Hybridization of Random Forest, LightGBM and Deep Learning for Maternal Health Risk Prediction
Abstract
Maternal health risk prediction is an important healthcare decision-support task, since the early identification of low-risk, mid-risk, and high-risk pregnancies supports timely monitoring, referral, and intervention. The benchmark MaternalNET-RF model combined an Artificial Neural Network (ANN) and a Random Forest (RF) through maximum probability voting and achieved 94.88% accuracy with an area under the receiver operating characteristic curve (AUC) of 0.9622, but it was limited by its use of only two classifiers, the absence of clinical feature engineering and data resampling, and an unweighted voting rule. This study developed an enhanced Hybrid Deep-Ensemble Framework for Maternal Health Risk prediction (HDEF-MHR) to address these limitations using the Maternal Health Risk dataset of 1,014 records described by six clinical variables and a three-class target. The model applied outlier correction, label-noise correction, feature scaling, clinical feature engineering that expanded the six raw variables into fifteen features, and Synthetic Minority Over-sampling Technique (SMOTE) class balancing on the training data, and it integrated three algorithmically diverse base learners, namely RF, Light Gradient Boosting Machine (LightGBM), and a deep ANN, combined through optimized soft voting and clinical threshold adjustment. Performance was evaluated across 20 stratified random seeds using accuracy, precision, recall, F1-score, AUC, balanced accuracy, the Matthews correlation coefficient (MCC), and specificity, and was reported as the mean ± standard deviation. The proposed model achieved a mean accuracy of 95.42%, a mean weighted precision of 95.58%, a mean weighted F1-score of 95.43%, and a mean AUC of 0.9910, compared with the benchmark's 94.88% accuracy and 0.9622 AUC. Although the accuracy gain was not statistically significant (p = 0.238), the substantial improvement in AUC from 0.9622 to 0.9910 indicates stronger risk discrimination. The study concludes that combining clinical feature engineering, class balancing, three diverse base learners, optimized soft voting, and threshold adjustment provides a stronger and more stable framework for maternal health risk classification, which can support future maternal healthcare decision-support systems after validation on larger and more diverse datasets.