Enhancing the Prediction of Clinically Significant Prostate Cancer Through Feature Engineering

Abstract

Purpose: Adverse pathological features in clinically significant prostate cancer (csPCa) indicate a more aggressive disease and higher mortality risk. This study develops a biopsy-based feature engineering model to assess csPCa risk and compares its performance to traditional PSA-based clinical models. Methods: This retrospective single-center study analyzed data from 824 patients undergoing transrectal prostate biopsy. The data preprocessing steps included one-hot encoding, standardization, interaction terms, and squared variables. Feature selection was performed using the Boruta algorithm. The dataset was divided into training and test sets, with patients classified into no cancer, very low-, low-, intermediate-, and high-risk groups. A binary Generalized Linear Model (GLM) was used to evaluate the relationship with csPCa. Model performance was assessed using the area under the receiver operating characteristic curve (AUC), and predictive capability was compared in terms of discrimination, calibration, and clinical utility. Results: Of the 824 patients, 320 (38.8%) were diagnosed with PCa, including 189 (22.9%) with csPCa. The GLM exhibited improved performance metrics, achieving an AUC of 0.833 compared to 0.721 for the PSA model and 0.794 for the combination of the PSA and digital rectal examination (DRE) model. The GLM showed a good fit and provided a greater netbenefit. The most significant predictors identified were PSA density, DRE, and PSA. Conclusion: This study employed feature engineering to identify clinical characteristics that predict csPCa in biopsy patients. The model demonstrated strong discriminatory ability and clinical utility; however, large-scale, multicenter studies are necessary to validate its effectiveness for clinical application.