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

Sleep staging underpins sleep medicine and neuroscience research, forming the clinical basis for sleep disorder diagnosis, sleep quality assessment, and neurological disease monitoring. Manual PSG scoring is time-intensive and subject to inter-rater variability, motivating the development of automated staging systems. Existing approaches often rely on uniform feature representations across sleep stages and fixed ensemble configurations, limiting their stage-specific adaptability. These observations motivate SHAE, which combines SHAP-guided stage-specific feature ranking with AE-based adaptive optimization of feature subset sizes and ensemble weights. This paper proposes SHAE, an adaptive heterogeneous ensemble framework combining SHapley Additive Explanations (SHAP)-based stage-specific feature selection with Alpha Evolution (AE) joint optimization. SHAE comprises five one-vs-rest binary classifiers - each targeting a single sleep stage - and a global five-class classifier. SHAP constructs stage-specific feature rankings for each binary task; AE jointly optimizes feature subset sizes and ensemble weights without requiring gradient information, enabling data-driven configuration learning. On SleepEDF-20, SleepEDF-78, and DREAMS, SHAE achieved accuracies of 89.94%, 87.74%, and 87.36%, macro-F1 scores of 85.80%, 82.78%, and 83.52%, and Kappa coefficients of 0.88, 0.83, and 0.84, respectively, outperforming all baselines across datasets. Ablation experiments confirm the contribution of both stage-specific feature selection and adaptive weight optimization.