Detecting impaired movements of stroke patients in bimanual training from motion sensor data

Stroke-induced hemiparesis is associated with loss of mobility and independence, preventing survivors from participation in activities of daily living. Survivors can recover motor function of their paretic limbs by adhering to a rehabilitation regimen, consisting of repetitive, high-intensity exercises. In telerehabilitation, information and communication technologies are leveraged to deliver such physical therapy to patients’ homes. However, monitoring motor performance remotely remains a challenging task, especially in light of high variability of motor impairments among patients. In order to evaluate motor performance, therapists require the aid of technicians, who would analyze sensor data and produce meaningful metrics. The therapists would then provide patients with feedback, after a few days at best. To automate this process and offer patients real-time feedback, we propose to train machine learning algorithms that detect impaired movements. We test this approach with ten healthy participants who interact with a low-cost telerehabilitation platform we have previously developed. The platform engages users in bimanual training, where movement of the affected arm is supported by the unaffected arm, and relies on a Microsoft Kinect sensor to record user movement. We report the accuracy of a classification algorithm in distinguishing movements of simulated of disability from normal ones. This effort constitutes a significant step toward programmed assessment of upper-limb movements in authentic telerehabilitation paradigms.