A new family of AI models trained on over 166,500 hours of sleep recordings outperforms existing systems on sleep staging and disease prediction — and the researchers documented exactly why prior models kept failing.
Polysomnography, the clinical gold standard for sleep assessment, produces wildly inconsistent data depending on the recording device and the patient population. A team of researchers curated recordings from nine public sources, built an open-source benchmark called SleepBench, and used it to stress-test four families of self-supervised training approaches. The result is OSF, a set of foundation models that hit state-of-the-art numbers across nine datasets. Their training recipe identified three problems with previous efforts: models broke down when sensors were missing at inference time, they failed to learn features that generalize across different recording setups, and they did not scale predictably with more data or bigger architectures.
The missing-channel finding is the most practically significant. Real clinical deployments rarely have perfect sensor coverage — patients move, leads fall off, and equipment varies by hospital. A model that collapses under those conditions is not useful outside a controlled lab. OSF's channel-invariant approach is a direct response to that gap, and the fact that scaling data, model size, and source diversity all reliably improved performance gives researchers a clearer roadmap than the field has had before.
Sleep AI has attracted serious attention from wearables companies and hospital systems alike, but most prior work has been narrow, proprietary, or both. OSF and SleepBench being fully open-source is either a genuine contribution to reproducible science or, at minimum, a way to set a benchmark that commercial competitors will have to beat in public.