Automated Staging of Knee Osteoarthritis Using Radiographic Image Features and SVM Algorithm: A Clinical Decision Support Approach

Abstract

Background: Knee osteoarthritis is one of the most prevalent joint disorders in the elderly, and accurate classification of its severity plays a critical role in therapeutic decision-making. Objectives: This study aimed to develop an automated classification model for assessing knee osteoarthritis severity using radiographic images and clinical features, based on the Support Vector Machine (SVM) algorithm. Methods: In this applied, retrospective, and observational research, 44 radiographic images of the left knee from patients aged 39 to 72 were collected from the radiology department of Imam Ali Hospital in Bojnourd. Four key clinical features — namely, the angle between the femoral and tibial axes, the joint space width (JSW) ratio, the extent of subarticular erosion, and osteophyte structure — were extracted from the images. All features were normalized and evaluated using SVM models with both linear and nonlinear kernels. Model performance was assessed using k-fold cross-validation and analyzed through classification accuracy, sensitivity, and specificity. Osteoarthritis severity was determined using the Kellgren-Lawrence (KL) grading system, as assessed by an orthopedic specialist. Results: The classification accuracy using all features and the radial kernel reached 79.89%. With the radial basis function (RBF) kernel at σ = 0.85, the highest accuracy of 83.53% was achieved. The femur-tibia angle feature alone yielded a reasonably high performance [74.14% with the multilayer perceptron (MLP)], while the osteophyte feature resulted in the lowest classification accuracy (59.22%). Comparative chart analyses revealed that nonlinear kernels had superior discriminatory power compared to linear kernels. Conclusions: The proposed SVM-based model, utilizing interpretable structural features, successfully classified the severity of knee osteoarthritis with acceptable accuracy. The achieved classification accuracy (~84%) suggests potential clinical utility, although direct comparison with human expert performance was not conducted. This approach is recommended as a diagnostic support system, particularly in resource-limited clinical settings. Future research can enhance the model's generalizability and accuracy by incorporating additional clinical data and multi-source imaging modalities.