MP31-02 MACHINE LEARNING OUTPERFORM TRADITIONAL APPROACHES IN PREDICTING CLINICALLY SIGNIFICANT PROSTATE CANCER

Le résumé fourni par la source

You have accessJournal of UrologyProstate Cancer: Detection & Screening III (MP31)1 May 2024MP31-02 MACHINE LEARNING OUTPERFORM TRADITIONAL APPROACHES IN PREDICTING CLINICALLY SIGNIFICANT PROSTATE CANCER Flavio Vasconcelos Ordones, Lodewikus Vermeulen, Ali Hooshyari, David Scholtz, Paulo Kawano, Gustavo Modelli de Andrade, Abner Barros, and Peter Gilling Flavio Vasconcelos OrdonesFlavio Vasconcelos Ordones , Lodewikus VermeulenLodewikus Vermeulen , Ali HooshyariAli Hooshyari , David ScholtzDavid Scholtz , Paulo KawanoPaulo Kawano , Gustavo Modelli de AndradeGustavo Modelli de Andrade , Abner BarrosAbner Barros , and Peter GillingPeter Gilling View All Author Informationhttps://doi.org/10.1097/01.JU.0001008936.35187.0b.02AboutPDF ToolsAdd to favoritesDownload CitationsTrack CitationsPermissionsReprints ShareFacebookLinked InTwitterEmail Abstract INTRODUCTION AND OBJECTIVE: Machine Learning models offer the potential to improve clinical predictions. While magnetic resonance images (MRI) have enhanced the prediction of clinically significant prostate cancer (CSPCa), accuracy remains a challenge. This study aims to predict CSPCa more accurately using machine learning models. METHODS: This is a comprehensive evaluation with 354 patients who underwent transperineal prostate biopsy (TPPB) at a single center in Tauranga-NZ. Demographics and clinical data (PSA , PSA density and PiRADS Score) were analyzed . Patients with PiRADS 4 and 5 had their index lesion size measured in millimeters. For internal validation, we divided the cohort into training (80%) and testing (20%) sets. Our predictive variables included PSA density (PSAd) , PI-RADS score, PI-RADS index lesion size, previous biopsy results, BMI, and age. CSPCa (Gleason≥7) was the outcome. We employed three machine learning models: Lasso regression, XGBoost, and LightGBM. To optimize performance, we employed 10-fold cross-validation for hyperparameter tuning. Evaluation of model performance was done and relevant accuracy metrics assessed. For robust results, we calculated the confidence interval (CI) for the area under the curve (AUC) using 2000 stratified bootstrap replicates. R version 4.1.2 software was utilized. RESULTS: 56% (n=199) of patients were diagnosed CSPCa. Comparing patients with CSPCa to those without, the former group was characterized by older age (68 [64-74] versus 64 [60-68] years, p<0.001), higher PSAd values (0.24 [0.16-0.36] versus 0.15 [0.12-0.20], p<0.001), and higher PI-RADS scores. Our machine learning models yielded the following results (accuracy and ROC-AUC): Lasso regression - accuracy 0.73 and ROC-AUC 0.80 [95% CI: 0.70-0.90]; XGBoost - accuracy 0.74 and ROC-AUC 0.80 [95% CI: 0.69-0.90]; LightGBM - accuracy 0.75 and ROC-AUC 0.83 [95% CI: 0.72-0.91]. Although ROC-, LightGBM had superior discriminant and calibration metrics and was selected as the final model. The best predictors identified were the size of the MRI index lesion size, PSA density, age, and PI-RADS Score. When classifying CSPCa, using PIRADS alone and PSA density resulted in ROC-AUC values of 0.79 [95% CI: 0.69-0.89] and 0.65 [95% CI: 0.52-0.78], respectively. CONCLUSIONS: Combining clinical parameters with PiRads (Score and lesion size) and PSAd significantly improves the performance of machine learning models in predicting patients with CSPCa. Our model achieved higher ROC-AUC values and demonstrated excellent calibration, indicating its superior predictive ability. Source of Funding: Nil © 2024 by American Urological Association Education and Research, Inc.FiguresReferencesRelatedDetails Volume 211Issue 5SMay 2024Page: e503 Advertisement Copyright & Permissions© 2024 by American Urological Association Education and Research, Inc.Metrics Author Information Flavio Vasconcelos Ordones More articles by this author Lodewikus Vermeulen More articles by this author Ali Hooshyari More articles by this author David Scholtz More articles by this author Paulo Kawano More articles by this author Gustavo Modelli de Andrade More articles by this author Abner Barros More articles by this author Peter Gilling More articles by this author Expand All Advertisement PDF downloadLoading ...

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.