WHEN SHOULD PRE-SLEEP MONITORING BEGIN? AN EXPLAINABLE EVALUATION OF TEMPORAL WINDOWS FOR SLEEP FRAGMENTATION PREDICTION

Prediction of sleep fragmentation is increasingly important for early sleep quality assessment and personalized sleep monitoring systems; however, most existing studies rely on arbitrarily selected pre-sleep windows without systematically evaluating their temporal relevance. This study examines how pre-sleep observation window length affects predictive performance and when informative physiological signals begin to emerge before sleep onset. Using the Sleep-EDF Expanded dataset, we extracted EEG, EOG, and EMG features across four pre-sleep windows (15, 30, 60, and 90 min before sleep onset). Sleep fragmentation was operationalized as a binary outcome based on Wake After Sleep Onset (WASO > 30 min). Subject-independent cross-validation was used to evaluate Logistic Regression, Random Forest, and XGBoost models. Results showed that predictive performance improved progressively with increasing pre-sleep window duration, with the 90-minute window achieving the highest performance (Random Forest AUC = 0.792). SHAP analysis consistently identified EEG Delta and Theta spectral power as dominant contributors to fragmented sleep prediction within the optimal temporal window. These findings suggest that predictive physiological markers of fragmented sleep emerge progressively during the pre-sleep transition period and that temporal window selection is a critical factor in sleep monitoring system design. This study provides a systematic and explainable framework for optimizing pre-sleep monitoring in physiological sleep analysis.