Comparative Performance Assessment of HyCIRRF and Traditional Machine Learning Models for Preterm Pregnancy Risk Prediction

Abstract

Nowadays, high-risk maternal pregnancy is a serious healthcare limitation by the combined influence of gestational, demographic, clinical, and physiological factors. Precision and timely prediction of maternal healthcare risks is crucial for reducing neonatal and maternal complications. Moreover, existing prediction method often exhibit limited capability in integrating heterogeneous maternal fetal data and providing personalize risk assessment. This research study proposes a hybrid crow search and independent component analysis optimized regression random forest (HyCIRFF) model for maternal health risk prediction and compares its performance with tradition machine learning (ML) methods, as support vector machine (SVM), XGBoost, logistic regression (LR), decision tree (DT) with particle swarm optimization (PSO), BiLTCN, and decision tree (DT). The proposed outline was calculated using the publicly available Kaggle Maternal Health Risk Dataset, which comprises clinical metrics like age, heart rate, temperature, BP, and blood glucose, etc., collected through Internet-of-Things (IoT) based on maternal monitoring systems. The database was divided into 183 training and 203 testing samples. The proposed HyCIRRF framework integrates Fast Independent Component Analysis (FastICA) for feature transformation, CatBoost for feature ranking, Crow Search Optimization (CSO) for optimal feature selection, and a Regression Random Forest Classifier with a soft-voting ensemble plan for classifying low-medium, and high-risk pregnancies. Simulation outcomes demonstrate that the proposed model attain superior prediction performance, attaining 93.0% precision, 94.0% recall, 93.0 % F1-Score, 93.0% accuracy, and a 7.0% error rate. In comparison, the optimized SVM-RBF with PSO model attained 86.0% accuracy, 83.3 % precision, 81.7% recall, 83.3% F1-score, and 14.0% error rate, while other baseline techniques that merging feature transformation, meta-heuristic optimization, and ensemble learning (EL) significantly improves maternal health risk prediction while maintaining clinical interpretability and supporting early decision-making in real-time healthcare applications.