medRxivJClub
A Unified Flexible Large Polysomnography Model for Sleep Staging and Downstream Clinical Screening
medRxiv · · Preprint · Open access
Deng, Niu + 15 more
Abstract ↗AI summary
The abstract is read at the publisher; the summary is JClub's.
LPSGM, trained on 220,500 hours of PSG from 16 public datasets, achieves expert-level sleep staging with kappa = 0.845 ± 0.066 versus inter-expert kappa = 0.850 ± 0.102.
- Why it matters: PSG scoring is labor-intensive and models often fail to generalize across centers, limiting reliable sleep staging and PSG-based screening for conditions like narcolepsy, OSA, and depression.
- What they did: LPSGM uses domain-adaptive pre-training, flexible channel configurations, and a unified architecture, then is validated prospectively and on independent private cohorts for staging and downstream screening.
- The result: It reaches 84.2% accuracy (kappa = 0.761) for three-class narcolepsy, AUC 0.879 for normal-versus-abnormal hypersomnia transfer, AUC 0.874 for severe OSA without respiratory signals, and AUC 0.801/0.965 for depression screening across APPLES and a clinically diagnosed cohort.
The findingWhy it mattersWhat they didThe result
- Open access
- 4 cites