Utilisation of maternal, newborn and child healthcare services in three sub-Saharan African countries (DRC, Kenya, and Tanzania) using Demographic Health Surveys data from 2007-2016: Application of Generalised Structural Equation and Machine Learning Models

Abstract

Background: The risk of child deaths within the first month of life is elevated than the later stages of childhood. Globally, Sub-Saharan Africa (SSA) has the highest neonatal mortality. Majority of the countries in SSA including the DRC, Kenya and Tanzania are struggling to meet Sustainable Development Goal (SDG) 3.2 of reducing the neonatal mortality rate to 12 deaths per 1,000 live births by 2030 (2). Most causes of neonatal deaths are preventable and treatable. Universal coverage, timely and effective utilisation of maternal, newborn, and child healthcare (MNCH) services during pregnancy, delivery, and postpartum has the potential to save many lives of newborns in high-burden countries. vii Antenatal care (ANC) is the first service offered to pregnant women in MNCH. The timing and frequency of ANC visits is critical for the mother and her unborn child. The WHO recommends that women initiate ANC within 16 weeks of pregnancy and attend a minimum of four ANC visits for timely and optimum care before delivery (3, 4). The WHO also recommends that pregnant women receive assistance from a skilled worker during delivery and get postnatal checks with their newborns within 6 weeks of delivery (5, 6). Furthermore, utilising the Continuum of Care (CoC) for MNCH could significantly reduce maternal and newborn deaths in SSA. In the context of MNCH, the CoC is an approach that ensures continuous care from the period of pregnancy, through to childbirth, postnatal period, infancy, and the childhood period (7). Despite the recognition of the use of vital services in MNCH, timely and adequate uptake of MNCH services remains poor and the coverage of MNCH is far from universal in SSA. Most pregnant women initiate ANC after 16 weeks and hence fail to receive timely ANC interventions (8). Uptake of ANC visits, skilled birth attendance (SBA) and postnatal care (PNC) is suboptimal (8-11). Studies in SSA have explored various factors associated with MNCH services utilisation, however, our understanding of MNCH services utilisation in SSA is still limited. Trends in utilisation of MNCH services over time such as late ANC uptake have not been thoroughly assessed. Late uptake of ANC is still a common problem in SSA. Tracking women’s progress in the timing of ANC will ascertain if they are any changes in women’s late uptake of ANC and the contributing factors. This information will guide future policies and programmes which focus on improving the timely uptake of ANC in the SGD era. There is also a dearth of empirical evidence on the factors associated with the utilisation of ANC, skilled delivery and postnatal care in the CoC using nationally representative data. The CoC views both the mother and child as a collective rather than as separate/ individual entities. Understanding factors that viii contribute to the full utilisation of drop out from the CoC is essential for the formulation of interventions than enhance the CoC. Furthermore, studies which investigated either the individual utilisation of MNCH services such as timing of ANC, ANC visits, SBA and PNC services or the CoC have tended to use more of the traditional analysis methods such as the logistic regression. The application of more versatile analysis methods such as Machine Learning is not common. Machine Learning methods are capable of extracting information that commonly used methods (logistic regression) fail to do by uncovering hidden patterns and relationships, particularly in large data sets (12). The application of Machine Learning methods can offer opportunities of enhancing existing methods (conventional regression methods) for predicting and classifying MNCH utilisation leading to more effective interventions to improve MNCH utilisation. There is also a limited understanding on the interrelationships between MNCH services utilisation and neonatal outcomes. The associations between MNCH services utilisation and newborn outcomes such as neonatal mortality are commonly assessed using traditional approaches that assume direct associations. Specific analytical methods, such as Generalised Structural Equation Modeling (GSEM) can be used to model complex relationships such as interrelated links between utilisation of different MNCH services and neonatal outcomes. GSEM gives a clear understanding of how different services of MNCH are related to one another with neonatal outcomes by estimating both direct and indirect paths associations for more effective targeted interventions. Given the critical role of MNCH in ending preventable neonatal mortality, the overarching aim of this study was to describe the utilisation of MNCH services and their associations with neonatal mortality using GSEM and Machine Learning models in three sub-Saharan African countries: the DRC, Kenya, and Tanzania. ix Methods: The study utilised cross-sectional secondary data of reproductive-age women from the Democratic Republic of Congo (DRC) (2007-2013/14), Kenya (2008-2014) and Tanzania (2010-2015/16) Demographic Health Surveys. Firstly, the multivariate logistic regression analysed factors associated with late ANC initiation accounting for clusters, survey weights and stratification for the different rounds of the Demographic Health Surveys. Trends in late initiation of ANC over time in each country were assessed by comparing the earlier and later surveys using differences in prediction scores (prediction probabilities generated after running the multivariate logistic regression models). Secondly, the study assessed the main predictors of non-utilisation of PNC using the Decision Tree. The model performance of the Decision Tree was compared to the Logistic Regression using Accuracy, Sensitivity, Specificity and area under the Receiver Operating Characteristics. Thirdly, factors associated with the drop out from the MNCH continuum, defined as not fully utilising either ANC, SBA, or PNC services, were analysed using multivariate logistic regression accounting for clusters, survey weights and stratification. Machine Learning analysis was used to predict the drop out from the MNCH continuum using features (predictors) that were found significant in the multivariate logistic regression. Five classification Machine Learning models were built and developed including the Artificial Neural Network, Decision Tree, Logistic Regression, Random Forest and Support Vector Machine to predict the drop out from the MNCH continuum. The prediction accuracies of the models were then compared using parameters including Accuracy, Precision, Recall, Specificity, F1 score and area under the Receiver Operating Characteristics. Fourthly, the Generalised Structural Equation Modeling (GSEM) was used to assess the mediatory role of MNCH services utilisation on neonatal mortality. The endogenous variables x were ANC attendance, SBA and PNC attendance, low birth weight and neonatal mortality. The GSEM analysis also accounted for survey weights and considered cluster random effects. Results: The findings showed a reduction in late ANC initiation (67.8%-60.5%) between 2008-2014 in Kenya as well as in Tanzania (60.9%-49.8%) between 2010-2016, but an increase was observed in the DRC (56.8%-61.0%) between 2007-2014. In the DRC, higher birth order was associated with ANC initiation delays from 2007-2014, whilst rural residency, lower maternal education and household income was linked to ANC initiation delays in 2014. In Kenya, lower maternal education and household income was associated with ANC initiation delays from 2008-2014, whilst rural residency and increased birth order were linked to ANC initiation delays in 2014. In Tanzania, higher birth order and larger households were linked to ANC delays from 2010-2016, whilst ANC initiation delays were associated with lower maternal education in 2010 and lower-income households in 2016. The results also showed that the Decision Tree models had higher prediction accuracy of non- utilisation of PNC than the Logistic Regression models. Using the Decision Tree, low quality of ANC, home deliveries and unemployment were associated with the highest probability of not utilising PNC (92.0%) in the DRC. In Kenya, home deliveries, unemployment and lack of access to mass media were associated the highest likelihood of not utilising PNC (87.0%). In Tanzania, home deliveries, low quality of ANC and unwanted pregnancies exhibited the highest likelihood of not utilising PNC (100.0%). The results also revealed very high rates of dropping out from the MNCH continuum in the DRC (91.0%), in Kenya (72.3%) and Tanzania (93.7%). Rural residence, lower maternal education and non-exposure to mass media were common predictors of dropping out from the MNCH continuum across the three countries. Further, the influence of factors such as xi household wealth, household size, access to money for medication, travel distance to health facilities, and parity and maternal age varied by country. Results from the Machine Learning analysis showed that the Logistic Regression had the least prediction accuracy, while the Random Forest exhibited the highest prediction accuracy. Using the Random Forest, the study further ranked the most important predictors of the drop out from the MNCH continuum. Household wealth, place of residence, maternal education and exposure to mass media were the top four most important predictors. The results also showed direct and indirect associations between MNCH services utilisation and neonatal mortality. ANC attendance mediated the total effects of PNC attendance on neonatal mortality by 8.8% in Kenya and 5.5% in Tanzania. ANC attendance and SBA also sequentially mediated the total effects of PNC attendance on neonatal mortality by 1.9% in Kenya and 1.0% in Tanzania. The results in Tanzania also showed ANC attendance mediated 2.8% of the total effects of LBW on neonatal mortality. No presence of mediation was observed in the DRC; however, ANC attendance moderated the relationship between parity and neonatal mortality. Conclusions: The study found that late uptake of ANC decreased between the two survey rounds in Kenya and Tanzania but increased in the DRC. Women from various geographic, educational, parity, and economic groups showed varying levels of late ANC uptake. Increasing women’s access to information platforms and strengthening initiatives that enhance female education, household incomes, and localise services may enhance early ANC uptake. The Decision Tree models showed higher prediction accuracy of non-utilisation of PNC than the Logistic Regression models in the DRC, Kenya and Tanzania. Using the Decision Tree, women who had poor quality of ANC, home deliveries, unemployment, unplanned pregnancies, and no mass media access were identified as high-risk subpopulations of non- xii utilisation of PNC. Improving access and quality of care, incorporation of TBAs into the formal health systems, government health financing, increasing access to mass media and integrating maternal healthcare services with family planning services should be considered as top priority interventions to improving the utilisation of PNC. Most women and children drop out of the MNCH continuum in the DRC, Kenya and Tanzania. Rural residence, lower maternal education and non-exposure to mass media were common factors linked to the high dropout in the MNCH continuum. The use of Machine Learning can help support evidence-based decisions in MNCH interventions. Rapid response mechanisms such as web-based applications can also be developed through the use Machine Learning whereby a pregnant woman’s future utilisation of the services in CoC is assessed and monitored in real-time. The GSEM findings showed interconnections between MNCH services utilisation such as timing of ANC, ANC visits, SBA, PNC and neonatal mortality. This suggests that more than direct and indirect factors are accountable for the associations between MNCH services utilisation and neonatal mortality. The mediation role of MNCH services on neonatal mortality indicates critical areas for targeted interventions to reduce neonatal mortality. Overall, the study aimed to describe the utilisation of the MNCH services and associations with neonatal mortality in the DRC, Kenya and Tanzania. The study showed declines in late ANC uptake in two countries, however, early uptake of ANC is far is still not universal. The study also showed very low levels of retention in the CoC, and most women and children drop out in the CoC at postpartum period. The findings also showed the existence of social, health system and individual inequalities in MNCH and their impact on early childhood survival. Women who are vulnerable to unequal and poor MNCH services utilisation are characterised by poverty, rural residence, long travel distances to health facilities, unaffordable medical expenses, home deliveries, low quality of xiii care, low education, high parity, younger age, unemployment, limited exposure to mass media, and unplanned pregnancies. Context-specific intervention programs such as female education, government health financing, MNCH promotion programs through mass media and improved accessibility and quality of care in health facilities, particularly for the most vulnerable groups of the populations such as women of low socioeconomic status and women from underserved rural areas are essential to improve the overall health of mothers and children and meeting the SDG-3 goals. Modern biostatistical models like Machine Learning provide essential tools to understand public health problems. These techniques should be applied to complement the conventional statistical methods, particularly the tree-based models like the Decision Tree and Random Forest for predicting and classifying the utilisation of MNCH services. The GSEM established interconnections between timing of ANC, ANC visits, SBA and PNC and neonatal mortality. The timing of the first ANC contact is an important starting point to a continuation through the COC. It makes women better informed about pregnancy and the subsequent use of MNCH services. All stakeholders should work more on promoting early uptake of ANC by setting up initiatives that increase women’s access to information platforms, enhance female education, improve household incomes, and bring services closer to communities.