MRI-based machine learning model for predicting early relapse in osteosarcoma following neoadjuvant chemotherapy.

11519 Background: Early relapse (ER, ≤ 1 year) in osteosarcoma frequently occurs following neoadjuvant chemotherapy (NAC) and surgical resection. This study developed the machine learning models integrating radiomics, clinical, and pathological features to predict ER risk in osteosarcoma. Methods: 142 osteosarcoma patients were retrospectively analyzed and preoperative MRI images (T1- and FST2-weighted) were obtained after NAC. After feature extraction and selection, 5 machine learning classifiers—random forest (RF), support vector machine, logistic regression, decision tree, and gradient boosting tree—were implemented to construct radiomics, clinicopathological, and multimodal models. The performance of these models were assessed and compared using receiver operating characteristic curves, decision curve analysis (DCA), and Kaplan-Meier survival analysis, with the components and structure of the best-performing model subsequently visualized. Results: The RF algorithm outperformed the other classifiers, forming optimal radiomics and multimodal models. The RF-based multimodal model, combining 14 radiomics features, alkaline phosphatase (ALP) levels, and tumor necrosis rate, achieved the highest performance, with an area under the curve (AUC) of 0.978 in the training cohort and 0.913 in the testing cohort. The corresponding radiomics model, designed for real-time preoperative evaluation, showed a slight reduction in performance but still performed well. DCA and Kaplan-Meier curves indicated significant clinical utility of these two models. Conclusions: The RF-based pipeline, which includes radiomics and multimodal models, could facilitate personalized chemotherapy by identifying high-risk patients, optimizing treatment decisions, and improving outcomes. Performance of the different radiomics models, clinical models and combined models in the training and testing cohorts. Cohorts Model Classifier AUC (95% CI) Training cohort Radiomics models RF 0.963 (0.928-0.998) SVM 0.811 (0.723-0.899) LR 0.802 (0.714-0.891) DT 0.876 (0.809-0.942) GBT 0.973 (0.942-0.999) Clinical models LR 0.709 (0.614-0.803) Combined models RF 0.978 (0.956-0.999) SVM 0.898 (0.836-0.960) LR 0.813 (0.728-0.898) DT 0.896 (0.834-0.958) GBT 0.931 (0.884-0.977) Testing cohort Radiomics models RF 0.857 (0.751-0.963) SVM 0.784 (0.650-0.917) LR 0.794 (0.662-0.927) DT 0.814 (0.686-0.942) GBT 0.728 (0.561-0.896) Clinical models LR 0.684 (0.525-0.843) Combined models RF 0.913 (0.833-0.994) SVM 0.818 (0.671-0.965) LR 0.812 (0.659-0.965) DT 0.881 (0.773-0.989) GBT 0.853 (0.731-0.974) RF: Random forest; SVM: Support vector machine; LR: Logistic regression; DT: Decision tree; GBT: Gradient Boosting Tree; CI: Confidence interval.