Analysis of Detecting Compensation for Robotic Stroke Rehabilitation Therapy using Imbalanced Learning and Outlier Detection

Stroke therapy is essential to reduce impairments and improve the motor movements of stroke survivors, however sessions can be expensive, time consuming, and geographically limited. Robotic stroke therapy seeks to remedy the limitations of traditional stroke therapy, but it is hampered by incorrect movements during the session. Incorrect usage of muscles, called as compensatory movements, can cause problems that can hamper the recovery of the patients. Thus, there is a need to develop tools to automatically detect compensatory movements to assist patients doing autonomous therapy sessions. Previous studies on automatic detection using depth sensors did not yield satisfactory results. This study explores class imbalance as a possible reason for the low F1-score results on machine learning classifiers. Different methods to address class imbalance were employed to improve and to analyze the performance of the classifiers. The methods employed allowed the classifiers to sometimes detect compensatory movements however this degraded the performance of detecting the correct movements. Adjusting the decision thresholds of outlier detection algorithms shows this explicitly. Since addressing class imbalance only marginally improves the performance of the classifiers, other possible methods can be explored in conjunction with it. This study shows the possibility of detecting compensations in stroke patients.