Predicting total knee arthroplasty in osteoarthritis: comparative performance of radiomics, clinical, and combined models
Author Block: A. Pooyan1, E. Alipour2, M. Nyflot2, M. Chalian2; 1Philadelphia, PA/US, 2Seattle, WA/US
Purpose: Knee osteoarthritis (OA) is a prevalent disabling disease. Total knee arthroplasty (TKA) treats end-stage OA. Early identification of patients at high risk for TKA may improve understanding of disease progression. This study aimed to find the features associated with TKA and compared the performance of clinical, radiomics, and combined models for predicting TKA in OA patients.
Methods or Background: We analyzed 507 knees from the OAI-ZIB dataset with manually segmented baseline MRIs. The outcome was TKA within 9 years in the segmented knee. Radiomics features (n=400) were extracted from femoral bone, femoral cartilage, tibial bone, and tibial cartilage using PyRadiomics. Clinical variables (n=13) included demographics, comorbidities, medication use, and baseline WOMAC scores. Univariate logistic regression with Benjamini-Hochberg correction identified significant features. Three model types: logistic regression, Least Absolute Shrinkage and Selection Operator (LASSO), and Extreme Gradient Boosting (XGBoost), were trained on three feature sets: radiomics only, clinical only, and combined. Performance was assessed by AUC-ROC, AUC-PR, and F1-score with 95% confidence intervals. Feature importance for XGBoost was evaluated using SHAP values.
Results or Findings: 45 radiomics features and three clinical variables (WOMAC pain, disability, stiffness) were significantly associated with TKA in univariate analysis. The combined XGBoost model performed best, achieving an AUC-ROC of 0.95 (95% CI: 0.92-0.97), AUC-PR of 0.72 (0.58-0.82), and F1 of 0.56 (0.46-0.64) in cross-validation, and an AUC-ROC of 0.72 (0.57-0.84), AUC-PR of 0.30 (0.16-0.47), and F1 of 0.48 (0.31-0.63) on the test set. Final model included 38 of 413 features, with femoral bone sphericity, WOMAC disability, and WOMAC pain as top predictors.
Conclusion: Combining radiomics with clinical features improves TKA prediction in OA, with XGBoost showing the strongest performance.
Limitations: Study was limited by using TKA as a surrogate outcome influenced by non-disease factors.
Funding for this study: No funding was received for this study.
Has your study been approved by an ethics committee? Not applicable
Ethics committee - additional information: