Ensemble Machine Learning Models for Predicting Bone Metastasis in Bladder Cancer

Zhan jiang YuXiang Da XuXin Chang ZouPei De SuHai Chao ChaoTao Zeng^*

• Nanchang University Second Affiliated Hospital, Nanchang, China

Background and Purpose: The occurrence of bone metastasis in advanced bladder cancer often signifies a poor prognosis. Currently, accurately predicting bone metastasis in bladder cancer remains a challenge. This study develops predictive models using machine learning algorithms to forecast bladder cancer bone metastasis (BCBM) and aid in personalized clinical decisions. Patients and Methods: We reviewed and analyzed data from patients diagnosed with bladder cancer (BC) between 2010 and 2015 in the Surveillance, Epidemiology, and End Results (SEER) database. In addition, we included 327 patients treated at the Second Affiliated Hospital of Nanchang University and Jiangxi Cancer Hospital as an external validation cohort. Independent risk factors for bone metastasis (BM) in BC patients were identified through univariate and multivariate logistic regression analyses. These features were then integrated into seven machine learning algorithms to build predictive models, including logistic regression (LR), support vector machine (SVM), gradient boosting machine (GBM), neural network (NN), random forest (RF), extreme gradient boosting (XGB), and K-nearest neighbors (KNN). The performance of these models was evaluated using the area under the receiver operating characteristic curve (AUC), along with accuracy, sensitivity (recall), and specificity. Results: A total of 22,114 patients diagnosed with BC were included in this study, with 537 (2.4%) developing BM. The identified independent risk factors for BCBM included age, race, tumor histology, tumor grade, T stage, N stage, presence of brain metastasis, liver metastasis, lung metastasis, and history of radiotherapy. Among the seven developed machine learning models, the tree-based GBM model exhibited the best performance on the test set, achieving AUC, accuracy, sensitivity, and specificity values of 0.855, 0.813, 0.733, and 0.815, respectively. The GBM model also demonstrated robust performance in the external validation set, achieving an AUC of 0.766 and accuracy of 0.945. According to Shapley Additive Explanations (SHAP), the most significant feature in the GBM prediction model was the T stage, followed by the N stage and radiotherapy. Conclusion: The GBM model offers a precise and personalized approach to predicting BCBM, potentially enhancing clinical decision-making and the efficiency of bone metastasis screening in bladder cancer patients.