On Sleep-Wake Disorder Identification from Polysomnographic Representations

Alvise Dei Rossi; Luigi Fiorillo; Giuliana Monachino; Alessia Calzoni; Julia van der Meer; Markus H. Schmidt; Claudio L. A. Bassetti; Silvia Santini

cs.LGOct 12, 2026

On Sleep-Wake Disorder Identification from Polysomnographic Representations

Authors: Alvise Dei Rossi, Luigi Fiorillo, Giuliana Monachino, Alessia Calzoni, Julia van der Meer, Markus H. Schmidt, Claudio L. A. Bassetti, Silvia Santini,

Organizations: Faculty of Informatics, Università della Svizzera Italiana · MeDiTech, SUPSI · Ente Ospedaliero Cantonale (EOC) · Department of Information Engineering, University of Brescia ·

Abstract

Self-supervised polysomnography foundation models (PSG FMs) aim to capture clinically useful information beyond sleep staging, yet are rarely compared with sleep-staging encoders pretrained at comparable scale and modality coverage. Using a common whole-night, missingness-aware state-space downstream modeling framework, we compared frozen representations from prominent large-scale, multi-cohort sleep-stage-supervised encoders and self-supervised PSG FMs for multilabel identification of six sleep-wake and circadian disorder groups in 7,906 PSGs from an unseen registry. Although PSG FMs outperformed an EEG-only, single-cohort supervised baseline, this ordering reversed when supervised pretraining approached comparable scale and modality coverage. Transfer learning from neural sleep scorers achieved higher overall performance while exhibiting greater downstream data efficiency. These findings underscore the need for scale- and modality-matched pretrained supervised baselines when assessing PSG FMs' downstream potential.

Figures & tables

Figure 1: Recordings are processed by a pretrained encoder, yielding representations that are jointly modeled across time and streams over the complete night using a mask-aware state-space network.

A. Without covariates
Method	Insomnia	SDB	CDH	CRSWD	Parasomnia	SRMD	Macro
Random Forest (Expert)	.695	.706	.682	.592	.524	.544	.624
Random Forest (SOMNUS)	.676	.699	.708	.598	.618	.553	.642
U-Sleep	.726	.854	.774	.625	.766	.642	.731
SleepTransformer	.738	.874	.787	.616	.803	.644	.744
SleepFM	.709	.875	.759	.636	.726	.625	.722

Table 1: Pooled out-of-fold AUROC without (A) and with (B) demographic and anthropometric covariates. Best and second-best results per panel are bolded and underlined, respectively.

Figure 2: Downstream performance without covariates. Left: labeled-data efficiency; points and error bars denote mean out-of-fold AUROC and one standard deviation across participant-level draws. Right: macro-AUROC using linear probing and missingness-aware state-space modeling.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Characteristic	Value
Patients, $n$	7,195
Patients with repeated PSGs, $n$ (%)	574 (8.0)
PSG recordings, $n$	7,906
Age, mean (SD), years	48.8 (18.0)
Female, $n$ (%)	2,652 (33.5)
Insomnia, $n$ (%)	1,131 (14.3)

Appendix

Table 2: Characteristics of the BSWR analysis cohort. Demographic and diagnostic distributions are reported at the recording level, corresponding to the unit of prediction. Diagnostic categories are non-mutually exclusive.

Modality	Channels and derivations
EEG	Central, frontal, and occipital channels (C3, C4, F3, F4, O1, O2, and Cz), including recordings referenced to A1/A2 or M1/M2.
EOG	Left and right electro-oculography (E1 and E2), available either as pre-referenced channels or derived against the corresponding A1/A2 or M1/M2 reference.
EMG	Central, upper, left, and right chin EMG, left and right tibialis or leg EMG.
ECG	ECG/EKG channels
RESP	Thoracic and abdominal (THX, ABD) respiratory effort; nasal pressure, cannula, or airflow; oral or thermistor airflow; snoring; oxygen saturation and pulse.

Appendix

Table 3: Canonical PSG signal families supplied to at least one evaluated encoder.

Model	Pretraining objective	Pretraining scale	# Params.	Native BSWR inputs	Embedding dimension	Resolution
U-Sleep	Sleep-stage classification	$\sim$ 220k h 17 cohorts	15.6M	EEG: available C3/C4/Cz, EEG: F3/F4, O1/O2 derivations EOG: available E1/E2 EOG: derivations EMG: Chin/ChinL/ChinR ECG: EKG RESP: ABD/THX	$\mathbb{R}^{6}$ per modality	$1\,\mathrm{Hz}$
SleepTransformer	Sleep-stage classification	$\sim$ 220k h 17 cohorts	19.2M	EEG: available C3/C4/Cz, EEG: F3/F4, O1/O2 derivations EOG: available E1/E2 EOG: derivations EMG: Chin/ChinL/ChinR ECG: EKG RESP: ABD/THX	$\mathbb{R}^{128}$ per modality	$\frac{1}{30}\,\mathrm{Hz}$
SleepFM	Leave-one-modality-out contrastive learning	432k h 4 cohorts	4.8M	BAS: available EEG/EOG BAS: derivations EKG: EKG EMG: Chin/ChinL/ChinR RESP: ABD/THX, airflow, RESP: snoring, SpO 2 /pulse	$\mathbb{R}^{128}$ per modality	$0.2\,\mathrm{Hz}$
OSF	DINO self-distillation with channel and temporal masking	$\sim$ 95k h 5 cohorts	85.3M	EEG: C3-(A2/M2), EEG: C4-(A1/M1) EOG: E1-(A2/M2), EOG: E2-(A1/M1) EMG: chin and bilateral leg ECG: EKG RESP: ABD/THX, nasal pressure, RESP: snoring	$\mathbb{R}^{768}$ fused	$\frac{1}{30}\,\mathrm{Hz}$
Hypnos	Autoregressive next-token prediction over quantized signals	$\sim$ 135k h 8 cohorts	450.7M	EEG: C3-(M2/A2), EEG: C4-(M1/A1) EOG: E1-(M2/A2), EOG: E2-(M1/A1) EMG: bipolar chin ECG: EKG RESP: ABD/THX	$\mathbb{R}^{768}$ per stream	$1\,\mathrm{Hz}$

Appendix

Table 4: Specifications of the evaluated frozen pretrained encoders.

Pretraining and inputs	Insomnia	SDB	CDH	CRSWD	Parasomnia	SRMD	Macro
Single-cohort; EEG	.700	.786	.753	.617	.764	.603	.704
Multi-cohort; EEG	.709	.797	.759	.631	.795	.603	.716
Multi-cohort; EEG/EOG/EMG/ECG	.712	.826	.778	.604	.792	.630	.724
Multi-cohort; EEG/EOG/EMG/ECG/RESP	.737	.874	.787	.609	.800	.649	.743

Appendix

Table 5: Performance of SleepTransformer downstream modeling across pretraining and modality-coverage conditions. Values are pooled out-of-fold AUROC without covariates. Best and second-best results are bolded and underlined.

A. Without covariates
Method	Insomnia	SDB	CDH	CRSWD	Parasomnia	SRMD	Macro
U-Sleep	.712	.814	.772	.629	.751	.635	.719
SleepTransformer	.718	.827	.783	.608	.802	.639	.730
SleepFM	.686	.782	.751	.608	.689	.594	.685
OSF	.697	.810	.748	.637	.709	.603	.701
Hypnos	.702	.816	.758	.620	.715	.597	.701

Appendix

Table 6: Pooled out-of-fold AUROC under common-core input coverage, without (A) and with (B) demographic and anthropometric covariates. Best and second-best available results per panel are bolded and underlined.

SleepFM inputs	SDB AUROC
All modality groups	.879
BAS/ECG/EMG only	.769
RESP group only	.869
BAS/ECG/EMG + respiratory effort	.865
BAS/ECG/EMG + airflow	.864
BAS/ECG/EMG + oximetry/pulse	.847

Appendix

Table 7: SleepFM SDB performance across full, non-respiratory-core, RESP-only, and core-plus-respiratory-subset conditions.

Method	Train (%)	Insomnia	SDB	CDH	CRSWD	Parasomnia	SRMD	Macro
U-Sleep	1	$\mathbf{.614}{\scriptstyle\pm.008}$	$.635{\scriptstyle\pm.011}$	$\underline{.598}{\scriptstyle\pm.042}$	$\mathbf{.557}{\scriptstyle\pm.033}$	$.533{\scriptstyle\pm.016}$	$\mathbf{.527}{\scriptstyle\pm.016}$	$\underline{.577}{\scriptstyle\pm.016}$
	10	$\underline{.672}{\scriptstyle\pm.003}$	$.749{\scriptstyle\pm.006}$	$\underline{.691}{\scriptstyle\pm.006}$	$\underline{.570}{\scriptstyle\pm.036}$	$\underline{.644}{\scriptstyle\pm.011}$	$\mathbf{.562}{\scriptstyle\pm.004}$	$\underline{.648}{\scriptstyle\pm.006}$
	25	$\underline{.696}{\scriptstyle\pm.016}$	$.799{\scriptstyle\pm.003}$	$\underline{.726}{\scriptstyle\pm.008}$	$\underline{.605}{\scriptstyle\pm.027}$	$\underline{.708}{\scriptstyle\pm.010}$	$\mathbf{.590}{\scriptstyle\pm.007}$	$\underline{.687}{\scriptstyle\pm.006}$
	50	$.700{\scriptstyle\pm.004}$	$.830{\scriptstyle\pm.003}$	$.748{\scriptstyle\pm.011}$	$.608{\scriptstyle\pm.017}$	$\underline{.736}{\scriptstyle\pm.005}$	$\underline{.604}{\scriptstyle\pm.005}$	$.704{\scriptstyle\pm.000}$
	100	.726	.854	.774	.625	.766	.642	.731
SleepTransformer	1	$\underline{.606}{\scriptstyle\pm.010}$	$\mathbf{.727}{\scriptstyle\pm.020}$	$\mathbf{.631}{\scriptstyle\pm.034}$	$\underline{.549}{\scriptstyle\pm.039}$	$\mathbf{.567}{\scriptstyle\pm.016}$	$.507{\scriptstyle\pm.015}$	$\mathbf{.598}{\scriptstyle\pm.012}$

Appendix

Table 8: Downstream data efficiency under each encoder’s full physiological input coverage, without demographic and anthropometric covariates. Reduced-data values are mean

\pm

standard deviation across participant draws. Best and second-best mean AUROCs within each training fraction and outcome are bolded and underlined, respectively.

Method	Train (%)	Insomnia	SDB	CDH	CRSWD	Parasomnia	SRMD	Macro
U-Sleep	1	$\mathbf{.605}{\scriptstyle\pm.014}$	$.614{\scriptstyle\pm.005}$	$\underline{.598}{\scriptstyle\pm.046}$	$\underline{.542}{\scriptstyle\pm.022}$	$\underline{.546}{\scriptstyle\pm.006}$	$\underline{.507}{\scriptstyle\pm.014}$	$\underline{.569}{\scriptstyle\pm.012}$
	10	$\underline{.662}{\scriptstyle\pm.004}$	$.715{\scriptstyle\pm.006}$	$\underline{.697}{\scriptstyle\pm.007}$	$\underline{.572}{\scriptstyle\pm.044}$	$\underline{.641}{\scriptstyle\pm.016}$	$\underline{.546}{\scriptstyle\pm.011}$	$\underline{.639}{\scriptstyle\pm.008}$
	25	$\mathbf{.692}{\scriptstyle\pm.008}$	$\underline{.764}{\scriptstyle\pm.002}$	$\underline{.727}{\scriptstyle\pm.004}$	$\underline{.590}{\scriptstyle\pm.022}$	$\underline{.707}{\scriptstyle\pm.012}$	$\mathbf{.581}{\scriptstyle\pm.007}$	$\underline{.677}{\scriptstyle\pm.005}$
	50	$\underline{.700}{\scriptstyle\pm.006}$	$\underline{.791}{\scriptstyle\pm.000}$	$\underline{.748}{\scriptstyle\pm.005}$	$.592{\scriptstyle\pm.011}$	$\underline{.733}{\scriptstyle\pm.001}$	$\underline{.599}{\scriptstyle\pm.009}$	$\underline{.694}{\scriptstyle\pm.003}$
	100	.712	.814	.772	.629	.751	.635	.719
SleepTransformer	1	$\underline{.591}{\scriptstyle\pm.013}$	$\mathbf{.685}{\scriptstyle\pm.017}$	$\mathbf{.624}{\scriptstyle\pm.050}$	$\mathbf{.550}{\scriptstyle\pm.016}$	$\mathbf{.596}{\scriptstyle\pm.011}$	$\mathbf{.511}{\scriptstyle\pm.012}$	$\mathbf{.593}{\scriptstyle\pm.012}$

Appendix

Table 9: Downstream data efficiency under non-respiratory core, without demographic and anthropometric covariates. Reduced-data values are mean

\pm

standard deviation across participant-level draws. Best and second-best reported mean AUROCs within each training fraction and outcome are bolded and underlined, respectively.

Method	Insomnia	SDB	CDH	CRSWD	Parasomnia	SRMD	Macro
U-Sleep	.672	.695	.669	.614	.517	.514	.613
SleepTransformer	.725	.845	.763	.629	.785	.640	.731
SleepFM	.685	.809	.722	.537	.687	.615	.676
OSF	.706	.848	.750	.624	.717	.648	.716
Hypnos	.698	.827	.738	.575	.721	.646	.701

Appendix

Table 10: Pooled out-of-fold AUROC obtained using the global-mean linear probe without demographic and anthropometric covariates. Best and second-best results are bolded and underlined.

Self-Supervised Pre-Training

Self-supervised pre-training learns transferable representations from unlabeled data to reduce the labels needed for downstream tasks.

See all 251 papers

OSF: On Pre-training and Scaling of Sleep Foundation Models
Feb 27, 2026Zitao Shuai, Zongzhe Xu, David Yang +2Self-Supervised Pre-Training Scaling Laws
University of California, Los Angeles · Emory University
SleepMaMi: A Universal Sleep Foundation Model for Integrating Macro- and Micro-structures
Feb 7, 2026Keondo Park, Younghoon Na, Yourim Choi +3Self-Supervised Time Series Representation Learning
Graduate School of Data Science, Seoul National University, Seoul, South Korea · Department of Biomedical Sciences, Seoul National University College of Medicine, Seoul, Republic of Korea · Obstructive Upper Airway Research (OUaR) Laboratory, Department of Pharmacology, Seoul National University College of Medicine, Seoul, Republic of Korea +1
Comparative Analysis of State-of-the-Art Foundation Models for Sleep Analysis Under Channel Reduction
Aug 14, 2026Hassan Mehdi, Riku Klen, Ayse Kosal Bulbul +5Sleep Stage Classification EEG Foundation Models
Institute of Biomedicine, University of Turku, Finland · Turku PET Centre, University of Turku and Turku University Hospital, Finland · University College Cork, Ireland +3

Feb 27, 2026J/K move · Enter open · S save

OSF: On Pre-training and Scaling of Sleep Foundation Models

Zitao Shuai, Zongzhe Xu, David Yang, Wei Wang+1

University of California, Los Angeles · Emory University

Self-Supervised Pre-Training Scaling Laws Time-Series Foundation Models

Open paper AskPDF arXiv

Polysomnography (PSG) provides the gold standard for sleep assessment but suffers from substantial heterogeneity across recording devices and cohorts. There have been growing efforts to build general-purpose foundation models (FMs) for sleep physiology, but lack an in-depth understanding of the pre-training process and scaling patterns that lead to more generalizable sleep FMs. To fill this gap, we curate a massive corpus of 166,500 hours of sleep recordings from nine public sources and establish SleepBench, a comprehensive, fully open-source benchmark. Leveraging SleepBench, we systematically evaluate four families of self-supervised pre-training objectives and uncover three critical findings: (1) existing FMs fail to generalize to missing channels at inference; (2) channel-invariant feature learning is essential for pre-training; and (3) scaling sample size, model capacity, and multi-source data mixture consistently improves downstream performance.With an enhanced pre-training and scaling recipe, we introduce OSF, a family of sleep FMs that achieves state-of-the-art performance across nine datasets on diverse sleep and disease prediction tasks. Further analysis of OSF also reveals intriguing properties in sample efficiency, hierarchical aggregation, and cross-dataset scaling.

Abstract

Self-Supervised Pre-Training

Explore similar work

OSF: On Pre-training and Scaling of Sleep Foundation Models

SleepMaMi: A Universal Sleep Foundation Model for Integrating Macro- and Micro-structures

Comparative Analysis of State-of-the-Art Foundation Models for Sleep Analysis Under Channel Reduction