Machine learning-integrated explainable artificial intelligence for survival prediction in urothelial carcinoma with enfortumab vedotin: an exploratory real-world analysis

Background: Enfortumab vedotin (EV) has transformed treatment for advanced urothelial carcinoma (aUC), but outcomes vary. Machine learning (ML) with explainable artificial intelligence (XAI) may improve survival prediction. Methods: Data from 544 aUC patients receiving EV after platinum chemotherapy and immunotherapy (51 centers, 24 countries) were analyzed. Four ML algorithms (Random Survival Forest, XGBoost, Elastic Net-regularized Cox, Support Vector Machine) were trained (80%) and tested (20%) to predict overall survival (OS). SHAP analysis (on best performing ML model) provided interpretability. Performance was assessed by C-index and time-dependent AUC. Results: XGBoost (C-index 0.59) and Elastic Net (C-index 0.60) showed best discrimination. XGBoost achieved highest time-dependent AUCs (0.77, 0.87, 0.93 at 1, 2, 3 years). SHAP identified prior immunotherapy (pembrolizumab, atezolizumab/nivolumab), radiotherapy, and upper tract tumors with lower mortality risk; lung, liver, bone, soft tissue metastases increased risk. ECOG performance status and metastatic distribution were key predictors. Conclusion: ML with XAI identifies clinically plausible survival predictors in EV-treated aUC. XGBoost and Elastic Net offer modest risk stratification, that are hypothesis generating but does not support routine clinical use. Functional status, metastatic pattern, and treatment context are key drivers, providing a foundation for externally validated prognostic tools.