Leveraging machine learning for fetal health classification

Abstract

Early diagnosis of fetal health complications is crucial to ensure maternal and infant safety. However, conventional tabular machine learning approaches often suffer from severe majority-class bias and a lack of clinical interpretability. In this research, an optimized Python-based framework was developed using a 2,126-record dataset containing 21 predictive features to classify fetal health status. To resolve structural class imbalance without data leakage, the synthetic minority over-sampling technique (SMOTE) was dynamically applied strictly within stratified 5-fold cross-validation loops. Six configurations—naive Bayes, logistic regression, support vector machines, a multi-tiered deep multi-layer perceptron (deep MLP), random forest, and extreme gradient boosting (XGBoost)—were systematically tuned using randomized grid searches. Among the benchmarked models, the upgraded XGBoost ensemble demonstrated the highest performance, leading with a diagnostic accuracy of 94.87%, a macro F1-score of 91.37%, and an area under the curve (AUC) of 98.77%. While the deep MLP required a high training latency of 13.8179 seconds, the optimal XGBoost layout achieved superior power with a training runtime of only 1.8325 seconds and an inference speed of 0.0046 seconds. Game-theoretic Shapley additive explanations (SHAP) values confirmed abnormal short-term variability as the primary risk driver, ensuring a transparent, high-throughput decision-support tool.