Integrated Machine Learning Approaches Highlight the Heterogeneity of Human Myeloid-Derived Suppressor Cells in Acute Sepsis

Abstract

Highly heterogeneous cell populations require multiple flow cytometric markers for appropriate phenotypic characterization. This exponentially increases the complexity of 2D scatter plot analysis and exacerbates human errors due to variations in manual gating of flow data. We describe a workflow involving the stepwise integration of several, newly available machine learning tools for the analysis of myeloid-derived suppressor cells (MDSCs) in septic and non-septic critical illness. Unsupervised clustering of flow cytometric data showed good correlation with, but significantly different numbers of, MDSCs as compared with the cell numbers obtained by manual gating. However, both quantification methods revealed a significant difference between numbers of PMN-MDSC at day 1 in healthy volunteers and critically ill patients having septic or non-septic illness. Numbers of PMN-MDSC obtained by machine learning positively correlated with 30 days hospital readmission following critical illness, whereas manual gating of this cell population distinguished between septic and non-septic critical illness. Neither gating strategy found a correlation between number of MDSCs and 30-day mortality or hospital length of stay.