SleepLM: Natural-Language Intelligence for Human Sleep

We present SleepLM, a family of sleep-language foundation models that enable human sleep alignment, interpretation, and interaction with natural language. Despite the critical role of sleep, learning-based sleep analysis systems operate in closed label spaces (e.g., predefined stages or events) and fail to describe, query, or generalize to novel sleep phenomena. SleepLM bridges natural language and multimodal polysomnography, enabling language-grounded representations of sleep physiology. To support this alignment, we introduce a multilevel sleep caption generation pipeline that enables the curation of the first large-scale sleep-text dataset, comprising over 100K hours of data from more than 10,000 individuals. Furthermore, we present a unified pretraining objective that combines contrastive alignment, caption generation, and signal reconstruction to better capture physiological fidelity and cross-modal interactions. Extensive experiments on real-world sleep understanding tasks verify that SleepLM outperforms state-of-the-art in zero-shot and few-shot learning, cross-modal retrieval, and sleep captioning. Importantly, SleepLM also exhibits intriguing capabilities including language-guided event localization, targeted insight generation, and zero-shot generalization to unseen tasks. To support reproducibility and future work, we open-source the captioning pipeline, pretrained checkpoints, and the model architectures at https://github.com/yang-ai-lab/SleepLM.

Sleep is a vital part of human health, but the signals recorded during a sleep study are complex and usually require expert interpretation. Most current AI systems for sleep can only assign fixed labels, such as sleep stages or apnea events, rather than explain what is happening in natural language. We introduce SleepLM, a family of AI models that connects overnight sleep recordings with text, allowing sleep patterns to be described, searched, and summarized using everyday language. To train these models, we built a large sleep-text dataset by generating multi-level descriptions for over 100,000 hours of sleep recordings from more than 10,000 people. SleepLM learns from brain, heart, breathing, muscle, and is trained to both match signals with descriptions and generate sleep-related captions. In experiments, SleepLM performed better than strong language and vision-language models on sleep staging, event detection, caption generation, retrieval, and few-shot learning. It can also localize events in time and generalize to some sleep concepts that were not explicitly seen during training. This work shows a path toward more interpretable sleep AI systems that could help researchers and clinicians explore long sleep recordings more efficiently. SleepLM is not a clinical diagnostic tool, but with proper validation, similar systems may eventually support clearer sleep reports and broader access to sleep health insights.