Foundation Models in Sleep Research: Opportunities and Limitations

Foundation Models (FMs) for sleep are large-scale, self-supervised models pretrained on extensive datasets and adapted for various downstream tasks. This Commentary assesses the current state of sleep FMs, emphasizing their benefits while also critically examining whether they are ready for clinical use and the standards needed to evaluate them properly.

We provide an overview of recently published sleep FMs, evaluating their training cohorts, assessment frameworks, and reported performance. Furthermore, we discuss key challenges related to data bias, interpretability, and scientific communication of findings. To illustrate practical limitations, we apply an existing sleep FM, without fine-tuning, to an independent cohort of patients with Narcolepsy Type 1 (n = 51) and healthy controls (n = 28).

Training cohorts across reviewed models were consistently biased toward older, predominantly mono-ethnic populations with established comorbidities. Evaluation frameworks are inconsistent, supervised comparisons are scarce, and disease prediction claims are difficult to interpret without proper demographic ablations.The illustrative example demonstrates that, as expected for an untuned FM, the zero-shot sleep-staging performance was modest and lower than that of supervised methods on the same cohort. Additionally, PSG-derived embeddings offered minimal improvement in disorder classification beyond demographic baselines.

Sleep FMs have the potential to advance sleep medicine by providing scalable, transferable representations across diverse datasets and clinical tasks. However, they are not yet suitable for clinical deployment. The field requires standardized evaluation methods, transparent reporting of limitations, and careful communication of results. Addressing these challenges is essential for integrating them into clinical routine.