Sleep Staging Foundation Models Encode Neural Disorder-Related EEG Representations that Generalize to Wakefulness

ObjectiveTo leverage sleep foundation models trained on large datasets of polysomnography for neurological disorder detection during an awake state. MethodsThree public resting-state electroencephalography (EEG) datasets from cognitively normal participants and patients with Alzheimers disease (AD), frontotemporal dementia (FTD), schizophrenia (SZ), and depression (MDD) were used. Embed-dings from three pre-trained sleep staging foundation models were extracted and compared against a spectral model for AD, FTD, SZ, and MDD detection. Analyses used a subset of the available EEG channels, enabling comparisons between each EEG channel and the detection capability of the models. ResultsAD, FTD, and SZ classifications were best performed by the spectral model (AUROC=0.931, 0.878, 0.964), followed closely by SleepFM (AUROC=0.889, 0.863, 0.821), which was consistently the top-performing sleep staging model for wakeful disorder detection. All classifiers struggled to detect MDD and distinguish between AD and FTD (AUROC <0.80). For MDD classification, SleepFM (AUROC=0.737) outperformed the spectral model (AUROC=0.710). Between the top classifiers, the T4 and O2 channels were selected as part of the most informative group of electrodes for two-thirds of the disorder classification tasks. ConclusionThe embeddings obtained from the sleep-staging foundation models captured meaningful information for detecting neurological disorders, even when using EEG from wakeful states. AD and FTD pathologies were the most sensitive to sleep staging analyses. SignificanceOur findings suggest that there is information in wakeful states proximal to sleep that reveals disease-induced neurological dysfunction, which can be represented by sleep staging models with no prior understanding of existing disorders.