AI ‘Sleep Language’ Model Predicts Over 100 Health Conditions from One Night’s Data

Stanford Medicine researchers have developed SleepFM, a pioneering AI foundation model capable of forecasting risks for 130 diseases by analyzing a single night’s sleep. By "learning the language of sleep" through 600,000 hours of physiological data, the model identifies early markers for conditions ranging from dementia to cardiovascular disease.

By: AXL Media

Published: Feb 16, 2026, 10:13 AM EST

Source: Stanford Medicine

The "Untapped Gold Mine" of Sleep Data

A new study published in Nature Medicine reveals how Stanford scientists converted standard polysomnography a comprehensive sleep assessment into a powerful diagnostic tool. While traditional sleep medicine focuses on a narrow set of metrics like breathing interruptions, the SleepFM model analyzes the full spectrum of brain activity, heart rate, respiratory signals, and muscle movements recorded during an eight-hour session.

Emmanuel Mignot, MD, PhD, co-senior author, describes the sleeping patient as "completely captive" for eight hours, providing a data-rich environment that reflects general human physiology in real-time.

Learning the Language of Sleep

SleepFM functions similarly to Large Language Models (LLMs) like ChatGPT but replaces text with physiological signals. The model was trained on 585,000 hours of data from 65,000 participants. To master the relationships between different bodily functions, researchers used a "leave-one-out" contrastive learning technique, forcing the AI to predict missing data such as heart activity based solely on other signals like brain waves or breathing.

Forecasting Future Disease Onset