Foundation Models in Sleep Research: Opportunities and Limitations

It systematically reviews recently published sleep foundation models in terms of training cohorts, assessment frameworks, and reported performance, noting that training cohorts are consistently biased toward older, predominantly mono-ethnic populations with established comorbidities. It places separately published sleep foundation models under a single comparative lens, making training population composition and evaluation approach comparable objects rather than isolated result statements. This is a commentary-style review based on the training cohort descriptions and reported performance of the reviewed models, without new primary cohort statistics.

Through an illustrative example of an existing sleep foundation model applied without fine-tuning to an independent cohort, it shows modest zero-shot sleep-staging performance that is lower than supervised methods on the same cohort. It tests the question of whether a foundation model transfers directly on a concrete, independent cohort that includes a clinical population, rather than relying on models' self-reported benchmark scores. The illustrative cohort comprises 51 patients with Narcolepsy Type 1 and 28 healthy controls under a zero-shot, no-fine-tuning setting, and the authors explicitly note this result is expected for an untuned foundation model.

PSG-derived embeddings offer minimal improvement in disorder classification beyond demographic baselines. It suggests that when assessing the disease-prediction value of sleep foundation models, demographic variables should serve as a comparison baseline, otherwise it is hard to tell whether gains come from physiological signals or population characteristics. Based on a comparison of embeddings against demographic baselines in the same independent cohort, this is an illustrative analysis rather than large-scale validation.

It argues that the field requires standardized evaluation methods, transparent reporting of limitations, and careful communication of results, holding that sleep foundation models have potential but are not yet suitable for clinical deployment. It moves the discussion from single-model performance to field-level evaluation norms and result communication, touching on data bias, interpretability, and scientific communication. This is a commentary and position-oriented argument based on a synthesis of the preceding review and illustrative analysis.

This Commentary is aimed at researchers in sleep medicine and physiological-signal machine learning as well as those setting evaluation standards, and it applies to discussions of how ready sleep foundation models are for clinical deployment and what evaluation norms are needed; its illustrative conclusions pertain to a no-fine-tuning, zero-shot usage setting and to the specific cohort composition of Narcolepsy Type 1 patients and healthy controls.

A careful reader would still want to know the specific list of reviewed models and details of their reported performance, the concrete manifestations of inconsistent evaluation frameworks, and the full numerical comparison of embeddings against demographic baselines in the example; the loaded text is summary-level and does not include figures or complete data, so these details cannot be confirmed from the available text.