AI Predicts Disease Risks from One Night’s Sleep
News Context
At a glance
- Researchers at Stanford medicine have developed SleepFM, an artificial intelligence model that predicts the risk of developing over 100 diseases based on physiological data collected during a single...
- sleepfm was linked to long-term health data from approximately 35,000 patients (aged 2-96 years) with up to 25 years of follow-up data.
- The model identified 130+ conditions for which risk could be predicted with reasonable to high accuracy.
SleepFM: Predicting Disease Risk from Sleep
Table of Contents
Overview
Researchers at Stanford medicine have developed SleepFM, an artificial intelligence model that predicts the risk of developing over 100 diseases based on physiological data collected during a single night’s sleep. This represents a significant step towards predictive and personalized medicine.
Data & Training
- Polysomnography Data: sleepfm was trained on approximately 600,000 hours of polysomnography data.
- Participants: Data from roughly 65,000 participants was used.
- Data streams: The model combines and interprets multiple physiological signals concurrently, including:
- Brain activity
- Heart activity
- Breathing patterns
- Muscle activity
- Eye movements
- Training Method: A novel training method was used where signals were temporarily hidden, forcing the model to reconstruct them based on other signals.
- data Intervals: Nighttime measurements are divided into five-second intervals, analogous to words in a language model.
| Data Type | Description | Quantity |
|---|---|---|
| Polysomnography Hours | Total hours of sleep data used for training | ~600,000 |
| Number of participants | Total number of individuals contributing data | ~65,000 |
| Disease Categories Analyzed | Number of diseases the model attempted to predict | > 1,000 |
| Predictable Diseases | Number of diseases predicted with reasonable to high accuracy | 130+ |
Model Characteristics
- Foundation Model: SleepFM is a foundation model, trained on large datasets and adaptable to various tasks.
- “Language of Sleep”: the model learns to interpret the patterns and relationships within sleep data, effectively learning the “language of sleep.”
- Validation: SleepFM performed at least and also current state-of-the-art algorithms in existing clinical applications like sleep stage recognition and sleep apnea severity determination.
disease Prediction
sleepfm was linked to long-term health data from approximately 35,000 patients (aged 2-96 years) with up to 25 years of follow-up data.
The model identified 130+ conditions for which risk could be predicted with reasonable to high accuracy.
Predictions were particularly strong for:
- Neurological disorders (Parkinson’s, Dementia)
- Cardiovascular disease
- Various forms of cancer
- Mortality