Abstract
In prostate cancer (PCa), risk calculators have been proposed, relying on clinical parameters and magnetic resonance imaging (MRI) enable early prediction of clinically significant cancer (CsPCa). The prostate imaging-reporting and data system (PI-RADS) is combined with clinical variables predominantly based on logistic regression models. This study explores modeling using regularization techniques such as ridge regression, LASSO, elastic net, classification tree, tree ensemble models like random forest or XGBoost, and neural networks to predict CsPCa in a dataset of 4799 patients in Catalonia (Spain). An 80-20% split was employed for training and validation. We used predictor variables such as age, prostate-specific antigen (PSA), prostate volume, PSA density (PSAD), digital rectal exam (DRE) findings, family history of PCa, a previous negative biopsy, and PI-RADS categories. When considering a sensitivity of 0.9, in the validation set, the XGBoost model outperforms others with a specificity of 0.640, followed closely by random forest (0.638), neural network (0.634), and logistic regression (0.620). In terms of clinical utility, for a 10% missclassification of CsPCa, XGBoost can avoid 41.77% of unnecessary biopsies, followed closely by random forest (41.67%) and neural networks (41.46%), while logistic regression has a lower rate of 40.62%. Using SHAP values for model explainability, PI-RADS emerges as the most influential risk factor, particularly for individuals with PI-RADS 4 and 5. Additionally, a positive digital rectal examination (DRE) or family history of prostate cancer proves highly influential for certain individuals, while a previous negative biopsy serves as a protective factor for others.