Lymphoma discrimination by computerized triple matrix analysis of list mode data from three‐color flow cytometric immunophenotypes of bone marrow aspirates

• Published in: Cytometry (New York, N.Y.) Vol. 41; no. 1; pp. 9 - 18
• Main Authors: Bartsch, R., Arland, M., Lange, St, Kahl, Ch, Valet, G., Höffkes, H.‐G.
• Format: Journal Article
• Language: English
• Published: New York John Wiley & Sons, Inc 01.09.2000
• Subjects: Algorithms; Artificial Intelligence; Biopsy, Needle; Bone Marrow Cells - pathology; CLASSIF1; Color; data mining; Diagnosis, Computer-Assisted; Diagnosis, Differential; Flow Cytometry - methods; flow cytometry, list mode; Hodgkin Disease - pathology; Humans; Immunophenotyping - classification; Immunophenotyping - methods; lymphoma; Lymphoma, Non-Hodgkin - pathology; Multiple Myeloma - pathology
• ISSN: 0196-4763; 1097-0320
• DOI: 10.1002/1097-0320(20000901)41:1<9::AID-CYTO2>3.0.CO;2-T

Background The goal of this study was to evaluate a self‐learning algorithm for the computer classification of information extracted from flow cytometric immunophenotype list mode files from high‐grade non‐Hodgkin's lymphoma (NHL), Hodgkin's disease (HD), and multiple myeloma (MM). Materials and Methods Bone marrow aspirates (BMA) were obtained from untreated NHL (n = 51), HD (n = 9), or MM (n = 13) patients. Bone marrow aspirates were not infiltrated in NHL and HD patients as confirmed by thorough histologic and cytologic investigation; however, MM patients showed an infiltration rate >50% by malignant myeloma cells. Peripheral blood leukocyte (PBL) samples were taken from age‐matched healthy volunteers (n = 44) as easily available control material. A second control group of 15 healthy volunteers, from whom BMA and PBL samples were available, allowed us to differentiate whether the observed classification results on malignant samples were due to the malignant process or simply to the inherent differences between BMA and PBL. Bone marrow aspirates and PBL were analyzed by the same immunophenotyping antibody panel (CD45/14/20, CD4/8/3, kappa/CD19/5, lambda/CD19/5). The acquired list mode data files were analyzed and classified by the self‐learning triple matrix classification algorithms CLASSIF1 following a priori separation of the data into a learning set and unknown test set. After completion of the learning phase, known patient samples were reclassified and unknown samples prospectively classified by the algorithm. Results Highly discriminatory information was extracted for the various lymphoma entities. The most discriminating information was encountered in antibody binding, antibody binding ratios, and relative antibody surface density parameters of leukocytes rather than in percentage frequencies of discrete leukocyte subpopulations. Samples from healthy controls were classified as normal in 97.2% of the cases, whereas those of NHL, HD, and MM patients were on average correctly classified in 80.8% of the cases. Conclusions Although no detectable lymphoma cells were present in BMA of NHL and HD patients, the CLASSIF1 classification of the immunophenotypes of morphologically normal cells provided a surprisingly good disease discrimination equal or better than that obtained by examining pathological lymph nodes according to the respective literature. The results are suggestive for a lymphoma‐related and disease‐specific antigen expression shift on normal hematopoietic bone marrow cells that can be used to discriminate the underlying disease (specificity of unspecific changes), i.e., in this case NHL from HD. Multiple myeloma patients were discriminated by changes on malignant as well as on normal bone marrow cells. Cytometry 41:9–18, 2000 © 2000 Wiley‐Liss, Inc.