Presentation Authors: Daniele Panfilo, Trieste, Italy, Cosimo De Nunzio*, rome, Italy, Antonio Luigi Pastore, Latina, Italy, Sebastiano Saccani, Boudewijnb Alexander, Pietro Tortella, Trieste, Italy, Manuela Mattioli, Riccardo Lombardo, rome, Italy, Antonio Carbone, Andrea Fuschi, Latina, Italy, Lorenzo Dutto, Glasgow, United Kingdom, Joern Witt, Gronau, Germany, Eric Medvet, Trieste, Italy, Andrea Tubaro, rome, Italy
Introduction: Several nomograms have been developed to predict prostate cancer upgrading however very little artificial intelligence (AI) tools are available for this purpose. Aim of our study was to test the feasibility of AI to predict upgrading in patients with PCa undergoing robotic radical prostatectomy. Materials & methods
Methods: Materials & methods Between 2012 and 2017, a consecutive series of patients with prostate cancer were treated with robotic radical prostatectomy (RRP) and lymph node dissection in a single center. Demographic, clinical and histological characteristics of the patients were recorded. Histological specimens were graded according to the new PGG classification. Two different methods of machine learning (AI) were used (binary tree classifier and random forrest) to predict upgrading risk using the following variables: age, BMI, number of cores, number of positive cores, percentage of cancer, primary gleason, secondary gleason, total PSA, PSA density, ASA score, prostate volume, transitional zone volume and clinical stage. Area under the curve and net benefit of classification tree and random forrest were then compared to the logistic model. Clinically significant upgrading was defined as: Gleasonâ‰¤ 3+4 to Gleasonâ‰¥ 4+3 and from Gleason â‰¤4+3 to Gleasonâ‰¥8.
Results: Overall 8357 patients were enrolled. Median age was 64 (60/70) years, median BMI was 27 (25/29) kg/m^2, mean prostate volume was 38 (29/50) ml and median PSA was 7.4 (5.5/11) ng/ml. Overall On RRP 1629/8357 (19%) presented a clinically significant upgrading. On RRP 3133 /8357 presented highr grade cancer (GSï‚³4+3) and 3223 advanced pathological stage (ï‚³pT3a). Overall the classification tree (AUC:0,76) and the random forrest (AUC: 0,78) outperformed the logistic model (0,67) in terms of accuracy (Figure 1)
Conclusions: Upgrading is a common issue in our cohort, arising in one out of five treated patients. Powerful predictive tools are therefore needed to help identify patients with an increased risk of upgrading. Such tools can play a pivotal role in patientsâ€™ diagnostic and treatment pathways. In this study, we have shown that artificial intelligence can outperform traditional Nomograms. Nonetheless, external validation is still required before clinical implementation.