An Explainable Hybrid TabNet–Residual MLP Framework for Robust Fetal Health Prediction Using Focal Loss and Leakage-Aware Cross-Validation

Effective fetal health prediction is crucial for early diagnosis of fetal distress and avoiding negative perinatal effects. Cardiotocography (CTG), which measures fetal heart rate and uterine contractions, is a popular method for prenatal examination; however, conventional interpretation is arbitrary and unreliable. Fetal health prediction continues to face challenges due to scarce and imbalanced CTG datasets and data leakage during model evaluation, which can lead to poor generalization and inaccurate performance estimates. Moreover, the absence of explainable AI reduces model clarity and restricts clinical utilization. This study offers a hybrid deep learning model that uses TabNet, Residual Multi-Layer Perceptron (Residual MLP), and Focal Loss to classify normal, suspect, and pathological fetal states. TabNet allows for attention feature learning from CTG data, Residual MLP increases predictive robustness, and Focal Loss aids minority abnormal case diagnosis. To achieve a reliable evaluation, data splitting is used to create a pipeline designed to eliminate conventional train–test data leakage by performing data partitioning before model training and evaluation. SHAP and LIME enhance interpretability by offering clear global and local explanations. The suggested model obtains 99.30% accuracy, 99.10% balanced accuracy, and 98.65% pathological recall on the augmented dataset. A comparative evaluation shows that fixing data leakage drops overinflated baseline accuracy from 96.80% to 93.14%, emphasizing the necessity of a robust experimental design. The results show that the proposed framework outperforms cutting-edge fetal health prediction approaches, offering a robust, understandable, and clinically reliable alternative for CTG-based fetal health evaluation.