Predicting Knee osteoarthritis progression using explainable machine learning and clinical imaging data.
Authors
Affiliations (5)
Affiliations (5)
- Department of Radiology, Mass General Brigham, Harvard Medical School, Boston, MA, USA.
- Harvard Data Science Initiative, Harvard University, Cambridge, MA, USA.
- Quantitative Musculoskeletal Imaging Group, Department of Radiology, Mass General Brigham, Boston, MA, USA.
- Department of Radiology, Brigham and Women's Hospital, Mass General Brigham, Harvard Medical School, Boston, MA, USA.
- Department of Orthopaedic Surgery, Brigham and Women's Hospital, Mass General Brigham, Harvard Medical School, Boston, MA, USA.
Abstract
To evaluate explainable AI models for predicting Knee osteoarthritis (KOA) progression using quantitative MRI and clinical data, for two outcomes: a composite endpoint (radiographic plus pain progression) and a radiographic-only progression endpoint. We analyzed 600 participants from the Foundation for the National Institutes of Health Osteoarthritis Biomarkers Consortium (FNIH), part of case-control cohort in the Osteoarthritis Initiative (OAI). The participants were grouped as composite progressors (n = 194), radiographic-only (n = 103), pain-only (n = 103), and non-progressors (n = 200). Input features included demographic data, Kellgren-Lawrence grade, joint space width, WOMAC pain score, and quantitative volume MRI measurements of KOA-related, including cartilage, bone marrow lesions, osteophytes, effusion-synovitis (ES), and Hoffa's synovitis (HS), at baseline and as 24-month change. Data were split into stratified 80% training and 20% held-out test sets, with 10-fold cross-validation used for model tuning within the training set. Five classifiers (random forest, XGBoost, logistic regression, decision tree, and multilayer perceptron). We applied multiple explainability methods, including Gini importance, SHAP values, regression coefficients, and permutation importance. Radiographic progression was most accurately predicted using longitudinal change features (random forest AUC = 0.87). Baseline features alone also yielded strong performance (AUC = 0.80). Composite progression was more difficult to predict (AUCs = 0.66-0.70). Across models, key expandability factors included medial femoral cartilage loss, BMLs in the medial tibia and femur, medial osteophyte, and ES volumes. Explainable machine learning using quantitative MRI enable interpretable prediction of KOA progression. This is the first study in the FNIH/OAI cohort to integrate longitudinal quantitative MRI features with model-agnostic explanations across multiple classifiers.