Multi-Modal Machine Learning Framework with Spatial Alignment of Quantitative MRI and Molecular Priors for Predicting Early Cartilage Degeneration in Osteoarthritis

To develop and validate a spatially informed multi-modal machine learning framework that integrates quantitative magnetic resonance imaging (qMRI) biomarkers with molecular pathway priors for early prediction of focal cartilage degeneration in knee osteoarthritis. This retrospective study utilized publicly available datasets. Quantitative MRI data from 620 subjects with early-stage knee osteoarthritis (Kellgren–Lawrence grades 0–2) were obtained from the Osteoarthritis Initiative, including T1ρ and T2 relaxation mapping, delayed gadolinium-enhanced MRI of cartilage (dGEMRIC), and cartilage thickness measurements with longitudinal follow-up of 12–36 months. Molecular pathway signatures representing extracellular matrix remodeling, inflammatory signaling, and chondrocyte stress were derived from transcriptomic datasets in the Gene Expression Omnibus and incorporated as population-level priors. Spatial alignment was performed at anatomically defined cartilage regions without assuming subject-level correspondence. A gradient boosting model was trained for classification and longitudinal prediction with post-hoc interpretability. Baseline imaging-only models were evaluated under identical conditions. External validation was performed using 430 subjects from the Multicenter Osteoarthritis Study. The framework achieved an accuracy of 86.8% (95% CI: 83.2–90.1), sensitivity of 88.1%, specificity of 85.2%, and AUC of 0.91 (95% CI: 0.88–0.94) for classification of early degenerative cartilage regions. For longitudinal prediction, the model achieved an AUC of 0.88 (95% CI: 0.84–0.92). Compared with imaging-only models, incorporation of molecular priors improved predictive performance, particularly in regions with subtle abnormalities. This approach may assist in identifying high-risk cartilage regions in clinical research settings. The proposed framework enables accurate prediction of early cartilage degeneration with post-hoc interpretability. Molecular priors are incorporated as population-level features and do not establish direct biological associations. The approach may support early risk stratification in knee osteoarthritis. • Spatial integration of quantitative MRI and molecular pathway priors for osteoarthritis prediction. • Region-specific quantitative MRI biomarkers enable early detection of focal cartilage degeneration. • Multi-modal fusion improves predictive performance compared with unimodal approaches. • External validation on the MOST cohort demonstrates model robustness and generalizability. • Interpretable imaging–molecular associations support biologically informed osteoarthritis assessment.