A visualized machine learning model using noninvasive parameters to differentiate men with and without prostatic carcinoma before biopsy

Abstract

Abstract This study aimed to create a visualized extreme gradient boosting (XGBOOST) model to distinguish prostatic carcinoma (PCA) from non-PCA using noninvasive prebiopsy parameters before biopsy. This was a cross-sectional study of 310 Chinese men who underwent prostate biopsy and were divided into PCA (n = 126) and non-PCA (n = 184) groups. The non-PCA patients were diagnosed with benign prostatic hyperplasia (BPH) based on biopsy results. The XGBOOST model was used to analyze 15 noninvasive prebiopsy parameters. The model performance was assessed by the area under the receiver operating characteristic curve (AUC) and compared with four other machine learning models (decision tree learning, lasso, neural network (NNET), and support vector machine (SVM)) and a logistic model. The logistic model identified serum thymidine kinase 1 (STK1p), total prostate-specific antigen (TPSA), and age as key prognostic factors. In the Lasso procedure, free prostate-specific antigen (FPSA) and free-to-total prostate-specific antigen (FTPSA) were also added to machine learning models. The XGBOOST model achieved an AUC of 0.965, which was significantly greater than those of other models (AUC = 0.708–0.817) and the logistic model (AUC = 0.813) (P < 0.001). The 49 decision trees generated by the XGBOOST model were visualized to aid in decision making. This study successfully developed a visualized XGBOOST model with high accuracy in differentiating PCA from non-PCA using eight noninvasive predictors. This model could aid in the precise selection of high-risk PCA patients for biopsy, potentially minimizing unnecessary procedures and their associated costs.