A REFINED RISK GROUP SYSTEM USING MACHINE LEARNING FOR PROSTATE CANCER TREATED WITH RADIATION THERAPY

Background: Risk adapted treatment guided by risk group stratification have been broadly accepted to treated prostate cancer (PCa). We developed an Artificial Neural Network (ANN) to refine the risk stratification group from NCCN in patients with PCa treated with radiotherapy (RT). Methods: We used information on 485 patients with PCa treated by RT (74-78Gy), during 2010-2017. The ANN were used for identifying crucial predictive factors for biochemical recurrence (BCR) and to refine the NCCN risk group. After selecting the major and minor drivers from ANN analysis, a risk group was built, and tested by univariate;multivariate analysis. Results: The median follow-up time was 56 months, the 5-year bRFS for the entire cohort was 86%. ANN identified PSA>50, Gleason 9-10 and T3-T4 as major drivers for high-risk (HR) PCa. For intermediate risk (IR) two majors (Gleason 7 (4+3); PSA 10-20) and minor drivers (positive % of biopsy (PPB)>50% and T2b-c) were identified. The accuracy to detect the BCR was 94% and the AUC was 97%. Combining these drivers, five subgroups: low-, low intermediate risk (LIR), high intermediate risk (HIR), standard high risk (SHR) and very high-risk (VHR) PCa were created. The 5-years bRFS for LR-, LIR-, HIR-, SHR- and VHR-PCa 96%, 90%, 79%, 75%, and 63% (p<0.0001). In multivariate analysis the new system was the only independent prognostic factor for bRFS (p=0.027). Conclusion: ANN was capable to stratify PCa with excellent accuracy in five risk groups with markedly distinct prognoses. This finding can be useful for risk adapted treatment.