Identification of biomarkers for non-invasive diagnosis and risk stratification in prostate cancer using NMR-based metabolomics and machine learning

The non-invasive diagnosis and risk stratification of prostate cancer (PCa) remain clinically challenging due to the limited specificity of prostate-specific antigen (PSA). In this retrospective study, we applied a comparative machine learning (ML) framework to rank and select biomarkers from serum 1H nuclear magnetic resonance (NMR)-based metabolomics, subsequently developing three sequential metabolite panels. These metabolite-based models effectively distinguished overall PCa from benign prostatic hyperplasia (BPH), clinically non-significant prostate cancer (cnsPCa) from BPH, and clinically significant prostate cancer (csPCa) from cnsPCa. All models achieved areas under the curve (AUC) consistently above 0.9 in both discovery and validation cohorts, without incorporating clinical variables such as age or PSA. Decision curve analysis (DCA) further confirmed their superior clinical utility over the current PSA-based strategy. This study underscores the potential of ML-driven metabolomics for accurate, non-invasive diagnosis and effective risk stratification of PCa, which could significantly improve patient management and reduce unnecessary interventions.