Using Machine Learning to unravel health utilization drivers in Bangladesh

Globally, countries have taken several exemplary pathways to improve reproductive, maternal, neonatal, child health, and nutrition (RMNCH-N) outcomes over the last decades.  Foremost is the role of timely and appropriate RMNCH-N service utilization investments and its determinants. Recent findings show that 95 percent of deaths from diarrhea and 67 percent of deaths from pneumonia in under-5 children globally, by 2025, could be prevented by prioritized interventions. In addition, global evidence also highlight both the equal importance of demand and supply-side determinants, and changes in the relative importance of determinants at different phases in a country’s progress trajectory. For example, while certain cultural factors may be more dominant in the early stages of progress, new factors like mass media influence may emerge later on. A recent global assessment has shown that countries which address a handful of context and time-specific RMNCH-N service determinants could reduce the fertility rate to 2.5 in certain Asian settings compared to generic health interventions.

Identifying and predicting the role and magnitude of demand and supply-side determinants, however, is difficult due to their variability and complexity, and requires regular tracking.

To solve this challenge, a recent policy research working paper endeavored to use Machine Learning (ML) methods to identify priority investments that could help Bangladesh accelerate progress towards RMNCH-N utilization. Notwithstanding noticeable improvements in the RMNCH-N landscape, uptake of certain key services such as institutional delivery, skilled birth attendance, and postnatal care visits have not been adequate to reach the country’s RMNCH-N goals. Inequities, for example, persist in the areas of service utilization and health status among different socio-economic groups. Accelerating further progress, therefore, hinges on strategic targeted investments in prioritized determinants.

To inform this, supervised ML algorithms were developed to compare the relative importance of over 30 demand- and supply-side determinants of 19 key RMNCH-N indicators related to service utilization, quality of care and health/nutrition outcomes. ML, a subset of Artificial Intelligence, imitates human learning processes and can effectively and efficiently analyze historic data and complex relations for decision making and predictions. As such, this method enabled comparative analysis of large, combined data sets from health facility surveys and the demographic and health surveys over the span of a decade.

Findings indicate that key supply-side determinants could provide a thrust towards further increases in utilization in contrast with earlier findings when demand-side determinants (e.g., age, and birth order) were more dominant. Supply-side determinants of particular importance include availability of skilled staff, functional readiness of health facilities, and quality of care. Demand-side awareness has improved substantially, with women, overall, being more influenced by service availability and quality than by cultural barriers. This shift could be considered progressive since women increasingly choose to seek care if quality services are perceived to be available at health facilities. Findings also revealed heavy reliance on the private for-profit sector for RMNCH-N care (except for postnatal care).

Wealth and education status remained as two relatively important demand-side determinants for predicting outcomes. As such, findings also point to the regressive role of wealth status on utilization, indicating that perhaps the prevailing user fees exemption alone might not be adequate to improve RMNCH-N service uptake. Rather, it may be pertinent to also consider addressing the direct and indirect costs of care through demand-side financial incentives. Equally important will be the need to improve the readiness of public facilities for RMNCH-N care provision given the relatively lower influence of public facilities (vis a vis the private for-profit sector). This will ensure that the current momentum of care-seeking is maintained. Conversely, the sincere care-seeking of pregnant women and mothers may not result in proportionate improvements in RMNCH-N status and reductions in mortality.

Strategies that improve the engagement of community health workers (CHW) in RMNCH-N utilization could also help to boost utilization patterns.  The influence of CHWs in maternal and childcare utilization (except family planning) was found to be low. Interestingly, findings revealed that women who have access to mass media stand a better chance of utilizing RMNCH-N services. These findings underscore the potential of mobile health technology in augmenting women’s awareness, and in supporting CHW capacity and connectivity with communities. 

Areas for future research

Relative to more traditional methods, machine learning methods increased the efficiency and precision of this analysis by better capturing non-linear relationships more clearly. Notwithstanding, the supervised learning algorithm in this study is a subcategory machine learning class designed to reduce biases, nevertheless it still faces the limitation of data, and, although informative, it does not provide a full picture. Additional analytical work is therefore needed to better understand the causal underpinnings between, for example, financial access of poor women, the role of CHWs and improved RMNCH-N outcomes.