[1728] Flow Cytometric Evaluation of Reactive and Dysplastic Granulocyte Maturation by a Novel Method of High Dimensional Data Analysis

WG Finn, KM Carter, R Raich, A Harrington, SH Kroft, AO Hero. University of Michigan, Ann Arbor, MI; Oregon State University, Corvallis, OR; Medical College of Wisconsin, Milwaukee, WI

Background: Traditional flow cytometry methods do not fully exploit the dimensionality of the data. We recently demonstrated the analysis of flow cytometry datasets as single high-dimensional objects via a novel method called FINE (Fisher information non-parametric embedding), that uses principles of information geometry and statistical manifolds to determine information distances between multidimensional datasets (Cytometry B 2008, July 18; epub ahead of print) and effectively distinguished leukemic lymphoblasts from physiologic lymphoid precursors (hematogones) in marrow. We now apply this method to the comparison of granulocyte maturation patterns in normal and dysplastic granulopoeisis, since published data are not consistent on the objective role of flow cytometry for this purpose.
Design: 28 marrow samples (11 non-neoplastic [NN] and 17 myelodysplastic [MDS]) were evaluated with a standard 4-color (6-parameter) tube measuring forward scatter (FS), side scatter (SS), CD11b, CD16, CD45, and CD56. Non-blastic granulocytes/precursors were selected by CD45 vs SS, and listmode datasets were analyzed by FINE, which 1)converts the 6-parameter listmode data in each case into 6-dimensional probability density functions via a kernel density estimate, 2)compares information distances via an entropy measurement, the Kullback-Liebler (K-L) divergence, and 3)constructs high-dimensional neighborhood maps of K-L divergence among cases, projecting these maps onto 2-dimensional (2D) plots for visual comparison.
Results: 10 MDS cases (4 RAEB, 3 RCMD, 3 MDS unclassified) formed a discrete cluster in FINE space. 7 MDS cases (2 RARS, 3 RCMD, 2 MDS unclassified) showed some overlap with the NN samples. One NN sample (maturation arrest due to early recovery from drug induced agranulocytosis) embedded within the MDS cluster.
Conclusions: Our analysis effectively separated high-grade MDS (all RAEB and half of RCMD cases) from NN marrow samples, but showed some overlap between NN and MDS overall. These data provide a demonstration of principle that granulocyte maturation patterns may be objectively compared by an algorithm (FINE) that treats flow cytometry data as single high-dimensional objects rather than as a series of 2D histograms. The clustering of a case of regenerative maturation arrest with MDS cases warrants additional study of similar cases.
Category: Techniques

Tuesday, March 10, 2009 1:45 PM

Platform Session: Section G, Tuesday Afternoon