Can a sleep foundation model learn transferable human physiology from two million hours?

A frozen encoder trained on unlabeled sleep signals

SleepFM-2 is a foundation model: a network trained on a large body of unlabeled recordings so that its internal representation can later be reused for many tasks. The pretraining corpus was 235,865 polysomnography recordings, and the full development and evaluation set covered 282,511 recordings from 26 cohorts and more than two million hours of brain, cardiac, muscle and respiratory signals. Polysomnography, or PSG, is the clinical gold standard for sleep measurement and records these modalities simultaneously through the night.

The key design choice is that most evaluations read out a frozen encoder. No encoder weights are updated; a small task-specific head is trained on top of the embeddings. This matters for interpretation: when the model predicts a disease or a sleep event, the result reflects what the pretrained representation already contains, not what a task-specific fine-tuning pass could add. The paper also reports that SleepFM-2 reduces model size relative to SleepFM and improves temporal resolution from five seconds to one second, a change that lets it address events occurring at finer time scales than sleep stages.

Disease prediction that survives an unseen health system

The central disease result is a comparison, not a single number. A model combining the PSG representation with age, sex and BMI met a prespecified discrimination and significance criterion for 215 subsequently recorded EHR phenotypes in two held-out cohorts. For 155 of those phenotypes, the PSG representation added reproducible information beyond demographics alone. The inclusion of a health system that was not part of pretraining is the demanding part of this test, because medical models can learn site-specific structure that fails to transfer.

The paper also reports that SleepFM-2 outperformed a 480-feature baseline derived from the same recordings. The margin was narrow in one cohort and roughly three times larger at the health system the encoder never saw, so the advantage is clearest where generalization is hardest. These are predictions of subsequent diagnoses, and the authors state explicitly that this does not establish causality or biological disease onset: the signal may also reflect shared risk factors, treatment effects, or prodromal and previously unrecognized disease.

One shared axis explains most disease-score variance

The disease scores were not independent. A single component explained approximately 55 to 56% of the variance in disease-specific predictions. That component had interpretable associations with reduced sigma-band spatial coupling and increased hypnodensity entropy, two properties of sleep EEG organization. The authors interpret this as consistent with a latent dimension of physiological aging or frailty captured from a single night, while cautioning that its meaning should not yet be equated with any specific construct.

This is the paper's most consequential mechanistic claim, and it is an interpretation rather than a demonstrated causal pathway. What the evidence supports is narrower: the predictions share a dominant common direction, and that direction correlates with measurable EEG features. Whether that direction is best described as aging, frailty, or something else remains open, and the authors frame a general physiological-risk measure as a motivation for future work rather than a validated clinical tool.

Event scoring at one-second resolution rivals expert range

Sleep staging is only part of clinical PSG interpretation. Cortical arousals, limb movements and respiratory events occur at finer temporal scales and determine the arousal index, the periodic limb movement index and the apnea–hypopnea index. The earlier SleepFM tokenized signals at five-second resolution, too coarse for these events, and left them out of what the model could score. SleepFM-2's one-second resolution was introduced to close that gap, and the paper reports support for arousal, limb movement and respiratory event detection.

On sleep events, the frozen encoder performed within the observed range of expert scorers. That phrasing matters: it is a comparison against the spread among human scorers, not a claim of superiority over them. The paper also reports improvements over SleepFM in both disease prediction and sleep scoring, so the gain is not confined to the new event types.

Transfer to wearables without updating encoder weights

The transfer tests ask whether the representation survives a change of sensor. With the encoder frozen, SleepFM-2 transferred to wakeful EEG, headband and in-ear EEG, wrist PPG and wrist accelerometry. It improved sleep staging across six accelerometry cohorts and achieved disease-prediction performance in UK Biobank similar to models pretrained directly on accelerometry. The wrist accelerometry result is notable because that input shares no channel with PSG, so the transfer cannot rest on matching sensor physics.

The comparison in UK Biobank is a parity result, not an advantage: similar performance to models pretrained directly on accelerometry. Read together with the six-cohort staging improvement, the pattern suggests the PSG representation carries information that survives aggressive changes in what is measured, while the size of the benefit depends on the target task and cohort.

Subjective sleep is recoverable for the recorded night only

The subjective-sleep analyses set a boundary that is easy to miss. In the BioSerenity cohort, recordings were paired with questionnaires about both the study night and habitual sleep. SleepFM-2 was the best-performing method on all six post-sleep questionnaire items about the recorded night, with AUROC improvements of +0.011 to +0.034 over the combined baseline (mean AUROC 0.750 versus 0.725); the improvement was nominally significant on every endpoint and survived Benjamini–Hochberg correction on four of six.

For the six matched sleep-history items describing habitual sleep, the picture weakens. SleepFM-2 was best on five items but significantly better on only two, and on the composite habitual poor-sleep endpoint the two approaches were essentially indistinguishable (0.680 versus 0.677). The authors explain this as a scope limit of one-night physiology: habitual complaints reflect multi-night patterns, psychological state, comorbidity and environment. The paper states that SleepFM-2 should not be treated as a substitute for longitudinal recording or questionnaire assessment, a limitation that is especially relevant to insomnia, where chronicity is intrinsic to the diagnosis.

Subjective–objective discrepancy is detectable but not diagnostic

Two endpoints measured how far a patient's report diverged from the PSG measurement: absolute difference in total sleep time, thresholded at 90 minutes with 35.5% prevalence, and absolute difference in sleep onset latency, thresholded at 45 minutes with 29.3% prevalence. SleepFM-2 reached AUROC 0.646 for the total-sleep-time discrepancy and 0.748 for the sleep-onset-latency discrepancy, outperforming the combined tabular baseline by +0.024 and +0.020 respectively.

The authors are direct that absolute performance is modest, particularly for total sleep time, and that the signal does not determine discrepancy accurately enough for standalone clinical diagnosis. The useful reading is that raw physiology contains information about the gap between reported and measured sleep that conventional PSG summaries and individual microstructural features miss, while the magnitude of that information does not support replacing patient-reported assessment.

The advantage is largest where generalization is hardest

Across these analyses, the paper's own priority ordering is consistent: the representation is most valuable for endpoints tied to the recorded night and to a health system the encoder never saw, and least valuable for habitual complaints that a single night cannot contain. The disease-prediction margin over the 480-feature baseline was narrow in one cohort and roughly three times larger at the unseen health system. The subjective-sleep margin was clear for the recorded night and absent for the habitual composite.

That split is the practical takeaway. A reader deciding whether this representation is worth adopting should ask which of those two situations they are in: predicting outcomes in a new site or scoring events and last-night experience from raw signals, where the evidence supports a real advantage, or inferring long-term sleep complaints from one recording, where the paper's own comparison shows the representation adds little beyond clinical and conventional PSG features. The authors position automated PSG interpretation as the nearer-term direction and a general physiological-risk measure as a motivation for future work, not a validated clinical output.

Sources used in this answer