An Adaptive Hand Exoskeleton Rehabilitation Training System Integrating Virtual Reality and an AI-Based Assessment Engine

Cui Junshuo^*

• School of Mechanical and Electrical Engineering, Shandong Jianzhu University, Jinan, China

Introduction Post-stroke hand motor impairment affects 70% of survivors, with conventional rehabilitation limited by low adherence and insufficient personalization. We propose a bio–AI–VR integrated system fusing biosignal sensing, AI assessment, and virtual reality for adaptive closed-loop training. Integration rationale includes: (i) multimodal data fusion—combining heterogeneous data through AI; (ii) closed-loop adaptive control—balancing challenge and capability; (iii) neuroplasticity multisensory enhancement—coordinating visual, proprioceptive, and motor pathways. This work tests three hypotheses: (RQ1) Can multimodal fusion achieve real-time assessment (R² ≥ 0.65, latency <50 ms)? (RQ2) Does integration yield FMA-UE improvement ≥ 6 points with d ≥ 0.8? (RQ3) Does ablation cause ≥ 15% degradation? Methods A lightweight exoskeleton (<400 g) integrates IMU (100 Hz) and 16-channel sEMG (1 kHz) to capture kinematics and muscle activation. A hybrid random forest and support vector regression model outputs real-time score St∈[0,1] via multi-task learning. FMA-UE proxy labels combined linear interpolation (80%), biomechanical anchoring (15%), and expert annotation (5%, κ=0.78). The assistance-as-needed algorithm adjusts exoskeleton torque and VR difficulty using St. Twenty-four stroke survivors (3–12 months post-stroke, FMA-UE 15–50) underwent 4-week training (5 sessions/week, 20 min/session). Analysis employed paired t-tests with Hedges' correction and leave-one-subject-out cross-validation. Results End-to-end latency: 38 ms [33–42]. Model: R²=0.72, Spearman ρ=0.68 (p<0.001); LOSOCV R²=0.68±0.09; ICC(2,1)=0.84. AAN reduced assist torque 62%→45%, increased VR difficulty +64%, improved success rate 61%→82%. Clinical outcomes: FMA-UE +9.1 [6.7, 11.5], d=0.98; ARAT +7.6, d=0.93; grip +4.1 kg. Moderate-to-severe patients showed greater gains (+10.7 vs +7.2, p<0.05). Ablation confirmed synergy: Bio-AI (R²=0.70, +7.2, compliance 68%) vs complete system (R²=0.72, +9.1, 88%). SUS 84±6; no adverse events. Discussion All hypotheses validated: multimodal fusion exceeded targets; clinical efficacy surpassed MCID (d=0.98), exceeding spontaneous recovery (2–4 points); ablation demonstrated ≥15% degradation, confirming synergistic effects. The system achieves paradigm shift from motor imagery to execution-based rehabilitation with direct motor intent capture. Limitations include single-arm design, small sample (n=24), and short duration (4 weeks). Future work requires RCTs with active controls and extended follow-up. The bio–AI–VR system demonstrates feasibility of data-driven rehabilitation for post-stroke hand recovery.