AI Sleep Co-Scientist predicts Parkinson's and Alzheimer's risk from sleep data

A team led by Rahul Thapa, James Zou, and others at Stanford University has introduced the AI Sleep Co-Scientist—an expert-guided multi-agent environment that lets human researchers direct specialist AI agents through hypothesis generation, signal preprocessing, and statistical analysis. The system was applied to four cohorts comprising roughly 124,000 polysomnography (PSG) recordings and over 50 terabytes of raw sleep signals. In five case studies, the AI uncovered novel connections between sleep physiology and disease. Diminished network-level physiological coupling during sleep was associated with a 1.48x hazard ratio for incident Parkinson's disease and 1.38x for Alzheimer's. A physiologically structured sleep-age model outperformed early-fusion approaches, and its age residual predicted disease across multiple organ systems.

Beyond disease risk, the AI Sleep Co-Scientist shed light on clinical classifications and sleep regulation. Arousal dynamics characterized comorbid insomnia and sleep apnea as an intermediate phenotype dominated by obstructive sleep apnea, distinguished by prolonged wakefulness after arousal. It also found that REM bout duration tracks preceding non-REM sleep more closely than intervening wakefulness, and transient-oscillation analysis revealed a fast-sigma deficit and excess centrofrontal theta activity in narcolepsy type 1. The system's design ensures full reproducibility—every reported result links directly to the executable code that produced it. This work demonstrates how agentic AI can scale multimodal discovery in sleep medicine, potentially leading to earlier diagnosis and personalized interventions for neurological and sleep disorders.