SleepFM: AI Model Predicts Disease Risks Based on Sleep Data

Contents

Scientists at Stanford University have developed a new, multimodal AI model called “SleepFM.” This “Foundation Model” uses vast datasets of sleep recordings to not only analyze the current sleep state but also predict the risk of developing over 130 different medical conditions. The study, published in Nature Medicine, shows that sleep measurements can serve as a valuable tool for general health assessment, going far beyond the diagnosis of sleep disorders.

Training an AI on the “Language of Sleep”

In the medical community, polysomnography (PSG) is considered the gold standard for sleep assessment. A PSG recording contains various signals such as heart activity (electrocardiogram; ECG), brain waves (electroencephalography; EEG), muscle movements (electromyography; EMG), and respiratory values. Due to the complexity and volume of this data, previous studies have mostly focused on the manual evaluation of sleep stages or the identification of specific events like apneas.

To analyze the vast amount of data, the scientists in their study adopted an approach similar to modern language models that form the basis for AI applications like ChatGPT. While large language models learn how words relate to each other in sentences, SleepFM was trained on over 585,000 hours of PSG data from more than 65,000 individuals. This allowed the model to learn the fundamental “physiology of sleep”.

Reviewing Basic Sleep Functions

Before investigating disease risks, the researchers validated SleepFM on classic sleep medicine tasks. The AI demonstrated an accuracy in automatically determining sleep stages comparable to specialized expert models. Furthermore, the model is capable of estimating a person's biological age from sleep data alone with an average deviation of only about 7.33 years. It also detected apneas with an accuracy of up to 87 percent.

Predicting diseases and Mortality Risk

However, SleepFM's most impressive capability lies in long-term prognostication. Based on a single night's data, the model could predict whether a patient would develop certain diseases within the next few years. Researchers use the so-called “C-index” to measure predictive power, a standard measure for predictive strength. In medical research, a C-index of 0.70 to 0.80 is considered “good,” and anything above 0.80 is considered “very good” to “excellent” (a value of 1.0 is a perfect result).

“SleepFM” was able to calculate with high accuracy the likelihood of developing many different diseases (including cardiovascular diseases, dementia, cancer, metabolic diseases) within 6 years of the sleep measurement.

The model was particularly good at predicting Alzheimer's (0.91), prostate cancer (0.89), heart failure (0.97), diabetes (0.87), and also general mortality risk (0.84).

A key finding of the study: SleepFM significantly outperforms traditional predictive models that rely solely on data such as age, sex, and body mass index (BMI). This implies that hidden physiological patterns exist in sleep that provide far more precise information about health status than classic medical indicators.

Overcoming Lack of Data Standardization in Sleep Medicine

One of the biggest challenges in analyzing sleep data is the lack of data standardization in PSG. Different clinics use varying numbers of sensors and different recording settings. SleepFM employs a “channel-agnostic” architecture and a novel “leave-one-out” approach for AI training, utilizing different PSG signals (cardiac, brain, respiratory). This allows the model to deliver reliable results even when individual sensors are missing or data quality fluctuates.

To investigate the model's reliability, the researchers also tested SleepFM on entirely new datasets that were not included in the training. Here too, the predictive power remained stable, demonstrating that the model can be used in various environments.

Potential Significance for Preventive Medicine

Sleep measurements can thus provide a wealth of valuable data for overall health status and disease risks, which can now also be interpreted using AI. In the future, sleep studies could therefore be used for routine health check-ups, for example. By detecting early “warning signs” for dementia, cardiovascular diseases, or cancer, timely investigations could be initiated, potentially years before symptoms appear.

With the further development of wearables (such as smartwatches), the insights from AI models like SleepFM could even find their way from the sleep lab into everyday life in the future. This would enable continuous, non-invasive health monitoring.

The scientists conclude that their approach offers a scalable path towards personalized, preventive healthcare, where sleep can be used as a basis for assessing disease risks.

(mki)