# Training and external validation of machine learning supervised prognostic models of upper tract urothelial cancer (UTUC) after nephroureterectomy

## TL;DR

This study develops and validates machine learning models to predict prognosis in upper tract urothelial cancer patients after surgery, showing promising results for clinical use.

## Contribution

The study introduces the first ML-based prognostic model for UTUC that accounts for differences between Asian and European patient populations.

## Key findings

- Logistic regression models achieved the best performance in predicting cancer-specific survival at 3 and 5 years.
- The model showed strong external validation results, particularly for cancer-specific and overall survival outcomes.
- The study highlights ML's potential in UTUC prognosis and emphasizes the need for further clinical validation.

## Abstract

The European association of Urology (EAU) suggests a prognostic stratification of Upper Tract Urothelial Cancer (UTUC) based on high and low risk patients, with Radical nephroureterectomy (RNU) and bladder cuff resection being the gold standard for the treatment of non-metastatic High risk UTUC. However, no consensus on post-operative patient management or tools that predict who would benefit the most from a close follow-up rather than adjuvant chemotherapy regimen exist. in Machine Learning (ML) is gaining interest in Urology providing models for prognostic prediction purpose; It’s role in UTUC has not yet been investigated. We aim to develop and validate multiple supervised ML models based on patient- and tumor- related features to predict prognosis in patients with preoperative Histological or Imaging proved UTUC treated with RNU within a multiethnic large cohort. Data from an international multicenter large cohort of histologically proven UTUC patients from Asia and Europe treated with RNU were retrospectively collected. Twenty different ML-supervised predictive models were first trained and then external validate with two separate set. Nomograms were constructed based on 8 independent prognostic factors (age, gender, grading, pT, pN, presence of Carcinoma in Situ (CIS), multifocality and Lymphovascular invasion(LVI)) to predict 6 Outcomes (Overall Survival (OS), Cancer Specific Survival (CSS) and Disease Free Survival (DFS) at 3 and 5 year). Performances were compared using Area-under-curve (AUC) of Receiver-Operating Characteristics (ROC). A total of 3129 patients were enrolled: 637 Asian Patients (training cohort) and 2492 European patients (validation cohort). Upon training assessment, LR models achieved the best results, being the best model for prediction of 4/6 outcomes, with the best result in CSS both at 3 and 5 years (AUC: 0.85, 0.84, 0.81 for CSS-3y, CSS-5y and DFS-3y respectively). Upon external validation, LR(CSL) models achieve the best results, being the number 1 model for prediction of 3/6 outcomes (AUC: 0.84, 0.79, 0.77 for CSS-3y, OS-3y and OS-5y respectively). ML is a promising technology in the field of UTUC. Our model achieve favorable results in terms of prediction of prognosis after RNU, especially in terms of CSS at 3 and 5 years, moreover is the first model of prognosis taking into account the differences in epidemiology existing between European and Asian patients. Further clinical validation and verification of its reliability for the case selection of adjuvant therapy are needed to assess its use in clinical practice linked to clinical decision making. ML is an advancing technology in the field of medicine and urology, which can also be applied to the definition of the prognosis of patients with UTUC undergoing RNU. Our study represents the first experience investigating this potential.

The online version contains supplementary material available at 10.1038/s41598-025-29043-w.