A Stacked LSTM-Ensemble Framework for Real-Time Maternal Health Risk Stratification System

Abstract

Maternal mortality and morbidity remain significant global public health challenges, particularly in low-resource settings where timely clinical intervention is often hindered by delayed risk detection. Traditional risk assessment methods rely heavily on manual triage and static clinical thresholds, which frequently fail to capture the complex, non-linear physiological patterns associated with pregnancy-related complications. The research proposes a stacked LSTM-ensemble framework for a real-time Maternal Health Risk Stratification System (MHRSSS). The system utilizes a dataset of maternal vitals, including age, blood pressure, blood glucose, and heart rate. To address inherent data imbalances, the Synthetic Minority Over-sampling Technique (SMOTE) was employed, ensuring robust sensitivity toward high-risk cases. The core architecture implements a hybrid approach, combining a stacking ensemble (utilizing random forest and gradient boosting) with a Long ShortTerm Memory (LSTM) neural network. The final classification is determined through a soft-voting probability fusion mechanism. Experimental results demonstrate that the proposed hybrid framework significantly outperforms standalone models, achieving an accuracy, precision, recall, and F1 score of 99.03%, compared to 88.23% for the Extra Trees Classifier (ETC) and 70.36% for the Random Forest Classifier (RFC). The system is deployed via a user-friendly Tkinter GUI, providing healthcare practitioners with a reliable, real-time tool for objective risk stratification, thereby facilitating early medical intervention and improving maternal health outcomes.