SleepGPT: A Sleep Stage Language Model for Efficient Sleep Assessment

Accurate and scalable sleep assessment is crucial for diagnosing sleep disorders and advancing personalized medicine. However, current approaches heavily rely on manual scoring of polysomnography (PSG) recordings, which is labor-intensive, costly, and difficult to scale. Here, we introduce SleepGPT, the first sleep-specific language model trained on over 5.8 million sleep stage annotations from 5,793 whole-night recordings to model the sequential dynamics of sleep macrostructure. We demonstrate that SleepGPT consistently enhances the performance of multiple state-of-the-art sleep staging models across diverse independent datasets (total N=1,320,654), and enables high staging accuracy from low-density wearable EEG recordings (N=120,095). We further show that SleepGPT supports robust diagnosis of sleep disorders when embedded into a sequence-level Transformer framework. The model accurately identifies abnormal sleep stage dynamics and Type-1 narcolepsy across multiple clinical datasets (total N=685). Attention-based visualization of model outputs revealed interpretable macrostructural patterns associated with diagnostic labels. Together, these results demonstrate that SleepGPT serves as a general-purpose foundation model for sleep sequence modeling, offering a scalable and clinically translatable tool for automated sleep staging and sleep disorder diagnosis.