Automated Gating of CD34 + Cells in Cord Blood: Performance Evaluation of a Machine Learning-Based ISHAGE Protocol

Abstract

Precise quantification of cellular subsets is fundamental for qualifying grafts and supporting emerging therapies. CD34⁺ enumeration in cord blood using the ISHAGE protocol exemplifies the operator variability inherent to manual gating. We evaluated whether a machine-learning approach could provide standardized automated enumeration and reduce variability. A machine-learning–based automatic gating algorithm was trained on 29 manually gated FCS files and applied to raw flow cytometry data. Performance was compared with manual gating from nine laboratories from a previously published multicenter study using Z-scores, rank positioning, absolute deviation, correlations, Bland–Altman analysis, and intraclass correlation coefficients. Across 12 samples, AI1 remained within ± 2 SD of the human consensus in all cases, whereas AI2 exceeded this threshold in two. AI1 consistently ranked closer to the human median and showed narrower deviations. Both models correlated strongly with manual gating (AI1: r = 0.991; AI2: r = 0.968). Bland–Altman analysis showed minimal bias and narrow limits of agreement for AI1 versus its human reference, while AI2 and human–human comparisons displayed greater variability. ICCs indicated high reliability across all comparisons, with the strongest agreement observed for AI1 versus Lab1 (ICC = 0.995). A machine learning–based automatic gating approach can reproduce expert CD34⁺ enumeration with high fidelity. By reducing operator-dependent variability, this method may strengthen cytometry standardization across cord blood banking and broader cellular therapy workflows.