Classification of Shoulder Rehabilitation Exercises by Using Wearable Systems and Machine Learning Algorithms

• Published in: IEEE sensors journal Vol. 24; no. 14; pp. 22934 - 22944
• Main Authors: Sassi, Martina, Carnevale, Arianna, Mancuso, Matilde, Schena, Emiliano, Pecchia, Leandro, Giuseppe Longo, Umile
• Format: Journal Article
• Language: English
• Published: New York IEEE 15.07.2024; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Acceleration; Activity recognition; Algorithms; Angular velocity; Classification; Classification algorithms; Data acquisition; Decision trees; Effectiveness; Feasibility studies; Feature recognition; inertial measurement unit; Inertial sensing devices; Machine learning; machine learning (ML); Performance evaluation; Performance prediction; Protocols; Rehabilitation; rehabilitation exercises; Remote sensors; Sensor phenomena and characterization; Sensors; Shoulder; Support vector machines; Task analysis; Wearable sensors; Wearable technology
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2024.3406138

Shoulder rehabilitation is considered one of the most effective treatments for restoring functional abilities, reducing shoulder pain, and enabling the leading of an active life, improving mobility, strength, and endurance. However, the burdens of travel and time may prevent patients from taking part in such rehabilitation programs. The increased availability of wearable sensors and the development of machine learning (ML) algorithm has shown the feasibility of remote home-based rehabilitation therapy. In this study, we proposed a wearable system based on three magneto-inertial sensors to classify shoulder rehabilitation exercises. The classification has been performed by five different supervised ML algorithms [i.e., k-nearest neighbors (k-NNs), support vector machine (SVM), Naïve Bayes (NB), decision tree (DT), and random forest (RF)] to find out the most performant one. The feasibility of the wearable system was assessed on 19 healthy subjects during six rehabilitation exercises. Each exercise was performed six times, for a total of 684 samples. The data were analyzed and classified using the five mentioned classification models. Performances of the algorithms in accurately classifying exercise activity were evaluated with the k-fold cross-validation method and the nested validation method. The results demonstrated the effectiveness of the proposed algorithms in recognizing all the exercises. Features derived from acceleration, angular velocity, and orientation data were shown to reach the optimal predictive accuracies. Future work should focus on evaluating the performance of such systems on data acquired on patients with musculoskeletal disorders and on the inclusion of more shoulder rehabilitation exercises in the protocol.