SleepJEPA

A night of at-home polysomnography records six to twelve hours of simultaneous brain, eye, chin, cardiac, oximetry and respiratory-effort signal, and clinical scoring compresses all of it into a short list of summaries: total sleep time, sleep efficiency, stage percentages, an apnea-hypopnea index. Those summaries were designed to diagnose sleep disorders, not to forecast a myocardial infarction fifteen years later, and whatever detail might carry that longer signal is discarded before anyone looks at the record. SleepJEPA keeps the whole night instead, encoding it into a latent representation that downstream models read.

The encoder is trained by joint embedding prediction. A context encoder sees an unmasked subset of the signal's patches, a small predictor network estimates the embeddings of the patches that were held out, and an exponentially-weighted moving-average copy of the encoder supplies the targets those estimates are regressed against. The loss lives entirely in embedding space and never touches the waveform, so the model is not obliged to spend capacity reproducing sensor noise — the cost that masked-reconstruction approaches pay on physiological recordings. Related sleep foundation models take the other two routes: SleepFM is contrastive across modalities, and PFTSleep, the group's own earlier in-lab model, is a masked autoencoder.

SleepJEPA was developed at the Icahn School of Medicine at Mount Sinai and posted as a preprint in March 2026. After pretraining, the encoder is frozen; every downstream capability — staging, age, daytime sleepiness, narcolepsy, and ten long-horizon disease risks — comes from separately trained lightweight heads reading those fixed features.

#Key Features

• Prediction in embedding space: Targets are the EMA encoder's own embeddings rather than raw signal, suppressing reconstruction of channel noise and reducing representational collapse during self-supervision.
• Three-second patches over a whole night: A per-channel convolution tokenizes each channel into 3-second patches, the smallest input unit reported for sleep staging, and PyTorch nested tensors let studies of differing length pass through without padding.
• Channel-independent embedding: Each of the seven channels is embedded on its own and cross-channel structure is left to attention, so the encoder accepts variable channel input rather than one rigid montage.
• Discrete hazard risk heads: Long-horizon risk is fit with a discrete hazard loss rather than a Cox proportional-hazards head, avoiding both the proportional-hazards assumption over 15-year horizons and the batch-size dependence of partial-likelihood training.

#Technical Details

The context encoder is an adapted PatchTST time-series transformer — 8 layers, 8 heads, feature dimension 768, with rotary positional embeddings — and the predictor is a 4-layer, 8-head PatchTST of dimension 384. Together they hold 66,464,256 parameters, of which the 58,772,736 in the context encoder generate downstream representations. Inputs are EEG (C4-M1 or C3-M2), left EOG, chin EMG, augmented lead II ECG, SpO2, and thoracic and abdominal respiratory inductance plethysmography, resampled to 128 Hz. Pretraining ran across three nodes of four NVIDIA H100 GPUs with 20–80% random masking, and epoch 16 was selected by linear-probe sleep-stage AUC. In total the model was trained, validated and tested on 422,035 hours from 55,518 sleep studies: 34,291 pretraining studies from 27,067 patients and 21,227 finetuning studies from 18,839 patients, drawn from the Human Sleep Project, a Mount Sinai polysomnography database, SHHS, the Wisconsin Sleep Cohort, MrOS, MESA, APPLES and the Mignot narcolepsy cohort. On a 2,531-study held-out set a bidirectional GRU head over frozen features reaches macro-AUC 0.97, Cohen's kappa 0.80 and macro F1 0.77 for five-stage scoring, and 0.92 macro-AUC on 4,050 independently held studies. Attentive classifiers estimate 15-year risk at iAUC 0.85 for congestive heart failure and coronary heart disease death, 0.83 for cardiovascular death and 0.78 for stroke, with 5-year obstructive sleep apnea at 0.86 and 10-year diabetes at 0.82; age is recovered with a mean absolute error of 5.51 years.

#Applications

The model suits settings where full-night home recordings already exist and more than a scored hypnogram is wanted from them: retrospective epidemiological cohorts, sleep clinics standardizing technician labels across sites, and prognostic screening that flags patients whose overnight physiology implies elevated cardiometabolic risk. Because the heads read frozen features, adding an outcome means training a small classifier rather than re-pretraining, and the released inference script consumes ordinary EDF files.

#Impact

SleepJEPA is the first joint embedding predictive architecture applied to at-home polysomnography, and it shows that a single frozen sleep representation supports staging competitive with task-specific models such as U-Sleep while also carrying 15-year prognostic signal that conventional scored metrics do not. Its limits are clear: the cohorts are predominantly older adults referred for sleep complaints, incident outcome labels come from follow-up visits and survey responses, and integrated-gradients attributions spread diffusely across channels rather than isolating interpretable features. Objective daytime sleepiness is modest (AUC 0.64, 95% CI 0.57–0.71), narcolepsy rests on a wide interval (AUC 0.88, 0.68–0.97), and independent-cohort stroke risk falls to iAUC 0.60 at ten years. The work remains a preprint that has not been peer reviewed.