Clustering and classification of early knee osteoarthritis using machine-learning analysis of step-up and down test kinematics in recreational table tennis players

Abstract

Objective: Early detection of knee osteoarthritis is crucial for improving patient outcomes. While conventional imaging methods often fail to detect early changes and require specialized expertise for interpretation, this study aimed to investigate the use of frontal plane kinematic data during step-up (SU) and step-down (SD) tests to classify and predict early osteoarthritis (EOA) using machine-learning techniques. Methods: Forty-three recreational table tennis players (eighty-six legs: 42 with EOA and 44 without EOA) underwent SU and SD tests. Frontal plane kinematics was analyzed using two-dimensional video analysis with markers placed at five key anatomical landmarks. Horizontal displacement measurements were compared between groups using independent t-tests. Unsupervised learning (Louvain clustering) was used to identify distinct movement patterns, whereas supervised learning algorithms were employed to classify EOA status. The feature importance was assessed using feature permutation importance (FPI). Results: Significant differences were observed between EOA and non-EOA groups in frontal plane kinematics during SU and SD tests (p < 0.001 for most variables). Louvain clustering identified four distinct kinematic profiles with varying proportions of EOA (ranging from 41.2% to 70.7%). Supervised learning models achieved high performance in classifying EOA status, with Random Forest, gradient boosting, and decision tree algorithms achieving 100% classification accuracy (AUC = 1.000) on the test dataset. FPI consistently highlighted the horizontal displacements of the ankle and femur during SU and of the pelvis and femur during SD as the most influential predictors. Conclusions: Machine-learning analysis of frontal plane kinematics during SU and SD tests showed promising potential for EOA detection and classification, offering a cost-effective and accessible alternative to conventional imaging-based approaches.