Machine learning models can accurately predict survival outcomes after surgery for upper tract urothelial cancer, offering a potential tool to guide post operative management and follow up intensity. A large international analysis shows that data driven models may help identify patients most likely to benefit from closer surveillance rather than immediate adjuvant chemotherapy following radical nephroureterectomy.
Unmet Need in Post Operative Risk Stratification
Upper tract urothelial cancer is commonly stratified into high and low risk categories, with radical nephroureterectomy and bladder cuff resection considered standard treatment for non metastatic high risk disease. However, there is currently no consensus on post operative management, particularly with regard to selecting patients for close follow up versus adjuvant chemotherapy. Existing clinical tools provide limited guidance, creating uncertainty in treatment planning after surgery.
Study Design and Machine Learning Approach
Investigators retrospectively collected data from a large multiethnic cohort of 3,129 patients with histologically confirmed upper tract urothelial cancer treated with radical nephroureterectomy across centres in Asia and Europe. A total of 637 Asian patients formed the training cohort, while 2,492 European patients were used for external validation. Twenty supervised machine learning models were trained and tested to predict overall survival, cancer specific survival, and disease-free survival at three and five years.
Nomograms were constructed using eight independent prognostic factors: age, sex, tumour grade, pathological tumour stage, pathological nodal status, presence of carcinoma in situ, multifocality, and lymphovascular invasion. Model performance was assessed using area under the receiver operating characteristic curve.
Predictive Performance and Clinical Implications
During model training, logistic regression-based approaches achieved the strongest performance, ranking highest for four of six outcomes. The best predictive accuracy was observed for cancer specific survival at three and five years, with area under the curve values of 0.85 and 0.84 respectively, and for disease free survival at three years with an area under the curve of 0.81. On external validation, logistic regression models continued to perform well, ranking first for three of six outcomes, including cancer specific survival at three years with an area under the curve of 0.84.
The findings suggest that machine learning can provide robust prognostic estimates after surgery for upper tract urothelial cancer and may account for epidemiological differences between European and Asian populations. While further clinical validation is required, these models could support personalised decision-making regarding adjuvant therapy and follow up strategies in routine practice.
Reference
Nicoletti R et al. Training and external validation of machine learning supervised prognostic models of upper tract urothelial cancer (UTUC) after nephroureterectomy. Scientific Reports. 2026;https://doi.org/10.1038/s41598-025-29043-w.
