Philipp Korn, Maximilian Pallauf, Soum Lokeshwar, Giacomo Musso, Zhenje Wu, Linhui Wang, Alex Kutikov, Randall Lee, Francesco Ditonno, Alessandro Veccia, Riccardo Bertolo, Alessandro Antonelli, Lin Lin, Vitaly Margulis, Alireza Ghoreifi, Hooman Ebrahimi, Hooman Djaladat, Hidefumi Kinoshita, Takashi Yoshida, Gabriele Tuderti, Flavia Proietti, Giuseppe Simone, Andrea Mari, Luca Lambertini, Andrea Minervini, Marco Tozzi, Matteo Ferro, Roberto Contieri, Sisto Perdona, Daniele Amparore, Gabriele Bignante, Francesco Porpiglia, Omri Falik Nativ, Mark L Gonzalgo, Alfonso Santangelo, Carlo Silvani, Alex Stephens, Firas Abdollah, Melinda Fu, Saum Ghodoussipour, Alec Zhu, Reza Mehrazin, Daniel A Sidhom, Chandru P Sundaram, Takashi Matsumoto, Masaki Shiota, Alireza Dehghanmanshadi, Soroush Rais-Bahrami, Karim Daher, Ithar H Derweesh, Riccardo Autorino, Nirmish Singla
We identified 3,179 patients with ≥pT2 stage (636 designated as the test cohort) and 1493 patients with pN+ disease (299 designated as the test cohort). In the test cohort, discrimination was moderate and comparable across models. For ≥pT2 prediction, AUC ranged from 0.73 to 0.74 (best model: stacking, AUC 0.74, 95% confidence interval [CI]: 0.70-0.78, and Brier 0.21). For pN+, AUC ranged from 0.73 to 0.75 (best model: stacking, AUC 0.75, 95% CI 0.69-0.81, and Brier 0.18). Established clinical predictors, particularly biopsy grade and clinical staging parameters dominated risk estimation. Limitations include retrospective design and lack of external validation.
BACKGROUND: Accurate preoperative staging and identification of adverse pathological features in upper tract urothelial carcinoma (UTUC) remain contemporary challenges. We aimed to develop and internally validate machine-learning (ML) models to predict muscle-invasive (≥pT2) and lymph node positive (pN+) disease using preoperative variables.
MATERIAL AND METHODS: We analyzed a large, multi-institutional registry of patients treated surgically for clinically nonmetastatic UTUC (ROBUUST 3.0). We separately identified patients with ≥pT2 stage and pN+ stage at surgery. Predictors were selected a priori based on clinical relevance, and missing data were imputed using Hyperimpute. Five classifiers (elastic net logistic regression, random forest, LightGBM, support vector machine with radial-basis-function kernel and a stacking ensemble) were trained using an 80/20 stratified train-test split. Discrimination (measured by area under the curve (AUC)), calibration and Brier score were assessed in the independent test cohort.
RESULTS: We identified 3,179 patients with ≥pT2 stage (636 designated as the test cohort) and 1493 patients with pN+ disease (299 designated as the test cohort). In the test cohort, discrimination was moderate and comparable across models. For ≥pT2 prediction, AUC ranged from 0.73 to 0.74 (best model: stacking, AUC 0.74, 95% confidence interval [CI]: 0.70-0.78, and Brier 0.21). For pN+, AUC ranged from 0.73 to 0.75 (best model: stacking, AUC 0.75, 95% CI 0.69-0.81, and Brier 0.18). Established clinical predictors, particularly biopsy grade and clinical staging parameters dominated risk estimation. Limitations include retrospective design and lack of external validation.
CONCLUSIONS AND CLINICAL IMPLICATIONS: ML models using preoperative variables achieved moderate discrimination for predicting adverse pathological features in UTUC. External validation and prospective assessment of clinical utility are necessary before informing preoperative patient selection.