Estimated at 2.6 million annually, stillbirths worldwide have stayed alarmingly high, in contrast to neonatal and under-five mortality rates. It is a neglected public health challenge globally, with less attention to its social determinants. We examined spatial patterns of country-level stillbirth rates and determined the influence of social determinants of health on spatial patterns of stillbirth rates. We also estimated probabilistic relationships between stillbirth rates and significant determinants from the spatial analysis. Using country-level aggregated data from the United Nations databases, it employed ecological spatial analysis and artificial intelligence modeling based on Bayesian network among 194 World Health Organization member countries. From the spatial analysis, thirty-seven countries formed a cluster of high values (hot-spots) for stillbirth and 13 countries formed a cluster of low values (cold-spots). In the multivariate regression, gender inequality and anaemia in pregnancy were significantly associated with spatial patterns of higher stillbirth rates, while higher antenatal care (ANC) coverage and skilled birth attendants during delivery were associated with clusters of lower stillbirth rates. The Bayesian network model suggests strong dependencies between stillbirth rate and gender inequality index, geographic regions and skilled birth attendants during delivery. The Bayesian network predicted that the probability of low stillbirth rate increased from 56% to 100% when the percentage of countries with high skilled birth attendants during delivery increased from 70% to 88%, high ANC coverage increased from 55% to 70%, high prevalence of anaemia in pregnancy decreased from 27% to 11% and high gender inequality index decreased from 43% to 21%. Recognizing the urgency in reducing stillbirths globally, multi-pronged strategies should be designed to promote gender equality and strengthen the reproductive and maternal health services in Africa, Eastern Mediterranean, South Eastern Asia, and other countries with disproportionately high stillbirth rates.
This is an ecological spatial and Bayesian network analyses that utilized country-level aggregated data which are publicly available from the United Nations (UN) databases (Supplementary Table S1). All the 194 World Health Organization (WHO) member countries17 were included in this study. According to WHO region categories, there were 46 African (AFR), 35 American (AMR), 22 Eastern Mediterranean (EMR), 53 European (EUR), 11 South East Asian (SEAR), and 27 Western Pacific (WPR) countries. The dependent variable—stillbirth rate was re-estimated by Blencowe and colleagues18 as the ratio of annual number of foetal deaths after 28 weeks to the total number of live births in a year2. The last estimation used for this study was conducted in 2015. The details of the calculations can be found at the WHO Global Health Observatory data repository website2. There were 23 independent variables, being measured at the country-level and obtained from the World Bank19, United Nations Development Programme (UNDP)20, and WHO21. These independent variables were selected to represent the key aspects of the Social Determinants of Health (SDH) Framework22: socioeconomic and cultural factors, lifestyle, healthcare resources, maternal infections and conditions, and maternal and reproductive health service coverage indices. As shown in Fig. 1, the conceptual model hypothesized the different classes of variables influencing stillbirth along various pathways. At one end of the spectrum, the nations with good demographic and socioeconomic indices would have better healthcare system (including MNCH services) and health behaviour. The culture, social organization, and health behaviour would have a profound influence on maternal health conditions, which mediate stillbirth. Specifically, skilled birth attendants during delivery and delivery by caesarean section were proposed as potential modifiers/mediators for the relationship between maternal health conditions and stillbirth. The conceptual model for the determinants of stillbirth (adapted from Social Determinants of Health Framework)22. As shown in Supplementary Table S1, the national demographic, socioeconomic and cultural factors examined were Gross National Income (GNI) per capita, poverty rate, urban residence, female educational status, crude birth rate, gender inequality index, and income inequality. Gender inequality index (GII) measures gender disparity between men and women in the 3 dimensions; reproductive health, empowerment and labor market. The lowest possible score is zero (equality) and highest possible score is one (inequality)20. For health behaviour/lifestyle, we used age-standardized prevalence of current smoking among adult females and total alcohol per capita consumption in adults. We also included current health expenditure as percentage of gross domestic product (GDP) and density of skilled health personnel (per 10 000 population) as proxies for healthcare resources. Maternal health conditions included prevalence of anaemia in pregnancy (adjusted for altitude and smoking), syphilis seropositivity among antenatal care (ANC) attendees, age-standardized prevalence of obesity among female adults, prevalence of underweight among female adults as a proxy for maternal nutrition, age-adjusted prevalence of hypertension among adult females and age-adjusted prevalence of diabetes mellitus in female adults. Indicators for healthcare service coverage included percentage of deliveries by caesarean section (measurement for access to emergency health care during childbirth), ANC coverage—at least 4 visits, skilled birth attendants during delivery, and proportion of women of reproductive age (married/in-union) who have their family planning needs satisfied with modern methods, adolescent birth rate and prevalence of child marriage—proxies for sexual and reproductive health. The reporting dates differ for independent variables, ranging from 2006 to 2017. GNI per capita, poverty rate, prevalence of current smoking among female adults, syphilis seropositivity, prevalence of underweight among female adults, density of skilled health personnel, and adolescent birth rates were observed to be skewed, hence they were log-transformed to ensure normality of the variables. The philosophical perspectives for this study were drawn from postpositivist and transformative worldviews. The postpositivist view was employed to verify our theory and determine the relationship of key independent variables on stillbirth rate (i.e. deterministic and reductionistic lens). In addition, transformative perspective was used to raise the consciousness of stakeholders for political/programmatic changes to mitigate health disparities, especially in LMICs. The spatial scale and unit of analysis is country. With GeoDa v. 1.12 software23, a distance based spatial weight matrix with a threshold distance of 5720 km was generated in a geographic coordinate shapefile obtained from ArcGIS24. This distance was the most appropriate after various calibrations with different distances; all the 194 countries were interlinked—a condition for determining spatial dependence25. Furthermore, the adequacy of spatial weights was confirmed with the symmetry of connectivity histogram and connectivity map25. To examine spatial dependence, cluster analyses were performed. The global and local spatial autocorrelations were measured with Global Moran’s I index and Local Moran’s I index, respectively. Moran’s I index value ranges from +1 to −1, indicating strong positive autocorrelation (perfect clustering) to negative autocorrelation (perfect dispersion)26. Correspondingly, the positive and negative indices suggest aggregation of neighbourhoods with similar and different values across geographical space than the expected random distribution, while zero implies no autocorrelation (i.e. perfect randomness). The local indicator of spatial autocorrelation (LISA) cluster maps were generated to show hot- and cold-spots of statistically significant spatial clusters of neighbouring countries with high and low stillbirth rates, respectively. Randomizations were set at 999 Monte Carlo permutation to ensure adequate statistical power for p-value 30)25. For the purpose of minimizing data loss, pairwise deletion was utilized to treat missing data. Data imputation was not considered because of its tendency to underestimate standard errors and overestimate test statistic, hence producing biased results. Where the OLS model diagnostics (Lagrange multiplier lag and Lagrange multiplier error tests)25 indicated spatial dependence, the models were fitted with spatial lag or spatial error regression, as appropriate; otherwise OLS result was reported. Multiple block entries in multivariate spatial regression were performed using parsimonious backward approach. For the final multivariate model, we selected covariates that significantly predicted spatial pattern of stillbirth rate from the multiple regression models (models 1 to 5). The spatial regression model with the best goodness-of-fit was determined by R-squared, log likelihoods of the maximum likelihood estimations (MLE), Bayesian Information Criterion (BIC) and Akaike’s Information Criterion (AIC). To visually represent the key determinants of spatial pattern of stillbirth rates in high-dimensional geometry, a multivariate parallel coordinate plot was generated. As spatial models could only show the strengths of association but not interrelationships among the determinants studied, we further generated Bayesian network model with an artificial intelligence modeling and machine learning software—GeNIe v. 2.4.4 software27 to test our theories with the aggregated dataset for the 194 WHO countries. The network structure was generated based on the author’s programmatic knowledge of the temporal precedence of the variables. With significant level parameter set to p-value = 0.05 in PC structural learning algorithm, the variables were assigned as follows; tier 1 (region), tier 2 (gender inequality index), tier 3 (prevalence of anaemia in pregnancy and ANC coverage), tier 4 (skilled birth attendants during delivery), and tier 5 (stillbirth rate). The network structure was assessed to ensure that there was no residual confounding (through back-door path) and to show plausible paths among the significant variables from the final regression model. The network model learnt the parameters by using Expectation-Maximization algorithm to account for the missing values. The model accuracy was evaluated by using leave-one-out cross-validation method. The strengths of influence (corresponding contributions) for the variables were denoted by the thickness of Euclidean weighted and normalized arc widths. In addition, we assessed the direction of influence by using arc coloring in dynamic influence mode. The dynamic influence mode was used because it is context-specific and adjusts indirect influences, unlike static mode28. The green, red, grey and purple colors correspond to positive, negative, null and ambiguous influences of various factors on the probability distribution of stillbirth rate, respectively28. The aggregated, continuous data were grouped into categories based on either the median or mean values. The categorizations are as follows: stillbirth rate (high >12, low ≤12 per 1000 births); gender inequality index (high >0.4, low ≤0.4); ANC coverage (high >75%, low ≤75%); skilled birth attendants during delivery (high >85%, low ≤85%); and anaemia in pregnancy (high >40%, low ≤40%). By using supervised machine learning, we generated prediction scenario that reflected the probability of each parameter, given that the proportion of countries with low stillbirth rates increased to 100% (desired state) from baseline for each WHO regions. We performed sensitivity analysis to determine the extent to which stillbirth rate is accounted for by the explanatory variables. To validate the model calibration with the dataset, logarithmic loss, quadratic loss (Brier score) and spherical payoff were evaluated. Logarithmic loss values should be between zero and infinity, with zero indicating the best goodness-of-fit29. The quadratic loss value should lie between zero and two, where zero is the best performance29. Also, the spherical payoff should be between zero and one, with one indicating a perfect fit29. Similarly, the goodness-of-fit was assessed with area under receiver operating characteristics (ROC) and sensitivity tornado plot (Supplementary Fig. S1). This study was exempted from ethical review by the full complement of the University of Saskatchewan Behavioural Ethics Committee (ID# 1066) because it relied on publicly available aggregated de-identified dataset.