Trends in maternal and neonatal mortality in South Africa: a systematic review

This review aimed to synthesise estimates of maternal and neonatal mortality for the period 1990 to 2015 in South Africa and to determine temporal trends during this period.

The review protocol was registered with the PROSPERO database in 2016 with a registration number CRD42016042769 (http://​www.​crd.​york.​ac.​uk/​PROSPERO/​display_​record.​php?​ID=​CRD42016042769) and has already been published [31]. The presentation and reporting of results in this review followed the systematic review reporting standard (PRISMA-P) [32]. To ensure transparency, a PRISMA flow chart was used and a table indicating all included studies was presented [33].

The population for eligible studies included pregnant women and neonates for ascertaining maternal and neonatal mortality, respectively. All studies that are nationally representative, reports providing national-level data (and trends thereof) and vital registration data were eligible for this review. Searches were restricted to studies written in English and being conducted in South Africa or which have used South African data, and multicentre studies including South Africa, reporting data covering the period 1990 to 2015. No restrictions on the date of publication were made in order to include articles reporting data from 1990 to 2015 which are published beyond 2015.

Separate searches for the two outcomes (maternal and neonatal mortality) were conducted in the following electronic databases: Medline, Africa-Wide Information, Scopus, Web of Science and CINAHL. The last search was carried out on 18 August 2017. No restrictions on the date of publication were made. Additional searches for conference abstracts and proceedings were made. Reference lists of retrieved articles were also screened for additional publications. Reports by the government or other agencies were included based on publications, and a number of data sources reported by them were included. Contacts with experts in the field of study were made to identify additional relevant articles.

The searches in the forementioned electronic databases were conducted from August 2016 to August 2017. All searches were restricted to articles written in English and reporting data covering the period from 1990 to 2015. In particular, the search strategy used in Medline database was as follows: ((“mothers”[MeSH Terms] OR “mothers”[All Fields] OR “maternal”[All Fields]) OR (“infant, newborn”[MeSH Terms] OR (“infant”[All Fields] AND “newborn”[All Fields]) OR “newborn infant”[All Fields] OR “neonatal”[All Fields])) AND ((“mortality”[Subheading] OR “mortality”[All Fields] OR “mortality”[MeSH Terms]) OR (“death”[MeSH Terms] OR “death”[All Fields])) AND (estimation[All Fields] OR estimates[All Fields]) AND (“South Africa”[Mesh] OR (“south africa”[MeSH Terms] OR (“south”[All Fields] AND “africa”[All Fields]) OR “south africa”[All Fields])) AND ((“1990/01/01”[PDAT]: “3000/12/31”[PDAT]) AND “humans”[MeSH Terms] AND English[lang]) Additional file 1.

Search outputs were managed in EndNote reference manager. Any duplicate records were removed before the screening process takes place. When the same article was captured in different journals or the same results were presented with different main authors, the most detailed publications were selected for review. Three independent reviewers were involved in the screening and selection of articles to be included in a quantitative (narrative) synthesis. This involved an assessment of articles based on titles and abstracts and full-text review using Covidence software (https://​www.​covidence.​org/​). For an article to be eligible for inclusion in the systematic review, two reviewers had to agree to include it. A third reviewer was consulted in case of any difference of opinion between the two reviewers. This followed when they failed to reach a consensus after a joint examination of the different views.

Analysis of the full text was conducted for all eligible articles. Two authors extracted data independently using a pre-agreed data abstraction template. In the case of discrepancies in the extracted data between authors, consensus was sought before involving a third author for resolving the disparities. During the data extraction process, study authors/investigators were contacted to provide extra information when there were insufficient information/data reported in the article.

The following information was extracted for eligible studies: first author’s name; year of publication; year of death (maternal and neonatal); number of pregnant women; number of live births; maternal deaths; neonatal deaths; definition of maternal death; definition of neonatal death; maternal mortality ratio/rate (if reported, and by year); neonatal mortality rate (if reported, and by year); sampling method; estimation method used; and an indicator variable whether the records are complete.

The main outcomes in this review were maternal and neonatal mortality. Maternal death/mortality was defined as the death of a woman while pregnant or within 42 days of termination of pregnancy, irrespective of the duration and site of the pregnancy, from any cause related to or aggravated by the pregnancy or its management but not from accidental or incidental causes [34]. Maternal mortality ratio (MMR) was defined as the number of maternal deaths per 100,000 live births.

Neonatal death/mortality was also defined as the death of live-born within the first 28 days of life. Neonatal mortality rate was defined as the number of infant deaths within the first 28 days of life per 1000 live births.

Assessment of risk of bias was done at a study and outcome level. Two authors assessed the study quality based on the following quality assessment criteria: (1) definition of maternal mortality, (2) definition of neonatal deaths, (3) completeness of ascertainment of maternal and neonatal mortality, (4) completeness of ascertainment of live births, (5) sampling technique/design and (6) data quality. Studies were assessed based on each criterion and were rated as “high risk of bias” or “low risk of bias” accordingly. Studies rated as high risk of bias on any criterion were assigned an overall rating of high risk of bias while the overall rating of low risk of bias was only assigned in studies with low risk of bias in all criteria. For model-based estimations, risk of bias was assessed based on the input data used. Reports by the government and other agencies such as Stat SA, National Department of Health and WHO were assessed using similar criteria as empirical studies. Table 1 shows the assessment criteria of risk of bias in individual studies.

Data were presented as ratios for maternal mortality and rates for neonatal mortality with their corresponding confidence or uncertainty intervals.

Data were entered and analysed using STATA software version 14.1 (Stata Corp, College Station, Texas). Data were presented as MMR or NMR in tables and graphs to depict trends over time. The reasons for study exclusions were clearly documented.

As presented in Fig. 1 below, a total of 948 studies were identified through the literature search and 21 additional studies were identified through screening of reference lists. After removing the duplicates, 670 studies were screened for eligibility. A total of 608 studies were excluded after screening the titles and abstracts as they reported irrelevant information. Sixty-two abstracts were shortlisted for full-text review and 39 studies met the inclusion criteria for analysis. Of studies included in the review, 25 reported data on maternal mortality [28, 35‐55] and 14 on neonatal mortality [14, 23, 40, 45, 55‐64].

Table 4 below presents the assessment of risk of bias of the individual studies. A total of 11 studies reporting maternal mortality data [28, 38, 49, 51‐55, 65‐67] and six studies reporting neonatal mortality data [23, 55, 56, 60, 62, 64] had overall high risk of bias. Among studies reporting maternal mortality data, three studies did not use the ICD-10 definition of maternal death [52, 65, 66], eight studies used data which were not population-representative [28, 38, 48, 49, 53‐55, 67], one study used sisterhood estimation methods [55] and the sampling technique was unclear in one study [53]. Of seven studies reporting neonatal mortality data having an overall high risk of bias, four were not population-representative [23, 62, 64] and three were surveys based on recall of neonatal deaths more than 6 months previously [55, 56, 60].

This systematic review aimed to provide an overview of maternal and neonatal mortality from 1990 to 2015 for monitoring purposes, tracking progress and to advocate for resources and policy attention. In general, the estimates derived from all studies and reports indicated that South Africa did not achieve the MDG 4a and 5a goals of reducing under-five mortality by two thirds and maternal mortality by three quarters between 1990 and 2015, respectively. Despite the country struggling to achieve the MDG goals for maternal and neonatal mortality in the last two decades, recent reports showed significant progress made in reducing these outcomes [45, 68, 73].

Although maternal and neonatal mortality are highly researched by both local and international authors or institutions in South Africa, there are considerable uncertainties around these estimates in the country. The possible reason for this might be high reliance on only a few data sources and limited empirical work. Accounting for the uncertainties about the actual levels of MMR and NMR in the country, estimates from both the institutional reports and global metrics indicated an upward trend in MMR and NMR until around 2006 and 2009. However, the increase in MMRs between 2001 and 2006 might specifically be explained by a consistent increase in HIV prevalence among pregnant women in the same period [74]. In addition, the downward trends in MMRs and NMRs from 2009 can be linked with the massive uptake of HIV treatment and an increased coverage of essential interventions, in particular the prevention of mother-to-child transmissions (PMTCT) of HIV which currently stands at over 90% [29, 30]. Nonetheless, all recent estimates are much more closely grouped indicating convergence over time.

Evidence from the literature indicated that empirical methods, i.e. vital registration, household surveys and censuses, are subjected to misclassification and under-reporting of maternal deaths, thus leading to wide uncertainty intervals. Furthermore, estimating neonatal mortality from census and household surveys in high HIV prevalence settings is known to provide a biased estimate of child mortality due to correlations between HIV deaths in mother and death of her child [70].

Large margins of uncertainty associated with the estimated MMR and NMR highlight the need of interpreting these estimates with caution as well as not using them for monitoring trends over a short duration. The reasons for variations in the estimates of maternal and neonatal mortality remain poorly researched over the past two decades. In this review, we have observed a substantial discrepancy in the consistency of definitions used in the estimation of these outcomes, such as differentiating maternal deaths from pregnancy-related deaths [52, 65, 66]. Thus, uncertainties in estimates of MMR might be partly explained by differences in definitions used. Different estimation techniques used to obtain MMR and NMR necessitated the use of different data inputs, i.e. empirical data versus modelled estimates, which likely contributed to the divergent estimates of these outcomes.

For these reasons, cross-country comparisons, comparisons based on data from different sources and assessments of the overall burden become difficult. These comparisons should in many cases be interpreted with considerable caution due to different strategies being employed to derive such estimates. Evidence from recent studies focusing on estimating child mortality have revealed that methodological differences bias and compromise international comparisons of perinatal mortality [75‐77]. Moreover, divergent estimates of MMR and NMR by different sources compromise interpretation of trends over time.

Over the past three decades, efforts have been made to improve the quality of maternal and neonatal mortality data due to the incompleteness of vital registration systems as well as the lack of reliable population surveys collecting detailed information on birth histories in the country. This included the introduction of modules about sibling history in national household surveys (e.g. Demographic and Health Survey (DHS)), including questions in censuses about whether a woman’s death was related to pregnancy, and the use of mixed methods cross-referencing facility records to determine the extent of under-registration of maternal deaths in vital registration system [4, 7, 78, 79]. However, these improvements could also have contributed to the increasing maternal and neonatal mortality over time.

Despite improvements in the completeness of death registrations in the last decade, the completeness of death registration has been reported to be lower in children as compared to adults and in rural areas than urban [79]. This might potentially explain some of the variability in estimates of both maternal and neonatal mortality in the country.

Generally, this review has revealed divergent estimates of MMR and NMR obtained from vital registration, household surveys, censuses and modelling over time. To obtain more accurate estimates, there is a need for applying additional estimation techniques which utilise available multiple data sources to correct for the underreporting of these outcomes, perhaps the capture-recapture method. This method is useful in resolving uncertainties in estimating conditions that have diverse estimates by operationalising statistically overlapping information from multiple data sources [80‐84].