Development and interpretability of a machine learning-derived model for the identification of latent prostate cancer in patients initially diagnosed with benign prostatic hyperplasia: a retrospective cohort study.

BACKGROUND: Prostate cancer (PCa) is characterized by pronounced intratumoral heterogeneity and multifocality, presenting ongoing challenges for early and precise diagnosis. In clinical practice, a substantial proportion of patients initially diagnosed with benign prostatic hyperplasia (BPH) are subsequently found to harbor latent PCa during follow-up. However, current prostate-specific antigen (PSA)-based screening lacks the specificity required to resolve this diagnostic overlap. This study aimed to develop and internally validate an interpretable machine learning (ML) model as an adjunct to PSA testing for identifying latent PCa in patients initially diagnosed with BPH. METHODS: This retrospective cohort study utilized data from 1,071 patients to develop a machine learning framework aimed at identifying pre-existing occult PCa diagnosed at least 6 months after the initial BPH assessment. The cohort was partitioned into training (70%) and test (30%) subsets. Within the training set, Recursive Feature Elimination (RFE) was utilized to identify a concise biomarker signature. Six machine learning algorithms were evaluated: logistic regression (LR), random forest (RF), extreme gradient boosting (XGB), support vector machine (SVM), k-nearest neighbors (KNN), and multilayer perceptron (MLP). Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), Brier score, and decision curve analysis (DCA). Model interpretability was facilitated through SHapley Additive exPlanations (SHAP) and Locally Weighted Scatterplot Smoothing (LOWESS) analysis. RESULTS: A concise four-biomarker signature comprising total prostate-specific antigen (tPSA), lactate dehydrogenase (LDH), red blood cell count (RBC), and creatine kinase-MB (CK-MB) was identified. The MLP and LR models exhibited high diagnostic performance, with test AUROC values of 0.949 and 0.946, respectively. CONCLUSION: This research developed and internally validated an interpretable machine learning framework designed to complement traditional PSA screening methodologies. The tool is intended exclusively for research purposes.