Adherence monitoring of rehabilitation exercise with inertial sensors: A clinical validation study

• Published in: Gait & posture Vol. 70; no. NA; pp. 211 - 217
• Main Authors: Bavan, Luckshman, Surmacz, Karl, Beard, David, Mellon, Stephen, Rees, Jonathan
• Format: Journal Article
• Language: English
• Published: England Elsevier B.V 01.05.2019
• Subjects: Activity classification; Inertial sensor; Rehabilitation; Shoulder; Inertial sensor; Rehabilitation; Shoulder; Activity classification
• ISSN: 0966-6362; 1879-2219; 1879-2219
• DOI: 10.1016/j.gaitpost.2019.03.008

•There is a lack of objective tools to measure compliance with home rehabilitation.•Inertial sensors offer a practical approach to monitoring exercise activity.•Machine learning algorithms can classify exercise activity with sensor data.•Trained classifiers predict exercise type with up to 97.2% accuracy. Rehabilitation has an established role in the management of a wide range of musculoskeletal conditions. Much of this treatment relies on self-directed exercises at home, where adherence of execution is unknown. Demonstrating treatment fidelity is necessary to draw conclusions about the efficacy of rehabilitation interventions in both clinical and research settings. There is a lack of tools and methods to achieve this. This study aims to evaluate the feasibility of using a single inertial sensor to recognise and classify shoulder rehabilitation activity using supervised machine learning techniques. Twenty patients with shoulder pain were monitored performing five rehabilitation exercises routinely prescribed for their condition. Accelerometer, gyroscope and magnetometer data were collected via a device mounted onto an arm sleeve. Non-specific motion data was included in the analysis. Time and frequency domain features were calculated from labelled data segments and ranked in terms of their predictive importance using the ReliefF algorithm. Selected features were used to train four supervised learning algorithms: decision tree, k-nearest neighbour, support vector machine and random forests. Performance of algorithms in accurately classifying exercise activity was evaluated with ten-fold cross-validation and leave-one-subject-out-validation methods. Optimal predictive accuracies for ten-fold cross-validation (97.2%) and leave-one-subject-out-validation (80.5%) were achieved by support vector machine and random forests algorithms, respectively. Time domain features derived from accelerometer, magnetometer and orientation data streams were shown to have the highest predictive value for classifying rehabilitation activity. Classification models performed well in differentiating patient exercise activity from non-specific movement and identifying specific exercise type using inertial sensor data. A clinically useful account of home rehabilitation activity will help guide treatment strategies and facilitate methods to improve patient engagement. Future work should focus on evaluating the performance of such systems in natural and unsupervised settings.