Diagnosis of chronic B-cell lymphoproliferative disease in peripheral blood = how machine learning may help to the interpretation of flow cytometry data

Flow cytometry (FCM) has become a method of choice for immunologic characterization of chronic lymphoproliferative disease (CLPD). To reduce the potential subjectivities of FCM data interpretation, we developed a machine learning random forest algorithm (RF) allowing unsupervised analysis. This assay relies on 16 parameters obtained from our FCM screening panel, routinely used in the exploration of peripheral blood (PB) samples (mean fluorescence intensity values (MFI) of CD19, CD45, CD5, CD20, CD200, CD23, HLA-DR, CD10 in CD19-gated B cells, ratio of kappa/Lambda, and different ratios of MFI B-cells/T-cells [CD20, CD200, CD23]). The RF algorithm was trained and validated on a large cohort of more than 300 annotated different CLPD cases (chronic B-cell leukemia, mantle cell lymphoma, marginal zone lymphoma, follicular lymphoma, splenic red pulp lymphoma, hairy cell leukemia) and non-tumoral selected from PB samples. The RF algorithm was able to differentiate tumoral from non-tumoral B-cells in all cases and to propose a correct CLPD classification in more than 90% of cases. In conclusion the RF algorithm could be proposed as an interesting help to FCM data interpretation allowing a first B-cells CLPD diagnostic hypothesis and/or to guide the management of complementary analysis (additional immunologic markers and genetic).