Wearable-based Frozen Shoulder Rehabilitation Exercise Recognition using Machine Learning Approaches

Abstract

Frozen shoulder is a disease that causes shoulder pain and stiffness. It limits the range of movement of the shoulder and has a great impact on the quality of daily life. One of the common treatment methods is to do frozen shoulder rehabilitation exercises. However, patients often fail to follow the instructions of the physical therapists during the program of home-based rehabilitation. Furthermore, clinical professionals are unavailable to track and monitor the home-based rehabilitation exercise performance of patients. To support clinical monitoring, we develop a wearable-based frozen shoulder rehabilitation exercise recognizer using different machine learning models and deep learning models. The proposed methods can automatically identify movement/silence segments from continuous signals and classify types of frozen shoulder rehabilitation exercises. Besides, we propose a finite state machine and fragmentation revision mechanism for error correction. Twenty subjects are invited to perform six types of rehabilitation exercises. The proposed methods achieve the best result of 95.6% accuracy, 95.83% F-score for the identification of movement/silence and 95.58% accuracy, 95.49% F-score for classification of exercise type, respectively. The results demonstrate the feasibility of the proposed method to automatically monitor the frozen shoulder rehabilitation exercise, which has the potential to provide objective, continuous and quantitative information for telerehabilitation.