AI-Powered Urine Test Reduces Unnecessary Prostate Biopsies by Up to 23%
A multicentre study evaluated EGPS, an AI-based urine test using extracellular vesicle gene expression to detect clinically significant prostate cancer without digital rectal examination. Across training, internal, and external validation cohorts the model achieved AUCs of 0.811-0.838 with sensitivity above 95%, reducing unnecessary biopsies by 15-23%. The DRE-free, non-invasive approach could streamline screening in men with PSA 0-15 ng/mL.
The original study
Multicentre Evaluation of an AI-Assisted Urine Test for Clinically Significant Prostate Cancer in Men Undergoing Initial Biopsy.
- Authors
- Jiang S, Yang C, Huang Z, Guo Z, Lu F, Nian X, et al.
- Journal
- Journal of extracellular vesicles
- PMID
- 41877634
Original abstract
The Extracellular Vesicles Gene-based Prostate Score (EGPS), powered by DeepSeek, is an artificial intelligence (AI) diagnostic tool that enhances the detection of clinically significant prostate cancer (csPCa) using urinary EV-derived gene expression, without requiring digital rectal examination (DRE). To address overdiagnosis resulting from the limited specificity of prostate-specific antigen (PSA) and reduce unnecessary biopsies, this study evaluated the clinical utility and generalizability of EGPS in men undergoing initial biopsy with PSA levels ranging from 0 to 15 ng/mL. A total of 645 patients were retrospectively enrolled: 586 from three centres were divided into training (70%) and internal validation (30%) cohorts, and 59 from two centres served as the external validation cohort. EVs were isolated using the EXODUS platform, and gene expression was measured by RT-qPCR. Ten machine learning algorithms were evaluated for constructing the EGPS model with selected genes. Diagnostic efficacy was assessed by ROC analysis, DeLong tests, and decision curve analysis. An AI diagnostic system using DeepSeek was also developed. The EGPS model, incorporating AMACR, HOXB13, and PSGR, achieved AUCs of 0.838, 0.825, and 0.811 in the training, internal validation, and external validation cohorts, respectively, outperforming PSA. At a cut-off value of 0.22, the model demonstrated sensitivity above 95%, with a missed diagnosis rate of 3.81% in the training cohort and 0% in the validation cohorts. The model reduced unnecessary biopsies by 79 (23.37%), 27 (18.62%) and 9 (15.25%) cases across the three cohorts, thereby lowering biopsy-related risks. A DeepSeek-powered AI diagnostic system integrating EGPS was developed to support csPCa diagnosis and minimize unnecessary biopsies. EGPS, derived from multicentre Chinese cohorts, enables accurate, DRE-free, non-invasive prediction of csPCa in men with PSA levels of 0-15 ng/mL. When integrated into an AI system, EGPS supports early screening and personalized clinical decision-making by reducing unnecessary biopsies.