A Machine Learning Framework for Maternal Health Risk Prediction Using Vital Signs and Class Imbalance Correction

Abstract

Maternal health complications often develop gradually and are difficult to detect using traditional manual screening methods. This study proposes a machine learning framework for early materna/l risk prediction using six routine clinical vitals: maternal age, systolic and diastolic blood pressure, blood sugar, body temperature, and heart rate. A dataset of 1014 patient records was used to develop a multi-class classifier capable of identifying low, mid, and high-risk cases. A major challenge was class imbalance, particularly the underrepresentation of mid-risk cases. To address this, the Synthetic Minority Oversampling Technique (SMOTE) was applied only to training folds within a stratified 10-fold cross-validation pipeline to prevent data leakage. Multiple models were evaluated, with Random Forests and shallow neural networks achieving the strongest performance, nearly 89% accuracy. Results show substantial improvement in detecting minority classes after balancing. The proposed system demonstrates that simple vital sign data can support reliable, low-cost maternal risk assessment in limited-resource settings.