LogicSleep combines neural models with physiological rules

LogicSleep uses a CNN–graph U-Net, conditional random fields and differentiable logic rules to discourage implausible sleep-stage transitions, regularise dwell times and support subject-specific test-time adaptation. Authors report consistent gains across ISRUC and MASS datasets.

Why it matters: Neurosymbolic constraints can make clinical sequences more plausible and explanations more useful than unconstrained classification.

Impact: Multi-dataset results are promising, but prospective clinical workflow, demographic robustness and diagnostic benefit are not established. npj Artificial Intelligence

Scientific AI · United Kingdom