Related Experiment Video
Updated: Aug 21, 2026
Machine Learning Algorithms for Early Detection of Bone Metastases in an Experimental Rat Model
Published on: August 16, 2020
COMBINING MOAKS AND MRI RADIOMICS IMPROVES MACHINE-LEARNING PREDICTION OF RADIOGRAPHIC KNEE OA PROGRESSION: DATA FROM
N Eshraghi1, P Mirghaderi1, Y Xu1
1Department of Radiology, University of Washington, Seattle, WA, USA.
Introduction:
Radiographic knee OA progression is heterogeneous and may not be well predicted by clinical and biospecimen biomarkers alone. Baseline MRI-derived features, including expert semiquantitative MOAKS and quantitative radiomics, may capture complementary whole-joint structural information relevant to progression risk.
Objective:
To develop and compare multimodal machine-learning models for predicting 24-month radiographic knee OA progression using clinical variables, biospecimen biomarkers, semiquantitative MRI features, and MRI radiomics from the FNIH Osteoarthritis Biomarkers Consortium.
Methods:
This nested case-control analysis included 600 OAI participants (600 knees) with baseline KL grade 1-3, baseline knee MRI, biospecimen biomarkers, and baseline and 24-month radiographs. Radiographic progression was defined as a decrease of at least 0.7 mm in minimum medial tibiofemoral JSW from baseline to 24 months. Baseline predictors included 12 clinical variables, 18 biospecimen biomarkers (Serum/urine), 78 MOAKS features, and 400 radiomics features extracted from femoral/tibial cartilage and bone on sagittal 3D DESS MRI after automated segmentation. Four feature sets were tested: BioClinical (M1), MOAKS (M2), Radiomics (M3), and Combined (M4). Logistic regression, XGBoost, random forest, and SVM were evaluated using nested cross-validation with pooled out-of-fold AUC, sensitivity, and specificity. AUC differences were tested using bootstrap resampling.
Results:
Among 600 participants, 297 (49.5%) developed radiographic progression; the BioClinical model showed modest discrimination (AUC 0.57-0.59). Adding MOAKS features improved AUC to 0.71-0.74, and adding radiomics improved AUC to 0.69-0.71; both outperformed BioClinical models across all algorithms (all p<0.0001). MOAKS and Radiomics models did not significantly differ from each other (all p>0.05). The Combined model achieved the best overall performance (AUC 0.75-0.77), with XGBoost and SVM reaching AUC 0.77. Combined models improved sensitivity/specificity for logistic regression from 0.64/0.50 to 0.70/0.71 and for SVM from 0.49/0.57 to 0.70/0.73. Decision curve analysis showed that Combined models provided the highest or near-highest net benefit across most threshold probabilities.
Conclusion:
Baseline MRI-derived features substantially improved machine-learning prediction of radiographic knee OA progression beyond clinical and biospecimen biomarkers alone. MOAKS and radiomics provided broadly comparable predictive performance, while their integration achieved the best overall discrimination, supporting complementary roles for expert semiquantitative scoring and automated quantitative MRI radiomics in OA progression risk stratification.
