Seonglae ChoThe input consists of multiple signals from PSG (polysomnography): EEG/EOG, ECG, EMG, respiratory signals, etc. Self-supervised pretraining was performed on approximately 585,000 hours of sleep study data from over 65,000 individuals. The model showed meaningful predictive power for 130 conditions, with particularly high performance for all-cause mortality, dementia, myocardial infarction, heart failure, chronic kidney disease, stroke, atrial fibrillation, etc.
- mortality C-index 0.84
- dementia 0.85
- myocardial infarction 0.81
- heart failure 0.80
- CKD 0.79
- stroke 0.78
Performance was well maintained even when transfer learning was applied to the external cohort SHHS. This is an attempt to use sleep PSG not simply for diagnosing sleep disorders, but as a broad disease risk biomarker.
www.nature.com
https://www.nature.com/articles/s41591-025-04133-4