A hybrid stacked ensemble and Kernel SHAP-based model for intelligent cardiotocography classification and...

Area of Science:

• Medical Informatics
• Machine Learning
• Obstetrics

Background:

• Intelligent cardiotocography (CTG) classification aids fetal health evaluation.
• Complex machine learning (ML) models offer high performance but lack interpretability, hindering clinical adoption.
• A gap exists in balancing accuracy and interpretability for ML-based CTG classification in prenatal care.

Purpose of the Study:

• To enhance CTG classification performance and prediction interpretability.
• To develop a hybrid model integrating stacked ensemble methods with feature analysis.
• To validate the model's effectiveness and interpretability using public and private datasets.

Main Methods:

• A stacked ensemble classifier was built using Support Vector Machines (SVM), Extreme Gradient Boosting (XGB), and Random Forests (RF) as base learners.
• A Backpropagation (BP) meta-learner integrated CTG features with base learner outputs.
• Kernel SHapley Additive exPlanations (SHAP) framework was employed for feature contribution analysis.

Main Results:

• The hybrid model achieved high accuracy (0.9539 public, 0.9201 private) and F1 scores (0.9249 public, 0.8926 private) with 10-fold cross-validation.
• Key determinants for CTG classification were identified as accelerations (AC) and the percentage of time with abnormal short-term variability (ASTV).
• Increased ASTV correlated with higher abnormality probability, while increased AC correlated with higher normal status likelihood.

Conclusions:

• The proposed hybrid model demonstrates strong classification performance for intelligent fetal monitoring.
• The model provides reasonable interpretability, identifying crucial features influencing fetal state predictions.
• This approach addresses the accuracy-interpretability trade-off, facilitating potential clinical integration.