Development and validation of an AI-enabled prediction of prostate cancer (PCa) using urine-based liquid biopsy.

Abstract

3080 Background: Prostate cancer (PCa) remains a major cause of malignancy-related mortality among men. Current diagnostic techniques, including PSA testing, lack accurate early detection capabilities, while global barriers include limited access to specialized facilities and cultural sensitivities around transrectal biopsy and digital rectal examination. This study evaluates a non-invasive, urine-based liquid biopsy assay for diagnosing PCa through disease-specific biochemical profiles using an artificial intelligence pipeline. Methods: We collected urine from men scheduled for prostate biopsy (biopsy-positive PCa n=197) and healthy controls (n=84). Samples were processed using NUTEC slides, underwent heat cycling, and were converted to digital images for AI analysis. Using 5x2 cross-validation with a random forest classifier, we evaluated cancer detection performance and analyzed cohorts with specific Gleason scores (Gle): Gle 6 (n=70), Gle 7 (3+4) (n=55), Gle 7 (4+3) (n=34), and Gle 8,9,10 (n=38). Results: Our classifier demonstrated strong overall performance in distinguishing cancer versus non-cancer subjects (F1=0.843) with notably high recall (R=0.967). Importantly, performance remained robust across Gleason score cohorts (F1=0.799-0.838), maintaining high recall (R>0.89) while preserving clinically relevant precision. The classifier showed particular strength in detecting intermediate- (Gle 7 (3+4): F1=0.838) and low- (Gle 6: F1=0.822) Gleason grade cancers. Conclusions: AI-enabled prediction of PCa using urine-based liquid biopsy demonstrates accurate, rapid, and accessible early cancer detection, with consistent performance across disease grades. This non-invasive approach addresses both clinical and cultural barriers to prostate cancer diagnostics. TASK F1 P R AUC ACC Cancer v. Controls 0.843 0.748 0.967 0.768 0.748 Gle6 v. Controls 0.822 0.770 0.893 0.776 0.746 Gle347 v. Controls 0.838 0.757 0.940 0.752 0.754 Gle437 v. Controls 0.800 0.715 0.913 0.695 0.691 Gle8910 v. Controls 0.799 0.722 0.900 0.695 0.698 AI-enabled prediction of PCa using urine-based liquid biopsy demonstrates accurate, rapid, and accessible early cancer detection, with consistent performance across disease grades. This non-invasive approach addresses both clinical and cultural barriers to prostate cancer screening.