MINING THIGH MUSCLE INFORMATION FROM MRI USING ARTIFICIAL INTELLIGENCE TO PREDICT INCIDENT KNEE OSTEOARTHRITIS

Knee osteoarthritis (OA) is the most common chronic joint disease with a substantial public health burden, and early diagnosis of knee OA is crucial for timely intervention and effective disease management. Thigh muscle changes play an important role in knee OA pathogenesis and are potential targets for intervention, which makes thigh muscles a promising target for developing predictive models. Here we developed MKOTIR (muscle-oriented knee osteoarthritis tailored incidence risk), a fully automated artificial intelligence (AI) model based on thigh MRI to predict incident knee OA over 8 years.

A total of 4070 thigh MRI scans were collected from five clinical centers, constituting a training set, a validation set, an internal testing set, and two external testing sets. MKOTIR processes original MRI scans and directly outputs the predicted probability of developing knee OA over 8 years. Model performance was evaluated using the area under the curve (AUC), accuracy, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and F1 score. A clinical model and a radiomics model were also constructed as benchmarks for comparison. To enhance interpretability, we correlated MKOTIR with key clinical characteristics and classical radiomics features. Furthermore, the independent prognostic value of MKOTIR was evaluated in the clinical context.

MKOTIR outperformed the benchmark models and demonstrated excellent predictive performance in two external testing sets, with AUC values of 0.977 and 0.965, accuracy of 0.982 and 0.982, sensitivity of 0.911 and 0.902, specificity of 0.991 and 0.997, PPV of 0.935 and 0.982, and NPV of 0.988 and 0.982, respectively. The F1 scores were 0.923 and 0.940, further supporting its robust and well-balanced performance. MKOTIR prediction was significantly associated with some clinical characteristics, confirming its clinical relevance, and was independent of potential confounders. Although MKOTIR prediction correlated with certain radiomics features, the radiomics model alone failed to achieve comparable performance, indicating that MKOTIR learned additional imaging patterns beyond those captured by the conventional radiomics approach. Moreover, the performance of MKOTIR remained robust in subgroup analyses.

The application of AI to thigh MRI enables mining muscle information for predicting incident knee OA over 8 years. Our developed MKOTIR model demonstrates excellent performance and therefore holds promise for the early initiation of muscle-targeted therapies.