UP-1 Predicting risk of disease progression during active surveillance for prostate cancer: analysis comparison of patient clinical features with a machine learning algorithm to the CANARY risk calculator
Thursday June 27, 2019 from
TBD
Presenter

M. Eric Hyndman, Canada

Urology

Department of Surgery

University of Calgary

Abstract

Predicting risk of disease progression during active surveillance for prostate cancer: Analysis comparison of patient clinical features with a machine learning algorithm to the CANARY risk calculator

Robert Paproski1, Stephanie Garbowski2, Tarek A Bismar3, Catalina Vasquez4, Perrin H Beatty1, John D Lewis4, M. Eric Hyndman2,5.

1Research and Development, Nanostics, Inc., Edmonton, AB, Canada; 2Prostate Cancer Centre, University of Calgary, Calgary, AB, Canada; 3Pathology & Laboratory Medicine and Oncology, University of Calgary, Cumming School of Medicine, Calgary, AB, Canada; 4Oncology, University of Alberta, Edmonton, AB, Canada; 5Surgery, University of Calgary, Calgary, AB, Canada

APCaRI. Alberta Cancer Foundation. Bird Dogs. Motorcycle Ride for Dad. Prostate Cancer Centre. Prostate Cancer Canada. University Hospital Foundation.

Introduction: Active surveillance (AS) is a common treatment for men with GG1 prostate cancer. Having accurate tools to identify men at risk for disease progression may reduce the number of biopsies needed. The Canary Prostate Active Surveillance Study Risk Calculator (PASS-RC) predicts the risk of reclassification from a biopsy. Machine learning (ML) algorithms, such as XGBoost, often have improved predictive accuracy over logistic regression(LR). Our aim was to determine the value of XGBoost and MRI features for predicting the progression of GG1 to GG2 PCa in AS.

Methods: A selected cohort of 139 men on an AS program in were included, with 2-10 year follow up.  Patients underwent an annual PSA, DRE, an MRI, and prostate biopsies at varying intervals depending on clinical risk. ML was performed using nested cross-validation repeated 10-times with different patient randomization using LR, and XGBoost algorithms using six predictive features: age, PSA, years since PCa diagnosis, the proportion of cores with PCa, number of PCa-free biopsies, and prostate volume. Additional MRI features were included in models and all models were compared to the Canary PASS-RC to predict progression to GG2 PCa.

Results: Using the six primary clinical features with ultrasound prostate volume, XGBoost outperformed logistic regression (Figure 1, area under the receiver operating characteristic curve (AUC) values 0.66 versus 0.72, p < 0.05). XGBoost models were further improved using MRI prostate volume versus ultrasound prostate volume (Figure 2, AUC 0.75 versus 0.72, p < 0.05). Incorporating prostate imaging reporting and data scoring (PI-RADS) and PSA density further improved XGBoost models (Figure 3; AUC 0.76) which were higher than the Canary PASS model (AUC 0.74).

Conclusions: Clinical risk calculators are useful tools in predicting PCa upgrading in AS and our risk calculator provided highly accurate prediction and may help reduce the number of serial biopsies performed for men with GG1 disease.


© 2022 CUA 74th Annual Meeting