Abstract 918: Using machine learning to predict the risk of either having an aggressive form of prostate cancer (PCa) or lower-grade PCa/benign prostatic hyperplasia (BPH) based upon the flow cytometry immunophenotyping of myeloid-derived suppressor cells (MDSCs) and lymphocyte cell populations

Abstract

Abstract The goal of this study was to create a non-invasive confirmatory test for prostate biopsies that objectively analyzes flow cytometry data using machine learning to predict whether a subject is at higher risk for having an aggressive form of prostate cancer (PCa; Gleason ≥ 4+3). The commonly used assay for prostate screening is a prostate specific antigen (PSA) blood test, but due to prostate physiology, PSA testing results in a large frequency of false positives leading to numerous men each year undergoing unnecessary prostate biopsy procedures. Here, we use machine learning to create a neural network (NN) to predict whether a subject has a greater probability in having an aggressive form of prostate cancer (HR-PCa) or is at lower risk (LR-PCa; Gleason < 4+3) or no risk (benign prostatic hyperplasia [BPH]/healthy donor [HD]) based upon the immunophenotyping of myeloid-derived suppressor cells (MDSCs) and various lymphocyte cell populations. MDSCs are known to be key contributors in supporting tumor progression and escape through their ability to suppress anti-tumor responses mediated through T cell and natural killer (NK) cell activity. We used standard multiparametric flow cytometry techniques to immunophenotype the MDSCs and lymphocyte cell populations found in the peripheral blood of 114 biopsy-confirmed PCa and 89 biopsy-confirmed BPH subjects along with 116 healthy donors (HD). Subjects were recommended for biopsy by their physician and excluded if they had a previous history of cancer (not including subjects under active surveillance), had a medical intervention for prostate cancer, or were receiving a dihydrotestosterone (DHT) or alpha-1 blocker for active treatment of benign prostatic hyperplasia (BPH). Machine learning and pattern recognition were used to create a two-network approach to predict whether a subject should be recommended for biopsy (HR-PCa) or be actively monitored (HD or BPH/LR-PCa). Initially, NN1 predicts whether the sample looks more like HD or HR-PCa; if the sample is predicted to be a possible HR-PCa, then it is tested by NN2 and predicted as either BPH/LR-PCa or HR-PCa. The final decision for biopsy would be made by the clinician. By combining immunophenotyping data with machine learning, we achieved a final 90% sensitivity for predicting whether subjects should undergo a prostate biopsy procedure based upon their immunophenotyping. For this study, out of the 203 subjects recommend for biopsy, approximately 105 procedures could have been avoided. In a clinical setting, we believe that this technology, in use with other known clinical risk factors, would allow for clinicians to have a more informed decision when recommending their patients for a prostate biopsy procedure. Citation Format: George A. Dominguez, John Roop, Alexander Polo, Anthony Campisi, Dmitry I. Gabrilovich, Amit Kumar. Using machine learning to predict the risk of either having an aggressive form of prostate cancer (PCa) or lower-grade PCa/benign prostatic hyperplasia (BPH) based upon the flow cytometry immunophenotyping of myeloid-derived suppressor cells (MDSCs) and lymphocyte cell populations [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2019; 2019 Mar 29-Apr 3; Atlanta, GA. Philadelphia (PA): AACR; Cancer Res 2019;79(13 Suppl):Abstract nr 918.