Machine learning-based study of the predictors of clinically important change in patient-reported outcomes in bilateral upper-limb function in patients receiving robotic stroke rehabilitation.

Introduction: Robotic therapy is an effective approach for poststroke rehabilitation of upper-limb function. There is a need for research on predictors of clinically meaningful change in patient-reported outcome measures of bilateral arm function relevant for daily life function in stroke patients receiving robotic therapy. Methods: This secondary analysis included data from 123 participants who received robotic therapy. We constructed machine learning classification models to predict the achievement of meaningful recovery of patient-perceived outcome of bilateral upper limb function based on the ABILHAND questionnaire. Clinically meaningful recovery was defined using the minimal clinically important difference (MCID). The prediction models included 14 potential predictors of three categories: demographic characteristics, stroke characteristics, and baseline assessment scores. The prediction models were built using four algorithms: logistic regression, k-nearest neighbors, support vector machine, and the random forest. Full models were built with all 14 potential predictors, and parsimonious models were built with the most important predictors identified by feature selection. Results: The prediction accuracy of the best-performing full models was 0.76 for the anchor-based MCID and 0.80 for the distribution-based MCID; the corresponding areas under the receiver operating characteristic curve were 0.75 and 0.84, respectively. The parsimonious models performed worse overall. For both MCID values, the most important predictors of recovery were time since stroke, the Wolf Motor Function Test-Time score, and the Stroke Impact Scale-Physical function score. When recovery was defined by the anchor-based MCID, the baseline ABILHAND and Chedoke Arm and Hand Activity Inventory scores were among the most important predictors. When recovery was defined by the distribution-based MCID, sex and stroke diagnosis were predictive of recovery. Conclusions: The time post onset of stroke, the speed of performing functional tasks, and self-perceived physical function were identified as the most important predictors of improvements in self-perceived function of bilateral upper limbs in daily activities after robotic therapy. The findings may inform clinicians about characteristics of patients with stroke who are more likely to benefit from robotic therapy.