An AI research team has built a model that can describe, query, and interpret human sleep physiology using plain language rather than rigid diagnostic categories.
SleepLM is a family of foundation models trained on a new dataset built by the same team: over 100,000 hours of polysomnography recordings from more than 10,000 people, paired with text captions generated through a multilevel pipeline the researchers designed specifically for this project. The model combines three training objectives - contrastive alignment, caption generation, and signal reconstruction - to link raw physiological signals with natural-language descriptions. In benchmarks, it outperformed existing approaches on zero-shot and few-shot tasks, cross-modal retrieval, and sleep captioning. Code and data are set to be released publicly.
The practical gap SleepLM targets is real: current sleep-analysis software forces clinicians into fixed label sets, meaning anything outside the predefined categories simply cannot be captured or queried. A model that accepts free-form language questions and returns grounded interpretations of sleep signals could change how researchers study novel sleep phenomena and how clinicians document edge cases. The zero-shot generalization claim is the one worth watching - if it holds under independent evaluation, it meaningfully lowers the cost of applying AI to rare or poorly labeled sleep conditions.
Sleep medicine has been slow to attract the foundation-model treatment that has reshaped radiology and pathology; SleepLM is an early entry, and the open-source release will determine whether the benchmark numbers survive contact with messier real-world data.