A Unified Flexible Large Polysomnography Model for Sleep Staging and Downstream Clinical Screening (LPSGM)

This medRxiv preprint introduces LPSGM, a flexible-channel large polysomnography model for sleep staging and downstream clinical screening. The model handles one to nine standard EEG, EOG, and chin channels, was pretrained on about 220,500 hours from 16 public PSG datasets, and reports prospective staging kappa near inter-expert agreement plus screening AUCs for narcolepsy, severe OSA, and depression symptoms. Official code is Apache-2.0; weights are shared without a separate licence statement.

• Form factor: clinical

Polysomnography (PSG) is central to sleep medicine, yet its analysis remains hindered by labor-intensive manual scoring and limited cross-center generalizability. We present LPSGM, a large-scale PSG model trained on 220,500 hours of data from 16 public datasets (24,000 recordings) that addresses two challenges: cross-center sleep staging generalization and PSG-based screening of narcolepsy, obstructive sleep apnea (OSA), and depression. LPSGM integrates domain-adaptive pre-training, flexible channel configurations, and a unified architecture to accommodate heterogeneous recording protocols. In prospective validation, LPSGM achieves expert-level sleep staging (κ = 0.845 ± 0.066 vs. inter-expert κ = 0.850 ± 0.102) and matches fully supervised models on independent private cohorts. For downstream screening, LPSGM achieves 84.2% accuracy (κ = 0.761) in three-class narcolepsy classification, with cross-institutional transfer for normal-versus-abnormal hypersomnia yielding an AUC of 0.879. Severe OSA screening reaches an AUC of 0.874 without access to respiratory signals. Depression symptom screening achieves an AUC of 0.801 in a surrogate-labeled APPLES cohort, and cross-dataset transfer to an independent clinically diagnosed cohort reaches an AUC of 0.965. These results suggest that LPSGM's pre-trained representations capture sleep-related signatures that generalize across institutions and support PSG-based screening of clinically relevant conditions. The code and pre-trained model are publicly available at https://github.com/Deng-GuiFeng/LPSGM.

Verbatim from the publisher's DOI record (opens in a new tab) · Publisher licence: http://creativecommons.org/licenses/by-nc-nd/4.0/

No results from this paper on ChatPPG leaderboards yet. Browse the leaderboards.

• Cleveland Family Study (CFS)NSRR family sleep-apnea PSG2,284 subjects · PPG · PSG
• MESA SleepNSRR polysomnography + wrist actigraphy (Multi-Ethnic Study of Atherosclerosis)2,237 subjects · PPG · ECG · IMU
• MrOS Sleep Study (Osteoporotic Fractures in Men, NSRR)ECG · SpO₂ · Resp
• Sleep Heart Health Study (SHHS, NSRR)ECG · SpO₂ · Resp
• STAGESNSRR Stanford Technology Analytics and Genomics in Sleep (clinic PSG + Huami actigraphy + ASQ)1,500 subjects · PPG · ECG · IMU