X-ray based radiomics machine learning models for predicting collapse of early-stage osteonecrosis of femoral head

Abstract This study aimed to develop an X-ray radiomics model for predicting collapse of early-stage osteonecrosis of the femoral head (ONFH). A total of 87 patients (111 hips; training set: n = 67, test set: n = 44) with non-traumatic ONFH at Association Research Circulation Osseous (ARCO) stage II were retrospectively enrolled. Following data dimensionality reduction and feature selection, radiomics models were constructed based on anteroposterior (AP), frog-lateral (FL), and AP + FL combined view using random forest (RF), support vector machine (SVM), and stochastic gradient descent (SGD). After the optimal radiomics model was selected based on areas under the curve (AUC), its performance on the test set was compared with that of orthopaedists using receiver operating characteristic (ROC) curves and confusion matrices. Among all radiomics models, the SVM-based AP + FL combined view model (AP + FL-Rad_SVM) achieved the highest individual performance demonstrating an AUC of 0.904 (95% CI 0.829 –0.978) in the test set, which was significantly better than that of three attending surgeons ( p = 0.014, 0.004, and 0.045, respectively). The SVM model based on AP + FL views of hip X-ray exhibited excellent ability in predicting the collapse of ONFH and showed superior performance compared with less experienced orthopaedic surgeons. This model may inform clinical decision-making for early-stage ONFH.