Clinical Features Outperform MRI Radiomics for Predicting Intra-Articular Hyaluronic Acid Response in Knee...

Intra-articular hyaluronic acid (IAHA) injections are used for knee osteoarthritis, but clinical response is highly variable. This study evaluated whether quantitative magnetic resonance imaging (MRI) radiomics provides incremental predictive value over clinical variables and assessed the impact of cross-validation strategies on performance. Using data from the Osteoarthritis Initiative (OAI), we identified 262 participants (329 knees) treated with IAHA, of whom 106 participants (128 knees) had baseline 3D double-echo steady-state (DESS) MRI suitable for automated segmentation and radiomics extraction. We compared clinical-only, radiomics-only, and combined models. Under rigorous participant-level grouped cross-validation, the clinical-only model achieved the highest performance (ROC-AUC: 0.716; 95% CI: 0.689-0.742 in the full cohort; ROC-AUC: 0.718 in the MRI subset). The radiomics-only model failed to predict response (ROC-AUC: 0.499; 95% CI: 0.446-0.549), and the combined model (ROC-AUC: 0.693; 95% CI: 0.646-0.736) did not improve upon clinical predictors alone. A knee-level sensitivity analysis yielded higher but artificially inflated estimates (combined AUC: 0.773), confirming the necessity of participant-level grouping. We conclude that routine clinical variables, principally baseline pain and age, outperform DESS-MRI radiomics for predicting IAHA response. This study demonstrates that improper validation strategies significantly overestimate radiomics utility, highlighting the critical necessity of participant-level grouping in musculoskeletal imaging research to prevent data leakage.