Abstract
Abstract
Sleep architecture is characterized by dynamic transitions between non-rapid eye movement (NREM) and rapid eye movement (REM) stages, each defined by distinct oscillatory signatures. Although quantitative EEG has been widely used to study primary insomnia (PI), reported findings remain inconsistent, partly due to limited consideration of cortical regional differences, sleep-stage specificity, and sex-related variability. This study integrates spectral power analysis with explainable machine learning to provide a comprehensive characterization of neural dynamics in PI. EEG recordings from 100 adults (50 healthy sleepers and 50 PI patients) were analyzed across frontal (F3, F4), central (C3, C4), and occipital (O1, O2) regions. Stage-specific classifiers were developed for NREM and REM using LASSO, random forest, CatBoost and XGBoost models, and Shapley Additive Explanations (SHAP) were employed to quantify the contribution of each spectral and regional feature to model predictions. Models trained on NREM features demonstrated superior discrimination compared with those trained on REM features, suggesting that slow-wave neural activity captures insomnia-related alterations more reliably than REM dynamics. SHAP analysis identified reduced delta power and elevated beta activity most prominently in central regions as the dominant predictors of PI during NREM sleep. In contrast, REM-based models pointed to prefrontal beta1 and occipital alpha as secondary discriminative features. Sex-stratified analyses further revealed systematically higher spectral power in females and distinct SHAP feature-importance patterns between sexes. Collectively, this work establishes an interpretable, stage-specific framework for identifying EEG biomarkers of insomnia and demonstrates the value of explainable artificial intelligence in advancing the neurophysiological understanding of sleep disturbance.