Informatics and Data Science

Standardized Machine-Learning Prediction of Knee Osteoarthritis Progression: Comparative Value of Traditional and Radiomics MRI Data Over Clinical and Biospecimen Predictors

Location: Concourse Foyer Kiosk

First Author(s)

NE

Nasim Eshraghi, MD

Postdoctoral Scholar
University of Washington, United States

Co-Author(s)

PM

Peyman Mirghaderi, MD, MPH

Postdoctoral Scholar
University of Washington
Seattle, Washington, United States
FP

Farhad Pishgar, MD

Resident R4
University of Washington, United States
Ag

Ali guermazi, MD PhD

Professor
Boston University, United States
MC

Majid Chalian, MD

Associate Professor
University of Washington, United States

Purpose: To benchmark standardized, reproducible machine-learning (ML) models for predicting radiographic knee osteoarthritis (OA) progression and to quantify the incremental value of the MRI Osteoarthritis Knee Score (MOAKS) and MRI radiomics beyond routinely available clinical variables and biospecimen biomarkers.

Methods/Materials: OAI data were analyzed to predict progression (≥0.7 mm loss in minimum medial tibiofemoral JSW from baseline to 24 months). Four models were evaluated: (1) clinical (12 variables) + biospecimens (18 biomarkers); (2) Model 1 + MOAKS (34 features); (3) Model 1 + MRI radiomics (398 features from tibial/femoral bone and cartilage); (4) all predictors combined. LASSO logistic regression and XGBoost were assessed using stratified 10-fold cross-validation with nested hyperparameter tuning. Performance was summarized by pooled out-of-fold AUC; incremental AUC differences were tested via bootstrap resampling.

Results: The baseline clinical-plus-biospecimen model showed modest discrimination (AUC 0.60). Adding MOAKS improved performance to AUC 0.72 (ΔAUC +0.12; p< 0.0001), and adding radiomics improved performance to AUC 0.67 (ΔAUC +0.08; p< 0.05). The combined model achieved the highest performance (AUC 0.74 for both algorithms; ΔAUC +0.14 vs baseline; p< 0.0001). MOAKS-augmented and radiomics-augmented models did not differ significantly, whereas integrating both feature sets yielded the best overall discrimination.

Conclusions: In standardized machine-learning models, MOAKS and MRI radiomics features provided potentially informative markers for predicting radiographic knee OA progression. Incremental gains beyond clinical variables and biochemical biomarkers may enable earlier identification of high-risk individuals and support future efforts to target preventive interventions, reduce progression, and advance individualized OA management.