WHO analysis of causes of maternal death: a proposed protocol for a global systematic review R. Champaneria 1, J. Daniels 1,2, K.S. Khan 1, L. Say 3, A.M. Gulmezoglu 3, S. Cousens 4, A.J.Howman 2 1 Academic Department of Obstetrics & Gynaecology, University of Birmingham, Birmingham Womens Hospital, Metchley Park Road, Edgbaston, Birmingham B15 2TG 2 Birmingham Clinical Trials Unit, Division of Medical Sciences, Robert Aitken Institute, University of Birmingham, Edgbaston, Birmingham B15 2TT 3 UNDP/ UNFPA/ WHO World Bank Special Programme of Research, Development and Research Training in Human Reproduction, Department of Reproductive Health and Research, WHO, Geneva 27, CH-1211, Switzerland 4 London School of Hygiene and Tropical Medicine, Keppel Street, London, WC1E 7HT 1
Summary Background A systematic review was undertaken by the World Health Organisation (WHO) to determine the distribution of causes of maternal death (Khan K.S. et al. 1066-74). This review of maternal mortality covered data reported during 1997-2002 and whilst it generated important information for healthcare policy makers, it identified gaps in the data, particularly in regions where information is most needed to guide practice and policy. We intend to update the review by looking at data since 2002. Through this work, we will update the estimates of the distribution of causes of maternal mortality across world regions, identify gaps in regional coverage, and also explore to what extent a country s development status, geographical location and datasets methodological features explain the variable distribution of causes of maternal deaths. Through well established statistical methods we will explore the use of imputation, to generate data where gaps exist so as to provide as complete an information set as possible on worldwide causes of maternal mortality. Objective To determine the global distribution of causes for maternal deaths. Methods Following extensive bibliographic searching, datasets will be selected using prespecified criteria. Dataset characteristics, methodological quality features and causes of maternal deaths will be recorded on a piloted data extraction form. Joint causes of maternal deaths will be analysed from datasets reporting at least four of the major causes. This systematic review will use the same method used in a previous review. We also intend to extrapolate to countries where data is not available, using multinomial regression. 2
Introduction WHO conducted a systematic review of maternal morbidity and mortality rates that covered data reported during 1997-2002. An extensive search of electronic databases, relevant in-country studies and other grey literature revealed over 60,000 citations of which over 2500 were included in a database. This database included detailed information on primary studies that allowed critical appraisal of data and stratified analyses according to relevant study characteristics. The outputs from this first ever systematic analysis of the cause-distribution of maternal deaths were the geographical variation in distribution of major causes of death and the joint distribution of maternal deaths. The analysis was published in the Lancet in 2006 (Khan K.S. et al. 1066-74). Results on the global burden of a number of other conditions have been published since the completion of this systematic review in 2004 (Gulmezoglu A.M. et al. 16). WHO have initiated an update of this previous work as follows: updating the literature database by identifying and including studies/reports that were published after 2002. The literature search has been repeated and newly identified articles are currently being screened reassessing facility based datasets excluded from the previous review for possible inclusion under revised inclusion criteria re-analysis of joint distribution of causes of death using the techniques employed in the 2006 Lancet paper. generation of pie-charts of cause-distribution of maternal deaths at both global, regional and sub-regional levels incorporate the vital registration (VR) data from low-mortality countries with accurate cause-attribution The Child Health Epidemiology Reference Group (CHERG) neonatal group has used a multinomial modelling approach to estimate the cause specific proportions of neonatal deaths in countries for which recent, nationally representative data are not available. Two models were developed, one for low mortality countries using VR data as input data, and one for high mortality countries using research study data as input data. A similar approach could be applied to estimating the distribution of causes of maternal deaths. One advantage of the multinomial model approach developed for neonatal mortality is that it can be used with input data sets which do not report information on all of the separate causes included in the model (provided that one is prepared to make assumptions about which category any unreported causes would have been put into). 3
Methods Identification and selection of datasets The literature search strategy used in the previous systematic review will be repeated (Khan K.S. et al. 1066-74). Potentially eligible datasets will include journal articles, registries, published and unpublished information from governments and other agencies, whether available in print or online. The updated search will be complemented by searching the grey literature, reference lists and contacts in WHO regional offices and country s health departments. The searches will be done by LS, who will remove duplicate records and create a Reference Manager database of all the potential datasets. In this review update, we will extend the eligibility of datasets to include datasets reporting rates at a facility or organisational level, such as hospitals. However, these facillity based studies will only be eligible for inclusion if >75% of the deliveries locally happen at the facility. Reports and datasets excluded from the previous review will be reassessed for inclusion. Types of studies Studies providing estimates of maternal mortality derived from direct counting or special surveys are eligible. Estimates derived from modelling of other variables or extrapolation from other populations will be excluded. Summary of inclusion criteria Studies will be included if and only if all of the following apply: 1. the study includes data on maternal mortality 2. The dates for data collection are specified 3. The study contains some data post 1997, and no data pre 1990 o Note studies with data pre 1990 are eligible if this data can be stripped out 4. The study methodology is described and considered suitable o Non-suitable methodologies would include eg case-control studies, where the study sample is not representative 5. The study reports at least 25 maternal deaths 6. The study covers at least 4 of the main causes of death listed on p.xx 7. If the study is based at a facility, then >75% of deliveries for the locality are at the facility Types of participants Women either pregnant or within one year of termination of pregnancy or delivery of baby. 4
Types of outcome Maternal mortality and morbid conditions defined according to the International Classification of Diseases, 10 th revision (WHO). Earlier reports using the 9 th revision of the ICD classification system will not necessarily be excluded if causes of death can be ascertained with confidence. Selection of datasets Once the literature searches are complete, RC will assess the list of citations, on the basis of their titles and abstracts. Those seen to be irrelevant will be excluded and reasons for exclusion will be noted for reference. In situations where the title and abstract of a paper do not provide enough information to decide inclusion or exclusion, the full paper will be retrieved and evaluated. A random sample of citations will be screened by RC and JD independently to estimate the degree of disagreement. If either one of the reviewers is unsure about inclusion of a citation, the full paper or report will be obtained. Data extraction The full papers and reports of eligible studies/ datasets will be sent to RC, who will then begin data extraction on the newly identified datasets. Additionally, the datasets previously included in the original review will be reexamined as data that was not previously extracted will be included in the update. For all included datasets, predictor information (see below) will be extracted either directly from the paper/ report or else authors contacted to provide data specific for the location and timeframe of the dataset. If it is not possible to locate appropriate local predictor data, country level estimates will be obtained from the WHO or other sources. Where possible, data obtained from these sources will be for the same time period as that of the study. Data will be extracted on a specifically designed piloted data extraction form. The form will be separated into subsections, covering: 1. general characteristics of the study, such as design, population, and setting 2. quality assessment of studies reporting maternal mortality 3. maternal mortality 4. predictor data LS and AMG will extract descriptive country level demographic data from the WHO and World Bank datasets e.g. GDP, literacy rates, life expectancy etc. Hypertensive disorders of pregnancy has been chosen as being the reference standard for the modelling, as this is a common cause of death and present in the majority of datasets, although datasets will not be excluded if they do not report hypertensive causes of death. A manual providing definitions and instructions on how to extract data will be developed. Both screening and data extraction form will be tested on group of studies of different designs. Any revision to the form(s) will be made prior to them being used. 5
Data processing Variables of interest will be categorised, and codes developed for responses to open-ended questions. This will ensure consistent data entry and facilitate statistical analysis. The ICD-10 will be used to classify causes of maternal death (WHO). Extraction of raw data will be preferred, but only if percentages or rates are given. Data from graphs or figures will only be used if numbers or percentages are not mentioned in the text or given in the graph. Data will not be included if extrapolation has been used to calculate figures. Once data extraction is complete, data will be reviewed to identify any duplication, for example the same results published in more than one journal or published papers whose unpublished drafts have been previously identified. Data will then be entered into a specific database. Methodological quality Studies will be excluded if they do not mention the methodology used to obtain data. Evaluation of both methodological features and quality will be used to assess the reliability and accuracy of the data. For instance, the selection criteria and some participant characteristics such as economic status, ethnicity, age group or health status will allow assessment of external validity or generalisability of results, as well as presentation of stratified analysis for different categories. Similarly, internal validity of a study will be assessed by examining information on the proportions and characteristics of losses to follow-up, non-responders or those who were not included in the final analysis, after having been initially selected. If a study states the definitions of conditions and descriptions of diagnostic methods/ procedures, this study should be regarded as of higher quality. A study should also be considered as of higher quality if, in addition to categorising definitions of maternal mortality, it reported the efforts undertaken by the authors to capture all maternal deaths and the methods used to confirm the causes of death (eg confidential enquiry). For quality assessment, data will be extracted on (i) study design, sampling method, sources of data and completeness of follow-up or records, and (ii) reported definitions and diagnostic procedures regarding mortality. (1) Does the study/ publication report the definition of maternal death? This reduces bias in ascertainment of denominator data in total mortality rate (any published definition reported vs no definition) (2) Does the study employ adequate data sources to ascertain a capture of maternal mortality that is as complete as possible (use of multiple data sources, special surveys, or clinical studies vs routine registration systems, in which adequate attribution of cause of death has been shown to be questionable for maternal deaths, leading to substantial underreporting); (3) Does the study report use of a robust approach to ascertain that the cause of death is a representation of the underlying condition that is as true as 6
possible (confidential enquiries, verbal autopsies, use of multiple sources of death certification vs no special efforts to confirm cause) (4) Does the study report a sufficiently high proportion of cases with attributable cause of death established (no minimum has been specified as a cutoff for inclusion); (5) Does the study identify a sufficiently large number of causes within a series to ensure that a fuller range of conditions have been included to reduce the risk of misclassification of cause of death (>=4 causes was taken as a cutoff for inclusion). Database The Birmingham team have been given access to the existing database. They will examine this database and propose quick and effective ways to compile the new database. All extracted data will be added to the newly created SQL relational database. The fields of the database will correspond to the questions in the data extraction form. The Birmingham team will be responsible for the compilation of the database, import of the original database and entry of new datasets and new data items i.e. predictor variables. Two new flat file relational databases will be created, one will have all the data extracted from journal articles, the other will have vital registration data. So, if for example, there is a piece of data not reported in a journal article, we will be able to fill this gap by using the vital registration data. Statistical analysis plan The primary analyses of joint causes of death distribution will be analysed in the same manner as the original review. The set of distinct categories of causes of death used in the initial review will also be used here. A decision on whether indirect causes should be modelled separately from direct causes will be taken at the analysis stage. To distinguish between high and low mortality countries, two separate analyses will be done: one for high mortality countries, the other for low mortality countries, though estimates at regional level may be based on pooling data from the two separate models. The range of maternal mortality ratios (MMR) for countries with vital registration data will then be examined and if any outliers are found, they will be excluded. The UN (United Nations) list of countries/ regions will be used, Data will be modelled by the use of multinomial logistic regression. The following potential predictor variables have been decided upon, and may be included in the statistical model: 7
Mortality and Morbidity A B C D E Indicator Years Country coverage Data Source Overall maternal mortality 2005 All WHOSIS website ratio (per 100,000 live births) Adult mortality rate 1990, 2000, All WHOSIS website (female 15-60 years) 2006 Life expectancy at birth 1900, 2000, All WHOSIS website (female) 2006 Neonatal mortality ratio 1900, 2000, All WHOSIS website (per 100,000 live births, 2006 boys and girls) HIV infection (per 100,000 >15years) or HIV prevalence 2005 All UNAids Website http://www.unaids.org/en/default.asp 2001 2007 Most countries in some years F Malarial prevalence 2000-2003 Most countries We have dataset from Lale, WHO Malaria Report 8
Health Service Coverage G H Antenatal care coverage (>=4 visits) Proportion of births attended by skilled birth attendant (various definitions) Have dataset Most years Have dataset Not every year I Institutional delivery Have dataset Most years, some for more countries than others J K Proportion of Caesarean deliveries Contraceptive use prevalence (women of reproductive age) 2001 2006 (not all years) 1997-2006, but not every year in between Most countries Most countries Most countries All Most We have dataset from Lale, but not sure of source. Datasets for 1+ and 4+ visits We have dataset from Lale, but not sure of source We have dataset from Lale, but not sure of source We have dataset from Lale, but not sure of source We have dataset from Lale, United Nations Dept. of Economic and Social Affairs Population Division. World Contraceptive Use 2007 9
Health Systems Resources L Per capita total expenditure on health (PPP$) 1995 2006 All Demographics M Total fertility rate 1970-1975, 2000-2005, (and some years in between) 1990, 2000 and 2006 All We have dataset from Lale, World Fertility Patterns www.unpopulation.org N Adult literacy rate (overall) 1990-2005 Sparse UN STATS/ MDG O Gross net income per 1990-2006 All capita (PPP$) P Proportion of population in 1990, 2000, All urban area 2006 Q Registration coverage of births 2000-2006 Sparse http://www.childinfo.org/files/birthregi stration_table.pdf R Registration coverage of deaths 2000-2005 All in 2005 10
S Proportion of women in paid, non-agricultural employment Legal status T Abortion - proportion of maternal deaths due to unsafe abortion U Polygamy V Female genital mutilation Remaining suggested indicators sources unclear a. Essential drug list/ formulary for drugs relevant to preventing maternal mortality b. Quality or availability of blood transfusion services c. Safety of anaesthesia equipment Obesity/ BMI (in WHOIS risk factors dbase 11
Flowchart of timeline of activities (December 2008 April 2010) 12
Reference List 1. Khan K.S., Wojdyla D., Say L., Gulmezoglu A.M., Van Look P.F. WHO analysis of causes of maternal death: a systematic review. Lancet 2006;367:1066-74. 2. Gulmezoglu A.M., Say L., Betran A.P., Villar J., Piaggio G. WHO systematic review of maternal mortality and morbidity: methodological issues and challenges. BMC Med Res Methodology 2004;4:16. 3. WHO. WHO. International Classification of Diseases and related health problems. World Health Organisation 10th revision. 1992. 13
14
15
16
17