Comparison of Machine Learning approaches for Classifying Upper Extremity Tasks in Individuals Post-Stroke

After a stroke, individuals often exhibit upper extremity (UE) motor dysfunction, influencing the performance of everyday tasks. Characterizing UE movements is useful to track recovery and response to intervention. Yet, due to the complexity of the recovery process, UE movements may be extremely variable and person-specific. While this renders automatic recognition of these gestures challenging, machine learning methods could be used to classify UE movements in atypical populations. In the current study, we utilize data from 20 individuals post-stroke and 20 age-matched controls to identify an optimal set of sensor-extracted features for the classification of unimanual and bimanual gestures during task performance. We found that using fewer than 100 features along with a random forest classifier produced the best performance across both groups, with both user-dependent and user-independent models.