Temporal Ensemble Learning for Robust Maternal Health Risk Prediction in Dynamic Clinical Environments

Abstract

Maternal mortality and morbidity continue to pose serious public health concerns worldwide, especially in resource-constrained regions where delayed identification of complications limits timely medical intervention. Conventional risk assessment approaches often depend on manual evaluation and fixed clinical thresholds, which are insufficient for capturing the complex and dynamic physiological changes that occur during pregnancy. To address these limitations, this study introduces a real-time Maternal Health Risk Stratification System (MHRSSS) based on a stacked Long ShortTerm Memory (LSTM) ensemble framework. The proposed system leverages a comprehensive dataset of maternal health indicators, including age, blood pressure, blood glucose levels, and heart rate, to enable accurate risk prediction. To mitigate the impact of class imbalance commonly observed in medical datasets, the Synthetic Minority Over-sampling Technique (SMOTE) is applied, enhancing the model’s ability to detect high-risk cases effectively. The architecture integrates a hybrid ensemble strategy, combining traditional machine learning models such as Random Forest and Gradient Boosting with an LSTM network to capture both static and temporal patterns in the data. A softvoting mechanism is employed to aggregate model predictions and produce the final risk classification. Experimental evaluation demonstrates that the proposed framework significantly outperforms individual baseline models, achieving an accuracy, precision, recall, and F1-score of 99.03%. In comparison, the Extra Trees Classifier (ETC) and Random Forest Classifier (RFC) achieved notably lower performance levels. The system is implemented through a user-friendly Tkinter-based interface, enabling healthcare professionals to perform real-time, data-driven risk assessment and support early clinical decision-making, ultimately contributing to improved maternal health outcomes.