Concordance between clinical, genomic, and AI-based risk stratification in prostate cancer.

Abstract

388 Background: An expanding number of tests utilizing clinical criteria, genomic classifiers, and machine learning are available to assist clinicians on treatment options for patients with localized prostate cancer. The degree of concordance among these tools (Decipher, Artera AI) and traditional clinical stratification systems remains unclear. We conducted a retrospective analysis of 40 patients to evaluate the degree of concordance among these prognostic and risk stratification tools in a localized prostate cancer cohort. Methods: After IRB approval, we analyzed data from 40 patients with localized prostate cancer treated between January and October 2025 who had available NCCN risk classification, Decipher genomic classifier, and Artera AI predictive model results. Each patient was assigned a categorical risk level (low, intermediate, or high) according to each tool’s standard criteria. Concordance was defined as agreement in risk classification between two systems. Pairwise comparisons were performed (NCCN vs Artera AI, Artera AI vs Decipher, and NCCN vs Decipher). Cross-tabulation matrices were generated, and overall concordance rates were calculated as the proportion of patients classified in the same category by both tools. Results: 40 patients were included. Artera AI and NCCN showed an overall concordance of 47.5%, with perfect agreement in the high-risk group (12/12, 100%). Artera AI frequently up-classified NCCN intermediate-risk patients to high risk (15/19). Artera AI and Decipher demonstrated the highest concordance (60.0%), with most discordance when Artera AI assigned intermediate and Decipher assigned high risk (15/19). NCCN and Decipher had a concordance rate of 55.0%, with frequent reclassification of NCCN intermediate-risk patients to Decipher high risk (5/8). High-risk assignments were more consistent, aligning in 25 of 31 cases. Conclusions: Moderate concordance was observed among NCCN, Decipher, and Artera AI, with the highest agreement between Artera AI and Decipher. Discordance was most notable in NCCN intermediate-risk patients, who were often up-classified to high risk by genomic or AI-based tools. These findings suggest that each system captures distinct clinical and biological features, underscoring the limitations of relying on a single classifier. Integrating clinical, genomic, and AI-based assessments may enhance risk stratification and support more individualized treatment decisions.