PREDICTING KNEE OSTEOARTHRITIS PROGRESSION USING EXPLAINABLE MACHINE LEARNING AND CLINICAL IMAGING DATA

INTRODUCTION

Accurate prediction of knee osteoarthritis (KOA) progression remains a clinical challenge due to its heterogeneous nature and discordance between structural and symptomatic outcomes. Integrated imaging and machine learning (ML) approaches may enhance prognostic modeling but often suffer from limited interpretability or reliance on static features.

OBJECTIVE

We aim to develop explainable ML models for predicting KOA progression using baseline and longitudinal imaging and clinical features. This study also aims to identify key imaging biomarkers associated with structural and symptomatic progression.

METHODS

Data and 3T MRI measurements from 600 participants in the FNIH OA Biomarkers Consortium were analyzed. Participants were grouped into four progression categories based on 48-month joint space narrowing and WOMAC pain: (1) radiographic + pain progressors, (2) radiographic-only, (3) pain-only, and (4) non-progressors. Two binary classification frameworks were defined: (1) radiographic + pain vs. all others (primary), and (2) all radiographic progressors vs. pain-only + non-progressors (secondary). ML models included Random Forest, XGBoost, logistic regression, decision tree, and multilayer perceptron (MLP). The model used demographic information and imaging features from semi-automated segmentation software. We measured the volume of medial compartment femur cartilage (Cart), bone marrow lesion (BML) in the MF, LF, MT, LT, patella, and trochlea, osteophytes (Ost) in the MF, LF, MT, and LT, Hoffa’s synovitis (HS), and effusion/synovitis (ES). Longitudinal delta values were computed over 24 months. Performance was assessed via 10-fold stratified cross-validation (AUC, F1-score). Explainability tools included SHAP, Gini importance, coefficients, and permutation importance.

RESULTS

In the cross-sectional setting, the Random Forest classifier achieved the highest discrimination performance, with AUC values of 0.672 for the primary task (radiographic + pain progressors vs. others) and 0.791 for the secondary task (all radiographic progressors vs. others). The MLP model showed similar results in the secondary task (AUC = 0.743). AUC performance metrics for all models are shown in Table 1. Model performance improved notably when incorporating 24-month changes in imaging features. In the longitudinal analysis, Random Forest again performed best in the secondary task (AUC = 0.873), followed by XGBoost and MLP. The strongest predictors in these models were changes in medial femoral cartilage thickness, medial tibial bone marrow lesions, and osteophyte scores. To better understand the basis of model predictions, we applied four feature ranking methods. Among them, the SHAP method produced the most consistent and clinically interpretable results. As an example, shown in Figure 1 which show top 15 important features, SHAP highlighted 24-month reductions in cartilage thickness and increases in bone marrow lesion scores as the most influential variables, especially in the medial compartment.

CONCLUSION

Explainable ML models can identify individuals at risk of KOA progression using multimodal data. Longitudinal imaging features enhanced predictive power, and transparent interpretation techniques revealed important markers of joint deterioration.