Abstract
BACKGROUND: Tumor stage, surgery and age are positively correlated with cancer-specific mortality (CSM) in patients diagnosed with bladder cancer (BCa). In light of the successful application of machine learning to process big data in many fields outside of medicine, we aimed to establish and validate whether machine learning models could improve our ability to predict the development of CSM in elderly BCa patients after radical cystectomy (RC). METHODS: Data on eligible patients diagnosed with BCa were obtained from the Surveillance, Epidemiology, and End Results database (2000-2021) and divided into training and validation cohorts in a ratio of 7:3. First, risk factors for the development of CSM in patients were identified by Cox regression analysis. Then, iterative testing and tuning through automated hyperparameter optimization and ten-fold cross-validation were performed to generate stable extreme gradient boosting (XGBoost) models with optimal performance. Receiver operating characteristic (ROC) curve, area under the curve (AUC), calibration curve and confusion matrix were used to evaluate the performance of XGBoost model. RESULTS: There were 11,763 patients included, of which 5,788 died from BCa. By the comparison of different machine learning models, the final XGBoost model we constructed showed high accuracy and precision in predicting the development of CSM in BCa patients (6-month CSM: AUC =0.799, 12-month CSM: AUC =0.756, 36-month CSM: AUC =0.746, and 60-month CSM: AUC =0.745). The results of accuracy, precision, recall and F1 score confirmed the superior performance of the XGBoost model. The important scores for clinical characteristics and the Shapley Additive Explanations plots highlighted the importance of key factors: chemotherapy, tumor stage, marital status, and tumor size were the top four factors in all models. CONCLUSIONS: Our study validated and confirmed the feasibility and high performance of the XGBoost model in predicting CSM in elderly BCa patients after RC. The potential of machine learning contributes to accurately predict the prognosis of cancer.