Junshuo Cui
Introduction Post-stroke hand motor impairment is a major cause of long-term functional disability and reduced quality of life, with approximately 70% of stroke survivors experiencing persistent limitations in fine motor control. Conventional rehabilitation is constrained by low adherence, subjective assessment, and insufficient individualization, which limits exploitation of the neuroplasticity window for motor relearning. To address these challenges, we propose a bio–AI–VR integrated hand rehabilitation system that fuses biosignal sensing (bio), AI-based analysis, and virtual reality (VR) interaction to realize an efficient, adaptive, and quantifiable closed-loop training process. The integration rationale is grounded in three theoretical pillars: (i) multimodal data fusion theory—combining heterogeneous biosignal and behavioral data through AI to overcome single-modality limitations; (ii) closed-loop adaptive control theory—dynamically balancing challenge and capability via real-time feedback; (iii) neuroplasticity multisensory enhancement theory—coordinating visual, proprioceptive, and motor pathways to strengthen cortical reorganization. This work addresses three testable hypotheses: (RQ1) Can multimodal biosignal fusion achieve real-time assessment with R 2 ≥ 0.65 and latency < 50 ms? (RQ2) Does bio-AI-VR integration yield FMA-UE improvement ≥ 6 points (minimal clinically important difference) with effect size d ≥ 0.8 ? (RQ3) Are all three components (bio, AI, VR) necessary, with ablation causing ≥ 15 % performance degradation? Methods A lightweight hand exoskeleton ( < 400 g, 3 DoF/finger) integrates a 6-axis IMU (100 Hz) and 16-channel sEMG (1 kHz) to synchronously acquire kinematics and muscle activation. Extended Kalman filtering fuses sensor streams before AI processing. Features include range of motion (ROM), smoothness metrics (SPARC, LDLJ), sEMG root-mean-square (RMS), median frequency (MDF), and co-contraction index (CCI). A hybrid model combining random forests (200 trees, depth 8) and support vector regression (RBF kernel, γ = 0.01 , C = 10 ) outputs a real-time composite score S t ∈ [ 0 , 1 ] via multi-task learning with GroupKFold cross-validation, mapped to clinical scales through Sigmoid normalization. FMA-UE proxy labels for window-level training were constructed via linear interpolation (80%), biomechanical anchoring (15%), and expert annotation (5%, inter-rater κ = 0.78 ). A cloud AI engine communicates bidirectionally with Unity-based VR over MQTT to close the perception-assessment-assistance loop. The assistance-as-needed (AAN) algorithm adjusts exoskeleton torque ( u t ) and VR difficulty ( d t ) using S t as control input with hysteresis, dead zone, and rate limiting to ensure smooth adaptation. Twenty-four stroke survivors (3–12 months post-stroke, FMA-UE 15–50) underwent 4-week training (5 sessions/week, 20 min/session). Outcomes included FMA-UE (primary), ARAT, grip strength, normalized ROM, task success rate, and System Usability Scale (SUS). Statistical analysis employed paired t -tests with Hedges’ correction for effect sizes, Bonferroni adjustment for multiple comparisons, and leave-one-subject-out cross-validation (LOSOCV) to assess model generalization. Results All 24 participants completed the study with one missed session (479 of 480 scheduled sessions, 8,946 annotated segments); end-to-end latency median 38 ms (IQR 33–42 ms), decomposed as: sampling 8 ± 2 ms, preprocessing <mml:math xmlns:m