Learning transferable human physiology from two million hours of sleep with SleepFM-2
Rahul Thapa, Christopher Sun, William Theodor Lehn-Schioler, Sophia Claire Kivelson, Umaer Hanif, Hyatt Moore, Harrison G. Zhang, Hafsa Ahmed, et al.
cs.AI
Quick verdict
A 2M-hour sleep foundation model beating interpretable baselines and transferring to wearables; impressive scale, but retrospective and fairness-untested.
Summary
SleepFM-2 is a sleep foundation model pretrained on 235,865 polysomnography recordings (>2 million hours) from 24 cohorts and evaluated on 282,511 recordings from 26 cohorts. Using one-second tokens, per-channel representations, and a combined leave-one-out contrastive plus masked-autoencoding objective, the frozen encoder supports sleep staging, arousal/limb-movement/respiratory event detection, phenome-wide disease prediction, and subjective-sleep endpoint prediction via light task heads. On 939 phecodes in two health systems, 215 phenotypes meet a prespecified dual criterion (C-index ≥0.75 and Bonferroni-corrected P<0.01 in both SSC and the unseen HSP cohort); 155 additionally pass a demographics-conditioned likelihood-ratio test. SleepFM-2 outperforms SleepFM-1 (e.g., mean C-index 0.729 vs 0.696 in HSP), a demographics-only model, and a 480-feature interpretable baseline, with margins widening at the unseen site. It performs within the range of expert scorers for event detection and transfers to wakeful EEG, wearable EEG/PPG, and wrist accelerometry, matching accelerometry foundation models in UK Biobank (mean C 0.687 vs 0.688). Limitations: retrospective referral cohorts, death treated as censoring without competing-risk analysis, no subgroup fairness assessment, an engineered surrogate bridge for accelerometry transfer, and no code/data availability statement visible in the provided text.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 4
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A frozen encoder pretrained on 235,865 PSG recordings predicts 215 incident phenotypes at C-index ≥0.75 with Bonferroni-corrected P<0.01 in two held-out cohorts, and for 155 phenotypes adds signal beyond age, sex and BMI.
- SleepFM-2 outperforms SleepFM-1 and a 480-feature interpretable baseline derived from the same recordings, with the largest margins at the health system unseen during pretraining (mean C 0.729 vs 0.696 in HSP).
- Disease predictions share a reproducible principal component (≈57–58% of variance, loading correlation ρ=0.63 across cohorts) associated with reduced sigma-band spatial coupling and increased hypnodensity entropy.
- At one-second token resolution, the frozen encoder performs within the range of expert scorers for cortical arousal, limb-movement and respiratory event detection (F1 0.601/0.604/0.462 vs scorer means 0.546/0.530/0.404).
- Representations transfer without weight updates to wakeful EEG, wearable EEG/PPG and wrist accelerometry, performing comparably to accelerometry-pretrained foundation models on UK Biobank disease prediction (mean C 0.687 vs 0.688).
Strengths
Dual-cohort external validation with prespecified, dual-gated criteria
215 of 939 phecodes reach C-index ≥0.75 with Bonferroni-corrected P<0.01 in both SSC and HSP — the latter a health system absent from pretraining; 155 survive a stricter Cox likelihood-ratio test conditioned on age, sex and BMI, cleanly isolating incremental PSG signal.
A genuinely strong interpretable baseline
The 480-feature bank (129 spectral, 138 coupling, 84 microstructure, 56 cardiac, 40 hypnodensity, 22 respiratory) is read out with the same nonlinear head as the embedding; SleepFM-2 beats it with the margin roughly 3× larger at the unseen HSP site (Fig. 2b–d).
Transparent component-wise ablation
Expert-parity event detection with matched ground truth
Calibrated claim discipline on subjective sleep
Weaknesses
Retrospective referral populations, no utility evaluation
No subgroup fairness assessment
Accelerometry transfer rests on an engineered surrogate bridge
Unquantified minibatch Cox approximation
General-risk axis is largely a mean-score restatement
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
SleepFM-2 (arXiv:2609.06849) is a sleep foundation model pretrained on 235,865 polysomnography recordings from 24 cohorts (>2 million hours), evaluated on 282,511 recordings from 26 cohorts. Five headline claims: (1) improved disease prediction and sleep scoring over SleepFM-1; (2) phenome-wide incident prediction — 215 of 939 phecodes reach C-index ≥0.75 with Bonferroni-corrected P<0.01 in both the Stanford Sleep Cohort (SSC) and the Human Sleep Project (HSP), the latter unseen in pretraining, with 155 surviving a demographics-conditioned likelihood-ratio test; (3) superiority over a 480-feature interpretable bank from the same recordings; (4) expert-range detection of arousals, limb movements and respiratory events at one-second resolution; (5) frozen-encoder transfer to wakeful EEG, wearable EEG/PPG and wrist accelerometry, including UK Biobank (UKB) disease prediction comparable to accelerometry-pretrained models.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is SleepFM-2 and what data was it trained on?
SleepFM-2 is a self-supervised foundation model for sleep physiology. It was pretrained on 235,865 polysomnography recordings (over two million hours) from 24 clinical cohorts, with evaluation extending to 282,511 recordings from 26 cohorts, plus wearable datasets. Inputs are EEG, EOG, ECG, EMG and respiratory channels tokenized at one-second resolution; training combines a leave-one-out contrastive objective across modality summaries with masked autoencoding. After pretraining the encoder is frozen and all tasks use small task heads.
Can SleepFM-2 be used with consumer wearables like smartwatches?
In this paper, yes at research stage: the frozen encoder stages sleep from headband and in-ear EEG and wrist PPG, and predicts disease from wrist accelerometry in UK Biobank, where it matches foundation models pretrained directly on accelerometry (mean C-index 0.687 vs 0.688). Accelerometry is fed through an engineered bridge of respiratory/cardiac surrogate channels, and the evaluations use research-grade recordings rather than off-the-shelf consumer devices, so real-world consumer performance is untested.
Is SleepFM-2 clinically validated or approved for medical use?
No. All results are retrospective analyses of existing recordings and electronic health records in referral populations. The authors explicitly state that discrimination improvements do not establish clinical utility, that prospective evaluation is future work, and that the accompanying SleepFM-Interface web demo is research software not validated for clinical use.
How much better is SleepFM-2 than the original SleepFM?
Under the original SleepFM evaluation protocol with identical heads and splits, mean disease C-index rises from 0.709 to 0.718 in the Stanford cohort and from 0.696 to 0.729 in the unseen HSP cohort, and five-class staging macro-F1 improves on every cohort (e.g., 0.679 to 0.716 on SSC, 0.783 to 0.811 on MESA). A component-wise ablation attributes gains to the AdamW optimizer, the modern LLaMA-style backbone, and combining contrastive with masked-autoencoding objectives, while the encoder actually shrinks from 4.83M to 2.57M parameters.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- The provided text is a sampled excerpt with explicit '[...sampling gap...]' sections; portions of the subjective-sleep results, supplementary tables and back matter were not shown, and judgments touching those regions are based on the visible portions.
- No data or code availability statement appears in the provided excerpt; codeAvailable=false reflects the excerpt and may not hold for the full paper (the statement may sit in an unsampled section).
- Figure-level numbers were read from embedded caption text; rendered figures and supplementary tables (e.g., exact per-cohort splits in Supplementary Table 1) could not be independently verified.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.