Artificial Intelligence-Based Personalized Oncological Outcome Prediction Model for Upper Urinary Tract Urothelial Carcinoma after Radical Nephroureterectomy: A Development and Multicenter Validation - Beyond the Abstract
We developed and externally validated an artificial intelligence-based survival prediction model for patients with upper urinary tract urothelial carcinoma (UTUC) following radical nephroureterectomy (RNU). The model was trained on 627 patients from three Korean institutions and demonstrated C-indices of 0.73–0.82 for disease-free survival and overall survival across internal and external validation cohorts.
Clinical Need for Individual Survival Prediction
Prognostic counseling for UTUC patients after RNU remains challenging. Current tools predominantly rely on Cox proportional hazards (Cox-PH) regression, which has two fundamental limitations in clinical application.
First, Cox-PH provides relative risk estimates—a hazard ratio indicating how much higher or lower a patient's risk is compared to a reference group. However, without knowing the baseline survival probability of that reference group, clinicians cannot translate this into an absolute survival estimate for the individual patient. Second, Cox-PH assumes that the effect of all covariates remains constant over time (proportional hazards assumption). In oncological data, this assumption is violated in approximately 20% of studies—for example, the prognostic impact of pathological T stage may differ between early and late follow-up periods.
The consequence is a gap between what models provide (relative risks, population-level estimates) and what patients need (their individual probability of survival at specific time points).
Methodological Approach: Accelerated Failure Time Model
To address these limitations, we employed the Accelerated Failure Time (AFT) model instead of Cox-PH regression. The AFT model directly models survival time rather than hazard ratios, and does not require the proportional hazards assumption. This enables direct calculation of patient-specific survival probabilities at any time point.
Our model architecture combines XGBoost for risk score generation with a bootstrapped Weibull AFT model for survival distribution estimation. Hyperparameter optimization was performed using Optuna over 500 trials. This hybrid approach generates individual survival distributions (ISD): each patient receives personalized survival probabilities with confidence intervals at 1, 3, 5, and 10 years—not a single hazard ratio or a population average.
External Validation Across Heterogeneous Populations
The model was externally validated using datasets from Seoul National University Hospital (n=255) and Hallym University Medical Center (n=107). Notably, these cohorts differed significantly from the development cohort: the Asan Medical Center cohort had higher rates of node-positive disease (17.5% vs. 2.9–3.7%), tumor multifocality (18.3% vs. 10.3–12.9%), and concomitant carcinoma in situ (40.9% vs. 3.7–29.0%). Despite these differences, model discrimination and calibration remained robust, supporting generalizability across institutions.
Clinical Deployment
The model was deployed via Google Cloud Platform with a web-based interface. In clinical practice, we observed that the interactive, visual presentation of individualized survival curves facilitates more engaged patient participation in prognostic discussions compared to conventional numerical estimates.
Limitations
This model was developed using 2010–2020 data, prior to the integration of newer systemic agents into UTUC management. Additionally, lymphovascular invasion and adjuvant therapy information were not fully incorporated. Future model updates addressing these factors will be necessary.
Written by: Jungyo Suh, MD, Assistant Professor, Department of Urology, Asan Medical Center, University of Ulsan College of Medicine, Seoul, Korea
Read the Abstract