# Role of urinary leukocytes in the risk stratification of prostate cancer using nonlinear stacking learning strategy: a bi-cohort diagnostic study

## TL;DR

This study shows that urinary leukocytes can improve prostate cancer risk prediction, especially in patients with intermediate PSA levels, when added to machine learning models.

## Contribution

The novel contribution is demonstrating the incremental clinical utility of urinary leukocytes within a non-linear stacking learning framework for prostate cancer risk stratification.

## Key findings

- Incorporating urinary leukocytes improved model AUC by 0.003–0.02 in specific subgroups.
- The benefit of urinary leukocytes was most pronounced in patients with intermediate PSA levels (4–10 ng/mL).
- SHAP analysis showed urinary leukocytes complement PSA-derived indices without being a dominant driver.

## Abstract

This study aimed to develop a non-linear stacking ensemble learning framework to evaluate the incremental diagnostic contribution of urinary leukocytes (UL) in prostate cancer (PCa) risk stratification, with a primary focus on predictive performance and clinical utility.

We retrospectively included 492 men with elevated PSA levels from Tongji Hospital (n = 415) and Xiangyang Central Hospital (n = 77). All patients underwent transrectal ultrasound-guided prostate biopsy and were classified into low-, intermediate-, and high-risk PCa according to the Gleason score. Clinical variables, including age, BMI, tPSA, f/tPSA, p2PSA, PHI, PHID, and PSAD, were collected and standardized. Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression combined with bootstrap-based stability analysis. Based on the selected features, we constructed a non-linear stacking ensemble comprising decision tree, logistic regression, support vector machine (SVM), k-nearest neighbors (KNN), and gradient boosting as base learners. Three-class risk stratification models were trained under two scenarios: with and without incorporation of UL. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC), macro-F1 score, and the DeLong test. Calibration curves and decision curve analysis (DCA) were applied to quantify the incremental net clinical benefit associated with UL. Predefined subgroup analyses across PSA strata (<10, 10–20, >20 ng/mL) were conducted to examine the context-dependent contribution of UL.

In the baseline setting without UL, the non-linear stacking model achieved AUCs of 0.962 and 0.928 in the internal and external cohorts, respectively, indicating robust discriminative performance. After incorporating UL, several base learners—particularly decision tree, KNN, and gradient boosting—demonstrated center-specific AUC improvements ranging from 0.003 to 0.02 (p < 0.05), accompanied by consistently increased net clinical benefit on DCA. Subgroup analyses showed that the incremental value of UL was most evident in patients with intermediate PSA levels (4–10 ng/mL) and in those with clinical features suggestive of benign prostatic hyperplasia. Post hoc SHapley Additive exPlanations (SHAP) analyses performed on a representative base learner indicated that UL exerted a modest but directionally consistent influence on high-risk predictions, complementing established PSA-derived indices rather than acting as a dominant independent driver.

Within a stacking ensemble–based risk stratification framework primarily optimized for predictive performance, urinary leukocytes provide a clinically meaningful auxiliary signal that improves discrimination and net benefit in specific PSA-defined subgroups. These findings support the use of UL as a complementary inflammation-related marker in PCa risk assessment, while interpretability is best understood at the level of base learners and original clinical features rather than the full ensemble model.