SHAE: A SHAP-guided adaptive ensemble framework for automatic sleep staging

Area of Science:

• * Computational neuroscience
• * Biomedical engineering
• * Machine learning in healthcare

Background:

• * Manual sleep scoring (polysomnography) is laborious and inconsistent.
• * Existing automated systems lack stage-specific adaptability.
• * Need for improved accuracy in sleep disorder diagnosis and monitoring.

Purpose of the Study:

• * Introduce SHAE, an adaptive heterogeneous ensemble for automated sleep staging.
• * Enhance stage-specific feature selection and ensemble configuration.
• * Improve the accuracy and reliability of automated sleep staging.

Main Methods:

• * Developed SHAE framework combining SHapley Additive Explanations (SHAP) and Alpha Evolution (AE).
• * Utilized SHAP for stage-specific feature ranking and AE for joint optimization of feature subsets and ensemble weights.
• * Implemented a hybrid model with five one-vs-rest binary classifiers and one global five-class classifier.

Main Results:

• * SHAE achieved high accuracy (up to 89.94%) and macro-F1 scores (up to 85.80%) on benchmark datasets (SleepEDF-20, SleepEDF-78, DREAMS).
• * Outperformed existing baseline methods across all tested datasets.
• * Ablation studies validated the effectiveness of stage-specific feature selection and adaptive optimization.

Conclusions:

• * SHAE offers a robust and adaptive solution for automated sleep staging.
• * The proposed framework significantly advances the state-of-the-art in computational sleep analysis.
• * SHAE has potential applications in clinical sleep medicine and neuroscience research.