Applications of machine learning for immunophenotypic measurable residual disease assessment in acute myeloid leukemia

Abstract

Immunophenotypic detection and quantification of residual leukemic cells by multiparameter flow cytometry is increasingly adopted in the clinical practice of acute myeloid leukemia (AML) to assess measurable residual disease (MRD). However, MRD levels quantified by manual gating analysis can differ based on differences in gating strategy between trained operators and clinical centers. Manual gating requires extensive training, is time-consuming in daily practice, and faces a significant hurdle in analyzing data from next-generation cytometry platforms. To address these challenges, several computational approaches involving machine learning and artificial intelligence algorithms have been proposed to automate or aid the assessment of MRD. However, the immunophenotypic variability between patients and the relatively low proportions of residual leukemic cells in AML challenge most algorithms and require innovative approaches. This review provides an overview of recent efforts in using computational methods for immunophenotypic AML-MRD assessment. We first explain the technical and conceptual background of the different algorithms that have been explored. Next, we discuss their strengths and limitations in the disease-specific context of AML. Finally, we highlight how computational approaches offer a unique opportunity to standardize or even outperform current manual gating analyses, and ultimately, improve the treatment of AML patients.

Figure 1

Heterogeneity of leukemia‐associated immunophenotypes (LAIPs) identified after two cycles of chemotherapy in AML patients (n  = 455) in the HOVON‐SAKK‐132 trial. A binary representation of observed (red) or not‐observed (gray) LAIPs was used for simplicity. Patients (rows) were hierarchically clustered based on the common LAIPs between rows (Jaccard index). LAIPs (columns) were organized according to the primitive marker (CD34, CD117, CD133, or mature) first and by frequency second.

Figure 2

Theoretical overview of computational approaches for AML‐MRD assessment. (A) Four samples are simulated for two hypothetical leukemic markers (LAIP1 and LAIP2). (B) Supervised classification of the three AML samples using AML1 as training data. Cells are classified as leukemic if the majority of its five closest (i.e., “nearest neighbor”) cells in AML1 are leukemic. (C) Semi‐supervised “cluster‐with‐normal” approach using K = 5. K means clusters. Cells in a cluster are classified as leukemic if more than 95% of cells belong to the test sample. (D) Semi‐supervised novelty detection approach. Cells are classified as leukemic if the median distance to its five closest (i.e., “nearest neighbor”) cells exceeds a certain distance (0.5). An interactive environment online to explore this simulation is available at: https://github.com/AUMC-HEMA/cMRD-review.