Machine learning-based integration of systemic immune-inflammation and nutritional signatures for predicting disease-free survival in upper tract urothelial carcinoma: a multicenter study

Background: Current staging for Upper Tract Urothelial Carcinoma (UTUC) fails to capture host biological heterogeneity. We aimed to develop and validate a machine - learning based prognostic signature integrating systemic immune - inflammatory and nutritional markers to enhance UTUC risk stratification. Methods: A total of 606 UTUC patients from four centers were divided into a training set (n = 263), an internal validation set (n = 114), and two external validation sets (n = 113, n = 116). Thirteen preoperative hematological markers were screened using LASSO - Cox regression. A Random Survival Forest (RSF) algorithm was utilized to construct a prognostic "ML Score", and SHAP analysis visualized the nonlinear relationships. A composite nomogram integrating the ML Score with clinical factors (age, grade, pT stage) was developed and comprehensively evaluated. Results: Seven principal predictors were identified: RDW, PLT, NEUT, NLR, SIRI, SII, and AISI. SHAP analysis revealed distinct nonlinear threshold effects of RDW and PLT on mortality risk. The ML Score served as an independent prognostic indicator, successfully identifying patients with significantly poorer disease - free survival (DFS) across all four cohorts (P < 0.01). The integrated nomogram demonstrated outstanding predictive accuracy, with a C - index of 0.762 in the training set and maintaining robust performance in all validation cohorts. Decision curve analysis confirmed its superior clinical net benefit. Conclusion: We developed and validated a robust ML Score that reflects the host's systemic immune - inflammatory and nutritional status. It offers substantial incremental prognostic value and serves as an accurate, non - invasive tool for personalized risk assessment in UTUC.