Artificial intelligence and identifying prognostic patterns in upper tract urothelical carcinoma (UTUC).

• DOI: 10.1200/jco.2024.42.16_suppl.e16616
• Publication Date: 2024-06-06T17:50:54Z
• Authors: Michael Staehler; Troja Ivanova; Iulia Blajan; Can Aydogdu; Annabel Graser; Marina Hofmann; Gerald Bastian Sc...
• Keywords: 03 medical and health sciences; 0302 clinical medicine

e16616 Background: Artificial intelligence tools (AI) are increasingly used in all scientific fields. UTUC is rare form of kidney cancer and prognostic factors seem unclear. We sought to understand if AI could help identify prognostic parameters for overall survival in UTUC and aid in the development of a prognostic score. Methods: n=231 consecutive patients who underwent surgery for UTUC between 2005 and 2020 were included. De-personalized data included were age, gender, ECOG performance status, smoking habits, baseline creatinine and platelets, histology, laterality, pathological and clinical T-stage, grading, cis, TNM stage, necrosis, lymphovascular invasion, vascular invasion, resection margin, tumor size, duration of surgery, surgeon, surgical approach, blood loss, complications (graded by Clavien-Dindo), complete removal of the ureter including bladder cuff, duration until recurrence and duration until death and reason of death. Several publicly accessible AI tools (chatGPT, Tableau, Julius AI, Microsoft Power BI, Polymer) were asked to first identify prognostic parameters and develop a prognostic score for RCC outcome. Statistical analysis was done using chi-square test, cox regression and Kaplan-Meier survival estimation. AI was asked to analyze the data set based on Kaplan-Meier survival analysis and cox proportional log rank testing was performed to estimate score accuracy. Results: Median age was 70.9 years (43.7-81.8). 61% (n=160) were male. ECOG performance stats was 0 in 42.1% (n=82), 1 in 27.2% (n=52). 53.6% (n=105) were non-smoker and 24.5% (n=48) smoker and 21.9% (n=43) were former smokers. 10.3% (n=12) had synchronous bladder cancer. 60.4% (n=131) suffered from gross hematuria and 19.7% (n=42) had flank pain. Median follow-up was 34.5 months (0-174). Median overall survival (OS) was 58.9 months (95% CI 46.2-71.5). Bellmunt risk factors, BMI and baseline creatine and CRP levels correlated significantly with adverse overall survival (p>0.001). Tumors localized in the ureter, larger than 5cm, carcinoma in situ and high grade had a significant worse OS. Resecting a bladder cuff improved OS to 86.7 months (95% CI 65-108.4) compared to 39.9 months (95% CI 27.9-51.9, p=0.012) without. AI revealed ECOG performance status, nodal status, grade, surgical expertise and resecting a bladder cuff to be prognostic factors and did not identify conventional prognostic factors to be relevant for a scoring system to predict OS and PFS. Conclusions: AI can help identify parameters to predict overall survival outcomes in patients with UTUC based on clinical factors. Human adjustment still is necessary. Interestingly, surgeon, complete resection of the ureter including a bladder cuff are factors that can be influenced and should be considered in future evaluations of prognostic models besides tumor characteristics like histology, grading and nodal status that can’t be medically influenced.