From Model to App: Gender-Specific Bayesian-Optimized Support Vector Machine Predicts Early Structural Progression in Knee Osteoarthritis

Background Knee osteoarthritis (OA) is a leading cause of disability, yet predicting its structural progression remains challenging. We aimed to improve our developed gender-specific machine learning model by integrating Bayesian Optimization within a Support Vector Machine (SVMBO) to enhance early prediction of OA structural progression. Methods Building on our previous gender-specific MATLAB SVM model, which included age, BMI, CRP/MCP-1, and leptin/CRP, SVMBO models were trained on 677 participants (364 men, 313 women). External validation comprised 44 patients (20 men, 24 women). Gender-specific models were compared with all-gender models using accuracy, F1-score, and AUC, as well as confusion matrices for the classification of progressors. Predictor influence was evaluated using SHAP values. A standalone, open-access application with user documentation was developed. Results Gender-specific SVMBO models achieved higher performance than the original SVM across all metrics and phases. In external validation, both gender models achieved ≥95% accuracy and F1-score, with high AUCs ≥99%, and confusion matrices showed ≥94% correct classification of progressors. The men-specific model consistently outperformed the all-gender model, whereas the women-specific model showed superior performance in external validation. SHAP analysis identified age and BMI as dominant predictors, with age contributing more in men and BMI in women, whereas CRP/MCP-1 and leptin/CRP refined risk estimation. The gender-specific standalone SVMBO application, which does not require a MATLAB license or account, is publicly available via GitHub. Conclusion Gender-specific SVMBO models provide robust and generalizable predictions of knee OA structural progression. The accompanying open-source software enables accessible implementation for clinical risk stratification, with extensibility to broader musculoskeletal prognostic applications.