UP-9 A machine learning approach to predicting progression on active surveillance for prostate cancer
Thursday June 27, 2019 from
TBD
Award Winner
Madhur Nayan, United States has been granted the
Presenter

Madhur Nayan, United States

Fellow in Urologic Oncology

Department of Urology

Massachusetts General Hospital

Abstract

A machine learning approach to predicting progression on active surveillance for prostate cancer

Madhur Nayan1, Keyan Salari1,2, Anthony Bozzo3, Wolfgang Ganglberger4, Gordon Lu1, Filipe Carvalho1, Andrew Gusev1, Brandon Westover4, Adam S Feldman1.

1Department of Urology, Massachusetts General Hospital, Boston, MA, United States; 2Broad Institute of Harvard and MIT, Cambridge, MA, United States; 3Division of Orthopaedic Surgery, Department of Surgery, McMaster University, Hamilton, MA, United States; 4Department of Neurology, Massachusetts General Hospital, Boston, MA, United States

Introduction: To date, studies that have developed models to predict progression on AS for prostate cancer have invariably used traditional statistical approaches. We evaluated whether a machine learning approach could improve prediction of progression on AS.

Methods: We performed a retrospective institutional cohort study of 790 very-low or low-risk prostate cancer patients managed with AS. The sample was split into a training and test set (ratio 80%/20%). In the training set, we developed a traditional logistic regression classifier (LRC), and alternate machine learning classifiers (MLCs) (support vector machine, random forest, and a full connected artificial neural network) to predict grade progression. Features considered for inclusion were clinical and biopsy characteristics measured at diagnosis, as well as time between diagnostic biopsy and last biopsy, and number of biopsies on surveillance. We used backward elimination to select features for the multivariable LRC. For the MLCs, all features were included in model development. We tuned the hyperparameters of the MLCs. Model performance was evaluated in the test set. The primary performance metric was the F1 score. Other performance metrics included sensitivity, specificity, positive predictive value, and negative predictive value.

Results: With a median follow-up of 6.3 years, 234 developed grade-progression. In descending order, the F1 scores were: support vector machine 0.600 (95% CI 0.593–0.605), artificial neural network 0.507 (95% CI 0.500–0.511), random forest 0.413 (95% CI 0.400–0.418), traditional LRC 0.182 (95% CI 0.151–0.185). All MLCs had a significantly higher F1 score than the traditional LRC (all p<0.001). Compared to the MLCs, the traditional LRC had relatively lower sensitivity and negative predictive value, but higher specificity and positive predictive value.

Conclusions: Alternative MLCs significantly outperformed a traditional LRC in predicting progression on AS for prostate cancer. 

 


© 2022 CUA 74th Annual Meeting