A hybrid explainable artificial intelligence framework for equitable prediction and analysis of postnatal care uptake in Ghana: Integrating ensemble learning, counterfactuals, and causal inference

Objectives To develop and validate a hybrid, equity-focused artificial intelligence (AI) framework for predicting and analysing postnatal care (PNC) uptake among women in Ghana, with emphasis on minority-class sensitivity, model interpretability, and actionable intervention insights. Study design Cross-sectional, secondary analysis of nationally representative survey data. Methods Using data from the 2022 Ghana Demographic and Health Survey (n = 3314), we developed a hybrid AI framework combining ensemble learning (Logistic Regression and Support Vector Machine with soft voting and threshold tuning), explainable AI (SHAP analysis), counterfactual reasoning (DiCE), and causal inference (doubly robust estimators for average and conditional average treatment effects). The Synthetic Minority Over-Sampling Technique (SMOTE) was applied to address class imbalance and stratified 5-fold cross-validation was used for model evaluation. Doubly robust causal inference was conducted using the econml library, with propensity score modelling used to account for treatment assignment, and positivity and exchangeability assumptions assessed through covariate overlap diagnostics. Results The ensemble model calibrated at a threshold of 0.45 achieved a Class 0 recall of 0.31 and an area under the precision-recall curve (AP) of 0.97, outperforming all individual baseline classifiers. SHAP analysis identified maternal age, education level, wealth status, caesarean section, and sex of household head as the most influential predictors of PNC non-utilization. Counterfactual analysis revealed that minimal, actionable changes including increased media exposure, improved educational attainment, or change of residence could shift individual predictions from non-utilization to utilization of PNC. Causal inference demonstrated significant heterogeneity in treatment responsiveness, with women aged 25–34, those with secondary education or above, urban residents, and those in female-headed households exhibiting the highest conditional average treatment effects. Conclusions This interpretable, fairness-aware AI framework advances the prediction and analysis of PNC uptake in Ghana. By prioritizing sensitivity to underserved women and integrating explainability, counterfactual reasoning, and causal inference, the framework provides empirical evidence for targeted, equitable maternal health interventions and is transferable to other low- and middle-income country settings.