Automated identification of phenotypically aberrant cells in acute myeloid leukemia MRD: An unsupervised machine learning approach.

Abstract

e18560 Background: Measurable residual disease (MRD) serves as a robust prognostic marker in acute myeloid leukemia (AML), guiding therapeutic stratification and treatment response assessment. Invivoscribe's AML MRD Assay is a multiparametric flow cytometry (MFC) 12-color panel of 21 biomarkers to characterize potential AML blast cells using a leukemia-associated immunophenotype (LAIP) based and different-from-normal (DfN) approach. We developed an unsupervised algorithm (FlowER) to automatically identify phenotypically aberrant cells from flow cytometry data without relying on or training on human-drawn gates. This retrospective internal validation study assessed FlowER's concordance with clinical MRD gates in distinguishing aberrant from normal cell populations. Methods: We analyzed bone marrow samples from 59 AML patients and 25 normal controls processed using one tube from our panel (Tube 2: CD15, CD13, 7-AAD, CD33, CD34, CD45, CD117, HLA-DR, CD5, CD7, CD2). The 25 controls served as a healthy reference pool to establish expected phenotypic distributions. For each patient sample, after density-aware subsampling, FlowER assigned anomaly scores to all cells using k-nearest neighbor distances to the reference pool in dimensionality-reduced feature space, agnostic to immunophenotypic lineage. To validate this unsupervised approach, clinical MRD gates were independently defined by expert cytometrists using standard LAIP-based approaches. We compared median anomaly scores between these clinically-defined aberrant cells and CD45 Dim non-MRD background cells within each patient. The primary outcome was the magnitude and significance of score separation (using a Wilcoxon signed-rank test). Results: Total samples analyzed: 59 patients. Median difference in anomaly scores: MRD cells +1.328 higher (35.6% increase) vs. CD45 Dim background; p = 1.78 × 10⁻¹⁰ (one-sided Wilcoxon signed-rank test). Samples with elevated anomaly scores in MRD cells: 52/59 (88%). Cohen’s d (effect size): 1.204. Separation was consistent across CD34+, CD117+, and CD34-/CD117- LAIP populations. Conclusions: FlowER demonstrated highly significant concordance with clinical MRD gating, successfully distinguishing aberrant from normal cells within immunophenotypic populations (p = 1.78 × 10⁻¹⁰). The consistent median score elevation in clinically-aberrant cells (88% of patients) with large effect size (Cohen's d = 1.204) indicates the algorithm captures phenotypic abnormalities without requiring labeled training data or manual gate definitions. This unsupervised, reference-based approach could enable standardized MRD quantification across institutions with reduced turnaround time, complement expert-driven assessment, and reduce inter-observer variability. Prospective validation with clinical outcome correlation is warranted to assess benefit for clinicians.