Remote Monitoring of Algorithm-Detected Compensation Cues During Long-Term Home-Based Upper-Limb Training After Stroke: A Case Series

This retrospective case series examined whether a webcam application can monitor algorithm-detected compensation cues during long-term home upper-limb training after stroke. Rules on MediaPipe landmarks, not validated in patients, trigger real-time cues for elbow flexion, shoulder elevation, lateral trunk lean, and head tilt. In a technical check on 621 clinician-labelled repetitions by healthy adults, the 90° abduction rules were specific but missed most simulated compensations. Six patients with post-stroke hemiparesis, selected for long, regular records, trained at home on 33-96% of days over 3.8-9.9 months. The primary outcome, the cue rate (share of repetitions with the cue), was modelled per patient across 2277 control sessions (first daily executions of 180° and 90° abduction and a hand-to-mouth task). The elbow cue was most frequent in 180° abduction and the shoulder cue in the hand-to-mouth task. Only four of 1s6 categorised leading-cue trajectories (all elbow cue) held in every applicable sensitivity analysis; their directions differed. Long-term markerless home monitoring of the cues was feasible in these selected patients. Cue rates reflect pace, cue priority, and programme adjustment, and clinical assessment was not repeated, so their changes are not evidence of clinical improvement or deterioration; validation in patients and controlled studies should follow.