The detection of clinically significant prostate cancer (csPca) poses a critical quality-of-care challenge, with multiparametric MRI recognized for enhancing csPca detection. Recent studies suggest that micro-ultrasound (micro-US) is equally effective in identifying csPca compared to MRI. This study aimed to develop a machine learning (ML) model to predict the probability of csPca, thereby aiding clinical decision-making. We analyzed a prospective database of biopsy-naïve patients who underwent biopsy at our tertiary referral center, where micro-US sonography was employed. The type of biopsy—whether target-only or combined target and systematic—was determined by the surgeon’s expertise and imaging reports. CsPca was defined as any Gleason group greater than 1. Our primary objective was to create a preoperative model for predicting csPca, while the secondary goal was to compare ML model performance against traditional logistic regression (LR). We evaluated three ML algorithms: random forest, extreme gradient boosting (XGBoost), and light gradient boosting machine (LightGBM). In our analysis of 848 patients, the LR model, incorporating factors like PSA density, suspicious lesions detected via micro-US and MRI, age, positive digital rectal examination, and biopsy type, achieved an accuracy of 71% and an AUC of 0.77. Notably, all ML models outperformed the LR model, with random forest achieving the highest accuracy of 86% and an AUC of 0.95. By integrating clinical features, serum biomarkers, and imaging findings, we have established a robust point-of-care model for predicting csPca, which could significantly improve clinical decision-making regarding biopsy approaches. Further external validation is necessary to confirm these findings.
November 4, 2024
