p=0·186 p=0·051 95% CIs are given in parenthesis. p values for pairwise comparisons are given in the last three columns (ANOVA for transfer ratio and ANCOVA for geometric mean antibody concentrations and GMRs). GMC=geometric mean antibody concentrations. GMR=ratio of GMCs to baseline. * Transfer ratio=infant GMC in blood collected within 72 h of birth divided by maternal GMC in blood collected at delivery.

Introduction More than 35 million people worldwide are living with HIV-1.1 There is no effective vaccine and therefore control of the HIV pandemic relies heavily on combination antiretroviral therapy (cART). WHO treatment guidelines for adult HIV-1 infection recommend the nucleotide reverse-transcriptase inhibitor (NRTI) tenofovir for first-line ART, in combination with lamivudine or emtricitabine and the non-nucleoside reverse-transcriptase inhibitor (NNRTI) efavirenz.2 Older NRTIs such as the thymidine analogue drugs are being replaced by tenofovir and the NNRTI nevirapine, although mentioned in WHO guidelines, is being phased out from first-line regimens.2 The global scale-up of cART has now reached 15 million treated individuals.1 The administration of cART at the time individuals with HIV-1 are initially diagnosed prevents immunological deterioration as early as possible and interrupts the spread of HIV-1 from newly diagnosed individuals.3 This strategy, referred to as treatment as prevention, is being studied especially in high-incidence regions and nearly always includes the use of first-line tenofovir-containing ART regimens. Likewise, the strategy of pre-exposure prophylaxis (PrEP) depends entirely on the administration of tenofovir or tenofovir and emtricitabine to uninfected individuals at high risk of HIV-1 infection.4

specially in high-incidence regions and nearly always includes the use of first-line tenofovir-containing ART regimens. Likewise, the strategy of pre-exposure prophylaxis (PrEP) depends entirely on the administration of tenofovir or tenofovir and emtricitabine to uninfected individuals at high risk of HIV-1 infection.4 In individuals receiving tenofovir, HIV-1 develops phenotypically and clinically significant resistance usually as a result of one mutation at position 65 (lysine to arginine; K65R) in the reverse transcriptase (RT) gene.5 Data from clinical trials and cohorts in high-income settings using tenofovir combined with NNRTI have reported low prevalence of tenofovir resistance at viral failure,6, 7, 8 in stark contrast with reports from low-income and middle-income countries where prevalence seems to be much higher.9, 10 Similarly, high-level resistance to NNRTI and the cytosine analogue component (emtricitabine and lamivudine) arise through changes to one aminoacid, which suggests a low genetic barrier to resistance for these drugs as well. In view of the pivotal role of tenofovir-containing ART as both treatment and prophylaxis, and the striking potential for drug resistance, we did a global assessment of drug resistance after virological failure with first-line tenofovir-containing ART. Research in context Evidence before this study

In individuals receiving tenofovir, HIV-1 develops phenotypically and clinically significant resistance usually as a result of one mutation at position 65 (lysine to arginine; K65R) in the reverse transcriptase (RT) gene.5 Data from clinical trials and cohorts in high-income settings using tenofovir combined with NNRTI have reported low prevalence of tenofovir resistance at viral failure,6, 7, 8 in stark contrast with reports from low-income and middle-income countries where prevalence seems to be much higher.9, 10 Similarly, high-level resistance to NNRTI and the cytosine analogue component (emtricitabine and lamivudine) arise through changes to one aminoacid, which suggests a low genetic barrier to resistance for these drugs as well. In view of the pivotal role of tenofovir-containing ART as both treatment and prophylaxis, and the striking potential for drug resistance, we did a global assessment of drug resistance after virological failure with first-line tenofovir-containing ART. Research in context Evidence before this study We searched PubMed for studies of the prevalence of tenofovir resistance after failure of first-line antiretroviral therapy with efavirenz or nevirapine (non-nucleoside reverse-transcriptase inhibitors [NNRTIs]) in patients with HIV-1, published between January, 1999, and June, 2015, using the search terms “HIV” AND “tenofovir” AND “resistance”. We identified studies done in untreated adults (age >15 years) in which either efavirenz or nevirapine was combined with tenofovir and either emtricitabine or lamivudine as first line antiretroviral therapy. Several studies reported resistance data for tenofovir when the drug was started after initial use of stavudine or zidovudine; these studies were not reviewed further. We also excluded studies that reported tenofovir use without NNRTI because standard first-line antiretroviral therapy under a public health approach is based on NNRTI in adults.

resistance data for tenofovir when the drug was started after initial use of stavudine or zidovudine; these studies were not reviewed further. We also excluded studies that reported tenofovir use without NNRTI because standard first-line antiretroviral therapy under a public health approach is based on NNRTI in adults. We identified randomised controlled trials and a meta-analysis comparing NNRTI with protease inhibitors, in combination with tenofovir, which reported resistance data. Patients in high-income settings reported tenofovir resistance in 0–25% of virological failures and those in sub-Saharan Africa in 28–50%. The only other prospective study in sub-Saharan Africa was PASER-M, and was limited by few resistance data for patients given tenofovir plus NNRTI-based combination antiretroviral therapy (cART). The remaining studies were largely from South Africa and reported a wide range of prevalence (between 23% and 70%) of tenofovir resistance after virological failure. In west Africa, one study reported that 57% of virological failures were tenofovir resistant in a very small sample of 23 patients. Although aforementioned studies also reported NNRTI and cytosine analogue resistance, they were unable to quantify to what extent tenofovir resistance was a marker for high-level compromise of the regimen. We found no studies that specifically reported resistance data for patients given first-line tenofovir in east Africa. No study reported resistance data from more than one continent, and none seemed adequately powered to establish the effect of co-administered reverse-transcriptase inhibitors on the emergence of tenofovir resistance.

dies that specifically reported resistance data for patients given first-line tenofovir in east Africa. No study reported resistance data from more than one continent, and none seemed adequately powered to establish the effect of co-administered reverse-transcriptase inhibitors on the emergence of tenofovir resistance. Added value of this study This study reports the most comprehensive assessment of HIV-1 drug resistance after scale-up of first-line WHO recommended tenofovir-based antiretroviral regimens, showing that tenofovir resistance is surprisingly common in patients with treatment failure across many studies in all low-income regions. Importantly, these individuals also have notable resistance to other drugs in their regimen, leading to almost complete compromise of combination treatment. Challenging current perceptions in the specialty, our findings show that tenofovir resistant viruses have substantial transmission potential. Furthermore, our results show that viral strain affects tenofovir resistance in Europe but is not the main driver for resistance in viruses circulating in sub-Saharan Africa. Newly identified risk factors for resistance to tenofovir and NNRTI drugs include pre-treatment CD4 cell count (but not viral load) and co-administered antiretrovirals. Implications of all the available evidence

This study reports the most comprehensive assessment of HIV-1 drug resistance after scale-up of first-line WHO recommended tenofovir-based antiretroviral regimens, showing that tenofovir resistance is surprisingly common in patients with treatment failure across many studies in all low-income regions. Importantly, these individuals also have notable resistance to other drugs in their regimen, leading to almost complete compromise of combination treatment. Challenging current perceptions in the specialty, our findings show that tenofovir resistant viruses have substantial transmission potential. Furthermore, our results show that viral strain affects tenofovir resistance in Europe but is not the main driver for resistance in viruses circulating in sub-Saharan Africa. Newly identified risk factors for resistance to tenofovir and NNRTI drugs include pre-treatment CD4 cell count (but not viral load) and co-administered antiretrovirals. Implications of all the available evidence Improvements in the quality of HIV care and viral load monitoring could mitigate the emergence and spread of tenofovir resistance, thereby prolonging the lifetime of tenofovir-containing regimens for both treatment and prophylaxis. Surveillance of tenofovir and NNRTI resistance should be a priority both in untreated and treated populations.

ality of HIV care and viral load monitoring could mitigate the emergence and spread of tenofovir resistance, thereby prolonging the lifetime of tenofovir-containing regimens for both treatment and prophylaxis. Surveillance of tenofovir and NNRTI resistance should be a priority both in untreated and treated populations. Methods Study population and design The TenoRes collaboration comprises adult HIV treatment cohorts and clinical trials from Europe, Latin and North America, sub-Saharan Africa, and Asia. Cohorts and trials were identified by RWS and RKG as those known to do genotypic resistance testing through previous collaborations, the WHO HIV Drug Resistance Network, and through the International HIV Drug Resistance Workshop. Moreover, we did a systematic review using the keywords “HIV”, AND “tenofovir” AND “resistance” in PubMed for articles published between January, 1999, and June, 2015. We identified 44 studies suitable for the reported analysis after applying the following inclusion criteria: documented virological failure after first-line ART comprising tenofovir plus either lamivudine or emtricitabine plus either efavirenz or nevirapine (virological failure was defined by local viral load thresholds or surveillance protocols); a successful resistance test result associated with virological failure of cART; tenofovir-based ART for at least 4 months before virological failure; and absence of thymidine analogue mutations at resistance testing (appendix). Exclusion criteria were: studies reporting resistance data after tenofovir that was started after initial use of stavudine or zidovudine; and studies reporting tenofovir use without NNRTI. Data were extracted and harmonised by RWS, RKG, MT, and JG and stored in a central database.

ions at resistance testing (appendix). Exclusion criteria were: studies reporting resistance data after tenofovir that was started after initial use of stavudine or zidovudine; and studies reporting tenofovir use without NNRTI. Data were extracted and harmonised by RWS, RKG, MT, and JG and stored in a central database. We collected individual-level data for a predefined set of covariates: age at first-line ART initiation, sex, frequency of viral load monitoring (number of tests per year), urban versus rural setting for HIV clinics, viral load threshold for virological failure and genotyping, co-administered antiretrovirals, duration of treatment, viral load and CD4 cell count before the start of first-line ART (baseline) and at time of viral failure, and resistance mutations based on the Stanford HIV Drug Resistance Database.

ing for HIV clinics, viral load threshold for virological failure and genotyping, co-administered antiretrovirals, duration of treatment, viral load and CD4 cell count before the start of first-line ART (baseline) and at time of viral failure, and resistance mutations based on the Stanford HIV Drug Resistance Database. Statistical analysis Our primary outcome was tenofovir resistance, defined as presence of K65R/N or K70E/G/Q mutations in the RT gene. Our secondary outcomes were resistance to first generation NNRTI (efavirenz and nevirapine), defined as specific mutations at aminoacid positions 100, 103, 106, 108, 181, 188, 190, and 225,11 and cytosine analogue resistance, defined as presence of M184V/I. Our main exposures of interest were baseline CD4 cell count (<100 vs ≥100 cells per μL), baseline viral load (<100 000 vs ≥100 000 copies HIV-1 RNA per mL; this cutoff was chosen because of findings from previous studies12), nevirapine versus efavirenz, and lamivudine versus emtricitabine. For our primary analysis, we estimated the odds ratios (ORs) for tenofovir resistance within each study before pooling estimates across studies using a random-effects meta-analysis with DerSimonian-Laird weighting and estimates of heterogeneity taken from the Mantel-Haenszel model. We chose this method to ensure that we only compared patients in the same study and country, thereby minimising confounding by differences in care at the study or country level. Findings were not sensitive to the choice of method used for the meta-analysis (ie, fixed or random effects). We also used a continuity correction of 0·5 for counts of 0, although findings were not sensitive to this choice.

nd country, thereby minimising confounding by differences in care at the study or country level. Findings were not sensitive to the choice of method used for the meta-analysis (ie, fixed or random effects). We also used a continuity correction of 0·5 for counts of 0, although findings were not sensitive to this choice. We did sensitivity analyses to investigate whether associations changed when adjusted for possible confounders. Because of the sparseness of data in many studies, we were unable to adjust within-study associations for potential confounders. Instead, we did additional analyses using logistic regression models with a random effect at study level to estimate associations before and after adjustment for possible confounders in a common subset of participants. To build the adjusted model, we included each of our main exposures and HIV subtype. We also considered for inclusion individual-level information about age, sex, year of treatment initiation, and length of time on tenofovir, but rejected these covariates because of a lack of any univariate association with tenofovir resistance. We chose to only use these models for working out the likely extent of confounding, because estimated associations from these models are partly derived from between-study comparisons.

nd length of time on tenofovir, but rejected these covariates because of a lack of any univariate association with tenofovir resistance. We chose to only use these models for working out the likely extent of confounding, because estimated associations from these models are partly derived from between-study comparisons. To clarify whether the association between baseline CD4 or baseline viral load and tenofovir resistance was linear (ie, followed a dose-response pattern), we categorised participants into four categories based on baseline CD4 cell count (<100, 100–200, 201–300, >300 cells per μL reference category) or baseline viral load (<25 000 [reference]; 25 001–100 000; 100 001–300 000; >300 000 copies HIV-1 RNA per mL). We assessed associations by plotting the estimated OR against the mean level of baseline CD4 (or baseline viral load), in a random-effects logistic regression model adjusted for region, co-administered drugs, and baseline viral load (or baseline CD4). To assess the potential transmissibility of mutant viruses, we graphically compared the distribution of plasma HIV-1 RNA concentrations of patients from the same study with and without tenofovir resistance.

To clarify whether the association between baseline CD4 or baseline viral load and tenofovir resistance was linear (ie, followed a dose-response pattern), we categorised participants into four categories based on baseline CD4 cell count (<100, 100–200, 201–300, >300 cells per μL reference category) or baseline viral load (<25 000 [reference]; 25 001–100 000; 100 001–300 000; >300 000 copies HIV-1 RNA per mL). We assessed associations by plotting the estimated OR against the mean level of baseline CD4 (or baseline viral load), in a random-effects logistic regression model adjusted for region, co-administered drugs, and baseline viral load (or baseline CD4). To assess the potential transmissibility of mutant viruses, we graphically compared the distribution of plasma HIV-1 RNA concentrations of patients from the same study with and without tenofovir resistance. We did not use multiple imputation to adjust for missing data because most missing data were the result of a lack of availability at the study level. Instead, we restricted analyses to the subset of participants with information available about all relevant covariates for each specific analysis. The appendix presents the amount of missing data and which studies contributed towards specific analyses. We used Stata (version 11.2) for all analyses.

at the study level. Instead, we restricted analyses to the subset of participants with information available about all relevant covariates for each specific analysis. The appendix presents the amount of missing data and which studies contributed towards specific analyses. We used Stata (version 11.2) for all analyses. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. RKG and JG had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results The TenoRes collaboration included 1926 individuals from 36 countries (figure 1 and appendix). Table 1 summarises the median size and year of ART initiation for the cohorts comprising the collaboration. Viral load monitoring was done in about 50% of the cohorts including nearly all of cohorts from upper-income regions and from a small proportion of the cohorts in low-income and middle-income countries (appendix shows income status for each cohort; table 1).

ART initiation for the cohorts comprising the collaboration. Viral load monitoring was done in about 50% of the cohorts including nearly all of cohorts from upper-income regions and from a small proportion of the cohorts in low-income and middle-income countries (appendix shows income status for each cohort; table 1). The region-level pre-ART median CD4 cell count ranged from 44 to 104  cells per μL in sub-Saharan Africa, Asia, and Latin America (table 2). As expected, in north America pre-ART median CD4 cell count was 144 cells per μL and 190  cells per μL in Europe. The proportion of individuals using emtricitabine (vs lamivudine) and efavirenz (vs nevirapine) varied significantly by region. Emtricitabine was used significantly more than lamivudine in Europe, North America, and west and central Africa, and efavirenz was used significantly more than nevirapine in all regions apart from east and west and central Africa. The median duration of ART ranged from 11 to 26 months. Pre-treatment viral load ranged between 4·80 and 5·58 log copies per mL and was significantly higher in eastern and western and central Africa and Latin America than the other regions (table 2).

evirapine in all regions apart from east and west and central Africa. The median duration of ART ranged from 11 to 26 months. Pre-treatment viral load ranged between 4·80 and 5·58 log copies per mL and was significantly higher in eastern and western and central Africa and Latin America than the other regions (table 2). Crude prevalence of tenofovir resistance in patients with treatment failure was highest in low-income and middle-income regions (figure 1). Prevalence of cytosine analogue resistance (M184V/I) was highest in sub-Saharan Africa and Latin America and lowest in western Europe. By contrast, resistance to NNRTI did not show this pattern (figure 1). Furthermore, the M184V/I mutation was less common than NNRTI resistance across all regions except in eastern Africa. Of the 700 patients with tenofovir resistance in the dataset, 457 (65%) had resistance to both remaining drugs. Participants with tenofovir resistant viruses were likely to be resistant to one or both accompanying drugs and therefore have profound compromise of their regimen, as compared with those without tenofovir resistance (figure 1).

with tenofovir resistance in the dataset, 457 (65%) had resistance to both remaining drugs. Participants with tenofovir resistant viruses were likely to be resistant to one or both accompanying drugs and therefore have profound compromise of their regimen, as compared with those without tenofovir resistance (figure 1). Low baseline CD4 cell count was consistently associated with a higher prevalence of tenofovir resistance across regions. The pooled OR for tenofovir in individuals with a CD4 cell count of less than 100 cells per μL versus 100  cells per μL was 1·50 (95% CI 1·27–1·77; figure 2). By contrast, a high baseline viral load was only associated with a small, not significant increase in tenofovir resistance (OR for viral load ≥100 000 copies per mL vs <100 000 copies per mL was 1·17, 95% CI 0·94–1·44; appendix). We compared tenofovir resistance by use of co-administered antiretrovirals with tenofovir as first-line therapy. Use of lamivudine rather than emtricitabine (NRTIs) was associated with a higher prevalence of tenofovir resistance (OR 1·48, 95% CI 1·20–1·82), as was use of the NNRTI nevirapine rather than efavirenz (OR 1·46, 1·28–1·67; appendix). Subgroup analysis showed that as well as associations being consistent across regions, they were also generally similar across a range of study settings (appendix), although there was some evidence of a greater effect size of baseline CD4 when efavirenz was co-administered with tenofovir, as compared with nevirapine.

. Subgroup analysis showed that as well as associations being consistent across regions, they were also generally similar across a range of study settings (appendix), although there was some evidence of a greater effect size of baseline CD4 when efavirenz was co-administered with tenofovir, as compared with nevirapine. When considering the effect of baseline CD4, baseline viral load (figure 3), and co-administered antiretrovirals (appendix) on cytosine analogue and NNRTI resistance, we noted that the magnitude of associations were smaller than those recorded for tenofovir resistance. We also assessed the relation between viral subtype C on acquisition of tenofovir resistance. We restricted this analysis to western European studies in view of the consistent standard of care available in this region and relatively lower level of subtype diversity in other regions (figure 1A). We also limited the comparison to subtypes found in immigrant populations to minimise bias due to socioeconomic factors (thereby excluding subtype B infections mainly recorded in participants born in western Europe). Tenofovir resistance was higher in subtype C compared with non-C, non-B infections with a pooled OR of 2·44 (1·66–3·59).

he comparison to subtypes found in immigrant populations to minimise bias due to socioeconomic factors (thereby excluding subtype B infections mainly recorded in participants born in western Europe). Tenofovir resistance was higher in subtype C compared with non-C, non-B infections with a pooled OR of 2·44 (1·66–3·59). As a sensitivity analysis we studied risk factors for tenofovir resistance using univariate (adjusted only for region) and multivariate logistic regression analyses (appendix). We noted a dose-response relationship for baseline CD4, which was not markedly altered by adjustment for baseline viral load, viral subtype, or type of co-administered drug used (appendix). Baseline viral load of 100 000 or more copies of HIV-1 RNA per mL was not significantly associated with tenofovir resistance (OR 1·31, 95% CI 0·91–1·91) and we noted no clear trend across increasing viral loads (appendix). Adjustment for several risk factors also had little effect on associations with tenofovir resistance of emtricitabine versus lamivudine and nevirapine versus efavirenz.

ot significantly associated with tenofovir resistance (OR 1·31, 95% CI 0·91–1·91) and we noted no clear trend across increasing viral loads (appendix). Adjustment for several risk factors also had little effect on associations with tenofovir resistance of emtricitabine versus lamivudine and nevirapine versus efavirenz. Finally, we compared the viral load at treatment failure in the presence and absence of tenofovir-associated mutations. The mean plasma viral load at treatment failure was not different in the presence or absence of tenofovir associated mutations (145 700 copies HIV RNA per mL [SE 12 480] vs 133 900 copies [SE 16 650]; p=0·626; figure 4 shows the within-study viral load by region). These results did not change when analysis was restricted to individuals who had evidence of the K65R mutation, either with or without M184V/I (appendix). Mutations at aminoacids K65 and M184 in the RT gene have been associated with suboptimum replication.13

0·626; figure 4 shows the within-study viral load by region). These results did not change when analysis was restricted to individuals who had evidence of the K65R mutation, either with or without M184V/I (appendix). Mutations at aminoacids K65 and M184 in the RT gene have been associated with suboptimum replication.13 Discussion Our study has three main findings relating to the prevalence, risk factors for, and transmissibility of tenofovir resistance. First, we noted that levels of tenofovir resistance in individuals with viral failure ranged from 20% in Europe to more than 50% in sub-Saharan Africa. Second, a CD4 cell count of less than 100 cells per μL, treatment with nevirapine rather than efavirenz, and treatment with lamivudine rather than emtricitabine, were consistently associated with a 50% higher odds of tenofovir resistance in those with viral failure. Third, we noted that in patients with viral failure, viral loads were similar in the presence or absence of tenofovir resistance.

apine rather than efavirenz, and treatment with lamivudine rather than emtricitabine, were consistently associated with a 50% higher odds of tenofovir resistance in those with viral failure. Third, we noted that in patients with viral failure, viral loads were similar in the presence or absence of tenofovir resistance. Our findings are important in view of the fact that following WHO recommendations,2 tenofovir is replacing thymidine analogues (zidovudine and stavudine) as part of the NRTI backbone in first-line regimens in resource-limited settings. Every drug in these regimens can be compromised by one aminoacid mutation, and the combination therapy is therefore potentially fragile. In view of the crucial role of tenofovir-containing ART in both treatment and prevention of new infections, restriction of drug resistance in high-burden settings is of paramount importance. Understanding how common tenofovir resistance is, and how and why it varies, is key to its prevention. Although our risk factors are only associated with a modest 50% increase in odds, this translates to a roughly 10% increase in resistance in those who fail when the overall tenofovir resistance prevalence is about 50% (as recorded in sub-Saharan Africa).

sistance is, and how and why it varies, is key to its prevention. Although our risk factors are only associated with a modest 50% increase in odds, this translates to a roughly 10% increase in resistance in those who fail when the overall tenofovir resistance prevalence is about 50% (as recorded in sub-Saharan Africa). We hypothesise that the regional differences in tenofovir resistance are due to the frequency of viral load monitoring with close patient follow-up and feedback of results. For example, although viral load monitoring is not routinely done in most low-income and middle-income countries, in high-income countries viral load is tested three to four times per year with close patient follow-up and adherence support. Such an approach is likely to lead to earlier detection of viral failure, before selection of drug resistance mutations against tenofovir has occurred.14 This view is supported by the uncommon detection of drug resistance mutations in specimens with low viral load (400–1000 copies per mL) from patients given tenofovir in both high-income settings (figure 1; see higher prevalence of tenofovir resistance where viral load >1000 copies per mL is used as threshold in western Europe)15 and sub-Saharan Africa (Chunfu Yang, Centres for Disease Control, Atlanta, GA, USA, personal communication). Tenofovir resistance could be limited by viral load monitoring,16 with rapid feedback to clinicians followed by adherence counselling to preserve first line, or switch to second line when this approach fails. Furthermore, pre-ART (baseline) resistance testing for key NNRTI mutations could potentially protect against tenofovir resistance by avoiding use of partly active treatment regimens. In our report, transmitted NNRTI resistance was low in the regions studied (<10%),17 and therefore not likely to be a major driver of wide variation in drug resistance across income settings.

key NNRTI mutations could potentially protect against tenofovir resistance by avoiding use of partly active treatment regimens. In our report, transmitted NNRTI resistance was low in the regions studied (<10%),17 and therefore not likely to be a major driver of wide variation in drug resistance across income settings. Other factors that vary geographically could also affect success of ART and should be noted. Treatment failure is associated not only with drug resistance, but also side-effects. Efavirenz is associated with CNS side-effects such as sleep disturbance and is associated with treatment discontinuation.18 Furthermore, drug stock-outs and other indicators of quality of HIV services that have shown geographic variation would also predispose to treatment failure.19 The issue of regional variation in adherence levels has received considerable attention, with data from several studies suggesting that adherence is not worse in sub-Saharan Africa compared with North America.20, 21

ality of HIV services that have shown geographic variation would also predispose to treatment failure.19 The issue of regional variation in adherence levels has received considerable attention, with data from several studies suggesting that adherence is not worse in sub-Saharan Africa compared with North America.20, 21 With regards to increased tenofovir resistance in individuals with low baseline CD4 counts, this finding is consistent with results from the ACTG 5202 trial22 suggesting higher frequency of RT mutations in patients given ART with low CD4 cell counts, and offer a benefit of CD4 cell count measurement after diagnosis of HIV infection beyond establishing prophylaxis against opportunistic infections.23 Lamivudine warrants further study in first-line regimens in view of data presented in our study and the conflicting reports regarding virological efficacy of lamivudine versus emtricitabine.24, 25, 26 Of note, the differences between lamivudine and emtricitabine might become less important in high-income regions where implementation of the second generation integrase inhibitor dolutegravir occurs, in view of the fact that this agent has not been associated with any cytosine analogue resistance at virological failure.27

note, the differences between lamivudine and emtricitabine might become less important in high-income regions where implementation of the second generation integrase inhibitor dolutegravir occurs, in view of the fact that this agent has not been associated with any cytosine analogue resistance at virological failure.27 Viral load has been associated with transmission risk.28 Despite evidence for diminished replication of tenofovir resistant viruses (containing the K65R mutation in the RT gene) in vitro, we noted similar viral loads in participants with and without tenofovir resistance. Therefore, there might be substantial potential for onward transmission to uninfected individuals,29 despite little evidence of K65R transmission up to now.30 This finding reinforces the need for drug resistance surveillance activities in both untreated and treated HIV-positive individuals.

ithout tenofovir resistance. Therefore, there might be substantial potential for onward transmission to uninfected individuals,29 despite little evidence of K65R transmission up to now.30 This finding reinforces the need for drug resistance surveillance activities in both untreated and treated HIV-positive individuals. There are several important limitations of our study. First, because we only included patients with virological failure related to existing study cohorts,1 our estimates of the prevalence of tenofovir resistance might not be representative in certain high-burden regions. Although this situation might have biased our findings on absolute prevalences of tenofovir resistance, it is unlikely to have affected associations with baseline CD4 or co-administered drugs. Second, we only included patients at failure so were unable to assess overall rates of tenofovir resistance in all patients starting first-line treatment. We used this method because many of the contributing studies had no clear denominator, especially those done in resource-limited settings. However, extensive WHO-led analysis reported that 15–35% (on treatment vs intention to treat) of patients in sub-Saharan Africa have virological failure by 12 months.31 Therefore, using a conservative 50% prevalence of tenofovir resistance at failure from our analysis, we suggest that it is likely that 7·5–17·5% of individuals given tenofovir plus cytosine analogue plus efavirenz will develop tenofovir resistance within 1 year of treatment initiation under present practices in sub-Saharan Africa.

sing a conservative 50% prevalence of tenofovir resistance at failure from our analysis, we suggest that it is likely that 7·5–17·5% of individuals given tenofovir plus cytosine analogue plus efavirenz will develop tenofovir resistance within 1 year of treatment initiation under present practices in sub-Saharan Africa. Third, our findings on risk factors for tenofovir resistance were derived from an unadjusted meta-analysis involving very different study populations. Although this enhances the generalisability of results, it has the potential to lead to biased comparisons. However, we took measures to minimise biases. We exclusively used within-study and within-country comparisons for our primary analyses, thereby ensuring that comparisons were for participants undergoing similar treatment monitoring practices. We tested associations between risk factors and found that they were generally weak. For example, baseline CD4 cell count and viral load were only weakly associated with one another and neither was strongly associated with type of co-administered drug. Additionally, we undertook sensitivity analyses, which suggested that adjustment for other covariates had minimum effect on estimated associations. Lastly, our data tended to be consistent with previous studies—eg, our findings of higher resistance in subtype C patients are consistent with in-vitro data suggesting subtype C viruses are more susceptible to developing the K65R mutation.32

adjustment for other covariates had minimum effect on estimated associations. Lastly, our data tended to be consistent with previous studies—eg, our findings of higher resistance in subtype C patients are consistent with in-vitro data suggesting subtype C viruses are more susceptible to developing the K65R mutation.32 Fourth, despite our analysis being the largest drug resistance study ever undertaken after failure of first-line tenofovir-containing cART, patient numbers were somewhat limited by the slow uptake of tenofovir-based regimens in west and central Africa, eastern Europe, and Asia (in particular China and Russia), and information about baseline viral load in these settings was uncommon. As a result, European countries, Thailand, and South Africa contributed substantially to the analysis. In summary, extensive drug resistance emerges in a high proportion of patients after virological failure on a tenofovir-containing first-line regimen across low-income and middle-income regions. Optimisation of treatment programmes and effective surveillance for transmission of drug resistance is therefore crucial. Correspondence to: Dr Ravindra K Gupta, UCL, Department of Infection, London WC1E 6BT, UK ravindra.gupta@ucl.ac.uk or Prof Robert W Shafer, Department of Medicine, Stanford University, Stanford, CA 94305, USA rshafer@stanford.edur Supplementary Material Supplementary appendix

In summary, extensive drug resistance emerges in a high proportion of patients after virological failure on a tenofovir-containing first-line regimen across low-income and middle-income regions. Optimisation of treatment programmes and effective surveillance for transmission of drug resistance is therefore crucial. Correspondence to: Dr Ravindra K Gupta, UCL, Department of Infection, London WC1E 6BT, UK ravindra.gupta@ucl.ac.uk or Prof Robert W Shafer, Department of Medicine, Stanford University, Stanford, CA 94305, USA rshafer@stanford.edur Supplementary Material Supplementary appendix Acknowledgments We thank the following groups: ACTG 5208 study team; the Lazio and Emilia Romagna Cohorts, Italy; Uganda Virus Research Institute/Ministry of Health (UVRI/MoH) Uganda surveillance study team; the Uganda HIV Drug Resistance Working group; participants and study teams from HIV treatment centres at Masaka and Mbale regional referral hospitals and Nsambya Home-Care; The ClinSurv Study Group; EuResist Network; Swiss HIV Cohort Study; the Athena cohort, Netherlands; CoRIS, Spain; Honduras and Nicaragua cohorts; RFVF, South Africa; Sinikithemba Clinic at McCord Hospital in Durban, South Africa; Tanzanian, Kenyan, and Ugandan Ministries of Health; Harvard/AIDS Prevention Initiative in Nigeria (APIN) prevention, treatment, and care programme; Stichting HIV Monitoring; Tanzanian, Nigeria, and Kenyan Ministries of Health. Infectious Disease Institute, Uganda and Tropical Disease Research Centre, Zambia; PEPFAR; The Cross Sectional Survey of Acquired Drug Resistance Study at Sentinel Sites Study Team; Kenya National HIVDR working group; CDC-Kenya; CDC-Tanzania; CDC-Atlanta; and Andrew Hill for helpful discussions.

n Ministries of Health. Infectious Disease Institute, Uganda and Tropical Disease Research Centre, Zambia; PEPFAR; The Cross Sectional Survey of Acquired Drug Resistance Study at Sentinel Sites Study Team; Kenya National HIVDR working group; CDC-Kenya; CDC-Tanzania; CDC-Atlanta; and Andrew Hill for helpful discussions. Contributors RKG and RWS conceived the study; JG, RKG, and RWS designed the study; MT, SYR, RLH, VCM, LD, IM, KB, NN, KT, TFRdeW, MA, FG, SM, JNT, HFG, CH, PK, NK, BK, OM, CC, ET, CR, LG, EKH, HS, DDC, AA, AM, AL, CM, NG, CVV, AB, AA, AS, UN, WJF, CFP, SA, MMS, CY, JLB, JJM, GH, LM, DS, CW, JA, WK, AT, TEH, NC, RC, TdeO, DP, CS, DD, PK, ER, RK, RKG, RWS, JG, SAR, GRT, AMO, SS, KR, and SM generated and analysed data; and JG, RWS, and RKG wrote the first draft.

NT, HFG, CH, PK, NK, BK, OM, CC, ET, CR, LG, EKH, HS, DDC, AA, AM, AL, CM, NG, CVV, AB, AA, AS, UN, WJF, CFP, SA, MMS, CY, JLB, JJM, GH, LM, DS, CW, JA, WK, AT, TEH, NC, RC, TdeO, DP, CS, DD, PK, ER, RK, RKG, RWS, JG, SAR, GRT, AMO, SS, KR, and SM generated and analysed data; and JG, RWS, and RKG wrote the first draft. TenoRes Study Group members John Gregson, Michele Tang*, Nicaise Ndembi*, Raph L Hamers*, Soo-Yon Rhee, Vincent C Marconi, Lameck Diero, Katherine Brooks, Kristof Theys, Tobias F Rinke de Wit, Monica Arruda, Frederico Garcia, Susana Monge, Huldrych F Günthard, Christopher J Hoffmann, Phyllis J Kanki, Nagalingeshwaran Kumarasamy, Bernard Kerschberger, Orna Mor, Charlotte Charpentier, Eva Todesco, Casper Rokx, Luuk Gras, Elias K Halvas, Henry Sunpath, Domenico Di Carlo, Antonio Antinori, Massimo Andreoni, Alessandra Latini, Cristina Mussini, Avelin Aghokeng, Anders Sonnerborg, Ujjwal Neogi, William J Fessel, Simon Agolory, Chunfu Yang, Jose L Blanco, James M Juma, Erasmus Smit, Daniel Schmidt, Christine Watera, Juliet Asio, Wilford Kirungi, Anna Tostevin, Tal El-Hay, Nathan Clumeck, Dominique Goedhals, Cloete van Vuuren, Philip Armand Bester, Caroline Sabin, Irene Mukui, Maria M Santoro, Carlo F Perno, Gillian Hunt, Lynn Morris, Ricardo Camacho, Tulio de Oliveira, Deenan Pillay, Eugene Schulter, Akio Murakami-Ogasawara, Gustavo Reyes-Terán, Karla Romero, Santiago Avila-Rios, Sunee Sirivichayakul, Kiat Ruxrungtham, Suwanna Mekprasan, David Dunn, Pontiano Kaleebu, Elliot Raizes, Rami Kantor, Robert W Shafer**, Ravindra K Gupta**, Department of Statistics, London School of Hygiene & Tropical Medicine (J Gregson PhD); Department of Medicine, Stanford University, Stanford, CA, USA (M Tang MD, S-Y Rhee PhD, R W Shafer MD); Institute of Human Virology Nigeria, Abuja, Federal Capital Territory, Nigeria (N Ndembi PhD); Amsterdam Institute for Global Health and Development, Department of Global Health and Department of Internal Medicine, Academic Medical Center of the University of Amsterdam, The Netherlands (R L Hamers MD, T F Rinke de Wit MD); Department of Global Health, Emory University Rollins School of Public Health, Atlanta, GA, USA (V C Marconi MD); Division of Infectious Diseases, Emory University School of Medicine, Atlanta, GA, USA (V C Marconi); Moi University and the Academic Model Providing Access to Healthcare, Eldoret, Kenya (L Diero MD); Division of Infectious Diseases Brown University Alpert Medical School, USA (K Brooks BA); KU Leuven–University of Leuven

vision of Infectious Diseases, Emory University School of Medicine, Atlanta, GA, USA (V C Marconi); Moi University and the Academic Model Providing Access to Healthcare, Eldoret, Kenya (L Diero MD); Division of Infectious Diseases Brown University Alpert Medical School, USA (K Brooks BA); KU Leuven–University of Leuven , Department Microbiology and Immunology, Rega Institute for Medical Research, B-3000 Leuven, Belgium (K Theys PhD, R Camacho, R Kantor MD); Laboratório de Virologia Molecular-LVM Instituto de Biologia - Universidade Federal do Rio de Janeiro (M Arruda PhD); Complejo Hospitalario Universitario de Granada, Granada, Spain (F Garcia); Universidad de Alcalá, Spain; CIBERESP, Spain (S Monge); Division of Infectious Diseases and Hospital Epidemiology, University of Zurich, Zurich, Switzerland (H F Günthard MD); Institute of Medical Virology, University of Zurich, Zurich Switzerland (H F Günthard); Johns Hopkins University, Baltimore, USA (C J Hoffmann MD); Aurum Institute, Johannesburg, South Africa (C J Hoffmann MD); Department of Immunology and Infectious Disease, Harvard T H Chan School of Public Health, Boston, MA, USA (P J Kanki MD); YRGCARE Medical Centre, VHS, Chennai, India (N Kumarasamy MD); Medecins Sans Frontieres (Operational Centre Geneva), Mbabane, Swaziland (B Kerschberger); Central Virology Laboratory, Public Health Services, Israel Ministry of Health (O Mor PhD); IAME, UMR 1137, Univ Paris Diderot, Sorbonne Paris Cité, Paris, France (C Charpentier PhD); IAME, UMR 1137, INSERM, Paris, France (C Charpentier); AP-HP, Hôpital Bichat-Claude Bernard, Laboratoire de Virologie, F-75018 Paris, France (C Charpentier PhD); Hôpital Pitié-Salpêtrière, Laboratoire de Virologie, Paris (E Todesco PhD); Department of Internal Medicine—Infectious Diseases, Erasmus University Medical Center, Rotterdam, Netherlands (C Rokx MD); Stichting HIV monitoring, Amsterdam, Netherlands (L Gras MSc); University of Pittsburgh, Pittsburgh, PA, USA (E K Halvas); Ethekwini District Health Office, KwaZulu-Natal, South Africa (H Sunpath MD); University of Rome Tor Vergata, Department of Experimental Medicine and Surgery, Rome, Italy (D Di Carlo MRes, M M Santoro PhD); INMI L.

itoring, Amsterdam, Netherlands (L Gras MSc); University of Pittsburgh, Pittsburgh, PA, USA (E K Halvas); Ethekwini District Health Office, KwaZulu-Natal, South Africa (H Sunpath MD); University of Rome Tor Vergata, Department of Experimental Medicine and Surgery, Rome, Italy (D Di Carlo MRes, M M Santoro PhD); INMI L. Spallanzani, Infectious Disease Unit, Rome, Italy (A Antinori); University Hospital Tor Vergata, Clinical Infectious Diseases, Rome, Italy (M Andreoni); San Gallicano Dermatological Institute, HIV/AIDS Unit, Rome, Italy (A Latini); Azienda Ospedaliero-Universitaria Policlinico, Clinic of Infectious Disease, Modena, Italy (C Mussini MD); Virology Laboratory CREMER-IMPM, Yaoundé, Cameroon (A Aghokeng PhD); Division of Clinical Microbiology and Unit of Infectious Diseases, Karolinska Institutet and Karolinska University Hospital, Stockholm, Sweden (A Sonnerborg MD, U Neogi PhD); Kaiser Permanente Medical Care Program - Northern California, San Francisco, CA, USA (W J Fessel); Division of Global HIV/AIDS, Center for Global Health, Centers for Disease Control and Prevention, Atlanta, GA, USA (S Agolory MD, E Raizes MD); International Laboratory Branch, Division of Global HIV/AIDS, Center for Global Health, Centers for Disease Control and Prevention, Atlanta, GA, USA (C Yang PhD); Clinic Universitari–Institut d'Investigacions Biomèdiques August Pi i Sunyer, University of Barcelona, Barcelona, Spain (J L Blanco); Ministry of Health and Social Welfare, Tanzania (J M Juma); Public Health Laboratory, Birmingham, Public Health England (E Smit MD); Department of Infectious Disease Epidemiology, HIV/AIDS, STI and Blood Born Infections, Robert Koch-Institute, Berlin, Germany (D Schmidt PhD); Uganda Research Unit on AIDS, Entebbe, Uganda (C Watera MSc, J Asio MSc, P Kaleebu PhD); Ministry of Health, Uganda (W Kirungi MD); MRC Clinical Trials Unit at UCL, London, UK (A Tostevin PhD, D Dunn PhD); IBM Haifa Research Lab, Israel (T El-Hay PhD); Saint-Pierre University Hospital, Université Libre de Bruxelles, Belgium (N Clumeck MD); Department of Medical Microbiology and Virology, National Health Laboratory Service/University of the Free State, Bloemfontein, South Africa (D Goedhals PhD, C van Vuuren MD); Infection and Population Health, UCL, London, UK (C Sabin PhD); National AIDS & STI Control Programme, Ministry of Health, Nairobi, Kenya (I Mukui); INMI L Spallanzani, Antiretroviral Drugs Monitoring Unit, Rome, Italy (C F Perno MD); National Institute for Communicable Diseases,

ica (D Goedhals PhD, C van Vuuren MD); Infection and Population Health, UCL, London, UK (C Sabin PhD); National AIDS & STI Control Programme, Ministry of Health, Nairobi, Kenya (I Mukui); INMI L Spallanzani, Antiretroviral Drugs Monitoring Unit, Rome, Italy (C F Perno MD); National Institute for Communicable Diseases, Sandringham, South Africa (G Hunt PhD, L Morris PhD); Wellcome Trust Africa Centre for Health and Population Studies, South Africa (T de Oliveira PhD, D Pillay FRCP); College of Health Sciences, University of KwaZulu-Natal, Durban, South Africa (T de Oliveira); Department of Infection, UCL, London, UK (D Pillay, R K Gupta MRCP); Institute of Virology, University of Cologne, Cologne, Germany (E Schulter DIP); Centre for Research in Infectious Diseases, National Institute of Respiratory Diseases, Mexico City, Mexico (A Murakami-Ogasawara MD); Department of Medicine, Chulalongkorn University, Bangkok, Thailand (S Sirivichayakul PhD, K Ruxrungtham PhD, S Mekprasan BSc); and MRC/UVRI Uganda Research Unit on AIDS, Entebbe, Uganda (P Kaleebu) *These authors contributed equally. **Equal second author contributions.

Sandringham, South Africa (G Hunt PhD, L Morris PhD); Wellcome Trust Africa Centre for Health and Population Studies, South Africa (T de Oliveira PhD, D Pillay FRCP); College of Health Sciences, University of KwaZulu-Natal, Durban, South Africa (T de Oliveira); Department of Infection, UCL, London, UK (D Pillay, R K Gupta MRCP); Institute of Virology, University of Cologne, Cologne, Germany (E Schulter DIP); Centre for Research in Infectious Diseases, National Institute of Respiratory Diseases, Mexico City, Mexico (A Murakami-Ogasawara MD); Department of Medicine, Chulalongkorn University, Bangkok, Thailand (S Sirivichayakul PhD, K Ruxrungtham PhD, S Mekprasan BSc); and MRC/UVRI Uganda Research Unit on AIDS, Entebbe, Uganda (P Kaleebu) *These authors contributed equally. **Equal second author contributions. Declaration of interests CR has received personal fees from ViiV Healthcare, personal fees from MSD/Gilead outside of the submitted work. RG has received personal fees from BMS and Janssen-Cilag outside of the submitted work. HG reports personal fees from BMS, Gilead Sciences, Janssen-Cilag, ViiV Healthcare, Abbvie, and Merck outside the submitted work. AA reports grants and personal fees from BMS, Gilead Sciences, Janssen-Cilag, ViiV Healthcare, Abbvie, and Merck outside the submitted work. CS has received personal fees from BMS, Gilead, ViiV outside of the submitted work. RC reports personal fees from ViiV Healthcare and personal fees and grants from Abbvie outside the submitted work. CC reports personal fees from outside the submitted work. RWS reports grants from Gilead Sciences, Merck, Celera, Siemens Health care and Roche molecular diagnostics outside the submitted work. FG reports personal fees from MSD, Gilead Sciences, Janssen-Cilag, ViiV Healthcare, and Abbvie outside the submitted work. CvV reports personal fees from Pfizer and Mylan. AS reports fees from MSD, Gilead Sciences, Janssen-Cilag, ViiV Healthcare, and Abbvie outside the submitted work.

ostics outside the submitted work. FG reports personal fees from MSD, Gilead Sciences, Janssen-Cilag, ViiV Healthcare, and Abbvie outside the submitted work. CvV reports personal fees from Pfizer and Mylan. AS reports fees from MSD, Gilead Sciences, Janssen-Cilag, ViiV Healthcare, and Abbvie outside the submitted work. Figure 1 (A) Countries contributing data to resistance analysis and HIV-1 subtype distribution, (B) prevalence of drug resistance by mutation and by region NNRTI=non-nucleotide reverse-transcriptase inhibitor. TDF=tenofovir disoproxil fumarate. *24% (n=462) of participants had tenofovir resistance when genotypes from viral load >1000 copies HIV-1 RNA per mL were considered. Figure 2 Pooled odds ratios for tenofovir resistance after viral failure for baseline CD4 cell count <100 vs ≥100 × 106 cells per μL TDF+ denotes presence of tenofovir resistance. TDF=tenofovir disoproxil fumarate. Figure 3 Odds ratios for NNRTI resistance for (A) baseline CD4 cell count <100 vs ≥100 cells per μL, (B) viral load ≥100 000 vs <100 000 copies HIV-1 RNA per mL NNRTI=non-nucleotide reverse-transcriptase inhibitor. Figure 4 Boxplot of log viral load by presence (TDF-positive) or absence (TDF-negative) of tenofovir resistance at viral failure in studies with at least ten patients with TDF resistance and a viral load measurement at treatment failure

Figure 3 Odds ratios for NNRTI resistance for (A) baseline CD4 cell count <100 vs ≥100 cells per μL, (B) viral load ≥100 000 vs <100 000 copies HIV-1 RNA per mL NNRTI=non-nucleotide reverse-transcriptase inhibitor. Figure 4 Boxplot of log viral load by presence (TDF-positive) or absence (TDF-negative) of tenofovir resistance at viral failure in studies with at least ten patients with TDF resistance and a viral load measurement at treatment failure We restricted to studies with at least ten TDF-resistant mutations to help with graphical clarity, although the pattern of similar distributions of failure viral load in the presence or absence of TDF resistance was true for all studies. TDF=tenofovir disoproxil fumarate. Blue dots represent outliers. Table 1 Characteristics of resistance studies included in analysis

We restricted to studies with at least ten TDF-resistant mutations to help with graphical clarity, although the pattern of similar distributions of failure viral load in the presence or absence of TDF resistance was true for all studies. TDF=tenofovir disoproxil fumarate. Blue dots represent outliers. Table 1 Characteristics of resistance studies included in analysis Countries Studies* Mean study size Median year of initiation of cART (range) Studies in which frequent viral load monitoring was done (>2 viral loads per year) Studies in which genotypic resistance testing done at viral load <1000 copies per mL Studies in which baseline resistance testing was done Rural clinics Eastern Africa (n=143) 3 7 24 2011 (2005–12) 0 1 (14%) 1 (14%) 2 (29%) Asia (n=356) 4 5 71 2010 (2005–13) 2 (40%) 2 (40%) 2 (40%) 1 (20%) Eastern Africa (n=143) 3 7 24 2011 (2005–12) 0 1 (14%) 1 (14%) 2 (29%) Latin America (n=68) 5 6 11 2008 (2000–15) 4 (67%) 2 (67%) 2 (67%) 0 (100%) North America (n=94) 2 3 47 2008 (2000–14) 3 (100%) 3 (100%) 3 (100%) 3 (100%) Southern Africa (n=404) 6 15 45 2010 (2005–12) 4 (27%) 4 (27%) 4 (27%) 5 (33%) West and central Africa (n=107) 5 10 12 2008 (2005–13) 1 (10%) 0 0 0 Western Europe (n=754) 11 20 69 2008 (1998–2013) 20 (100%) 20 (100%) 20 (100%) 0 All (n=1926) 36 66 29 2008 (1998–2015) 34 (52%) 32 (49%) 32 (49%) 11 (17%) Data are n, range, or n (%). cART=combination antiretroviral therapy. * Multinational studies were treated as separate studies within each country. Table 2 Participant characteristics and details of antiretroviral therapy

Countries Studies* Mean study size Median year of initiation of cART (range) Studies in which frequent viral load monitoring was done (>2 viral loads per year) Studies in which genotypic resistance testing done at viral load <1000 copies per mL Studies in which baseline resistance testing was done Rural clinics Eastern Africa (n=143) 3 7 24 2011 (2005–12) 0 1 (14%) 1 (14%) 2 (29%) Asia (n=356) 4 5 71 2010 (2005–13) 2 (40%) 2 (40%) 2 (40%) 1 (20%) Eastern Africa (n=143) 3 7 24 2011 (2005–12) 0 1 (14%) 1 (14%) 2 (29%) Latin America (n=68) 5 6 11 2008 (2000–15) 4 (67%) 2 (67%) 2 (67%) 0 (100%) North America (n=94) 2 3 47 2008 (2000–14) 3 (100%) 3 (100%) 3 (100%) 3 (100%) Southern Africa (n=404) 6 15 45 2010 (2005–12) 4 (27%) 4 (27%) 4 (27%) 5 (33%) West and central Africa (n=107) 5 10 12 2008 (2005–13) 1 (10%) 0 0 0 Western Europe (n=754) 11 20 69 2008 (1998–2013) 20 (100%) 20 (100%) 20 (100%) 0 All (n=1926) 36 66 29 2008 (1998–2015) 34 (52%) 32 (49%) 32 (49%) 11 (17%) Data are n, range, or n (%). cART=combination antiretroviral therapy. * Multinational studies were treated as separate studies within each country. Table 2 Participant characteristics and details of antiretroviral therapy Men Age (years) Efavirenz Emtricitabine Baseline CD4 cell count (× 106cells per μL) Pre-treatment log10baseline viral load Number of months on TDF Asia (n=356) 229 (67%) 35 (30–39) 300 (84%) 73 (21%) 100 (45–229) 5·00 (4·55–5·68) 14 (9–21) Eastern Africa (n=143) 57 (40%) 36 (29–44) 56 (39%) 53 (37%) 104 (42–210) 5·58 (5·30–5·83) 14 (12–26) Latin America (n=68) 19 (70%) 34 (26–44) 65 (96%) 44 (65%) 44 (14–86) 5·47 (5·00–5·93) 26 (11–57) North America (n=94) 78 (84%) 41 (35–48) 81 (87%) 61 (66%) 144 (25–303) 5·00 (4·59–5·53) 11 (6–24) Southern Africa (n=404) 147 (36%) 34 (28–40) 290 (72%) 89 (22%) 98 (40–169) 4·80 (3·81–5·47) 18 (12–28) West and central Africa (n=107) 45 (42%) 36 (30–42) 39 (36%) 79 (74%) 89 (37–166) 5·32 (4·92–5·81) 13 (11–18) Western Europe (n=754) 571 (76%) 38 (32–44) 653 (87%) 633 (84%) 199 (91–300) 5·00 (4·28–5·46) 12 (7–26) All (n=1926) 1146 (62%) 37 (30–44) 1485 (77%) 1032 (54%) 139 (53–250) 5·06 (4·45–5·56) 14 (9–27) Data are n (%) or median (IQR). TDF=tenofovir disoproxil fumarate.

Introduction The associations between first-trimester malaria, treatment, and miscarriage remain poorly documented because these events often occur before women present to antenatal care.1, 2 A single first-trimester malaria episode is associated with miscarriage,3 and women with first-trimester malaria who are not adequately treated are at high risk of placental malaria.4 Because women are usually not protected by preventive interventions until the second trimester,3 early diagnosis and effective treatment of first-trimester malaria are essential to limit the deleterious effects of malaria.5

h first-trimester malaria who are not adequately treated are at high risk of placental malaria.4 Because women are usually not protected by preventive interventions until the second trimester,3 early diagnosis and effective treatment of first-trimester malaria are essential to limit the deleterious effects of malaria.5 Artemisinin derivatives (hereafter referred to as artemisinins) are the most effective antimalarials available. Artemisinin-based combination therapies are recommended by the WHO for first-line treatment of falciparum malaria, except during the first trimester of pregnancy.6 Animal studies have raised concerns about the safety of artemisinins in the first trimester, but data for human beings are scarce. In animals, artemisinins are embryotoxic and teratogenic because they deplete embryonic erythroblasts, causing miscarriage and congenital malformations (mainly cardiovascular and skeletal).7 If artemisinins are also embryotoxic or teratogenic in human beings, the embryo-sensitive period is predicted to be between 6 weeks' and 13 weeks' gestation when erythroblasts are the primary form of circulating red blood cells.8 Because of these safety concerns, quinine is still recommended for uncomplicated first-trimester falciparum malaria rather than artemisinins, despite being an inferior treatment.6 Available data for first-trimester artemisinin safety comes from observational studies of inadvertent treatments, which are common but rarely documented.2 No specific adverse effects have been noted in human beings in 935 documented first-trimester artemisinin treatments (appendix p 1),9, 10, 11, 12, 13, 14, 15, 16, 17, 18 which although reassuring, has not been sufficient to change treatment recommendations.6

dvertent treatments, which are common but rarely documented.2 No specific adverse effects have been noted in human beings in 935 documented first-trimester artemisinin treatments (appendix p 1),9, 10, 11, 12, 13, 14, 15, 16, 17, 18 which although reassuring, has not been sufficient to change treatment recommendations.6 The Shoklo Malaria Research Unit (SMRU) screens pregnant women frequently for malaria because there are no effective preventive interventions (appendix p 3).19 Since 1986, prospective data have been collected on confirmed malaria infections, antimalarial treatment, and pregnancy outcomes of women attending SMRU antenatal clinics, providing an important source of observational evidence on first-trimester artemisinin safety. In this setting, a single first-trimester malaria episode (falciparum or vivax) increased the odds of miscarriage, but first-trimester artemisinin treatment was not associated with miscarriage.3 However, for analytic clarity in this earlier study women with recurrent infections were excluded, which reduced the number of artemisinin treatments to 44 and overestimated the effect of malaria, because a recurrent infection in pregnancy depends on the fetus surviving the initial infection. Here, we extend this seminal study by including women with recurrent malaria, which might be either novel, recrudescent, or a relapse in the case of vivax malaria, and added 3 further years of data. Assessment of the safety of artemisinins requires weighing the risks of malaria and its treatment. Therefore, we sought to assess the effect of both first-trimester malaria and artemisinin treatment on miscarriage and major congenital malformations.

lapse in the case of vivax malaria, and added 3 further years of data. Assessment of the safety of artemisinins requires weighing the risks of malaria and its treatment. Therefore, we sought to assess the effect of both first-trimester malaria and artemisinin treatment on miscarriage and major congenital malformations. Research in context Evidence before this study

lapse in the case of vivax malaria, and added 3 further years of data. Assessment of the safety of artemisinins requires weighing the risks of malaria and its treatment. Therefore, we sought to assess the effect of both first-trimester malaria and artemisinin treatment on miscarriage and major congenital malformations. Research in context Evidence before this study We searched Scopus and PubMed for articles published up to Oct 5, 2015, in any language, that addressed the association between first-trimester artemisinin treatment and miscarriage using the search terms: “malaria or plasmodium”, “pregnan*”, “*artemisinin* OR ACT* OR artesunate OR artemether OR Coartem”, “first-trimester OR ‘first trimester’ OR ‘early pregnancy’”, and “miscarriage* OR abortion”. One study reported on the association between a single first-trimester malaria episode and miscarriage. No randomised controlled trials of first-trimester artemisinin treatment were identified. No studies have reported on the association between recurrent first-trimester malaria and miscarriage. Ten observational studies of first-trimester artemisinin treatment were identified totalling 935 documented treatments, and a systematic review published in 2007. These studies showed no evidence of an increased risk of miscarriage or major congenital malformations associated with first-trimester artemisinin treatment. Importantly, only one published study examining the association between first-line artemisinin treatment and miscarriage accounted for left truncation, which is necessary when women present at varying gestations due to the declining risk of miscarriage as a pregnancy progresses, and few were able to account for confounding by indication and disease severity.

amining the association between first-line artemisinin treatment and miscarriage accounted for left truncation, which is necessary when women present at varying gestations due to the declining risk of miscarriage as a pregnancy progresses, and few were able to account for confounding by indication and disease severity. Added value of this study Assessment of the safety of artemisinin derivatives requires weighing the risks of falciparum malaria against those of its treatment. We noted that first-trimester falciparum malaria increases the risk of miscarriage, especially after recurrence. However, there was no evidence that first-line treatment with an artemisinin derivative in the first trimester was associated with an increased risk of miscarriage or congenital malformations compared with first-line quinine, which is currently recommended by the WHO. We compared first-line treatment with an artemisinin derivative with first-line quinine in women with first-trimester falciparum malaria in an area of low seasonal transmission, and accounted for confounding by indication and disease severity, thereby separating the effects of infection from the effects of treatment. To the best of our knowledge, this study is the first to estimate the association between recurrent first-trimester malaria and miscarriage, and contributes a further 183 documented first-trimester artemisinin treatments. Left truncation, which adjusts for the temporally changing risks of miscarriage and varying gestation at presentation, was also accounted for and is essential to avoid significant bias.

en recurrent first-trimester malaria and miscarriage, and contributes a further 183 documented first-trimester artemisinin treatments. Left truncation, which adjusts for the temporally changing risks of miscarriage and varying gestation at presentation, was also accounted for and is essential to avoid significant bias. Implications of all the available evidence Effective treatment of first-trimester falciparum malaria is imperative. Our results add to a growing body of observational evidence that artemisinins, the most effective antimalarials available, are safe in the first trimester of pregnancy. Methods Study design and participants In this observational study, we assessed data from antenatal clinics on the Thai–Myanmar border between Jan 1, 1994, and Dec 31, 2013. We included women who presented to antenatal clinics during their first trimester with a viable fetus. Women were screened for malaria, and data for malaria, antimalarial treatment, and birth outcomes were collected. The Oxford Tropical Research Ethics Committee granted ethical approval for audits of SMRU clinical records (OXTREC 28-09), and the Tak Province Community Ethics Advisory Board provided local permission (T-CAB-4/1/2015). Data for first-trimester malaria from some of the records included in this analysis have been published previously.3, 19, 20, 21, 22

ics Committee granted ethical approval for audits of SMRU clinical records (OXTREC 28-09), and the Tak Province Community Ethics Advisory Board provided local permission (T-CAB-4/1/2015). Data for first-trimester malaria from some of the records included in this analysis have been published previously.3, 19, 20, 21, 22 Procedures At SMRU antenatal clinics, women are encouraged to present early and return weekly throughout their pregnancy for malaria screening, consisting of a finger-prick blood sample that is examined by trained microscopists using Giemsa stained thick and thin blood films (appendix p 3).5 Women are also encouraged to present if they feel unwell, and to deliver at SMRU clinics. The first consultation involves taking obstetric and medical histories, a detailed clinical examination, and gestational age estimation.23 With each positive screen, information about species, parasitaemia, symptoms, and treatment are recorded. Women are also asked about recent antimalarial treatments at outpatient clinics, and these treatments (usually mefloquine–artesunate [MAS] for P falciparum) are recorded retrospectively. Presumptive malaria treatment is not used, and pregnancy termination is unavailable.

asitaemia, symptoms, and treatment are recorded. Women are also asked about recent antimalarial treatments at outpatient clinics, and these treatments (usually mefloquine–artesunate [MAS] for P falciparum) are recorded retrospectively. Presumptive malaria treatment is not used, and pregnancy termination is unavailable. First-trimester non-malaria febrile morbidity was defined as fever (temperature ≥37·5oC) not associated with malaria. Malaria was defined as the presence of asexual stages of plasmodia parasites in the peripheral blood, counted per 500 white blood cells or 1000 red blood cells. Hyperparasitaemia was defined as 4% parasitaemia or greater, and severe malaria was defined according to signs of vital organ dysfunction. Symptomatic malaria was defined as patent parasitaemia and a history of fever (past 48 h) or temperature 37·5oC or greater. Vivax malaria was treated with oral chloroquine. Falciparum malaria was treated with oral quinine in the first trimester, or an artemisinin-based treatment in the second and third trimester (either artesunate, artemether–lumefantrine, dihydroartemisinin–piperaquine, or mefloquine–artesunate). Mefloquine monotherapy was given for falciparum malaria until 1996. Clindamycin was added to quinine and artesunate 7-day treatments in 2007 to augment efficacy. According to WHO recommendations, artemisinins were given in the first trimester for quinine failures, hyperparasitaemia, severe malaria, or if the fetus was no longer viable.6 Details on treatment regimens and drug manufacturers are given in the appendix (p 4).

artesunate 7-day treatments in 2007 to augment efficacy. According to WHO recommendations, artemisinins were given in the first trimester for quinine failures, hyperparasitaemia, severe malaria, or if the fetus was no longer viable.6 Details on treatment regimens and drug manufacturers are given in the appendix (p 4). Outcomes Primary exposures were malaria and first-line artemisinin treatment in the first trimester, defined as less than 14 weeks' gestation. The primary outcome was miscarriage, defined as fetal death before 28 weeks' gestation because infant respiratory support is unavailable. The ability to determine gestation and fetal viability at SMRU improved after ultrasound was introduced in 2002 (appendix p 3).23, 24 The date of miscarriage was recorded consistently as the date of expulsion of the uterine contents, either spontaneously or through surgical intervention, which can occur some time after intrauterine death. The secondary outcome was major congenital malformations. A surface examination was done on all newborns by trained staff; a physician verified all malformations, except for some early neonatal deaths. Artemisinin-based treatments were first deployed in the general population in 1994. Therefore, we included women who presented to antenatal clinics during their first trimester with a viable fetus between Jan 1, 1994, and Dec 31, 2013.

taff; a physician verified all malformations, except for some early neonatal deaths. Artemisinin-based treatments were first deployed in the general population in 1994. Therefore, we included women who presented to antenatal clinics during their first trimester with a viable fetus between Jan 1, 1994, and Dec 31, 2013. Statistical analysis We used Cox proportional hazards models accounting for left truncation (appendix p 5) and time-varying exposures for all miscarriage analyses, with censoring at the gestation time of miscarriage, gestation time when last seen, or 28 weeks' gestation. To assess the association between malaria and miscarriage, women entered the analysis at the gestation time of their first antenatal visit. Multivariable models adjusted for year of first consultation, gravidity, smoking, and first-trimester non-malaria febrile morbidity. To assess the association between first-line artemisinin treatment and miscarriage (primary analysis), we included women with first-trimester falciparum malaria, and compared first-line quinine treatment (including quinine plus clindamycin) with first-line mefloquine monotherapy, artemisinin treatment (all derivatives) following quinine failure (ie, artemisinin rescue), and first-line artemisinin treatment (all derivatives). Women entered the analysis at the gestation time of their first falciparum malaria episode. Treatments given after determination of fetal non-viability were excluded. Multivariable models adjusted for year of first consultation, disease severity pertaining to the first falciparum malaria episode (asymptomatic, symptomatic, or hyperparasitaemic or severe), and first-trimester non-malaria febrile morbidity. The prevalence of major congenital malformation was described by first-trimester falciparum malaria and first-line treatment. Malformations were grouped by organ system to increase the likelihood of detecting teratogenic signals.25 Data were analysed with Stata version 13 (StataCorp, College Station, TX, USA).

rbidity. The prevalence of major congenital malformation was described by first-trimester falciparum malaria and first-line treatment. Malformations were grouped by organ system to increase the likelihood of detecting teratogenic signals.25 Data were analysed with Stata version 13 (StataCorp, College Station, TX, USA). Role of the funding source The funding sources (The Wellcome Trust and The Bill & Melinda Gates Foundation) had no role in the study design, data collection, data analysis, data interpretation, writing of the report, or the decision to submit for publication. The corresponding author had full access to all data and had the final responsibility in the decision to submit for publication. Results Between Jan 1, 1994 and Dec 31, 2013, 55 636 pregnant women presented to SMRU clinics, of whom 25 485 (46%) presented during their first trimester with a viable fetus (figure 1). Of these, 2257 (10%) of 23 118 miscarried, 2367 (9%) of 25 485 were lost to follow-up before 28 weeks gestation, and 2558 (10%) of 25 485 had first-trimester malaria (figure 1). Women with first-trimester malaria were more likely to miscarry or be lost to follow-up and tended to present for antenatal care earlier, be younger, be primigravid, and smoke compared with women without first-trimester malaria (all p<0·0001; table 1).

n, and 2558 (10%) of 25 485 had first-trimester malaria (figure 1). Women with first-trimester malaria were more likely to miscarry or be lost to follow-up and tended to present for antenatal care earlier, be younger, be primigravid, and smoke compared with women without first-trimester malaria (all p<0·0001; table 1). Of the 2558 women with first-trimester malaria, 1207 (47%) had falciparum malaria, 1532 (60%) had vivax malaria, and 181 (7%) had both vivax and falciparum (either separate or mixed infections). Recurrent first-trimester falciparum malaria occurred in 162 (13%) of 1207 women, and recurrent first-trimester vivax malaria in 139 (9%) of 1532. Most (971 [80%] of 1207) women with first-trimester falciparum malaria were treated initially with quinine and 183 (15%) of 1207 were treated initially with artemisinin (ie, first-line artemisinin treatment; table 1). Of the 971 women who received first-line quinine treatment, 129 (13%) were rescued with artemisinin (usually artesunate monotherapy or artesunate plus clindamycin) following recurrence. Of the 183 first-line artemisinin treatments, 37 (20%) were for hyperparasitaemia (administered orally) or severe disease (administered parenterally). First-line treatment of first-trimester falciparum malaria occurred at a median of 8·2 gestation weeks (IQR 5·3–11·1). Loss to follow-up was similar between antimalarial treatment groups, except women receiving mefloquine–artesunate were less likely to be lost (p=0·0417; appendix p 6). Rates of falciparum malaria during pregnancy and miscarriage and the frequency of first-line quinine and artemisinin treatments in first trimester over time are shown in figure 2.

ar between antimalarial treatment groups, except women receiving mefloquine–artesunate were less likely to be lost (p=0·0417; appendix p 6). Rates of falciparum malaria during pregnancy and miscarriage and the frequency of first-line quinine and artemisinin treatments in first trimester over time are shown in figure 2. Of 1207 women with first-trimester falciparum malaria, 165 (17% of 983 followed until 28 weeks') miscarried and 224 (19%) were lost to follow-up compared with 1963 (9%) of 20 978 and 1949 (9%) of 22 927 in women with no first-trimester malaria, respectively. In multivariable analyses, the hazard of miscarriage increased 1·61-fold (95% CI 1·32–1·97; p<0·0001) with an initial first-trimester falciparum malaria episode, and 3·24-fold (2·24–4·68; p<0·0001) with recurrent first-trimester falciparum malaria (figure 3). This association was stronger in women with symptomatic falciparum malaria than in women with asymptomatic falciparum malaria (figure 2). A single first-trimester hyperparasitaemic or severe falciparum malaria episode increased the hazard of miscarriage 4·21-fold (95% CI 2·43–7·29; p<0·0001; figure 3). An initial first-trimester vivax malaria episode, either asymptomatic or symptomatic, increased the hazard of miscarriage slightly (figure 3). Recurrent symptomatic first-trimester vivax malaria increased the hazard of miscarriage 2·44-fold (95% CI 1·01–5·88; p=0·0473; figure 3).

(95% CI 2·43–7·29; p<0·0001; figure 3). An initial first-trimester vivax malaria episode, either asymptomatic or symptomatic, increased the hazard of miscarriage slightly (figure 3). Recurrent symptomatic first-trimester vivax malaria increased the hazard of miscarriage 2·44-fold (95% CI 1·01–5·88; p=0·0473; figure 3). Of the 1207 women with first-trimester falciparum malaria, 1179 (98%) had a known first-line antimalarial treatment and a viable fetus at the time of treatment (figure 1). Most (842 [71%] of 1179) received first-line quinine (including quinine plus clindamycin), 129 (11%) of 1179 received first-line quinine followed by artemisinin (artemisinin rescue), 25 (2%) of 1179 received first-line mefloquine monotherapy, and 183 (16%) of 1179 received first-line artemisinin. First-line artemisinin treatment was not associated with miscarriage when compared with women who received first-line quinine only (HR 0·78 [95% CI 0·45–1·34]; p=0·3645; figure 4). Five (3%) of 183 women received two first-trimester artemisinin treatments; one miscarried, and four delivered.

ved first-line artemisinin. First-line artemisinin treatment was not associated with miscarriage when compared with women who received first-line quinine only (HR 0·78 [95% CI 0·45–1·34]; p=0·3645; figure 4). Five (3%) of 183 women received two first-trimester artemisinin treatments; one miscarried, and four delivered. Because animal studies suggest a theoretical embryo-sensitive window in human beings of 6–13 weeks' gestation, we also estimated the association between first-line artemisinin treatment and miscarriage before, during, and after this window.9 First-line artemisinin treatments before the embryo-sensitive window were associated with a non-significant decrease in the hazard of miscarriage (HR 0·54 [95% CI 0·25– 1·15]; p=0·1108), whereas treatments during the embryo-sensitive window were not associated with a changed hazard of miscarriage (1·15 [0·46–2·87]; p=0·7602; figure 4).

sinin treatments before the embryo-sensitive window were associated with a non-significant decrease in the hazard of miscarriage (HR 0·54 [95% CI 0·25– 1·15]; p=0·1108), whereas treatments during the embryo-sensitive window were not associated with a changed hazard of miscarriage (1·15 [0·46–2·87]; p=0·7602; figure 4). Of note, a high proportion of women who received first-line mefloquine–artesunate miscarried (15 [21%] of 71), and most miscarriages in the artemisinin treatment group (15 [63%] of 23) were in women who received mefloquine–artesunate specifically. Further, a third of women who received mefloquine–artesunate during the embryo-sensitive window miscarried (eight [33%] of 24). However, this might be explained by the circumstances of this treatment: women received mefloquine–artesunate from outpatient clinics (rather than antenatal clinics) where they presented because of illness before they became aware of their pregnancy. Additionally, mefloquine–artesunate treatments at outpatient clinics were given at earlier gestations than for the other treatments when women are at greater risk of miscarriage (mefloquine–artesunate: 3·8 weeks [IQR 1·9–7·4]; quinine: 10·1 [7·5–11·8]; other artemisinins: 12·4 [9·0– 13·3]; appendix p 7).

eir pregnancy. Additionally, mefloquine–artesunate treatments at outpatient clinics were given at earlier gestations than for the other treatments when women are at greater risk of miscarriage (mefloquine–artesunate: 3·8 weeks [IQR 1·9–7·4]; quinine: 10·1 [7·5–11·8]; other artemisinins: 12·4 [9·0– 13·3]; appendix p 7). 20 628 women presented for antenatal care between 1994 and 2013 and gave birth to a singleton newborn, of whom 175 (1%) of 20 628 (95% CI 0·73–0·98) had newborns with a major congenital malformation (figure 1). The prevalence of congenital malformations in the newborn babies of women with no first-trimester malaria was 0·84% (158/18 803 [95% CI 0·71–0·98]; table 2). The prevalence of congenital malformations was similar in the newborns of women with first-trimester vivax malaria (0·59% [95% CI 0·22–1·27]). In the newborn babies of women with uncomplicated first-trimester falciparum malaria, malformation prevalence was slightly higher (1·29% [10/773]; [95% CI 0·62–2·37]), but did not differ between the newborn of women who received first-line quinine (1·25% [8/641; [95% CI 0·54–2·44]) and those of women who received first-line artemisinin (2/109 [microphthalmia; imperforate anus]; 1·83% [95% CI 0·22–6·47]; p=0·7551; table 2). Two other newborns of mothers who had hyperparasitaemic or severe first-trimester falciparum malaria had a malformation (syndactyly; cleft lip and palate; 9·09% [2/22]; [95% CI 1·12–29·16]); both were of mothers who received first-line artemisinin, but only eight women received first-line quinine (table 2). No newborns of mothers who received artemisinin in first trimester had the skeletal or cardiovascular malformations as reported in animal studies.

eft lip and palate; 9·09% [2/22]; [95% CI 1·12–29·16]); both were of mothers who received first-line artemisinin, but only eight women received first-line quinine (table 2). No newborns of mothers who received artemisinin in first trimester had the skeletal or cardiovascular malformations as reported in animal studies. Discussion First-trimester falciparum malaria increases the risk of miscarriage, especially after recurrence, but this large prospective observational study found no evidence that first-line treatment with an artemisinin derivative was associated with an increased risk of miscarriage or congenital malformations. Assessment of the safety of artemisinin treatment during pregnancy requires weighing the risks of falciparum malaria against those of its treatment. This is the first study to estimate the effects of initial and recurrent first-trimester malaria, its symptomatology, and its treatment on miscarriage.

ital malformations. Assessment of the safety of artemisinin treatment during pregnancy requires weighing the risks of falciparum malaria against those of its treatment. This is the first study to estimate the effects of initial and recurrent first-trimester malaria, its symptomatology, and its treatment on miscarriage. Legitimate ethical concerns regarding randomised-controlled trials of first-trimester artemisinin treatment have meant that only observational studies have been done to date, and these have not adjusted for confounding by indication and disease severity in assessing risks and benefits.26 A major strength of this study is that it was possible to adjust for these important confounders by comparing with nearly 1000 women who received quinine treatment (appendix p 5). Left truncation, which adjusts for the temporally changing risks of miscarriage and varying gestation at presentation, was also accounted for since this is essential to avoid bias.27 Nevertheless, this study still has limitations common to observational designs. Data were collected over a long period of time, relatively few first-trimester artemisinin treatments were given, toxicities other than miscarriage and major malformations detectable at birth from surface examination were not captured, and all artemisinin derivatives were analysed together. Several associations of considerable magnitude had wide CIs that crossed null, and we cannot rule out potential confounding effects of time and unmeasured variables, or residual confounding by disease severity. Furthermore, women with first-trimester malaria were more likely to be lost to follow-up, raising the possibility of informative right censoring, but this would underestimate the effect of malaria (appendix p 5).

otential confounding effects of time and unmeasured variables, or residual confounding by disease severity. Furthermore, women with first-trimester malaria were more likely to be lost to follow-up, raising the possibility of informative right censoring, but this would underestimate the effect of malaria (appendix p 5). We noted no evidence that first-line treatment with an artemisinin derivative increased the rate of miscarriage compared with first-line treatment with quinine. There was a higher risk of miscarriage in women who received an artemisinin derivative during the putative embryo-sensitive window, but this might be explained at least in part by the administration of mefloquine–artesunate at earlier gestations to symptomatic women in the routine outpatient clinics compared with the active surveillance of antenatal clinics. In rats, embryotoxicity of artesunate was attenuated when co-administered with mefloquine.28 Primates, including human beings, might be less sensitive to the effects of artemisinins because of differences in placentation and the visceral yolk sac, which could result in different levels of embryonic exposure to artemisinins.7, 29 Additionally, a 3-day artemisinin regimen means that the exposure period is relatively short in human beings because organogenesis is 3 days in rats but 3 months in human beings.7, 29 Therefore, artemisinin-induced depletion of embryonic erythroblasts severe enough to cause miscarriage in rats might not translate to human beings, but could still cause congenital malformations.

posure period is relatively short in human beings because organogenesis is 3 days in rats but 3 months in human beings.7, 29 Therefore, artemisinin-induced depletion of embryonic erythroblasts severe enough to cause miscarriage in rats might not translate to human beings, but could still cause congenital malformations. We cannot draw firm conclusions on the possible effects of first-trimester artemisinin treatment on congenital malformations because of relatively small numbers of treatments and cases. Furthermore, the prevalence of major congenital malformations is most likely an underestimation because only those detectable at birth from surface examination and heart auscultation were recorded routinely, and major malformations (particularly cardiovascular) are often not detected or confirmed until later in life. Only four newborns whose mother received first-line artemisinin treatment during first trimester had a major congenital malformation, and the organ systems involved were inconsistent with the types of malformations induced by artemisinins in animal studies.

ular) are often not detected or confirmed until later in life. Only four newborns whose mother received first-line artemisinin treatment during first trimester had a major congenital malformation, and the organ systems involved were inconsistent with the types of malformations induced by artemisinins in animal studies. These results have important implications for malaria treatment and control policies, and future studies of artemisinin safety. Recurrent first-trimester vivax malaria is associated with miscarriage, yet radical cure is not possible during pregnancy with currently available drugs. First-trimester falciparum malaria is strongly associated with miscarriage, especially after recurrence. We noted no evidence of harm associated with first-line artemisinin treatment of first-trimester falciparum malaria. Quinine is comparatively poorly tolerated and associated with a shorter time to recurrence than artemisinin in pregnant women.30 Furthermore, women who received artemisinins following quinine failure were more likely to miscarry than those who received first-line artemisinin treatment. Early and effective antimalarial treatment is imperative, especially because current preventive measures do not adequately cover early pregnancy.3 Artemisinins are the most effective antimalarials available and have been recommended as first-line treatment in the general population by the WHO since 2006. Yet, artemisinin safety in first trimester is still a concern. This study contributes a further 183 well-documented first-trimester artemisinin treatments, and adds to a growing body of observational evidence supporting the use of artemisinins in the first trimester of pregnancy.3, 11, 12, 14, 15, 16, 17, 18, 22, 31 In view of the wide availability of artemisinin-based combination therapies, their excellent tolerability and efficacy, the likely reduced future availability of quinine, and the rarity of congenital malformations, now might be the time to endorse the use of artemisinin derivatives for the treatment of first-trimester falciparum malaria, accompanied by robust pharmacovigilance.

ation therapies, their excellent tolerability and efficacy, the likely reduced future availability of quinine, and the rarity of congenital malformations, now might be the time to endorse the use of artemisinin derivatives for the treatment of first-trimester falciparum malaria, accompanied by robust pharmacovigilance. This online publication has been corrected. The corrected version first appeared at thelancet.com/infection on April 18, 2016 Supplementary Material Supplementary appendix Contributors KAM, JAS, RM, FJIF, and FN developed the analytical plan. RM, JW, MKP, MP, MJR, and PJ collected the data. KAM analysed the data. KAM, JAS, RM, FJIF, NJW, and FN interpreted the data. KAM drafted the report. All authors read and critically revised the draft report, and approved the final report. All authors agreed to be accountable for all aspects of the work. Declaration of interests We declare no competing interests.

Contributors KAM, JAS, RM, FJIF, and FN developed the analytical plan. RM, JW, MKP, MP, MJR, and PJ collected the data. KAM analysed the data. KAM, JAS, RM, FJIF, NJW, and FN interpreted the data. KAM drafted the report. All authors read and critically revised the draft report, and approved the final report. All authors agreed to be accountable for all aspects of the work. Declaration of interests We declare no competing interests. Acknowledgements The pregnant refugee and migrant women and SMRU staff made this study possible. We thank Paul Agius for statistical advice and Campbell Aitken for technical assistance. The Shoklo Malaria Research Unit is part of the Wellcome Trust Mahidol University Oxford Tropical Medicine Research Programme supported by the Wellcome Trust of Great Britain (Major Overseas Programme–Thailand Unit Core Grant). KAM is supported by an Australian Postgraduate Award funded by the Commonwealth Government of Australia. FJIF is supported by a Future Fellowship funded by the Australian Research Council. KAM and FIJF are supported by an Operational Infrastructure Support grant awarded to the Burnet Institute and funded by the Victorian State Government. The data extraction and analysis was supported by a grant funded by the Bill & Melinda Gates Foundation (ID 46589). Figure 1 Study profile P vivax=Plasmodium vivax. P malariae=Plasmodium malariae. P ovale=Plasmodium ovale. P falciparum=Plasmodium falciparum. Figure 2 Frequency of first-line quinine and artemisinin treatments in first trimester and rates of falciparum malaria during pregnancy and miscarriage over time

Acknowledgements The pregnant refugee and migrant women and SMRU staff made this study possible. We thank Paul Agius for statistical advice and Campbell Aitken for technical assistance. The Shoklo Malaria Research Unit is part of the Wellcome Trust Mahidol University Oxford Tropical Medicine Research Programme supported by the Wellcome Trust of Great Britain (Major Overseas Programme–Thailand Unit Core Grant). KAM is supported by an Australian Postgraduate Award funded by the Commonwealth Government of Australia. FJIF is supported by a Future Fellowship funded by the Australian Research Council. KAM and FIJF are supported by an Operational Infrastructure Support grant awarded to the Burnet Institute and funded by the Victorian State Government. The data extraction and analysis was supported by a grant funded by the Bill & Melinda Gates Foundation (ID 46589). Figure 1 Study profile P vivax=Plasmodium vivax. P malariae=Plasmodium malariae. P ovale=Plasmodium ovale. P falciparum=Plasmodium falciparum. Figure 2 Frequency of first-line quinine and artemisinin treatments in first trimester and rates of falciparum malaria during pregnancy and miscarriage over time The increase in the rate of falciparum malaria in 1998 was due to the establishment of Shoklo Malaria Research Unit antenatal clinics in migrant communities. Figure 3 Association between initial and recurrent first-trimester malaria and miscarriage

Figure 2 Frequency of first-line quinine and artemisinin treatments in first trimester and rates of falciparum malaria during pregnancy and miscarriage over time The increase in the rate of falciparum malaria in 1998 was due to the establishment of Shoklo Malaria Research Unit antenatal clinics in migrant communities. Figure 3 Association between initial and recurrent first-trimester malaria and miscarriage Models include women lost to follow-up before 28 weeks (until gestation time last seen), but percentage calculations for delivered or miscarried do not. Models for falciparum malaria (1–4) include women that might have also had first-trimester vivax, malariae, or ovale malaria. See appendix p 8 for associations in women with only first-trimester falciparum malaria. Models for vivax malaria (5–6) exclude women who also had first-trimester falciparum malaria. Models 2 and 5 exclude women with symptomatic malaria. Models 3 and 6 exclude women with asymptomatic infections. Model 4 excludes women with uncomplicated infections. Models were adjusted for year (by stratification due to non-proportional hazards [p<0·001]), gravidity, current smoking status, and non-malaria febrile morbidity in first trimester. Age and previous miscarriage were omitted from multivariable models due to collinearity with gravidity. Adjusted results for gravidity, current smoking status, and febrile morbidity in first trimester are shown from Model 1. 146 women had recurrent first-trimester falciparum malaria (136 had two and ten had three episodes). 13 women had recurrent (two) asymptomatic first-trimester falciparum malaria. 81 women had recurrent symptomatic first-trimester falciparum malaria (75 had two, and six had three episodes). 17 women had recurrent (two episodes) asymptomatic fi0rst-trimester vivax malaria, and none miscarried. 38 women had recurrent (two) symptomatic first-trimester vivax malaria. See appendix p 8 for a table version of this figure, including univariable associations.

parum malaria (75 had two, and six had three episodes). 17 women had recurrent (two episodes) asymptomatic fi0rst-trimester vivax malaria, and none miscarried. 38 women had recurrent (two) symptomatic first-trimester vivax malaria. See appendix p 8 for a table version of this figure, including univariable associations. Figure 4 Association between first-line treatment of first-trimester falciparum malaria and miscarriage (n=1179)

parum malaria (75 had two, and six had three episodes). 17 women had recurrent (two episodes) asymptomatic fi0rst-trimester vivax malaria, and none miscarried. 38 women had recurrent (two) symptomatic first-trimester vivax malaria. See appendix p 8 for a table version of this figure, including univariable associations. Figure 4 Association between first-line treatment of first-trimester falciparum malaria and miscarriage (n=1179) Models were adjusted for severity of the first falciparum malaria episode (asymptomatic, symptomatic, or hyperparasitaemic or severe), non-malaria febrile morbidity in the first trimester, and year of first consultation. See appendix p 9 for a table version of this figure, including univariable associations. *Hazard ratios are shown for treatments occurring before (<6 weeks’ gestation) and during (≥6 and <13 weeks’ gestation) the embryosensitive window. †Artemisinin rescue after quinine failure refers to artemisinin-based treatment in first trimester following failure of first-line treatment with quinine or quinine plus clindamycin. No miscarriages occurred in women who received artemisinin treatment after the embryo-sensitive window (≥13 and <14 weeks’ gestation). We did a subgroup analysis excluding women with asymptomatic malaria (n=919); the association between artemisinin treatment and miscarriage changed by <5% (appendix p 9). We did a subgroup analysis in women attending before 2007 whose gestational age was estimated from ultrasound biometry, because the accuracy of gestational age estimates affects the accuracy of the gestation time of antimalarial treatment, and quinine plus clindamycin succeeded quinine monotherapy in 2007 (n=469); associations were in the same direction but of greater magnitude (appendix p 9).

age was estimated from ultrasound biometry, because the accuracy of gestational age estimates affects the accuracy of the gestation time of antimalarial treatment, and quinine plus clindamycin succeeded quinine monotherapy in 2007 (n=469); associations were in the same direction but of greater magnitude (appendix p 9). Table 1 Cohort demographics

age was estimated from ultrasound biometry, because the accuracy of gestational age estimates affects the accuracy of the gestation time of antimalarial treatment, and quinine plus clindamycin succeeded quinine monotherapy in 2007 (n=469); associations were in the same direction but of greater magnitude (appendix p 9). Table 1 Cohort demographics No first-trimester malaria (n=22 927) First-trimester malaria (n=2558) Miscarried* 1963 (9%) 294 (14%) Lost to follow-up before 28 weeks' gestation† 1949 (9%) 418 (16%) Gestation at first consultation (weeks) 9·0 (7·2–11·3) [0·0–14·0] 8·4 (6·6–10·6) [0·1–14·0] Maternal age (years) 26 (21–31) [13–51] 23 (19–30) [13–46] 13–20 5443 (24%) 909 (36%) 21–25 5959 (26%) 628 (25%) 26–30 5709 (25%) 482 (19%) ≥31 5816 (25%) 539 (21%) Primigravid 5628 (25%) 821 (32%) Current smoker 5362 (26%) 764 (35%) History of miscarriage 6200 (27%) 758 (30%) Haematocrit (first consultation; %) 36 (33–38) [9–52] 34 (31–37) [13–48] Severe anaemia (haematocrit <20%) 9 (0%) 13 (1%) Non-malaria febrile morbidity in first trimester 310 (1%) 38 (1%) Number of antenatal malaria screens 23 (14–28) [1–40] 22 (15–28) [1–38] Estimated gestational age from ultrasonography scans 16 714 (73%) 1648 (64%) Details of initial first-trimester malaria Symptoms Asymptomatic NA 919 (36%) Symptomatic NA 1639 (64%) First-line treatment of first-trimester falciparum malaria Quinine NA 971 (81%) Mefloquine monotherapy NA 25 (2%) Artemisinin derivative NA 183 (15%) Mefloquine–artesunate NA 71 (6%) Artemether–lumefantrine NA 10 (1%) Artesunate plus clindamycin NA 50 (4%) Artesunate monotherapy NA 49 (4%) Dihydroartemisinin–piperaquine NA 3 (0%) Other or unknown NA 8 (1%) Died before administration NA 20 (2%) Data are median (IQR) [range] or n (%). Missing: gravidity, ten; smoking status, 2853; history of miscarriage, nine; haematocrit, 969; and miscarriage, 2367 (ie, lost to follow up before 28 weeks' gestation). Continuous variables were compared between groups using the Student's t test for normal distribution or the Mann-Whitney U test for skewed distribution. Categorical variables were compared with the χ2 test.

of miscarriage, nine; haematocrit, 969; and miscarriage, 2367 (ie, lost to follow up before 28 weeks' gestation). Continuous variables were compared between groups using the Student's t test for normal distribution or the Mann-Whitney U test for skewed distribution. Categorical variables were compared with the χ2 test. * In women followed until 28 weeks; women who miscarried presented for antenatal care 1·4 weeks earlier (p<0·0001), because early attendance increases the chances that a miscarriage will be detected. † Women lost to follow-up were slightly younger (p<0·0001), and were more likely to be primigravid (p<0·0001) and have first-trimester malaria (p<0·0001; appendix p 6). NA=not applicable. Table 2 Major congenital malformations by first-line treatment of first-trimester falciparum malaria by organ system No malaria (n=18 803) Uncomplicated falciparum malaria (n=773)* Hyperparasitaemic or severe falciparum malaria (n=31)* Quinine (n=641) Artemisinin (n=109) Quinine (n=8) Artemisinin (n=22) Multiple 26 (17%) 1 (13%) 0 0 0 Syndromic 4 (3%) 0 0 0 0 CNS 33 (21%) 1 (13%)† 0 0 0 Ears, eyes, face, or neck 24 (15%) 0 1 (50%)‡ 0 0 Circulatory 13 (8%) 1 (13%)§ 0 0 0 Respiratory 2 (1%) 0 0 0 0 Digestive 59 (37%) 2¶ (25%) 1‖ (50%) 0 1** (50%) Genital 19 (12%) 0 0 0 0 Renal 6 (4%) 0 0 0 0 Musculoskeletal 42 (27%) 4†† (50%) 0 0 1‡‡ (50%) Skin 4 (3%) 0 0 0 0 Other 1 (1%) 0 0 0 0 Data are n (%). * 24 women received first-line mefloquine or had unknown first-line treatment (zero malformations). † Anencephaly (treated at 10 weeks' gestation). ‡ Microphthalmia (8 weeks'). § Heart defect (7 weeks').

No malaria (n=18 803) Uncomplicated falciparum malaria (n=773)* Hyperparasitaemic or severe falciparum malaria (n=31)* Quinine (n=641) Artemisinin (n=109) Quinine (n=8) Artemisinin (n=22) Multiple 26 (17%) 1 (13%) 0 0 0 Syndromic 4 (3%) 0 0 0 0 CNS 33 (21%) 1 (13%)† 0 0 0 Ears, eyes, face, or neck 24 (15%) 0 1 (50%)‡ 0 0 Circulatory 13 (8%) 1 (13%)§ 0 0 0 Respiratory 2 (1%) 0 0 0 0 Digestive 59 (37%) 2¶ (25%) 1‖ (50%) 0 1** (50%) Genital 19 (12%) 0 0 0 0 Renal 6 (4%) 0 0 0 0 Musculoskeletal 42 (27%) 4†† (50%) 0 0 1‡‡ (50%) Skin 4 (3%) 0 0 0 0 Other 1 (1%) 0 0 0 0 Data are n (%). * 24 women received first-line mefloquine or had unknown first-line treatment (zero malformations). † Anencephaly (treated at 10 weeks' gestation). ‡ Microphthalmia (8 weeks'). § Heart defect (7 weeks'). ¶ Cleft lip and palate (10 weeks'); cleft lip and palate (9 weeks'). ‖ Imperforated anus (6 weeks'). ** Cleft lip and palate (14 weeks'). †† Polydactyly (12 weeks'); polydactyly (12 weeks'); syndactyly and talipes (11 weeks'); syndactyly (9 weeks'). ‡‡ Syndactyly (13 weeks').

Introduction Tuberculosis is a major global public health challenge.1 In 2014, 6·3 million cases of tuberculosis worldwide were reported to WHO, with India accounting for over a quarter of these cases, the highest of any country.1 Although standardised tuberculosis treatment in India is delivered by the public sector through the Revised National TB Control Programme (RNTCP), early diagnosis and treatment are hampered by the presence of a vast and unregulated private health-care sector.2, 3, 4, 5 Poor diagnostic practices in this sector prolong tuberculosis transmission by delaying diagnosis,3, 5, 6 whereas a general lack of counselling and support of treatment adherence hampers successful, relapse-free cure.4 Moreover, most cases treated in the private sector are never notified to public health authorities.7 Estimating the numbers of patients being treated in the private sector is important for several reasons: it provides information about the performance of a public system in detecting tuberculosis cases, while also helping in planning for government intervention in the private sector.8 Overall, it is crucial to know the scale of the problem: the undetected burden that exists outside the public health system. However, with a lack of systematic data on the private sector, arriving at these estimates has proven difficult.9 Instead, alternative approaches—such as that used by WHO—draw from expert opinion on the proportion of cases that are detected by the public sector.

cted burden that exists outside the public health system. However, with a lack of systematic data on the private sector, arriving at these estimates has proven difficult.9 Instead, alternative approaches—such as that used by WHO—draw from expert opinion on the proportion of cases that are detected by the public sector. In this work, we present an alternative approach. We build on earlier, innovative work that addressed the private market for tuberculosis drugs using comprehensive data on the sales of these drugs in the private sector.10 In the present study, using corresponding data for 2013 and 2014, we explored systematically the implications for tuberculosis burden (numbers of patients) being managed by the private sector in India, and compared this burden directly against that managed by the public sector. Methods Overview We drew from a large, nationally representative dataset for private sector drug sales across the country, collected by the organisation IMS Health. We limited the analysis to 189 drugs containing rifampicin, which have fewer non-tuberculosis indications than, for example, fluoroquinolones. These 189 products capture all rifampicin-containing drugs being sold in India between 2013 and 2014. All products were fixed-dose combinations or branded drugs: they were thus sold only in the private sector and not the public sector, which uses different product forms (including loose pills) under non-proprietary names. Research in context Evidence before this study

Methods Overview We drew from a large, nationally representative dataset for private sector drug sales across the country, collected by the organisation IMS Health. We limited the analysis to 189 drugs containing rifampicin, which have fewer non-tuberculosis indications than, for example, fluoroquinolones. These 189 products capture all rifampicin-containing drugs being sold in India between 2013 and 2014. All products were fixed-dose combinations or branded drugs: they were thus sold only in the private sector and not the public sector, which uses different product forms (including loose pills) under non-proprietary names. Research in context Evidence before this study Not all incident cases of tuberculosis are reported to public health authorities: WHO estimates overall tuberculosis incidence in India by estimating the proportion of incident cases that are notified (the case detection rate [CDR]), and dividing published tuberculosis notifications by this fraction. In 2014, this approach suggested that over 800 000 tuberculosis cases in India escaped diagnosis by the public health-care system: most of these cases are assumed to have been treated in the private sector. However, CDR estimates are based on expert opinion, with the most recent estimate varying substantially from previous years. We searched PubMed for all studies with keywords “India”, “tuberculosis”, “private”, and “burden”, finding 25 studies for articles published in English from inception until May 31, 2016. Most of these related to the quality of tuberculosis care, whereas one study from 2011 assessed the amount of drug sales in the private sector in India and nine other countries. With a focus on market size estimates, this study also presented an illustrative estimate for how many patients were on treatment for tuberculosis.

of these related to the quality of tuberculosis care, whereas one study from 2011 assessed the amount of drug sales in the private sector in India and nine other countries. With a focus on market size estimates, this study also presented an illustrative estimate for how many patients were on treatment for tuberculosis. Added value of this study There is a need for systematic estimates of private sector tuberculosis burden that are independent of expert opinion. We used updated data from 2013 and 2014 for anti-tuberculosis drug sales in the private sector in India, adjusted for indication of use and data capture. With that empirical data, we built on previous work by systematically exploring the effect of assumptions of duration of treatment, and the extent of over-diagnosis of tuberculosis, on the number of patients treated in the private sector. Although there is limited evidence for either of these parameters, we modelled a range of scenarios to assess the feasibility of current estimates based on expert opinion. Implications of all the available evidence Tuberculosis treatment in the private sector is considerably greater than previous estimates suggest, and estimates of tuberculosis disease burden for India are implausibly low. This study illustrates the need to address the burden of tuberculosis treated by the private sector and improve surveillance. This study also raises an urgent need to revise current estimates of tuberculosis burden, informed by more systematic evidence relating to tuberculosis management in the private sector.

bly low. This study illustrates the need to address the burden of tuberculosis treated by the private sector and improve surveillance. This study also raises an urgent need to revise current estimates of tuberculosis burden, informed by more systematic evidence relating to tuberculosis management in the private sector. We aimed to estimate the treatment volume, or the total patient-months of treatment for tuberculosis in the private sector, taking account of both the proportion of prescriptions for a given drug that are for tuberculosis, and the proportion of total drug sales that are captured by IMS Health data. We found estimates at the state level in India, as well as on the national level, for 2013 and 2014. We also estimated 95% credible intervals, informed by uncertainty in the input parameters. Calculating volume (patient-months) of treatment Each product is uniquely identified by its product code, indexed i in the analysis. We define the following parameters for data in a given state and year: Ni is the IMS data for total packs of product i sold, ci is the proportion of total sales of product i that are captured by IMS data, mi is the total months of tuberculosis treatment represented by one pack of product i, and pi is the proportion of prescriptions containing product i that are for tuberculosis. When each of these quantities is specified, the total number of patient-months of treatment (PM), in a given state and year, is then given by a sum over all product codes i: PM=∑iNiCimipi

ment represented by one pack of product i, and pi is the proportion of prescriptions containing product i that are for tuberculosis. When each of these quantities is specified, the total number of patient-months of treatment (PM), in a given state and year, is then given by a sum over all product codes i: PM=∑iNiCimipi That is, adjusting sales data (Ni) for IMS data coverage (ci), the duration of treatment associated with each product form (mi) and the indications for tuberculosis versus other diseases (pi). Ni is measured directly. In practice, each of the remaining parameters carries some uncertainty, which we captured by modelling them as random variables, using distributions described below. Using Latin hypercube sampling, we took 10 000 samples for each of the parameters (ci, mi, and pi over all product codes i), and then calculated PM for each sample using the equation. From the resulting ensemble of 10 000 estimates for PM, we then obtained the point estimate for the patient-months using the median, and the uncertainty intervals using the 2·5th and 97·5th percentiles. We then repeated this process for each state and year.

i), and then calculated PM for each sample using the equation. From the resulting ensemble of 10 000 estimates for PM, we then obtained the point estimate for the patient-months using the median, and the uncertainty intervals using the 2·5th and 97·5th percentiles. We then repeated this process for each state and year. Data sources and probability distributions for input parameters For the total drug sales Ni we used state-specific data from the IMS Health Drug Sales Audit. These are monthly drug sales data reported to IMS Health by a recruited panel of stockists. We collected monthly drug sales data using invoices raised for sales of goods to retailers and sub-stockists, hospitals and hospital retailers, and dispensing doctors. Overall, IMS Health's combined drug sales audit in the retail, hospital, and dispensing doctors sectors was estimated to account for over 87% of the total Indian pharmaceutical market in 2014.11

invoices raised for sales of goods to retailers and sub-stockists, hospitals and hospital retailers, and dispensing doctors. Overall, IMS Health's combined drug sales audit in the retail, hospital, and dispensing doctors sectors was estimated to account for over 87% of the total Indian pharmaceutical market in 2014.11 For the proportion of prescriptions pi of product i that are for tuberculosis, we drew from the IMS Medical Health Audit, consisting of monthly prescription data from a panel of 4600 doctors following internationally recognised medical practice, and translating to over 800 000 prescriptions every month. The panel of doctors is recruited through a sampling exercise that takes into account the region, specialty type, and patient turnover. In these data, if product i has P prescriptions of which T are for tuberculosis, then we modelled pi as a β-distributed random variable, with shape and scale parameters T + 1 and P – T + 1, respectively. Data are available at the regional level, but not at the state level. Accordingly, for each state we selected the data from the relevant region.

has P prescriptions of which T are for tuberculosis, then we modelled pi as a β-distributed random variable, with shape and scale parameters T + 1 and P – T + 1, respectively. Data are available at the regional level, but not at the state level. Accordingly, for each state we selected the data from the relevant region. For mi, we again drew from the Medical Health Audit data. In particular, prescriptions of product code i have a certain frequency distribution, available from the audits. Putting this together with the duration of treatment associated with each dose, we constructed the probability distribution for the number of months associated with each prescription of product code i. Again, since these data are only regionally stratified, for each state we used the corresponding, region-specific estimates.

g this together with the duration of treatment associated with each dose, we constructed the probability distribution for the number of months associated with each prescription of product code i. Again, since these data are only regionally stratified, for each state we used the corresponding, region-specific estimates. Finally for ci, we used data from IMS Health data validation studies. In brief, at the end of each year, pharmaceutical companies subscribed to IMS are supplied with IMS estimates for their yearly sales volume, for comparison with their actual sales volume. Not all rifampicin-containing products are included in these studies. Accordingly, for each product code in the present work (analysis products), we estimated ci using those products in the validation study (validation products) having a comparable volume of sales. In particular, we grouped validation products by volume (whether high, moderate, or low volume products), found the mean and variance for each volume category, and then modelled IMS coverage within each category as a normal distribution. By categorising analysis products in the same way, we modelled ci for each required product using the normal distribution from the relevant volume category.

erate, or low volume products), found the mean and variance for each volume category, and then modelled IMS coverage within each category as a normal distribution. By categorising analysis products in the same way, we modelled ci for each required product using the normal distribution from the relevant volume category. Implications of treatment volume for burden (numbers of patients) Given an estimate of PM in a given year, the corresponding number of patients receiving tuberculosis treatment is given by PM / D, where D is the average duration (in months) for which patients take tuberculosis treatment in the private sector. However, not all of these patients might genuinely have tuberculosis. To adjust for potential overdiagnosis, we incorporated the positive predictive value (PPV) of tuberculosis diagnosis in the private sector (ie, the proportion of people diagnosed with tuberculosis in the private sector who genuinely have tuberculosis). Therefore, overall the number of patients with tuberculosis receiving private-sector treatment in a given year is estimated simply as PM × (PPV / D). In the absence of systematic, quantitative estimates for these parameters, we present results for a range of scenarios for PPV and D.

or who genuinely have tuberculosis). Therefore, overall the number of patients with tuberculosis receiving private-sector treatment in a given year is estimated simply as PM × (PPV / D). In the absence of systematic, quantitative estimates for these parameters, we present results for a range of scenarios for PPV and D. Patient-months of treatment in the public sector To compare against the amount of treatment in the public sector, we used RNTCP notifications and, for simplicity, assumed 6 months of treatment for new cases and 9 months of treatment for retreatment. Because some patients might not complete treatment even in the public sector, this approach yields an upper bound for patient-months of treatment. Consequently, this approach would tend to be conservative with respect to the relative amount of treatment in the private versus public sectors (ie, tending to underestimate this quantity). Role of the funding source This work was funded by the Bill & Melinda Gates Foundation. PD is affiliated with the Bill & Melinda Gates Foundation and was involved in the conception of the study, preparation of the manuscript, and interpretation of results, but had no role in the data analysis. The funder otherwise had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

f results, but had no role in the data analysis. The funder otherwise had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Table 1 shows estimates for the total patient-months of treatment (PM) in the private sector in 2013 and 2014, with a comparison against corresponding numbers in the public sector. Overall, estimates are stable between the years, although there is noticeable variation between states in the relative amount of treatment between private and public sectors. At one extreme, Orissa shows the public sector having 1·5–2·8 times as many PM as the private sector (taking the inverse of the ratios shown). At the other extreme, the private sector in Bihar provides over three times as many PM as the public sector. Overall, on a national level in both years, there was roughly twice as much tuberculosis treatment in the private sector as in the public sector. Although the analysis focuses on rifampicin-containing drugs, other tuberculosis drugs (isoniazid and ethambutol) showed similar sales volumes on a national level over this period (appendix).

ional level in both years, there was roughly twice as much tuberculosis treatment in the private sector as in the public sector. Although the analysis focuses on rifampicin-containing drugs, other tuberculosis drugs (isoniazid and ethambutol) showed similar sales volumes on a national level over this period (appendix). To translate these population estimates to numbers of patients being treated (whether or not they are genuine tuberculosis cases), table 2 shows estimates for 2014 under different scenarios ranging from 3 months to 9 months, with a comparison against numbers of patients registered for treatment under RNTCP. The appendix shows corresponding estimates for 2013.

o numbers of patients being treated (whether or not they are genuine tuberculosis cases), table 2 shows estimates for 2014 under different scenarios ranging from 3 months to 9 months, with a comparison against numbers of patients registered for treatment under RNTCP. The appendix shows corresponding estimates for 2013. Finally, to estimate actual burden of tuberculosis cases in the private sector, the figure shows estimates for 2014, under a range of scenarios for the PPV of tuberculosis diagnosis in the private sector, and for the average duration of treatment in the private sector. For illustration, and in the absence of systematic data on either of these parameters, the diamond marks a moderate set of parameter values: if a patient diagnosed with tuberculosis in the private sector undergoes 4 months of treatment on average, and if 50% of tuberculosis diagnoses in the private sector are genuine cases of tuberculosis, then these figures suggest that 2·2 million genuine cases of tuberculosis were treated in the private sector in 2014 (compared with 1·42 million patients treated in the public sector in the same year). This estimate increases when assuming higher values for PPV, and when assuming shorter average treatment duration.

is, then these figures suggest that 2·2 million genuine cases of tuberculosis were treated in the private sector in 2014 (compared with 1·42 million patients treated in the public sector in the same year). This estimate increases when assuming higher values for PPV, and when assuming shorter average treatment duration. Discussion The vast and fragmented private health-care sector is a prominent feature in the health landscape of India. In the context of tuberculosis, this sector is difficult to study and systematically characterise, yet remains crucial for understanding and managing the overall burden of tuberculosis. In this work, we took advantage of systematic collection of drug sales data in the private sector to address this gap, presenting new estimates that suggest the burden of tuberculosis might be considerably higher than previously recognised. The key output of this approach is the volume (patient-months) of patient treatment in the private sector, which is twice as much as that provided in the public sector. On any given day, this translates on average to 1·46 million people being on tuberculosis treatment, more than 0·12% of the country's population. Moreover, tuberculosis treatment in the private sector is typically paid for by out-of-pocket expenditure; if a 6-month course of first-line, anti-tuberculosis medication costs US$20, our estimates imply that in 2014, over $59 million was spent in out-of-pocket expenditure on first-line tuberculosis drugs alone.

ation. Moreover, tuberculosis treatment in the private sector is typically paid for by out-of-pocket expenditure; if a 6-month course of first-line, anti-tuberculosis medication costs US$20, our estimates imply that in 2014, over $59 million was spent in out-of-pocket expenditure on first-line tuberculosis drugs alone. Estimates by WHO use expert opinion for case detection rates to project from notifications to overall incidence.9 The most recent estimates imply that in 2014 about 800 000 patients went untreated by the public sector.1 Reconciling these estimates with 17·8 million patient-months of private-sector treatment in 2014 would require either a very low PPV (27% if assuming a 6-month treatment duration) or a long treatment duration (over 11 months if assuming a PPV of 50%). Instead, taking plausible ranges of 40–60% for PPV and 2–6 months for treatment duration suggests that in 2014 alone, 1·19–5·34 million tuberculosis patients received private-sector tuberculosis treatment. The midpoint in this range corresponds to a private-sector tuberculosis burden of 2·2 million cases, more than twice the burden suggested by previous assumptions.

PV and 2–6 months for treatment duration suggests that in 2014 alone, 1·19–5·34 million tuberculosis patients received private-sector tuberculosis treatment. The midpoint in this range corresponds to a private-sector tuberculosis burden of 2·2 million cases, more than twice the burden suggested by previous assumptions. Our findings have implications for the tuberculosis strategy in India. First, the vast disorganised private health-care sector poses major challenges to tuberculosis control. India's RNTCP has committed to providing free, high-quality tuberculosis care to patients in the private sector.12 Initiatives such as private-sector engagement to improve tuberculosis care in this sector, offer potential mechanisms for realising these goals.8 In this context, our results suggest that the scale of the challenge is substantially larger than has hitherto been appreciated. These findings underscore the need for redoubled efforts to reach patients being treated in the private sector, to deliver the highest possible standards of tuberculosis care. Second, our work points to the urgent need for further strengthening of tuberculosis surveillance in the private sector. Although there has been increasing notification of tuberculosis cases by the private sector to public health authorities, these accounted in 2014 for 106 414 patients13—a level far below that estimated here. Emerging initiatives, such as the proposed provision of free, daily-dosed tuberculosis treatment to all those needing it in the private sector, could bring about important steps in this direction.14

ector to public health authorities, these accounted in 2014 for 106 414 patients13—a level far below that estimated here. Emerging initiatives, such as the proposed provision of free, daily-dosed tuberculosis treatment to all those needing it in the private sector, could bring about important steps in this direction.14 Third, our findings highlight uncertainty around the true burden of tuberculosis in India. Methods for estimating this burden should be complemented by independent approaches generating primary data. In addition to the surveillance needs mentioned above, a national prevalence survey would provide direct evidence for the numbers of patients receiving treatment in the private sector. Moreover model-based approaches, such as the Global Burden of Disease study,15 offer the capability to collate disparate but important sources for estimating tuberculosis burden. In future, findings such as those presented here could constitute an additional source of evidence for refining these and other analytical approaches.

model-based approaches, such as the Global Burden of Disease study,15 offer the capability to collate disparate but important sources for estimating tuberculosis burden. In future, findings such as those presented here could constitute an additional source of evidence for refining these and other analytical approaches. Previous work on the role of the private sector used interviews of patients diagnosed with tuberculosis in 30 districts in India, to estimate that nearly half of patients were on treatment outside RNTCP.2 Relying as it does on self-reported tuberculosis, these estimates can be interpreted as a lower bound of the amount of tuberculosis treatment in the private sector. Another study, also using drug sales in the private sector, cast valuable light on the private market for different tuberculosis drugs.10 Our findings for overall treatment volumes in India are broadly consistent. Moreover, Delhi features prominently in our results for the amount of private-sector treatment relative to the public sector (table 1). This result is consistent with other findings in the city,16 where engagement with private sector providers led to a greater than ten times increase in tuberculosis notifications, indicative of a large tuberculosis burden being managed by this sector. Although it might be tempting to hold India's large informal health sector responsible for the observed high usage of tuberculosis drugs, recent work from India, using standardised patients, show that anti-tuberculosis drugs are rarely dispensed by pharmacists, informal providers, and practitioners of alternative medical systems.6, 17 Thus, qualified, allopathic doctors in India are the primary source of anti-tuberculosis drug prescriptions, and should be the target of engagement and antimicrobial stewardship efforts.

ti-tuberculosis drugs are rarely dispensed by pharmacists, informal providers, and practitioners of alternative medical systems.6, 17 Thus, qualified, allopathic doctors in India are the primary source of anti-tuberculosis drug prescriptions, and should be the target of engagement and antimicrobial stewardship efforts. Our approach has some limitations. First, we do not have reliable estimates for the PPV of tuberculosis diagnosis in the private sector, nor for the mean duration of treatment (D) in the private sector. The estimates that we present for numbers of patients with tuberculosis, under different scenarios, should thus be taken as illustrative, and not definitive. Estimating such parameters in a systematic way is a real challenge. Nonetheless, new methods are emerging, such as the use of standardised patients to assess the quality of tuberculosis care in this sector.6 In future, these and other approaches could be valuable in quantifying PPV and D more precisely, and more broadly for systematically studying the private health-care sector. Second, in the simple estimates in the figure, we neglect complexities such as the potential for a patient to receive treatment in the private sector first, and subsequently in the public sector. However, a nationally representative study18 in 2010 estimated that such patients accounted for about 8% of all tuberculosis cases that were notified in that year. These findings suggest that the numbers are not so great as to considerably bias our estimates. Further work could aim to extend these findings to more recent years.

a nationally representative study18 in 2010 estimated that such patients accounted for about 8% of all tuberculosis cases that were notified in that year. These findings suggest that the numbers are not so great as to considerably bias our estimates. Further work could aim to extend these findings to more recent years. Third, there are several types of patient with tuberculosis that the data do not capture. For example, those who could be receiving treatment for tuberculosis in the informal health-care sector, those who have not contacted the health-care system, or those being treated for multidrug-resistant tuberculosis in the private sector. Moreover, there is evidence to suggest that some patients could be treated for tuberculosis with other drugs such as fluoroquinolones in the private sector,6 although there is no systematic evidence for the proportion of patients receiving these drugs in combination with rifampicin-containing products. Nonetheless, taken together, all these factors would suggest that the true burden of tuberculosis is even greater than suggested in the present analysis.

the private sector,6 although there is no systematic evidence for the proportion of patients receiving these drugs in combination with rifampicin-containing products. Nonetheless, taken together, all these factors would suggest that the true burden of tuberculosis is even greater than suggested in the present analysis. Overall, the approach described here cannot replace traditional approaches to surveillance, including routine notifications and periodic surveys. There remains a pressing need to strengthen and widen these systems. Nonetheless, the implications of this analysis could offer additional perspectives on such a vast and complex health-care system as in India. In future these and other approaches, in combination with existing and improved sources of data, could help to build a truly comprehensive picture of the burden of tuberculosis in India. Supplementary Material Supplementary appendix Contributors PD and SAN conceived the study. NA, DB, TV, NM, and LS did the analysis. NA, DB, SK, SAN, and PD interpreted the results. NA wrote a first draft of the manuscript, and all authors contributed towards the final version of the manuscript. Declaration of interests We declare no competing interests. Figure Implications of treatment volume in 2014 for tuberculosis burden managed by the private sector in India

Contributors PD and SAN conceived the study. NA, DB, TV, NM, and LS did the analysis. NA, DB, SK, SAN, and PD interpreted the results. NA wrote a first draft of the manuscript, and all authors contributed towards the final version of the manuscript. Declaration of interests We declare no competing interests. Figure Implications of treatment volume in 2014 for tuberculosis burden managed by the private sector in India Estimates are shown for the number of patients with tuberculosis treated by the private sector (see colour bar for numbers in millions), under different scenarios for the average duration of treatment in the private sector, and the proportion of private-sector tuberculosis diagnoses that genuinely have tuberculosis. The diamond illustrates a moderate parameter regime, in which 50% of diagnoses in the private sector genuinely have tuberculosis, and the average treatment duration is for 4 months. This corresponds to an estimated 2·2 million patients being treated in the private sector in 2014. Table 1 Patient-months of treatment in 2013 and 2014 across India

Estimates are shown for the number of patients with tuberculosis treated by the private sector (see colour bar for numbers in millions), under different scenarios for the average duration of treatment in the private sector, and the proportion of private-sector tuberculosis diagnoses that genuinely have tuberculosis. The diamond illustrates a moderate parameter regime, in which 50% of diagnoses in the private sector genuinely have tuberculosis, and the average treatment duration is for 4 months. This corresponds to an estimated 2·2 million patients being treated in the private sector in 2014. Table 1 Patient-months of treatment in 2013 and 2014 across India Patient-months in 2013 Patient-months in 2014 Private sector (thousands) Public sector (thousands) Ratio, private to public Private sector (thousands) Public sector (thousands) Ratio, private to public Andhra Pradesh 1020 (793–1328) 683 1·5 (1·2–1·9) 947 (736–1258) 709 1·3 (1·0–1·8) Assam & North East 344 (277–458) 348 1·0 (0·8–1·3) 375 (298–518) 364 1·0 (0·8–1·4) Bihar 1560 (1357–1892) 441 3·5 (3·1–4·3) 1567 (1356–1950) 446 3·5 (3·0–4·4) Chhattisgarh 300 (244–380) 163 1·8 (1·5–2·3) 266 (218–341) 183 1·5 (1·2–1·9) Delhi 1175 (934–1504) 335 3·5 (2·8–4·5) 1108 (880–1496) 356 3·1 (2·5–4·2) Goa 18 (14–26) 11 1·6 (1·2–2·3) 19 (13–28) 10 1·8 (1·3–2·7) Gujarat 1044 (837–1292) 501 2·1 (1·7–2·6) 976 (790–1231) 525 1·9 (1·5–2·3) Haryana 357 (290–452) 258 1·4 (1·1–1·8) 353 (286–459) 266 1·3 (1·1–1·7) Himachal Pradesh 48 (38–67) 90 0·5 (0·4–0·7) 54 (41–78) 95 0·6 (0·4–0·8) Jammu & Kashmir 180 (145–240) 72 2·5 (2·0–3·3) 133 (108–180) 68 2·0 (1·6–2·7) Jharkhand 309 (265–392) 225 1·4 (1·2–1·7) 377 (314–495) 231 1·6 (1·4–2·1) Karnataka 556 (409–744) 406 1·4 (1–1·8) 558 (395–776) 404 1·4 (1·0–1·9) Kerala 220 (168–293) 154 1·4 (1·1–1·9) 174 (133–237) 149 1·2 (0·9–1·6) Madhya Pradesh 1166 (985–1413) 599 1·9 (1·6–2·4) 1008 (837–1241) 644 1·6 (1·3–1·9) Maharashtra 1639 (1296–2074) 906 1·8 (1·4–2·3) 1623 (1257–2063) 890 1·8 (1·4–2·3) Orissa 123 (103–171) 288 0·4 (0·4–0·6) 141 (115–188) 292 0·5 (0·4–0·6) Punjab 461 (386–587) 265 1·7 (1·5–2·2) 403 (329–521) 268 1·5 (1·2–1·9) Rajasthan 1063 (900–1307) 595 1·8 (1·5–2·2) 1039 (865–1274) 596 1·7 (1·5–2·1) Tamilnadu 672 (508–891) 530 1·3 (1·0–1·7) 619 (467–825) 559 1·1 (0·8–1·5) Uttar Pradesh 4942 (4214–6232) 1615 3·1 (2·6–3·9) 5041 (4292–6601) 1600 3·2 (2·7–4·1) Uttaranchal 328 (278–445) 85 3·8 (3·3–5·2) 331 (275–431) 96 3·4 (2·9–4·5) West Bengal 390 (317–552) 602 0·6 (0·5–0·9) 470 (373–680) 578 0·8 (0·6–1·2) National 18 118 (16 993–19 717) 9180 2·0 (1·9–2·1) 17 793 (16 709–19 841) 9340 1·9 (1·8–2·1) Data in parentheses are 95% credible intervals. Private sector represents estimates from IMS data.

445) 85 3·8 (3·3–5·2) 331 (275–431) 96 3·4 (2·9–4·5) West Bengal 390 (317–552) 602 0·6 (0·5–0·9) 470 (373–680) 578 0·8 (0·6–1·2) National 18 118 (16 993–19 717) 9180 2·0 (1·9–2·1) 17 793 (16 709–19 841) 9340 1·9 (1·8–2·1) Data in parentheses are 95% credible intervals. Private sector represents estimates from IMS data. Public sector numbers are obtained using Revised National TB Control Programme notifications and assuming treatment durations of 6 months and 9 months for new and retreatment cases, respectively. For conciseness, the smallest states have been aggregated as follows: North East includes Arunachal Pradesh, Manipur, Meghalaya, Mizoram, Nagaland, and Tripura; Gujarat includes Gujarat and Daman & Diu; Kerala includes Kerala and Lakshadweep; Maharashtra includes Maharashtra and Dadar and Nagar Haveli; Punjab includes Punjab and Chandigarh; Tamil Nadu includes Tamil Nadu, Pondicherry, and Andaman & Nicobar; West Bengal includes West Bengal and Sikkim. Table 2 Estimated numbers of patients receiving tuberculosis treatment in 2014 across India

Public sector numbers are obtained using Revised National TB Control Programme notifications and assuming treatment durations of 6 months and 9 months for new and retreatment cases, respectively. For conciseness, the smallest states have been aggregated as follows: North East includes Arunachal Pradesh, Manipur, Meghalaya, Mizoram, Nagaland, and Tripura; Gujarat includes Gujarat and Daman & Diu; Kerala includes Kerala and Lakshadweep; Maharashtra includes Maharashtra and Dadar and Nagar Haveli; Punjab includes Punjab and Chandigarh; Tamil Nadu includes Tamil Nadu, Pondicherry, and Andaman & Nicobar; West Bengal includes West Bengal and Sikkim. Table 2 Estimated numbers of patients receiving tuberculosis treatment in 2014 across India Patients in private sector (thousands) Patients in public sector (thousands) 3 month duration 6 month duration 9 month duration Andhra Pradesh 315 (245–419) 157 (122–209) 105 (81–139) 107 Assam 125 (99–172) 62 (49–86) 41 (33–57) 55 Bihar 522 (452–650) 261 (226–325) 174 (150–216) 67 Chhattisgarh 88 (72–113) 44 (36–56) 29 (24–37) 28 Delhi 369 (293–498) 184 (146–249) 123 (97–166) 53 Goa 6 (4–9) 3 (2–4) 2 (1–3) 1 Gujarat 325 (263–410) 162 (131–205) 108 (87–136) 77 Haryana 117 (95–153) 58 (47–76) 39 (31–51) 39 Himachal Pradesh 18 (13–26) 9 (6–13) 6 (4–8) 14 Jammu & Kashmir 44 (36–60) 22 (18–30) 14 (12–20) 10 Jharkhand 125 (104–165) 62 (52–82) 41 (34–55) 35 Karnataka 186 (131–258) 93 (65–129) 62 (43–86) 61 Kerala 58 (44–79) 29 (22–39) 19 (14–26) 23 Madhya Pradesh 336 (279–413) 168 (139–206) 112 (93–137) 99 Maharashtra 541 (419–687) 270 (209–343) 180 (139–229) 133 Orissa 47 (38–62) 23 (19–31) 15 (12–20) 45 Punjab 134 (109–173) 67 (54–86) 44 (36–57) 40 Rajasthan 346 (288–424) 173 (144–212) 115 (96–141) 90 Tamilnadu 206 (155–275) 103 (77–137) 68 (51–91) 85 Uttar Pradesh 1680 (1430–2200) 840 (715–1100) 560 (476–733) 245 Uttaranchal 110 (91–143) 55 (45–71) 36 (30–47) 14 West Bengal 156 (124–226) 78 (62–113) 52 (41–75) 88 National 5931 (5569–6613) 2965 (2784–3306) 1977 (1856–2204) 1421 Data in parentheses are 95% credible intervals. In the private sector, estimates are shown under different assumptions for the average duration of treatment, ranging from 3 months to 9 months. In the public sector, the total number of cases registered for treatment by the Revised National TB Control Programme in 2014 are shown. In the private sector, not all patients receiving tuberculosis treatment might genuinely have tuberculosis: the figures are adjusted for potential overdiagnosis in the private sector.

Introduction Antimicrobial resistance is a global health emergency,1, 2 and the indiscriminate use of antibiotics is a major driver.3, 4 Although India ranks first in total antibiotic use worldwide,2 the absence of data linking antibiotic use to underlying illnesses makes it hard to assess the appropriateness of such use in view of India's high infectious disease burden. With some of the highest incidences of drug-resistant bacterial pathogens in the world, identification of the sources and circumstances of antibiotic abuse as opposed to use is a crucial first step to understanding what can be done about antibiotic overuse in India.4 Here, we develop a unique method to address this gap, focusing our attention on a specific illness, tuberculosis, and a specific source of health care—pharmacies.

sources and circumstances of antibiotic abuse as opposed to use is a crucial first step to understanding what can be done about antibiotic overuse in India.4 Here, we develop a unique method to address this gap, focusing our attention on a specific illness, tuberculosis, and a specific source of health care—pharmacies. Our choice of tuberculosis, a disease that affects 2·2 million Indians every year, as a lens through which to investigate antibiotic use is driven by several factors. The symptoms of early pulmonary tuberculosis are common, non-specific, non-severe, and persistent. In this case, assessment of pharmacist behaviour provides a realistic and externally valid estimate of unnecessary antibiotic use. Further, indiscriminate drug use can harm both the patient and the efficacy of existing anti-tuberculosis treatments. For instance, tuberculosis symptoms subside temporarily with the use of fluoroquinolones or corticosteroids, delaying diagnosis and leading to the possibility that patients receive several antibiotic courses for the wrong diagnosis.5 Partial courses of anti-tuberculosis drugs can result in drug resistance.6 Finally, international and national guidelines for the optimum management of tuberculosis cases7, 8 allow the assessment of the extent of antibiotic misuse.

e possibility that patients receive several antibiotic courses for the wrong diagnosis.5 Partial courses of anti-tuberculosis drugs can result in drug resistance.6 Finally, international and national guidelines for the optimum management of tuberculosis cases7, 8 allow the assessment of the extent of antibiotic misuse. Our focus on pharmacies is premised on the belief that their practices contribute to the availability and use of antibiotics in the population.4 This premise is partly the result of their widespread availability—more than 750 000 private retail pharmacies provide easy access to drugs.9 However, our premise also reflects the willingness of pharmacists to provide prescription-only drugs to patients. Despite clear guidelines on the use of over-the-counter versus prescription-only drugs,10 enforcement is widely believed to be suboptimum.11, 12 Pharmacies are thought to be dispensing antibiotics and anti-tuberculosis drugs without prescriptions. Many tuberculosis patients do seek medical advice and drugs from pharmacies,13 driven by the ease of access and the possibility of avoiding consultation charges by doctors.14 Research in context Evidence before this study

Our focus on pharmacies is premised on the belief that their practices contribute to the availability and use of antibiotics in the population.4 This premise is partly the result of their widespread availability—more than 750 000 private retail pharmacies provide easy access to drugs.9 However, our premise also reflects the willingness of pharmacists to provide prescription-only drugs to patients. Despite clear guidelines on the use of over-the-counter versus prescription-only drugs,10 enforcement is widely believed to be suboptimum.11, 12 Pharmacies are thought to be dispensing antibiotics and anti-tuberculosis drugs without prescriptions. Many tuberculosis patients do seek medical advice and drugs from pharmacies,13 driven by the ease of access and the possibility of avoiding consultation charges by doctors.14 Research in context Evidence before this study Antimicrobial resistance is a global health emergency, and, as the largest consumer of antibiotics, India is at highest risk. The standardised patient method can help to assess the extent and appropriateness of antibiotic use because such use can be directly related to the underlying illness of the patient. To identify previous research on this topic, we searched PubMed and Google Scholar using a combination of the terms “standardized patients” (“mystery clients”, “fake patients”, or “simulated patients”), “pharmacy” (“pharmacist” or “chemist”), and “tuberculosis” with and without the keyword “India” for articles published in English until March 31, 2015. Our search showed that previous studies of physician management of standardised patients in India have reported unnecessary antibiotic prescribing for various conditions, including tuberculosis, diarrhoea, asthma, and angina. However, these studies have not addressed antibiotic abuse by pharmacists, who respond to the health-care needs of a substantial proportion of India's population.

tandardised patients in India have reported unnecessary antibiotic prescribing for various conditions, including tuberculosis, diarrhoea, asthma, and angina. However, these studies have not addressed antibiotic abuse by pharmacists, who respond to the health-care needs of a substantial proportion of India's population. Added value of this study We used standardised patients to quantify the extent of antibiotic overuse in pharmacists for patients with tuberculosis. We developed two standardised patient cases: first, a patient presenting with 2–3 weeks of pulmonary tuberculosis symptoms (Case 1); and second, a patient with microbiologically confirmed pulmonary tuberculosis (Case 2). Across all interactions, 319 (27%) of 1200 (95% CI 24–29) resulted in the use of an antibiotic although no pharmacy dispensed first-line anti-tuberculosis drugs. Ideal case management, defined as referrals without the use of antibiotics or steroids, was much lower in Case 1 interactions (13%) than Case 2 interactions (62%). Our study results add to the growing evidence on antibiotic abuse, but also underscore that the use and misuse of antibiotics are mediated by drug category and the information that patients present. Although antibiotic use is high and such use can delay diagnosis, none of the pharmacies dispensed first-line anti-tuberculosis drugs, and the use of stronger fluoroquinolone antibiotics and heavily restricted drug classes was low. Furthermore, the use of all antibiotics decreased sharply when the patient's diagnosis was made available to the pharmacists.

uch use can delay diagnosis, none of the pharmacies dispensed first-line anti-tuberculosis drugs, and the use of stronger fluoroquinolone antibiotics and heavily restricted drug classes was low. Furthermore, the use of all antibiotics decreased sharply when the patient's diagnosis was made available to the pharmacists. Implications of all the available evidence Our findings suggest that non-adherence to regulatory standards is higher when the patient's condition is unknown, and that pharmacies prefer to treat in such cases rather than refer the patient to appropriate care. These findings can inform interventions to engage pharmacies in tuberculosis control and antimicrobial stewardship. Tuberculosis is a major problem in all three cities studied (Delhi, Mumbai, and Patna), with notification rates (officially reported) of 294, 210, and 77 per 100 000, respectively.15 However, these rates are probably underestimated because many cases treated in the private sector are not notified.16 All three cities are experiencing rising rates of drug-resistant tuberculosis, especially in the city of Mumbai,17 and it is widely believed that pharmacists are a key component of the dispensing landscape and often a first contact for primary care.

d because many cases treated in the private sector are not notified.16 All three cities are experiencing rising rates of drug-resistant tuberculosis, especially in the city of Mumbai,17 and it is widely believed that pharmacists are a key component of the dispensing landscape and often a first contact for primary care. Guidelines for pharmacies are specified under the Ministry of Health and Family Welfare's Drugs and Cosmetics Rules Act, 1945.10 All antibiotics and steroids are listed under two different schedules—Schedule H and Schedule H1. Schedule H drugs cannot be given to patients without a prescription from a qualified medical practitioner. In 2013, regulations were further tightened, with anti-tuberculosis drugs (isoniazid, rifampicin, ethambutol, and pyrazinamide) and some fluoroquinolones (such as moxifloxacin and levofloxacin, used in the treatment of tuberculosis) listed on a newly created Schedule H1. For H1 drugs, pharmacies require both a prescription from a qualified medical practitioner and a separate register to record the name and address of the prescriber, the patient, the names of the drugs and the quantity supplied.18 Schedule X, the most restrictive list, includes drugs such as narcotics, which require a prescription from a qualified provider to be retained by the retailer for 2 years.19

Contributors MZ, ASD, DMC, CG, KF, RH, SH, NI, MKi, LR, SR-G, WAW, and KW contributed to the study design. NA, SA, AMC, ADr, MD, ZH, AHusa, AHuss, MKa, MM, SVO, MSe, AS, and ST performed the laboratory testing. MZ and ASD analysed data, which was interpreted by all authors. MZ, ASD, KF, and MCR drafted the report with input from all other authors. All authors have seen and approved the final report. Declaration of interests We declare no competing interests. Table 1 Results of sequencing of the pncA gene (associated with pyrazinamide resistance) Azerbaijan Bangladesh Belarus (Minsk city) Pakistan South Africa (Gauteng) South Africa (KwaZulu Natal) New tuberculosis cases 530, 10·2% (7·6–12·8) 751, 2·6% (1·4–3·8) 144, 30·0% (22·6–37·3) 1299, 2·1% (1·3–2·9) 648, 2·8% (1·5–4·1) 444, 2·0% (0·8–3·3) Previously treated tuberculosis cases 218, 17·9% (12·7–23·0) 203, 13·8% (8·8–18·8) 57, 69·9% (58·4–81·4) 201, 8·9% (5·1–12·8) 145, 4·7% (1·5–7·8) 128, 10·5% (5·1–15·8) All tuberculosis cases 748, 12·6% (10·1–15·0) 955, 5·1% (3·4–6·8) 201, 42·1% (35·4–48·8) 1500, 3·0% (2·0–4·0) 877, 3·1% (1·9–4·4) 691, 3·9% (2·4–5·4) Rifampicin susceptible 619, 2·2% (1·1–3·4) 892, 2·5% (1·3–3·6) 103, 4·2% (0·1–8·3) 1397, 0·5% (0·1–0·8) 838, 1·3% (0·4–2·2) 657, 1·3% (0·4–2·3) Rifampicin resistant 129, 59·9% (51·0–68·9) 63, 36·7% (25·9–47·4) 98, 81·3% (73·7–88·9) 103, 39·5% (30·1–48·9) 39, 39·1% (22·9–55·3) 34, 49·1% (32·7–65·5) Resistance in rifampicin resistant vs rifampicin susceptible <0·0001 <0·0001 <0·0001 <0·0001 <0·0001 <0·0001 Resistance in newly vs previously treated 0·004 <0·0001 <0·0001 <0·0001 0·169 <0·0001 Data are number tested, % resistant (95% CI) or p value.

itioner and a separate register to record the name and address of the prescriber, the patient, the names of the drugs and the quantity supplied.18 Schedule X, the most restrictive list, includes drugs such as narcotics, which require a prescription from a qualified provider to be retained by the retailer for 2 years.19 We have previously assessed the quality of tuberculosis care in India by health-care providers using standardised patients and use a similar method to study the practices of staff at pharmacies.20 Although standardised patients are routinely used to assess pharmacy practices in low-income and high-income countries,21 to our knowledge, no study has used standardised patients to assess pharmacy practices for tuberculosis in India. In our previous study, we validated the use of standardised patients for tuberculosis and showed the viability and accuracy of this method for measuring quality of tuberculosis care along several dimensions, including very low likelihood of detection, minimum to no study participation risk for either standardised patients or health-care providers, and high levels of accurate recall of the clinical interaction among standardised patients. This study complements our previous validation study by extending the method to pharmacists. The method developed here addresses the dual objectives of, first, assessment of pharmacists' behaviour and drug use for a patient with a complaint, but no prescription. Second, it allows us to assess how case management and drug use differs when the diagnosis is unknown versus confirmed.

ing the method to pharmacists. The method developed here addresses the dual objectives of, first, assessment of pharmacists' behaviour and drug use for a patient with a complaint, but no prescription. Second, it allows us to assess how case management and drug use differs when the diagnosis is unknown versus confirmed. Methods Study design and setting This cross-sectional study was done in Delhi, Mumbai, and Patna.

ing the method to pharmacists. The method developed here addresses the dual objectives of, first, assessment of pharmacists' behaviour and drug use for a patient with a complaint, but no prescription. Second, it allows us to assess how case management and drug use differs when the diagnosis is unknown versus confirmed. Methods Study design and setting This cross-sectional study was done in Delhi, Mumbai, and Patna. Through this multi-site study we aimed to assess the medical advice and drug dispensing practices of pharmacies for standardised patients presenting with either presumptive tuberculosis (Case 1) or microbiologically confirmed tuberculosis (Case 2). By assessing the difference in antibiotic use across the two cases for the same pharmacists, we broke down the relative importance of antibiotic misuse arising from the lack of diagnosis (Case 1) versus antibiotic use despite a confirmed diagnosis for which antibiotics are contraindicated (Case 2). To set the benchmark for what pharmacists should do when faced with such patients, we used guidelines from the Government of India and the Indian Pharmaceutical Association. These guidelines specify that pharmacies should counsel patients about tuberculosis, identify and refer persons with tuberculosis symptoms to the nearest public health facilities for tuberculosis testing, and play a part in the provision of tuberculosis treatment.22 Therefore, pharmacists adhering to these guidelines should have referred the standardised patients to health-care providers without dispensing either antibiotics or steroids, both of which require a prescription.

c health facilities for tuberculosis testing, and play a part in the provision of tuberculosis treatment.22 Therefore, pharmacists adhering to these guidelines should have referred the standardised patients to health-care providers without dispensing either antibiotics or steroids, both of which require a prescription. Standardised patients The two cases of standardised patients used in our study were adapted from our validation study in Delhi.15 Standardised patients trained as Case 1 presented with 2–3 weeks of cough and fever and were directly seeking drugs from a pharmacy. Differential diagnosis for this case included upper respiratory tract infection, pneumonia, asthma and acute or chronic bronchitis; antibiotic use might be warranted for some of these conditions although not without a prescription from a doctor.

weeks of cough and fever and were directly seeking drugs from a pharmacy. Differential diagnosis for this case included upper respiratory tract infection, pneumonia, asthma and acute or chronic bronchitis; antibiotic use might be warranted for some of these conditions although not without a prescription from a doctor. Standardised patients trained as Case 2 presented with 1 month of cough and fever and a tuberculosis-positive laboratory report from a recent sputum smear test at a government dispensary. In this case, tuberculosis was confirmed, although the standardised patients, who presented as uninformed patients, made it clear that they did not fully understand what the report said. In this situation, the pharmacist plausibly could know the correct diagnosis and could recognise that short-term antibiotics would not help, but also could realise that the patient would still purchase antibiotics if offered because of their ignorance of the test results. Standardised patients did not present with drug prescriptions; table 1 shows their opening statements and case scenarios. After each pharmacy visit, standardised patients were debriefed with a structured questionnaire within 1 h of the visit. The accompanying appendix (pp 5,6) provides more details on the development of the cases and the recruitment and characteristics of the standardised patients in the study. Cases are available from the authors by request.

isit, standardised patients were debriefed with a structured questionnaire within 1 h of the visit. The accompanying appendix (pp 5,6) provides more details on the development of the cases and the recruitment and characteristics of the standardised patients in the study. Cases are available from the authors by request. Selection of pharmacies, standardised patient visits, and study size Standardised patients visited 54 pharmacies in Delhi using a convenience sample from 28 low-income localities in April, 2014. This phase of the study validated the approach and provided key parameter estimates for power calculations employed in Mumbai and Patna. Based on these power calculations we sent standardised patients to 308 randomly sampled pharmacies in Mumbai and 260 in Patna between Nov 5, 2014, and Nov 29, 2015. 1200 (96%)of 1244 interactions were completed as planned, and we completed both cases for a sampled pharmacy in 1156 (93%) of 1244 scheduled interactions. The appendix discusses the sample and sampling weights, case development, standardised patient recruitment, sample size calculations, drug identification, and deviations from the sampling scheme (pp 2–8).

pleted as planned, and we completed both cases for a sampled pharmacy in 1156 (93%) of 1244 scheduled interactions. The appendix discusses the sample and sampling weights, case development, standardised patient recruitment, sample size calculations, drug identification, and deviations from the sampling scheme (pp 2–8). We obtained approvals from the ethics committees of McGill University Health Centre in Montreal, Canada, and the Institute of Socio-Economic Research on Development and Democracy (ISERDD) in New Delhi. Both ethics committees approved a waiver from obtaining informed consent from pharmacies in Mumbai and Patna. All individuals who participated as standardised patients were hired as staff and trained to protect themselves from any harmful medical interventions, such as avoiding injections, invasive tests, or consuming any drugs at the pharmacy. Statistical analysis Our unit of analysis was a pharmacy-standardised patient interaction irrespective of who (pharmacy owners, pharmacists, or pharmacy assistants) the standardised patient interacted with. Whether the case was correctly managed was assessed from a tuberculosis perspective, consistent with Standards for Tuberculosis Care in India and International Standards for Tuberculosis Care.7, 8 We regarded ideal management for both cases as verbal or written referral to a health-care provider (public or private), without dispensing any antibiotics, including anti-tuberculosis drugs and fluoroquinolones, or steroids (table 1).

or Tuberculosis Care in India and International Standards for Tuberculosis Care.7, 8 We regarded ideal management for both cases as verbal or written referral to a health-care provider (public or private), without dispensing any antibiotics, including anti-tuberculosis drugs and fluoroquinolones, or steroids (table 1). We calculated the proportion and 95% CI for our primary outcome, the proportion of interactions that resulted in ideal management, as well as the proportion of interactions resulting in antibiotic, fluoroquinolone, and steroid use with appropriate sampling weights (appendix p 3). To assess the difference in case management and the use of drugs across the two cases, we used a random intercept logit model with indicator variables for each city as additional controls. In view of the study design and since every sampled pharmacy was attempted by both cases, the choice of model (logit, logit with fixed effects, or logit with random intercepts) should have yielded similar unbiased estimates, with differences arising only from the small portion of pharmacies that received one case but not the other. However, coefficients from the random-intercepts model are more precisely estimated. The appendix (pp 10–13) provides a series of alternate estimates, with both marginal effects and odds ratios from different model specifications and confirm that the results are very similar across specifications. All analyses were done using Stata (version 13).

the random-intercepts model are more precisely estimated. The appendix (pp 10–13) provides a series of alternate estimates, with both marginal effects and odds ratios from different model specifications and confirm that the results are very similar across specifications. All analyses were done using Stata (version 13). Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

the random-intercepts model are more precisely estimated. The appendix (pp 10–13) provides a series of alternate estimates, with both marginal effects and odds ratios from different model specifications and confirm that the results are very similar across specifications. All analyses were done using Stata (version 13). Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results 96 (16%) of 599 pharmacies (95% CI 13–19) referred Case 1 interactions to health-care providers, but because in 16 (17%) of these 96 cases the standardised patient was also given an antibiotic or steroid (95% CI 11–25), ideal case management (referral to a health-care provider without any antibiotics and steroids) occurred in 80 (13%) of 599 Case 1 interactions (95% CI 11–16). Overall, antibiotics were used in 221 (37%; 95% CI 33–41) of 599 interactions, steroids in 45 (8%; 95% CI 6–10), and fluoroquinolones in 61 (10%; 95% CI 8–13). Because Schedule H drugs also include prescription-only drugs that are not antibiotics or steroids (eg, ibuprofen or cetirizine), the use of these drugs was higher (401 [67%] of 599 interactions, 95% CI 63–71). The use of Schedule H1 drugs was notably lower (37 [6%] of 599, 95% CI 4–8) and Schedule X drugs and anti-tuberculosis drugs were never given. Table 2 provides the mean proportions of the key outcome variables in all cities combined for Case 1 and Case 2. Since the sampling scheme was different for Delhi compared with Mumbai and Patna, we also provide results excluding Delhi (table 2), and for each city by case (appendix pp 8,9).

losis drugs were never given. Table 2 provides the mean proportions of the key outcome variables in all cities combined for Case 1 and Case 2. Since the sampling scheme was different for Delhi compared with Mumbai and Patna, we also provide results excluding Delhi (table 2), and for each city by case (appendix pp 8,9). By contrast with Case 1, 401 (67%) of 601 pharmacies (95% CI 63–70) referred Case 2 to a health-care provider (table 2). As before, some patients received antibiotics or steroids even with a referral, so ideal case management was recorded in 372 (62%) of 601 interactions (95% CI 58–66). Antibiotics, steroids, and fluoroquinolones were all used much less frequently, although Schedule H drugs were still given in 188 (31%) of 601 interactions (95% CI 28–35). As before, Schedule X and anti-tuberculosis drugs were never used.

eal case management was recorded in 372 (62%) of 601 interactions (95% CI 58–66). Antibiotics, steroids, and fluoroquinolones were all used much less frequently, although Schedule H drugs were still given in 188 (31%) of 601 interactions (95% CI 28–35). As before, Schedule X and anti-tuberculosis drugs were never used. Figure 1 uses the random-intercept model together with indicator variables for each city to estimate the difference in pharmacy behaviour for the main outcome variables as odds ratios. All these differences were significant and precisely estimated. For instance, the adjusted odds of pharmacies referring a standardised patient with a sputum smear-positive tuberculosis report to a health-care provider without dispensing antibiotics and steroids (ideal case management) was 21·03 (95% CI 12·33–35·86; p<0·0001) for Case 2 relative to Case 1; the odds ratio for antibiotic use was 0·21 (0·15–0·31; p<0·0001) and for fluoroquinolones 0·31 (0·18–0·53; p<0·0001). We also note that of the 497 referrals across the two cases, 301 (60%) were to doctors in the private sector and the remaining 40% were to the public sector (data not shown). In only three instances was the standardised patient referred specifically to a directly observed treatment, short-course (DOTS) centre.

0·0001). We also note that of the 497 referrals across the two cases, 301 (60%) were to doctors in the private sector and the remaining 40% were to the public sector (data not shown). In only three instances was the standardised patient referred specifically to a directly observed treatment, short-course (DOTS) centre. In terms of behaviour conditional on referral, the differences between Case 1 and Case 2 reflect, to a substantial degree, the large increase in referrals for Case 2. Figure 2 shows the proportion of interactions that received antibiotics or steroids, or both, or no drug separated by case and referral decision. Both for Case 1 and Case 2, the use of antibiotics or steroids and the total number of drugs fell when the pharmacist referred the patient (0·75 for Case 1, 95% CI 0·48–1·02 vs 0·38 for Case 2, 0·29–0·46; data not shown). However, conditioning on the decision to refer, the difference in pharmacist behaviour across the two cases was much smaller.

the use of antibiotics or steroids and the total number of drugs fell when the pharmacist referred the patient (0·75 for Case 1, 95% CI 0·48–1·02 vs 0·38 for Case 2, 0·29–0·46; data not shown). However, conditioning on the decision to refer, the difference in pharmacist behaviour across the two cases was much smaller. The practice of pharmacies varied across cities, although caution is warranted in interpreting these results in view of the different sampling methods used (appendix p 3). We noted similar patterns across the three cities of high use of Schedule H drugs, referrals, and ideal case management (figure 3). Two differences worth highlighting are that compared with Mumbai, the use of antibiotics, steroids, fluoroquinolones, and Schedule H1 drugs were all much higher in Patna; and that there was no fluoroquinolone use in Delhi and little use in Mumbai compared with Patna. These differences are robust to adjustment for differences in the standardised patients used across different cities, an analysis that we did by comparing outcomes only among the (smaller) group of standardised patients who were common to two or more cities (appendix pp 13,14).

little use in Mumbai compared with Patna. These differences are robust to adjustment for differences in the standardised patients used across different cities, an analysis that we did by comparing outcomes only among the (smaller) group of standardised patients who were common to two or more cities (appendix pp 13,14). In terms of type of drugs dispensed, for Case 1, pharmacies dispensed 2·09 drugs on average (95% CI 1·99–2·20; figure 4). The most common classes of drugs dispensed were analgesics such as paracetamol and nimesulide, antibiotics, cough syrups, and anti-allergy drugs. Among antibiotics, amoxicillin was the most common, and 61 (10%) of 599 (95% CI 8–13) pharmacies dispensed fluoroquinolones (eg, ciprofloxacin, levofloxacin, ofloxacin), whereas 45 (8%) of 599 gave steroids such as betamethasone and prednisolone (95% CI 6–10). For Case 2, pharmacies dispensed 0·98 drugs on average (95% CI 0·88–1·09). The classes of drugs dispensed for Case 2 were similar to Case 1, although the overall frequencies were much lower. This finding is again consistent with the result that the difference in behaviour between the two cases was driven, to a large extent, by the sharp increase in referrals for Case 2.

ge (95% CI 0·88–1·09). The classes of drugs dispensed for Case 2 were similar to Case 1, although the overall frequencies were much lower. This finding is again consistent with the result that the difference in behaviour between the two cases was driven, to a large extent, by the sharp increase in referrals for Case 2. Discussion To our knowledge, this is the first study that used standardised patients to examine how pharmacies in India treat patients with tuberculosis symptoms and diagnosed tuberculosis, complementing our recent study that assessed tuberculosis management by health-care providers.20 Because the standardised patient method standardises the presentation of the underlying condition across different providers,23 the results are reliable, valid, and comparable across pharmacies. The similar patterns we recorded across the three cities suggest that the results might be generalisable to other urban areas in India.

e standardised patient method standardises the presentation of the underlying condition across different providers,23 the results are reliable, valid, and comparable across pharmacies. The similar patterns we recorded across the three cities suggest that the results might be generalisable to other urban areas in India. A key finding is that none of the pharmacies in our study dispensed first-line anti-tuberculosis drugs. Concerns regarding the use of anti-tuberculosis drugs by pharmacies seem to be unfounded, at least in major cities, and pharmacies are unlikely sources of irrational drug use that contributes to multidrug-resistant tuberculosis. Why pharmacists do not dispense tuberculosis drugs requires further research, but the fact that tuberculosis drugs (unlike antibiotics such as amoxicillin) are considered toxic and that tuberculosis requires long-term treatment might play a part. Proactiveness of the Indian National Tuberculosis Control Program in including tuberculosis drugs under Schedule H1 and the requirement to document tuberculosis drug prescriptions might also have reduced abuse. However, our findings showed that 38% of the pharmacies dispensed antibiotics or steroids to people with tuberculosis symptoms but no test results. The use of fluoroquinolones in 7% and steroids in 5% of interactions is especially worrying because these drugs delay tuberculosis diagnosis.5, 24 Additionally, fluoroquinolones are also an essential part of multidrug-resistant tuberculosis treatment regimens and emerging regimens, so quinolone abuse is a concern.5

ts. The use of fluoroquinolones in 7% and steroids in 5% of interactions is especially worrying because these drugs delay tuberculosis diagnosis.5, 24 Additionally, fluoroquinolones are also an essential part of multidrug-resistant tuberculosis treatment regimens and emerging regimens, so quinolone abuse is a concern.5 The widespread use of antibiotics and steroids for respiratory symptoms also has implications for community-acquired infections more generally. Unnecessary use of fluoroquinolones is a major risk factor for creating highly resistant Gram-negative enteric bacteria (eg, extended spectrum beta-lactamase resistance) that might cause diarrhoeal illness, bacteraemia, and other infections, especially in India.25 The common use of aminopenicillins (eg, amoxicillin) and macrolides (eg, azithromycin) for respiratory symptoms identified in our study might contribute to resistant strains of common respiratory pathogens such as Streptococcus pneumoniae and Haemophilus influenzae.26 In addition to potentially delaying tuberculosis diagnosis, unnecessary use of steroids is associated with an increased risk of developing lower respiratory tract infection, cellulitis, herpes zoster, and candidiasis.27

rains of common respiratory pathogens such as Streptococcus pneumoniae and Haemophilus influenzae.26 In addition to potentially delaying tuberculosis diagnosis, unnecessary use of steroids is associated with an increased risk of developing lower respiratory tract infection, cellulitis, herpes zoster, and candidiasis.27 Our results also clearly show that a first-order problem both in the management of tuberculosis and antimicrobial resistance is the information that patients present to the pharmacist. Confirmed diagnoses discipline what pharmacists do, with sharp increases in ideal management and large decreases in antibiotic use. This dramatic difference suggests that the main challenge faced by pharmacists is confusion about the likely diagnosis, in which case better training regarding tuberculosis symptoms and encouraging early referrals for patient with tuberculosis symptoms might help. Lastly, our study shows the value of the standardised patient method in tracking inappropriate antibiotic use.28 Although prescription audits can be used, prescriptions do not capture the off-prescription use of drugs and often do not include diagnoses.

Our results also clearly show that a first-order problem both in the management of tuberculosis and antimicrobial resistance is the information that patients present to the pharmacist. Confirmed diagnoses discipline what pharmacists do, with sharp increases in ideal management and large decreases in antibiotic use. This dramatic difference suggests that the main challenge faced by pharmacists is confusion about the likely diagnosis, in which case better training regarding tuberculosis symptoms and encouraging early referrals for patient with tuberculosis symptoms might help. Lastly, our study shows the value of the standardised patient method in tracking inappropriate antibiotic use.28 Although prescription audits can be used, prescriptions do not capture the off-prescription use of drugs and often do not include diagnoses. Although the behaviour change in Case 2 suggests that pharmacists substantially decrease the use of unnecessary drugs when the diagnosis is known, it is unknown why some pharmacists give antibiotics and others do not; neither can we uncover the reasons why pharmacists are unwilling to follow regulations regarding drug use in these three cities. It is unclear whether the variation in our data is explained by the competence and qualification of the person providing advice in pharmacies, which we did not track in the study. Qualitative evidence suggests that a combination of other factors might also be at play, including pharmaceutical industry marketing techniques, business models followed by local providers, and active demand from patients for medicines.11, 29 Pharmacists in Delhi have described overstock, near-expiry, and undersupply as further factors precipitating misuse of antibiotics and restricted drugs.11

be at play, including pharmaceutical industry marketing techniques, business models followed by local providers, and active demand from patients for medicines.11, 29 Pharmacists in Delhi have described overstock, near-expiry, and undersupply as further factors precipitating misuse of antibiotics and restricted drugs.11 Second, we noted significantly higher use of antibiotics and quinolones in Patna than in Mumbai pointing to some differences across cities. We are able to rule out that these differences reflect the composition of standardised patients deployed across cities (appendix p 13), but with an effective sample size of only three cities, we cannot explain this variation. Also, our study does not provide evidence on how pharmacists in rural areas manage patients with tuberculosis or tuberculosis symptoms. Third, our study reflects what happens when pharmacists receive a completely unknown patient as opposed to a known, regular client, or a client who returns to the pharmacist after one round of ineffective treatment. We note though that only 5–6% of pharmacists asked the patient to return (if they did not feel better; appendix p 9). Fourth, differences between Case 1 and Case 2 could reflect variation in the standardised patient profile. Because different standardised patients were assigned to the two cases with no crossover, we cannot assess this possibility. Generally, the inclusion of standardised patient characteristics has little effect on estimated coefficients in previous standardised patient studies and our coefficients remain stable when we account for standardised patient sex, height, and weight (appendix pp 13,14).

no crossover, we cannot assess this possibility. Generally, the inclusion of standardised patient characteristics has little effect on estimated coefficients in previous standardised patient studies and our coefficients remain stable when we account for standardised patient sex, height, and weight (appendix pp 13,14). To conclude, our study adds to the growing evidence in India on antibiotic abuse, but also underscores that the use of antibiotics is mediated by drug category and the information that patients present. Although antibiotic use is high and such use can delay diagnosis, none of the pharmacies dispensed anti-tuberculosis drugs and the use of stronger fluoroquinolone antibiotics and heavily restricted drug classes was low. Furthermore, the use of all antibiotics decreased sharply when the patient's diagnosis was revealed to the pharmacists. These findings can inform interventions to engage pharmacies in tuberculosis control and antimicrobial stewardship. Supplementary Material Supplementary appendix

To conclude, our study adds to the growing evidence in India on antibiotic abuse, but also underscores that the use of antibiotics is mediated by drug category and the information that patients present. Although antibiotic use is high and such use can delay diagnosis, none of the pharmacies dispensed anti-tuberculosis drugs and the use of stronger fluoroquinolone antibiotics and heavily restricted drug classes was low. Furthermore, the use of all antibiotics decreased sharply when the patient's diagnosis was revealed to the pharmacists. These findings can inform interventions to engage pharmacies in tuberculosis control and antimicrobial stewardship. Supplementary Material Supplementary appendix Acknowledgments This study is funded by Grand Challenges Canada (grant ID: S5 0373-01), Bill & Melinda Gates Foundation (grant number: OPP1091843), Knowledge for Change Program, World Bank Development Research Group. SS is supported by a fellowship from the Canadian Thoracic Society and is also a senior operations research fellow at Center for Operational Research, The Union (Paris, France). RS is supported by a Fogarty Global Health Equity Scholars Fellowship (NIAID R25 TW009338). JD and BD received funds from the Knowledge for Change Program (The World Bank). MP is a recipient of Tier 1 Canada Research Chair from Canadian Institutes of Health Research. We thank Puneet Dewan, Sarang Deo, Nim Pathy, and Vaibhav Saria for useful input on analysis and interpretation; Rajan Singh, Purshottam, Chinar Singh, Geeta, Devender, Varun Kumar, Anand Kumar, Babloo, and Charu Nanda who supervised and implemented the ISERDD fieldwork; and all the standardised patients for their dedication and hard work. The findings, interpretations, and conclusions expressed here are those of the authors and do not necessarily represent the views of the World Bank, its Executive Directors, or the governments they represent.

sed and implemented the ISERDD fieldwork; and all the standardised patients for their dedication and hard work. The findings, interpretations, and conclusions expressed here are those of the authors and do not necessarily represent the views of the World Bank, its Executive Directors, or the governments they represent. Contributors JD and MP obtained funding and designed the study. JD, AK, SS, VD, and MP developed the standardised patient cases and scripts. AK and RKD collected data and supervised data collection. BD, SS, and RS coded the data. VD, MP, and AK trained the standardised patients. SS, JD, AK and BD analysed the data. SS, JD, BD, AK, RS, AM, and MP interpreted the data. The report was written by SS, JD, BD, MP, AK, and SB, and all authors provided critical review and comments to the revision of the report. Declaration of interests MP serves as a consultant for the Bill & Melinda Gates Foundation. He has no financial conflicts to disclose. All other authors declare no competing interests. Figure 1 Odds ratios for case management outcomes for Case 1 versus Case 2

Contributors JD and MP obtained funding and designed the study. JD, AK, SS, VD, and MP developed the standardised patient cases and scripts. AK and RKD collected data and supervised data collection. BD, SS, and RS coded the data. VD, MP, and AK trained the standardised patients. SS, JD, AK and BD analysed the data. SS, JD, BD, AK, RS, AM, and MP interpreted the data. The report was written by SS, JD, BD, MP, AK, and SB, and all authors provided critical review and comments to the revision of the report. Declaration of interests MP serves as a consultant for the Bill & Melinda Gates Foundation. He has no financial conflicts to disclose. All other authors declare no competing interests. Figure 1 Odds ratios for case management outcomes for Case 1 versus Case 2 Reported odds ratios are from a random-intercepts model using each pharmacy as its own control, with city fixed effects. Odds ratios greater than 1 favour Case 2. Referral is any instance in which the pharmacy staff recommended that the standardised patient seeks further care from a health-care provider. Ideal case management for both cases is defined as a referral without the dispensing of antibiotics or steroids. Schedule H, H1, and X drugs are defined as per the Drugs and Cosmetics Act, 1945, of the Ministry of Health and Family Welfare, Government of India and its amendments. Figure 2 Drug use by referral decisions for two standardised patient cases

Reported odds ratios are from a random-intercepts model using each pharmacy as its own control, with city fixed effects. Odds ratios greater than 1 favour Case 2. Referral is any instance in which the pharmacy staff recommended that the standardised patient seeks further care from a health-care provider. Ideal case management for both cases is defined as a referral without the dispensing of antibiotics or steroids. Schedule H, H1, and X drugs are defined as per the Drugs and Cosmetics Act, 1945, of the Ministry of Health and Family Welfare, Government of India and its amendments. Figure 2 Drug use by referral decisions for two standardised patient cases Each panel describes the use of drugs in each case; the first shows pharmacies that did not refer the standardised patient to another health-care provider (left panel) and the second shows those who did (right panel). Both cases are presented in percentages; the percentages making referral decisions are shown below the case labels in each panel. Percentages indicate the number of interactions within each case-referral category dispensing the indicated types of drugs; percentages may add to more than 100% due to rounding. Figure 3 Management of both Case 1 and Case 2 combined by city

Each panel describes the use of drugs in each case; the first shows pharmacies that did not refer the standardised patient to another health-care provider (left panel) and the second shows those who did (right panel). Both cases are presented in percentages; the percentages making referral decisions are shown below the case labels in each panel. Percentages indicate the number of interactions within each case-referral category dispensing the indicated types of drugs; percentages may add to more than 100% due to rounding. Figure 3 Management of both Case 1 and Case 2 combined by city Referral is any instance in which the pharmacy staff recommended that the standardised patient seek further care from a health-care provider. Ideal case management for both cases is defined as a referral without the dispensing of antibiotics or steroids. Schedule H, H1, and X drugs are defined as per the Drugs and Cosmetics Act, 1945, of the Ministry of Health and Family Welfare, Government of India and its amendments. Figure 4 Active ingredients in drugs given for each case The frequency with which each listed active ingredient was contained in drugs given to standardised patients for each case. The number in brackets is the number of interactions in which that active ingredient was recorded. Table 1 Standardised patient case descriptions

Figure 4 Active ingredients in drugs given for each case The frequency with which each listed active ingredient was contained in drugs given to standardised patients for each case. The number in brackets is the number of interactions in which that active ingredient was recorded. Table 1 Standardised patient case descriptions Case description Presentation of standardised patient Expected case management Case 1 Classic case of presumed tuberculosis with 2–3 weeks of cough and fever and directly seeking care from a chemist or pharmacist Case 1 presents with the opening statement, “Sir, I have cough and fever that is not getting better. Please give me some medicine.” At presentation, this case has had a 2–3 week cough, which occurred more during early morning and night, accompanied by a 2–3 week, on-and-off, low-grade fever. The patient was producing sputum that did not contain any blood. The case would admit to a loss of appetite and to his or her clothes becoming a bit loose if prompted by the chemist. If the chemist asked about taking medicines for this illness, the patient would say no Verbal or written referral to a DOTS centre or a health-care provider without dispensing any antibiotics (including anti-tuberculosis drugs and fluoroquinolones) or steroids Case 2 Chronic cough with a positive sputum smear report for tuberculosis from a government dispensary and directly seeking care from a chemist or pharmacist Case 2 presents with a positive sputum smear result visiting a chemist, presenting with the opening statement, “Sir, I am having cough for nearly a month now and also have fever.” While showing a positive sputum report to the chemist, the patient continues, “I went to the government dispensary and they asked me to get my sputum tested. I have this report. Can you please give me some medicine?” At presentation, this case has had a cough for 1 month and produces sputum without blood, accompanied by a 1 month, on-and-off, low-grade fever, which was more during evening times. Similar to Case 1, the case would admit to a loss of appetite and to his or her clothes becoming a bit loose if prompted by the chemist. If the chemist asked about taking medicines for this illness, the patient would say no Verbal or written referral to a DOTS centre or a health-care provider without dispensing any antibiotics (including anti-tuberculosis drugs and fluoroquinolones) or steroids DOTS=directly observed treatment, short-course.

he chemist. If the chemist asked about taking medicines for this illness, the patient would say no Verbal or written referral to a DOTS centre or a health-care provider without dispensing any antibiotics (including anti-tuberculosis drugs and fluoroquinolones) or steroids DOTS=directly observed treatment, short-course. Table 2 Management of Case 1 and Case 2 for all cities and for Patna and Mumbai only All cities (Delhi, Mumbai, and Patna) Patna and Mumbai only Case 1 Case 2 Case 1 Case 2 Number of interactions 599 601 548 548 Referral 96, 0·16 (0·13–0·19) 401, 0·67 (0·63–0·70) 75, 0·14 (0·11–0·17) 362, 0·66 (0·62–0·70) Ideal case management 80, 0·13 (0·11–0·16) 372, 0·62 (0·58–0·66) 64, 0·12 (0·09–0·14) 335, 0·61 (0·57–0·65) Drugs Number of drugs 2·09 (1·99–2·20) 0·98 (0·88–1·09) 2·07 (1·97–2·18) 0·97 (0·86–1·08) Antibiotic 221, 0·37 (0·33–0·41) 98, 0·16 (0·13–0·19) 200, 0·36 (0·32–0·41) 88, 0·16 (0·13–0·19) Steroid 45, 0·08 (0·05–0·10) 16, 0·03 (0·01–0·04) 37, 0·07 (0·05–0·09) 13, 0·02 (0·01–0·04) Antibiotic or steroid 230, 0·38 (0·34–0·42) 104, 0·17 (0·14–0·20) 208, 0·38 (0·34–0·42) 94, 0·17 (0·14–0·20) Fluoroquinolone 61, 0·10 (0·08–0·13) 23, 0·04 (0·02–0·05) 61, 0·11 (0·08–0·14) 23, 0·04 (0·03–0·06) Schedule H 401, 0·67 (0·63–0·71) 188, 0·31 (0·28–0·35) 367, 0·67 (0·63–0·71) 172, 0·31 (0·27–0·35) Schedule H1 37, 0·06 (0·04–0·08) 19, 0·03 (0·02–0·05) 31, 0·06 (0·04–0·08) 16, 0·03 (0·02–0·04) Schedule X 0 0 0 0 Anti-tuberculosis 0 0 0 0 Data are n, proportion (95% CI) or mean (95% CI).

Introduction With 9·6 million new cases and 1·5 million deaths estimated in 2014, tuberculosis represents a major global health problem and ranks alongside HIV as a leading cause of infectious-disease-related deaths.1 Although global incidence has been falling slowly during the past decade, the number of people affected every year remains daunting. Among the most serious obstacles to successful prevention and treatment of tuberculosis are the inadequate identification of individuals with latent tuberculosis infection who are at highest risk of developing the disease,2 insufficient capacity of health systems to rapidly identify and diagnose all tuberculosis cases (especially those with drug resistance),3 inappropriate management of contacts of infectious cases, long duration of treatment (especially for drug-resistant tuberculosis),4 concurrent infection with HIV, and worldwide spread of Mycobacterium tuberculosis strains that are resistant to the most effective antituberculosis agents. To accelerate global progress in the control of tuberculosis, new drugs and shorter, easily administered regimens are needed to treat all forms of tuberculosis, including multidrug-resistant and extensively drug-resistant tuberculosis.

Introduction With 9·6 million new cases and 1·5 million deaths estimated in 2014, tuberculosis represents a major global health problem and ranks alongside HIV as a leading cause of infectious-disease-related deaths.1 Although global incidence has been falling slowly during the past decade, the number of people affected every year remains daunting. Among the most serious obstacles to successful prevention and treatment of tuberculosis are the inadequate identification of individuals with latent tuberculosis infection who are at highest risk of developing the disease,2 insufficient capacity of health systems to rapidly identify and diagnose all tuberculosis cases (especially those with drug resistance),3 inappropriate management of contacts of infectious cases, long duration of treatment (especially for drug-resistant tuberculosis),4 concurrent infection with HIV, and worldwide spread of Mycobacterium tuberculosis strains that are resistant to the most effective antituberculosis agents. To accelerate global progress in the control of tuberculosis, new drugs and shorter, easily administered regimens are needed to treat all forms of tuberculosis, including multidrug-resistant and extensively drug-resistant tuberculosis. The use of a fourth-generation fluoroquinolone (ie, moxifloxacin or gatifloxacin) to shorten the treatment of drug-susceptible tuberculosis to 4 months has been recently assessed in three separate large trials (OFLOTUB,5 REMoxTB,6 and RIFAQUIN7). Unfortunately, none of these trial findings showed non-inferiority compared with the WHO-recommended 6-month standard regimen for the treatment of tuberculosis.8

rten the treatment of drug-susceptible tuberculosis to 4 months has been recently assessed in three separate large trials (OFLOTUB,5 REMoxTB,6 and RIFAQUIN7). Unfortunately, none of these trial findings showed non-inferiority compared with the WHO-recommended 6-month standard regimen for the treatment of tuberculosis.8 A few new antituberculosis drugs have undergone clinical evaluation over the past decade. These include bedaquiline (a diary quinoline) and delamanid (a nitroimidazole), which have been recently approved by national regulatory authorities and recommended by WHO1 for use in selected patients with multidrug-resistant tuberculosis. Additionally, pretomanid, another nitroimidazole, is under evaluation in short multidrug regimens for the treatment of drug-susceptible and drug-resistant tuberculosis.9 Research in context Evidence before this study

A few new antituberculosis drugs have undergone clinical evaluation over the past decade. These include bedaquiline (a diary quinoline) and delamanid (a nitroimidazole), which have been recently approved by national regulatory authorities and recommended by WHO1 for use in selected patients with multidrug-resistant tuberculosis. Additionally, pretomanid, another nitroimidazole, is under evaluation in short multidrug regimens for the treatment of drug-susceptible and drug-resistant tuberculosis.9 Research in context Evidence before this study The combination of pyrazinamide plus a fourth-generation fluoroquinolone (moxifloxacin or gatifloxacin) is considered essential in novel rifampicin-sparing regimens for the treatment of tuberculosis and in shorter regimens for the treatment of multidrug-resistant tuberculosis. Understanding the background prevalence at population level of resistance to these drugs is critical to assess the feasibility of introducing new and shorter regimens in tuberculosis control programmes and the need for drug-susceptibility testing to accompany the introduction of these new regimens. For the past 20 years, levels of resistance to the most powerful first-line antituberculosis drugs, rifampicin and isoniazid, have been monitored in more than 150 countries worldwide through routine surveillance or ad-hoc population-based surveys. Results of these studies are reported to WHO. Susceptibility testing to fluoroquinolones and pyrazinamide is not routinely performed on all tuberculosis cases as part of drug resistance surveillance efforts. Therefore, population-representative surveillance data on levels of resistance to pyrazinamide and fourth-generation fluoroquinolones (moxifloxacin or gatifloxacin) among all patients with tuberculosis do not exist at present. We searched MEDLINE (1966 to March 20, 2016) and Embase (1980 to March 20, 2016), using the terms “tuberculosis”, “drug”, “resistance”, and “surveillance”, with restriction to English, French, and Spanish results. We also searched the database of the WHO global project on antituberculosis drug resistance surveillance, containing results of all published and unpublished national population-based antituberculosis drug resistance surveys and surveillance conducted worldwide (1994 to March 20, 2016).

sh, French, and Spanish results. We also searched the database of the WHO global project on antituberculosis drug resistance surveillance, containing results of all published and unpublished national population-based antituberculosis drug resistance surveys and surveillance conducted worldwide (1994 to March 20, 2016). Added value of this study This study presents the results of the first population-based surveys investigating levels of resistance to pyrazinamide, ofloxacin, levofloxacin, moxifloxacin, and gatifloxacin among patients with tuberculosis in countries with high burden of tuberculosis and multidrug-resistant tuberculosis. In routine surveillance and patient management, testing for resistance to these drugs is restricted to certain patient groups, such as those with rifampicin resistance or a history of previous treatment for tuberculosis. This dataset therefore provides essential insight into the background proportions of resistance to these drugs at population level and in particular among newly diagnosed tuberculosis cases. Our work offers insight into the feasibility of introducing new tuberculosis treatment regimens and strategies for drug-susceptibility testing in these settings. Implications of all the available evidence

This study presents the results of the first population-based surveys investigating levels of resistance to pyrazinamide, ofloxacin, levofloxacin, moxifloxacin, and gatifloxacin among patients with tuberculosis in countries with high burden of tuberculosis and multidrug-resistant tuberculosis. In routine surveillance and patient management, testing for resistance to these drugs is restricted to certain patient groups, such as those with rifampicin resistance or a history of previous treatment for tuberculosis. This dataset therefore provides essential insight into the background proportions of resistance to these drugs at population level and in particular among newly diagnosed tuberculosis cases. Our work offers insight into the feasibility of introducing new tuberculosis treatment regimens and strategies for drug-susceptibility testing in these settings. Implications of all the available evidence Our results have programmatic implications for both treatment strategies and investment opportunities in the development and implementation of new diagnostic technologies. This work shows that the presence of rifampicin resistance, which currently is easily identified thanks to the wide availability of new rapid molecular technology, should prompt attention to the possibility of the simultaneous presence of resistance to pyrazinamide and, in some settings, the earlier generation fluoroquinolones. At the same time, resistance to the latest generation fluoroquinolones at the clinical breakpoint is still uncommon. This is a crucial finding for the design of standard regimens in different settings, and for guidance to regimen developers and diagnostic manufacturers. Our findings support public health policy makers in prioritisation and introduction of new regimens and algorithms for drug-susceptibility testing, and call for a rethinking of surveillance needs to ensure that more and better data are available to understand levels of resistance to pyrazinamide and fluoroquinolones in different epidemiological settings and patient groups.

oritisation and introduction of new regimens and algorithms for drug-susceptibility testing, and call for a rethinking of surveillance needs to ensure that more and better data are available to understand levels of resistance to pyrazinamide and fluoroquinolones in different epidemiological settings and patient groups. Although individual new drugs are important, there remains a programmatic need for effective and shorter regimens, particularly for drug-resistant tuberculosis.10, 11 Many of the potential new regimens being proposed or tested contain at least one or more of the existing antituberculosis drugs. Notably, the combination of pyrazinamide plus moxifloxacin or gatifloxacin is considered essential in novel rifampicin-sparing regimens for the treatment of tuberculosis9, 12 and in shorter regimens for the treatment of multidrug-resistant tuberculosis.10, 11

at least one or more of the existing antituberculosis drugs. Notably, the combination of pyrazinamide plus moxifloxacin or gatifloxacin is considered essential in novel rifampicin-sparing regimens for the treatment of tuberculosis9, 12 and in shorter regimens for the treatment of multidrug-resistant tuberculosis.10, 11 When resistance to an individual drug within a treatment regimen emerges, a reduction in treatment efficacy is usually the result. Additionally, there is greater potential for generating resistance to the remaining drugs in the regimen. Ideally, a new regimen would be introduced in a population that has little or no pre-existing resistance. Therefore, an understanding of the background prevalence of resistance to all drugs included in new regimens is needed to assess the feasibility of shorter regimens and the need for drug-susceptibility testing to accompany regimen introduction.13 Whereas levels of rifampicin and isoniazid resistance are routinely monitored in most tuberculosis-endemic countries, susceptibility testing to fluoroquinolones and pyrazinamide is not routinely performed as part of drug resistance surveillance.14 Therefore, little population-representative surveillance data about levels of resistance to these drugs are available.

istance are routinely monitored in most tuberculosis-endemic countries, susceptibility testing to fluoroquinolones and pyrazinamide is not routinely performed as part of drug resistance surveillance.14 Therefore, little population-representative surveillance data about levels of resistance to these drugs are available. In this Article, we present the results of the first population-based surveys investigating levels of resistance to pyrazinamide, ofloxacin, levofloxacin, moxifloxacin, and gatifloxacin among patients with tuberculosis in countries with a high burden of tuberculosis and multidrug-resistant tuberculosis. We also investigate levels of cross resistance among fluoroquinolones.

pulation-based surveys investigating levels of resistance to pyrazinamide, ofloxacin, levofloxacin, moxifloxacin, and gatifloxacin among patients with tuberculosis in countries with a high burden of tuberculosis and multidrug-resistant tuberculosis. We also investigate levels of cross resistance among fluoroquinolones. Methods Study design and participants Drug resistance surveys are specially designed studies to measure antituberculosis drug resistance among a representative sample of notified patients with pulmonary tuberculosis. Details about the design of drug resistance surveys are provided elsewhere.14 Data presented in this Article were gathered from isolates from patients with tuberculosis enrolled in national surveys in Azerbaijan (2013),15 Bangladesh (2011),16 Pakistan (2013),17 and in subnational surveys in Minsk city, Belarus (2010)18 and in Gauteng and KwaZulu Natal provinces of South Africa (2014).19 These countries were selected because they represent a variety of programmatic and epidemiological settings. They all have high tuberculosis burden, with tuberculosis incidence varying between 58 per 100 000 and 834 per 100 000 population and proportion of tuberculosis cases with rifampicin resistance ranging from 4·9% to 49·1% (appendix). In all of these countries, uncomplicated tuberculosis is treated in the public health sector using a standardised regimen recommended by WHO.8 In Bangladesh and Pakistan in particular, a substantial number of patients are treated by private practitioners with variable drug combinations and schemes.

to 49·1% (appendix). In all of these countries, uncomplicated tuberculosis is treated in the public health sector using a standardised regimen recommended by WHO.8 In Bangladesh and Pakistan in particular, a substantial number of patients are treated by private practitioners with variable drug combinations and schemes. Procedures Sputum samples collected from patients enrolled in these surveys underwent culture and susceptibility testing to first-line antituberculosis drugs at the National tuberculosis Reference Laboratory using either the LJ proportion method (Azerbaijan, Bangladesh, Pakistan) or MGIT 960 (Becton Dickinson, Sparks, MD, USA; Belarus and South Africa). Isolates were then sent to selected WHO tuberculosis supranational reference laboratories where testing for resistance to pyrazinamide, ofloxacin, levofloxacin, moxifloxacin, and gatifloxacin was performed. Laboratory methods were standardised and all laboratories successfully passed proficiency testing for pyrazinamide and fluoroquinolones before starting the project. Resistance to pyrazinamide was assessed by sequencing for the detection of resistance-conferring mutations in the pncA gene (Rv2043c) and the promoter, located in the Rv2044c–Rv2043c intergenic region. The technologies employed were those already in use at the supranational reference laboratories and included Sanger sequencing using 3730xl (Thermo Fisher Scientific, MA, USA), next generation sequencing using Hiseq 2500 and MiSeq platforms (Illumina, San Diego, CA, USA), and Ion Torrent PGM (Thermo Fisher Scientific), according to manufacturers' instructions. The role of mutations in conferring resistance was assigned on the basis of available scientific literature20, 21 and online databases.22, 23, 24 Additionally, phenotypic susceptibility testing to pyrazinamide was conducted on MGIT 960 at the concentration of 100·0 μg/mL using the MGIT-PZA kit25 on all isolates with previously unreported mutations. The role of these mutations in conferring resistance was assigned according to the observed phenotype.

s.22, 23, 24 Additionally, phenotypic susceptibility testing to pyrazinamide was conducted on MGIT 960 at the concentration of 100·0 μg/mL using the MGIT-PZA kit25 on all isolates with previously unreported mutations. The role of these mutations in conferring resistance was assigned according to the observed phenotype. Susceptibility testing to fluoroquinolones was conducted using the MGIT system and stock solutions of drugs prepared in house. All isolates were tested for susceptibility to ofloxacin at 2·0 μg/mL and moxifloxacin at 0·5 μg/mL. Any isolate found to be resistant to either ofloxacin at 2·0 μg/mL or moxifloxacin at 0·5 μg/mL was subsequently tested for resistance to levofloxacin at 1·5 μg/mL, and moxifloxacin and gatifloxacin at 2·0 μg/mL. Moxifloxacin was tested at two concentrations because 0·5 μg/mL is considered the epidemiological cutoff and 2·0 μg/mL is considered the clinical breakpoint,25 defined as the concentration of an antibiotic that can be achieved in body fluids or target sites during optimal therapy. Isolates found to be susceptible to both ofloxacin at 2·0 μg/mL and moxifloxacin at 0·5 μg/mL were considered to be susceptible to levofloxacin at 1·5 μg/mL, and moxifloxacin and gatifloxacin at 2·0 μg/mL.

as the concentration of an antibiotic that can be achieved in body fluids or target sites during optimal therapy. Isolates found to be susceptible to both ofloxacin at 2·0 μg/mL and moxifloxacin at 0·5 μg/mL were considered to be susceptible to levofloxacin at 1·5 μg/mL, and moxifloxacin and gatifloxacin at 2·0 μg/mL. Statistical analysis We analysed data using version 12.0 of the Stata package, stratified for new cases, previously treated cases, all cases combined, cases of rifampicin-susceptible tuberculosis, and cases of rifampicin-resistant tuberculosis to investigate the association of resistance with treatment history and rifampicin resistance status. Due to the failure to regrow and test all strains, multiple imputation of missing values was performed across all settings for all drugs. Final imputation models for each drug included the following variables: age, sex, history of previous treatment, and rifampicin resistance status. Proportions of resistance within each group and 95% CIs were calculated by logistic regression, specifying robust standard errors to account for the cluster-based survey design in some countries (Bangladesh, Pakistan, South Africa). For drugs without any resistant cases, 95% CIs were calculated using the normal approximation to the binomial. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Statistical analysis We analysed data using version 12.0 of the Stata package, stratified for new cases, previously treated cases, all cases combined, cases of rifampicin-susceptible tuberculosis, and cases of rifampicin-resistant tuberculosis to investigate the association of resistance with treatment history and rifampicin resistance status. Due to the failure to regrow and test all strains, multiple imputation of missing values was performed across all settings for all drugs. Final imputation models for each drug included the following variables: age, sex, history of previous treatment, and rifampicin resistance status. Proportions of resistance within each group and 95% CIs were calculated by logistic regression, specifying robust standard errors to account for the cluster-based survey design in some countries (Bangladesh, Pakistan, South Africa). For drugs without any resistant cases, 95% CIs were calculated using the normal approximation to the binomial. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results In this retrospective study, not all isolates collected during the original surveys could be successfully regrown. The proportions of the original survey strains that could not be retrieved due to poor growth varied across countries: 5% (41/789) in Azerbaijan, 29% (389/1344) in Bangladesh, 10% (23/224) in Belarus, 6% (89/1592) in Pakistan, and 6% (52/929) in Gauteng province and 9% (65/756) in KwaZulu Natal province in South Africa. Individuals whose sample could not be re-grown did not differ from those with viable samples in relation to age, sex, history of treatment or rifampicin resistance status (data not shown).

n Belarus, 6% (89/1592) in Pakistan, and 6% (52/929) in Gauteng province and 9% (65/756) in KwaZulu Natal province in South Africa. Individuals whose sample could not be re-grown did not differ from those with viable samples in relation to age, sex, history of treatment or rifampicin resistance status (data not shown). Proportions of resistance to pyrazinamide by setting, history of previous treatment and rifampicin susceptibility are reported in table 1. A total of 4972 patients were tested to investigate resistance to pyrazinamide in the five countries. Levels of resistance among all tuberculosis cases varied substantially (3·0–42·1% in the surveyed settings) and were lowest in Bangladesh, Pakistan, and South Africa. In all countries and for all patient groups, levels of resistance to pyrazinamide did not statistically differ from the levels of resistance to rifampicin (appendix); the only exception was Pakistan, where pyrazinamide resistance was significantly lower than rifampicin resistance in all patient groups (p<0·0001). Proportions of pyrazinamide resistance were significantly higher among patients with rifampicin-resistant tuberculosis (compared with patients with rifampicin-susceptible tuberculosis), and among patients previously treated for tuberculosis (compared with those never treated for tuberculosis) in almost all settings.

). Proportions of pyrazinamide resistance were significantly higher among patients with rifampicin-resistant tuberculosis (compared with patients with rifampicin-susceptible tuberculosis), and among patients previously treated for tuberculosis (compared with those never treated for tuberculosis) in almost all settings. 5015 patients were tested to investigate resistance to fluoroquinolones. Proportions of resistance to individual fluoroquinolones by setting, history of previous treatment, and rifampicin susceptibility are described in table 2 (ofloxacin), table 3 (levofloxacin), table 4 (moxifloxacin), and table 5 (gatifloxacin). Within a particular country, the proportion of all cases with resistance was similar for three drugs (ofloxacin, levofloxacin, and moxifloxacin 0·5 μg/mL). Comparing among countries, the resistance values for all cases ranged from 1·0% to 16·6% (ofloxacin), 0·5% to 12·4% (levofloxacin), and 0·9% to 14·6% (moxifloxacin 0·5 μg/mL). By contrast, the ranges for resistance among all cases for moxifloxacin 2 μg/mL and gatifloxacin were 0·0–5·1% and 0·0–2·5%, respectively, and thus were both lower and less variable.

nce values for all cases ranged from 1·0% to 16·6% (ofloxacin), 0·5% to 12·4% (levofloxacin), and 0·9% to 14·6% (moxifloxacin 0·5 μg/mL). By contrast, the ranges for resistance among all cases for moxifloxacin 2 μg/mL and gatifloxacin were 0·0–5·1% and 0·0–2·5%, respectively, and thus were both lower and less variable. The proportion of new cases of fluoroquinolone resistance was significantly lower than the proportion of rifampicin resistance cases in all countries except Bangladesh and Pakistan. Higher levels of ofloxacin resistance were found in all settings among rifampicin-resistant tuberculosis cases compared with rifampicin-susceptible tuberculosis cases, but this association was not significant in any setting when compared with the other fluoroquinolones. Finally, we observed higher resistance in retreatment than in new cases, but this finding was statistically significant in only half or fewer of the six settings. Of the 303 isolates that were resistant to ofloxacin in this study, drug-susceptibility testing results for all other fluoroquinolones were available for 282 (93%). Proportions of cross resistance among ofloxacin-resistant isolates are presented in table 6 and were high for levofloxacin (87%), and moxifloxacin when tested at 0·5 μg/mL (72%). However, cross resistance remained very low between ofloxacin and moxifloxacin (when tested at 2·0 μg/mL; 7%), and gatifloxacin (2%).

(93%). Proportions of cross resistance among ofloxacin-resistant isolates are presented in table 6 and were high for levofloxacin (87%), and moxifloxacin when tested at 0·5 μg/mL (72%). However, cross resistance remained very low between ofloxacin and moxifloxacin (when tested at 2·0 μg/mL; 7%), and gatifloxacin (2%). Discussion In this multicountry survey, we present the first population-based data for the prevalence of resistance to pyrazinamide and fluoroquinolones among patients with tuberculosis in high-burden countries. In routine surveillance and patient management, testing for resistance to these drugs is restricted to particular patient groups, such as those with rifampicin resistance or a history of previous tuberculosis treatment. This dataset therefore provides essential insight into background levels of resistance and the feasibility of introducing new tuberculosis treatment regimens and strategies for drug-susceptibility testing in these settings.

roups, such as those with rifampicin resistance or a history of previous tuberculosis treatment. This dataset therefore provides essential insight into background levels of resistance and the feasibility of introducing new tuberculosis treatment regimens and strategies for drug-susceptibility testing in these settings. Pyrazinamide is a crucial component of the most commonly used short-course regimen for the treatment of tuberculosis recommended by WHO worldwide,8 and also of second-line regimens for the treatment of multidrug-resistant tuberculosis.11 In the countries investigated, our study showed no significant difference in the overall levels of resistance to pyrazinamide and rifampicin, with the only exception being Pakistan where pyrazinamide resistance was significantly lower than rifampicin resistance. Additionally, pyrazinamide resistance was significantly associated with rifampicin resistance in all settings, confirming that the vast majority of the burden of pyrazinamide resistance is among patients with rifampicin resistance who can be rapidly identified using existing molecular tests. Nonetheless, for a substantial proportion of patients with rifampicin-resistant tuberculosis (19–63%, based on the findings of this study; table 1), pyrazinamide could still be effective. These findings have two important implications. First, rifampicin-sparing regimens that include pyrazinamide might not be more effective than the current first-line regimen for the treatment of tuberculosis at the population level, given that levels of resistance to rifampicin and pyrazinamide are similar. Second, it is crucial to rapidly distinguish between patients who could benefit from a pyrazinamide-containing regimen and those for whom inclusion of pyrazinamide would not be effective. This distinction requires rapid molecular tests for the diagnosis of pyrazinamide resistance, which do not currently exist.26

milar. Second, it is crucial to rapidly distinguish between patients who could benefit from a pyrazinamide-containing regimen and those for whom inclusion of pyrazinamide would not be effective. This distinction requires rapid molecular tests for the diagnosis of pyrazinamide resistance, which do not currently exist.26 As with pyrazinamide, levels of resistance to ofloxacin were significantly associated with rifampicin resistance in all settings. Levels of resistance to ofloxacin, levofloxacin, and moxifloxacin at 0·5 μg/mL were similar in all settings and significantly lower than those of rifampicin resistance in Azerbaijan and Belarus (despite the high prevalence of multidrug-resistant tuberculosis in both settings). By contrast, in Pakistan (and to a lesser extent in Bangladesh) resistance to fluoroquinolones was higher than rifampicin resistance, an alarming finding. This finding probably results from the extensive use of fluoroquinolones in many parts of Asia for the treatment of not only tuberculosis but also pneumonia and uncomplicated respiratory-tract infections generally.27 In eastern Europe and South Africa, fluoroquinolone resistance is mostly confined to patients with rifampicin resistance, reflecting use of this class of antibiotics for tuberculosis treatment only as second-line therapy. In South Africa, although fluoroquinolones are used for the treatment of pneumonia, this use is primarily in the private health sector which is often not accessible to patients with tuberculosis.

h rifampicin resistance, reflecting use of this class of antibiotics for tuberculosis treatment only as second-line therapy. In South Africa, although fluoroquinolones are used for the treatment of pneumonia, this use is primarily in the private health sector which is often not accessible to patients with tuberculosis. Resistance to the latest generation of fluoroquinolones (moxifloxacin and gatifloxacin at 2·0 μg/mL) was extremely low in all settings, even among patients with rifampicin resistance. This finding can be partly explained by the still-infrequent use of fourth-generation fluoroquinolones in most countries. However, it could also represent an underestimation of the real problem in view of the poor understanding of the association between the critical concentration of susceptibility testing of some fluoroquinolones in the laboratory and patient clinical outcomes. Recent data suggest that the breakpoint of 2·0 μg/mL in liquid media might in fact be too high.28, 29 The finding of extensive cross resistance between ofloxacin, levofloxacin, and moxifloxacin at 0·5 μg/mL (table 6) was expected and in line with the results of genetic studies.30 Importantly, cross resistance is still very limited between ofloxacin at 2·0 μg/mL and either moxifloxacin at 2·0 μg/mL or gatifloxacin, but this may partly be a consequence of the excessively high breakpoints as already discussed. This finding supports the current recommendations of using moxifloxacin or gatifloxacin in the treatment of multidrug-resistant tuberculosis.11

ween ofloxacin at 2·0 μg/mL and either moxifloxacin at 2·0 μg/mL or gatifloxacin, but this may partly be a consequence of the excessively high breakpoints as already discussed. This finding supports the current recommendations of using moxifloxacin or gatifloxacin in the treatment of multidrug-resistant tuberculosis.11 Our study has two main limitations. The first is related to its retrospective nature. Surveys were designed to investigate proportions of multidrug-resistant tuberculosis, not proportions of resistance to pyrazinamide or fluoroquinolones. As a consequence, when very low levels of resistance were detected (particularly to moxifloxacin at 2·0 μg/mL and gatifloxacin), the estimates have wide confidence intervals (Table 4, Table 5). Although higher levels of resistance to most drugs were evident among rifampicin-resistant and previously treated cases, statistically significant differences could not be detected in all settings due to insufficient power. Additionally, a proportion—generally below 10%, but 29% in Bangladesh—of the original surveys strains could not be regrown. Imputation of missing values was performed to address this.14

t and previously treated cases, statistically significant differences could not be detected in all settings due to insufficient power. Additionally, a proportion—generally below 10%, but 29% in Bangladesh—of the original surveys strains could not be regrown. Imputation of missing values was performed to address this.14 A second limitation is related to the laboratory component. Although critical concentrations recommended by WHO25 were used for phenotypic testing, recent evidence suggests that some of these thresholds might not be ideal.20, 28, 29, 30, 31 In particular, the low detected levels of resistance to moxifloxacin and gatifloxacin could be a consequence of excessively high breakpoints. To estimate levels of resistance to pyrazinamide, we largely relied on pncA mutations to avoid problems related to the suboptimal reproducibility of phenotypic testing and uncertainties around the most appropriate critical concentration.20, 25, 31 It is expected that pncA mutations identify around 85–90% of all existing resistance to pyrazinamide32, 33 and consequently our results may slightly underestimate the true burden of resistance. Additionally, although only methods recognised by WHO were used and all laboratories passed proficiency testing before starting the project, a level of variability in the results between laboratories can be expected.

razinamide32, 33 and consequently our results may slightly underestimate the true burden of resistance. Additionally, although only methods recognised by WHO were used and all laboratories passed proficiency testing before starting the project, a level of variability in the results between laboratories can be expected. Our findings show that the presence of rifampicin resistance, which can be easily identified due to the wide availability of new rapid molecular technologies, should draw attention to the possibility of the simultaneous presence of resistance to pyrazinamide and, in some settings, the earlier-generation fluoroquinolones. At the same time, resistance to the latest generation of fluoroquinolones at the clinical breakpoint is still uncommon. These findings are crucial for the design of standard regimens in different settings, guidance to regimen developers and diagnostic manufacturers, and the introduction of existing regimens for the treatment of drug-resistant tuberculosis (eg, such as shorter regimens).11 Choices about prioritisation and introduction of new regimens and algorithms for drug-susceptibility testing must take these data into consideration, and surveillance approaches need to be re-thought so that better data are available to understand levels of resistance to pyrazinamide and fluoroquinolones in different epidemiological settings and patient groups. Without this information, the risk of introducing ineffective regimens that are not curative and might amplify development of drug resistance, including to new agents, remains high. Progress towards drug-susceptibility testing in all cases and rapid development of sequencing technologies for detection of mutations expressing resistance to as many drugs as possible is crucial to optimise treatment and prevent the creation of additional drug resistance.

stance, including to new agents, remains high. Progress towards drug-susceptibility testing in all cases and rapid development of sequencing technologies for detection of mutations expressing resistance to as many drugs as possible is crucial to optimise treatment and prevent the creation of additional drug resistance. Supplementary Material Supplementary appendix Acknowledgments We thank Mohamed Shamim Hossain (National Institute of Diseases of the Chest and Hospital, Dhaka, Bangladesh), Paolo Miotto (IRCCS San Raffaele Scientific Institute, Milan, Italy), Mussarat Ashraf, Samreen Shafiq, and Joveria Farooqi (Department of Pathology and Laboratory Medicine, Aga Khan University, Karachi, Pakistan), Faisal M Khanzada (National Tuberculosis reference Laboratory, National Tuberculosis Control Programme, Islamabad, Pakistan), Ananta Nanoo (National Institute for Communicable Diseases, Sandringham, South Africa), Armand Van Deun (Mycobacteriology Unit, Institute of Tropical Medicine, Antwerp, Belgium), and Charalampos Sismanidis (WHO) for their support to data collection, analysis, and interpretation. The views and opinions expressed in this paper are those of the authors and not necessarily the views and opinions of WHO or of the United States Agency for International Development.

opical Medicine, Antwerp, Belgium), and Charalampos Sismanidis (WHO) for their support to data collection, analysis, and interpretation. The views and opinions expressed in this paper are those of the authors and not necessarily the views and opinions of WHO or of the United States Agency for International Development. Contributors MZ, ASD, DMC, CG, KF, RH, SH, NI, MKi, LR, SR-G, WAW, and KW contributed to the study design. NA, SA, AMC, ADr, MD, ZH, AHusa, AHuss, MKa, MM, SVO, MSe, AS, and ST performed the laboratory testing. MZ and ASD analysed data, which was interpreted by all authors. MZ, ASD, KF, and MCR drafted the report with input from all other authors. All authors have seen and approved the final report. Declaration of interests We declare no competing interests. Table 1 Results of sequencing of the pncA gene (associated with pyrazinamide resistance)

) 103, 39·5% (30·1–48·9) 39, 39·1% (22·9–55·3) 34, 49·1% (32·7–65·5) Resistance in rifampicin resistant vs rifampicin susceptible <0·0001 <0·0001 <0·0001 <0·0001 <0·0001 <0·0001 Resistance in newly vs previously treated 0·004 <0·0001 <0·0001 <0·0001 0·169 <0·0001 Data are number tested, % resistant (95% CI) or p value. Table 2 Ofloxacin 2·0 μg/mL susceptibility testing results Azerbaijan Bangladesh Belarus (Minsk city) Pakistan South Africa (Gauteng) South Africa (Kwazulu Natal) New tuberculosis cases 528, 3·4% (1·9–5·0) 736, 4·4% (2·6–6·2) 141, 7·0% (2·7–11·3) 1301, 11·2% (7·8–14·7) 716, 1·0% (0·1–1·8) 437, 1·0% (0·0–2·1) Previously treated tuberculosis cases 220, 8·6% (5·0–12·3) 196, 9·2% (5·1–13·3) 55, 38·8% (26·2–51·4) 202, 15·1% (10·0–20·3) 153, 0·8% (0·0–2·5) 125, 2·0% (0·0–4·2) All tuberculosis cases 748, 5·0% (3·4–6·6) 933, 5·5% (3·7–7·3) 196, 16·6% (11·4–21·9) 1503, 11·8% (8·7–14·9) 960, 1·0% (0·0–1·8) 675, 1·3% (0·4–2·2) Rifampicin susceptible 618, 0·7% (0·0–1·3) 873, 4·6% (3·0–6·2) 99, 2·9% (0·0–6·3) 1401, 11·1% (7·8–14·3) 919, 0·4% (0·0–0·9) 637, 0·2% (0·0–0·7) Rifampicin resistant 130, 25·0% (17·6–32·4) 60, 16·0% (6·3–25·7) 97, 30·7% (21·5–40·0) 102, 21·8% (13·1–30·5) 41, 12·3% (1·5–23·2) 33, 18·3% (5·7–31·0) Resistance in rifampicin resistant vs rifampicin susceptible <0·0001 0·001 <0·0001 0·009 <0·0001 <0·0001 Resistance in newly vs previously treated 0·004 0·014 <0·0001 0·186 0·868 0·383 Data are number tested, % resistant (95% CI) or p value. Table 3 Levofloxacin 1·5 μg/mL susceptibility testing results

Azerbaijan Bangladesh Belarus (Minsk city) Pakistan South Africa (Gauteng) South Africa (Kwazulu Natal) New tuberculosis cases 528, 3·4% (1·9–5·0) 736, 4·4% (2·6–6·2) 141, 7·0% (2·7–11·3) 1301, 11·2% (7·8–14·7) 716, 1·0% (0·1–1·8) 437, 1·0% (0·0–2·1) Previously treated tuberculosis cases 220, 8·6% (5·0–12·3) 196, 9·2% (5·1–13·3) 55, 38·8% (26·2–51·4) 202, 15·1% (10·0–20·3) 153, 0·8% (0·0–2·5) 125, 2·0% (0·0–4·2) All tuberculosis cases 748, 5·0% (3·4–6·6) 933, 5·5% (3·7–7·3) 196, 16·6% (11·4–21·9) 1503, 11·8% (8·7–14·9) 960, 1·0% (0·0–1·8) 675, 1·3% (0·4–2·2) Rifampicin susceptible 618, 0·7% (0·0–1·3) 873, 4·6% (3·0–6·2) 99, 2·9% (0·0–6·3) 1401, 11·1% (7·8–14·3) 919, 0·4% (0·0–0·9) 637, 0·2% (0·0–0·7) Rifampicin resistant 130, 25·0% (17·6–32·4) 60, 16·0% (6·3–25·7) 97, 30·7% (21·5–40·0) 102, 21·8% (13·1–30·5) 41, 12·3% (1·5–23·2) 33, 18·3% (5·7–31·0) Resistance in rifampicin resistant vs rifampicin susceptible <0·0001 0·001 <0·0001 0·009 <0·0001 <0·0001 Resistance in newly vs previously treated 0·004 0·014 <0·0001 0·186 0·868 0·383 Data are number tested, % resistant (95% CI) or p value. Table 3 Levofloxacin 1·5 μg/mL susceptibility testing results Azerbaijan Bangladesh Belarus (Minsk city) Pakistan South Africa (Gauteng) South Africa (KwaZulu Natal) New tuberculosis cases 527, 2·2% (1·2–4·0) 729, 3·3% (1·7–5·0) 141, 4·7% (1·2–8·1) 1299, 10·1% (6·7–13·4) 705, 0·5% (0·0–1·1) 419, 0·5% (0·0–1·3) Previously treated tuberculosis cases 220, 6·9% (3·6–10·3) 192, 5·4% (2·6–8·3) 55, 30·3% (18·7–41·9) 201, 14·9% (9·8–19·9) 151, 0·5% (0·0–1·2) 121, 0·9% (0·0–2·6) All tuberculosis cases 747, 3·9% (2·5–5·3) 921, 3·8% (2·3–5·3) 196, 12·4% (7·8–17·0) 1500, 10·7% (7·7–13·7) 945, 0·5% (0·0–1·1) 650, 0·6% (0·0–1·3) Rifampicin susceptible 618, 0·5% (0·0–1·1) 866, 3·6% (2·1–5·2) 99, 2·8% (0·0–6·0) 1401, 10·3% (7·1–13·5) 806, 0·2% (0·0–0·9) 620, 0·3% (0·0–1·4) Rifampicin resistant 129, 19·4% (12·5–26·3) 56, 5·5% (0·0–12·0) 97, 22·4% (14·2–30·6) 99, 21·8% (13·1–30·5) 39, 7·0% (0·0–15·1) 30, 7·9% (0·0–17·5) Resistance in rifampicin resistant vs rifampicin susceptible <0·0001 0·537 0·002 0·151 0·040 0·040 Resistance in newly vs previously treated 0·007 0·156 <0·0001 0·100 0·280 0·661 Data are number tested, % resistant (95% CI) or p value.

97, 22·4% (14·2–30·6) 99, 21·8% (13·1–30·5) 39, 7·0% (0·0–15·1) 30, 7·9% (0·0–17·5) Resistance in rifampicin resistant vs rifampicin susceptible <0·0001 0·537 0·002 0·151 0·040 0·040 Resistance in newly vs previously treated 0·007 0·156 <0·0001 0·100 0·280 0·661 Data are number tested, % resistant (95% CI) or p value. Table 4 Moxifloxacin 0·5 μg/mL and 2·0 μg/mL susceptibility testing results

97, 22·4% (14·2–30·6) 99, 21·8% (13·1–30·5) 39, 7·0% (0·0–15·1) 30, 7·9% (0·0–17·5) Resistance in rifampicin resistant vs rifampicin susceptible <0·0001 0·537 0·002 0·151 0·040 0·040 Resistance in newly vs previously treated 0·007 0·156 <0·0001 0·100 0·280 0·661 Data are number tested, % resistant (95% CI) or p value. Table 4 Moxifloxacin 0·5 μg/mL and 2·0 μg/mL susceptibility testing results Azerbaijan Bangladesh Belarus (Minsk city) Pakistan South Africa (Gauteng) South Africa (KwaZulu Natal) 0·5 μg/mL New tuberculosis cases 528, 2·2% (0·9–3·4) 736, 3·8% (2·2–5·4) 141, 6·3% (2·2–10·4) 1301, 7·5% (5·9–9·1) 709, 0·8% (0·0–1·6) 421, 0·6% (0·0–1·4) Previously treated tuberculosis cases 220, 7·0% (3·6–10·4) 196, 7·0% (3·7–10·4) 55, 33·5% (21·7–45·4) 202, 12·1% (7·3–16·8) 152, 0·9% (0·0–2·8) 123, 2·1% (0·0–4·3) All tuberculosis cases 748, 3·6% (2·3–5·0) 933, 4·5% (2·9–6·1) 196, 14·6% (9·6–19·5) 1503, 8·1% (6·7–9·6) 951, 0·9% (0·0–1·7) 654, 1·0% (0·2–1·7) Rifampicin susceptible 618, 0·5% (0·0–1·1) 873, 3·9% (2·4–5·3) 99, 2·7% (0·0–5·9) 1401, 7·7% (6·1–9·3) 910, 0·5% (0·0–1·1) 621, 0·3% (0·0–0·8) Rifampicin resistant 130, 17·9% (11·2–24·5) 60, 12·2% (3·7–20·7) 97, 26·8% (18·0–35·7) 102, 13·8% (6·3–21·4) 41, 8·4% (0·0–18·4) 33, 12·2% (2·2–22·2) Resistance in rifampicin resistant vs rifampicin susceptible <0·0001 0·007 <0·0001 0·075 <0·0001 <0·0001 Resistance in newly vs previously treated 0·002 0·044 <0·0001 0·053 0·919 0·174 2·0 μg/mL New tuberculosis cases 528, 0·4% (0·0–1·0) 732, 0·4% (0·0–1·0) 141, 1·2% (0·0–3·3) 1299, 0·4% (0·0–0·8) 707, 0·5% (0·0–1·4) 420, 0·0% (0·0–11·2) Previously treated tuberculosis cases 220, 0·5% (0·0–1·5) 192, 1·4% (0·0–3·1) 55, 14·0% (4·5–23·5) 202, 1·7% (0·0–4·0) 151, 0·7% (0·0–2·0) 121, 0·0% (0·0–3·0) All tuberculosis cases 748, 0·5% (0·0–1·0) 925, 0·7% (0·0–1·3) 196, 5·1% (1·7–8·6) 1501, 0·5% (0·0–1·1) 948, 0·5% (0·0–1·1) 651, 0·0% (0·0–0·6) Rifampicin susceptible 618, 0·9% (0·0–3·7) 869, 0·4% (0·0–0·9) 99, 4·5% (0·0–13·9) 1401, 0·5% (0·0–1·0) 908, 0·3% (0·0–0·7) 620, 0·0% (0·0–0·6) Rifampicin resistant 130, 2·0% (0·0–5·6) 56, 3·2% (0·0–8·3) 97, 9·2% (3·1–15·3) 100, 1·4% (0·0–4·3) 40, 3·8% (0·0–10·8) 31, 0·0% (0·0–11·2) Resistance in rifampicin resistant vs rifampicin susceptible 0·410 0·051 0·270 0·316 0·008 .. Resistance in newly vs previously treated 0·926 0·247 0·007 0·132 0·653 .. Data are number tested, % resistant (95% CI) or p value.

3·2% (0·0–8·3) 97, 9·2% (3·1–15·3) 100, 1·4% (0·0–4·3) 40, 3·8% (0·0–10·8) 31, 0·0% (0·0–11·2) Resistance in rifampicin resistant vs rifampicin susceptible 0·410 0·051 0·270 0·316 0·008 .. Resistance in newly vs previously treated 0·926 0·247 0·007 0·132 0·653 .. Data are number tested, % resistant (95% CI) or p value. Table 5 Gatifloxacin 2·0 μg/mL susceptibility testing results Azerbaijan Bangladesh Belarus (Minsk city) Pakistan South Africa (Gauteng) South Africa (KwaZulu Natal) New tuberculosis cases 528, 0·6% (0·0–1·4) 729, 0·0% (0·0–0·5) 141, 1·2% (0·0–3·2) 1299, 0·0% (0·0–0·3) 705, 0·0% (0·0–0·5) 419, 0·0% (0·0–0·9) Previously treated tuberculosis cases 220, 0·5% (0·0–1·5) 192, 0·0% (0·0–1·9) 55, 5·4% (0·0–12·4) 201, 0·0% (0·0–1·8) 151, 0·0% (0·0–2·4) 121, 0·0% (0·0–3·0) All tuberculosis cases 748, 0·6% (0·0–1·2) 922, 0·0% (0·0–0·4) 196, 2·5% (0·0–5·2) 1500, 0·0% (0·0–0·2) 945, 0·0% (0·0–0·4) 650, 0·0% (0·0–0·6) Rifampicin susceptible 618, 1·9% (0·0–6·9) 866, 0·0% (0·0–0·4) 99, 3·0% (0·0–7·2) 1401, 0·0% (0·0–0·3) 906, 0·0% (0·0–0·4) 620, 0·0% (0·0–0·6) Rifampicin resistant 130, 3·3% (0·1–6·5) 56, 0·0% (0·0–6·4) 97, 4·0% (0·0–8·5) 99, 0·0% (0·0–3·7) 39, 0·0% (0·0–9·0) 30, 0·0% (0·0–11·6) Resistance in rifampicin resistant vs rifampicin susceptible 0·310 .. 0·270 .. .. .. Resistance in newly vs previously treated 0·818 .. 0·143 .. .. .. Data are number tested, % resistant (95% CI) or p value. Table 6 Cross resistance among fluoroquinolones

Azerbaijan Bangladesh Belarus (Minsk city) Pakistan South Africa (Gauteng) South Africa (KwaZulu Natal) New tuberculosis cases 528, 0·6% (0·0–1·4) 729, 0·0% (0·0–0·5) 141, 1·2% (0·0–3·2) 1299, 0·0% (0·0–0·3) 705, 0·0% (0·0–0·5) 419, 0·0% (0·0–0·9) Previously treated tuberculosis cases 220, 0·5% (0·0–1·5) 192, 0·0% (0·0–1·9) 55, 5·4% (0·0–12·4) 201, 0·0% (0·0–1·8) 151, 0·0% (0·0–2·4) 121, 0·0% (0·0–3·0) All tuberculosis cases 748, 0·6% (0·0–1·2) 922, 0·0% (0·0–0·4) 196, 2·5% (0·0–5·2) 1500, 0·0% (0·0–0·2) 945, 0·0% (0·0–0·4) 650, 0·0% (0·0–0·6) Rifampicin susceptible 618, 1·9% (0·0–6·9) 866, 0·0% (0·0–0·4) 99, 3·0% (0·0–7·2) 1401, 0·0% (0·0–0·3) 906, 0·0% (0·0–0·4) 620, 0·0% (0·0–0·6) Rifampicin resistant 130, 3·3% (0·1–6·5) 56, 0·0% (0·0–6·4) 97, 4·0% (0·0–8·5) 99, 0·0% (0·0–3·7) 39, 0·0% (0·0–9·0) 30, 0·0% (0·0–11·6) Resistance in rifampicin resistant vs rifampicin susceptible 0·310 .. 0·270 .. .. .. Resistance in newly vs previously treated 0·818 .. 0·143 .. .. .. Data are number tested, % resistant (95% CI) or p value. Table 6 Cross resistance among fluoroquinolones No resistant strains No susceptible strains % resistant strains Ofloxacin (2·0 μg/mL) 282 0 100% Levofloxacin (1·5 μg/mL) 245 37 87% Moxifloxacin (0·5 μg/mL) 203 79 72% Moxifloxacin (2·0 μg/mL) 19 263 7% Gatifloxacin (2·0 μg/mL) 7 275 2%

Research in context Evidence before this study Immunisation of pregnant women against influenza has been common practice in industrialised nations since their vulnerability to severe disease and adverse outcomes was recognised. Nevertheless, this practice has not been adopted by low-income countries with constrained resources. When we undertook this trial, maternal immunisation as a strategy to prevent infant illness and avert associated morbidity and mortality was gaining traction. We searched PubMed between April 17, 1996, and March 8, 2016, for clinical trials with the terms “maternal influenza vaccination” and “maternal influenza immunization”. Our search yielded 36 publications. Two publications were of randomised clinical trials done in low-income and middle-income countries that measured the efficacy of maternal influenza vaccination in protection of infants. The first trial, done in Bangladesh, reported 63% efficacy in the reduction of laboratory-confirmed influenza in infants aged up to 24 weeks. Furthermore, infants born to women who received influenza vaccine were less likely to be small for gestational age and had a higher mean birthweight than did those born to women in the control group. The next trial was done in South Africa and reported 48·8% efficacy in infants aged up to 24 weeks; however, the other benefits to infants were not shown. Added value of this study

Immunisation of pregnant women against influenza has been common practice in industrialised nations since their vulnerability to severe disease and adverse outcomes was recognised. Nevertheless, this practice has not been adopted by low-income countries with constrained resources. When we undertook this trial, maternal immunisation as a strategy to prevent infant illness and avert associated morbidity and mortality was gaining traction. We searched PubMed between April 17, 1996, and March 8, 2016, for clinical trials with the terms “maternal influenza vaccination” and “maternal influenza immunization”. Our search yielded 36 publications. Two publications were of randomised clinical trials done in low-income and middle-income countries that measured the efficacy of maternal influenza vaccination in protection of infants. The first trial, done in Bangladesh, reported 63% efficacy in the reduction of laboratory-confirmed influenza in infants aged up to 24 weeks. Furthermore, infants born to women who received influenza vaccine were less likely to be small for gestational age and had a higher mean birthweight than did those born to women in the control group. The next trial was done in South Africa and reported 48·8% efficacy in infants aged up to 24 weeks; however, the other benefits to infants were not shown. Added value of this study Our trial represents the largest evaluation so far of maternal influenza vaccination as a strategy to prevent influenza in the youngest infants. Additionally, it is the first such study to be completed in west Africa, specifically Mali, one of the poorest countries in the world. Demonstrating the efficacy of maternal influenza vaccination in this setting is compelling. Moreover, establishing that efficacy is highest in the first 4 months of life (67·9%) is important as the duration of protection conferred through maternal vaccination and anticipated benefits are assessed. Finally, the absence of an effect on the incidence of low birthweight is consistent with findings shown in South Africa.

r, establishing that efficacy is highest in the first 4 months of life (67·9%) is important as the duration of protection conferred through maternal vaccination and anticipated benefits are assessed. Finally, the absence of an effect on the incidence of low birthweight is consistent with findings shown in South Africa. Implications of all the available evidence Our study unequivocally demonstrates efficacy of maternal influenza vaccination against laboratory-confirmed influenza in infants and mothers. Moreover, there was high acceptability and logistical feasibility. However, our trial and that done in South Africa did not corroborate the previously reported benefits on neonatal outcomes. Moreover, because these trials were not designed to measure an effect on severe, deadly disease, there remains a notable gap when assessing the cost-effectiveness of this intervention. Although the success of maternal tetanus immunisation programmes suggests that implementation of influenza vaccination would also be successful, the related cost would need to be justified by the gains of the benefits afforded so that local policy makers and donors could invest their restricted funds.

his intervention. Although the success of maternal tetanus immunisation programmes suggests that implementation of influenza vaccination would also be successful, the related cost would need to be justified by the gains of the benefits afforded so that local policy makers and donors could invest their restricted funds. Introduction Pregnant women and young infants are at increased risk of developing severe, complicated, and sometimes fatal influenza infection;1, 2, 3, 4 however, no influenza vaccines are approved for infants younger than 6 months.5, 6, 7 Maternal immunisation against influenza is a promising strategy to reduce disease in pregnant women and young infants.8, 9 Trials in Bangladesh10 and South Africa11 showed protection against laboratory-confirmed influenza in infants born to mothers who received trivalent inactivated influenza vaccine, but additional health benefits in those infants (eg, higher birthweight and reduced likelihood of being small for gestational age) have been inconsistent.11, 12 Remaining questions include more precise determination of the duration of protection for infants that can accrue from maternal immunisation,8 and the technical and logistical feasibility and effectiveness of implementation of programmes in resource-limited settings with high to moderate infant mortality rates.13

11, 12 Remaining questions include more precise determination of the duration of protection for infants that can accrue from maternal immunisation,8 and the technical and logistical feasibility and effectiveness of implementation of programmes in resource-limited settings with high to moderate infant mortality rates.13 We aimed to address these questions in the course of a post-licensure clinical trial of the safety, immunogenicity, and efficacy of maternal influenza immunisation for prevention of influenza in infants younger than 6 months in Mali, west Africa—one of the world's least developed countries, with the world's seventh highest infant mortality rate.14, 15 Mali, nevertheless, maintains a vaunted Expanded Program on Immunization (EPI) that includes immunisation of pregnant women with tetanus toxoid and the introduction of five new EPI vaccines for infants since 2005.16, 17, 18 Introduction of additional vaccines for pregnant women, particularly trivalent inactivated influenza vaccine, the composition of which changes annually, would be challenging, but would make use of an existing effective vaccine delivery platform. Methods Study design and participants We did this prospective, active-controlled, observer-blind, randomised phase 4 trial at six referral centres and community health centres in Bamako, Mali. In the year before starting the trial, influenza activity occurred from September to April, with peaks in October and February.

Methods Study design and participants We did this prospective, active-controlled, observer-blind, randomised phase 4 trial at six referral centres and community health centres in Bamako, Mali. In the year before starting the trial, influenza activity occurred from September to April, with peaks in October and February. Third-trimester pregnant women (≥28 weeks' gestation based on last menstrual period, ultrasound, or uterine height) presenting to participating health centres for prenatal care were eligible for inclusion. Participants had to be able to understand and comply with planned study procedures, provide written informed consent before initiation of any study procedures, and intend to reside in the study area until their newborn infants were 6 months old. Participants could not be members of a household that already had a woman who was participating or had participated in this study. Other exclusion criteria were a history of severe reactions following previous immunisation with influenza or meningococcal vaccines; Guillain–Barré syndrome; known allergy or hypersensitivity to eggs, egg proteins, latex, diphtheria toxoid, or any other components of trivalent inactivated influenza vaccine (Vaxigrip) and quadrivalent meningococcal conjugate vaccine (Menactra); known chronic medical disorder that, in the judgment of the investigator, could compromise assessment of the study vaccine or put the participant at risk; known active infection with HIV, hepatitis B virus, or hepatitis C virus; complications with the ongoing pregnancy, including preterm labour (with cervical change), placental abruption, premature rupture of membranes, known major congenital anomaly, or pre-eclampsia; acute illness or an oral temperature greater than or equal to 37·8°C within 72 h of vaccination (resulted in a temporary delay of vaccination); receipt of any other vaccine, excluding tetanus toxoid, within 2 weeks (for inactivated vaccines) or 4 weeks (for live vaccines and meningococcal A conjugate vaccine) before vaccination in this study; receipt of immunoglobulins or any blood products within 30 days before administration of study vaccines; chronic administration of immunosuppressants or other immune-modifying drugs within 90 days before administration of study vaccines; or any disorder that, in the opinion of the investigator, might compromise the wellbeing of the participant or compliance with study procedures, or interfere with the assessment of study vaccines.

dministration of immunosuppressants or other immune-modifying drugs within 90 days before administration of study vaccines; or any disorder that, in the opinion of the investigator, might compromise the wellbeing of the participant or compliance with study procedures, or interfere with the assessment of study vaccines. We additionally excluded women who intended to travel out of the study area in the 40 days after delivery. Enrolment continued until the requisite number of laboratory-confirmed influenza cases was detected in infants born to vaccinated women. Approval for the research was obtained from the University of Maryland, Baltimore Institutional Review Board; the ethics committee of the Faculté de Médecine, Pharmacie et Odonto-Stomatologie of Mali; and the Ministry of Health of Mali. Community sensitisation was achieved through community leaders, health centre representatives and community members who attended community-wide meetings. All participants provided informed consent. If the participant was illiterate, consent was obtained in the presence of a literate witness after listening to the audiotaped version of the consent form in Bambara, the local language.

ealth centre representatives and community members who attended community-wide meetings. All participants provided informed consent. If the participant was illiterate, consent was obtained in the presence of a literate witness after listening to the audiotaped version of the consent form in Bambara, the local language. Randomisation and masking Participants were randomly allocated (1:1), via a computer-generated, centre-specific list with alternate block sizes of six or 12, to receive trivalent inactivated influenza vaccine (Vaxigrip, Sanofi Pasteur, Lyon, France) or quadrivalent meningococcal conjugate vaccine (Menactra, Sanofi Pasteur, Lyon, France). At enrolment, consenting participants were assigned an identification number, which at vaccination was referenced on the randomisation list and the allocated treatment given. The identification numbers for ineligible participants or those who withdrew before vaccination were not reassigned. Study personnel who administered study vaccines and were aware of treatment allocation had no contact with participants after vaccination and were instructed not to reveal the identity of the study vaccines either to participants or to personnel masked to treatment allocation. Although the syringes used to administer the vaccines were different in appearance, participants were instructed to look away from the vaccinator and were unaware of the assigned intervention.

tructed not to reveal the identity of the study vaccines either to participants or to personnel masked to treatment allocation. Although the syringes used to administer the vaccines were different in appearance, participants were instructed to look away from the vaccinator and were unaware of the assigned intervention. Procedures Quadrivalent meningococcal conjugate vaccine, rather than placebo, was given to controls to provide a potential benefit for all participants in this poor, mostly illiterate, vulnerable population of pregnant women. Moreover, that vaccine was unlikely to interfere with the primary outcome of the trial, yet would provide protection against meningococcal disease. Although disease due to serogroup A has largely disappeared from the region, other serogroups continue to cause epidemics in Mali.

vulnerable population of pregnant women. Moreover, that vaccine was unlikely to interfere with the primary outcome of the trial, yet would provide protection against meningococcal disease. Although disease due to serogroup A has largely disappeared from the region, other serogroups continue to cause epidemics in Mali. The composition of trivalent inactivated influenza vaccine, supplied in prefilled syringes, changed during the trial. From September, 2011, to November, 2012, A/California/7/2009(H1N1[pandemic]-like), A/Perth/16/2009(H3N2)-like, and B/Brisbane/60/2008-like (2011 northern hemisphere formulation and then 2012 southern hemisphere formulation) were administered. From December, 2012, to April, 2013, A/California/7/2009 (H1N1[pandemic]-like), A/Victoria/361/2011(H3N2)-like strain, and B/Wisconsin/1/2010-like (2012 northern hemisphere formulation) were administered. Quadrivalent meningococcal conjugate vaccine, composed of 4 μg each of Neisseria meningitidis serogroup A, C, Y, and W-135 polysaccharides conjugated to diphtheria toxoid protein, was supplied in single-dose vials. A single 0·5 mL dose of trivalent inactivated influenza vaccine or quadrivalent meningococcal conjugate vaccine was injected into the deltoid muscle. Study vaccines were stored in secure, temperature-monitored refrigerators or cold rooms at 2–8°C.

jugated to diphtheria toxoid protein, was supplied in single-dose vials. A single 0·5 mL dose of trivalent inactivated influenza vaccine or quadrivalent meningococcal conjugate vaccine was injected into the deltoid muscle. Study vaccines were stored in secure, temperature-monitored refrigerators or cold rooms at 2–8°C. After vaccination, women were observed for 30 min. 7 days after vaccination, field personnel interviewed the women about any local and systemic reactions. 28 days after vaccination, participants were clinically evaluated. Additional visits to evaluate safety and immunogenicity in women and infants were done at delivery and when the infant was 3 months and 6 months old. Each evaluation included a physical examination and blood specimen collection. When available, the infant birth sample was cord blood; otherwise, the birth sample was collected within 7 days after birth. To determine gestational age at birth, the New Ballard Score was measured at delivery or within 7 days after birth.19 Serious adverse events were recorded throughout study participation.

lection. When available, the infant birth sample was cord blood; otherwise, the birth sample was collected within 7 days after birth. To determine gestational age at birth, the New Ballard Score was measured at delivery or within 7 days after birth.19 Serious adverse events were recorded throughout study participation. Besides safety follow-up visits, from enrolment to when the infant reached age 6 months, field personnel undertook weekly visits to detect influenza-like illness and severe acute respiratory infection. During each visit, the participating woman and infant (if already born) had their temperatures measured and were examined for influenza-like illness; women were additionally examined for severe acute respiratory infection. When case definitions for either disease were met (appendix p 5), nasopharyngeal and oropharyngeal swabs and a malaria blood smear were obtained. If influenza was detected by RT-PCR, the case was deemed to be laboratory-confirmed influenza. Standard-of-care treatment was offered.

for severe acute respiratory infection. When case definitions for either disease were met (appendix p 5), nasopharyngeal and oropharyngeal swabs and a malaria blood smear were obtained. If influenza was detected by RT-PCR, the case was deemed to be laboratory-confirmed influenza. Standard-of-care treatment was offered. Because the primary objective was to measure the efficacy of maternal immunisation for prevention of laboratory-confirmed influenza in their infants younger than 6 months, women were withdrawn from weekly surveillance of influenza-like illness following stillbirth, fetal death, infant death, or other events that precluded infant surveillance. Nevertheless, safety follow-up of women continued until 6 months after delivery. Appendix p 6 describes methods for sample collection, RT-PCR to detect influenza virus, virus subtyping, and haemagglutination inhibition antibody measurement.

fetal death, infant death, or other events that precluded infant surveillance. Nevertheless, safety follow-up of women continued until 6 months after delivery. Appendix p 6 describes methods for sample collection, RT-PCR to detect influenza virus, virus subtyping, and haemagglutination inhibition antibody measurement. Outcomes We assessed two coprimary objectives: vaccine efficacy in infants born to women vaccinated any time prepartum (intention-to-treat analysis), and vaccine efficacy in infants born to women vaccinated at least 14 days prepartum (per-protocol analysis). The primary outcome was the occurrence of a first case of laboratory-confirmed influenza by age 6 months. Secondary outcomes were the occurrence of a first case of laboratory-confirmed influenza in women (prepartum and post partum); occurrence of a first case of febrile influenza-like illness by age 6 months in infants; occurrence of a first case of febrile influenza-like illness in women (prepartum and post partum); occurrence of local and systemic reactogenicity after injection, related serious adverse events for the entire follow-up period, and all pregnancy complications; levels of influenza virus antibodies by haemagglutination inhibition before and 4 weeks after vaccination, at delivery, and 3 and 6 months after delivery. Tertiary outcomes included the frequency of each influenza virus type circulating in the study population, the levels of maternally derived influenza virus haemagglutination inhibition antibodies present in infants at birth and at ages 3 and 6 months, birthweights of infants born at a health centre, and the occurrence of severe acute respiratory infection in pregnant women. Appendix p 6 lists additional outcomes not included in the manuscript.

ally derived influenza virus haemagglutination inhibition antibodies present in infants at birth and at ages 3 and 6 months, birthweights of infants born at a health centre, and the occurrence of severe acute respiratory infection in pregnant women. Appendix p 6 lists additional outcomes not included in the manuscript. Statistical analysis We calculated vaccine efficacy with the formula: VE=1-hP(1-P) where VE is vaccine efficacy, h is the ratio of follow-up time up to age 6 months in infants born to recipients of quadrivalent meningococcal vaccine to the follow-up time in infants born to recipients of trivalent inactivated influenza vaccine, and P is the proportion of all cases of laboratory-confirmed influenza occurring by age 6 months in infants whose mothers received trivalent inactivated influenza vaccine. This calculation is equivalent to estimating vaccine efficacy as 1–R, where R is the ratio of laboratory-confirmed influenza incidence rates. We used the ratio of incidence rates, rather than the ratio of proportions of participants who had laboratory-confirmed influenza to account for infants lost to follow-up before age 6 months. We estimated vaccine efficacy in both the intention-to-treat and the per-protocol populations. Only infants' first laboratory-confirmed influenza episodes were counted. Follow-up time was time from birth to first case of laboratory-confirmed influenza, infants reaching age 6 months, or exiting the study. We calculated vaccine efficacy for each month of age (0–5 months) and cumulative to each month of age.

tions. Only infants' first laboratory-confirmed influenza episodes were counted. Follow-up time was time from birth to first case of laboratory-confirmed influenza, infants reaching age 6 months, or exiting the study. We calculated vaccine efficacy for each month of age (0–5 months) and cumulative to each month of age. For safety outcomes, we used Fisher's exact tests and Student's t tests to compare the proportion of participants who had each event per vaccine group. We did time-to-event analysis using Cox proportional hazards regression with laboratory-confirmed influenza as the outcome to establish whether year of vaccination or timing of vaccination relative to delivery had an effect on efficacy. Birthweight analysis was limited to weights that were either 500 g and more or 5000 g and less. We compared birthweight between vaccine groups both overall and within influenza seasons, defined as months with higher-than-average rates of laboratory-confirmed illness (February to April, September to October).

efficacy. Birthweight analysis was limited to weights that were either 500 g and more or 5000 g and less. We compared birthweight between vaccine groups both overall and within influenza seasons, defined as months with higher-than-average rates of laboratory-confirmed illness (February to April, September to October). Sample-size calculations were based on a comparison of the expected proportion, P, of all cases of laboratory-confirmed influenza that occurred by age 6 months in infants whose mothers received trivalent inactivated influenza vaccine to the null value, P0, using exact binomial calculations and assuming equal total follow-up time in each vaccine group (h=1). For the intention-to-treat analysis, we assumed a laboratory-confirmed influenza attack rate of 2·2% by age 6 months in infants born to recipients of quadrivalent meningococcal vaccine and a 55% reduction in the attack rate in infants of recipients of trivalent inactivated influenza vaccine, to 0·99%; therefore, p=0·31034 and P0=0·5. For a one-sided α of 0·025, 77 cases of laboratory-confirmed influenza were needed to ensure 90% power for the intention-to-treat analysis, implying a need for about 4828 participants. Allowing for a 10% loss to follow-up, the sample size calculated became about 5370 participants. For the per-protocol analysis, we assumed vaccine efficacy to be 60%—ie, a laboratory-confirmed influenza attack rate of 0·88% by age 6 months in infants born to recipients of trivalent inactivated influenza vaccine. To ensure 90% power to show a vaccine efficacy of more than 5%, 67 cases of laboratory-confirmed influenza were needed, implying a sample size of 4352 participants. Allowing for a 20% loss to follow-up, or for the mother receiving vaccine less than 14 days before delivery, the sample-size requirement became about 5440 participants.

power to show a vaccine efficacy of more than 5%, 67 cases of laboratory-confirmed influenza were needed, implying a sample size of 4352 participants. Allowing for a 20% loss to follow-up, or for the mother receiving vaccine less than 14 days before delivery, the sample-size requirement became about 5440 participants. Enrolment was closed once 77 cases of infant laboratory-confirmed influenza were recorded, but surveillance continued until the infants reached 6 months of age. A Data Safety Monitoring Board oversaw the study and reviewed data on a regular basis. We did analyses with Stata (version 14.0) and NCSS (version 10). We did power calculations with PASS (version 12). This trial is registered with ClinicalTrials.gov, number NCT01430689. Role of the funding source The funder of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Enrolment was closed once 77 cases of infant laboratory-confirmed influenza were recorded, but surveillance continued until the infants reached 6 months of age. A Data Safety Monitoring Board oversaw the study and reviewed data on a regular basis. We did analyses with Stata (version 14.0) and NCSS (version 10). We did power calculations with PASS (version 12). This trial is registered with ClinicalTrials.gov, number NCT01430689. Role of the funding source The funder of the study had no role in the study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results We did this trial from Sept 12, 2011, to Jan 28, 2014. Between Sept 12, 2011, and April 18, 2013, we randomly assigned 4193 women to receive trivalent inactivated influenza vaccine (n=2108) or quadrivalent meningococcal vaccine (n=2085; figure 1). Baseline characteristics were similar between groups (table 1). One (<1%) woman, who was inadvertently vaccinated twice (once with each vaccine), was followed up as part of her initial assignment group. 4087 (97%) women remained in the study until delivery; 3661 (87%) women were followed up until 6 months after delivery (figure 1). There were 4105 livebirths; 1797 (87%) of 2064 infants in the trivalent inactivated influenza vaccine group and 1793 (88%) of 2041 infants in the quadrivalent meningococcal vaccine group were followed up until age 6 months (figure 1). Due to political upheaval in Mali, study personnel were unable to do household visits for 1 week in March, 2012, and 2 weeks in May, 2012.

s in the trivalent inactivated influenza vaccine group and 1793 (88%) of 2041 infants in the quadrivalent meningococcal vaccine group were followed up until age 6 months (figure 1). Due to political upheaval in Mali, study personnel were unable to do household visits for 1 week in March, 2012, and 2 weeks in May, 2012. We recorded 5279 influenza-like illness episodes in 2789 infants younger than 6 months, of which 131 (2%) episodes were laboratory-confirmed influenza. 129 (98%) cases of laboratory-confirmed influenza were first episodes (n=77 in the quadrivalent meningococcal vaccine group vs n=52 in the trivalent inactivated influenza vaccine group). 116 (90%) first episodes of laboratory-confirmed influenza were in infants of women vaccinated at least 14 days prepartum. The 77 cases needed to complete vaccine efficacy analyses were reached by April 16, 2013; surveillance of post-partum women and their infants continued until infants reached age 6 months. One episode of laboratory-confirmed influenza was associated with malaria parasitaemia.

women vaccinated at least 14 days prepartum. The 77 cases needed to complete vaccine efficacy analyses were reached by April 16, 2013; surveillance of post-partum women and their infants continued until infants reached age 6 months. One episode of laboratory-confirmed influenza was associated with malaria parasitaemia. Overall infant vaccine efficacy was 33·1% (95% CI 3·7–53·9) in the intention-to-treat population, and 37·3% (7·6–57·8) in the per-protocol population (table 2). Notably, cumulative vaccine efficacy in infants in the intention-to-treat population was 67·9% in the first 4 months of follow-up, fell to 57·3% at the fifth month of surveillance, and dropped precipitously in the last month of follow-up, by which time protection was no longer evident (table 2). Cumulative vaccine efficacy in infants in the per-protocol population was 70·2% in the first 4 months of follow-up and 60·7% at the fifth month of surveillance (table 2). Within the trivalent inactivated influenza vaccine group, Cox regression analysis of the relative risk of laboratory-confirmed influenza showed that risk of influenza decreased when trivalent inactivated influenza vaccine had been given at least 15 days prepartum (p=0·02; appendix p 7). As long as the vaccine was given at least 15 days before delivery, no additional benefit was noted in women who had even longer intervals; women in neither the trivalent inactivated influenza vaccine group (Cox regression p=0·90) nor the quadrivalent meningococcal vaccine group (p=0·73) had a significant change in rates of laboratory-confirmed influenza as the time from delivery to vaccination increased above 14 days.

women who had even longer intervals; women in neither the trivalent inactivated influenza vaccine group (Cox regression p=0·90) nor the quadrivalent meningococcal vaccine group (p=0·73) had a significant change in rates of laboratory-confirmed influenza as the time from delivery to vaccination increased above 14 days. 102 first episodes of laboratory-confirmed influenza in infants were due to influenza type A, including 41 H1N1, 59 H3N2, and two non-subtypeable viruses, and to 27 influenza B viruses (table 3). In the first 5 months of life, vaccine efficacy against influenza A was 64·5% overall, and 66·6% for H1N1 and 62·9% for H3N2 (table 3). No vaccine was effective against influenza type B in infants (table 3). In the first 6 months of life, no vaccine was effective against influenza-like illness (1% efficacy, 95% CI −7·0 to 8·5).

rst 5 months of life, vaccine efficacy against influenza A was 64·5% overall, and 66·6% for H1N1 and 62·9% for H3N2 (table 3). No vaccine was effective against influenza type B in infants (table 3). In the first 6 months of life, no vaccine was effective against influenza-like illness (1% efficacy, 95% CI −7·0 to 8·5). During the 29 months of surveillance of infants for laboratory-confirmed illness, the 131 circulating influenza viruses detected by RT-PCR changed over time and included 41 H1N1 viruses, 59 H3N2 viruses, two additional influenza A viruses that could not be subtyped, and 29 influenza B viruses (appendix p 8). We attempted to culture influenza virus from these 131 RT-PCR-positive infant clinical specimens to enable more definitive typing of viruses. Two contaminated samples could not be processed. Of the remaining 129 specimens, 65 grew influenza viruses, which were typed as A/California/7/2009(H1N1[pandemic]-like; n=32) detected from March, 2012, to October, 2013; A/Victoria/361/2011(H3N2; n=11) detected from September, 2012, to October, 2013; and B/Brisbane/03/2007 (Yamagata lineage; n=22) detected from November, 2012, to November, 2013. The H1N1 strains matched those in the trivalent inactivated influenza vaccine formulations; the H3N2 strains matched those in the second formulation and the B strains were of the same lineage as that within the second formulation.

risbane/03/2007 (Yamagata lineage; n=22) detected from November, 2012, to November, 2013. The H1N1 strains matched those in the trivalent inactivated influenza vaccine formulations; the H3N2 strains matched those in the second formulation and the B strains were of the same lineage as that within the second formulation. From Sept 12, 2011, to Jan 28, 2014, we noted 1385 episodes of influenza-like illness in participating women; 52 (4%) episodes were laboratory-confirmed illness. 51 (96%) cases of laboratory-confirmed illness were first episodes. One episode was associated with malaria parasitaemia. Appendix p 9 summarises these cases. There were three episodes of severe acute respiratory infection, all of which were influenza negative. Overall vaccine efficacy was 70·3% (95% CI 42·2–85·8). Efficacy against influenza type A in women was 72·0% and against type B was 73·3% (table 3). Subtype-specific efficacy against H1N1 was 83·3% and could not be measured against H3N2 due to the few cases detected. Vaccine efficacy was 76·6% (95% CI 28·4–94·3) in pregnant women and 70·1% (28·0–89·1) in the post-partum period (appendix p 9). No vaccine was effective against maternal influenza-like illness (2·2% efficacy, 95% CI −10·5 to 13·5).

nst H1N1 was 83·3% and could not be measured against H3N2 due to the few cases detected. Vaccine efficacy was 76·6% (95% CI 28·4–94·3) in pregnant women and 70·1% (28·0–89·1) in the post-partum period (appendix p 9). No vaccine was effective against maternal influenza-like illness (2·2% efficacy, 95% CI −10·5 to 13·5). We measured haemagglutination inhibition antibody titres against influenza A/California/07/09 in 180 mother–infant pairs (figure 2). A subset of 43 pairs (plus one twin) constituted a nested case-control study in which we tested samples from 11 H1N1 cases (including a pair of twins) and 33 birthdate-matched controls (plus or minus 30 days). The remaining 137 pairs included 46 pairs with an infant with laboratory-confirmed illness, and 91 pairs with an infant who did not have laboratory-confirmed illness but might have had influenza-like illness; these pairs represented a convenience sample of participants who completed the study. By age 3 months, infant geometric mean titres (GMT) had decreased by more than 50%, although more infants in the trivalent inactivated influenza vaccine group had haemagglutination inhibition antibody titres of 40 or more (figure 2, appendix pp 10, 11). At age 6 months, haemagglutination inhibition antibody titres of 40 or more did not differ significantly between infants in either vaccine group, although maternal titres remained higher in the trivalent inactivated influenza vaccine group (appendix p 11). Of note, GMT increased in the quadrivalent meningococcal vaccine group at age 6 months, probably due to natural immunity acquired between ages 3 and 6 months. As haemagglutination inhibition antibody titres decreased with age in the trivalent inactivated influenza vaccine group, efficacy also decreased (figure 2).

11). Of note, GMT increased in the quadrivalent meningococcal vaccine group at age 6 months, probably due to natural immunity acquired between ages 3 and 6 months. As haemagglutination inhibition antibody titres decreased with age in the trivalent inactivated influenza vaccine group, efficacy also decreased (figure 2). The most frequently reported local and systemic reactions were pain at the injection site and febrile sensation (appendix p 12). Pain at the injection site was more commonly reported in women given quadrivalent meningococcal vaccine than in those given trivalent inactivated influenza vaccine (n=253 vs n=132; p<0·0001), although reactions were mostly mild (92%; appendix p 12). Rates of unrelated obstetrical and non-obstetrical serious adverse events in women were similar between groups (appendix p 13). The most commonly reported events were hypertensive disorders of pregnancy, which were equally common among both vaccine groups; 1% of participants in each group had pre-eclampsia (p=0·89; appendix p 13). No serious adverse event was related to study treatment. There were five (<1%) unrelated deaths among study participants (n=2 in the trivalent inactivated influenza vaccine group and n=3 in the quadrivalent meningococcal vaccine group); two (40%) patients died due to obstetrical complications and three (60%) patients died after the 42 day period after delivery.

dy treatment. There were five (<1%) unrelated deaths among study participants (n=2 in the trivalent inactivated influenza vaccine group and n=3 in the quadrivalent meningococcal vaccine group); two (40%) patients died due to obstetrical complications and three (60%) patients died after the 42 day period after delivery. Although rates of serious adverse events in infants were similar between groups, presumed neonatal infection was more common in infants in the trivalent inactivated influenza vaccine group than in those in the quadrivalent meningococcal vaccine group (n=60 vs n=37; p=0·02; appendix p 14). No serious adverse events in infants were related to maternal vaccination. 89 infants died: 52 (59%) infants in the trivalent inactivated influenza vaccine group and 37 (41%) infants in the quadrivalent meningococcal vaccine group (p=0·13; appendix p 14); no deaths were due to laboratory-confirmed influenza. Appendix p 15 summarises the timing and causes of death. Per-protocol analysis of the number of infants with a Ballard score less than 33 yielded an overall prematurity rate of 1·8%, which did not correlate with rates measured using date of last menstrual period or results of first-trimester ultrasounds (appendix p 16). 358 (9%) liveborn infants were born at a low birthweight; there was no difference in birthweight between vaccine groups (p=0·20). Furthermore, there was no difference in birthweight among infants born during influenza season.

using date of last menstrual period or results of first-trimester ultrasounds (appendix p 16). 358 (9%) liveborn infants were born at a low birthweight; there was no difference in birthweight between vaccine groups (p=0·20). Furthermore, there was no difference in birthweight among infants born during influenza season. Discussion Here we report results of the largest randomised controlled trial so far of trivalent inactivated influenza vaccine in pregnant women, which was undertaken in Mali, where infant and maternal mortality rates are among the world's highest.14 Trivalent inactivated influenza vaccine elicited robust antibody responses and women and their infants were significantly protected against laboratory-confirmed influenza, corroborating results from Bangladesh (63% vaccine efficacy, 95% CI 5–85) and South Africa (48·8%, 11·6–70·4),10, 11 and supporting WHO recommendations that pregnant women should be the highest priority target for influenza vaccination.20 Because pregnant women and infants are at high risk for severe and fatal influenza illness even in affluent countries, our findings showing efficacy of maternal immunisation in severely resource-constrained Mali, one of the world's least developed countries, constitute encouraging new information.

luenza vaccination.20 Because pregnant women and infants are at high risk for severe and fatal influenza illness even in affluent countries, our findings showing efficacy of maternal immunisation in severely resource-constrained Mali, one of the world's least developed countries, constitute encouraging new information. Maternal immunisation with trivalent inactivated influenza vaccine provided robust protection to infants during the first 4 months of life. Thereafter, as haemagglutination inhibition antibody titres diminished, efficacy decreased and was no longer evident at month 6 of follow-up. These observations support the contention that transplacental maternal antibody protects infants against laboratory-confirmed influenza. Haemagglutination inhibition antibody kinetics resembled those reported in Bangladesh21 and South Africa,22 and align with findings from seroepidemiological studies23, 24 showing that by age 6 months most Malian infants no longer have protective titres of maternally derived measles and Haemophilus influenzae type b antibodies.

utination inhibition antibody kinetics resembled those reported in Bangladesh21 and South Africa,22 and align with findings from seroepidemiological studies23, 24 showing that by age 6 months most Malian infants no longer have protective titres of maternally derived measles and Haemophilus influenzae type b antibodies. Influenza vaccine was well tolerated by pregnant women in our study, corroborating increasing evidence supporting the safety of trivalent inactivated influenza vaccine during pregnancy.25 The Bangladesh trial reported that infants born during the influenza season to women who received influenza vaccine had higher birthweights than did those born to women who received control vaccine during that period.12 In Mali and South Africa there was no beneficial effect of maternal immunisation with trivalent inactivated influenza vaccine on birthweight in infants born anytime during the study,11 including during influenza season. Exclusion of women with high-risk pregnancies from our study and inclusion of women late in pregnancy might have made it difficult to detect differences in birthweight due to maternal disorders. Moreover, infants born to women who had been in the study longer had higher birthweights than did those vaccinated closer to delivery, further decreasing the likelihood of detecting a difference between vaccine groups.

regnancy might have made it difficult to detect differences in birthweight due to maternal disorders. Moreover, infants born to women who had been in the study longer had higher birthweights than did those vaccinated closer to delivery, further decreasing the likelihood of detecting a difference between vaccine groups. While demonstrating the efficacy and safety of maternal influenza immunisation, we were also able to address the technical and logistical feasibility of implementation of such a programme in Mali. The trial was well received by the community as the study team worked at local health centres alongside routine prenatal care (that included the administration of tetanus toxoid) to enrol more than 4000 women. The workflow pattern for administration of study vaccine paralleled that of tetanus toxoid administration and was well accepted by local providers. Nevertheless, remaining aspects, such as the availability of an appropriate vaccine, access to prenatal care, and cost, would affect implementation of a maternal influenza immunisation programme.

r administration of study vaccine paralleled that of tetanus toxoid administration and was well accepted by local providers. Nevertheless, remaining aspects, such as the availability of an appropriate vaccine, access to prenatal care, and cost, would affect implementation of a maternal influenza immunisation programme. We noted seasonal influenza peaks with different influenza viruses from year to year, and these fluctuations affected vaccine efficacy. Notably, vaccinated mothers were significantly protected against influenza B, whereas infants were not. The probable explanation relates to when different B-virus lineages circulated. Cases of type B laboratory-confirmed influenza in infants born to mothers vaccinated with trivalent inactivated influenza vaccine were Yamagata lineage infections, whereas mothers had received B/Brisbane/60/2008 (Victoria lineage) vaccine. By the time infants born to women who received Yamagata lineage-containing vaccine (B/Wisconsin/1/2010) were exposed to influenza, little type B was circulating. This finding shows the complexity of vaccine selection and supports the use of quadrivalent influenza vaccines containing both type-B lineages. Timeliness of importation of newly formulated vaccine, promptness of initiation of vaccination of pregnant women and the types of circulating influenza viruses in relation to vaccine viruses, all affect vaccine efficacy. If maternal vaccination is to succeed in Mali, infants born in September to October will need to be protected. Because northern hemisphere influenza vaccine becomes available in August or September, immunisation will need to be implemented almost immediately upon vaccine importation. The second peak (February to April) does not present this logistical issue. An alternative strategy for countries with an influenza epidemiology similar to Mali is to use vaccine with an extended shelf-life throughout the year.26 This approach would allow immunisation of Malian pregnant women in the months leading up to the September to October peak before the newer formulation is available.

. An alternative strategy for countries with an influenza epidemiology similar to Mali is to use vaccine with an extended shelf-life throughout the year.26 This approach would allow immunisation of Malian pregnant women in the months leading up to the September to October peak before the newer formulation is available. Another factor influencing the overall effect and sustainability of maternal immunisation in countries such as Mali is access to health-care interventions. The 2014 Demographic and Health Survey reported that whereas 95·2% of pregnant women in Bamako and 91·8% in other urban areas had at least one prenatal visit during their most recent pregnancy,27 this was true for only 69·3% of pregnant women in rural Mali. Barriers limiting access to health care in rural areas should be overcome to achieve high maternal immunisation coverage.

s 95·2% of pregnant women in Bamako and 91·8% in other urban areas had at least one prenatal visit during their most recent pregnancy,27 this was true for only 69·3% of pregnant women in rural Mali. Barriers limiting access to health care in rural areas should be overcome to achieve high maternal immunisation coverage. As other similarly low-resourced countries consider the implementation of a maternal influenza vaccination programme, the cost will also affect the feasibility. Since we did not observe an effect of trivalent inactivated influenza vaccine on birthweight, the cost-effectiveness of implementation of the vaccine in pregnancy to prevent infant influenza in Mali will hinge on prevention of severe illness and infant deaths. However, our study was not powered to measure the efficacy of trivalent inactivated influenza vaccine in the prevention of severe laboratory-confirmed influenza. Furthermore, because we visited households of study participants weekly, and intervened when illnesses were encountered by treating and referring participants earlier than they might have sought care in our absence, we probably interrupted progression of illness in many infants. Addressing this gap will require a different trial design and a larger sample size.

of study participants weekly, and intervened when illnesses were encountered by treating and referring participants earlier than they might have sought care in our absence, we probably interrupted progression of illness in many infants. Addressing this gap will require a different trial design and a larger sample size. Our study unequivocally demonstrates efficacy of maternal immunisation against laboratory-confirmed influenza among infants and mothers, and shows high acceptability and logistical feasibility, thereby paving the way for a larger trial to assess prevention of severe laboratory-confirmed influenza leading to hospital admission in infants. Our findings support a vision that, in the future, developing countries might use the maternal immunisation platform to deliver vaccines to prevent respiratory syncytial virus,28 pertussis,29 influenza,13, 26 and tetanus.30 Supplementary Material Supplementary appendix Acknowledgments This study was funded by the Bill & Melinda Gates Foundation (grant OPP1002744). Sanofi Pasteur provided the vaccines for this trial. We thank the study and local health centre personnel and study participants who made this work possible; members of the Data Safety Monitoring Board (George Armah, Amadou Dolo, Glenda Gray, Mamadou Marouf Keita, Katherine O'Brien, Andrew Pollard, Geeta Swami, and Janet Wittes); and Niteen Wairagkar for his guidance and support throughout this project.

th centre personnel and study participants who made this work possible; members of the Data Safety Monitoring Board (George Armah, Amadou Dolo, Glenda Gray, Mamadou Marouf Keita, Katherine O'Brien, Andrew Pollard, Geeta Swami, and Janet Wittes); and Niteen Wairagkar for his guidance and support throughout this project. Contributors MDT, SOS, KLK, WHC, and MML participated in the study design. MDT, SOS, BT, IT, UO, SMT, FC, AT, AMK, FCH, FD, MD, DS, EWO, LAVO, and JT participated in data collection. MK was in charge of vaccine management and vaccination. BT, SMT, MFP, and JT were responsible for laboratory testing. UO, ED, and NHS participated in data cleaning. MDT, SOS, IT, MFP, WCB, NHS, AB, KLK, WHC, EWO, LAVO, JV, JB, JT, and MML participated in data analysis and interpretation. MDT and MML participated in the literature review and primary manuscript writing. All authors contributed to revision of the manuscript. Declaration of interests We declare no competing interests. Figure 1 Trial profile TIIV=trivalent inactivated influenza vaccine. MCV=quadrivalent meningococcal conjugate vaccine. *1886 (91%) infants were born to women vaccinated with TIIV 14 or more days prepartum. †1869 (92%) infants were born to women vaccinated with MCV 14 or more days prepartum. Figure 2 Vaccine efficacy and HAI antibody geometric mean titres in infants, by age and maternal vaccine group Error bars and data in parentheses show 95% CIs. TIIV=trivalent inactivated influenza vaccine. MCV=quadrivalent meningococcal conjugate vaccine. HAI=hemagglutination inhibition antibodies. Table 1 Baseline characteristics

Figure 2 Vaccine efficacy and HAI antibody geometric mean titres in infants, by age and maternal vaccine group Error bars and data in parentheses show 95% CIs. TIIV=trivalent inactivated influenza vaccine. MCV=quadrivalent meningococcal conjugate vaccine. HAI=hemagglutination inhibition antibodies. Table 1 Baseline characteristics TIIV group (n=2108) MCV group (n=2085) Age (years) 24·7 (5·9) 24·7 (6·02) Gravidity 3·2 (2·1) 3·3 (2·1) Parity 2·1 (2·05) 2·1 (2·03) Gestational age at enrolment (weeks) 32·6 (3·6) 32·6 (3·7) Available method to estimate gestational age at enrolment Early ultrasound (<15 weeks) 326 (15%) 322 (15%) Ultrasound after 15 weeks 667 (32%) 641 (31%) Date of last menstrual period 136 (6%) 134 (6%) Uterine height 979 (46%) 988 (47%) Completed HIV testing 716 (34%) 696 (33%) Time from vaccination to delivery (days) 53·7 (28·3) 53·3 (28·0) Delivered at health centre 1966 (93·3%) 1988 (95·3%) Delivery by cesarean section 128 (6%) 126 (6%) Livebirths 2064 (98%) 2041 (98%) Twin birth 36 (2%) 36 (2%) TIIV=trivalent inactivated influenza vaccine. MCV=quadrivalent meningococcal conjugate vaccine. Table 2 Maternal influenza vaccine efficacy against first episodes of laboratory-confirmed influenza in infants younger than 6 months born to women vaccinated at any time prepartum or 14 or more days prepartum

TIIV group (n=2108) MCV group (n=2085) Age (years) 24·7 (5·9) 24·7 (6·02) Gravidity 3·2 (2·1) 3·3 (2·1) Parity 2·1 (2·05) 2·1 (2·03) Gestational age at enrolment (weeks) 32·6 (3·6) 32·6 (3·7) Available method to estimate gestational age at enrolment Early ultrasound (<15 weeks) 326 (15%) 322 (15%) Ultrasound after 15 weeks 667 (32%) 641 (31%) Date of last menstrual period 136 (6%) 134 (6%) Uterine height 979 (46%) 988 (47%) Completed HIV testing 716 (34%) 696 (33%) Time from vaccination to delivery (days) 53·7 (28·3) 53·3 (28·0) Delivered at health centre 1966 (93·3%) 1988 (95·3%) Delivery by cesarean section 128 (6%) 126 (6%) Livebirths 2064 (98%) 2041 (98%) Twin birth 36 (2%) 36 (2%) TIIV=trivalent inactivated influenza vaccine. MCV=quadrivalent meningococcal conjugate vaccine. Table 2 Maternal influenza vaccine efficacy against first episodes of laboratory-confirmed influenza in infants younger than 6 months born to women vaccinated at any time prepartum or 14 or more days prepartum Born to women vaccinated at any time prepartum Born to women vaccinated ≥14 days prepartum TIIV group (n=2064) MCV group (n=2041) TIIV group (n=1886) MCV group (n=1869) By month Cumulative By month Cumulative By month Cumulative By month Cumulative Days of follow-up Cases* Days of follow-up Cases* Days of follow-up Cases* Days of follow-up Cases* Cumulative vaccine efficacy (95% CI) Days of follow-up Cases* Days of follow-up Cases* Days of follow-up Cases* Days of follow- up Cases* Cumulative vaccine efficacy (95% CI) <1 month 61 254 0 (0·0) 61 254 0 (0·0) 60 719 6 (0·10) 60 719 6 (0·10) 100%(15·8 to 100) 55 981 0 (0·00) 55 981 0 (0·00) 55 602 5 (0·09) 55 602 5 (0·09) 100% (−8·4 to 100) 1 month 60 251 2 (0·03) 121 505 2 (0·02) 59 562 3 (0·05) 120 281 9 (0·07) 78·0% (−6·3 to 97·7) 55 181 2 (0·03) 111 162 2 (0·02) 54 637 3 (0·05) 110 239 8 (0·07) 75·2% (−24.2 to 97) 2 months 58 886 4 (0·07) 180 391 6 (0·03) 58 675 10 (0·17) 178 956 19 (0·11) 68·7%(18·4 to 89·8) 53 979 3 (0·06) 165 141 5 (0·03) 53 764 8 (0·15) 164 003 16 (0·10) 69·0%(11·.3 to 91·1) 3 months 57 468 5 (0·09) 2378 59 11 (0·05) 57 017 15 (0·26) 235 973 34 (0·14) 67·9%(35·1 to 85·3) 52 638 4 (0·08) 217 779 9 (0·04) 52 212 14 (0·27) 216 215 30 (0·14) 70·2%(35·7 to 87·6) 4 months 55 600 14 (0·25) 293 459 25 (0·09) 54 913 24 (0·44) 290 886 58 (0·20) 57·3%(30·6 to 74·4) 50 893 12 (0·24) 268 672 21 (0·08) 50 200 23 (0·46) 266 415 53 (0·20) 60·7%(33·8 to 77·5) 5 months 48 485 27 (0·56) 341 944 52 (0·15) 47 608 19 (0·40) 338 494 77 (0·23) 33·1%(3·7 to 53·9) 44 434 24 (0·54) 313 106 45 (0·14) 43 539 18 (0·41) 309 954 71 (0·23) 37·3%(7·6 to 57·8) TIIV=trivalent inactivated influenza vaccine. MCV=quadrivalent meningococcal conjugate vaccine.

266 415 53 (0·20) 60·7%(33·8 to 77·5) 5 months 48 485 27 (0·56) 341 944 52 (0·15) 47 608 19 (0·40) 338 494 77 (0·23) 33·1%(3·7 to 53·9) 44 434 24 (0·54) 313 106 45 (0·14) 43 539 18 (0·41) 309 954 71 (0·23) 37·3%(7·6 to 57·8) TIIV=trivalent inactivated influenza vaccine. MCV=quadrivalent meningococcal conjugate vaccine. * Incidence per 1000 infant-days of follow-up. Table 3 Number of cases of influenza and influenza vaccine efficacy against first episodes of laboratory-confirmed influenza by type in women and their infants up to 5 months of age Women Vaccine efficacy (95% CI) Infants Vaccine efficacy (95% CI) TIIV group (n=2108) Incidence per 1000 person-days of follow-up MCV group (n=2085) Incidence per 1000 person-days of follow-up TIIV group (n=2064) Incidence per 1000 person-days of follow-up MCV group (n=2041) Incidence per 1000 person-days of follow-up Type A 7 0·03 25 0·09 72·0% (35·2 to 87·9) 17 0·06 48 0·17 64·5% (38·3 to 79·6) H3N2 4 0·01 7 0·03 42·8% (−95·4 to 83·3) 10 0·04 27 0·10 62·9% (23·4 to 82·0) H1N1 3 0·01 18 0·06 83·3% (43·4 to 95·1) 7 0·02 21 0·07 66·6% (21·5 to 85·8) Type B 4 0·01 15 0·05 73·3% (19·6 to 91·1) 8 0·03 10 0·04 19·9% (−103·0 to 68·4) TIIV=trivalent inactivated influenza vaccine. MCV=quadrivalent meningococcal conjugate vaccine.

INTRODUCTION Highly pathogenic avian influenza (HPAI) A(H5N1) virus was first isolated and characterised in a domestic goose in Guangdong province, China in 1996,1 and outbreaks have since been reported in domestic poultry, wild birds and humans in over 60 countries.2-4 The spread of HPAI H5N1 in poultry populations increases the risk of human infections.5-8 The first reported case of human illness with H5N1 virus infection occurred in May 1997 in Hong Kong Special Administrative Region (SAR) of China, with a total of 18 cases and 6 deaths.9-12 After an apparent five-year absence, two cases with a history of travel to southern China were reported in February 2003 in Hong Kong SAR.13 Following the pattern of spread and persistence of the virus in poultry, human cases of H5N1 virus infection with high mortality were subsequently detected in China,14,15 Southeast Asia,16,17 West Asia,18,19 and most recently Africa, with cases detected in Egypt every year from 2006 to 2015.20-23 Compared to previous years,24-26 the incidence of human H5N1 cases has remained at a low level globally between October 2012 and October 2014,27,28 while attention has been focused on the more recent emergence of variant swine influenza A(H3N2) in North America,29 novel avian influenza A(H7N9) in China,30-32 other avian influenza A(H5) subtypes in Asia, Europe and North America,27,33 and other emerging infections.34-36

y between October 2012 and October 2014,27,28 while attention has been focused on the more recent emergence of variant swine influenza A(H3N2) in North America,29 novel avian influenza A(H7N9) in China,30-32 other avian influenza A(H5) subtypes in Asia, Europe and North America,27,33 and other emerging infections.34-36 However, between 1 November 2014 and 30 April 2015, a total of 165 cases, including 48 deaths were reported to the World Health Organization (WHO).37 This is by far the highest number of human cases ever reported globally over a similar period.38 Moreover, the number of human H5N1 cases reported in Egypt with onset in February 2015 is the highest number reported by any country in a single month.39 The emergence of a novel cluster of H5N1 virus clade 2.2.1.2 has been found in poultry in Egypt since mid-2014 and has quickly become predominant.40 It is not yet known if this emerging phylotype has increased zoonotic potential and, thus, can be transmitted more efficiently to humans.39-41

country in a single month.39 The emergence of a novel cluster of H5N1 virus clade 2.2.1.2 has been found in poultry in Egypt since mid-2014 and has quickly become predominant.40 It is not yet known if this emerging phylotype has increased zoonotic potential and, thus, can be transmitted more efficiently to humans.39-41 There is a lack of comprehensive epidemiological analysis of global human cases of H5N1 for the 1997-2015 period,17,42-45 and few studies have presented in detail the changing epidemiology of human H5N1 cases in Egypt by comparing the cases before November 2014 to the most recent outbreaks from November 2014 through to April 2015.20,40,46 In order to improve understanding of the epidemiology of HPAI H5N1, we conducted a systematic review of individual case data to describe the magnitude and distribution of all human H5N1 cases globally with illness onset between 1 May 1997 and 30 April 2015, focusing on the characteristics of cases, seasonal and geographical patterns, and examining in more detail the epidemiology of human H5N1 cases in Egypt.

ic review of individual case data to describe the magnitude and distribution of all human H5N1 cases globally with illness onset between 1 May 1997 and 30 April 2015, focusing on the characteristics of cases, seasonal and geographical patterns, and examining in more detail the epidemiology of human H5N1 cases in Egypt. METHODS Search strategy and selection criteria Human H5N1 case data were identified and compiled according to the probable and confirmed case definitions described in the next section. Data on all human H5N1 cases in mainland China were downloaded from the online National Notifiable Infectious Disease Reporting Information System at the Chinese Center for Disease Control and Prevention (China CDC).30,47 Data on human H5N1 cases in Vietnam and Azerbaijan as of 30 April 2014 were provided by the Vietnam National Institute of Hygiene and Epidemiology and the Azerbaijan Ministry of Health, respectively. Data on human H5N1 cases in all other affected countries or regions were obtained from publicly available sources (Appendix Table 1), including the WHO’s Disease Outbreak News of the Global Alert and Response (GAR);48 the WHO’s Weekly Epidemiological Record;49 the WHO Western Pacific Region’s Avian Influenza Weekly Update;50 the FluTrackers website (www.flutrackers.com);51 and the websites of the Ministries of Health in individual countries or regions.

cluding the WHO’s Disease Outbreak News of the Global Alert and Response (GAR);48 the WHO’s Weekly Epidemiological Record;49 the WHO Western Pacific Region’s Avian Influenza Weekly Update;50 the FluTrackers website (www.flutrackers.com);51 and the websites of the Ministries of Health in individual countries or regions. We also searched in PubMed for related studies using a systematic review approach that followed the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines (Figure 1).52 The literature published from May 1, 1997 to April 30, 2015 was searched by the queries “(H5N1[Title] AND (PATIENT[Title] OR PATIENTS[Title] OR HUMAN[Title] OR HUMANS[Title] OR PERSON[Title] OR CASE[Title] OR CASES[Title])) AND ("1997/05/01"[Date - Publication] : "2015/04/30"[Date - Publication])”. Articles published in English and Chinese were included, and the full-text of Chinese articles was searched from China National Knowledge Infrastructure and Wanfang Data. Relevant articles and reports published between 1997 and 2015 were identified through searches in the reports from WHO and the ProMed-mail posts. Articles resulting from these searches and relevant references cited in those articles were reviewed.

searched from China National Knowledge Infrastructure and Wanfang Data. Relevant articles and reports published between 1997 and 2015 were identified through searches in the reports from WHO and the ProMed-mail posts. Articles resulting from these searches and relevant references cited in those articles were reviewed. WHO GAR updates and the WHO statistics on cumulative number of confirmed human H5N1 cases from November 2003 to April 2015 were used to establish a line list of human H5N1 cases.48,53 All cases from sources other than WHO updates were matched with the initial line list (Figure 1). The latest cases, which were not yet officially announced by WHO, were identified through ProMed-mail posts and FluTrackers, and confirmed by the announcements of Ministries of Health in individual countries/regions. When critical information was missing, additional information was sought from published literature, ProMed posts and English language news releases from the regional office of WHO and the relevant Ministry of Health (Appendix Table 1).18,50,54-57

he announcements of Ministries of Health in individual countries/regions. When critical information was missing, additional information was sought from published literature, ProMed posts and English language news releases from the regional office of WHO and the relevant Ministry of Health (Appendix Table 1).18,50,54-57 Case definition The WHO case definition was used.58 A confirmed case was defined as a human case of influenza A(H5N1) virus infection reported by WHO and with laboratory confirmation, i.e. a patient with defined clinical signs, epidemiological linkage and laboratory confirmation by an influenza laboratory accepted by WHO, as specified in the WHO case definition. Other reported cases were considered as probable cases if they had exposure to other confirmed human cases, or to sick or dead poultry, or the H5N1 infection was confirmed by the country or local institutions but not meeting WHO criteria or announced by WHO.

ory accepted by WHO, as specified in the WHO case definition. Other reported cases were considered as probable cases if they had exposure to other confirmed human cases, or to sick or dead poultry, or the H5N1 infection was confirmed by the country or local institutions but not meeting WHO criteria or announced by WHO. Data variables and extraction All probable and confirmed cases with illness onset by 30 April 2015 were included in the analysis. Individual data on cases included age, sex, country, type of diagnosis, year, month and day of onset, date of hospitalization, final outcome (fatal or non-fatal), date of outcome, and potential risk factors (Appendix Table 2). Information on exposure potentially related to the acquisition of H5N1 infections was collected (Box 1). Where contradictory information was found for a given variable, WHO and Ministry of Health data were given priority over journal articles, and journal articles were given priority over other sources of information (Appendix Table 1). The epidemic curves were plotted and the demographic characteristics were summarized by outcome and geographical region.

s found for a given variable, WHO and Ministry of Health data were given priority over journal articles, and journal articles were given priority over other sources of information (Appendix Table 1). The epidemic curves were plotted and the demographic characteristics were summarized by outcome and geographical region. Data on the clade or subclade of H5N1 virus isolated from human cases were also collated from the regular WHO reports.59 In total, 17 reports issued between August 2006 and February 2015 were reviewed, which provided information on H5N1 virus clades circulating and characterized from 1997 to February 2015.59 However, not all individual cases were reported with laboratory results of the clade or subclade, and thus, where this information was not available, the infection was presumed with the clade or subclade of H5N1 virus in the same period and area.17,40,42,60,61 All data used in this study were anonymized. Ethical approval The National Health and Family Planning Commission of China, the Ministry of Health of Vietnam, and the Ministry of Health of Azerbaijan determined that the collection of data from human cases of avian influenza A(H5N1) virus infection was part of the public health investigation of an outbreak and was exempt from institutional review board assessment. All data were supplied and analyzed in an anonymous format, without access to personal identifying information.

ed that the collection of data from human cases of avian influenza A(H5N1) virus infection was part of the public health investigation of an outbreak and was exempt from institutional review board assessment. All data were supplied and analyzed in an anonymous format, without access to personal identifying information. Role of the funding source The sponsor of the study had no role in the study design, data collection, data analysis, data interpretation, writing of the report, or the decision to publish. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. The views expressed are those of the authors and do not necessarily represent the policy of the China CDC or the institutions with which the authors are affiliated. RESULTS A total of 907 human H5N1 cases were reported globally during the 18-year period from May 1, 1997 through April 30, 2015, of which 94.6% were confirmed cases and 5.4% were probable cases (Table 1 and Figure 2). Annual case numbers displayed striking variations, with the highest numbers recorded in 2015 (Figure 3, Appendix Figures 1-2). The total number of cases (213) in 2014-2015 was greater than that (181 cases) of the four years from 2010-2013 (Appendix Figure 3), with the highest monthly number occurring in February 2015 when there were 55 cases in Egypt and one in China.

with the highest numbers recorded in 2015 (Figure 3, Appendix Figures 1-2). The total number of cases (213) in 2014-2015 was greater than that (181 cases) of the four years from 2010-2013 (Appendix Figure 3), with the highest monthly number occurring in February 2015 when there were 55 cases in Egypt and one in China. The overall male-to-female ratio was almost even (1:1.2) from 1997 to 2014, although this pattern was not uniform across regions (Table 1). The median age of cases was 19 years, with an inter quartile range (IQR) from 5 to 32 years, and 41.2% (363/881) were children under 15 years of age and 80.3% (707/881) were in people under 35y. The median age of non-fatal cases was younger in North Africa and older in East and Southeast Asia (median and IQR: 6 and 3-31 years vs. 18.5 and 6-30 years), but the median age of fatal cases was older in North Africa than in East and Southeast Asia (median and IQR: 30 and 20-36 years vs. 19 and 9-30 years) (Figure 4).

he median age of non-fatal cases was younger in North Africa and older in East and Southeast Asia (median and IQR: 6 and 3-31 years vs. 18.5 and 6-30 years), but the median age of fatal cases was older in North Africa than in East and Southeast Asia (median and IQR: 30 and 20-36 years vs. 19 and 9-30 years) (Figure 4). In total, 16 countries reported human cases between 1997 and 2015. The number of affected countries has risen between 2003 and 2008, with expansion from East Asia to Southeast Asia, then West Asia, North Africa and other regions, and apparent ongoing transmission and cases reported almost every year in China, Vietnam, Cambodia, Indonesia and Egypt (Appendix Figure 4A). The incidence in Asia remained at low levels in 2013-2015, while the number of cases in Egypt has increased in 2014-2015. During 1997-2015, 67.2% (594/884) of cases were reported from December to March, with a peak in January (20.9%) (Appendix Figure 5). However, compared to the countries in Southeast Asia and North Africa at lower latitudes, countries in East Asia and West Asia had fewer cases occurring in the warm/hot season from April to September (8.1% vs. 26.2%), and showed earlier peaks (December vs. January) and shorter epidemic periods, with cases occurring year round in Southeast Asia and North Africa, but from January to June and October to December in East Asia, and only from December to March in West Asia(Appendix Figure 4B and 5).

om April to September (8.1% vs. 26.2%), and showed earlier peaks (December vs. January) and shorter epidemic periods, with cases occurring year round in Southeast Asia and North Africa, but from January to June and October to December in East Asia, and only from December to March in West Asia(Appendix Figure 4B and 5). After excluding four cases with unknown outcome (two of Vietnam in 2005 and two of Egypt in 2015), the overall case fatality risk (CFR) was 53.5% (483/903), with a decrease from 70.7% (275/420) in 2003-2008 to 43.4% (202/465) in 2009-2015, and varied across geographical regions, with a CFR (69.4%, 349/503) in East and Southeast Asia more than twice that in North Africa (32.1%, 116/361) (Table 1). The age distribution of cases also differed by outcome, with a median age of 22 years (IQR: 11.5-32 years) for fatal cases and 10 years (IQR: 3-30 years) for cases who recovered (Figure 3C-D). The majority (95.8%, 748/781) of cases reported exposure to poultry including: 85.7% (439/512) exposed to sick or dead poultry; 61.4% (188/306) exposed to backyard poultry; 26.4% (82/311) exposed to LBMs; 4.7% (15/321) occupational exposure to live poultry. In addition, 6.2% (49/792) reported contact with a human H5N1 case before the onset of illness (Table 1, Appendix Table 3). Time from onset of illness to hospitalization was available for 79.7% (723/907) cases with a median of 4 days (IQR: 2-6 days). Generally, the survived cases had an earlier hospitalization than fatal cases (median and IQR: 3 and 1-6 days vs. 5 and 3-7) (Appendix Figure 6). Additionally, the cases in North Africa had a shorter time from onset to hospitalization than cases in East and Southeast Asia (median and IQR: 3 and 1-6 days vs. 5 and 3-7), but the median time from onset to outcome was the same (10 days) between cases in North Africa and cases in East and Southeast Asia.

x Figure 6). Additionally, the cases in North Africa had a shorter time from onset to hospitalization than cases in East and Southeast Asia (median and IQR: 3 and 1-6 days vs. 5 and 3-7), but the median time from onset to outcome was the same (10 days) between cases in North Africa and cases in East and Southeast Asia. The A(H5N1) viruses in human cases have been characterized as belonging to clade or subclade 0, 1, 2.1, 2.1, 2.3, and 7 (Table 1-2, Appendix Table 3). Clade 1 was first reported in Hong Kong SAR in 2003, and then reported in Southeast Asia each year from 2003 to 2014, but subclade 2.1 was only reported in Indonesia since 2005, and subclade 2.2 has circulated in Egypt since 2006 with sporadic reporting in Africa and West Asia. In addition, subclade 2.3 has been reported in East and Southeast Asia since 2005.

AR in 2003, and then reported in Southeast Asia each year from 2003 to 2014, but subclade 2.1 was only reported in Indonesia since 2005, and subclade 2.2 has circulated in Egypt since 2006 with sporadic reporting in Africa and West Asia. In addition, subclade 2.3 has been reported in East and Southeast Asia since 2005. Human cases of H5N1 in Egypt During March 2006 – April 2015, a total of 363 human cases with influenza A(H5N1) virus infection were reported in Egypt with 116 deaths (32%) (Appendix Figure 7), of which more than half (51%) of cases were reported during the 6 months of November 2014 – April 2015 (Appendix Table 4). The male-to-female ratio was not significantly different between cases before November 2014 and cases in the period November 2014 – April 2015, but the latter had an older median age (median and IQR: 26; 4-38 years) than the former (median and IQR: 16; 3.25-30 years), which was also different for both non-fatal and fatal cases (Figure 4E-F). However, the CFR was not significantly different at 36% (64/178) before November 2014 compared to 28.4% (52/183) during November 2014 – April 2015 (Appendix Table 4). For fatal cases, the median time and IQR (5; 3-6 days) for onset to hospital admission was the same between March 2006 – October 2014 and November 2014 – April 2015, but the time was different for non-fatal cases with a median of one day (IQR: 1-3 days) before November 2014 and four days (IQR: 2-6 days) during November 2014 – April 2015. Most cases reported a history of poultry exposure - 96.1% before November 2014 and 90.8% in November 2014 – April 2015.

ovember 2014 – April 2015, but the time was different for non-fatal cases with a median of one day (IQR: 1-3 days) before November 2014 and four days (IQR: 2-6 days) during November 2014 – April 2015. Most cases reported a history of poultry exposure - 96.1% before November 2014 and 90.8% in November 2014 – April 2015. DISCUSSION In this study, a global dataset spanning 18 years was systematically collated to investigate changes in the epidemiological characteristics of human H5N1 cases, and also focused on Egypt, given its unique situation of increasing incidence since November 2014.20,37,46 Our analyses suggest that the geographic extent of human H5N1 cases has expanded from East Asia to Southeast Asia, then to West Asia and North Africa during 2003-2009, which may be related to the global dispersal of the virus via bird migration.62-64 The bird migration network was shown to better reflect the observed viral gene sequence data than other networks and contributes to seasonal H5N1 epidemics in local regions.3,5,7 In addition, previous evidence demonstrated Siberia as a major hub for the dispersal of the virus via bird populations, and Southeast Asia and Africa as areas of local virus persistence and the major sources of genetically and antigenically novel strains.5,7,65,66 Therefore, the increasing range of virus dispersal and outbreaks among birds may also increase the risk of human exposure.3,67 However, some of the apparent geographical dispersal in cases may also be attributed to enhanced clinical and laboratory surveillance capacity in the past 15-20 years.

y novel strains.5,7,65,66 Therefore, the increasing range of virus dispersal and outbreaks among birds may also increase the risk of human exposure.3,67 However, some of the apparent geographical dispersal in cases may also be attributed to enhanced clinical and laboratory surveillance capacity in the past 15-20 years. Human H5N1 infections were found to exhibit seasonality, related to the cooler season from December to March and across diverse climate zones in the Northern Hemisphere (Appendix Figure 4B and 5), which may correlate with the migration patterns of wild birds and the activity of virus in winter or cooler seasons.3,7,43 A recent study found that the timing of H5N1 outbreaks and viral migrations were closely associated with bird migration networks in Asia.5 In addition, the lower temperatures in Asia and North Africa across diverse climates were associated with increasing human H5N1 virus infection in winter, which is consistent with increased poultry outbreaks and H5N1 virus transmission during cold and dry conditions, and also overlapped with human seasonal influenza epidemics.3,43,68,69

ower temperatures in Asia and North Africa across diverse climates were associated with increasing human H5N1 virus infection in winter, which is consistent with increased poultry outbreaks and H5N1 virus transmission during cold and dry conditions, and also overlapped with human seasonal influenza epidemics.3,43,68,69 Although most human populations are thought to have little or no immunity to influenza A(H5N1) viruses, most cases examined in this study were children and younger adults, and these age groups were also more likely to recover, whereas the fatality risk was higher in adults, which might be related to the immunological reaction of virus in different age groups.41 Consistent with previous reports,28,45 the cases with ≥3 days from onset of illness to hospitalization were more likely to be fatal than those admitted within 3 days of onset with a odds ratios (OR) of 3.6 and 95% confidence intervals (CI) of 2.5 - 5.1, which might be due to the early administration of antiviral treatment, or selection bias where the cases admitted later after onset were more likely to be severe.17 Compared with Indonesia, Vietnam, Cambodia, mainland China and Thailand , the lower CFR in Egypt (Chi-squared tests, p<0.001) might be related to a less virulent virus clade, less severe clinical disease, and earlier identification, hospitalization and early treatment with oseltamivir for H5N1 cases.20,44,70 However, the CFR might be underestimated because various government entities or reports may not have identified or updated which cases have died at the time we collated data. Additionally, almost all human cases of H5N1 infection had a recent history of close contact with infected live or dead birds, other human cases, or H5N1-contaminated environments, which reaffirmed reports that human H5N1 virus infection is typically preceded by exposure to sick or dead poultry in backyards, LBMs or farms.71-76

lly, almost all human cases of H5N1 infection had a recent history of close contact with infected live or dead birds, other human cases, or H5N1-contaminated environments, which reaffirmed reports that human H5N1 virus infection is typically preceded by exposure to sick or dead poultry in backyards, LBMs or farms.71-76 An increased number of animal-to-human infections has been reported by Egypt during November 2014 – April 2015 with the number of cases more than the total of the last 8 years from 2006-2014.20 The increase in the number of human cases in Egypt since November 2014 can be attributed to a mixture of factors, including increased circulation of H5N1 viruses in poultry, lower public health awareness of risks in middle and upper Egypt and seasonal factors, such as closer proximity to poultry because of cold weather and possible longer survival of the viruses in the environment.77 However, the increased numbers of human cases in Egypt are of major concern because of the continued potential pandemic threat from H5N1. A few cases of human-to-human transmission and family clusters have been reported in Egypt and other countries.40,46,78-82 Nevertheless, H5N1 viruses do not currently appear to transmit easily from person to person, and the risk of community level spread of these viruses remains low.20,27,39

pandemic threat from H5N1. A few cases of human-to-human transmission and family clusters have been reported in Egypt and other countries.40,46,78-82 Nevertheless, H5N1 viruses do not currently appear to transmit easily from person to person, and the risk of community level spread of these viruses remains low.20,27,39 H5N1 viruses have evolved from the 1996 progenitor strain and now comprise at least 10 clades, through a complexity of genetic changes, which have infected domestic poultry and wild birds in many countries.21,62,63,83,84 In this study, 4 clades (0, 1, 2, and 7) and 3 subclades (2.1, 2.2, and 2.3) of H5N1 virus strains have infected humans, all of which have been reported in human cases before 2006.41,85 Compared to clade 0, the cases with clade 1, subclade 2.1 and 2.3 were more likely to result in death with a crude OR of 2.8 (95% CI: 0.93, 9.6), 11.0 (95%CI: 3.5, 37.8) and 3.2 (95%CI: 1.0, 11.4) respectively (Appendix Table 3). However, the risk of death between cases with clade 0 and subclade 2.1 was not significantly different (OR: 1.0; 95% CI: 0.3, 3.3). Based on available information, the clades of viruses isolated from humans were the same as the clades circulating in local poultry.21,28 During the period from late 2003 to mid-2005, most H5N1 virus infections in humans were caused by clade 1 strains in Southeast Asia (i.e., Vietnam, Thailand, and Cambodia).85

, 3.3). Based on available information, the clades of viruses isolated from humans were the same as the clades circulating in local poultry.21,28 During the period from late 2003 to mid-2005, most H5N1 virus infections in humans were caused by clade 1 strains in Southeast Asia (i.e., Vietnam, Thailand, and Cambodia).85 Although the highly pathogenic H5N1 virus strains can be expected to continue evolving over time, preliminary laboratory investigation has not detected major genetic changes in the viruses isolated from the patients or animals in 2014-2015 compared to previously circulating isolates in the same regions,41,86 and the genetic diversity of the H5N1 virus decreased significantly between 1996 and 2011 in China, presumably under strong selective pressure associated with vaccination in poultry.56 However, other influenza A(H5) subtypes, such as influenza A(H5N2), A(H5N3), A(H5N6) and A(H5N8), have recently been detected in birds in Europe, North America, and Asia, and so far no human cases of infection have been reported, with the exception of three human infections with influenza A(H5N6) virus detected in China in 2014-15.39,77 However, the co-circulation of influenza A viruses in human and animal reservoirs can provide opportunities for these viruses to re-assort and acquire the genetic characteristics that facilitate sustained human-to-human transmission, a necessary trait of pandemic viruses.3,87

6) virus detected in China in 2014-15.39,77 However, the co-circulation of influenza A viruses in human and animal reservoirs can provide opportunities for these viruses to re-assort and acquire the genetic characteristics that facilitate sustained human-to-human transmission, a necessary trait of pandemic viruses.3,87 Vaccines and antivirals are the most effective approaches to prevent influenza virus infection and treat illness respectively.41,88,89 Vaccination of poultry has been implemented in many of the affected countries to control H5N1 in poultry, especially in those locations where H5N1 viruses have become enzootic in poultry and wild birds.90-92 Currently, 27 A(H5N1) candidate vaccine viruses for humans are available and a new candidate vaccine is in preparation to protect against the currently circulating H5 clade 2.2.1.2 of viruses, covering all the recent H5N1 virus isolates from Egypt.41,93 The first adjuvant vaccine for the prevention of H5N1 influenza has been approved by the United States Food and Drug Administration in November 2013, and this vaccine is being stockpiled for pandemic preparedness by the United States government.94 In addition, the antiviral drug oseltamivir can reduce the severity of illness and mortality when started soon after symptom onset and appears to benefit all age groups. Prompt diagnosis and early therapeutic intervention should therefore be considered for all H5N1 cases,89,95,96 though antiviral resistance continues to receive attention and there is a need for continued monitoring.97 The availability of antivirals and vaccines in the event of a H5N1 pandemic should be considered in advance.98

iagnosis and early therapeutic intervention should therefore be considered for all H5N1 cases,89,95,96 though antiviral resistance continues to receive attention and there is a need for continued monitoring.97 The availability of antivirals and vaccines in the event of a H5N1 pandemic should be considered in advance.98 There are some limitations to this study. First, the data used were collated from different sources. The data quality may be influenced by key steps in public health surveillance or reports including the case definitions, reporting methods, availability of health care and laboratory diagnostics, under reporting, and the completeness and accuracy of data reported or announced by different countries or organizations. Compared to the areas where many cases were seen in this study, some countries with few cases or without cases reported might be attributed to the low availability and capability of public health services, serological testing, and surveillance. Second, detailed data on case characteristics and clinical management were unavailable to assess the association between clinical manifestation, treatment and outcome, and this study did not include the cases with subclinical H5N1 virus infection, which have been occasionally reported.72,99-101 Third, the findings might be influenced by missing data on exposure, outcome and hospitalization, and the misclassification of cases with presumed clade or subclade. In addition, this study only included data sources in English or Chinese, which might neglect data on cases reported in other languages, including announcements or reports from Egypt.

influenced by missing data on exposure, outcome and hospitalization, and the misclassification of cases with presumed clade or subclade. In addition, this study only included data sources in English or Chinese, which might neglect data on cases reported in other languages, including announcements or reports from Egypt. In conclusion, the high-risk areas, population groups and seasonality of human HPAI H5N1 infections have been systematically reviewed here, providing evidence for the planning of prevention and control. The geographic distribution of countries with human H5N1 infections has expanded, especially between 2003 and 2008, with variations in outcome, demography, seasonality and the clade or subclade of viruses across the region. The incidence of human infections increased dramatically in Egypt from November 2014 to April 2015, but remained at a low level in other regions, and the CFR in Egypt has not significantly changed. However, since avian influenza A(H5N1) viruses present a continuous threat to human populations, echoing the recommendations of WHO and other organizations on influenza at the human-animal interface,41,89,102-104 there is a need for sustained efforts and close collaboration between the animal health and public heath sectors at community, national, and international levels to monitor the dynamics in human, poultry and wild birds, and to conduct early clinical management, while downstream research is encouraged to develop vaccines and antivirals, explore the driving factors behind the epidemic, and evaluate the potential for future pandemics.

community, national, and international levels to monitor the dynamics in human, poultry and wild birds, and to conduct early clinical management, while downstream research is encouraged to develop vaccines and antivirals, explore the driving factors behind the epidemic, and evaluate the potential for future pandemics. Supplementary Material 1 Appendix Table 1 The data source of human case with H5N1 virus infection in each country, May 1997 – April 2015. Appendix Table 2 The list of variables in the individual dataset of human case with H5N1 virus infection, May 1997 – April 2015. Appendix Table 3 Demographic and Epidemiologic characteristics of human case with H5N1 virus infection by outcomes, May 1997 – April 2015. Appendix Table 4 The characteristics of human case with H5N1 virus infection in Egypt before and since 1 November 2014. Appendix Figure 1 Epidemic curve of human cases with H5N1 virus infection by climate zones, May 1997 – April 2015. Appendix Figure 2 The number of human cases with H5N1 virus infection by year and geographic region, May 1997 – April 2015 (N=907). Appendix Figure 3 The number of human cases with H5N1 virus infection by year and country, May 1997 – April 2015 (N=907). Appendix Figure 4 Heat map of the reported data of human cases with H5N1 virus infection by country, sorted by geographical region and the date of the first cases illness onset, May 1997–April 2015. Appendix Figure 5 The seasonality of human cases with H5N1 virus infection by the month of illness onset, May 1997 – April 2015.

Appendix Figure 4 Heat map of the reported data of human cases with H5N1 virus infection by country, sorted by geographical region and the date of the first cases illness onset, May 1997–April 2015. Appendix Figure 5 The seasonality of human cases with H5N1 virus infection by the month of illness onset, May 1997 – April 2015. Appendix Figure 6 The distribution of days from onset to hospital admission of human H5N1 cases by outcome and geographic region, May 1997–April 2015. Appendix Figure 7 The geographic distribution of human cases with H5N1 virus infection by outcome in Egypt, March 2006–April 2015 (n=363). We thank staff members at county-, district-, prefecture-, and provincial- level CDCs at the provinces with human H5N1 cases occurred for providing assistance with field investigation, administration and data collection in China. We thank staff members from WHO, the Ministry of Health of Azerbaijan, and the Vietnam National Institute of Hygiene and Epidemiology for assistance in coordination of data collection. Funding

We thank staff members at county-, district-, prefecture-, and provincial- level CDCs at the provinces with human H5N1 cases occurred for providing assistance with field investigation, administration and data collection in China. We thank staff members from WHO, the Ministry of Health of Azerbaijan, and the Vietnam National Institute of Hygiene and Epidemiology for assistance in coordination of data collection. Funding This study was supported by the grants from the National Science Fund for Distinguished Young Scholars (No. 81525023), the US National Institutes of Health (Comprehensive International Program for Research on AIDS grant U19 AI51915); the Ministry of Science and Technology of China (2012 ZX10004-201, 2014BAI13B05); the Ministry of Health of China (201202006); China CDC’s Key Laboratory of Surveillance and Early-warning on Infectious Disease; the Harvard Center for Communicable Disease Dynamics from the National Institute of General Medical Sciences (grant no. U54 GM088558); a commissioned grant from the Health and Medical Research Fund of the Health, Welfare and Food Bureau of the Hong Kong SAR Government; and the University Grants Committee of the Hong Kong SAR of China (project no. T11-705/14N). SL is supported by the Flowminder Foundation and the University of Hong Kong for his postgraduate research on population movement and infectious disease dynamics in the University of Southampton. PWH is funded by the Wellcome Trust (grants 089276/Z/09/Z and 089276/B/09/Z) and the EU FP7 project PREPARE (602525). AJT is supported by funding from NIH/NIAID (U19AI089674), the Bill & Melinda Gates Foundation (OPP1106427, 1032350) and the RAPIDD program of the Science and Technology Directorate, Department of Homeland Security, and the Fogarty International Center, National Institutes of Health. NAW is supported by the UK Medical Research Council (MR/J012343/1).

m NIH/NIAID (U19AI089674), the Bill & Melinda Gates Foundation (OPP1106427, 1032350) and the RAPIDD program of the Science and Technology Directorate, Department of Homeland Security, and the Fogarty International Center, National Institutes of Health. NAW is supported by the UK Medical Research Council (MR/J012343/1). Contributors HY conceived, designed, and supervised the study. SL and YQ designed the study, collected data, finalized the analysis, wrote the drafts of the manuscript, and interpreted the findings. XR participated in the analysis and mapped the case geographic distribution. SL, TKT, LF, HJ, ZP, JZ, QL participated in the literature search, data collection and analysis. BJC, NAW, MG, WP, PWH, JJF, GFG, AJT interpreted the findings and commented on and revised drafts of the manuscript. All authors read and approved the final manuscript. Conflicts of interest BJC has received research funding from MedImmune Inc. and Sanofi Pasteur, and consults for Crucell NV. The authors report no other potential conflicts of interest. SL is an assistant investigator and medical epidemiologist of the Division of Infectious Diseases, China CDC, and is also a Ph.D. candidate of the Department of Geography and Environment at the University of Southampton, UK. HY is the director and medical epidemiologist of the Division of Infectious Disease, China CDC.

SL is an assistant investigator and medical epidemiologist of the Division of Infectious Diseases, China CDC, and is also a Ph.D. candidate of the Department of Geography and Environment at the University of Southampton, UK. HY is the director and medical epidemiologist of the Division of Infectious Disease, China CDC. This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. Figure 1 Flow chart of study selection and collection of individual case data on H5N1 cases Figure 2 The geographic distribution of human cases with H5N1 virus infection by outcome, May 1997–April 2015 (n=907) The data for China includes the cases reported by mainland China (52 cases) and Hong Kong SAR (23 cases).

This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain. Figure 1 Flow chart of study selection and collection of individual case data on H5N1 cases Figure 2 The geographic distribution of human cases with H5N1 virus infection by outcome, May 1997–April 2015 (n=907) The data for China includes the cases reported by mainland China (52 cases) and Hong Kong SAR (23 cases). Figure 3 Epidemic curve of human cases with H5N1 virus infection by region, May 1997–April 2015 (A) The epidemic curve of H5N1 human cases reported globally (884 cases). (B) East and Southeast Asia (484 cases) includes Indonesia (187), Viet Nam (134), Cambodia (58), mainland China (52), Thailand (27), Hong Kong SAR (23), Laos (2), and Myanmar (1). (C) North Africa (363 cases) includes Egypt (363). Twenty-three cases with unknown month of illness (21 cases of Indonesia in 2009 and two cases of Turkey in 2006) are excluded from this epidemic curve.

7), Viet Nam (134), Cambodia (58), mainland China (52), Thailand (27), Hong Kong SAR (23), Laos (2), and Myanmar (1). (C) North Africa (363 cases) includes Egypt (363). Twenty-three cases with unknown month of illness (21 cases of Indonesia in 2009 and two cases of Turkey in 2006) are excluded from this epidemic curve. Figure 4 The age distribution of human cases with H5N1 virus infection by gender, geographic regions and outcome, May 1997–April 2015 (A) The age distribution of all cases by male (n=401) and female (n=476). (B) The age distribution of all cases by death (n=463) and survive (n=416). (C) The age distribution of survive cases by North Africa (n=245), East and Southeast Asia (n=152). (D) The age distribution of death cases by North Africa (n=116), East and Southeast Asia (n=329). (E) The age distribution of survive cases in Egypt before (n=114) and since 1 November 2014 (n=131). (F) The age distribution of death cases in Egypt before (n=64) and since 1 November 2014 (n=52). Table 1 The characteristics of human case with H5N1 virus infection by geographic region, May 1997 – April 2015 Characteristics Total (n=907) East and Southeast Asia (n=505) North Africa (n=363) Other (n=39) Type of cases Confirmed case 858 (94.6%) 479 (94.9%) 343 (94.5%) 36 (92.3%) Probable case 49 (5.4%) 26 (5.1%) 20 (5.5%) 3 (7.7%) Sex Female 476 (52.5%) 246 (48.7%) 213 (58.7%) 17 (43.6%) Unknown 29 (3.2%) 21 (4.2%) 6 (1.7%) 2 (5.1%) Age Median (yrs, range) 19 (0.25, 86) 19 (0.3, 75) 20 (0.25, 86) 15 (1.3, 52) Final outcome Death 483 (53.3%) 349 (69.1%) 116 (32%) 18 (46.2%) Unknown 4 (0.4%) 2 (0.4%) 2 (0.6%) 0 (0)

Confirmed case 858 (94.6%) 479 (94.9%) 343 (94.5%) 36 (92.3%) Probable case 49 (5.4%) 26 (5.1%) 20 (5.5%) 3 (7.7%) Sex Female 476 (52.5%) 246 (48.7%) 213 (58.7%) 17 (43.6%) Unknown 29 (3.2%) 21 (4.2%) 6 (1.7%) 2 (5.1%) Age Median (yrs, range) 19 (0.25, 86) 19 (0.3, 75) 20 (0.25, 86) 15 (1.3, 52) Final outcome Death 483 (53.3%) 349 (69.1%) 116 (32%) 18 (46.2%) Unknown 4 (0.4%) 2 (0.4%) 2 (0.6%) 0 (0) Hospitalization Yes 819 (90.3%) 438 (86.7%) 353 (97.2%) 28 (71.8%) Unknown 82 (9%) 64 (12.7%) 9 (2.5%) 9 (23.1%) Median of time delay (days, range) Time from onset to hospital admission 4 (0, 90) 5 (0, 90) 3 (0, 33) 2 (0, 13) Unknown 184 (20.3%) 121 (24%) 46 (12.7%) 17 (43.6%) Time from hospital admission to death or discharge (recovery) 5 (0, 116) 4 (0, 116) 5 (0, 28) 5 (2, 25) Unknown 403 (44.4%) 166 (32.9%) 219 (60.3%) 18 (46.2%) Time from onset to death or discharge (recovery) 10 (0, 119) 10 (0, 119) 10 (2, 32) 9 (2, 32) Unknown 360 (39.7%) 124 (24.6%) 221 (60.9%) 15 (38.5%) Predominant clade or subclade 0 18 (2%) 18 (3.6%) 0 (0) 0 (0) 1 193 (21.3%) 193 (38.2%) 0 (0) 0 (0) 2.1 208 (22.9%) 208 (41.2%) 0 (0) 0 (0) 2.2 393 (43.3%) 0 (0) 363 (100%) 30 (76.9%) 2.3 89 (9.8%) 84 (16.6%) 0 (0) 5 (12.8%) 7 2 (0.2%) 2 (0.4%) 0 (0) 0 (0) Unknown 4 (0.4%) 0 (0) 0 (0) 4 (10.3%) Exposure history

Time from onset to hospital admission 4 (0, 90) 5 (0, 90) 3 (0, 33) 2 (0, 13) Unknown 184 (20.3%) 121 (24%) 46 (12.7%) 17 (43.6%) Time from hospital admission to death or discharge (recovery) 5 (0, 116) 4 (0, 116) 5 (0, 28) 5 (2, 25) Unknown 403 (44.4%) 166 (32.9%) 219 (60.3%) 18 (46.2%) Time from onset to death or discharge (recovery) 10 (0, 119) 10 (0, 119) 10 (2, 32) 9 (2, 32) Unknown 360 (39.7%) 124 (24.6%) 221 (60.9%) 15 (38.5%) Predominant clade or subclade 0 18 (2%) 18 (3.6%) 0 (0) 0 (0) 1 193 (21.3%) 193 (38.2%) 0 (0) 0 (0) 2.1 208 (22.9%) 208 (41.2%) 0 (0) 0 (0) 2.2 393 (43.3%) 0 (0) 363 (100%) 30 (76.9%) 2.3 89 (9.8%) 84 (16.6%) 0 (0) 5 (12.8%) 7 2 (0.2%) 2 (0.4%) 0 (0) 0 (0) Unknown 4 (0.4%) 0 (0) 0 (0) 4 (10.3%) Exposure history Any exposure to poultry 748 (82.5%) 382 (75.6%) 339 (93.4%) 27 (69.2%) Unknown 126 (13.9%) 94 (18.6%) 24 (6.6%) 8 (20.5%) Occupational exposure to live poultry 15 (1.7%) 12 (2.4%) 2 (0.6%) 1 (2.6%) Unknown 586 (64.6%) 289 (57.2%) 286 (78.8%) 11 (28.2%) Visit LBMs 82 (9%) 68 (13.5%) 11 (3%) 3 (7.7%) Unknown 596 (65.7%) 296 (58.6%) 286 (78.8%) 14 (35.9%) Exposure to sick or dead poultry 439 (48.4%) 242 (47.9%) 174 (47.9%) 23 (59%) Unknown 395 (43.6%) 217 (43%) 166 (45.7%) 12 (30.8%) Exposure to backyard poultry 188 (20.7%) 113 (22.4%) 64 (17.6%) 11 (28.2%) Unknown 601 (66.3%) 301 (59.6%) 286 (78.8%) 14 (35.9%) Human case contact 49 (5.4%) 35 (6.9%) 3 (0.8%) 11 (28.2%) Unknown 115 (12.7%) 86 (17%) 21 (5.8%) 8 (20.5%) Note: Data are presented as no. (%) of patients unless otherwise indicated. LBMs: Live bird markets. East and Southeast Asia (505 cases): Indonesia (208), Viet Nam (134), Cambodia (58), mainland China (52), Thailand (27), Hong Kong SAR (23), Laos (2), and Myanmar (1); North Africa (363 cases): Egypt (363); Other (39 cases): Turkey (12), Azerbaijan (9), Bangladesh (7), Pakistan (4), Iraq (3), Nigeria (2), Djibouti (1), and Canada (1). Data on H5N1 clade or subclade of Human cases was based on the reports from WHO website, or the literature, and the known geographic distribution of the viruses. No all cases were laboratory confirmed and reported with clade results, so we presumed that the case in each area was infected by the reported predominant clade or subclade of H5N1 virus in the same period and area. The clade or subclade in each area were clade 0 in Hong Kong SAR in 1997, clade 1 in Viet Nam, Cambodia, Thailand, and Hong Kong SAR, subclade 2.1 mainly in Indonesia, 2.2 in Egypt, Turkey, Azerbaijan, Bangladesh, Iraq, Nigeria and Djibouti, and 2.3 in Viet Nam, Viet Nam, Bangladesh, Laos, Canada and Myanmar, and 7 in mainland China. The data of clade was unavailable for 4 cases in Pakistan in 2007.

in Viet Nam, Cambodia, Thailand, and Hong Kong SAR, subclade 2.1 mainly in Indonesia, 2.2 in Egypt, Turkey, Azerbaijan, Bangladesh, Iraq, Nigeria and Djibouti, and 2.3 in Viet Nam, Viet Nam, Bangladesh, Laos, Canada and Myanmar, and 7 in mainland China. The data of clade was unavailable for 4 cases in Pakistan in 2007. Table 2 The clade or subclade and fatality of human case with H5N1 virus infection, May 1997 – April 2015 Clade or subclade Year first identified Locations identified Case fatality risk 0 1997 Hong Kong SAR 31.6% (6/18) 1 2003 Hong Kong SAR, Vietnam, Cambodia and Thailand 58.6% (112/191) 2.1 2005 Indonesia 84.6% (176/208) 2.2 2005 Turkey, Egypt, Azerbaijan, Djibouti, Iraq, Nigeria, and Bangladesh 33.2% (130/391) 2.3 2005 Mainland China, Laos, Myanmar, Vietnam, Hong Kong SAR, Bangladesh and Canada 61.8% (55/89) 7 2003 Mainland China 100% (2/2) Note: Data on H5N1 clade or subclade of Human cases was based on the reports from WHO website, or the literature, and the known geographic distribution of the viruses. No all cases were laboratory confirmed and reported with clade results, so we presumed that the case was infected by the reported clade or subclade of H5N1 virus in the same period and area. The data of clade was unavailable for four cases in Pakistan in 2007, and four cases with unknown outcome (two of Viet Nam in 2005 and two of Egypt in 2015) were also excluded. Box 1 Definition of exposures to poultry and humans Type of exposure Definition

Clade or subclade Year first identified Locations identified Case fatality risk 0 1997 Hong Kong SAR 31.6% (6/18) 1 2003 Hong Kong SAR, Vietnam, Cambodia and Thailand 58.6% (112/191) 2.1 2005 Indonesia 84.6% (176/208) 2.2 2005 Turkey, Egypt, Azerbaijan, Djibouti, Iraq, Nigeria, and Bangladesh 33.2% (130/391) 2.3 2005 Mainland China, Laos, Myanmar, Vietnam, Hong Kong SAR, Bangladesh and Canada 61.8% (55/89) 7 2003 Mainland China 100% (2/2) Note: Data on H5N1 clade or subclade of Human cases was based on the reports from WHO website, or the literature, and the known geographic distribution of the viruses. No all cases were laboratory confirmed and reported with clade results, so we presumed that the case was infected by the reported clade or subclade of H5N1 virus in the same period and area. The data of clade was unavailable for four cases in Pakistan in 2007, and four cases with unknown outcome (two of Viet Nam in 2005 and two of Egypt in 2015) were also excluded. Box 1 Definition of exposures to poultry and humans Type of exposure Definition Occupational exposure to live poultry refers to poultry related professional exposure (e.g. poultry breeding, trafficking, sale, and quarantine) in the two weeks prior to the onset of illness. Visiting live bird market (LBM) refers to the visit of a wholesale or retail market of live poultry or birds within the two weeks prior to the onset of symptoms. Exposure to sick or dead poultry refers to direct physical contact with, or in proximity to, sick or dead poultry or its production or faeces in the two weeks prior to onset.

Visiting live bird market (LBM) refers to the visit of a wholesale or retail market of live poultry or birds within the two weeks prior to the onset of symptoms. Exposure to sick or dead poultry refers to direct physical contact with, or in proximity to, sick or dead poultry or its production or faeces in the two weeks prior to onset. Exposure to backyard poultry refers to whether the case had been exposed to poultry raised in the backyard within two weeks prior to onset. Any exposure to poultry refers to direct contact, indirect contact or proximity to healthy, sick or dead poultry (including any kind of poultry or birds, e.g. chicken, ducks, goose, pet birds, pigeons, et al.) in LBMs, backyard, farm, or neighborhood, or consumption of improperly processed poultry products. Human case contact refers to a patient with a history of close contact with a confirmed or probable human H5N1 case (any time from the day before the onset of illness to death or during the period that case was hospitalized) in the two weeks before the onset of symptoms.

Introduction Malaria is a major, preventable cause of morbidity, mortality and adverse birth outcomes in sub-Saharan Africa.1, 2 Although malaria mortality has fallen as a result of the scale-up of insecticide-treated bed nets and artemisinin-based combination therapies (ACTs), additional efforts are needed.3 Intermittent preventive treatment (IPT) of malaria is a strategy for the control of malaria in pregnant women (IPTp), infants, children (seasonal malaria chemoprevention [SMC]),4 and potentially in high-risk subgroups of non-pregnant adults and schoolchildren. IPT involves the administration of curative doses of antimalarials at predefined intervals irrespective of malaria infection status. Of the available ACTs, dihydroartemisinin-piperaquine (DP) is one of the most attractive drugs for IPT. It is effective, with cure rates of 98% or more in non-pregnant and pregnant populations.5, 6, 7 The long half-life of piperaquine (about 23 days [range 19–28] in adults and 14 days [range 10–18] in children)6 provides 1–2 weeks' longer post-treatment prophylaxis than artemether-lumefantrine (AL, half-life 3–6 days),8 artesunate-amodiaquine (half-life 6–18 days),9 or sulfadoxine-pyrimethamine (SP, half-life 4–11 days),10 and a similar duration of post-treatment prophylaxis as mefloquine (half-life 10·5–14 days).11 It is well tolerated compared with other antimalarials: side-effects are typically limited to minor gastrointestinal adverse events, mild headache, and dizziness.12

,9 or sulfadoxine-pyrimethamine (SP, half-life 4–11 days),10 and a similar duration of post-treatment prophylaxis as mefloquine (half-life 10·5–14 days).11 It is well tolerated compared with other antimalarials: side-effects are typically limited to minor gastrointestinal adverse events, mild headache, and dizziness.12 DP can cause dose-dependent prolongation of the QT interval13 and is not recommended in patients with congenital long QT syndrome (about one in 2500 children)14 or who are taking other QT prolonging drugs.13 Numerous drugs have been associated with QT prolongation, including multiple classes of antibiotics (eg, erythromycin,15 quinolones,15 co-trimoxazole16) and antimalarials.17 Mild QT prolongation is clinically silent, but extreme prolongation can cause arrhythmias, including torsade de pointes, a potentially fatal polymorphic ventricular tachycardia occurring in roughly one of 10 000 exposures to QT prolonging drugs.18 Diagnosis of prolonged QT requires electrocardiograms (ECG); the normal range differs for men and women, as well as children and adults. Few studies of DP have assessed ECGs.19, 20 Research in context Evidence before this study

DP can cause dose-dependent prolongation of the QT interval13 and is not recommended in patients with congenital long QT syndrome (about one in 2500 children)14 or who are taking other QT prolonging drugs.13 Numerous drugs have been associated with QT prolongation, including multiple classes of antibiotics (eg, erythromycin,15 quinolones,15 co-trimoxazole16) and antimalarials.17 Mild QT prolongation is clinically silent, but extreme prolongation can cause arrhythmias, including torsade de pointes, a potentially fatal polymorphic ventricular tachycardia occurring in roughly one of 10 000 exposures to QT prolonging drugs.18 Diagnosis of prolonged QT requires electrocardiograms (ECG); the normal range differs for men and women, as well as children and adults. Few studies of DP have assessed ECGs.19, 20 Research in context Evidence before this study Malaria is a major, preventable cause of morbidity, mortality, and adverse birth outcomes in sub-Saharan Africa. Intermittent preventive treatment (IPT) of malaria, which involves curative doses of antimalarials at predefined intervals irrespective of malaria infection status, is a strategy for the control of malaria in pregnant women, infants, and children (seasonal malaria chemoprevention [SMC]). Dihydroartemisinin-piperaquine (DP) is an effective, well tolerated antimalarial, and the long half-life of piperaquine makes DP an attractive choice for IPT. However, DP is known to cause dose-dependent prolongation of the QT interval, and limited data exists on whether the risk of QT prolongation is increased with repeated dosing. We conducted a systematic review and meta-analysis to establish the efficacy and safety of repeated treatment with 3-day courses of DP.

for IPT. However, DP is known to cause dose-dependent prolongation of the QT interval, and limited data exists on whether the risk of QT prolongation is increased with repeated dosing. We conducted a systematic review and meta-analysis to establish the efficacy and safety of repeated treatment with 3-day courses of DP. We searched Medline, Embase, Web of Science, Scopus, CINAHL Plus, the Cochrane Library databases, WHO Global Health Library, the Malaria in Pregnancy Consortium (MiPc) Library, ‘grey literature’ databases (unpublished literature including ongoing clinical trials, ongoing PhDs, unpublished PhDs, aborted research, and any other unconventional unpublished literature on the topic), and conference abstracts for articles published before Sept 1, 2016 using the terms: “human” AND “dihydroartemisinin-piperaquine” OR “DHA-PPQ”, and restricting the language to English. There are several reviews on the safety and efficacy of a single course of DP for treatment (ie, case management), and one review of studies of IPT in children (now called SMC) (including two using DP or other piperaquine combinations) and one of IPT in schoolchildren (including one trial using DP). Added value of this study

We searched Medline, Embase, Web of Science, Scopus, CINAHL Plus, the Cochrane Library databases, WHO Global Health Library, the Malaria in Pregnancy Consortium (MiPc) Library, ‘grey literature’ databases (unpublished literature including ongoing clinical trials, ongoing PhDs, unpublished PhDs, aborted research, and any other unconventional unpublished literature on the topic), and conference abstracts for articles published before Sept 1, 2016 using the terms: “human” AND “dihydroartemisinin-piperaquine” OR “DHA-PPQ”, and restricting the language to English. There are several reviews on the safety and efficacy of a single course of DP for treatment (ie, case management), and one review of studies of IPT in children (now called SMC) (including two using DP or other piperaquine combinations) and one of IPT in schoolchildren (including one trial using DP). Added value of this study To our knowledge, this is the first review and meta-analysis to specifically assess the safety and efficacy with repeated courses of DP for case management, IPT, mass drug administration or seasonal malaria chemoprevention in all age groups when compared with placebo or other antimalarial interventions. Monthly DP was more effective than most other options for the prevention of malaria, and appeared to be well tolerated and safe, with less serious adverse events than many comparator interventions and no evidence for increased risk of adverse cardiac events. Nevertheless, data on cardiotoxicity is still scarce. Implications of all the available evidence

To our knowledge, this is the first review and meta-analysis to specifically assess the safety and efficacy with repeated courses of DP for case management, IPT, mass drug administration or seasonal malaria chemoprevention in all age groups when compared with placebo or other antimalarial interventions. Monthly DP was more effective than most other options for the prevention of malaria, and appeared to be well tolerated and safe, with less serious adverse events than many comparator interventions and no evidence for increased risk of adverse cardiac events. Nevertheless, data on cardiotoxicity is still scarce. Implications of all the available evidence DP is a valuable potential candidate for use as IPT and could greatly reduce malaria morbidity and mortality. Additional studies incorporating electrocardiogram measurements are needed to confirm the cardiac safety of repeated monthly dosing.

To our knowledge, this is the first review and meta-analysis to specifically assess the safety and efficacy with repeated courses of DP for case management, IPT, mass drug administration or seasonal malaria chemoprevention in all age groups when compared with placebo or other antimalarial interventions. Monthly DP was more effective than most other options for the prevention of malaria, and appeared to be well tolerated and safe, with less serious adverse events than many comparator interventions and no evidence for increased risk of adverse cardiac events. Nevertheless, data on cardiotoxicity is still scarce. Implications of all the available evidence DP is a valuable potential candidate for use as IPT and could greatly reduce malaria morbidity and mortality. Additional studies incorporating electrocardiogram measurements are needed to confirm the cardiac safety of repeated monthly dosing. Administration of DP with food, particularly fat, increases the bioavailability, leading to increased drug concentrations and a greater degree of QT prolongation, which persists for a longer duration.21 Additionally, piperaquine concentrations might also be increased when co-administered with drugs that are CYP3A4-inhibitors (eg, some protease inhibitors).13 For these reasons, the drug manufacturer recommends obtaining ECGs to monitor therapy when clinically indicated. However, this is not practical if DP is to be given as IPT in resource poor settings and studies assessing the cardiotoxicity of DP when provided for case management show the risk to be low.22 Furthermore, neither DP nor AL displayed an in-vitro signal for a significant pro-arrhythmic risk or appear to induce potential torsadogenic effects.23 However, piperaquine is eliminated slowly and theoretically this risk might be increased when repeated doses are given, especially when given monthly. We conducted a systematic review and meta-analysis to assess the efficacy, safety, and tolerability of repeated dosing of DP when used for case management, IPT, mass drug administration or seasonal malaria chemoprevention.

ically this risk might be increased when repeated doses are given, especially when given monthly. We conducted a systematic review and meta-analysis to assess the efficacy, safety, and tolerability of repeated dosing of DP when used for case management, IPT, mass drug administration or seasonal malaria chemoprevention. Methods Search strategy We did a systematic literature search according to PRISMA guidelines24 on Sept 1, 2016, using simple search terms “human” AND “dihydroartemisinin-piperaquine” OR “DHA-PPQ” (see appendix). Studies were eligible if they were randomised controlled trials (RCTs) or prospective cohort studies involving repeat exposures to standard 3-day courses of DP for either chemoprevention (IPT/SMC), mass drug administration, or treatment of clinical malaria, conducted at any time and in any geographic location. The search was restricted to the English language (appendix). Data management Two independent reviewers (SK, JG) screened titles, abstracts, and full texts and agreed on final study eligibility. Reviewers independently extracted data using a standardised form and database. If required, additional information was obtained from authors. Quality assessment The Cochrane Collaboration's tool was used to assess the quality and risk of bias of clinical trials.25 The quality of observational studies was assessed using the Newcastle Ottawa Scale.26

Data management Two independent reviewers (SK, JG) screened titles, abstracts, and full texts and agreed on final study eligibility. Reviewers independently extracted data using a standardised form and database. If required, additional information was obtained from authors. Quality assessment The Cochrane Collaboration's tool was used to assess the quality and risk of bias of clinical trials.25 The quality of observational studies was assessed using the Newcastle Ottawa Scale.26 Data analysis Random-effects meta-analysis was used to generate pooled incidence rate ratios (IRR) and relative risks, or risk differences when there were zero events in both study groups, to compare the effect of DP relative with other antimalarials or placebo on malaria incidence and tolerability; odds ratio (OR) was used for serious adverse events (SAEs) because they are rare events. To correct for studies reporting no SAEs in either the DP or comparison group, a fixed effects model with continuity correction of 0·5 was used to generate Mantel-Haenzel pooled ORs for each study to be informative.26

tolerability; odds ratio (OR) was used for serious adverse events (SAEs) because they are rare events. To correct for studies reporting no SAEs in either the DP or comparison group, a fixed effects model with continuity correction of 0·5 was used to generate Mantel-Haenzel pooled ORs for each study to be informative.26 Heterogeneity was expressed as I2 value and categorised as low if I2 was 0–40%, moderate if I2 was 30–60%, substantial if I2 was 50–90%, and considerable if I2 was 75–100%.24 Analyses were stratified by study type (case management vs IPT/SMC/mass drug administration) and geographic location (east Africa, west Africa, and Asia). Due to scarcity of data, we could not stratify on pregnancy status. The influence of study quality on results was assessed by sensitivity analyses. Publication bias was assessed through funnel plots. Two-tailed p-values <0·05 were considered statistically significant. Role of the funding source The sponsor of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Heterogeneity was expressed as I2 value and categorised as low if I2 was 0–40%, moderate if I2 was 30–60%, substantial if I2 was 50–90%, and considerable if I2 was 75–100%.24 Analyses were stratified by study type (case management vs IPT/SMC/mass drug administration) and geographic location (east Africa, west Africa, and Asia). Due to scarcity of data, we could not stratify on pregnancy status. The influence of study quality on results was assessed by sensitivity analyses. Publication bias was assessed through funnel plots. Two-tailed p-values <0·05 were considered statistically significant. Role of the funding source The sponsor of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results Our search identified 898 citations; after title review, 380 abstracts and 46 full text articles (29 distinct studies) were reviewed (figure 1). 11 studies were eligible: one cohort study in pregnant women (n=5288),30 one RCT of repeated treatments in children younger than 5 years (n=312),20 and nine RCTs with IPT/SMC (henceforth referred to as IPT). Of the nine RCTs, five were in children younger than 5 years (n=5481),6, 31, 32, 33, 34 one in schoolchildren (n=740),35 one in adult men at occupational risk of malaria (n=961),36 and two in pregnant women (n=1846;37, 38 table). In total, there were 14 628 participants; 4883 in DP groups, of whom 4511 were exposed to DP and 3935 received at least two courses of DP, including 762 pregnant women and 1913 children aged less than 5 years. The remaining 9745 were exposed to placebo or other comparator therapy (including 990 exposed to SP–piperaquine). The 4511 participants exposed to DP received a total of 18  873 courses, with 18 297 courses taken by the 3935 participants who received at least two doses, some of whom received as many as 18 monthly doses. Several different dosing intervals were studied, including monthly (including in pregnancy), every 2 months, quarterly, and three times during the second and third trimester of pregnancy. Comparator interventions included placebo, AL, SP, SP+amodiaquine, SP + piperaquine, SP + chloroquine, and piperaquine + co-trimoxazole. All studies were conducted in areas with no or low parasite resistance to piperaquine or the artemisinins.

arterly, and three times during the second and third trimester of pregnancy. Comparator interventions included placebo, AL, SP, SP+amodiaquine, SP + piperaquine, SP + chloroquine, and piperaquine + co-trimoxazole. All studies were conducted in areas with no or low parasite resistance to piperaquine or the artemisinins. The Cochrane Collaboration tool scored four RCTs as low risk of bias and six as moderate risk of bias (appendix). The Newcastle Ottawa Scale suggested a moderate risk of bias for the single cohort study. Protective efficacy Repeated first-line course of DP for case management was associated with a 16% lower risk of parasitological treatment failure by day 28 compared with AL, but only one trial provided data for analysis (IRR 0·84 95% CI 0·81–0·86).20 Monthly DP for IPT was associated with an 84% reduction in the incidence of malaria parasitaemia measured by microscopy compared with placebo (pooled IRR; figure 2). This was 75% in east Africa, 91% in west Africa, and 97% in adults in Thailand36 (appendix). Monthly IPT with DP provided similar efficacy to monthly SP+amodiaquine for preventing any parasitaemia, but inferior efficacy compared with monthly SP+primaquine (figure 3). Monthly IPT-DP was significantly better than daily co-trimoxazole, or monthly IPT-SP for the prevention of malaria infection. Dosing of IPT-DP on a less than monthly schedule (every 2 months36 or 3 months35) provided significantly less protection against any parasitaemia than monthly dosing (figure 3).

Monthly IPT with DP provided similar efficacy to monthly SP+amodiaquine for preventing any parasitaemia, but inferior efficacy compared with monthly SP+primaquine (figure 3). Monthly IPT-DP was significantly better than daily co-trimoxazole, or monthly IPT-SP for the prevention of malaria infection. Dosing of IPT-DP on a less than monthly schedule (every 2 months36 or 3 months35) provided significantly less protection against any parasitaemia than monthly dosing (figure 3). The considerable heterogeneity (I2>75%) among placebo controlled RCTs was partly explained by difference in quality of the trials as established by the Cochrane Collaboration's tool:25 there was no heterogeneity in the two RCTs classified as having low potential for bias (I2=0%) but high heterogeneity (I2=99·6%) among the four RCTs classified as having moderate potential for bias (appendix). Absence of variability in study quality within each comparator drug subgroup precluded further assessment of the influence of the risk of bias on the heterogeneity by comparator drug. Geographic stratification did not explain the heterogeneity (appendix).

CTs classified as having moderate potential for bias (appendix). Absence of variability in study quality within each comparator drug subgroup precluded further assessment of the influence of the risk of bias on the heterogeneity by comparator drug. Geographic stratification did not explain the heterogeneity (appendix). Among 3960 participants, after excluding arms of studies where most had received no or only 1 course of DP,30, 37 133 SAEs were reported, including 23 in patients receiving DP+co-trimoxazole. Including all 4883 participants in DP groups (3935 of whom received at least two courses), 233 SAEs were reported. An additional four SAEs were reported in 990 recipients of SP+piperaquine. Among 3180 participants receiving other treatments and 5575 receiving placebo, 287 and 186 SAEs were reported, respectively (table 1). After correction for zero events, repeated DP exposure was associated with a significantly lower odds of SAEs compared with placebo, co-trimoxazole, or IPT-SP (figure 4). IPT-DP was also associated with fewer hospital admissions than IPT-SP (appendix). Repeated case management with DP was associated with fewer hospital admissions compared with AL (appendix).

ed DP exposure was associated with a significantly lower odds of SAEs compared with placebo, co-trimoxazole, or IPT-SP (figure 4). IPT-DP was also associated with fewer hospital admissions than IPT-SP (appendix). Repeated case management with DP was associated with fewer hospital admissions compared with AL (appendix). None of the 11 studies reported SAEs was consistent with sudden cardiac death. Overall, 15 deaths were reported among those exposed to DP, two among those exposed to SP+piperaquine, 18 among those exposed to other comparator therapies, and four among those in placebo groups. No studies reported any sudden or unexplained deaths (figure 5, appendix). IPTp-DP was not associated with an increased risk of loss to follow-up (which could reflect undetected or unreported sudden death) compared with co-trimoxazole, SP, SP + piperaquine, or SP+amodiaquine, but was associated with a 47% higher risk of loss to follow-up compared with placebo (appendix).

(figure 5, appendix). IPTp-DP was not associated with an increased risk of loss to follow-up (which could reflect undetected or unreported sudden death) compared with co-trimoxazole, SP, SP + piperaquine, or SP+amodiaquine, but was associated with a 47% higher risk of loss to follow-up compared with placebo (appendix). The effect of DP on cardiac repolarisation was assessed in 19 HIV-unexposed31 and seven HIV-exposed children (Dorsey, unpublished) and 30 pregnant women.38 In the 26 children, 183 ECGs were conducted at baseline and follow-up (4–6 h after the third dose of DP with each monthly course); all of the baseline and follow-up ECGs had a QTc less than 450 ms with a mean QTc of 396 ms (SD 31·3, range 278–444). There were no differences in the mean QTc intervals measured after the third dose for children who had been prescribed three to five previous courses of DP (mean QTc 405 ms, SD 26), six to ten previous courses of DP (388 ms, 33), or 11–18 previous courses of DP (396 ms, 33). None of the 30 pregnant women who underwent ECG measurements at 28 weeks' gestational age pre-dosing and post-dosing had QTc intervals greater than 450 ms.38 The median increase in QTc from baseline to 4–6 hours after the third dose was 30 ms (range −30 to 50) and 20 ms (−10 to 50) in women randomised to receive monthly DP (n=13) and 3 doses of DP (n=17), respectively, compared with 5 ms (−40 to 60) in women who received three doses of SP (n=12, p=0·57 and 0·28 for monthly and three dose DP compared with three dose SP).

hours after the third dose was 30 ms (range −30 to 50) and 20 ms (−10 to 50) in women randomised to receive monthly DP (n=13) and 3 doses of DP (n=17), respectively, compared with 5 ms (−40 to 60) in women who received three doses of SP (n=12, p=0·57 and 0·28 for monthly and three dose DP compared with three dose SP). IPT-DP was associated with similar cumulative risk of any vomiting compared with placebo, SP, and SP+amodiaquine, and with a lower risk compared with SP+primaquine (appendix). In the single treatment study with tolerability data, DP was associated with a lower risk of vomiting compared with AL (RR 0·52, 95% CI 0·45–0·61).20 In children under 5 years of age, both IPT-DP and IPT-SP were associated with more vomiting during the first course than subsequent monthly courses (DP around 4% vs <2%; SP around 3·5% vs <2%).6 No vomiting was reported among school-aged (6–14 years) children receiving monthly IPT-DP after any of the three courses.35 Treatment of clinical malaria with DP was not associated with more vomiting than AL for the first and second courses, and for the third course, participants given DP vomited less than those given AL (2·8% vs 7·8% p=0·08).20 IPT-DP was not associated with an increased risk of diarrhoea compared with placebo, SP+amodiaquine, SP+primaquine, or IPT-SP in six studies (appendix). Only four studies provided data on rash or allergic reactions, and no study reported any SAEs due to allergic reactions. IPT-DP was not associated with an increased risk of rash compared with placebo, SP+primaquine or SP+amodiaquine (appendix).

IPT-DP was not associated with an increased risk of diarrhoea compared with placebo, SP+amodiaquine, SP+primaquine, or IPT-SP in six studies (appendix). Only four studies provided data on rash or allergic reactions, and no study reported any SAEs due to allergic reactions. IPT-DP was not associated with an increased risk of rash compared with placebo, SP+primaquine or SP+amodiaquine (appendix). Discussion This meta-analysis suggests that DP is as safe as other combinations assessed for IPT or the repeat treatment of clinical malaria, and that it was well tolerated. DP provided superior protection against malaria and resulted in fewer hospital admissions than comparators. In comparison with dosing every 2 or 3 months, monthly administration of DP provided much better protection from malaria, without increasing the risk of adverse events or adversely affecting tolerability.35, 36, 37, 38

vided superior protection against malaria and resulted in fewer hospital admissions than comparators. In comparison with dosing every 2 or 3 months, monthly administration of DP provided much better protection from malaria, without increasing the risk of adverse events or adversely affecting tolerability.35, 36, 37, 38 DP, like some other antimalarials, has been associated with dose dependent risk for QTc prolongation. A previous review assessing the risk of QTc prolongation following a single course of treatment found no difference in the risk for prolonged QTc between DP and AL, but DP was associated with more frequent prolongation of the QTc interval compared with mefloquine-artesunate.39 No cardiac arrhythmias or sudden death were reported for any of the drugs, although it is possible that sudden death due to a cardiac arrhythmia could have been incorrectly attributed to other causes. Similarly, in our meta-analysis, no cardiac events were reported among 3935 recipients of repeat courses of DP involving 18 297 courses of DP ranging from two to 18 courses per individual. As only three studies in different populations assessed the effect of DP on the ECG, it was not possible to do a meta-analysis to assess the risk of repeated courses of DP on the ECG; however, no significant QT prolongation was reported with repeat dosing in the individual studies. Furthermore, the risk of death was not significantly increased following receipt of repeat courses of DP, suggesting no significant increased risk of sudden cardiac death, although the rare nature of this event makes it difficult to rule out. It should be noted, however, that although DP was not associated with increased loss to follow-up compared with comparators, there was more loss to follow-up among participants in the DP group in the studies comparing DP against placebo. This was driven primarily by the high loss to follow-up in Lwin and colleagues,36 which was unrelated to the intervention since only four withdrew due to adverse events: two from the IPT group and two from the placebo group. The rest were lost due to other reasons.

DP group in the studies comparing DP against placebo. This was driven primarily by the high loss to follow-up in Lwin and colleagues,36 which was unrelated to the intervention since only four withdrew due to adverse events: two from the IPT group and two from the placebo group. The rest were lost due to other reasons. The clinical relevance of the dose dependent risk for QTc prolongation with DP is not clear, since the pro-arrhythmic potential of piperaquine in vitro appears lower than chloroquine and similar to AL.23 One post-marketing study, comparing a compressed 2-day regimen of DP with placebo, has reported clinically significant QT prolongation among participants exposed to DP.19 Given the potential for dose accumulation with monthly dosing,36 it was reassuring to find that the studies in children did not find evidence that repeat monthly courses were associated with greater degrees of QT prolongation than the first course, even among children that had received ten or more monthly courses31 (Dorsey, unpublished data).

dose accumulation with monthly dosing,36 it was reassuring to find that the studies in children did not find evidence that repeat monthly courses were associated with greater degrees of QT prolongation than the first course, even among children that had received ten or more monthly courses31 (Dorsey, unpublished data). It is possible that the absence of additional QTc prolongation with repeat courses reflects the finding that QTc interval returns to normal within approximately 12–48 h following the last dose after each course. Nevertheless, some increase in QTc prolongation with increasing number of courses could not be excluded in pregnant women,38 and the strength of the evidence to date is limited because ECGs were only done in three studies involving 56 participants receiving monthly courses, thus more studies are needed. Advances in mobile adapted technology, such as Smartheart or AliveCor, might allow for improved monitoring of patients in remote and resource poor settings in the future. Our search did not find any studies where repeat courses of DP were provided as part of malaria elimination campaigns that often involve multiple rounds of mass drug administration within a single year. However, our findings with DP as IPT are likely to be generalisable to mass drug administration since both strategies involve asymptomatic carriers of malaria parasites and individuals without malaria parasites at the time of drug administration.

involve multiple rounds of mass drug administration within a single year. However, our findings with DP as IPT are likely to be generalisable to mass drug administration since both strategies involve asymptomatic carriers of malaria parasites and individuals without malaria parasites at the time of drug administration. The few people exposed to multiple courses of DP to date precludes our ability to detect an increased risk of an arrhythmia such as torsadogenic event that occur in about one in 10 000 exposures to QT prolonging drugs; our overall sample size can only exclude (95% CI) such events in one in 6099 exposures. The fact that different patient populations were grouped is a potential weakness, as the QTc (and risk for cardiotoxicity) is affected by age, sex, and pregnancy status, and achieved drug concentrations might also vary by patient population and gestational age; however, the paucity of data precluded reviewing the groups individually. The included studies involving young children were conducted before WHO's dose increase for DP in children aged 1–4 years,4 and continued collection of safety data with the new dose is needed. One of the treatment studies included many participants who only received a single course of DP. However, no details were provided in the source study that allowed a breakdown of SAEs by number of courses. Finally, it is possible that restricting to English language excluded relevant studies published in other languages.

e of the treatment studies included many participants who only received a single course of DP. However, no details were provided in the source study that allowed a breakdown of SAEs by number of courses. Finally, it is possible that restricting to English language excluded relevant studies published in other languages. In this meta-analysis of nearly 4000 patients exposed to repeated courses of DP, IPT-DP was highly effective for the prevention of malaria and reduced all-cause hospital admission compared with other drugs, particularly when provided as monthly courses. Overall, DP was well tolerated, with no evidence of increased frequency of mild or serious adverse events with repeated dosing. The data do not suggest that the known risk of QT prolongation increases with repeated monthly courses, or an increased risk of cardiac events or death following repeated dosing. DP is a valuable potential candidate for use as IPT and ongoing monitoring for cardiac events is needed to provide further reassurance of its safety with repeat doses. This online publication has been corrected. The corrected version first appeared at thelancet.com/infection on January 5, 2017 Supplementary Material Supplementary appendix Acknowledgments We would like to thank Jeanne Rini Poespoprodjo, Joaniter Nankabirwa, and Issaka Zongo for providing additional data and Richard Kovacs for advice on the effects of QT prolongation.

This online publication has been corrected. The corrected version first appeared at thelancet.com/infection on January 5, 2017 Supplementary Material Supplementary appendix Acknowledgments We would like to thank Jeanne Rini Poespoprodjo, Joaniter Nankabirwa, and Issaka Zongo for providing additional data and Richard Kovacs for advice on the effects of QT prolongation. Funding AS and FOtK were partly funded by the Malaria in Pregnancy Consortium that is funded through a grant from the Bill & Melinda Gates Foundation to the Liverpool School of Tropical Medicine. FOtK is also grateful to the US Centers for Disease Control and Prevention (CDC) through a cooperative agreement between the Division of Parasitic Diseases and Malaria, CDC, and the Malaria Epidemiology Unit of the Child and Reproductive Health group, Liverpool School of Tropical Medicine held by FOtK. SK was supported in part by the National Center for Advancing Translational Sciences of the National Institutes of Health under Award Number TL1TR000422. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Disclaimer The findings and conclusions presented in this manuscript are those of the authors and do not necessarily reflect the official position of the US Centers for Disease Control and Prevention.

Funding AS and FOtK were partly funded by the Malaria in Pregnancy Consortium that is funded through a grant from the Bill & Melinda Gates Foundation to the Liverpool School of Tropical Medicine. FOtK is also grateful to the US Centers for Disease Control and Prevention (CDC) through a cooperative agreement between the Division of Parasitic Diseases and Malaria, CDC, and the Malaria Epidemiology Unit of the Child and Reproductive Health group, Liverpool School of Tropical Medicine held by FOtK. SK was supported in part by the National Center for Advancing Translational Sciences of the National Institutes of Health under Award Number TL1TR000422. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Disclaimer The findings and conclusions presented in this manuscript are those of the authors and do not necessarily reflect the official position of the US Centers for Disease Control and Prevention. Contributors SK and JG wrote the protocol with input from AS and FTK. SK and JG developed the search terms. JG and SK reviewed all abstracts and assigned bias scores independently; AS served as the tiebreaker. SK and JG abstracted all data independently. SK conducted all meta-analyses. JG and SK drafted the manuscript. FTK, AS, and GD reviewed and edited the manuscript. All authors read and approved the final version. Declaration of interests We declare no competing interests. Figure 1 PRISMA flow chart

Contributors SK and JG wrote the protocol with input from AS and FTK. SK and JG developed the search terms. JG and SK reviewed all abstracts and assigned bias scores independently; AS served as the tiebreaker. SK and JG abstracted all data independently. SK conducted all meta-analyses. JG and SK drafted the manuscript. FTK, AS, and GD reviewed and edited the manuscript. All authors read and approved the final version. Declaration of interests We declare no competing interests. Figure 1 PRISMA flow chart DP=dihydroartemisinin-piperaquine. AL=artemether-lumefantrine. A second trial29 reporting on the use of DP for rescue treatment among pregnant women included nine women who received at least two courses of DP (six received three courses and three received two courses), but all women had also received a preceding course of either quinine or intravenous artesunate with or without clindamycin, and there were no control women who had not received DP. *One trial28 comparing seasonal malaria chemoprevention (SMC) with sulphadoxine-pyrimethamine plus amodiaquine vs placebo SMC (passive case detection and case management with either DP or AL during the malaria transmissions season) was excluded because only 27 of 800 children (3·4%) in placebo SMC group (ie, the DP case management group) received two or more courses of DP and safety data by number of courses received were not available. Figure 2 Pooled incidence rate ratio for any parasitaemia, monthly dihydroartemisinin-piperaquine vs placebo

DP=dihydroartemisinin-piperaquine. AL=artemether-lumefantrine. A second trial29 reporting on the use of DP for rescue treatment among pregnant women included nine women who received at least two courses of DP (six received three courses and three received two courses), but all women had also received a preceding course of either quinine or intravenous artesunate with or without clindamycin, and there were no control women who had not received DP. *One trial28 comparing seasonal malaria chemoprevention (SMC) with sulphadoxine-pyrimethamine plus amodiaquine vs placebo SMC (passive case detection and case management with either DP or AL during the malaria transmissions season) was excluded because only 27 of 800 children (3·4%) in placebo SMC group (ie, the DP case management group) received two or more courses of DP and safety data by number of courses received were not available. Figure 2 Pooled incidence rate ratio for any parasitaemia, monthly dihydroartemisinin-piperaquine vs placebo DP=dihydroartemisinin-piperaquine. PYAR=person-years at risk. IR=incidence rate. IRR=incidence rate ratio. Lwin and colleagues36 and Zongo and colleagues6 did not report PYAR, instead they reported cumulative incidence over a year. PYAR was calculated based on the incidence rate and number of events. Zongo and colleagues'6 numbers are based on intent to treat. Figure 3 Pooled incidence rate ratio or relative risk for any parasitaemia, monthly dihydroartemisinin-piperaquine vs any other therapy

DP=dihydroartemisinin-piperaquine. PYAR=person-years at risk. IR=incidence rate. IRR=incidence rate ratio. Lwin and colleagues36 and Zongo and colleagues6 did not report PYAR, instead they reported cumulative incidence over a year. PYAR was calculated based on the incidence rate and number of events. Zongo and colleagues'6 numbers are based on intent to treat. Figure 3 Pooled incidence rate ratio or relative risk for any parasitaemia, monthly dihydroartemisinin-piperaquine vs any other therapy DP=dihydroartemisinin-piperaquine. PYAR=person-years at risk. IR=incidence rate. IRR=incidence rate ratio. CTX=co-trimoxazole. SP=sulfadoxine-pyrimethamine. SP+PQ=sulfadoxine-pyrimethamine piperaquine. SP+AQ=sulfadoxine-pyrimethamine amodiaquine. AL=artemether lumefantrine. Lwin and colleagues36 and Zongo and colleagues6 did not report PYAR. PYAR was calculated based on the incidence rate and number of events. Cisse and colleagues32 reported cumulative incidence. Kakuru37 reported detection of malaria parasites by LAMP at each visit as the prevalence of positive tests during pregnancy out of all tests. Zongo and colleagues'6 numbers are based on intention to treat. Figure 4 Pooled odds ratios for any serious adverse event after exposure to dihydroartemisinin-piperaquine stratified by comparator therapy

DP=dihydroartemisinin-piperaquine. PYAR=person-years at risk. IR=incidence rate. IRR=incidence rate ratio. CTX=co-trimoxazole. SP=sulfadoxine-pyrimethamine. SP+PQ=sulfadoxine-pyrimethamine piperaquine. SP+AQ=sulfadoxine-pyrimethamine amodiaquine. AL=artemether lumefantrine. Lwin and colleagues36 and Zongo and colleagues6 did not report PYAR. PYAR was calculated based on the incidence rate and number of events. Cisse and colleagues32 reported cumulative incidence. Kakuru37 reported detection of malaria parasites by LAMP at each visit as the prevalence of positive tests during pregnancy out of all tests. Zongo and colleagues'6 numbers are based on intention to treat. Figure 4 Pooled odds ratios for any serious adverse event after exposure to dihydroartemisinin-piperaquine stratified by comparator therapy DP=dihydroartemisinin-piperaquine. SAE=serious adverse event. CTX=co-trimoxazole. IPT=intermittent preventive treatment. IST=intermittent screening and treatment. SP=sulfadoxine-pyrimethamine. SP+PQ=sulfadoxine-pyrimethamine piperaquine. SP+AQ=sulfadoxine-pyrimethamine amodiaquine. SP+CQ=sulfadoxine-pyrimethamine chloroquine. AL=artemether-lumefantrine. Zongo and colleagues'6 numbers are based on actual drug exposures. Poespoprodjo and colleagues30: only 64 of 408 DP recipients received two or more courses of DP, but information of SAEs by number of courses received was not available. Figure 5 Pooled odds ratios for death after exposure to repeated courses of dihydroartemisinin-piperaquine stratified by comparator therapy

DP=dihydroartemisinin-piperaquine. SAE=serious adverse event. CTX=co-trimoxazole. IPT=intermittent preventive treatment. IST=intermittent screening and treatment. SP=sulfadoxine-pyrimethamine. SP+PQ=sulfadoxine-pyrimethamine piperaquine. SP+AQ=sulfadoxine-pyrimethamine amodiaquine. SP+CQ=sulfadoxine-pyrimethamine chloroquine. AL=artemether-lumefantrine. Zongo and colleagues'6 numbers are based on actual drug exposures. Poespoprodjo and colleagues30: only 64 of 408 DP recipients received two or more courses of DP, but information of SAEs by number of courses received was not available. Figure 5 Pooled odds ratios for death after exposure to repeated courses of dihydroartemisinin-piperaquine stratified by comparator therapy Comparisons with zero events in both groups were excluded from the analysis of the pooled OR. OR=odds ratio. DP=dihydroartemisinin-piperaquine. CTX=co-trimoxazole. IPT=intermittent preventive treatment. IST=intermittent screening and treatment. SP=sulfadoxine-pyrimethamine. SP+PQ=sulfadoxine-pyrimethamine piperaquine. SP+AQ=sulfadoxine-pyrimethamine amodiaquine. SP+CQ=sulfadoxine-pyrimethamine chloroquine. AL=artemether-lumefantrine. Zongo and colleagues'6 numbers are based on actual drug exposures. Poespoprodjo and colleagues: only 64 of 408 DP recipients received two or more courses of DP. Table Details of included studies

Comparisons with zero events in both groups were excluded from the analysis of the pooled OR. OR=odds ratio. DP=dihydroartemisinin-piperaquine. CTX=co-trimoxazole. IPT=intermittent preventive treatment. IST=intermittent screening and treatment. SP=sulfadoxine-pyrimethamine. SP+PQ=sulfadoxine-pyrimethamine piperaquine. SP+AQ=sulfadoxine-pyrimethamine amodiaquine. SP+CQ=sulfadoxine-pyrimethamine chloroquine. AL=artemether-lumefantrine. Zongo and colleagues'6 numbers are based on actual drug exposures. Poespoprodjo and colleagues: only 64 of 408 DP recipients received two or more courses of DP. Table Details of included studies Country Study type Study population Comparators* Efficacy data SAE† Deaths DOT,‡number of courses Bigira, 201431 Uganda Clinical trial-IPT Children under 5 years including HIV exposed infants DP 98 SP 98 CTX 99 No treatment 98 Monthly active detection of parasitaemia from 6–24 months of age DP 13 SP 52 CTX 29 No treatment 26 DP 0 SP 2 CTX 2 No treatment 1 First dose DOT, 1592 courses administered§ Bojang, 201032 The Gambia Clinical trial-IPT Children under 5 years DP 336 (335, 328) SP+AQ 336 SP+PQ 336 No treatment 286 Any malaria within 16-week rainy season (passive surveillance), active detection at study end DP 4 SP+AQ 2 SP+PQ 1 No treatment 0 DP 1 SP+AQ 0 SP+PQ 0 No treatment 0 All doses DOT, 952 courses administered Cisse, 200933 Senegal Clinical trial-IPT Children under 5 years DP 598 (578, 539) SP+AQ 607 SP+PQ 654 Passive detection of malaria during 4-month rainy season, active detection at study end DP 2 SP+AQ 2 SP+PQ 2 DP 2 SP+AQ 2 SP+PQ 2 First dose DOT, 1544 courses administered Desai, 201537 Kenya Clinical trial-IPT Pregnant women in second or third trimester IPT-DP 516 (516, 477) IST-DP 515 (167, 27) IPT-SP 515 Active detection of parasitaemia at each antenatal clinic visit during pregnancy IPT-DP 37 IST-DP 82 IPT-SP 85 IPT-DP 0 IST-DP 1 IPT-SP 2 First dose DOT, 1585 courses administered Kakuru,201538 Uganda Clinical trial-IPT Pregnant women in second or third trimester DP monthly 100 DP ×3 94 SP ×3 106 Monthly assessment with LAMP¶ DP monthly 4 DP ×3 9 SP ×3 6 DP monthly 0 DP ×3 0 SP ×3 0 First dose DOT, 1136 courses administered Kamya,201434 Uganda Clinical trial-IPT Children under 5 years DP 47 SP 46 CTX 47 No treatment 46 Monthly active detection of parasitaemia from age 4–5 months until age 24 months DP 10 SP 23 CTX 16 No treatment 21 DP 1 SP 2 CTX 2 No treatment 2 No DOT, drug intake recorded by parents, 561 courses administered§ Lwin,201236 Thailand Clinical trial-IPT Adults DP 387 DP Q2month 381 Placebo 193 Monthly active detection of parasitaemia for 36 weeks DP 1 DP Q2 month 0 Placebo 0 DP 1 DP Q2 month 0 Placebo 0 All doses DOT, 4089 courses administered§ Nankabirwa, 201435 Uganda Clinical trial-IPT School-age children (aged 6–14 years) DP

Lwin,201236 Thailand Clinical trial-IPT Adults DP 387 DP Q2month 381 Placebo 193 Monthly active detection of parasitaemia for 36 weeks DP 1 DP Q2 month 0 Placebo 0 DP 1 DP Q2 month 0 Placebo 0 All doses DOT, 4089 courses administered§ Nankabirwa, 201435 Uganda Clinical trial-IPT School-age children (aged 6–14 years) DP 244 DP quarterly 248 Placebo 248 Monthly active detection of parasitaemia for 12 months DP 6 DP quarterly 5 Placebo 2 DP 0 DP quarterly 1 Placebo 0 All doses DOT, 2648 courses administered Poespoprodjo, 201430 Indonesia Cohort study-treatment Pregnant women in second or third trimester DP 408 (408, 64) ‖ SP+CQ 24 Quinine 402 No treatment 4454 No DP 10 SP+CQ 0 Quinine 18 No treatment 134 DP 0 SP+CQ 0 Quinine 0 No treatment 0 First dose DOT, 486 courses administered Wanzira, 201420 Uganda Clinical trial-treatment Children under 5 years including HIV exposed infants DP (+/− CTX) 154 (154, 147)** AL (+/− CTX) 158 Passive detection of parasitaemia before age 5 years DP 13 DP+CTX 23 AL 39 AL+CTX 14 DP 0 DP+CTX 4 AL 1 AL+CTX 3 First dose DOT, 2218 courses administered Zongo, 20156 Burkina Faso Clinical trial-IPT Children under 5 years DP 750 (757)†† SP+AQ 749 (742)†† No treatment 250 Monthly active detection of parasitaemia for 4 months DP 6 SP+AQ 3 No treatment 2 DP 4 SP+AQ 2 No treatment 1 All doses DOT, 2063 courses administered SAE=serious adverse event.

218 courses administered Zongo, 20156 Burkina Faso Clinical trial-IPT Children under 5 years DP 750 (757)†† SP+AQ 749 (742)†† No treatment 250 Monthly active detection of parasitaemia for 4 months DP 6 SP+AQ 3 No treatment 2 DP 4 SP+AQ 2 No treatment 1 All doses DOT, 2063 courses administered SAE=serious adverse event. DOT=directly observed therapy. IPT=intermittent preventive treatment. IST=intermittent screening and treatment. DP=dihydroartemisinin-piperaquine. SP=sulfadoxine pyrimethamine. CTX=co-trimoxazole. AQ=amodiaquine, PQ=piperaquine. CQ=chloroquine. AL=artemether-lumefantrine. Q2month=every other month. CHW=community health worker. LAMP=loop-mediated isothermal amplification. * Numbers in brackets represent the number who received one or more and two or more courses of DP, if reported to be different from the overall sample size in the DP group.

DOT=directly observed therapy. IPT=intermittent preventive treatment. IST=intermittent screening and treatment. DP=dihydroartemisinin-piperaquine. SP=sulfadoxine pyrimethamine. CTX=co-trimoxazole. AQ=amodiaquine, PQ=piperaquine. CQ=chloroquine. AL=artemether-lumefantrine. Q2month=every other month. CHW=community health worker. LAMP=loop-mediated isothermal amplification. * Numbers in brackets represent the number who received one or more and two or more courses of DP, if reported to be different from the overall sample size in the DP group. † In addition to any other SAEs reported by the study, all hospital admissions and deaths were considered SAEs. SAEs were reported unrelated to study drugs unless otherwise noted: Bigira and colleagues reported 19 (4·5%) grade 3–4 AEs as possibly related to study drugs, with no significant differences between the intervention groups. Desai and colleagues reported one drug related SAE (an allergic reaction to DP); Kakuru and colleagues reported one patient who developed anaemia after both the first and second dose of DP, after which DP was stopped; Kamya and colleagues reported eight (5·6%) AEs possibly related to study drugs, with no significant differences between the intervention groups; Lwin and colleagues reported that four patients withdrew due to drug related AEs (two in the DP every other month group and two in placebo group). ‡ DOT by study staff, first dose=only the first dose of each course was administered as DOT. § The total number of doses was divided by three to estimate the number of courses and rounded to the nearest whole number.

† In addition to any other SAEs reported by the study, all hospital admissions and deaths were considered SAEs. SAEs were reported unrelated to study drugs unless otherwise noted: Bigira and colleagues reported 19 (4·5%) grade 3–4 AEs as possibly related to study drugs, with no significant differences between the intervention groups. Desai and colleagues reported one drug related SAE (an allergic reaction to DP); Kakuru and colleagues reported one patient who developed anaemia after both the first and second dose of DP, after which DP was stopped; Kamya and colleagues reported eight (5·6%) AEs possibly related to study drugs, with no significant differences between the intervention groups; Lwin and colleagues reported that four patients withdrew due to drug related AEs (two in the DP every other month group and two in placebo group). ‡ DOT by study staff, first dose=only the first dose of each course was administered as DOT. § The total number of doses was divided by three to estimate the number of courses and rounded to the nearest whole number. ¶ Reported prevalence over the course of pregnancy (incidence was not reported). ‖ Average duration between courses 4·2 months ** Average duration between courses 2·2 months †† Intention to treat included 750 in the DP group and 740 in the SP+AQ group, but due to allocation errors, 757 were given DP and 742 were given SP+AQ.

Introduction Yellow fever virus is a mosquito-borne flavivirus that causes infections in human beings, with symptoms ranging from mild non-specific illness to severe disease with jaundice, haemorrhage, and death.1 A single-dose vaccine has existed since the 1940s and has helped to control and reduce yellow fever virus transmission substantially.2, 3, 4 Complete eradication is, however, prevented by the sylvatic cycle of the virus within which non-human primates act as primary hosts and Aedes aegypti mosquitoes are responsible for occasional transmission to people.5, 6 From Dec 5, 2015, to November, 2016, a large yellow fever outbreak has affected Angola and the Democratic Republic of the Congo (DR Congo), with 7334 suspected cases, of which 962 have been confirmed, and 393 deaths reported to WHO as of Oct 28, 2016.7 Responses to such outbreaks rely mainly on reactive vaccination campaigns and pose various strategic and logistical challenges. For example, in response to the current outbreak in central Africa, the global yellow fever virus vaccine emergency stockpile (6 million doses) was exhausted after the initial mass vaccination campaign.7 In the context of finite resources, decisions about which geographic areas should be targeted first need to be informed by a detailed understanding of the determinants of the spatial spread of yellow fever virus and by predictions of where yellow fever virus is most likely to spread to in the future.

ination campaign.7 In the context of finite resources, decisions about which geographic areas should be targeted first need to be informed by a detailed understanding of the determinants of the spatial spread of yellow fever virus and by predictions of where yellow fever virus is most likely to spread to in the future. Such assessments need to capture a range of factors. First, human mobility could facilitate the introduction of the pathogen into disease-free areas, which has been reported for other outbreaks in Africa.8, 9 Such regional spread of a disease is largely governed by underlying population structures and transport networks, as well as by individuals' economic, cultural, and recreational activities.10, 11 Furthermore, in the context of a vector-borne disease such as yellow fever, the ecological landscape of the A aegypti mosquito that transmits the virus between people must also coincide with patterns of human movements for a successful viral transmission cycle to be established.12, 13 This ecological landscape has been shown to be strongly affected by temperature, precipitation, humidity, vegetation coverage, and degree of urbanisation.14 Research in context Evidence before this study

Such assessments need to capture a range of factors. First, human mobility could facilitate the introduction of the pathogen into disease-free areas, which has been reported for other outbreaks in Africa.8, 9 Such regional spread of a disease is largely governed by underlying population structures and transport networks, as well as by individuals' economic, cultural, and recreational activities.10, 11 Furthermore, in the context of a vector-borne disease such as yellow fever, the ecological landscape of the A aegypti mosquito that transmits the virus between people must also coincide with patterns of human movements for a successful viral transmission cycle to be established.12, 13 This ecological landscape has been shown to be strongly affected by temperature, precipitation, humidity, vegetation coverage, and degree of urbanisation.14 Research in context Evidence before this study We searched PubMed up to Oct 11, 2016, for publications about any yellow fever virus outbreaks in Angola and the Democratic Republic of the Congo (DR Congo), yellow fever virus outbreaks since the beginning of 2015, and the spatial spread of yellow fever virus, without any language restrictions. We used the search terms (“yellow fever”[MeSH Terms] OR “yellow fever”[All Fields]) AND ((“angola”[MeSH Terms] OR “angola”[All Fields]) OR (“congo”[MeSH Terms] OR “congo”[All Fields] OR “DRC”[All Fields]) OR ((“epidemic”[All Fields] OR “epidemics”[MeSH Terms] OR “disease outbreaks”[MeSH Terms]) AND (“2015/01/01”[PDAT]:”3000/12/31”[PDAT])) OR “spatial”[All Fields]). Almost no empirical information is available for the current yellow fever virus outbreak in Angola and DR Congo. Previous work has described the identification of the outbreak, discussed vaccination coverage and the apparent vaccine shortage from a policy perspective, and emphasised the risk of international spread outside the region.

cal information is available for the current yellow fever virus outbreak in Angola and DR Congo. Previous work has described the identification of the outbreak, discussed vaccination coverage and the apparent vaccine shortage from a policy perspective, and emphasised the risk of international spread outside the region. Added value of this study To our knowledge our study is the first to investigate key epidemiological parameters of the outbreak and show the important ecological and demographic determinants that govern transmission and spread of the virus in the region. Implications of all the available evidence Across districts in Angola and the DR Congo, the spread of yellow fever virus was governed by high population density, including locations that were distant from the origin of the outbreak in Luanda. Transmission is also more likely to be sustained in areas where people live in close proximity. Human movements in and out of the capital cities of Angola and the DR Congo have caused yellow fever virus to spread to almost all districts in Angola. Our model fits the expansion process of the pathogen well and allows for extrapolation into the future. Our approach can be used to help policy makers prioritise areas to be targeted, especially in the context of finite public-health resources.

have caused yellow fever virus to spread to almost all districts in Angola. Our model fits the expansion process of the pathogen well and allows for extrapolation into the future. Our approach can be used to help policy makers prioritise areas to be targeted, especially in the context of finite public-health resources. In this study, we jointly analysed datasets describing the epidemic and spatial spread of yellow fever virus, vector suitability, human demography, and mobility in central Africa, aiming to better understand, quantify, and predict the spread of yellow fever virus in this region. By identifying the areas that are at the highest risk of yellow fever virus transmission, such a framework could inform the response to ongoing outbreaks. Methods Epidemiological and vector data We obtained epidemiological data on confirmed and suspected yellow fever virus cases in Angola and the DR Congo (the most affected countries) from the WHO situation reports for the period Dec 4, 2015, to Aug 4, 2016. We extracted the first and last dates of disease onset for each district in Angola (n=163) and for each commune in the DR Congo (n=150). We used numbers of cases per week in Angola, but weekly case numbers were not available for the DR Congo at the time of data collection (Aug 4, 2016). We matched locations with administrative unit files available from the Database of Global Administrative Areas.

ola (n=163) and for each commune in the DR Congo (n=150). We used numbers of cases per week in Angola, but weekly case numbers were not available for the DR Congo at the time of data collection (Aug 4, 2016). We matched locations with administrative unit files available from the Database of Global Administrative Areas. To generate district-level ecological risk of disease transmission, we extracted estimated suitability values for the primary urban vector A aegypti at the lowest administrative level (ie, all 313 districts in Angola and communes in the DR Congo).14, 15 These estimates combine climatic and socioeconomic variables to produce an estimate of vector suitability and have previously been used to estimate the time-varying risk of the emergence and spread of arboviruses.16, 17 Human movement data and models We considered eight connectivity metrics in our model, all of which capture different aspects of connectivity between two districts, shown to be useful for inferring regular daily commuting patterns, longer-term movements, and the general human diffusion process.8, 18 The first set of metrics included distance and travel time distance. For distance, we calculated the great circle distance between the centroids of each pair of districts using data from using data from the Database of Global Administrative Areas. For travel time distance, we calculated the travel time between each pair of districts.19 This metric represents the underlying transport network, which has been shown to reflect population mixing (appendix p 7).16

ds of each pair of districts using data from using data from the Database of Global Administrative Areas. For travel time distance, we calculated the travel time between each pair of districts.19 This metric represents the underlying transport network, which has been shown to reflect population mixing (appendix p 7).16 We derived a second set of metrics from standard models describing human mobility. These were the gravity model, the radiation model, and adjacency network metrics. The gravity model assumes that the relative population flow between districts is a log-linear function of the population sizes of the districts and the distance between them (functional form shown in the appendix).20 This model therefore emphasises the attractive power of large population centres such as the capital cities Luanda and Kinshasa (appendix p 6). The radiation model additionally takes into account the draw from other populations within the same radius of the districts considered, as well as the population sizes and distance of the origin and destination locations (functional form shown in the appendix).18, 21 This model therefore reflects recurrent workplace commuting, assuming that every locality has a competing underlying attractiveness. Adjacency network metrics encode the number of district borders an individual would need to cross to move from one district to another. This metric thus reflects the effects of national and sub-national borders on movement within the region. We then disaggregated this adjacency matrix into three binary connectivity matrices with connectivity degrees of one (ie, districts that share a border), connectivity degrees of two (ie, districts that share a common neighbour), and connectivity degrees of more than two. A full list of the 11 variables used to predict the geographic expansion of yellow fever virus is available in the appendix (p 8).

s with connectivity degrees of one (ie, districts that share a border), connectivity degrees of two (ie, districts that share a common neighbour), and connectivity degrees of more than two. A full list of the 11 variables used to predict the geographic expansion of yellow fever virus is available in the appendix (p 8). To calibrate the gravity and radiation models, we used aggregated and deidentified mobile phone-derived mobility estimates at the constituency level from Namibia between Oct 1, 2010, and Sept 30, 2011. These data measure the proportion of time that unique subscriber identity module (SIM) cards in each constituency spend in all other constituencies, and have been described in detail by Ruktanonchai and colleagues.22 We used data from Namibia because it directly neighbours the study area and has a similar per capita gross domestic product (GDP) to both Angola and the DR Congo and because mobile phone-derived mobility estimates were not available from the study region at the time of analysis. We then used the models to predict between-district mobility for Angola and the DR Congo using the movement package in R. We computed national adjacency networks using administrative boundary data from the Database of Global Administrative Areas dataset. For each of the 313 administrative regions in Angola and the DR Congo, we calculated the total human population size using gridded population estimates based on data from WorldPop.

in R. We computed national adjacency networks using administrative boundary data from the Database of Global Administrative Areas dataset. For each of the 313 administrative regions in Angola and the DR Congo, we calculated the total human population size using gridded population estimates based on data from WorldPop. We tested the hypothesis that the rate of spatial expansion changed with time by introducing a time-varying binary variable that was equal to 0 before change point T and equal to 1 after T. In our baseline scenario, we assumed that the week of the change point corresponded to the rollout of mass vaccination in the region (week 9 of 2016), but we also explored alternative scenarios in a sensitivity analysis. Model of geographical spread of yellow fever virus In our model, xi(t) represents the infection status of district at time (ie, a binary variable that takes the value 1 if there were infections that time step and is 0 otherwise). We assume that districts were infected in all time-steps between the first and last reported cases described in the WHO reports. We used a standard logistic model to characterize the probability that district j will become infected at time t: logit(P(xj(t)=1|xj(t-1)=0))=β0+∑k=1nβkYj,tk+ɛ Yk corresponds to explanatory variable k and ε is an error distributed by the standard logistic distribution. Explanatory variables included in this analysis are vector suitability and whether infection occurred after vaccine rollout. Connectivity metrics that quantifies the connectivity between districts i and j were denoted as follows: Aij(k)

Yk corresponds to explanatory variable k and ε is an error distributed by the standard logistic distribution. Explanatory variables included in this analysis are vector suitability and whether infection occurred after vaccine rollout. Connectivity metrics that quantifies the connectivity between districts i and j were denoted as follows: Aij(k) For each of these metrics, we derived the global force of infection exerted from all infected districts to j: Yj,tk=∑iAij(k)xi(t-1) We used our model for the geographic spread of yellow fever virus to investigate the contribution of single variables in a univariate analysis and then in a multivariable framework. To assess the model accuracy, we calculated the probability of invasion pW predicted by the model for each district-week that had not yet reported cases. We then partitioned district–weeks into eight groups with predicted probability in the range 0–1%, 1–5%, 5–10%, 10–15%, 15–20%, 20–25%, 25–35%, and 35–100%. For each group, we calculated the mean predicted probability and compared it with the proportion of district–weeks in the group in which invasion effectively took place. Because there are 4·33 weeks per month, the monthly probability of invasion can be calculated with the following formula: pM=1-(1-pw)4.33

To assess the model accuracy, we calculated the probability of invasion pW predicted by the model for each district-week that had not yet reported cases. We then partitioned district–weeks into eight groups with predicted probability in the range 0–1%, 1–5%, 5–10%, 10–15%, 15–20%, 20–25%, 25–35%, and 35–100%. For each group, we calculated the mean predicted probability and compared it with the proportion of district–weeks in the group in which invasion effectively took place. Because there are 4·33 weeks per month, the monthly probability of invasion can be calculated with the following formula: pM=1-(1-pw)4.33 Application of the model In the context of finite resources, we assessed how this analytical framework could have helped to inform the prioritisation of districts to be targeted for intervention at the time when such insight was needed—ie, on week 9 of 2016 when the vaccination campaign started. We fitted the model to data available up until that week, then ran 1000 simulations of the model to infer the probability of invasion in the subsequent month for each district. In each simulation, a district can become infected in a certain week and then contribute to invasion risk in that particular simulation. We assumed that invaded districts would not become disease-free within the short simulation period. Assuming that only n districts can be targeted for intervention, we compared the performance of a prioritisation strategy that targeted the n districts with the highest predicted invasion probability versus a strategy that targeted n districts chosen at random.

ot become disease-free within the short simulation period. Assuming that only n districts can be targeted for intervention, we compared the performance of a prioritisation strategy that targeted the n districts with the highest predicted invasion probability versus a strategy that targeted n districts chosen at random. For forward prediction, we used the same approach to predict the possible expansion of yellow fever virus 1–2 months ahead. We fitted the model to the entire epidemic from week 49 in 2015 to late August, 2016, then used the model parameters to simulate the geographic spread forward in time. We estimated the exponential growth rate and the doubling time of the epidemic in the early phase when the number of cases was growing exponentially. We reconstructed the generation time distribution of yellow fever virus (ie, the time lag from the infection of a case to infection of subsequent cases in the chain of transmission) from previous knowledge of the natural history of yellow fever virus infection in people and in mosquitoes. We then derived the reproduction number from standard formula linking it to the exponential growth rate and the generation time distribution.23 Technical details are available in the appendix.

in of transmission) from previous knowledge of the natural history of yellow fever virus infection in people and in mosquitoes. We then derived the reproduction number from standard formula linking it to the exponential growth rate and the generation time distribution.23 Technical details are available in the appendix. Finally, we assessed the sensitivity of our model to time-varying reporting rates. Our spatial model is based on a detailed analysis of the dates when districts reported their first yellow fever virus case. If reporting of cases increased during the course of the epidemic, the delay between the first infection and the first reported case should have shortened with time. We therefore did a simulation study in which, for each district, the date of the first infection was reconstructed from the date of the first report under the assumption of an average delay of 4 weeks in the early phase of the epidemic (until week 9 of 2016) and of 1 week in subsequent weeks. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Finally, we assessed the sensitivity of our model to time-varying reporting rates. Our spatial model is based on a detailed analysis of the dates when districts reported their first yellow fever virus case. If reporting of cases increased during the course of the epidemic, the delay between the first infection and the first reported case should have shortened with time. We therefore did a simulation study in which, for each district, the date of the first infection was reconstructed from the date of the first report under the assumption of an average delay of 4 weeks in the early phase of the epidemic (until week 9 of 2016) and of 1 week in subsequent weeks. Role of the funding source The funder of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results The first cases of yellow fever were reported in Luanda, the capital of Angola, in the first week of December, 2015. Between week 1 and week 5 of 2016, the number of cases grew exponentially, with a doubling time of 6 days (95% CI 5–7; figure 1A). Under the assumptions that reporting remained stable during this period and that the generation time had a mean of 15 days (SD 6; appendix), we estimated that the reproduction number of yellow fever virus was 4·8 (95% CI 4·0–5·6).

mber of cases grew exponentially, with a doubling time of 6 days (95% CI 5–7; figure 1A). Under the assumptions that reporting remained stable during this period and that the generation time had a mean of 15 days (SD 6; appendix), we estimated that the reproduction number of yellow fever virus was 4·8 (95% CI 4·0–5·6). Although the epidemic was initially focused in Luanda, it quickly expanded to other districts, including Belas, Lobito, and Huambo (figure 1B and figure 2A). By the end of February, 49 districts had reported cases (figure 1B) with the proportion of cases from Luanda dropping from 75% in early February, 2016 (week 5), to about 30% in April, 2016 (week 14; figure 1A). In a simple univariate correlation analysis, we found that earlier timing of yellow fever virus invasion into a district was correlated with greater population density (Pearson's r 0·52, 95% CI 0·34–0·66; appendix pp 6, 9). Additionally, the further away districts were from Luanda by distance or travel time, the longer they took to be invaded (Pearson's r for distance 0·60, 0·52–0·66; Pearson's r for travel time 0·63, 0·56–0·69; appendix p 10). By contrast, early invasion was not correlated with A aegypti suitability, which is abundant in most of the study area (appendix p 13).14

ere from Luanda by distance or travel time, the longer they took to be invaded (Pearson's r for distance 0·60, 0·52–0·66; Pearson's r for travel time 0·63, 0·56–0·69; appendix p 10). By contrast, early invasion was not correlated with A aegypti suitability, which is abundant in most of the study area (appendix p 13).14 We also investigated the duration that the virus persisted in a district after invasion. In a survival analysis, the probability that a district was still reporting cases was 0·94 (95% CI 0·84–1·00) at 17 weeks after the first report, 0·88 (0·73–1·00) at 18 weeks, and 0·50 (0·26–0·98) at 20 weeks (appendix p 9). In a Cox model, we estimated that the hazard ratio for cases ceasing was 0·74, 95% CI 0·13–0·92, per log-unit increase in the population size of a district. The duration for which a district reported cases was not significantly associated with the date when it was first invaded.

(0·26–0·98) at 20 weeks (appendix p 9). In a Cox model, we estimated that the hazard ratio for cases ceasing was 0·74, 95% CI 0·13–0·92, per log-unit increase in the population size of a district. The duration for which a district reported cases was not significantly associated with the date when it was first invaded. We then fitted the geographic spread model to all data available (ie, Dec 4, 2015, to Aug 4, 2016). The terms retained in the model were, first, the adjacency metric (more than two degrees away); second, the binary variable defining whether the invasion occurred before or after vaccination rollout at week 9 of 2016; third, A aegypti suitability; fourth, the gravity human mobility metric; and fifth, the radiation human mobility metric. In our assessment of model accuracy, we found excellent agreement between the probability of invasion per week per district, as predicted by the model, and the observed proportion invaded (figure 3A). The model successfully discriminated district–weeks with a high probability of invasion from district–weeks with a lower probability of invasion (area under the curve 0·94, 95% CI 0·92–0·97; appendix p 16). For example, 31 (0·31%) district-weeks had a predicted weekly probability of invasion greater than 25%, corresponding to an average per month probability of invasion of 77% (figure 3A). By contrast, 8754 (88·5%) district–weeks had a weekly invasion probability less than 1% (figure 3A; appendix p 11). If we judge the expanding phase of the outbreak to be from late January to late March, then there was also a high correlation between the invasion of yellow fever virus and the invasion probability predicted by the model (0·78, 95% CI 0·74–0·82).

ad a weekly invasion probability less than 1% (figure 3A; appendix p 11). If we judge the expanding phase of the outbreak to be from late January to late March, then there was also a high correlation between the invasion of yellow fever virus and the invasion probability predicted by the model (0·78, 95% CI 0·74–0·82). We noted that spatial expansion slowed down significantly during the course of the epidemic (figure 1B). If we assumed that the change occurred on week 9 of 2016, when vaccination was rolled out, the odds ratio for a district being invaded after the change point T versus before T was 0·948 (95% CI 0·946–0·949; appendix pp 17, 18). However, the proportion of deviance explained was slightly higher if we used the assumption that the change occurred 2–3 weeks before vaccination roll out (33·6% for 2 or 3 weeks before week 9 vs 30·8% for week 9; appendix p 18).

ter the change point T versus before T was 0·948 (95% CI 0·946–0·949; appendix pp 17, 18). However, the proportion of deviance explained was slightly higher if we used the assumption that the change occurred 2–3 weeks before vaccination roll out (33·6% for 2 or 3 weeks before week 9 vs 30·8% for week 9; appendix p 18). In the context of finite resources, insights into district-specific real-time invasion risks might help to inform the prioritisation of districts to be targeted for interventions such as vaccination and vector control. As an example, we considered the situation in mid-February, 2016, when the epidemic was still expanding (32 districts were invaded in the next month). If, on the basis of data available up to mid-February, the 20 districts with the highest probability of invasion had been targeted for vaccination, then 13 (41%) of the 32 districts that were actually invaded would have been targeted, whereas this number would have increased to 17 (53%) if the 30 districts with the highest risk had been targeted and 27 (84%) if the 50 districts with the highest probability had been targeted (figure 3B; appendix p 12). The number of districts successfully targeted would drop to two (6%), three (9%), and six (19%), respectively, if prioritisation had been random.

(53%) if the 30 districts with the highest risk had been targeted and 27 (84%) if the 50 districts with the highest probability had been targeted (figure 3B; appendix p 12). The number of districts successfully targeted would drop to two (6%), three (9%), and six (19%), respectively, if prioritisation had been random. Univariate models tended to be worse than the full model when we compared the models' predicted probability of invasion versus the observed proportion of districts invaded (figure 3A; appendix p 11). However, we were encouraged to see that simpler models that included only travel distance or neighbourhood effects were able to discriminate similarly well in terms of their ranking of locations that might become affected (appendix p 14). We then used the full model to predict the future spread of yellow fever virus in Angola and the DR Congo (figure 4). The most recent confirmed cases of yellow fever are from Kinshasa in the DR Congo, and we predicted that future spread would mostly occur along the road east to Kananga and south to the border with Angola where cases have been identified previously (figure 4; appendix pp 6, 7). The two locations in Angola with the highest predicted risk of introduction are Uige and Luanda (figure 4). Rural regions in Angola and the DR Congo had low predicted probabilities of yellow fever virus introduction.

uth to the border with Angola where cases have been identified previously (figure 4; appendix pp 6, 7). The two locations in Angola with the highest predicted risk of introduction are Uige and Luanda (figure 4). Rural regions in Angola and the DR Congo had low predicted probabilities of yellow fever virus introduction. We investigated the robustness of the variables retained in the final model to the time when the analysis was done. The gravity metric was the variable selected most often at 21 out of 24 weeks compared with 17 times for the radiation metric and 19 times for the aedes suitability surface (appendix p 15). However, there was heterogeneity as to when these variables were selected. For example, the gravity metric seemed to be important in the early phase of the outbreak and radiation metric seemed to be important in the transitioning phase. In a sensitivity analysis, we noted that model parameters would remain almost identical and the predictive accuracy would remain very high if the delay from the first infection to the first reported case in a district had shortened from 4 weeks to 1 week during the course of the epidemic (appendix pp 16, 17).

ng phase. In a sensitivity analysis, we noted that model parameters would remain almost identical and the predictive accuracy would remain very high if the delay from the first infection to the first reported case in a district had shortened from 4 weeks to 1 week during the course of the epidemic (appendix pp 16, 17). Discussion In the context of rapidly spreading and potentially fatal infectious disease epidemics for which vaccine stockpiles are limited, such as the current yellow fever outbreak in central Africa, it is essential to establish which areas are at the greatest risk of infection to inform vaccine prioritisation decisions. This risk estimation requires an in-depth understanding of the determinants of disease spread. We integrated and analysed diverse datasets, including the size and mobility of human populations and detailed maps of vector suitability, to describe the epidemic and spatial spread of yellow fever virus. We found that the spatial spread of yellow fever virus was well explained by human mobility and vector suitability and that our approach could help to discriminate districts at high risk of invasion from others at lower risk.

maps of vector suitability, to describe the epidemic and spatial spread of yellow fever virus. We found that the spatial spread of yellow fever virus was well explained by human mobility and vector suitability and that our approach could help to discriminate districts at high risk of invasion from others at lower risk. We estimated that the reproduction number of yellow fever virus was 4·8, which is in line with other estimates available for yellow fever virus.24 This suggests a critical vaccination coverage for yellow fever virus of about 80%. However, we cannot rule out the possibility that in the early stage of the epidemic, part of the rise in case counts was caused by increased reporting, which could bias estimates of the reproduction number (and the critical vaccination coverage) upwards. Although we estimated the reproduction number from national data, it will be interesting to assess transmission dynamics at a refined local level.

ic, part of the rise in case counts was caused by increased reporting, which could bias estimates of the reproduction number (and the critical vaccination coverage) upwards. Although we estimated the reproduction number from national data, it will be interesting to assess transmission dynamics at a refined local level. We characterised various factors that affect the spread of yellow fever virus in central Africa. For example, we found that human mobility was an important predictor of the spatial expansion of yellow fever virus. The importance of large population centres in driving the expansion dynamics during the early stage of the epidemic was apparent in both the univariate and multivariate models, and the gravity metric played a key part early in the epidemic (figure 2A and 2B; appendix p 6). Our results from the univariate analysis also suggest that a substantial neighbourhood effect existed during this outbreak—ie, places closer to the outbreak were more likely to become infected than were those further away, which was exemplified by the adjacency and travel distance metric models performing better than the other univariate models (appendix p 11). Univariate models were not well suited to capture the dynamics of the spread that include short and long distance introductions of the pathogen, such as the cases reported in Kinshasa, DR Congo, most of which are connected to Luanda, Angola. The inclusion of the spatial distribution of A aegypti suitability substantially increased the fit of the full model, but this variable did not show significant associations with early yellow fever virus invasion (appendix p 11). Vector suitability is mostly constant across the region, which might result in a lower predictive power for that variable (appendix p 13).14 More studies are needed to measure the mosquito per human ratio in Angola and the DR Congo and to quantify their spatial heterogeneity (appendix p 13). Our data also suggested that epidemics might last longer in areas that are more densely populated (figure 2B; appendix p 6). This effect would support estimates of other arboviruses such as dengue virus, for which urban areas have been shown to confer substantially higher reproduction numbers because of a higher person-to-person contact rate.16

at epidemics might last longer in areas that are more densely populated (figure 2B; appendix p 6). This effect would support estimates of other arboviruses such as dengue virus, for which urban areas have been shown to confer substantially higher reproduction numbers because of a higher person-to-person contact rate.16 In our analysis, the rate of spatial expansion of yellow fever virus declined in February, 2016, which roughly coincides with the start of the vaccination campaign. Although the observation of such temporal association is interesting, our statistical framework does not allow us to show a causal link or to quantify the proportion of the decline that was attributable to vaccination versus other causes. There is weak evidence in the data that spatial expansion might have started to decline even before vaccination started (appendix p 18). However, the current absence of data describing the vaccination campaign makes more detailed and a definitive assessment of its effects on spread difficult. Future research should aim to further characterise this impact from more detailed data documenting both case counts and vaccine distribution over space and time. This future work should assess the impact of vaccination on both spatial expansion (ie, ability of yellow fever virus to spread from district to district) and on local transmission (ie, ability of yellow fever virus to generate outbreaks within districts).

both case counts and vaccine distribution over space and time. This future work should assess the impact of vaccination on both spatial expansion (ie, ability of yellow fever virus to spread from district to district) and on local transmission (ie, ability of yellow fever virus to generate outbreaks within districts). In addition to understanding the drivers of the spread of yellow fever virus transmission, we tested our statistical model's predictive power in real time. During the expanding phase of the outbreak, the model could help to predict the invasion of yellow fever virus in the region and to identify which districts should be targeted for intervention (figure 3B). This result is useful in a context in which resources and vaccine stock are finite. We emphasise the need to reparameterise the model because the importance of different ecological and mobility factors might vary depending on the week of the epidemic. Fortunately, the outbreak is now contained and no confirmed cases have been reported since July 12, 2016.7 Nevertheless, our approach presents a flexible framework that can be readily updated and could therefore be used to predict the geographic spread of future epidemics of yellow fever virus or other related infectious diseases. However, to translate the insight generated by such a modelling approach into a concrete vaccination strategy, complex logistic constraints would also need to be accounted for, such as supply and delivery. These important issues are beyond the scope of this study. Our estimates of characteristics of spatial spread might be affected if the reporting of cases varies spatially.

elling approach into a concrete vaccination strategy, complex logistic constraints would also need to be accounted for, such as supply and delivery. These important issues are beyond the scope of this study. Our estimates of characteristics of spatial spread might be affected if the reporting of cases varies spatially. During the past 10 years or so, mosquito-borne diseases such as those caused by dengue virus, chikungunya virus, and more recently Zika virus, have been expanding geographically. By contrast, yellow fever virus was previously thought to be on the decline because of the positive effects of vaccination campaigns. However, the recent transmission of yellow fever virus in central Africa has shown the continued threat that yellow fever virus poses to human populations, as well as the need for continued vaccination campaigns and close monitoring of sporadic outbreaks in rural regions. Previous work has also shown that vaccination coverage in the region analysed here is low and that there was no significant spatial heterogeneity between districts.6 In the wider context, increased urbanisation and human mobility in the region could lead to future increases in the risk of emergence and rapid expansion of such outbreaks as the current yellow fever epidemic.

in the region analysed here is low and that there was no significant spatial heterogeneity between districts.6 In the wider context, increased urbanisation and human mobility in the region could lead to future increases in the risk of emergence and rapid expansion of such outbreaks as the current yellow fever epidemic. At this stage, because of the absence of available data, our analysis cannot address the yellow fever virus outbreak from a genetic perspective. However, our results provide a baseline to which future viral genetic results could be directly added. Genomic surveillance could be used to extend our results, by helping to identify the origin of the outbreak,25 monitoring the diversity and spatial distribution of circulating viruses,26 and characterising signatures of host adaptation.27 Two past outbreaks of yellow fever virus have been recorded in Angola, in 1971 and 1988, and a single genetic isolate indicates that the 1971 outbreak strain belongs to the east and central African genotype. Although genetic data for the current outbreak in central Africa have yet to be reported, preliminary analysis of available yellow fever virus genomes from travellers returning from the affected region suggest that the current outbreak also belongs to the east and central African lineage.28 Whether the virus has been circulating in a sylvatic cycle in rural areas, or whether cryptic circulation has been occurring in people in the region for at least 28 years are questions that require further investigation.

ed region suggest that the current outbreak also belongs to the east and central African lineage.28 Whether the virus has been circulating in a sylvatic cycle in rural areas, or whether cryptic circulation has been occurring in people in the region for at least 28 years are questions that require further investigation. Because no weekly case count data are available from outside the capitals of Angola and the DR Congo, we were restricted to modelling the potential for location-specific yellow fever virus introduction and circulation, rather than predicting cumulative numbers of expected cases and the quantitative effect of intervention strategies and climatic variables such as rainfall and precipitation. We were also further restricted to using suspected and confirmed cases, because the proportion of cases that have been laboratory tested remained small.29 Although our model performs well in predicting the expansion of yellow fever virus in the region, estimates might be improved by the addition of country specific real-time mobility estimates from mobile phone data or other sources, such as travel surveys. Such data could inform which populations contribute most to the invasion process and might help the targeting of intervention strategies.

ver virus in the region, estimates might be improved by the addition of country specific real-time mobility estimates from mobile phone data or other sources, such as travel surveys. Such data could inform which populations contribute most to the invasion process and might help the targeting of intervention strategies. Beyond local spread, there is a risk of international spread of the virus via air travel, as seen with the yellow fever virus cases that have been detected in travellers returning from Angola to China.28 In the present study, we refrained from extrapolating our results to regions other than the two that are currently at the core of the affected countries because evidence of overland cross-border local transmissions has, so far, only been reported between Angola and the DR Congo; frequent exchange along the major road from Uige, Angola, to Kinshasa, DR Congo, has been recorded (figure 4). Additionally, the Angolan district of Cabinda is located entirely within the DR Congo, which results in substantial population flows through the western coastal corridor of the DR Congo. Spread to other neighbouring countries such as Namibia, Zambia, or the Central African Republic might be possible but has not been recorded as part of the current outbreak.

abinda is located entirely within the DR Congo, which results in substantial population flows through the western coastal corridor of the DR Congo. Spread to other neighbouring countries such as Namibia, Zambia, or the Central African Republic might be possible but has not been recorded as part of the current outbreak. The identification of locations expected to be invaded by the virus has broad implications for disease control, specifically for the prioritisation of where and when treatment and prevention measures would be best implemented to prevent subsequent rapid geographic spread of a pathogen.30 The importance of these techniques based on historical data has been shown for other diseases. However, our modelling techniques in this study allow the analysis of near real-time data to inform the control of an ongoing outbreak, constituting a methodological advance that is applicable to other diseases. For the WorldPop project see http://www.worldpop.org.uk/ Supplementary Material Supplementary appendix

The identification of locations expected to be invaded by the virus has broad implications for disease control, specifically for the prioritisation of where and when treatment and prevention measures would be best implemented to prevent subsequent rapid geographic spread of a pathogen.30 The importance of these techniques based on historical data has been shown for other diseases. However, our modelling techniques in this study allow the analysis of near real-time data to inform the control of an ongoing outbreak, constituting a methodological advance that is applicable to other diseases. For the WorldPop project see http://www.worldpop.org.uk/ Supplementary Material Supplementary appendix Acknowledgments MUGK receives funding from the International research Consortium on Dengue Risk Assessment Management and Surveillance (IDAMS; European Commission 7th Framework Programme [21893]). NG is supported by a University of Melbourne McKenzie fellowship. MAJ received partial support from the Models of Infectious Disease Agent Study program (Cooperative Agreement number 1U54GM088558). FMS acknowledges funding from the Rhodes Trust. SIH received a grant from the Research for Health in Humanitarian Crises (R2HC) Programme, managed by ELRHA (number 13468), which also supported MUGK. The Research for Health in Humanitarian Crises (R2HC) programme aims to improve health outcomes by strengthening the evidence base for public health interventions in humanitarian crises. The £8 million R2HC programme is funded equally by the Wellcome Trust and Department for International Development, with Enhancing Learning and Research for Humanitarian Assistance (ELRHA) overseeing the programme's execution and management. TdO is funded by a Flagship Grant from the Medical Research Council (MRC) of the Republic of South Africa (MRC-RFA-UFSP-01-2013/UKZN HIVEPI), by the VIROGENESIS project European Union's Horizon 2020 (number 634650), and a Royal Society Newton Advanced Fellowship. SIH is funded by a Senior Research Fellowship from the Wellcome Trust (number 095066), and grants from the Bill & Melinda Gates Foundation (OPP1119467, OPP1093011, OPP1106023, and OPP1132415). HHN and BSRP are funded by the European Research Council through the Advanced Investigator Grant Momentum (number 324247). EON is supported by a grant from the National Institutes of Health (number K01ES025438). DLS and AJT are funded by the National Institutes of Health and National Institute of Allergy and Infectious Diseases (number U10AI089674), and the Bill & Melinda Gates Foundation (AJT: number OPP1106427 and 1032350; DLS: number OPP1110495). AJT is also supported by a Wellcome Trust Sustaining Health Grant (number 10688/Z/15/Z). JSB acknowledges funding from the National Institutes of Health (#5R01LM010812-06).

and Infectious Diseases (number U10AI089674), and the Bill & Melinda Gates Foundation (AJT: number OPP1106427 and 1032350; DLS: number OPP1110495). AJT is also supported by a Wellcome Trust Sustaining Health Grant (number 10688/Z/15/Z). JSB acknowledges funding from the National Institutes of Health (#5R01LM010812-06). This study was made possible by the support of the American people through the United States Agency for International Development Emerging Pandemic Threats Program-2 PREDICT-2 (Cooperative Agreement number AID-OAA-A-14-00102), which supports OGP and MUGK. SC acknowledges funding from the French Government's Investissement d'Avenir program, Laboratoire d'Excellence “Integrative Biology of Emerging Infectious Diseases” (number ANR-10-LABX-62-IBEID), the NIGMS MIDAS initiative, the AXA Research Fund, and the European Union Seventh Framework Programme (FP7/2007-2013) under Grant Agreement 278433 (PREDEMICS, EU program ZikAlliance). Contributors MUGK and SC developed the idea and design of the study. MUGK, BN, SS, HS, OF, FMS, NRM, and MAJ gathered the data. MUGK, SC, RCR, and BN analysed the data. MUGK, SC, RCR, NRF, NG, GRWW, OF, SCH, MN, DB, RNT, HHN, BSRP, NT, EON, NRF, JSB, AJT, KK, TdO, DLS, OGP, AAS, IIB, and SIH contributed to the data interpretation. MUGK and SC wrote the first draft of the manuscript. All authors contributed to writing and approved the final version of the manuscript. Declaration of interests We declare no competing interests. Figure 1 Number of cases in Angola and geographic spread of the epidemic

Contributors MUGK and SC developed the idea and design of the study. MUGK, BN, SS, HS, OF, FMS, NRM, and MAJ gathered the data. MUGK, SC, RCR, and BN analysed the data. MUGK, SC, RCR, NRF, NG, GRWW, OF, SCH, MN, DB, RNT, HHN, BSRP, NT, EON, NRF, JSB, AJT, KK, TdO, DLS, OGP, AAS, IIB, and SIH contributed to the data interpretation. MUGK and SC wrote the first draft of the manuscript. All authors contributed to writing and approved the final version of the manuscript. Declaration of interests We declare no competing interests. Figure 1 Number of cases in Angola and geographic spread of the epidemic (A) Epidemic curve for suspected and confirmed cases in Angola and Luanda from Dec 1, 2015, to Aug 25, 2016. (B) Fit of exponential growth during the early phase of the epidemic. (C) Number of districts affected during each week over the course of the outbreak. Figure 2 Timing of the introduction of yellow fever virus and duration of infection (A) Timing of the introduction of yellow fever virus to each district starting from the origin of the outbreak in Luanda, Angola. Colouring shows the weeks until the first case was reported. (B) Duration of transmission. Figure 3 Model accuracy and real-time prediction of the yellow fever virus invasion model

Figure 2 Timing of the introduction of yellow fever virus and duration of infection (A) Timing of the introduction of yellow fever virus to each district starting from the origin of the outbreak in Luanda, Angola. Colouring shows the weeks until the first case was reported. (B) Duration of transmission. Figure 3 Model accuracy and real-time prediction of the yellow fever virus invasion model (A) Model prediction accuracy as assessed by comparing the predicted invasion probability from the geographic spread model with the observed proportion of districts that became invaded; numbers represent district–weeks. (B) Comparisons between district targeting based on real-time modelling analysis vs random targeting during the expansion phase of the outbreak between mid-March and mid-April, during which 32 districts were newly invaded. Figure 4 Model-based predictions of yellow fever virus spread Maps show model-based predictions for the invasion of yellow fever virus in central Africa originating from Kinshasa, the location with the latest reported cases, at 4 weeks (A) and 8 weeks (B) ahead of the last case onset date, July 12, 2016. Colours represent weekly probability of invasion.

Introduction Tackling malaria remains a high priority internationally, with elimination and eradication back on the global agenda.1 In the past century, more than 50 countries have succeeded in eliminating the disease. Nevertheless, although malaria is no longer endemic in these nations, increasing travel to endemic areas in recent decades means that the malaria-endemic world is becoming increasingly connected by population movements,2 and that imported malaria cases remain common.3 Such cases continue to pose challenges for diagnosis and management, with malaria remaining an infrequently encountered disease for many physicians in non-endemic areas,4 where it can be expensive to treat5 and result in high mortality.6 Moreover, with anopheles vectors still present in many non-endemic countries, imported cases can also cause secondary transmission,7 although the chances of resumption of endemic transmission are very small.8 Finally, although the effects of imported malaria on malaria-free countries are problematic, data on the features of imported cases can also provide valuable information about both the epidemiology of malaria in endemic regions where surveillance systems are weak, and on how malaria moves around the world.2

ery small.8 Finally, although the effects of imported malaria on malaria-free countries are problematic, data on the features of imported cases can also provide valuable information about both the epidemiology of malaria in endemic regions where surveillance systems are weak, and on how malaria moves around the world.2 The timing, number, and origin of imported malaria cases into non-endemic regions vary by country and are a function of several factors including the transmission intensity of the origin location, the number of people visiting that location, the activities undertaken in the location, and prophylaxis availability and adherence.3, 9, 10, 11 Depending on the country, some demographic groups have substantially higher infection rates. For instance, malaria imported to Europe is often reported in travellers returning from (or migrants coming from) endemic areas and migrants living in Europe returning from visiting friends and relatives, with children who are visiting friends and relatives being particularly at risk.12

antially higher infection rates. For instance, malaria imported to Europe is often reported in travellers returning from (or migrants coming from) endemic areas and migrants living in Europe returning from visiting friends and relatives, with children who are visiting friends and relatives being particularly at risk.12 Information about imported malaria is to an extent captured by national authorities where, for most high-income countries, malaria is a notifiable disease. However, under-reporting is probably common; for instance, WHO reported that 6244 cases of malaria were imported to Europe in 2010, but the true number might be six times higher.13 Such deficiencies have prompted the initiation of surveillance networks such as GeoSentinel14 and EuroTravNet,15, 16 which now play a key part in the surveillance of imported malaria, in the identification of changing trends in malaria importation, in tracking drug-resistance patterns, and in establishing the changing profile of malaria risk at traveller destinations. Nevertheless, nationally reported data continue to be widely collected and still provide valuable information about the trends, composition, and drivers of imported malaria for most non-endemic countries, with annual data compilations and analyses of statistics providing the main source of information guiding national policies on imported malaria. However, a contemporary global assembly of such nationally reported data, and assessment of patterns and variations has not previously been undertaken. Research in context Evidence before this study

Information about imported malaria is to an extent captured by national authorities where, for most high-income countries, malaria is a notifiable disease. However, under-reporting is probably common; for instance, WHO reported that 6244 cases of malaria were imported to Europe in 2010, but the true number might be six times higher.13 Such deficiencies have prompted the initiation of surveillance networks such as GeoSentinel14 and EuroTravNet,15, 16 which now play a key part in the surveillance of imported malaria, in the identification of changing trends in malaria importation, in tracking drug-resistance patterns, and in establishing the changing profile of malaria risk at traveller destinations. Nevertheless, nationally reported data continue to be widely collected and still provide valuable information about the trends, composition, and drivers of imported malaria for most non-endemic countries, with annual data compilations and analyses of statistics providing the main source of information guiding national policies on imported malaria. However, a contemporary global assembly of such nationally reported data, and assessment of patterns and variations has not previously been undertaken. Research in context Evidence before this study The rise of air travel in the past century has resulted in a highly interconnected world, where geographical distance is becoming less of a barrier to pathogen dispersal. Malaria-free countries that were once endemic to the disease still report thousands of cases every year through importation, resulting in deaths, health system burden, and occasional secondary transmission. The disease is notifiable in most high-income countries, providing data on case numbers and characteristics to direct mitigation strategies. Multiple individual studies at national scales over the past decade have highlighted the substantial heterogeneities in imported malaria numbers, rates, and risks that exist across geographies, demographics, and malaria species, but have not been brought together in a single study.

tics to direct mitigation strategies. Multiple individual studies at national scales over the past decade have highlighted the substantial heterogeneities in imported malaria numbers, rates, and risks that exist across geographies, demographics, and malaria species, but have not been brought together in a single study. Added value of this study This study reports the first global collection of nationally reported imported malaria statistics in 20 years. Moreover, it presents the first global assessment of the geographical features of imported malaria, including patterns of flows, connectivity, species distributions, and diagnostic capacities. Our findings show how limits to malaria dispersal remain and how clear patterns in movements exist that have never been quantified before, with the architecture of the air network, historical ties, demographics of travellers, and malaria endemicities all contributing to highly heterogeneous patterns of numbers, routes, and species compositions of parasites transported. Implications of all the available evidence

This study reports the first global collection of nationally reported imported malaria statistics in 20 years. Moreover, it presents the first global assessment of the geographical features of imported malaria, including patterns of flows, connectivity, species distributions, and diagnostic capacities. Our findings show how limits to malaria dispersal remain and how clear patterns in movements exist that have never been quantified before, with the architecture of the air network, historical ties, demographics of travellers, and malaria endemicities all contributing to highly heterogeneous patterns of numbers, routes, and species compositions of parasites transported. Implications of all the available evidence With global malaria eradication on the international agenda, the threat of spreading drug resistance, and the continued burden of imported cases to non-endemic countries, understanding and measuring the patterns of malaria connectivity and their drivers is increasing in importance for designing mitigation, control, and elimination strategies. Prioritising surveillance and control efforts to high-traffic routes and highly connected locations, as well as coordinating elimination efforts around highly connected regional groupings of countries is likely to be the most effective and efficient approach.

r designing mitigation, control, and elimination strategies. Prioritising surveillance and control efforts to high-traffic routes and highly connected locations, as well as coordinating elimination efforts around highly connected regional groupings of countries is likely to be the most effective and efficient approach. Here, we describe the first global assembly of nationally reported imported malaria data in 20 years,11 and the geographic analysis of these data from 40 non-endemic countries in the past 10 years. We map the system of transmission from endemic to non-endemic areas to provide insights into the underlying dynamics of the system. Exploration of the driving factors behind these patterns is beyond the scope of this study. We examine the rates of flow of cases from endemic to non-endemic countries and their inherent spatial patterns. Moreover, we map the species composition of cases by sending and receiving regions. Finally, we discuss candidate factors that shape the recorded patterns and likely future trends. Methods Full details of the process of constructing a library of imported malaria statistics and data extractions are provided in appendix (pp 2–8). Here we provide brief details of the steps taken in assembly and analysis of the data.

Here, we describe the first global assembly of nationally reported imported malaria data in 20 years,11 and the geographic analysis of these data from 40 non-endemic countries in the past 10 years. We map the system of transmission from endemic to non-endemic areas to provide insights into the underlying dynamics of the system. Exploration of the driving factors behind these patterns is beyond the scope of this study. We examine the rates of flow of cases from endemic to non-endemic countries and their inherent spatial patterns. Moreover, we map the species composition of cases by sending and receiving regions. Finally, we discuss candidate factors that shape the recorded patterns and likely future trends. Methods Full details of the process of constructing a library of imported malaria statistics and data extractions are provided in appendix (pp 2–8). Here we provide brief details of the steps taken in assembly and analysis of the data. National imported malaria statistics Most non-endemic countries compile their notified cases of malaria into annual summary reports. These reports were assembled from the national laboratories and agencies that compile imported malaria statistics for as many years as available (appendix). We complemented these data with searches on PubMed, Web of Science, Google Scholar, and standard Google search for “imported malaria” and the name of the non-endemic country in question, in both English and the primary language of the country where this was not English. These searches identified a set of additional academic papers and reports documenting imported cases. For each country, where available, we extracted data on the numbers of confirmed cases reported, the year, their likely origin regions or countries, the species of parasite, and the method of diagnosis.

as not English. These searches identified a set of additional academic papers and reports documenting imported cases. For each country, where available, we extracted data on the numbers of confirmed cases reported, the year, their likely origin regions or countries, the species of parasite, and the method of diagnosis. Data processing We constructed a set of broad rules to facilitate data summarisation, exclusion, and processing. These were based on achieving a balance between maintaining a wide representation of data from several countries, time periods, and sources, and implementing some quality control to ensure comparability between datasets and avoid double-counting. We minimised double-counting through examination of data obtained from academic publications—if they were obtained from nationally reported statistics covering the same period as already extracted from national laboratory reports, then the data were not included in analyses. Throughout, we prioritised statistics from national agencies over academic papers, which were used as supplemental data sources to cover missing periods. We did different hierarchies of analysis to enable presentations of outputs where data inclusion criteria were relaxed to enable comparisons across many countries, and criteria were tightened or data were aggregated to facilitate the production of more robust, but less detailed, conclusions.

s to cover missing periods. We did different hierarchies of analysis to enable presentations of outputs where data inclusion criteria were relaxed to enable comparisons across many countries, and criteria were tightened or data were aggregated to facilitate the production of more robust, but less detailed, conclusions. Although the datasets assembled extended from 1960 to the end of 2015, to obtain a contemporary picture while still including a large number of countries and regions, analyses were undertaken only for the most recent 10 years of data. Therefore, data were restricted to 2005–15, using an annual mean of cases across the full 10 years when available, although for some countries, data were only available for less than 5 years of this period. For each endemic exporting country, we aggregated all reported annual mean case numbers exported to non-endemic reporting countries to obtain estimates of the proportions of each parasite species (or mixed infections, when documented as such) exported. Although this averaging masked temporal trends in the data, clear trends over time were not apparent for most countries, and in view of the gaps in publicly available data (appendix), this time window facilitated the inclusion of many more countries than a more constrained one. We constructed origin–destination matrices for the average number of cases per year imported from endemic to non-endemic countries. Many data sources reported exported cases only by large regions, therefore we also constructed a regional version of this matrix to enable the inclusion of more data and thus identify geographical patterns more robustly. We also analysed these data to estimate the aggregate malaria species compositions being exported from endemic countries and imported to non-endemic countries. Where species breakdowns of imported cases were reported, they were aggregated and summarised across the reporting period.17 Similar to the origin–destination matrices, data for species composition for many countries were reported only by origin region; thus data were also aggregated by region to provide larger sample sizes and thus more confidence in estimates of differences between regions by composition.

sed across the reporting period.17 Similar to the origin–destination matrices, data for species composition for many countries were reported only by origin region; thus data were also aggregated by region to provide larger sample sizes and thus more confidence in estimates of differences between regions by composition. Network community detection Communities in a network reflect a group of nodes that are densely connected and separated from the other nodes in the network, and thus they share common properties and have similar roles within the network. By mapping communities on the imported malaria network defined here, we aimed to identify groups of countries that show strong links in terms of movements of infected travellers. Newman and Girvan17 define a modularity score, which is a measure of the strength of a division of a network into communities (groups of countries in this case). The analysis uses a multilevel algorithm for community detection,18 which uses an iterative approach that merges communities to maximise the modularity.

lers. Newman and Girvan17 define a modularity score, which is a measure of the strength of a division of a network into communities (groups of countries in this case). The analysis uses a multilevel algorithm for community detection,18 which uses an iterative approach that merges communities to maximise the modularity. Additional datasets The construction of modelled global Plasmodium falciparum19 and Plasmodium vivax20 parasite prevalence maps enabled simple comparisons to be made with the imported malaria statistics. The datasets were obtained from the Malaria Atlas Project and summarised to a national level using gridded population data from the WorldPop project and the Global Rural Urban Mapping Project to produce a population-weighted mean P falciparum and P vivax prevalence for each country. Additionally, we obtained bilateral data21 for migrations between each pair of endemic and non-endemic countries to enable further comparisons to be made with the number of cases of imported malaria. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication.

Additional datasets The construction of modelled global Plasmodium falciparum19 and Plasmodium vivax20 parasite prevalence maps enabled simple comparisons to be made with the imported malaria statistics. The datasets were obtained from the Malaria Atlas Project and summarised to a national level using gridded population data from the WorldPop project and the Global Rural Urban Mapping Project to produce a population-weighted mean P falciparum and P vivax prevalence for each country. Additionally, we obtained bilateral data21 for migrations between each pair of endemic and non-endemic countries to enable further comparisons to be made with the number of cases of imported malaria. Role of the funding source The funders of the study had no role in study design, data collection, data analysis, data interpretation, or writing of the report. The corresponding author had full access to all the data in the study and had final responsibility for the decision to submit for publication. Results The movement of people with malaria in 2005–15 followed specific routes (figure 1). Of those cases with origin location recorded between 2005 and 2015, more cases were reported in France (2169 cases per year on average) and the UK (n=1898) than any other country (figure 1A), with the USA (n=1511), Italy (n=637), and Germany (n=401) close behind. Most (22 946/24 941 [92%]) exported cases to non-endemic countries originate in west Africa (13 947 [56%]), India (4988 [20%]), east Africa (3242 [13%]), and Papua New Guinea (748 [3%]; figure 1A). Figure 1B shows that the connections between the UK and west Africa, and between France and west Africa are the strongest in terms of annual average numbers of cases moving from endemic to non-endemic countries (2492 cases on average per year, 2005–15), but that many other routes produce an annual average of more than 50 cases reported in non-endemic countries. These include the movements between the USA and India (149 cases per year on average), the USA and west Africa (n=716), the USA and Haiti (n=52), Australia and Papua New Guinea (n=97), and the UK and Pakistan (n=69). By defining the origin–destination pairs of endemic–non-endemic countries and the cases originating and reported in each as a weighted network, the community detection analyses identified sections of this global imported malaria flow matrix that were more strongly connected than others. Although the matrix is incomplete, with country-level imported malaria data unavailable for some non-endemic countries, it is clear that the structure of these mapped network communities is not mainly geographically determined (figure 2), with historical, economic, language, and cultural ties evident. For example, the UK community includes the English-speaking African nations connected to the UK that were its former colonies (including Nigeria, Ghana, Kenya, Uganda, and The Gambia), and French colonial ties are also evident (including Mali, Niger, Chad, CÔte d'Ivoire, Burkina Faso, Benin, Togo, and Madagascar).

ent. For example, the UK community includes the English-speaking African nations connected to the UK that were its former colonies (including Nigeria, Ghana, Kenya, Uganda, and The Gambia), and French colonial ties are also evident (including Mali, Niger, Chad, CÔte d'Ivoire, Burkina Faso, Benin, Togo, and Madagascar). Clear associations exist between the average annual number of outgoing P falciparum cases from endemic to non-endemic countries, P falciparum prevalence in the endemic countries, and the migration flows to non-endemic countries (figure 3). Moreover, the countries with the most migrants residing in non-endemic countries (eg, India, Pakistan, and Nigeria) typically fall further towards the upper-left side of the plot, as larger numbers of travellers (particularly those visiting friends and relatives in endemic countries) are related to larger numbers of imported cases to non-endemic countries. Although there are associations between the numbers of cases and both malaria prevalence in endemic countries and numbers of migrants in non-endemic countries, many other factors play a part, including demographics, levels of prophylaxis and protection use, and travel activities.

mported cases to non-endemic countries. Although there are associations between the numbers of cases and both malaria prevalence in endemic countries and numbers of migrants in non-endemic countries, many other factors play a part, including demographics, levels of prophylaxis and protection use, and travel activities. We analysed the species compositions of cases reported in non-endemic countries by region (figure 4) and nation (figure 5 and table). Reports of cases in non-endemic countries often present species type broken down by origin region rather than country, thus initially we did the species composition analyses at pooled regional level to ensure larger and more stable sample sizes (figure 4). The results emphasise the large variation in the species composition of malaria cases travelling between the different regions of the world. Although these outputs represent a pooling of data of varying numbers, time periods, treatment-seeking behaviours, and diagnostic capacities, the clear and geographically consistent patterns suggest a robustness in the outputs. The dominance of P falciparum from African and Caribbean sources (mean percentage of cases across regions 74·4%) compared with those originating in Central and South America (13·1%) and Asia and Oceania (17·6%) is clear, although no single species in any region has total dominance. This finding is also reflected at national levels (figure 5), with strong geographically coherent patterns recorded, but also a mixed picture in many places, especially in southeast Asia and central America. The species compositions of cases received in each non-endemic country (figure 4 and table) are indicative of each country's connections to endemic regions. For example, the high P falciparum percentage for France results from its strong ties to west Africa (figure 1). European proximity and ties to Africa result in more cases of P falciparum (mean percentage of cases of 65·8%) than in the Americas (41·7%) or the Asia-Pacific region (32·9%), although a divide is clear, with higher proportions of P vivax in eastern compared with western Europe evident (figure 4). Finally, analyses of diagnostic capacities in European countries (appendix pp 9, 10) highlight the growth in capacities in the past decade, with an increasing use of PCR and rapid tests. However, substantial geographical differences remain, with the range of methods and their reporting higher in western Europe than eastern Europe.

lyses of diagnostic capacities in European countries (appendix pp 9, 10) highlight the growth in capacities in the past decade, with an increasing use of PCR and rapid tests. However, substantial geographical differences remain, with the range of methods and their reporting higher in western Europe than eastern Europe. Discussion The substantial growth in the reach and rates of human travel, in particular the air traffic network, in recent decades, has had a major effect on global disease epidemiology, including malaria.9 Rising rates of travel to and from endemic areas has resulted in imported malaria being frequently reported in malaria-free countries, with occasional secondary transmission.7 However, this travel expansion has not been ubiquitous, with historical and economic ties driving growth along certain routes far more than others, and resulting in uneven malaria movement.2 Moreover, substantial investment in malaria control in recent decades has resulted in overall decreased prevalence in endemic areas, with some areas noting especially large decreases,22 further contributing to variations in importation to non-endemic countries. Here, we have presented unique analyses of a global assembly of publicly available contemporary data for the national reporting of imported malaria to capture these variations and quantify the broad geographic features.

ially large decreases,22 further contributing to variations in importation to non-endemic countries. Here, we have presented unique analyses of a global assembly of publicly available contemporary data for the national reporting of imported malaria to capture these variations and quantify the broad geographic features. We noted clear and consistent patterns despite differences in data quality, completeness, and temporality, and data being indicative of the different surveillance systems and diagnostic capacities of the reporting countries (appendix pp 9, 10). Our results underline the substantial geographical heterogeneities that exist in reported malaria case numbers and compositions in non-endemic countries. Moreover, certain routes from endemic to non-endemic countries carry substantially more infections than others, with evidence of tight couplings that reflect historical ties. These communities of countries can serve to guide surveillance, develop mitigation strategies, and highlight likely routes of drug-resistant malaria movement.23 The tight coupling of locations also highlights risks for secondary transmission following imported cases, such as through immigrant labour in the Middle East24 or Chinese labourers returning from Africa.25 Further, the species compositions highlight P vivax, P ovale, and Plasmodium malariae as potential malaria parasites in areas of the world where they are rarely considered, such as much of Africa. Coupled with policy shifts towards species-specific diagnostics and reporting, this finding could prompt a robust assessment of the more neglected non-falciparum parasites that can still cause severe clinical illness and require specific control interventions.

d where they are rarely considered, such as much of Africa. Coupled with policy shifts towards species-specific diagnostics and reporting, this finding could prompt a robust assessment of the more neglected non-falciparum parasites that can still cause severe clinical illness and require specific control interventions. We have endeavoured to minimise the uncertainties and errors that arise through the analysis of data from such a large range of sources. However, many factors, including the opportunistic and highly varied nature of the available data, affect our ability to compare between countries and draw precise conclusions. First, the data represent a small proportion of a possibly larger pool of cases, with some estimates suggesting that national statistics might capture just one-sixth of all imported cases.13 Variations in health system reporting mechanisms and diagnostic capacities between countries also probably mean that some countries capture more cases than others, some cities and regions within countries capture more than others, and some countries have a greater capacity to undertake reliable speciation through using PCR or having more experienced and well trained microscopists. Microscopic examination is widely available in most non-endemic countries (appendix pp 9, 10); however, misdiagnoses or late diagnoses can still be common because of the failure of medical personnel to relate the febrile symptoms to a disease that is rarely reported in their region. Malaria symptoms are non-specific and cannot easily be distinguished from other febrile disorders on clinical grounds alone. Moreover, changes in reporting standards, practices, and capacity over time within nations can affect the comparability of data over time, and affect outcomes and representativeness when data are pooled over many years. Microscopic diagnosis is often slow and inaccurate in non-specialised laboratories.26, 27 In some cases, molecular assays can become insufficient to make a correct diagnosis, especially to detect all species in mixed infections or in cases when parasitaemia is low, which is often the case in non-immune patients who complied with chemoprophylaxis.

en slow and inaccurate in non-specialised laboratories.26, 27 In some cases, molecular assays can become insufficient to make a correct diagnosis, especially to detect all species in mixed infections or in cases when parasitaemia is low, which is often the case in non-immune patients who complied with chemoprophylaxis. Moreover, one or more species in mixed species infections are easily overlooked, and some species are more difficult to classify than others, with, for example, morphological similarities between P vivax and P ovale potentially a source of misclassification.28 In relation to these classification challenges, the confidence in the reports regarding imported malaria varies between studies depending on the method used to detect the infection. The large percentages of unknown malaria types for some countries are likely to be indicative of a lack of diagnostic capacity. National health statistics often do not report the techniques used, and therefore it is necessary to refer to the academic publications describing the summarised national data in which this information is provided to assess the overall precision of the diagnosis at the country level. Finally, intervention scale-up,22 urbanisation,29 changing wealth,30 and improved health systems are all likely to have affected the prevalence and species composition of malaria in endemic regions in the past decade, subsequently affecting the comparability of imported case data in non-endemic regions between years. Nevertheless, the clear, consistent, and coherent patterns we recorded within regions and between countries and the congruence with results through dedicated surveillance networks,3, 14, 15, 16 suggest that the data presented here form a representative sample.

of imported case data in non-endemic regions between years. Nevertheless, the clear, consistent, and coherent patterns we recorded within regions and between countries and the congruence with results through dedicated surveillance networks,3, 14, 15, 16 suggest that the data presented here form a representative sample. Second, several factors relating to differences in traveller type and activity between countries contribute to the representativeness, comparability between nations, and uncertainties in outputs. Rates of chemoprophylaxis, prescription, use, and antimalarial adherence vary by country and by demographic group,31, 32 as does the use of protective measures while travelling.33 Further, the demographics and ethnic composition of traveller groups vary by country; for example, nations that have large migrant populations originating from endemic countries probably contribute to more cases arising from those visiting friends and relatives.6 The proportion of imported malaria cases due to migrants in Europe has increased in the past 15 years,34 with those visiting friends and relatives travelling to endemic areas of Africa more than eight times more likely to be diagnosed with malaria compared with tourists,35 and their children being especially at risk.36 Activities in endemic regions might also contribute to differences recorded; for instance, people travelling to urban areas and staying in hotels are likely to be at lower risk. Differences in demographics and health systems also translate to differences in treatment seeking as well as whether case importation occurs principally through visitors or travelling residents. Some demographic groups are more likely to seek treatment than others for travel-related health issues.34

o be at lower risk. Differences in demographics and health systems also translate to differences in treatment seeking as well as whether case importation occurs principally through visitors or travelling residents. Some demographic groups are more likely to seek treatment than others for travel-related health issues.34 Our study is an ongoing effort. Summaries of national malaria surveillance data are not made publicly available for all countries and years and many additional relevant datasets probably remain unpublished, so we welcome input from those who have access to datasets not included here to enable continual updates. Our study provides a global picture of malaria importation to non-endemic countries, but does not extend to exploration of the driving factors behind these patterns. However, our future work will focus on building datasets and a modelling framework for understanding what drives the patterns noted here.

ntinual updates. Our study provides a global picture of malaria importation to non-endemic countries, but does not extend to exploration of the driving factors behind these patterns. However, our future work will focus on building datasets and a modelling framework for understanding what drives the patterns noted here. The associations with malaria endemicity and migration flows suggest two key drivers, but further data for travel patterns and volumes, malaria transmission, demographics, health system efficiency, diagnostic capacities, treatment-seeking behaviours, and prophylaxis compliance and availability, among other factors, need to be collated to better explain and model the malaria importation patterns recorded, with a goal of predictive modelling. Further, such analyses could be extended to other commonly imported infectious diseases3, 9, 37 and the effects of seasonal variations in these drivers could be incorporated.38, 39 We have focused on broad comparisons through pooling across years to provide sufficient data. This approach has probably ignored changes that have occurred across time, and future work will have to focus on augmenting and breaking these data down to explore temporal trends. Finally, our results match closely those found through analysis of data collected by surveillance networks such as GeoSentinel14 and EuroTravNet,15, 16 but future work should focus on undertaking quantitative comparisons.

ure work will have to focus on augmenting and breaking these data down to explore temporal trends. Finally, our results match closely those found through analysis of data collected by surveillance networks such as GeoSentinel14 and EuroTravNet,15, 16 but future work should focus on undertaking quantitative comparisons. As many countries move towards national malaria elimination, global eradication moves up the international agenda,1 and the threat of spreading drug resistance grows,23 there is an increasing focus on malaria importation and the vulnerability of countries to resurgence.40 This study forms part of wider efforts to understand patterns of human and malaria parasite movement and how such information can guide control and elimination efforts. Malaria parasites do not respect national borders, and with human mobility continuing to increase in its volumes and reach, increasing global connectivity,2 control, and treatment strategies should account for the continued globalisation of malaria. Supplementary Material Supplementary appendix

As many countries move towards national malaria elimination, global eradication moves up the international agenda,1 and the threat of spreading drug resistance grows,23 there is an increasing focus on malaria importation and the vulnerability of countries to resurgence.40 This study forms part of wider efforts to understand patterns of human and malaria parasite movement and how such information can guide control and elimination efforts. Malaria parasites do not respect national borders, and with human mobility continuing to increase in its volumes and reach, increasing global connectivity,2 control, and treatment strategies should account for the continued globalisation of malaria. Supplementary Material Supplementary appendix Acknowledgments Funding from the Bill & Melinda Gates Foundation supports AJT (OPP1106427, 1032350, OPP1134076), DLS (OPP1110495), SIH (OPP1119467, OPP1093011, OPP1106023, OPP1132415), and PWG (OPP1068048, OPP1106023). AJT and DLS are also supported by NIH/NIAID (grant number U19AI089674). AJT is supported by a Wellcome Trust Sustaining Health Grant (106866/Z/15/Z). All authors also acknowledge funding support from the RAPIDD program of the Science and Technology Directorate, Department of Homeland Security, and the Fogarty International Center, National Institutes of Health, USA. SIH is a Wellcome Trust Senior Research Fellow (grant number 095066), whose fellowship also supports RH. PWG is a Career Development Fellow (grant number K00669X) jointly funded by the UK Medical Research Council (MRC) and the UK Department for International Development (DFID) under the MRC/DFID Concordat agreement, also part of the EDCTP2 programme supported by the European Union, and receives support from the Bill & Melinda Gates Foundation (grant numbers OPP1068048 and OPP1106023). This work forms part of the output of the Vector-borne Disease Airline Importation risk project, Flowminder Foundation, and WorldPop project, and part of the output of the Malaria Atlas Project. We thank the many employees of national health agencies around the world who guided us to datasets on imported malaria statistics and who took the time to answer questions and review drafts of the report.

ion risk project, Flowminder Foundation, and WorldPop project, and part of the output of the Malaria Atlas Project. We thank the many employees of national health agencies around the world who guided us to datasets on imported malaria statistics and who took the time to answer questions and review drafts of the report. Contributors AJT conceived the study and designed the analyses. PJ, DO, MF, and AJT undertook collection of the reports and references and extraction of data. AJT, PJ, and DO implemented the data processing and analysis. All authors contributed to the writing and editing of the report. Declaration of interests We declare no competing interests. Figure 1 Origins, destinations, and flows of imported cases of malaria from endemic to non-endemic countries (A) Of the non-endemic countries that reported the origin country of imported cases, the average annual number of malaria cases (all species) between 2005 and 2015 exported from endemic to non-endemic countries (red) and imported cases to non-endemic from endemic countries (blue). (B) Malaria endemic to non-endemic country connectivity through cases imported to the non-endemic country. Of the non-endemic countries that reported the origin country of imported cases, the average annual number of malaria cases (all species) between 2005 and 2015 moving from endemic to non-endemic country pairs are mapped as flow lines. Only average annual flows of >50 cases are mapped, with >200 in red, 100–200 in pink, and 50–100 in yellow. The flow lines are overlaid on a map of Plasmodium falciparum prevalence.19

l number of malaria cases (all species) between 2005 and 2015 moving from endemic to non-endemic country pairs are mapped as flow lines. Only average annual flows of >50 cases are mapped, with >200 in red, 100–200 in pink, and 50–100 in yellow. The flow lines are overlaid on a map of Plasmodium falciparum prevalence.19 Figure 2 Results of community detections on the network of malaria endemic–non-endemic imported malaria pairs Countries mapped in the same colour belong to a unique community, with imported malaria case movements being larger within the communities than between them. Fewer cases in Asia and the Americas had origin–destination information available, making the community detection results less robust. Moreover, with substantially fewer non-endemic countries outside of Europe with strong reporting, analyses simply show the regions as homogeneous single communities. Figure 3 Plasmodium falciparum prevalence in 2010 versus average annual number of P falciparum cases imported to non-endemic countries for all endemic countries, 2005–15 The circles are coloured by region (green=Americas, pink=Africa, blue=Asia), and their sizes correspond to numbers of outgoing migrants to non-endemic countries. Linear model fit to P faciparum parasite rate (PfPR) against P faciparum cases: r2=0·32; p<0·01. Figure 4 Pooled data for imported malaria species breakdown to non-endemic regions from endemic ones The figure shows proportions of total imported case numbers, rather than absolute numbers, which are shown in figure 1A.

The circles are coloured by region (green=Americas, pink=Africa, blue=Asia), and their sizes correspond to numbers of outgoing migrants to non-endemic countries. Linear model fit to P faciparum parasite rate (PfPR) against P faciparum cases: r2=0·32; p<0·01. Figure 4 Pooled data for imported malaria species breakdown to non-endemic regions from endemic ones The figure shows proportions of total imported case numbers, rather than absolute numbers, which are shown in figure 1A. Figure 5 Species composition of reported imported malaria cases to non-endemic countries mapped by endemic country of origin of the cases Table Species composition of reported imported malaria cases to non-endemic countries from available data with species composition recorded between 2005 and 2015

The figure shows proportions of total imported case numbers, rather than absolute numbers, which are shown in figure 1A. Figure 5 Species composition of reported imported malaria cases to non-endemic countries mapped by endemic country of origin of the cases Table Species composition of reported imported malaria cases to non-endemic countries from available data with species composition recorded between 2005 and 2015 Average number of cases per year P falciparum P vivax P malariae P ovale Other or unknown Australia 222 44·8 44·4 0·0 0·0 10·8 Austria 84 56·0 34·7 2·2 0·0 7·1 Bahrain 158 14·3 85·7 0·0 0·0 0·0 Belgium 227 63·0 23·0 5·0 9·0 0·0 Bulgaria 40 68·6 23·9 1·8 2·3 3·3 Canada 20 37·9 47·1 0·7 5·7 8·6 Croatia 11 64·8 19·9 1·9 0·5 12·8 Czech Republic 20 56·7 43·3 0·0 0·0 0·0 Denmark 104 76·8 17·0 2·3 2·9 1·1 Estonia 3 61·2 25·8 1·2 1·8 10·0 Finland 31 71·7 19·2 2·3 5·7 1·1 France 2169 85·7 6·5 2·1 5·7 0·0 Germany 401 82·0 8·0 2·9 2·6 4·6 Greece 41 42·6 50·0 2·5 0·8 4·0 Hong Kong 40 21·2 52·9 8·7 8·7 8·7 Ireland 54 74·7 7·1 1·9 4·5 11·7 Israel 60 45·0 52·1 1·3 1·6 0·0 Italy 637 83·4 8·4 1·6 6·5 0·0 Japan 45 47·2 48·7 0·9 3·2 0·0 Lithuania 4 62·5 12·5 12·5 0·0 12·5 Morocco 58 86·5 1·7 1·7 10·2 0·0 Netherlands 366 75·0 25·0 0·0 0·0 0·0 New Zealand 44 33·6 58·8 1·6 2·0 4·0 Norway 49 68·1 17·0 4·0 1·4 9·4 Poland 25 71·5 21·9 1·3 2·0 3·3 Portugal 178 71·5 0·0 0·0 0·0 28·5 Qatar 146 13·7 40·0 0·0 0·0 46·3 Réunion Island 156 85·8 10·7 1·9 1·6 0·0 Romania 29 65·0 0·0 0·0 0·0 35·0 Serbia 15 62·4 16·8 0·0 0·0 20·8 Singapore 148 29·6 67·3 1·2 0·0 1·9 Slovakia 5 58·0 42·0 0·0 0·0 0·0 Slovenia 7 48·5 43·8 0·0 4·7 3·1 Spain 374 54·5 17·8 1·1 11·1 15·6 Sweden 72 63·8 27·6 1·4 7·2 0·0 Switzerland 225 80·1 11·7 3·2 3·3 1·7 UK 1898 76·4 15·1 1·8 5·9 0·7 USA 1511 46·9 16·8 2·4 2·4 31·5

Introduction Pregnant women with cholera are at risk of complications, leading to fetal losses in 2–36% of cases if not treated promptly.1, 2 Killed, whole-cell, oral cholera vaccines are recommended by WHO to reduce the risk of cholera. Licensed oral cholera vaccines include Dukoral (Valneva, Lyon, France), Shanchol (Shantha Biotechnics, Hydrabad, India), mORCVAX (Vabiotech, Hanoi, Vietnam), and Euvichol (EuBiologic Co, Ltd, Chuncheon, South Korea). Shanchol is used most commonly in outbreak response in low-income countries, and has a cumulative efficacy of 65% over 5 years.3 Shanchol and Euvichol are available through the global stockpile of oral cholera vaccine.4 Findings from clinical trials with non-pregnant participants have shown that oral cholera vaccine is safe.5 However, cholera vaccination campaigns often exclude pregnant women because of insufficient data about the safety of the vaccine during pregnancy. WHO recommends vaccination of pregnant women in cholera-endemic settings, for whom the risk of cholera infection can be high.6 The package inserts for Dukoral and Shanchol are cautious about the use of these vaccines during pregnancy, because definitive evidence of safety during pregnancy is not available. In Tanzania, the risk of pregnancy loss was not significantly higher among pregnant women who were inadvertently vaccinated with Dukoral during the mass vaccination campaign in 2009.7 Findings from a retrospective cohort study in Guinea showed no evidence of increased risk of pregnancy loss after receiving Shanchol.8 However, such retrospective studies are subject to biases and represent low-quality evidence.