Improved predictive performance of prostate biopsy collaborative group risk calculator when based on automated machine learning

Abstract

Background: The Prostate Biopsy Collaborative Group risk calculator (PBCG RC) has a moderate discriminatory capability. This study aimed to create automated machine learning (AutoML) PBCG RC for predicting the probability of any-grade and high-grade prostate cancer (PCa). Methods: This retrospective, single-center study was carried out using the database with 832 patients who were subject to transrectal ultrasound-guided prostate biopsy with prostate-specific antigen (PSA) values from 2 to 50 ng/ml. Information about PBCG RC predictors was gathered for all patients. We used H2O, as an open-source platform for AutoML, where the set of 20 base learning algorithms were trained. The AutoML PBCG RC was compared in terms of discrimination, calibration, and clinical utility with the original PBCG RC. Results: PCa was detected in 341 (41%) men, and 159 (19.1%) of them had high-grade PCa. Our AutoML models demonstrated better discriminative ability than the original PBCG RC for detection of PCa (area under the curve [AUC]: 0.703 vs 0.628; P = 0.023) and high-grade PCa (AUC: 0.990 vs 0.717; P < 0.001). The decision curve analyses showed that AutoML models performed better. For high-grade PCa the PSA was the most important feature. Conclusions: We applied ensemble techniques to create a freely available online PCa risk tool based on PBCG RC predictors and AutoML algorithms. The AutoML models drastically improved original model performance and the predictions of high-grade PCa were nearly perfect. However, new models should be used with a reserve, because external validation has not been performed yet.