An Automatic Diagnosis of Several Types of Lymphoma by Flow Cytometry Data
Ming-Chih Shih, Rachel Donohue, Lixin Zhang, Chun-Che (Jeff) Chang, Shou-Hsuan Stephen Huang, Youli Zu. The Methodist Hospital, Houston, TX; University of Houston, Houston, TX; University of Central Florida, Orlando, FL
Background: Flow cytometry is a valuable tool in the diagnosis and management of hematologic malignancies. Flow cytometry platforms generate multi-dimensional datasets, posing a challenge for manual gating on bi-dimensional plots. Automated multivariate clustering is also stymied by the identification of rare populations that form small clusters and the computational challenges posed by the large size and dimensionality. Promising new research has demonstrated the utility of model-based analysis in identification of cell subtypes. We propose a novel five dimensional (5D) model designed to detect most of the B-cell leukemias/lymphomas by profiling the expression of five biomarkers.
Design: Based on our previous research we used a Gaussian mixture model and an expectation maximization algorithm to build a 5D model represented by ellipsoids. Because of their usefulness in identifying B-cell lymphomas we chose CD19, CD5, CD10, kappa light chain, and lambda light chain as our five parameters in building both a normal and chronic lymphocytic leukemia (CLL) 5D profile (Figure 1). We designed a fitting algorithm to compare a test sample profile with the pre-programmed profiles to determine if they fit.
Results: The Multi-Profile Lymphoma Detection system was able to correctly identify 10/10 CLL test subjects and 16/16 normal test subjects with 100% accuracy based on the 5D model. We used a support vector machine with 30 training cases (20 normal training cases and 10 CLL training cases) to draw a separate line for our final result (Figure 2).
Conclusions: We have demonstrated that one can obtain an automated diagnosis of chronic lymphocytic leukemia by flow cytometry data. In addition, study of other types of lymphomas and leukemias by this model is currently ongoing and findings will be discussed.
Tuesday, March 20, 2012 1:00 PM
Poster Session IV # 211, Tuesday Afternoon