Explainable AI For Predictive Modelling of Maternal Health Behavior Under PMMVY: Accessibility, Benefit Adequacy, and District Heterogeneity

Abstract

This research proposes an explainable predictive model to evaluate maternal health behavior of beneficiaries of Pradhan Mantri Matru Vandana Yojana (PMMVY) in the districts of Karauli and Dholpur of Rajasthan, India. The analytical sample was cross-sectional of 350 beneficiaries in rural areas, equally distributed between the two districts. Maternal health behaviour was assessed by a three item composite score, and predictor variables were the demographic characteristics, four indicators of PMMVY accessibility, and perceived adequacy of programme benefits. Conventional benchmark models were evaluated using linear regression, Ridge regression, Elastic Net and a dummy regressor and nonlinear predictive models were assessed with repeated nested cross validation. The final Gradient Boosting model had an MAE of 0.208, an RMSE of 0.535 and an R² of 0.865, which is better than the best conventional model. Digitalisation related ease of enrolment was the most important predictor according to Permutation Feature Importance and SHAP analysis, followed by education and ease of documentation. The additional predictive value of perceived benefit adequacy was small. Pooled performance was relatively stable across predictor-profile grouping in robustness analyses, but there was considerable district heterogeneity and poor cross-district transportability. Generally, accessibility factors were more predictive than perceived benefit adequacy. The results highlight the importance of explainable predictive analysis for programme level maternal health assessment and the need for external validation prior to wider geographic application.