Fetal health state detection method based on parameters efficient ensembling of deep learning.

Background: The classification of cardiotocography (CTG) can assist obstetricians in assessing the health status of the fetus. However, traditional fetal heart rate monitoring data has the problem of strong subjectivity in manual interpretation, and deep learning models have poor representation ability on tabular data. Methods: This study proposed PLE-TabM, a tabular deep learning model combined piecewise linear encoding (PLE) and efficient weight integration. We used the public CTG dataset as the research subject. The PLE was used to improve the perception ability of feature segment intervals and TabM integrates multiple Multi-Layer Perceptron (MLP) weak classifiers. Results: The experimental results of fetal health status classification demonstrated that the performance of the PLE-TabM algorithm exceeding traditional machine learning methods. Its accuracy reached 95.77% and macro averaged F1 score reached 93.83%. Meanwhile, Gradient SHapley Additive exPlanations (Gradient SHAP) was used to analyze the feature importance that affects the classification. Finally, the algorithm was verified on 50 clinical patients. Conclusions: This study combines efficient tabular learning with interpretability analysis and applies it to the CTG classification problem, providing a reliable and objective tool to assist obstetricians in fetal monitoring and clinical decision making.