A THIGH MUSCLE-INFORMED MULTI-TASK DEEP LEARNING SYSTEM FOR PREDICTING KNEE OSTEOARTHRITIS STRUCTURAL AND SYMPTOMATIC...

Knee osteoarthritis (OA) is a prevalent chronic joint disease whose progression imposes a substantial healthcare burden. Accurate individualized prediction is crucial for implementing targeted interventions. Thigh muscle quality is a key contributor to knee OA progression and can be improved through various interventions. Here, we developed MKOTPR (muscle-oriented knee osteoarthritis tailored progression risk), a multi-task deep-learning-based system that enables end-to-end analysis of raw thigh MRI and outputs probabilities of structural and symptomatic progression of knee OA.

A total of 4023 thigh MRI scans were collected from five clinical centers, forming training, validation, internal testing, and two external testing sets. Model performance was assessed using the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. Time-dependent receiver operating characteristic (ROC) analyses were conducted to evaluate performance at specific follow-up time points, and Kaplan-Meier curves were generated to illustrate outcomes between risk groups. To enhance interpretability, MKOTPR was correlated with established clinical risk factors. Its independent predictive value was evaluated in the clinical context. Grad-CAM was used to elucidate decision-relevant information captured by MKOTPR.

For predicting structural progression, MKOTPR yielded AUCs of 0.935, 0.908, and 0.935 in the internal and two external testing sets, respectively, and effectively stratified patients into low- and high-risk groups. For predicting symptomatic progression, MKOTPR achieved AUCs of 0.957, 0.945, and 0.960 across testing sets, with similarly effective risk stratification. The MKOTPR also exhibited favorable accuracy, sensitivity, specificity, PPV, NPV, and F1 scores for both outcomes. Time-dependent ROC analyses further demonstrated its good discriminative performance at years 1-4. The observed alignment between model predictions and known clinical factors confirmed MKOTPR's clinical relevance. MKOTPR emerged as an independent prognostic factor for knee OA progression. The Grad-CAM visualized important regions that informed MKOTPR in making predictions. Moreover, model performance remained robust in subgroup analyses.

By performing entire-thigh muscle analysis on MRI, MKOTPR simultaneously predicted structural and symptomatic progression of knee OA. Such a risk stratification tool may help guide muscle-targeted interventions to optimize patient outcomes.