Predicting biochemical recurrence after prostatectomy: Can machine learning beat CAPRA score? Results of a multicentric retrospective analysis on 4,700 patients.

Abstract

343 Background: Recent advances in machine learning algorithms and deep learning solutions paved the way for improved accuracy in survival analysis. We aim to investigate the accuracy of conventional machine learning and deep learning methods in the prediction of 3-year biochemical recurrence (BCR) as compared to CAPRA score prediction. Methods: A total of 5043 men who underwent RP between 2000 and 2015 for clinically localized PCa iwere analyzed retrospectively. Three-year BCR was predicted using the following models: CAPRA score, Cox regression analysis, logistic regression, k-nearest neighbor, random forest and densely connected feed-forward neural network classifier. The discrimination of the models was quantified using the C-index or the area under the receiver operating characteristics curve. Results: Patients with CAPRA score 2 and 3 accounted for 64% of the population. C-index measuring performance for the prediction of the three-year BCR for CAPRA score was 0.63. C-index values for k-neighbor classifier, logistic regression, Cox regression analysis, random forest classifier and densely optimized neural network were respectively 0.55, 0.63, 0.64, 0.64 and 0.70 (pairwise, adjusted p-value < 0.01). After inclusion of available post-surgical variables, C-index value reached respectively 0.58, 0.77, 0.74, 0.75 and 0.84 (pairwise, adjusted p-value < 0.05). Conclusions: Our results show that CAPRA score performed poorly in intermediate-risk patients undergoing RP. Densely connected neural networks with simple architecture further increased predictive power with low computational cost. In order to predict 3-years BCR, adding post-surgical features to the model greatly enhanced its performance.