Long-term markerless webcam monitoring of algorithm-detected compensation cues during home upper-limb training was feasible in six selected post-stroke patients, but cue-rate changes are not evidence of clinical improvement or deterioration and the rules have not been validated in patients.
In a technical check on healthy adults, the 90° abduction compensation rules were specific but missed most simulated compensations.
Six selected post-stroke patients with hemiparesis trained at home over several months with substantial but variable adherence.
Cue rates (share of repetitions with a cue) were modelled per patient across 2277 control sessions.
The most frequent cue differed by task: the elbow cue in 180° abduction and the shoulder cue in the hand-to-mouth task.
Few leading-cue trajectories were robust to sensitivity analyses, and their directions differed between patients.
Long-term markerless home monitoring of compensation cues was feasible in these selected patients.
Changes in cue rates cannot be interpreted as clinical improvement or deterioration.
This study looked at whether a simple webcam program could track unhelpful 'compensation' movements during home arm exercises after a stroke. Compensation means using other body parts, such as bending the elbow, hiking the shoulder, leaning the trunk, or tilting the head, to make up for a weak arm. The software uses body-tracking landmarks to give real-time cues when it detects these movements. Six stroke patients who had long, regular training records used it at home for about 4 to 10 months, training on roughly a third to nearly all days. The researchers found that continuous home monitoring was feasible in these patients. The types of cues differed by exercise: elbow cues were most common when raising the arm fully overhead, and shoulder cues were most common in a hand-to-mouth task. However, only four of 16 trends in the most frequent cue held up across all robustness checks, and these trends went in different directions for different patients. In a test with healthy adults, the rules for the 90° arm raise rarely raised false alarms but missed most simulated compensations. This research suggests such webcam tools could be used for remote tracking, but they are not yet proven. The rules were not validated in patients, the sample was small and selected, and clinical assessments were not repeated. Changes in cue rates also depended on exercise pace, which cues were prioritised, and programme adjustments. They should therefore not be taken as signs that a patient was getting better or worse. Further validation and controlled studies are needed.
Skvortsov D, Lobunko D, Petrosyan V, Agapov V, Pavlikov A, Gorodnichev M. (2026). Remote Monitoring of Algorithm-Detected Compensation Cues During Long-Term Home-Based Upper-Limb Training After Stroke: A Case Series.. Sensors (Basel, Switzerland). https://doi.org/10.3390/s26206651