New interpretable AI framework predicts knee pain trajectories with 0.91 accuracy

Researchers led by Jincheng Yu (University of Arizona and collaborators) have published a paper on arXiv presenting an interpretable AI framework designed for large-scale longitudinal studies of knee structure-pain associations. The framework uses deep learning to predict MRI Osteoarthritis Knee Score (MOAKS) features directly from knee MRIs, but its key innovation is the integration of conformal prediction—a statistical method that quantifies prediction uncertainty. This allows the model to filter out low-confidence predictions, keeping only high-confidence MOAKS outputs for downstream analysis.

Using this uncertainty-aware filtering, the team expanded the sample size for statistical modeling to 2,175 knees from the Osteoarthritis Initiative (OAI). They then applied a longitudinal latent class mixed model (LCMM) to examine associations between three key structural abnormalities—bone marrow lesions (BML), cartilage loss (CART), and meniscal extrusion (ME)—and four complementary knee pain measurements. The results were striking: Matthews correlation coefficient (MCC) improved dramatically—BML from 0.69 to 0.91, CART from 0.45 to 0.80, ME from 0.59 to 0.89.

The LCMM analysis revealed two distinct pain trajectories: rapid and stable progression. The odds ratios for belonging to the rapid progression group were 1.62 for BML, 1.83 for cartilage loss, and 2.50 for meniscal extrusion (all statistically significant with 95% confidence intervals). This quantifies how much each structural abnormality raises the risk of faster pain worsening in osteoarthritis patients. The framework's interpretability—combining deep learning predictions with classical statistical modeling—makes it particularly valuable for clinical decision-making, as clinicians can understand which specific abnormalities are driving pain progression over time.