Abstract

Background: Prostate-specific antigen (PSA) and multiparametric magnetic resonance imaging (MP-MRI) screening for prostate cancer (Pca) frequently results in unnecessary prostate biopsy (PB), especially in patients with non-clinically significant Pca (NCS-Pca). This study aimed to validate the performance of the Prostate Biopsy Risk Score (PBRS) and machine learning (ML) models in predicting Pca and to streamline the PB and subsequent treatment workflow.
Methods: A retrospective and prospective study was conducted across multiple hospital branches, enrolling 3,467 participants. Cohort I (retrospective training set, Jan 201-Dec 2018, n = 1,078) was used to develop PBRS and ML models. Cohort II (retrospective internal validation, Dec 2018-Dec 2023, n = 2,190) validated model performance and established PBRS thresholds for biopsy avoidance. Cohort III (prospective validation, Jan-May 2025, n = 199) validated these results. Diagnostic performance was assessed using area under the curve (AUC), decision curve analysis, and net reclassification improvement.
Results: PBRS achieved an AUC of 0.89 for Pca prediction, comparable to ML models (0.88-0.89) and superior to PSA (0.76) and MP-MRI (0.84). In retrospective Cohort II, PBRS ≤5 or ≥18 yielded high predictive accuracy (95-100%). PBRS ≤5 showed negative predictive values of 100% (scores 1-4) and 95% (score 5), while PBRS ≥18 demonstrated positive predictive values of 95-99%. These results were confirmed prospectively in Cohort III, with 100% predictive values at the defined thresholds. Based on high prediction accuracy, a one-stop clinical pathway was implemented, incorporating 3D visual ultrasound-targeted biopsy and immediate radical prostatectomy (RP) for patients with PBRS ≥18. Additionally, for RP patients without prior biopsy, analysis confirmed that all patients with PBRS ≤5 had NCS-Pca, whereas all with PBRS ≥18 had clinically significant Pca requiring intervention.
Conclusion: PBRS performs comparably to ML models in predicting Pca and effectively identifies candidates for biopsy-free (PBRS ≤5 or ≥18), enabling non-invasive risk stratification and personalized management in men with suspected Pca.

Keywords

Prostate cancer, non-invasive diagnosis, predictive scale, machine learning models, personalized treatment