MP49-02 DEVELOPMENT AND EXTERNAL VALIDATION OF INTERPRETABLE MACHINE LEARNING MODELS FOR CLINICALLY SIGNIFICANT PROSTATE CANCER DIAGNOSIS IN PATIENTS WITH LESIONS OF PI-RADS V2.1 SCORE ≥3

Mingjian Ruan,Yi Liu,Kaifeng Yao,Kexin Wang,Xiaoying Wang,Shiliang Wu

DOI: https://doi.org/10.1097/01.ju.0001008696.31772.28.02

2024-01-01

Abstract:You have accessJournal of UrologyProstate Cancer: Detection & Screening IV (MP49)1 May 2024MP49-02 DEVELOPMENT AND EXTERNAL VALIDATION OF INTERPRETABLE MACHINE LEARNING MODELS FOR CLINICALLY SIGNIFICANT PROSTATE CANCER DIAGNOSIS IN PATIENTS WITH LESIONS OF PI-RADS V2.1 SCORE ≥3 Mingjian Ruan, Yi Liu, Kaifeng Yao, Kexin Wang, Xiaoying Wang, and Shiliang Wu Mingjian RuanMingjian Ruan , Yi LiuYi Liu , Kaifeng YaoKaifeng Yao , Kexin WangKexin Wang , Xiaoying WangXiaoying Wang , and Shiliang WuShiliang Wu View All Author Informationhttps://doi.org/10.1097/01.JU.0001008696.31772.28.02AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: For patients with lesions of PI-RADS v2.1 score≥3, prostate biopsy is almost a given. However, the low positive rate suggests that clinically significant prostate cancer (csPCa) risk assessments for those patients needs improvement. Our aim is to develop and validate machine learning (ML) models based on clinical and imaging parameters for the prediction of csPCa in patients with lesions of PI-RADS v2.1 ≥3. METHODS: A total of 1083 cases with PI-RADS v2.1 ≥3 lesion included retrospectively, randomly split into training (70%, N=759) and internal validation (30%, N=324). And 147 eligible cases were enrolled prospectively for external validation. Five different ML models were developed using the results of univariate and multivariate logistic regression analyses for independent csPCa predictors: logistic regression, extreme gradient boosting, random forest (RF), decision tree, and support vector machines. The area under the curve (AUC) and Brier score were utilized to evaluate the performance of various models. RESULTS: The csPCa detection rates were 67.5% (512/759), 64.8% (210/324) and 57.8% (85/147) in the training, internal and external validation cohorts, respectively. Based on age, prostate specific antigen, prostate volume, lesion diameter, PI-RADS v2.1 score and lesions location, five ML models were developed. In the three cohorts, the RF model displayed the largest mean AUC (0.926) and the lowest mean Brier score (0.103). At threshold of 0.450 (99% sensitivity of the RF model in the training cohort), the sensitivity and negative predictive values were 93.8-97.6% and 82.7-95.1% in the internal and external validation cohorts, respectively. And prostate biopsies could be avoided in 23.1-27.9% of cases. CONCLUSIONS: The RF model is a superior tool for the prediction of csPCa among patients with lesions of PI-RADS v2.1≥3 compared to other ML models. This improvement in csPCa prediction might help to avoid unnecessary prostate biopsies. Download PPTDownload PPT Source of Funding: Interdisciplinary clinical research project of Peking University First Hospital (2022IR27) © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e782 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Mingjian Ruan More articles by this author Yi Liu More articles by this author Kaifeng Yao More articles by this author Kexin Wang More articles by this author Xiaoying Wang More articles by this author Shiliang Wu More articles by this author Expand All Advertisement PDF downloadLoading ...