The Autonomous Clinic: Navigating AI, Sensors and the Equity Gap in Modern Physiotherapy

Severe motor impairment is the clinical situation in which post-stroke rehabilitation performs least well, because most active therapies presuppose a degree of residual voluntary movement that the most affected survivors do not possess. Two technologies have been proposed to close this gap. Brain–computer interfaces (BCIs) detect movement-related cortical activity, most often modulation of sensorimotor rhythms recorded by electroencephalography (EEG), and therefore provide an output channel that does not depend on residual muscle activation. Robotic exoskeletons apply torque across anatomically aligned joints and can deliver structured, repeatable and proprioceptively rich movement to a limb that cannot move itself. Coupling the two creates a closed loop in which decoded intention is rewarded immediately with a somatosensory consequence. This review asks whether that convergence constitutes a coherent therapeutic framework for severe stroke, or an assembly of components whose separate limitations persist when they are combined. Literature was identified through structured searching of major open scholarly sources up to 17 June 2026, supplemented by backward and forward citation tracking and by appraisal of methodological quality. The available evidence indicates that contingent closed-loop training alters cortical and corticospinal measures in ways that non-contingent conditions do not, and that trials coupling interface control to peripheral actuation generally report favourable impairment-level effects. Confidence in the magnitude and durability of those effects is limited by small samples, inconsistent comparator design, incomplete dose matching, and outcome batteries that capture impairment more reliably than everyday limb use. Evidence specific to severe impairment is thinner still, because the participants with the greatest deficit are the participants most often excluded at screening, most likely to yield unstable decoding, and least well served by activity-level instruments. Exoskeleton-specific evidence is not interchangeable with evidence from end-effector robots or electrical stimulation, although reviews frequently pool them. Priorities include severity-stratified trials with contingency-controlled comparators, decoding strategies validated in lesioned brains, outcome measures sensitive at the floor, and economic evaluation conducted at service scale.