Automated Flow Cytometric MRD Assessment in Childhood Acute B‐ Lymphoblastic Leukemia Using Supervised Machine Learning. Issue 9 (7th July 2019)

• Record Type: Journal Article
• Authors: Reiter, Michael; Diem, Markus; Schumich, Angela; Maurer‐Granofszky, Margarita; Karawajew, Leonid; Rossi, Jorge G.; Ratei, Richard; Groeneveld‐Krentz, Stefanie; Sajaroff, Elisa O.; Suhendra, Susanne; Kampel, Martin; Dworzak, Michael N.
• Is Part Of: Cytometry. Volume 95:Issue 9(2019)
• Page Start: 966
• Page End: 975
• Publication Date: 2019-07-07
• Subjects: acute lymphoblastic leukemia -- B‐ALL -- minimal residual disease -- multiparameter flow cytometry -- automated gating -- machine learning -- gaussian mixture model -- algorithm; Flow cytometry -- Periodicals; Imaging systems in biology -- Periodicals; Imaging systems in medicine -- Periodicals; Diagnostic imaging -- Periodicals; 571.605
• DOI: 10.1002/cyto.a.23852
• Languages: English
• ISSNs: 1552-4922

Abstract

Abstract: Minimal residual disease (MRD) as measured by multiparameter flow cytometry (FCM) is an independent and strong prognostic factor in B‐cell acute lymphoblastic leukemia (B‐ALL). However, reliable flow cytometric detection of MRD strongly depends on operator skills and expert knowledge. Hence, an objective, automated tool for reliable FCM‐MRD quantification, able to overcome the technical diversity and analytical subjectivity, would be most helpful. We developed a supervised machine learning approach using a combination of multiple Gaussian Mixture Models (GMM) as a parametric density model. The approach was used for finding the weights of a linear combination of multiple GMMs to represent new, "unseen" samples by an interpolation of stored samples. The experimental data set contained FCM‐MRD data of 337 bone marrow samples collected at day 15 of induction therapy in three different laboratories from pediatric patients with B‐ALL for which accurate, expert‐set gates existed. We compared MRD quantification by our proposed GMM approach to operator assessments, its performance on data from different laboratories, as well as to other state‐of‐the‐art automated read‐out methods. Our proposed GMM‐combination approach proved superior over support vector machines, deep neural networks, and a single GMM approach in terms of precision and average F 1 ‐scores. A high correlation of expert operator‐based and automated MRD assessment was achieved with reliable automated MRD quantification ( F 1 ‐scores >0.5 in more than 95% of samples) in the clinically relevant range. Although best performance was found, if test and training samples were from the same system (i.e., flow cytometer and staining panel; lowest median F 1 ‐score 0.92), cross‐system performance remained high with a median F 1 ‐score above 0.85 in all settings. In conclusion, our proposed automated approach could potentially be used to assess FCM‐MRD in B‐ALL in an objective and standardized manner across different laboratories. © 2019 International Society for Advancement of Cytometry Abstract :