Dual-center development and external validation of machine learning models integrating PSA-derived and peripheral inflammatory markers for prostate cancer diagnosis.

Background: Prostate-specific antigen (PSA) has limited specificity for prostate cancer (PCa), particularly within the PSA gray zone (4-10 ng/mL), where benign and malignant conditions overlap. This study aimed to develop and evaluate logistic regression, decision tree, and extreme gradient boosting (XGBoost) models that combine PSA-derived variables with routinely available peripheral inflammatory markers, with a prespecified subgroup analysis of patients in the PSA gray zone. Methods: This retrospective dual-center study consecutively enrolled 317 newly diagnosed, treatment-naïve patients at Hainan General Hospital between January 2019 and December 2024, including 151 patients with PCa and 166 with benign prostatic hyperplasia (BPH). The PSA gray-zone subgroup comprised 116 patients (46 PCa and 70 BPH). Total PSA (tPSA), free PSA (fPSA), and routine blood-cell parameters were obtained from the first blood draw after admission and before biopsy, surgery, or anticancer treatment; the free-to-total PSA ratio (f/t PSA ratio), neutrophil-to-lymphocyte ratio (NLR), and monocyte-to-lymphocyte ratio (MLR) were calculated. Logistic regression, decision tree, and XGBoost models were developed using stratified 8:2 training/internal-validation splits. A separate cohort of 60 patients (30 PCa and 30 BPH) from a different institution was used for external validation. Direct area under the curve (AUC) comparisons were performed on a common internal-validation set using paired DeLong tests. Results: Among individual indicators, tPSA showed the highest AUC in the overall cohort (0.801), whereas the f/t PSA ratio showed the highest AUC in the gray-zone subgroup (0.683). The decision tree achieved AUCs of 0.903, 0.867, and 0.724 in the training, internal-validation, and external-validation sets, respectively; the corresponding XGBoost AUCs were 0.930, 0.933, and 0.781. On the common internal-validation set used for paired comparison, the AUCs were 0.811 for logistic regression, 0.867 for the decision tree, and 0.851 for XGBoost, with no significant between-model differences after multiple-comparison correction. In the gray-zone subgroup, the decision tree used the f/t PSA ratio as its primary split and had the highest optimism-corrected AUC (0.757), compared with XGBoost (0.649) and logistic regression (0.609). Conclusions: Models combining PSA-derived variables with peripheral inflammatory markers showed potential as complementary tools for PCa risk stratification, also within the PSA gray zone. No multivariable model demonstrated statistically superior discrimination in the paired comparison. The decision tree offered explicit threshold-based rules, but prospective multicenter evaluation and calibration assessment are required before clinical use.