A Machine Learning-Based Longitudinal Assessment of Maternal Health Service Utilisation in Karnali Province, Nepal (2001-2022).
Abstract
This study presents a machine learning-based longitudinal assessment of maternal health service utilisation across remote districts in Karnali Province, Nepal, covering a 22-year period (2001-2022). It utilises a harmonized longitudinal dataset (n=1,063) combined with Demographic and Health Surveys (DHS) spatial points and DHIS2 health facility surveillance data, a Spatial Friction of Distance Model is constructed establishing an empirical 40 km accessibility threshold. The research trained Spatial Random Forest classifiers achieving an AUC of 0.795 and F1-score of 0.707 to predict service delivery barriers. Furthermore, Causal Double Machine Learning (EconML) is applied with Random Forest nuisance models to evaluate health decentralization policy, deriving an overall Average Treatment Effect (ATE) of +21.8% and a Conditional Average Treatment Effect (CATE) of +24.0% in isolated mountain regions. Model interpretability is derived using SHAP (SHapley Additive exPlanations), identifying facility density and maternal education as primary levers for intervention. Master's Thesis in Data Science for York St John University, UK. Code Repository: https://github.com/prajwallyova/Manoj-Kumar-Aryal
