摘要
Introduction & Background. Functional imaging signatures of sleep loss rely on classifiers trained on labelled scans. Foundation fMRI models pretrained without sleep labels may already encode this state.
Methods. Brain-Semantoks (Gijsen et al., ICLR 2026) was pretrained by self-supervision on 39,139 UK Biobank resting-state fMRI scans, without sleep labels. Its frozen encoder converted each resting-state or task fMRI run into an embedding. In a Sleep Restriction study (N=35; scans at baseline, after five nights with 5-h sleep opportunities, and after one night of recovery sleep), the sleep-loss direction was the mean embedding change from baseline to restriction. Without refitting, this direction was applied to Sleep Deprivation study 1 (N=31, after one night awake: 13 following test-retest baselines, 18 following 5-h restriction and one recovery night) and study 2 (N=38: 19 before and after one night awake, 19 rested test-retest controls). Network contributions were measured by activation patching. Global-signal amplitude, autocorrelation and motion were erased by linear concept erasure.
Results. Embeddings moved toward the sleep-loss side in 26 of 35 after restriction (dz = 0.75) and in 11 of 13, 16 of 18 and 19 of 19 after one night awake (dz = 1.51, 1.07, 1.62), more than between test-retest scans without sleep loss (dz = −0.38, 0.42). A direction learned from acute deprivation reproduced these results. Somatomotor input carried the largest share in all three sleep-loss cohorts (31%, 28%, 21%); subcortical input ranked second after chronic restriction (24%). After erasure, recovery sleep moved the sleep-restriction embeddings back toward the rested side (dz = −0.48, p = 0.008).
Conclusions. Without training on sleep/wake information, a pretrained fMRI model captures sleep loss across cohorts, anchors the difference in somatomotor activity, with a subcortical contribution after chronic restriction, and tracks recovery sleep.
