Development and Internal Validation of Machine-Learning Models to Predict Adverse Pathological Features in Upper Tract Urothelial Carcinoma Using Data from the ROBUUST 3.0 Collaborative
Recommended Citation
Korn P, Lokeshwar S, Musso G, Wu Z, Djaladat H, Lee R, Margulis V, Antonelli A, Simone G, Minervini A, Abdollah F, Takashi Y, Shiota M, Sundaram C, Gonzalgo M, Perdonà S, Ferro M, Rais-Bahrami S, Porpiglia F, Mehrazin R, Ghodoussipour S, Derweesh I, Autorino R, Singla N. Development and Internal Validation of Machine-Learning Models to Predict Adverse Pathological Features in Upper Tract Urothelial Carcinoma Using Data from the ROBUUST 3.0 Collaborative. Eur Urol 2026; 89(5s).
Document Type
Conference Proceeding
Publication Date
5-1-2026
Publication Title
Eur Urol
Keywords
creatinine, adult, cancer staging, classifier, clinical feature, conference abstract, controlled study, cross validation, cytology, diagnosis, elastic tissue, female, histology, human, hydronephrosis, lymph node dissection, machine learning, major clinical study, male, neoadjuvant chemotherapy, prediction, predictive model, random forest, risk factor, support vector machine, surgery, transitional cell carcinoma, tumor volume
Abstract
Introduction & Objectives: Reliable preoperative identification of adverse pathological features in upper tract urothelial carcinoma (UTUC) remains challenging and upstaging on final pathology occurs in up to 60% of patients. Current guidelines differ in preoperative risk stratification, and most established risk factors are mainly based on postoperative findings. We aimed to develop and internally validate machine-learning (ML) models for preoperative prediction of muscle-invasive (≥pT2) and node positive disease (pN+) in UTUC. Materials & Methods: Patients who underwent extirpative surgery without neoadjuvant chemotherapy from the ROBUUST 3.0 registry were identifed. Routinely available preoperative variables, including demographics, biopsy pathology (grade, variant histology, cytology), clinical features (cT-stage, cN-stage, tumor size, hydronephrosis, multifocality, serume creatinine) were processed with missing data imputed using Hyperimpute. Four classifiers (elastic-net logistic regression, random forest, LightGBM and SVM-RBF) were optimized via five-fold cross validation and hyperparameter optimization via Optuna. A logistic meta-learner combined base models into a stacking ensemble. Performance was evaluated using area under the curve (AUC) and Brier score in an 80/20 train-validation split. Results: A total of 3584 patients were identified. Among these, 1596 patients who additionally underwent lymph-node dissection were avilable for the pN+ prediction model. Muscle-invasive disease was present in 1979 (55.2%) of cases, and node-positive disease in 1595 (31.6%) cases. The stacking model achieved the best discrimination for pN+ prediction (AUC 0.78 [95% CI: 0.72 – 0.83], Brier 0.16), while for ≥pT2 prediction random forest outperformed other single learners and the stacking ensemble (AUC 0.66 [95% CI: 0.63 – 0.70], Brier 0.23). Significant predictors for pN+ disease were cT-stage, cN-stage and cytology. Significant predictors for >pT2 were cN-stage, creatinine and tumor size. (Figure Presented).Conclusions: ML models achieved robust accuracy for predicting nodal metastases but only moderate discrimination for muscle-invasive disease, underscoring the limitations of accurate preoperative tumor staging. External validation is needed to confirm clinical utility and inform preoperative risk-stratification and multimodal approaches to treatment.
Volume
89
Issue
5s
