The application of multiclass SVM to the detection of knee pathologies using kinetic data: a preliminary study

Knee pathologies such as patellofemoral pain syndrome (PFPS) and knee osteoarthritis (OA) can lead to pain and disability. PFPS limits mobility and potentially leads to knee osteoarthritis. A common treatment for knee OA is joint replacement surgery. In this paper, we investigated the application of multiclass support vector machines (SVM) to classify gait patterns between three knee pathology groups using ground reaction force (GRF) data. Results indicate that the multiclass SVM could identify the different knee pathologies with a maximum leave one out (LOO) accuracy of 78%-88% on the testing set. When feature selection was applied, the accuracy improved to 85%-92% and accuracy on the test set improved from 37% to 62%. The SVM detected 7 and 8 GRF features related to the peak GRFs and their relative timing as being sensitive to distinguish between patients with knee replacement and both knee OA and PFPS, respectively. Only 2 features (peak anterior GRF and time to heel strike transient) were required to discriminate knee OA from PFPS group. The SVM classifier was able to effectively recognize gait parameters that were altered due to the various knee pathology conditions. This suggests that GRF information is indicative of abnormal joint loading and can be detected using the multiclass SVM.