Development and External Validation of a Motor Intention–Integrated Prediction Model for Upper Extremity Motor Recovery After Intention-Driven Robotic Hand Training for Chronic Stroke

Abstract

ObjectiveTo derive and validate a prediction model for minimal clinically important differences (MCIDs) in upper extremity (UE) motor function after intention-driven robotic hand training using residual voluntary electromyography (EMG) signals from affected UE. DesignA prospective longitudinal multicenter cohort study. We collected preintervention candidate predictors: demographics, clinical characteristics, Fugl-Meyer assessment of UE (FMAUE), Action Research Arm Test scores, and motor intention of flexor digitorum and extensor digitorum (ED) measured by EMG during maximal voluntary contraction (MVC). For EMG measures, recognizing challenges for stroke survivors to move paralyzed hand, peak signals were extracted from 8 time windows during MVC-EMG (0.1-5s) to identify subjects’ motor intention. Classification and regression tree algorithm was employed to predict survivors with MCID of FMAUE. Relationship between predictors and motor improvements was further investigated. SettingNine rehabilitation centers. ParticipantsChronic stroke survivors (N=131), including 87 for derivation sample, and 44 for validation sample. InterventionsAll participants underwent 20-session robotic hand training (40min/session, 3-5sessions/wk). Main Outcome MeasuresPrediction efficacies of models were assessed by area under the receiver operating characteristic curve (AUC). The best effective model was final model and validated using AUC and overall accuracy. ResultsThe best model comprised FMAUE (cutoff score, 46) and peak activity of ED from 1-second MVC-EMG (MVC-EMG 4.604 times higher than resting EMG), which demonstrated significantly higher prediction accuracy (AUC, 0.807) than other time windows or solely using clinical scores (AUC, 0.595). In external validation, this model displayed robust prediction (AUC, 0.916). Significant quadratic relationship was observed between ED-EMG and FMAUE increases. ConclusionsThis study presents a prediction model for intention-driven robotic hand training in chronic stroke survivors. It highlights significance of capturing motor intention through 1-second EMG window as a predictor for MCID improvement in UE motor function after 20-session robotic training. Survivors in 2 conditions showed high percentage of clinical motor improvement: moderate-to-high motor intention and low-to-moderate function; as well as high intention and high function.