Detecting Myelodysplastic Syndromes on Peripheral Blood Using a Machine Learning Approach To Analyze Multiparameter Hematology Analyzer Data
Dick G Hwang, David M Dorfman, Debra A Briggs, Ricardo Silverio, Olga Pozdnyakova. Virginia Mason Medical Center, Seattle, WA; Brigham and Women's Hospital, Boston, MA; Dana Farber Cancer Institute, Boston, MA
Background: In the absence of clear dysplasia or specific cytogenetic abnormalities, the diagnosis of myelodysplastic syndromes (MDS) can be challenging. NEUT-X, a measurement of neutrophil side scatter, is one of many parameters measured by the Sysmex XE-5000 hematology analyzer on peripheral blood samples but not routinely reported. NEUT-X has been found to be associated with MDS but is insensitive when used alone. We investigate whether simultaneous analysis of multiple hematology analyzer parameters can improve detection of MDS.
Design: 516 peripheral blood samples from BWH and DFCI patients with confirmed MDS-related neoplasms were analyzed on the Sysmex XE-5000 analyzer and compared to 28,216 control samples that included sick patients but excluded MDS or cancer patients. Support vector machine (SVM), a linear classification algorithm that operates on multi-parameter data, was used to analyze several parameter sets (see Table). MDS and control cases were equally partitioned into training and validation sets. The SVM algorithm was run on the training set to generate a mathematical model, which was then used to assign an "MDS-SVM" score to each sample in the validation set. To confirm robustness, the sensitivity and specificity shown are averages over 10,000 repetitions of random subsampling validation.
Results: The histogram (Figure A) demonstrates separation of MDS samples and control samples using the MDS-SVM score based on the "Full" parameter set. Figure B shows sensitivity of the MDS-SVM score in detecting MDS at various levels of specificity for the different parameter sets.
Conclusions: The results demonstrate that using the SVM algorithm to simultaneously analyze routine CBC + differential parameters can detect many MDS cases with high specificity. The sensitivity is further increased by including unreported machine-specific service and research parameters, some of which represent structural information of the cells. The MDS-SVM score, which can be computed by entering the parameters into a simple equation, can therefore be of practical use in the diagnosis of MDS.
Tuesday, March 5, 2013 9:30 AM
Poster Session III # 170, Tuesday Morning