Bavan L1, Surmacz K2, Mellon S1, Beard D1, Rees J1
1University of Oxford, Oxford, United Kingdom, 2McLaren Applied Technologies, Woking, United Kingdom
Background: Musculoskeletal disorders are prevalent amongst all age groups and are a major contributor to global disability. Exercise-based rehabilitation has an established role in the prevention and treatment of a wide range of musculoskeletal problems. In many cases, patients are required to undergo self-directed or unsupervised therapy. Currently there are no approved tools for objectively monitoring rehabilitation adherence or performance in the home setting, and commonly used self-reported measures lack validity and reliability. This has implications for clinicians involved in the assessment of patient recovery, as well as researchers attempting to accurately quantify and evaluate the benefits of exercise activity.
Purpose: The use of inertial sensors to remotely monitor patient activity is becoming increasingly popular for a range of clinical applications. This study aims to evaluate the performance of supervised machine learning techniques for recognition and classification of shoulder rehabilitation activity using inertial data acquired through a single wireless sensor.
Methods: Twenty patients with shoulder pain suitable for a course of physiotherapy were monitored performing five routinely prescribed rehabilitation exercises. Accelerometer, gyroscope and magnetometer data was collected via a small sensing device mounted onto an arm sleeve. Collected motion data was pre-processed, segmented and labelled. Time and frequency domain features were calculated from labelled data segments and ranked in terms of their predictive importance. Selected features were used to train four supervised learning algorithms: decision tree, ensemble decision tree, k-nearest neighbour and support vector machine. Performance of algorithms in accurately classifying exercise activity was evaluated using two distinct validation techniques: a conventional five-fold cross-validation and a subject stratified cross-validation method.
Results: A selection of both time and frequency domain features derived from accelerometer, gyroscope and magnetometer data were shown to have high predictive value for classifying shoulder rehabilitation activity. Five-fold cross-validation yielded predictive accuracies between 85% and 95% for all trained models. Subject stratified validation, evaluating model performance on unseen subject data, yielded predictive accuracies of up to 85%.
Conclusion(s): Trained classification models were capable of making accurate predictions on new data, accounting for the inter-subject and intra-subject variation in exercise pace and performance seen in real patients. This study demonstrates the feasibility of remote patient monitoring and automated classification of rehabilitation activity with wireless sensor technology.
Implications: The affordability and practicality of inertial based systems will make this form of monitoring attractive to both healthcare providers and patients. A clinically useful and reliable account of home rehabilitation activity will help guide treatment strategies and facilitate methods to improve patient engagement and performance. Tools to accurately and objectively classify movement will also support research aimed at evaluating specific rehabilitative approaches.
Keywords: Rehabilitation, inertial sensor, machine learning
Funding acknowledgements: This work was supported by NIHR Oxford Biomedical Research Centre and the Nuffield Department of Orthopaedics, Rheumatology and Musculoskeletal Sciences.
Purpose: The use of inertial sensors to remotely monitor patient activity is becoming increasingly popular for a range of clinical applications. This study aims to evaluate the performance of supervised machine learning techniques for recognition and classification of shoulder rehabilitation activity using inertial data acquired through a single wireless sensor.
Methods: Twenty patients with shoulder pain suitable for a course of physiotherapy were monitored performing five routinely prescribed rehabilitation exercises. Accelerometer, gyroscope and magnetometer data was collected via a small sensing device mounted onto an arm sleeve. Collected motion data was pre-processed, segmented and labelled. Time and frequency domain features were calculated from labelled data segments and ranked in terms of their predictive importance. Selected features were used to train four supervised learning algorithms: decision tree, ensemble decision tree, k-nearest neighbour and support vector machine. Performance of algorithms in accurately classifying exercise activity was evaluated using two distinct validation techniques: a conventional five-fold cross-validation and a subject stratified cross-validation method.
Results: A selection of both time and frequency domain features derived from accelerometer, gyroscope and magnetometer data were shown to have high predictive value for classifying shoulder rehabilitation activity. Five-fold cross-validation yielded predictive accuracies between 85% and 95% for all trained models. Subject stratified validation, evaluating model performance on unseen subject data, yielded predictive accuracies of up to 85%.
Conclusion(s): Trained classification models were capable of making accurate predictions on new data, accounting for the inter-subject and intra-subject variation in exercise pace and performance seen in real patients. This study demonstrates the feasibility of remote patient monitoring and automated classification of rehabilitation activity with wireless sensor technology.
Implications: The affordability and practicality of inertial based systems will make this form of monitoring attractive to both healthcare providers and patients. A clinically useful and reliable account of home rehabilitation activity will help guide treatment strategies and facilitate methods to improve patient engagement and performance. Tools to accurately and objectively classify movement will also support research aimed at evaluating specific rehabilitative approaches.
Keywords: Rehabilitation, inertial sensor, machine learning
Funding acknowledgements: This work was supported by NIHR Oxford Biomedical Research Centre and the Nuffield Department of Orthopaedics, Rheumatology and Musculoskeletal Sciences.
Topic: Musculoskeletal; Human movement analysis; Information management, technology & big data
Ethics approval required: Yes
Institution: Health Research Authority
Ethics committee: South West Central Bristol Research Ethics Committee
Ethics number: 17/SW/0217
All authors, affiliations and abstracts have been published as submitted.
