Machine learning-based identification of high-risk bone metastasis factors after radical prostatectomy in prostate cancer.

Background: Bone metastasis is a serious complication following radical prostatectomy in prostate cancer patients, significantly affecting their long-term survival. This study aims to develop a clinical predictive model utilizing Magnetic Resonance Imaging (MRI) and advanced machine learning algorithms to identify key factors that increase the risk of bone metastasis (BM). Patients and methods: The study analyzed a cohort of 1161 prostate cancer patients, including 38 who developed bone metastasis. Preoperative T2-weighted images (T2WI) were obtained, and tumor lesions were manually delineated to extract relevant features from the imaging data. Spearman correlation analysis, the least absolute shrinkage and selection operator (LASSO) algorithm, and logistic regression were used to select and construct the model. Four machine learning algorithms-extreme gradient boosting (XGBoost), random forest (RF), support vector machine (SVM), and k-nearest neighbor (KNN)-were employed to predict BM occurrence, integrating these with clinical information. Results: Among the four prognostic models evaluated, the XGBoost algorithm performed the best. In the training dataset, the XGBoost model achieved an AUC of 0.926 (0.870-0.982), an accuracy of 0.847 (0.773-0.921), a sensitivity of 0.880 (0.835-0.926), and a specificity of 0.829 (0.755-0.904). In the validation dataset, the XGBoost model attained an AUC of 0.706 (0.586-0.826), an accuracy of 0.687 (0.661-0.713), a sensitivity of 0.693 (0.557-0.829), and a specificity of 0.664 (0.505-0.822). The external validation dataset yielded an AUC of 0.91, demonstrating the robust predictive capabilities of the XGBoost model. Conclusion: The predictive model for bone metastasis in prostate cancer, developed using the XGBoost machine learning algorithm, shows high accuracy and significant clinical relevance. This model provides a valuable tool for identifying high-risk patients, potentially informing better management and treatment strategies.