Abstract
Prostate cancer (PCa) is the second most prevalent malignancy among men worldwide. However, current screening tools such as serum prostate-specific antigen (PSA) tests and digital rectal examination (DRE) are limited by low specificity and high false-positive rates, often leading to unnecessary biopsies and overtreatment. To address this clinical challenge, we developed a novel diagnostic framework termed PCASSO (Prostate CAncer diagnosis using Sensitive and Sophisticated ML classifiers based on nOn-invasive urinary RNA biomarkers), which integrates machine learning (ML) algorithms with non-invasive urinary RNA biomarker profiles obtained from DRE-free whole urine. A total of 163 urine samples (112 PCa, 51 benign prostatic hyperplasia [BPH]) were analyzed using quantitative PCR for 20 RNA biomarkers, including 2 long noncoding RNAs, 1 fusion gene, and 17 miRNAs. Among six ML classifiers evaluated, a Gradient Boosting model using an optimized 9-biomarker panel achieved the highest diagnostic performance (AUC: 0.99), with robust cross-validation results (Stratified-K-Fold: 0.912; LOOCV: 0.890). Notably, this classifier retained high accuracy in patients within the PSA gray zone (3-10 ng/mL), where clinical decision-making is often ambiguous. Our results demonstrate that ML-based classifiers using DRE-free urinary RNA biomarkers showed improved performance through robust internal validation, providing a basis for future validation studies.