Abstract
Prostate cancer (PCa) screening and detection relies heavily upon prostate-specific antigen (PSA) testing, but PSA testing has a high rate of false positives, leading to increased risks for overdiagnosis and overtreatment; thus, additional blood-based biomarkers for PCa detection are needed. Flow cytometry-based immunophenotyping of peripheral blood is an accessible and noninvasive technology, but as more parameters are included, new computational methods must be developed for the efficient analysis and utilization of these large datasets for clinical applications. Machine learning algorithms, specifically pattern recognition neural networks (PRNNs), have the potential to assist in these types of analyses, but the flow cytometry data need to be transformed into a usable input format. The goal of this study was to use our newly developed “hypervoxelation of cytometry events” computational technique, or HyperVOX, to transform flow cytometry data into a useable format for input into a series of PRNNs to detect PCa of all Gleason scores (GS) from circulating immune cells. We used standard multiparametric flow cytometry techniques to measure 16 different myeloid and lymphoid cell populations found in the peripheral blood of 156 biopsy-confirmed PCa (GS6 n = 59, GS7 n = 68, GS8 n = 12, GS9 n = 16, and GS10 n = 1; median age = 68 ± 8.7 years) along with 99 male healthy donors (HD) (median age = 53 ± 8.5 years). Flow cytometry data were then transformed using HyperVOX in order to create hypervoxels that can be used as the common feature across all samples. Briefly, each channel was used as an axis in a multidimensional space and divided into four segments, with each event being defined by its location within each segment of each axis. The resulting count of events that fall within each hypervoxel for each sample is then used as the input for the PRNN. With this, a screening-type assay was developed to detect PCa compared against HD. PRNNs were trained using raw flow cytometry data processed using HyperVOX from 97 PCa patients and 67 HD controls. Predictions were evaluated using the performance of the trained PRNNs on 59 PCa patients and 32 HD that were not used for PRNN training (holdout test set). The PRNN classified 28 out of 32 HD and 57 out of 59 PCa samples correctly, resulting in a sensitivity of 96.6% (95% CI, 88.3–99.6), specificity of 87.5% (95% CI, 71.0–96.5), positive predictive value (PPV) of 93.4% (95% CI, 85.1–98.2), negative predictive value (NPV) of 93.3% (95% CI, 78.1–98.2), and an AUC of 0.9656 (95% CI, 0.9202–1). Upon Gleason score stratification, the NN classified 27 out of 28 GS6, 18 out of 19 GS7, and 11 out of 11 >GS7 samples correctly. In a clinical setting, this technology would improve PCa detection and allow clinicians to have a more informed decision when recommending their patients for a prostate biopsy procedure and subsequent medical interventions to help reduce overdiagnosis and overtreatment.
Citation Format: George A. Dominguez, John Roop, Alexander Polo, Anthony Campisi, Dmitry I. Gabrilovich, Amit Kumar. Using pattern recognition neural networks to detect prostate cancer: A new method to analyze flow cytometry-based immunophenotyping using machine learning [abstract]. In: Proceedings of the AACR Special Conference on Advances in Liquid Biopsies; Jan 13-16, 2020; Miami, FL. Philadelphia (PA): AACR; Clin Cancer Res 2020;26(11_Suppl):Abstract nr B50.