PD38-08 AVOIDING UNNECESSARY TARGETED PROSTATE BIOPSIES USING MACHINE LEARNING

Abstract

You have accessJournal of UrologyCME1 Apr 2023PD38-08 AVOIDING UNNECESSARY TARGETED PROSTATE BIOPSIES USING MACHINE LEARNING Roxana Esmaili, Ghazal Khajir, Michael Leapman, Preston Sprenkle, Darryl Martin, and John Onofrey Roxana EsmailiRoxana Esmaili More articles by this author , Ghazal KhajirGhazal Khajir More articles by this author , Michael LeapmanMichael Leapman More articles by this author , Preston SprenklePreston Sprenkle More articles by this author , Darryl MartinDarryl Martin More articles by this author , and John OnofreyJohn Onofrey More articles by this author View All Author Informationhttps://doi.org/10.1097/JU.0000000000003336.08AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Magnetic resonance imaging (MRI)-targeted transrectal ultrasound (TRUS) guided needle biopsy is the current gold standard for diagnosing prostate cancer in patients with suspicion for prostate cancer. This leads to a frequent needle biopsy to determine the presence of clinically significant prostate cancer (csPCa) with a high rate of invasive overdiagnosis in patients with non-aggressive cancer. Supervised machine learning (ML) algorithms provide a mathematical framework to make predictions about csPCa in order to make personalized decisions to perform a biopsy or not and avoid unnecessary targeted needle biopsies. METHODS: From January 2016 to February 2020, 1099 patients underwent multiparametric MRI before TRUS fusion biopsy. We included PI-RADS score (version 2), gland volume, lesion region of interest volume before biopsy, PSA value, PSA density, and patient age at the time of biopsy as input variables to calculate the probability of csPCa for each lesion for each patient before biopsy using different ML classification algorithms: logistic regression (LR); random forest (RF); naïve Bayes (NB); and nearest neighbor (NN). Gleason Grade Group ≥2 was used as a reference for the presence of csPCa. Ten-fold cross-validation stratified over patients validated the ML approaches and compared to univariate models. RESULTS: Area under the receiver operating characteristic curves (AUC) and net reduction curves show that multivariate LR (AUC 81.3%) outperformed RF (AUC 79.2%), NN (AUC 71.8%) and NB (AUC 78.3%) and outperformed standard univariate LR using PSA density (AUC 71.4%) and PI-RADS (AUC 73.9%). The algorithms were evaluated at a point on the ROC curve for high sensitivity. Using the multivariate LR model, 636 (41.38%) targeted biopsies could have been avoided while missing only 90 (5.86%) csPCa, which is reasonable with a threshold probability of 25% and a net benefit of 0.21. CONCLUSIONS: Our findings suggest that multivariate ML models using clinical data can be used to avoid unnecessary targeted biopsy cores among patients with suspicion of prostate cancer undergoing prostate biopsy. Source of Funding: JAO supported by NIH R42 CA224888. DTM supported by DoD W81XWH-19-PCRP-IDA. © 2023 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 209Issue Supplement 4April 2023Page: e996 Advertisement Copyright & Permissions© 2023 by American Urological Association Education and Research, Inc.MetricsAuthor Information Roxana Esmaili More articles by this author Ghazal Khajir More articles by this author Michael Leapman More articles by this author Preston Sprenkle More articles by this author Darryl Martin More articles by this author John Onofrey More articles by this author Expand All Advertisement PDF downloadLoading ...