Abstract Sleep medicine has rapidly adopted artificial intelligence, but most applications still prioritize automated sleep staging or single summary indices, limiting clinical translation when symptoms arise from within-stage dynamics. This review proposes a physiology-grounded framework in which artificial intelligence targets sleep microstructure and nocturnal motor activity as temporally structured expressions of sleep–wake control. We discuss how transient arousals and cyclic alternating pattern activity can be modeled as time-resolved instability trajectories rather than reduced to hourly counts, and why grounding models in established constructs improves interpretability and trust. We then examine motor events across the continuum from leg movements to periodic limb movements and large muscle group movements, emphasizing that periodicity, clustering, state dependence, and coupling to cortical and autonomic activation convey more clinical information than event counts alone. Because autonomic surges are measurable outside the laboratory, we highlight multimodal approaches integrating electroencephalography, electromyography, actigraphy, cardiopulmonary signals, and wearable photoplethysmography to infer instability and movement–autonomic coupling in ambulatory settings. Finally, we translate these outputs into clinician-readable phenotypes that may refine diagnosis, prognosis, and treatment stratification, and we define priorities for the field: harmonized labeling standards, multi-center external validation, calibration across age and comorbidity, explainable artificial intelligence approaches, and deployment as decision-support tools that complement expert judgment.
Mogavero et al. (2026) studied this question.