Identifying the Presence of Prostate Cancer in Individuals with PSA Levels <20 ng ml-1 Using Computational Data Extraction Analysis of High Dimensional Peripheral Blood Flow Cytometric Phenotyping Data

Georgina Cosma; Stéphanie E McArdle; Stephen Reeder; Gemma A Foulds; Simon Hood; Masood Khan; A Graham Pockley

doi:10.3389/fimmu.2017.01771

Identifying the Presence of Prostate Cancer in Individuals with PSA Levels <20 ng ml^-1 Using Computational Data Extraction Analysis of High Dimensional Peripheral Blood Flow Cytometric Phenotyping Data

Front Immunol. 2017 Dec 18:8:1771. doi: 10.3389/fimmu.2017.01771. eCollection 2017.

Authors

Georgina Cosma¹, Stéphanie E McArdle², Stephen Reeder², Gemma A Foulds², Simon Hood², Masood Khan³, A Graham Pockley²

Affiliations

¹ School of Science and Technology, Nottingham Trent University, Nottingham, United Kingdom.
² John van Geest Cancer Research Centre, School of Science and Technology, Nottingham Trent University, Nottingham, United Kingdom.
³ University Hospitals of Leicester NHS Trust, Leicester, United Kingdom.

Abstract

Determining whether an asymptomatic individual with Prostate-Specific Antigen (PSA) levels below 20 ng ml^-1 has prostate cancer in the absence of definitive, biopsy-based evidence continues to present a significant challenge to clinicians who must decide whether such individuals with low PSA values have prostate cancer. Herein, we present an advanced computational data extraction approach which can identify the presence of prostate cancer in men with PSA levels <20 ng ml^-1 on the basis of peripheral blood immune cell profiles that have been generated using multi-parameter flow cytometry. Statistical analysis of immune phenotyping datasets relating to the presence and prevalence of key leukocyte populations in the peripheral blood, as generated from individuals undergoing routine tests for prostate cancer (including tissue biopsy) using multi-parametric flow cytometric analysis, was unable to identify significant relationships between leukocyte population profiles and the presence of benign disease (no prostate cancer) or prostate cancer. By contrast, a Genetic Algorithm computational approach identified a subset of five flow cytometry features (CD8⁺CD45RA^-CD27^-CD28^- (CD8⁺ Effector Memory cells); CD4⁺CD45RA^-CD27^-CD28^- (CD4⁺ Terminally Differentiated Effector Memory Cells re-expressing CD45RA); CD3^-CD19⁺ (B cells); CD3⁺CD56⁺CD8⁺CD4⁺ (NKT cells)) from a set of twenty features, which could potentially discriminate between benign disease and prostate cancer. These features were used to construct a prostate cancer prediction model using the k-Nearest-Neighbor classification algorithm. The proposed model, which takes as input the set of flow cytometry features, outperformed the predictive model which takes PSA values as input. Specifically, the flow cytometry-based model achieved Accuracy = 83.33%, AUC = 83.40%, and optimal ROC points of FPR = 16.13%, TPR = 82.93%, whereas the PSA-based model achieved Accuracy = 77.78%, AUC = 76.95%, and optimal ROC points of FPR = 29.03%, TPR = 82.93%. Combining PSA and flow cytometry predictors achieved Accuracy = 79.17%, AUC = 78.17% and optimal ROC points of FPR = 29.03%, TPR = 85.37%. The results demonstrate the value of computational intelligence-based approaches for interrogating immunophenotyping datasets and that combining peripheral blood phenotypic profiling with PSA levels improves diagnostic accuracy compared to using PSA test alone. These studies also demonstrate that the presence of cancer is reflected in changes in the peripheral blood immune phenotype profile which can be identified using computational analysis and interpretation of complex flow cytometry datasets.

Keywords: PSA level; computational analysis; flow cytometry; genetic algorithm; immunophenotyping data; machine learning; predictive modeling; prostate cancer.