Learning classifier to evaluate movement quality in unassisted pick-and-place exercises for post-stroke patients: A preliminary study

The use of service robots is gaining increasing interest in stroke rehabilitation research. In autonomous robot therapy sessions, it is important to monitor the qualitative motor performance of the patient and to evaluate if therapeutically desirable movements are reinforced as per the therapist's instructions. Research on computational means to address this issue has received little attention, despite the acknowledged importance thereof. This paper poses the question of whether the subjective evaluation of the therapist on the qualitative motor performance of the individual patient can be modelled. In an attempt to answer this question, we frame it as a supervised machine learning problem and investigate if kinematic sensor readings to the subjective evaluation of the therapist can be correlated. We further examine if the personalized classifier trained with the data of the individual patient should be favored over the general classifier. Our preliminary results suggest that, given kinematic observation, trained classifiers may be able to imitate the evaluation of the therapist for the movement quality of the post-stroke patient. In our experiments, the general classifier outperformed the personalized classifiers, however, further investigation is necessary to generalize this finding.