Named ‘SleepFM’, this model was developed by a team that includes researchers from Stanford University in the US and was trained on approximately 600,000 hours of sleep data gathered from 65,000 participants.
The AI system, detailed in a publication in the journal Nature Medicine, was initially evaluated through standard sleep analysis tasks, such as monitoring various sleep stages and assessing the severity of sleep apnea.
Subsequently, the model was utilized to foresee the future onset of diseases by examining sleep data, while health record information was obtained from a sleep clinic.
Researchers analyzed over 1,000 disease categories from the health records, finding that 130 could be predicted with a reasonable degree of accuracy using a patient’s sleep data.
“We capture an incredible array of signals while studying sleep. It allows us to observe general physiology for an extended period in a subject who is fully engaged. It’s a treasure trove of data,” stated senior author Emmanual Mignot, a professor in sleep medicine at Stanford University’s psychiatry and behavioral sciences department.
Polysomnography, regarded as the gold standard in sleep studies, is a widely used method for collecting sleep data. It employs sensors to measure brain activity, heart function, respiratory signals, and eye movements, among other metrics.
The AI model demonstrated the ability to integrate various data streams—such as electroencephalography (brain electrical activity), electrocardiography, electromyography (muscle electrical activity), pulse rate, and respiratory airflow—and analyze the relationships among them, according to the researchers.
The team introduced a novel training technique for the AI, termed ‘leave-one-out’ contrastive learning. This method obscures one data stream and prompts the model to reconstruct the missing information using the remaining signals.
Predictions from the AI system were particularly robust for conditions such as cancer, pregnancy complications, circulatory issues, and mental health disorders, achieving a C-index score exceeding 0.8.
The C-index, or concordance index, serves as a standard measure of an AI’s predictive capability—specifically, its proficiency in determining which individual in a group will encounter an event first, as noted by the researchers.
“From just one night of sleep, SleepFM can accurately predict 130 conditions with a C-index of at least 0.75, including all-cause mortality (C-index 0.84), dementia (0.85), myocardial infarction (0.81), heart failure (0.80), chronic kidney disease (0.79), stroke (0.78), and atrial fibrillation (0.78),” the authors stated.
“While performance varies across different categories, SleepFM shows promising results in several areas, such as neoplasms (tumors), pregnancy complications, circulatory disorders, and mental health issues,” they added.
The AI model also showcased strong predictive performance for the risk of Parkinson’s disease—where sleep issues are among the early warning signs—as well as for developmental delays and disorders.
