LPSGM (Large PSG Model; flexible-channel EEG/EOG/EMG sleep staging + clinical screening; ZJU / Hangzhou Seventh People's Hospital)
- Year
- 2026
- Form factor
- Clinical
Activity counts measure attention, not model quality. Missing counts are unknown. Publication status follows the linked source; a listed venue alone does not confirm peer review.
Prospective sleep staging κ 0.845 ± 0.066 against expert consensus (inter-expert κ 0.850 ± 0.102). Three-class narcolepsy 84.2% accuracy (κ 0.761); cross-institution hypersomnia normal-vs-abnormal AUC 0.879; severe OSA AUC 0.874 without respiratory channels; depression symptoms AUC 0.801 on surrogate-labelled APPLES and 0.965 on an external clinically diagnosed cohort.
- Signals
- EEG, EMG, EOG
- Form factor
- Clinical
- Year
- 2026
- Architecture
Epoch Encoder (dual-branch CNN, small + large kernels, per 30-s epoch and channel) + Sequence Encoder (N Transformer blocks over 20-epoch sequences) + classifier; padding/masking over 1-9 standard channels (F3, F4, C3, C4, O1, O2, E1, E2, Chin); domain-adaptive pre-training across 16 public PSG corpora
- Pretraining data
220,500 h / 24,000 recordings from 16 public PSG datasets (README lists APPLES, DCSM, DOD, HMC, ISRUC, SVUH/UCDDB, PhysioNet 2018, STAGES, ABC, NCHSDB, HomePAP, CHAT, CCSHS, CFS, MrOS, SHHS, MASS SS1+SS3, MESA)
- License
Code Apache-2.0 (LICENSE.txt in repo). Weights: no separate terms on the Google Drive folder; trained on NSRR/PhysioNet/other corpora whose DUAs may restrict derived use
- Type
- Foundation model
- Form factor
- Clinical
- Parameters
- Not reported
- Architecture
- Epoch Encoder (dual-branch CNN, small + large kernels, per 30-s epoch and channel) + Sequence Encoder (N Transformer blocks over 20-epoch sequences) + classifier; padding/masking over 1-9 standard channels (F3, F4, C3, C4, O1, O2, E1, E2, Chin); domain-adaptive pre-training across 16 public PSG corpora
- Venue
- medRxiv preprint (health informatics), v5 2026-10-27 (v1 2024-12-11); not peer reviewed
- Year
- 2026
- arXiv
- Not reported
- DOI
- Not reported
- Weights
- Open
- Licence
- Code Apache-2.0 (LICENSE.txt in repo). Weights: no separate terms on the Google Drive folder; trained on NSRR/PhysioNet/other corpora whose DUAs may restrict derived use
- Caveats
- Not reported
- Datasets
- Added
- Oct 27, 2026
- Updated
- Oct 27, 2026
Original description
Figure 1: Overview of the LPSGM framework. Panel (a): Data harmonization schematizes the aggregation of PSG data from 16 public datasets (220,500 hours), the public MNC dataset and 2 independent clinical cohorts, spanning diverse geographic populations and recording protocols. Panel (b): Cross-center generalization outlines the training-evaluation pipeline: LPSGM is pre-trained on multi-center public datasets, validated for cross-domain sleep staging on two unseen private datasets, and fine-tuned for downstream screening tasks covering narcolepsy, obstructive sleep apnea (OSA), and depression. Panel (c): Analytical validation details the study's three-pronged evaluation: (1) a prospective clinical trial benchmarking LPSGM against expert consensus, (2) interpretability analysis to decode decision-making patterns, and (3) ablation studies quantifying the contribution of key components.
Verbatim fromGitHub· Description from the repository README; licence: Apache-2.0
Key results
How it works
Datasets linked in the hub; pretraining vs evaluation roles follow the paper.
- Pretraining data
- 220,500 h / 24,000 recordings from 16 public PSG datasets (README lists APPLES, DCSM, DOD, HMC, ISRUC, SVUH/UCDDB, PhysioNet 2018, STAGES, ABC, NCHSDB, HomePAP, CHAT, CCSHS, CFS, MrOS, SHHS, MASS SS1+SS3, MESA)
- Architecture
- Epoch Encoder (dual-branch CNN, small + large kernels, per 30-s epoch and channel) + Sequence Encoder (N Transformer blocks over 20-epoch sequences) + classifier; padding/masking over 1-9 standard channels (F3, F4, C3, C4, O1, O2, E1, E2, Chin); domain-adaptive pre-training across 16 public PSG corpora
- Inputs
- EEG, EMG, EOG
- Approach
- foundation model · sleep · PSG · sleep staging · flexible channel · narcolepsy
Datasets· 5
- Sleep Heart Health Study (SHHS, NSRR)ECG · SpO₂ · Resp
- 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
- Cleveland Family Study (CFS)NSRR family sleep-apnea PSG2,284 subjects · PPG · PSG
- STAGESNSRR Stanford Technology Analytics and Genomics in Sleep (clinic PSG + Huami actigraphy + ASQ)1,500 subjects · PPG · ECG · IMU
Cite
APA
LPSGM (Large PSG Model; flexible-channel EEG/EOG/EMG sleep staging + clinical screening; ZJU / Hangzhou Seventh People's Hospital). (2026). medRxiv preprint , v5 2026-10-27.
BibTeX
@misc{lpsgm2026lpsgm,
title = {{LPSGM (Large PSG Model; flexible-channel EEG/EOG/EMG sleep staging + clinical screening; ZJU / Hangzhou Seventh People's Hospital)}},
year = {2026},
howpublished = {medRxiv preprint , v5 2026-10-27}
}Built from the hub record. Check it against the publisher before submitting.
History
Added to the ChatPPG research hub
Connections
Open in the graph explorerPapers
A Unified Flexible Large Polysomnography Model for Sleep Staging and Downstream Clinical Screening (LPSGM)Also uses Sleep Heart Health Study (SHHS, NSRR)medRxiv preprint · 2024
SleepMaMi: A Universal Sleep Foundation Model for Integrating Macro- and Micro-structuresAlso uses Sleep Heart Health Study (SHHS, NSRR)ICML 2026 · 2026
