Challenges of Utilizing Artificial Intelligence in Identifying Prognostic Patterns in Localized Renal Cell Carcinoma

ABSTRACT Artificial intelligence (AI) tools are increasingly being adopted across a wide range of scientific disciplines, offering transformative potential in data analysis and interpretation and can enhance the understanding and optimization of medical databases. Specifically, in the context of renal cell carcinoma (RCC), prognostic models such as UISS, SSIGN, and ASSURE scores are commonly utilized to estimate recurrence risk and overall survival. However, these models rely heavily on conventional statistical methods and predefined clinical parameters. Our aim was to investigate the feasibility of leveraging readily available AI tools to identify novel prognostic factors for overall survival in RCC patients undergoing surgery and to aid in developing an enhanced, data-driven prognostic scoring system. MATERIAL AND METHODS Data from 1,273 consecutive patients who underwent surgery for renal cell carcinoma between 2010 and 2020 at a German tertiary care center were analyzed. The de-identified dataset included demographics, clinical and pathologi- cal variables, surgical details, clinical scores such as R.E.N.A.L. score, and time to event endpoints such as recurrence and overall survival. Five publicly accessible AI tools (chatGPT, Tableau, Julius AI, Microsoft Power BI, Polymer) were utilized to identify key prognostic parameters and develop a prognostic scoring system for overall survival in RCC pa- tients. These bots were asked to first identify prognostic parameters and then develop a multi-variate prognostic model in response to the question: “Build a prognostic score for overall survival based on a multivariate analysis of the most predictive parameters”. The dataset was divided into a training (n=746) and test (n=527) subsets for score development and validation. Kaplan-Meier survival analysis and log rank testing were performed to evaluate the prognostic accuracy of the generated scoring systems. RESULTS Julius AI proved to be the simplest assistant to utilize, with workflows and programming derived from natural speech that are automatically converted into code. Key factors in predicting outcomes with this AI included duration until recurrence (r=0.75), surgery duration (r=0.35), initial metastasis status (r=0.34), tumor size (r=0.28), female gender (r=0.27), presence of necrosis (r=0.25), T-stage (r=0.24), grading (r=0.24), surgeon (r=0.24), and histology type (clear cell versus non-clear cell) at r=0.21. The AI automatically selected fields to construct a score, incorporating age, gender, ECOG performance status, surgical duration, surgeon, histology, pT stage, tumor size, necrosis presence, resection status, grading, type of nephrectomy (partial vs radical), and R.E.N.A.L. score. Patients were stratified into high-, intermediate-, and low-risk groups, and Kaplan-Meier analysis demonstrating significant survival differences among the groups (log-rank test p < 0.0001). However, one major limitation of the tools was that repeated analyses often yielded varying results, indicating that human judgment remains necessary. CONCLUSIONS AI displays potential in identifying parameters that predict survival outcomes in renal cell carcinoma (RCC) patients based on clinical factors. Nonetheless, human oversight remains crucial. In addition to established scoring systems like SSIGN, UISS, and ASSURE, key factors such as surgeon expertise, surgical duration, and the RENAL score, which have been historically underutilized, should be integrated into future prognostic models. These factors, combined with tumor characteristics such as histology, size, grading, and necrosis, offer an opportunity to enhance the accuracy and utility of prognostic tools. Refining AI systems for greater reliability and reproducibility is critical for advancing their application in clinical decision-making and personalized patient care.