O004 Incorporating Arousals into Sleep vs. Wakefulness Classification Outperforms Traditional Binary Classification at 1-Second Epoch Resolution

Abstract Introduction 30-second epoch-based sleep scoring of polysomnography has key limitations in capturing physiological sleep-state dynamics. Particularly, epoch boundaries may not align with physiological sleep-state transitions (or transition frequencies); and the separate scoring of wakefulness and arousals artificially separates these phenomena. Method A deep-learning-based algorithm was developed for sleep vs. wakefulness/arousal classification with a 1-second epoch resolution. The Multi-Ethnic Study of Atherosclerosis Sleep dataset (n=2034) was used to first train and evaluate traditional sleep vs. wakefulness classification (Model 1). Afterwards, it was used to classify sleep vs. wakefulness/arousals, incorporating arousals with a heuristic correction for sleep-to-wake transitions when required (Model 2). Additionally, we analysed how segments with arousals are treated by both models. Results Model 1 achieved an overall concordance of 90.38% (Cohen’s kappa κ=0.80, global F1 score=0.89) with manual scoring in the evaluation set (n=306), while Model 2 attained a concordance of 91.04% (κ=0.82, F1=0.90). Model 2 had slightly improved performance compared to Model 1 in classifying both wakefulness (91.9% vs. 90.1%) and sleep (92.9% vs. 92.7), p<0.05 (McNemar’s test). Model 2 also more accurately classified arousal segments as a form of wakefulness (61.1% vs. 51.5%). Discussion Despite training on a conventionally annotated dataset with 30-second epochs, Model 2 with arousals incorporated outperformed Model 1’s traditional binary classification concordance at 1-second epoch resolution. These results highlight the potential of using shorter epochs for sleep classification by treating arousals similarly to wakefulness. This approach may capture nuanced sleep-wake transitional characteristics but requires validation against datasets scored on a continuous basis.