Abstract
Abstract Background To address the problem of low recognition efficiency caused by redundant plantar pressure features in gait analysis of patients with knee osteoarthritis (KOA), this study proposes a key feature screening method based on the iterative forward feature selection (IFFS) algorithm to enable the auxiliary diagnosis of KOA. Methods This study collected vertical ground reaction force (vGRF) and center of pressure (COP) displacement data from 95 volunteers during natural walking, including 69 patients with KOA and 26 healthy controls, and constructed a multidimensional feature set comprising spatiotemporal parameters, nonlinear dynamic features, and bilateral symmetry indices. Subsequently, the IFFS algorithm was applied to remove redundant variables weakly associated with disease labels through statistical difference analysis, thereby preliminarily revealing gait-related statistical characteristics associated with KOA. By maximizing classification accuracy while suppressing feature redundancy, the most discriminative feature subset was selected, and machine learning was used to achieve KOA recognition. Results The experimental results showed that the optimal feature subset, consisting of the maximum Lyapunov exponent, Hurst exponent, and root mean square of bilateral vGRF, as well as the fractal dimension and standard deviation of COP in the anterior–posterior direction, achieved the best discriminative performance. Based on this feature subset, the random forest model achieved an accuracy of 89.47% and an F1-score of 95.71% in a single test. Conclusions The results indicate that the IFFS method can effectively reduce the dimensionality of plantar pressure features, providing a reference for the auxiliary identification of KOA based on plantar pressure gait features.