深層学習を用いた時間的連続自動睡眠覚醒分類

Segmenting sleep into fixed 30-second epochs remains central to current sleep scoring practice, yet it imposes rigid boundaries that may not accurately reflect the true temporal sleep dynamics. We aimed to develop a deep learning-based, high-temporal-resolution sleep-wake classifier leveraging temporally continuous manual reference scoring without fixed epoch boundaries and transfer learning techniques to facilitate progress toward a more physiologically consistent sleep assessment.

Three independent datasets were utilized, of which two included sleep-wake scoring manually conducted in a temporally continuous manner. A U-Net based model was initially trained on a large dataset scored using 30-second epochs, with post hoc scoring modifications (n=2034). It was then fine-tuned via transfer learning using a subset of one of the datasets with temporally continuous scoring (n=39) and validated on both its holdout portion (n=40) and the other independent temporally continuous scoring dataset (n=20). Wakefulness and arousals were consolidated, acknowledging their shared physiological characteristics. Prediction confidence estimates were also generated.

The model achieved overall concordance of 88.96% (κ=0.78) and 88.23% (κ=0.76) in the holdout and second independent evaluation dataset, respectively, with temporally continuous scoring. Correlation between 1-second automatic predictions and temporally continuous manual scoring was r=0.93 (p<0.001) for total sleep time and r=0.67 (p<0.001) for sleep-to-wake transition index.

These findings support the utility of our model in addressing key limitations of 30-second epoch-based scoring and progressing toward more physiologically consistent sleep-wake assessment by providing a practical basis for subsequent analyses. Misclassifications generally showed lower confidences, indicating additional value for targeted review.