Characterization of the Sequence of Cytopenia after Total-body Irradiation using Unsupervised Machine Learning.

Abstract

Hematopoietic acute radiation syndrome (H-ARS) elicits multidimensional effects, as total-body irradiation (TBI) induced myelosuppression results in distinct patterns of loss and recovery across blood cell lineages. This study aimed to characterize the sequence of pancytopenia after TBI in male and female non-human primates (NHPs) with and without sargramostim treatment using unsupervised machine learning (ML). Longitudinal blood cell measures from three 60-day randomized, controlled trials of NHPs (N = 501) exposed to TBI and administered either vehicle (n = 173) or sargramostim (n = 328) were pooled for analysis. Hierarchical clustering of blood cells was performed using dynamic time warping (DTW) distances. Summary metrics of longitudinal trends were compared across treatment groups and sex. All analyses were performed in Pumas 2.0. Trends in each cluster showed that lymphocytes and reticulocytes are the first two cell populations to deplete rapidly, almost immediately after TBI, followed by white blood cell (WBC) related cells, platelets, and red blood cell (RBC) related cells. Reticulocytes reach nadir at day 8, followed by WBC-related cells at day 12, platelets at day 13, and RBC-related cells at day 17. The characterized pancytopenia sequence aligns with the typical lifespan of impaired blood cells, except for RBCs. Comparison across clusters between treated and untreated NHPs showed that sargramostim-treated animals started to recover earlier or at the same time as untreated animals. Female NHPs generally experienced deeper nadirs than males. In summary, characterizing pancytopenia using unsupervised machine learning improves understanding of the pathogenesis of H-ARS, which can enable improved therapeutic care, countermeasure development, and efficient medical triage.