Contemporary Machine Learning Approaches Towards Biomechanical Analysis in the Diagnosis and Prognosis Prediction of Knee Osteoarthritis: A Systematic Review

Abstract

Introduction: Knee Osteoarthritis (KOA) is the second most reported condition for persons 50 years and up; approximated by the continuous degradation of the knee, and eventually extending to the debilitation of biomechanical gait parameters. Inconsistencies with existing diagnostic methods mean that Machine Learning (ML) has been leveraged in creating gait-based predictive models in relation to KOA. The purpose of this study is to explore existing literature with camera and sensor-based methodologies, along with the employed algorithms in the diagnosis and prognosis prediction of KOA.

Methods: Searches for literature were accomplished on Google Scholar and PUBMED databases using relevant keywords, within a time frame of 2010 - 2023. Information pertaining to the data collection method, algorithm used, and model performance was collected.

Results: After the initial search of 1132 articles, the selection process yielded 22 articles for further review. Of the 22 articles, 10% focused on the prediction of patient outcomes and disease prognosis, while 90% focused on the initial diagnosis or severity prediction of KOA. 28% of the reviewed literature utilized sensor-based technology for biomechanical gait parameter collection, while the remainder utilized a more traditional camera-based approach. While evaluatory metrics varied between studies, of the studies with reported accuracy metrics (n=11), camera-based models had on average a higher accuracy compared to sensor-based algorithms, 92.05% compared to 67.96%, respectively.

Discussion: Support Vector Machine (SVM) was found to be the most common algorithm used within the reviewed studies, and had the highest accuracy on average, possibly attributed to the ability of the algorithm to manage small yet high dimensional datasets. The difference in accuracy between camera-based and sensor-based approaches was determined to be statistically significant through application of a Mann-Whitney U Test. While sensors have a reduced quantity of features capable of being measured, it is a more applicable technology for clinical application, indicating an area for future development.

Conclusion: Overall, literature concerning the binary classification of symptomatic KOA provided high accuracy, yet further validation to minimize overfitting is required. Furthermore, areas for prognosis prediction and multiclass classification of KOA severity remain as areas for further development.