Using a machine learning algorithm to predict muscle invasiveness in upper tract urothelial carcinoma: Insights from the ROBUUST collaborative group
Recommended Citation
Bignante G, Orecchio C, Checcucci E, Amparore D, Alladio E, Sundaram CP, Derweesh IH, Margulis V, Abdollah F, Ferro M, Djaladat H, Simone G, Mehrazin R, Gonzalgo ML, Wu Z, Correa AF, Antonelli A, Rais-Bahrami S, Singla N, Perdonà S, Yoshida T, Fiori C, Autorino R, Porpiglia F. Using a machine learning algorithm to predict muscle invasiveness in upper tract urothelial carcinoma: Insights from the ROBUUST collaborative group. Eur Urol 2025; 87(S1):540.
Document Type
Conference Proceeding
Publication Date
3-1-2025
Publication Title
Eur Urol
Abstract
Introduction & Objectives: In Upper Tract Urothelial Carcinoma (UTUC), as with most solid tumors, tumor grade and stage are the primary independent predictors of patient outcomes. While tumor grade can now be assessed through endoscopic biopsies, available surgical and radiological techniques still lack the precision needed to reliably determine the clinical tumor stage. This study aimed to predict muscle invasiveness (MI) in UTUC using a machine learning (ML) algorithm. Materials & Methods: The study population was drawn from the ROBUUST (ROBotic surgery for Upper Tract Urothelial Cancer STudy) dataset, which includes data from 18 high-volume tertiary centers worldwide. This dataset contains information on patients diagnosed with UTUC who underwent surgical treatment between 2015 and 2024. Patients included in this analysis had undergone open, laparoscopic or robot-assisted surgery for UTUC. Exclusion criteria included clinical nodal involvement, distant metastases, or prior neoadjuvant chemotherapy. To predict a categorical outcome (non-MI vs. MI UTUC), we applied a Random Forest (RF) model, an ensemble of individual classification trees. Each tree within the RF generates a class prediction, with the majority vote determining the model's final output. After curating the dataset, we randomly split it into a calibration set (80% of patients) for model training and an evaluation set (20%) used only at the end to assess the model's performance, ensuring generalizability to new data. Results: A total of 755 patients and 18 clinical tumor-related variables met the inclusion criteria. Using the RF model on the evaluation set, we assessed the model's performance through three key metrics: the confusion matrix, the ROC curve, and the classification report. The RF model demonstrated an accuracy of 62% and an AUC of 0.65. Preliminary analyses identified the most important predictors distinguishing non-MI from MI tumors: tumor size, symptoms, high-grade (HG) findings on biopsy, and hydronephrosis were the primary risk factors associated with MI tumors. Conclusions: ML algorithms like the RF model may aid in preoperatively identifying MI UTUC cases, guiding more tailored treatment approaches. Further studies on larger cohorts are needed to confirm these findings and improve predictive accuracy. 40th Annual EAU Congress
Volume
87
Issue
S1
First Page
540
