SleepJEPA: Learning the latent world of sleep with at-home sleep data to estimate disease risk

SleepJEPA is a foundational deep learning model trained on over 422,000 hours of at-home sleep data that effectively learns latent sleep representations to accurately predict long-term risks for ten major cardiovascular, metabolic, and neurological diseases, as well as estimate sleep stages and specific sleep disorders.

The Gist

Imagine your sleep is a complex, symphony-like performance. Every night, your brain, heart, lungs, and muscles play different instruments together. For decades, doctors have only been able to listen to this symphony in a sterile, expensive recording studio (the sleep lab). But now, we have high-quality microphones that can record this symphony right in your own bedroom.

The problem? We have terabytes of these bedroom recordings, but we don't really know how to "read" the music to predict what might happen to your health 10 or 15 years down the road.

Enter SleepJEPA. Think of it as a super-smart, AI-powered "music critic" that has listened to over 422,000 hours of sleep recordings from more than 55,000 people. It doesn't just count how many times you toss and turn; it learns the deep, hidden language of sleep to predict your future health risks.

Here is a breakdown of how it works and what it found, using simple analogies:

1. The "Sleep Translator" (How it Learns)

Most AI models are like students who memorize answers to specific questions. If you ask them a slightly different question, they get confused.

SleepJEPA is different. It uses a technique called JEPA (Joint Embedding Predictive Architecture). Imagine you are trying to learn a new language. Instead of memorizing a dictionary, you listen to a conversation, cover up a word, and try to guess what it was based on the context of the sentence.

• The Analogy: SleepJEPA looks at your sleep data, hides a small chunk of it (like a 3-second snippet of your heart rhythm or brain waves), and tries to predict what that missing piece should have been based on the rest of the night.
• The Result: By doing this millions of times, it doesn't just memorize patterns; it learns the essence of what a healthy or unhealthy sleep "feels" like in a hidden, mathematical space. It becomes a master translator of sleep physiology.

2. The Crystal Ball (Predicting Disease)

Once the AI has learned the "language" of sleep, the researchers asked it a big question: "Based on this one night of sleep, what is the risk of this person getting sick in the next 1 to 15 years?"

They tested it on 10 different health conditions, acting like a crystal ball for:

• Heart Trouble: Heart attacks, heart failure, and high blood pressure.
• Brain Health: Cognitive decline (memory loss) and stroke.
• Metabolic Issues: Diabetes and sleep apnea.

The Results:
SleepJEPA was surprisingly accurate. For example:

• It predicted heart failure with an accuracy score of 0.83 (where 1.0 is perfect).
• It predicted heart disease death with a score of 0.86.
• It even predicted diabetes and stroke with high accuracy, looking at data collected years before the disease actually appeared.

It's like the AI noticed that the "rhythm" of your sleep was slightly off, and that tiny off-beat was a warning sign that your heart might struggle a decade later.

3. The "Sleep Staging" Scorecard

Before predicting diseases, the AI had to prove it could do the basics. Doctors usually spend hours manually labeling your sleep into stages: Awake, Light Sleep, Deep Sleep, or REM (Dreaming).

SleepJEPA did this automatically with 77% accuracy (which is very high for a computer). It's like a referee that can instantly tell if you are in a deep dream or just dozing off, without needing a human to watch the screen. This is crucial because it means the AI understands the structure of your night, not just the raw numbers.

4. The "Sleepiness" Detector

The researchers also tested if the AI could tell if you were dangerously sleepy during the day (a sign of narcolepsy or severe sleep apnea).

• It successfully identified people with Type 1 Narcolepsy with high accuracy (88%).
• It could spot objective daytime sleepiness, which is a major risk factor for accidents and early death.

5. Why This Matters (The "So What?")

Currently, if you want to know your sleep health, you have to go to a lab, wear a bunch of wires, and pay a lot of money. This is hard to do for millions of people.

SleepJEPA changes the game because:

• It works with home data: It was trained on "at-home" sleep studies, which are cheaper, easier, and less stressful.
• It's a long-term view: Most health checks tell you what's wrong today. SleepJEPA tells you what might go wrong tomorrow (or 15 years from now).
• It's a "Foundational Model": Just like a foundation supports a whole building, this AI model supports many different tasks. You don't need to build a new AI for every disease; you just use this one smart model and ask it different questions.

The Bottom Line

Think of SleepJEPA as a preventive health guardian. It listens to the quiet, nightly symphony of your body and whispers a warning: "Hey, your heart rhythm during deep sleep looks a bit shaky. If we don't fix your sleep habits now, you might be at risk for heart trouble in 10 years."

By turning a single night of sleep data into a long-term health forecast, this technology could help doctors catch diseases early, saving lives and money, all while you sleep in your own bed.

Technical Summary

1. Problem Statement

Sleep disturbances are a global epidemic linked to significant risks for cardiovascular (CV), metabolic, and neurological diseases. While in-lab polysomnography (PSG) is the gold standard for sleep analysis, it is expensive, resource-intensive, and not scalable. At-home PSG (hPSG) offers a scalable alternative with similar data richness (EEG, EOG, EMG, ECG, SpO2, respiratory signals), but the link between hPSG measurements and future long-term disease risk remains poorly understood.

Existing AI approaches in sleep medicine face several limitations:

• Data Utilization: Many models rely on single-modality data or aggregated features, failing to capture subtle, multi-channel dynamics.
• Generalizability: Most studies are trained and tested on single cohorts, limiting out-of-distribution performance.
• Methodology: Traditional supervised models focus on single outcomes. Recent self-supervised models often use Masked Autoencoders (MAE) or contrastive learning, which can be noisy or prone to representational collapse.
• Risk Horizon: Existing foundational models often limit risk estimation to short timeframes (e.g., 5–6 years), failing to capture long-term disease trajectories (up to 15 years).

2. Methodology: SleepJEPA Architecture

The authors propose SleepJEPA, the first Joint Embedding Predictive Architecture (JEPA) designed for sleep signal representation.

• Architecture Design:

  • Input: Variable-length hPSG recordings (6–12 hours) from 7 signal channels: EEG (C4-M1/C3-M2), EOG, Chin EMG, ECG (Lead II), SpO2, and Respiratory Inductance Plethysmography (RIP) for thorax and abdomen.
  • Tokenization: Signals are tokenized into 3-second patches.
  • JEPA Mechanism: Unlike MAE which reconstructs raw inputs, JEPA learns representations in a latent space.
    1. Context Encoder: Processes unmasked input tokens to create context embeddings.
    2. Predictor: A smaller network that predicts the embeddings of masked target tokens based on the context.
    3. Target Encoder: A copy of the context encoder with Exponential Moving Average (EMA) weight updates, generating the ground-truth target embeddings.
    4. Loss Function: A regression loss compares the predictor's output to the target embeddings. This approach aims to learn semantic physiological states while reducing noise and preventing representational collapse.
  • Backbone: Utilizes Patch Time Series Transformers (Patch-TST) for the context, target, and predictor networks.

• Training Strategy:

  • Pretraining: Self-supervised training on 34,291 sleep studies (27,067 unique patients) totaling ~270,000 hours.
  • Monitoring: Linear probing on sleep staging (Wake, N1, N2, N3, REM) was used to select the optimal pretraining epoch (Epoch 16) to avoid collapse.
  • Finetuning: Frozen representations were used to train lightweight heads for:
    • Sleep Staging: Bidirectional GRU.
    • Age Estimation: Attentive regression.
    • Daytime Sleepiness/Narcolepsy: Attentive classifiers (using MSLT scores).
    • Long-term Disease Risk: Attentive classifiers using a Discrete Hazard Loss function. This was chosen over Cox Proportional Hazards to handle non-proportional hazards, varying observation effects over time, and batch-size dependency issues.

• Datasets:

  • Pretraining: Human Sleep Project, Mount Sinai PSG, WSC, APPLES, MNC.
  • Finetuning/Testing: SHHS (Sleep Heart Health Study), MrOS (Osteoporotic Fractures in Men), MESA (Multi-Ethnic Study of Atherosclerosis), WSC.
  • Total Data: 422,035 hours of sleep data from 55,518 studies.

3. Key Contributions

1. First JEPA for Sleep: Introduces a foundational model specifically for multi-channel, variable-length hPSG data using joint embedding predictive architecture.
2. Long-Horizon Risk Estimation: Successfully estimates cumulative disease risk up to 15 years for 10 distinct conditions, a significant extension over previous 5–6 year models.
3. Discrete Hazard Modeling: Implements a discrete hazard loss function to better model time-to-event data without the strict assumptions of Cox models.
4. Comprehensive Evaluation: Validates performance across held-out test sets and independent cohorts (e.g., MESA, MrOS, WSC), demonstrating robust generalizability.
5. Multi-Task Capability: The same latent representations effectively support sleep staging, age prediction, daytime sleepiness, narcolepsy detection, and diverse disease risk assessments.

4. Key Results

A. Sleep Staging & Physiological Tasks

• Sleep Staging: Achieved a macro-F1 of 0.77 on held-out tests and 0.61 on independent tests (MrOS Visit 2), performing comparably to state-of-the-art models like SleepFM and L-SeqSleepNet, despite using frozen representations.
• Age Prediction: Strong correlation with chronological age (Pearson $r=0.80$ on held-out, $0.61$ on independent).
• Daytime Sleepiness & Narcolepsy:
  • Objective daytime sleepiness (MSLT $\le$ 8 min): AUC = 0.64.
  • Type 1 Narcolepsy: AUC = 0.88 (held-out), dropping to 0.56 on independent tests, suggesting distinct feature sets for narcolepsy vs. general sleepiness.

B. Long-Term Disease Risk Estimation

SleepJEPA+ (SleepJEPA + Age/Sex) significantly outperformed demographics-only baselines and matched or exceeded recent foundational models (SleepFM, He et al.).

• 15-Year Cumulative Risk (iAUC):
  • Coronary Heart Disease Death: 0.86
  • Congestive Heart Failure: 0.83
  • CV Disease Death: 0.81
  • Myocardial Infarction: 0.82
  • Obstructive Sleep Apnea (5-year): 0.86
  • Diabetes (10-year): 0.82
  • Cognitive Decline (10-year): 0.79
• Generalization: In independent cohorts (e.g., MESA), the model maintained strong performance (e.g., CV Death iAUC = 0.80 at 10 years), outperforming demographics-only models.

C. Interpretability

• Feature Correlations: Risk scores correlated strongly with age, sleep efficiency, SpO2, and Apnea-Hypopnea Index (AHI).
• Attribution Analysis (Integrated Gradients):
  • Respiratory (RIP) and EMG channels were top contributors for most CV and metabolic outcomes.
  • ECG was highly attributed for Coronary Heart Disease Death and Hypertension.
  • SpO2 and EEG were critical for Cognitive Decline and Diabetes.
  • The model distributed attention across channels, suggesting it learns complex, multi-modal physiological states rather than relying on a single signal.

5. Significance and Impact

• Scalability: By utilizing at-home sleep data, SleepJEPA offers a scalable, non-invasive method for long-term disease risk stratification, potentially democratizing access to sleep-based prognostics.
• Clinical Utility: The ability to estimate risk up to 15 years allows for early intervention in younger populations with low short-term risk but high long-term trajectory risk.
• Methodological Advancement: Demonstrates that JEPA architectures are superior to MAE/Contrastive learning for physiological time-series data, reducing noise and improving latent space representation.
• Future Directions: The authors suggest combining SleepJEPA with other clinical features and exploring reduced signal sets (e.g., wearable devices) to further lower barriers to clinical deployment.

Conclusion: SleepJEPA represents a significant leap in sleep medicine AI, transforming raw at-home sleep signals into a powerful, generalizable foundation for predicting long-term cardiovascular, metabolic, and neurological disease risks.