SleepLM: Natural-Language Intelligence for Human Sleep

A research team led by Zongzhe Xu has introduced SleepLM, a groundbreaking family of foundation models designed to bridge natural language with human sleep physiology data. The system addresses a critical limitation in traditional sleep analysis, which operates within closed label spaces (like predefined sleep stages) and cannot describe or query novel sleep phenomena. SleepLM enables natural language interaction with multimodal polysomnography (PSG) data, creating language-grounded representations of sleep. To achieve this, the team developed a multilevel sleep caption generation pipeline, which was used to curate the first large-scale sleep-text dataset—comprising over 100,000 hours of data from more than 10,000 individuals—providing the necessary training corpus for this novel alignment.

The model employs a unified pretraining objective combining contrastive alignment, caption generation, and signal reconstruction to capture both physiological fidelity and cross-modal interactions. Extensive experiments demonstrate that SleepLM outperforms current state-of-the-art methods in zero-shot and few-shot learning, cross-modal retrieval, and sleep captioning tasks. Importantly, it exhibits emergent capabilities including language-guided event localization, targeted insight generation, and zero-shot generalization to unseen diagnostic tasks. The researchers have committed to open-sourcing all code and data, which could democratize advanced sleep analysis and enable new applications in clinical sleep medicine, personalized health tracking, and neuroscience research.