Session: MP55: Prostate Cancer: Detection & Screening III
MP55-07: Can Neuronal Network-Based Machine Learning Predict Pathologically Significant Prostate Cancer in Patients Diagnosed with Gleason Grade 6 on Pre-operative Prostate Biopsy.
Introduction: The management of PCa is based on a risk stratification where in the case of localized disease the patients are elected for either active surveillance or an invasive treatment. This risk stratification system is based on three main parameters: PSA level, Gleason grade and the clinical stage of the disease. Despite the clinical definition of a low-risk disease, in certain cases the disease biology is actually aggressive with a potential to become a significant disease or in other cases a sampling error has undegraded the true disease pathology. The aim of our study was to determine the probability of Gleason score upgrading to a clinically significant disease using neuronal network machine learning. Methods: Using our local radical prostatectomy registry, we have retrospectively reviewed patients' medical records for Age, BMI, PSA level at diagnosis, prostate biopsy findings, prostate volume, PSA-D, D'Amico risk classification, Gleason grade group and pathological staging following prostate biopsy and surgery. Clinically significant (CS) PCa was defined as Gleason 7(3+4) and higher (= GG 2). Univariant and multivariant analyses were used to predict significant PCa at radical prostatectomy specimens. Fisher exact and Mann-Whitney U tests were used for categorical and continuous variables, respectively. Next, a neural network (NN) machine learning model was trained using the above variables in order to predict the probability of pathologically upgrading to a CS disease. A back propagation algorithm was used for validation of the NN. Results: 428 patients underwent robotic-assisted radical prostatectomy between 2012 and 2021 in our institution. 72 patients (17%) had a clinically insignificant disease determined by prostate biopsy. of which 34 patients (72%) were upgraded to a CS disease on the final pathology. On univariate and multivariate analysis, the only preoperative variable that predicted a CS final pathology was the PSA-density (PSAD) with a sensitivity and specificity of 36% and 74% respectively. When using the NN machine learning the probability to predict disease upgrading was more accurate with a sensitivity and specificity of 94.1% and 97.1% respectively. Conclusions: NN machine learning can be trained and used for predicting the probability of PCa upgrading in patients with a biopsy proven Gleason 6 disease. In our cohort the NN sensitivity and specificity were superior compared to the traditional descriptive statistics. SOURCE OF Funding: None
