Development of an Explainable Ai-Driven Early Warning System for Predicting Postpartum Haemorrhage in Low-Resource Environments

Abstract

Postpartum haemorrhage (PPH) is a leading cause of maternal morbidity and mortality worldwide, with a disproportionate burden in low-resource settings where timely diagnosis is often difficult. This study presents an interpretable AI-based early warning framework for predicting PPH using machine learning. A retrospective dataset of 223 anonymized obstetric records from Mpilo Central Hospital, Zimbabwe, was used for model development. Preprocessing involved missing-value imputation, outlier capping, feature engineering, and normalization, generating clinically relevant features such as labour duration in minutes, binary obstetric risk indicators, and a composite maternal risk score. Random Forest and Multilayer Perceptron (MLP) models were trained and evaluated on an 80:20 split. Random Forest achieved superior performance, with 86.67% accuracy, 89.47% precision, 80.95% recall, an F1-score of 85.00%, and ROC-AUC of 0.8730 for the PPH class. Labour duration and delivery method were the most influential predictors. SHAP (Shapley Additive Explanations) was used to enhance interpretability and clinical transparency. The findings show that interpretable machine learning can deliver reliable predictive performance even on small clinical datasets from resource-constrained settings, supporting proactive obstetric intervention and improved decision-making.