Longitudinal Camera-Based Upper Limb Motor Recovery Monitoring in Stroke Survivors Using MediaPipe Hand Landmark Detection: A Home-Deployed Kinematic Classification System

Background: Upper limb motor impairment affects approximately 80% of stroke survivors. The Fugl-Meyer Assessment requires clinician administration and cannot be performed daily at home. Markerless motion capture using artificial intelligence offers a low-cost alternative for kinematic monitoring. Methods: A longitudinal browser-based upper limb kinematic monitoring system was implemented using Next.js 14, TypeScript, and MediaPipe Tasks Vision 0.10.14. The Hand Landmarker model extracts 21 landmarks per frame from a standard webcam. Eight kinematic features were computed per session. A three-class trend classifier (Improving, Stable, Declining) was implemented using threshold-based rules calibrated against published kinematic ranges. An FMA-UE proxy score (0–66) was estimated from kinematics. All processing occurs client-side with no video transmission. Results: Mean landmark extraction frame rate was 24.3 fps (SD = 2.1) across four exercises on a standard laptop webcam. All eight kinematic features were computed within 30-second exercise windows. Weekly report generation was validated across 20 simulated multi-session profiles with 100% trend classification accuracy. Network analysis confirmed no video data left the browser. Conclusion: A browser-based longitudinal upper limb kinematic monitoring system is technically feasible using MediaPipe Hand Landmarker. The system addresses a specific gap in longitudinal home-based upper limb monitoring. Clinical validation with stroke patients under ethical approval is required.