Risk prediction of prostate biopsy outcomes in patients with PI-RADS ≥ 4 lesions using serum lipoprotein(a) and inflammatory markers based on interpretable machine learning: a multicenter cohort study with prospective validation

In patients with PI-RADS ≥ 4 lesions, false-positive MRI findings can lead to unnecessary prostate biopsies, particularly when inflammatory changes resemble malignancy. We developed an interpretable model that combines routine clinical variables with inflammatory and lipid-related markers to predict benign and malignant biopsy outcomes. This multicenter study integrated retrospective model development with prospective validation. The retrospectively collected training cohort included 280 patients from the First Affiliated Hospital of Anhui Medical University, while the prospective internal validation cohort from the same institution and the prospective external multicenter validation cohort included 97 and 123 patients, respectively. Candidate predictors were screened by univariate testing, multicollinearity assessment, and Least Absolute Shrinkage and Selection Operator (LASSO) regression with 10-fold cross-validation. Nine machine-learning models were trained and compared by receiver operating characteristic analysis. SHapley Additive exPlanations (SHAP) were used to interpret the selected model, and a nomogram was constructed from the final logistic regression model. A total of 280 patients were included in the training cohort. Nine candidate variables differed significantly between the BPB and MPB groups. In the LASSO analysis, five predictors were retained: age, fPSA/TPSA ratio, PLR, Lp(a), and PI-RADS category. Among nine machine-learning models developed using these predictors, logistic regression demonstrated the best and most stable discrimination in internal validation (area under the curve [AUC] = 0.824). SHAP analysis identified PI-RADS category and fPSA/TPSA ratio as the two most influential predictors. The resulting nomogram showed good discrimination and calibration in the training (AUC = 0.777; Hosmer–Lemeshow P = 0.46), internal validation (AUC = 0.824; P = 0.58), and external validation (AUC = 0.830; P = 0.24) cohorts. The five-variable logistic nomogram provided interpretable and externally validated prediction of biopsy outcomes in patients with PI-RADS ≥ 4 lesions. By integrating imaging category with clinical, inflammatory, and lipid-related markers, the model may support individualized biopsy decision-making; however, its ability to reduce unnecessary biopsies requires confirmation in prospective clinical implementation studies. This study was registered at the Chinese Clinical Trial Registry (http//www.chictr.org.cn/) under registration number ChiCTR2500114315 on 2025/12/10. Retrospectively registered. • LASSO regression identified a five-variable predictor set comprising age, fPSA/TPSA ratio, PLR, Lp(a), and PI-RADS. • Nine machine-learning models were compared, with logistic regression providing the most stable and interpretable performance. • The five-variable nomogram was validated in independent internal and external cohorts, with AUCs of 0.824 and 0.830, respectively.