Explainable machine learning estimates clinically significant prostate cancer risk in men with PSA 4–20 ng/mL

Abstract Men with prostate-specific antigen (PSA) values of 4–20 ng/mL often have overlapping risks of benign disease, clinically insignificant prostate cancer, and clinically significant prostate cancer (csPCa). We retrospectively studied 859 men who underwent pathological evaluation; 76 had csPCa. Twelve prediction algorithms were assessed with 500 bootstrap resamples. In every resample, LASSO feature selection, hyperparameter tuning, and model fitting used only the bootstrap training sample, and performance was estimated from patient-level out-of-bag predictions. C5.0 had the highest AUROC point estimate (0.843, 95% CI 0.801–0.886), an AUPRC of 0.421, and a Brier score of 0.066. Calibration was broadly concordant at low-to-moderate predicted risks but uncertain in the sparse high-risk tail. At a 0.10 risk threshold, the model would avoid 544.8 biopsies per 1,000 men while missing 10.5 csPCa cases per 1,000. PSA-derived measures supplied most of the linear predictive signal, whereas the clearest additional gain arose with nonlinear integration. C5.0 may support pre-biopsy triage, but multicenter external validation, recalibration assessment, and prospective impact studies are required before clinical use.