Predicting Robot-Based Tracking Performance After Stroke Using Lesion Features and Clinical Assessments: A Pilot Study

Background: Following stroke, motor and cognitive impairments often limit functional independence. Personalized rehabilitation requires comprehensive clinical assessments, which are costly and time-consuming. Imaging scans are routinely collected, but their predictive value in determining post-stroke function remains unclear. Objectives: To determine the ability of clinical assessments, lesion features with medical chart information, and a combined clinical and biomedical model to predict kinematic outcomes from a robot-based tracking task in chronic stroke individuals. Methods: Eighteen participants (70+/-6 years) completed a robot-based tracking task and a battery of motor and cognitive assessments. Three elastic net regression models were developed: 1) clinical assessments only; 2) lesion metrics and chart information only; and 3) a combined model. Models were evaluated using leave-one-subject-out cross-validation, bootstrapping, and permutation testing with false discovery rate correction. Robot-based outcomes included tracking error, normalized distance, velocity error response, and delay. Results: The clinical assessment model significantly predicted tracking error (R^2=0.45), delay (R^2=0.22), velocity error response (R^2=0.18), and normalized distance (R^2=0.12). The combined model significantly predicted tracking error (R^2=0.37), delay (R^2=0.19), and velocity error response (R^2=0.11). The imaging model did not significantly predict any outcome. Box and Blocks Test and Color Trails 1 were stable features. The right dorsal attention network was a stable predictor in several combined models. Conclusion: Clinical assessments outperformed lesion features in predicting tracking performance after chronic stroke. Attention- and motor-related measures were the strongest predictors, highlighting the importance of cognitive-motor interactions in function. A larger sample and more granular lesion features may improve prediction from imaging biomarkers.