5297. Explainable Radiomics Model to Predict Total Knee Replacement in Patients with Osteoarthritis: Data from Osteoarthritis Initiative (OAI)
Authors * Denotes Presenting Author
  1. Atefe Pooyan *; University of Washington
  2. Ehsan Alipour; University of Washington
  3. Arash Azhideh; University of Washington
  4. Matthew Nyflot; University of Washington
  5. Firoozeh Shomal Zadeh; University of Washington
  6. Negar Firoozeh; University of Washington
  7. Majid Chalian; University of Washington
Knee osteoarthritis (OA) is a debilitating disease affecting millions worldwide. It causes significant pain and loss of mobility and can necessitate knee replacement surgery, which is associated with potential complications and high healthcare costs. Therefore, early identification of high-risk patients is crucial for effectively managing severe osteoarthritis and improving the prognosis, including the need for knee replacement. We aimed to predict knee replacement based on the baseline MRI radiomics.

Materials and Methods:
We used a dataset of manually segmented knee structures in 507 patients (80% train, 20% test) with various grades of knee osteoarthritis. Knee replacement in the segmented knee within 9 years was the outcome of interest. Original radiomics features were retrieved from baseline knee magnetic resonance imaging (MRI) for four locations of interest (femoral bone, femoral cartilage, tibial bone, and tibial cartilage) using the Pyradiomics package. The Extreme Gradient Boosting (XGBoost) model was developed based on radiomics features, patient demographics (age, sex), whether they had basic insurance, family history of knee replacement in a first-degree relative, comorbidities (history of diabetes, body mass index (BMI) history of knee replacement in the other knee), history of statin, and frequent nonsteroidal anti-inflammatory drug (NSAID) use over 1 year, as well as WOMAC pain, disability, and stiffness score in the baseline visit. A random grid search with 10-fold cross-validation was used to select the best model parameters, and the model was calibrated using actual probabilities from the train set. The completed model was tested on the hold-out test set, and the feature relevance was determined using Shapley values. Data was obtained from the publicly available OAI-ZIB dataset, which contains manually segmented knee MRIs for a subset of Osteoarthritis Initiative (OAI) subjects.

The developed model achieved AUC of 82.1% in cross-validation and 71.3% in test predicting total knee replacement in patients with knee OA. Sensitivity and specificity were 77.7% and 70.2%, respectively. The final model included 27 features out of 413 radiomics, demographics, and clinical variables. Femoral bone sphericity, WOMAC pain score, and femoral cartilage first order skewness were the three most important features in the model.

We proposed an explainable multimodal radiomics model that can aid in early detection of patients with OA who are at risk for total knee replacement.