2026 ASH®
Annual Meeting & Exposition December 12-15, 2026 ▪ New Orleans, Louisiana

Paper Details

Abstracts, authors, affiliations, and disclosures.

Saturday, December 12, 2026
05:30 p.m. - 07:30 p.m.
MCCNO - Hall F
Access, Representation, Population Health, Technology and Procedures, Imaging, Molecular Testing, Omics Technologies, Pathology, Bioinformatics, Blood Banking, Computational Biology, Emerging Technologies, Profiling, Treatment Considerations, Drug Development, Supportive Care, Therapy Sequence, Biological Therapies, Chimeric Antigen Receptor, (CAR)-T Cell Therapies, Women’s Health, Workforce, Adverse Events, Artificial Intelligence (AI), Computer Vision, Deep Learning, Generative AI, Machine Learning, Natural Language Processing, Biological Processes, Apoptosis, Genomics, Hematopoiesis, Molecular Biology, Clinical Practice (Health Services and Quality), Clinical Guidelines, Quality Improvement, Quality of Care, Systems-Based Hematology, Diseases, Bleeding and Clotting, Thromboembolism, Bone Marrow Failure Syndromes, Inherited Marrow Failure Syndromes, Paroxysmal Nocturnal Hemoglobinuria, Lymphoid Malignancies, Lymphoid Leukemias, Lymphomas, Plasma Cell Disorders, Myeloid Malignancies, Acute Myeloid Malignancies, Chronic Myeloid Malignancies, Metabolic Disorders, Iron Deficiency, Measurable Residual Disease, Non-Hodgkin Lymphoma, B Cell Lymphoma, Aggressive Lymphoma, Research, Fundamental Science, AML, APL, Translational Research, MDS, MPN, Clinical Research, Clinical Trials, Real-World Evidence, Study Population, Human, Adult, Pediatric
PAPER

2574 Automated identification of normal hematopoietic cell populations by machine learning: Reproduction of expert flow cytometric gating using radial kernel support vector machines

Parnaz Daneshpajouhnejad, MD, Sarah Blain, MD2, Jon Richardson3, Andrew J. Menssen, Ph.D., HCLD(ABB)4, Michael Loken, PhD5
1 Hematologics, Seattle, Washington, United States
2 Hematologics, Seattle, Washington, United States
3 Hematologics, Seattle, Washington, United States
4 Hematologics, Seattle, Washington, United States
5 Hematologics, Seattle, Washington, United States

Background

Flow cytometry is an integral part of acute myeloid leukemia monitoring response to therapy. Post-therapy specimens contain regenerating normal cells, which serve as a reference for measuring hematopoietic regeneration. In this study, we evaluated whether supervised radial basis function (RBF) support vector machines (SVMs) can quantify normal hematopoietic bone marrow populations to facilitate the identification of minimal residual disease (MRD).

Methods

Myeloid MRD-negative regenerative post-therapy bone marrow samples (100 training, 100 validation) were randomly selected from routine clinical cases obtained from 5 separate standardized four-color flow cytometers over a period spanning 1.5 years.

The pipeline was implemented in Python using scikit-learn library for RBF SVM classification. Raw flow cytometry events, including compensation data, were processed, and events were labeled as inside-gate (positive) or outside-gate (negative) to identify normal cell populations by the expert using Boolean gating (WinList, Verity Software House, Topsham, ME). Fluorescent intensities and side scatter were analyzed in log space, while linear data were used for forward scatter.

Binary SVM classifiers were trained for 30 normal populations using randomly sampled, expert-labeled positive and negative events from each case, with class balancing and sampling limits to prevent dominant gates from biasing the model's training. For each population, negative training events were sampled from its biological parent population (e.g., lymphocytes for T cells). This ensures that negative cells used for training are phenotypically closer to positive cells and thus, most informative for learning the decision boundary. For each gate, hyperparameters (C and γ) and decision thresholds were optimized using cross-validation, with model selection based on Matthews correlation coefficient (MCC).

Models were evaluated by comparing predicted gate membership with expert multidimensional gating using sensitivity, specificity, PPV, F1 score, MCC (a metric well-suited to imbalanced classes in rare populations), and coefficient of determination (R²) of predicted population frequencies and marker fluorescence intensity across gates.

Results

The SVM demonstrated excellent agreement with expert manual gating across all gates, with mean sensitivity, specificity, PPV, F1 score, and MCC of 98.5%, 99.7%, 88.3%, 0.93, and 0.93, respectively. Classification accuracy exceeded 98% for every gate, with sensitivity and specificity exceeding 99% across many major populations, including myeloid cells, lymphocytes, monocytes, CD34+ progenitors, doublets, basophils, and dendritic cells. Lower performance (MCC: 0.71-0.81) was observed primarily in a small number of rare populations, including early committed myeloid progenitors (CD34+, CD13+), immature NK cells, early monocytes, erythroid progenitors, and eosinophils.

SVM also preserved immunophenotypic characteristics of expert-defined populations. Across 480 marker-population combinations, median fluorescence intensities exhibited R² values of 0.99, indicating nearly identical marker-population characteristics.

Canonical lineage markers and scatter properties consistently demonstrated strong discriminatory performance with large effect sizes (AUC >0.99, absolute Cohen’s d values = 2-5), including CD16 for mature neutrophils, CD123 for basophils and dendritic cells, CD14 for monocytes, CD117 for mast cells, CD38 for plasma cells, CD34 for CD34+ progenitors, SSC for myeloid cells, and FSC for doublets, indicating that SVM learned biologically meaningful immunophenotypic relationships.

Population frequency estimates generated by the SVM closely reproduced expert gating, with a median population frequency R² of 0.97, and 65% of gates exceeding R² of 0.91. Lower agreement was primarily observed in rare populations with limited event counts, including basophils, myeloid progenitors, early monocytes, and erythroid progenitors.

Conclusion

Supervised RBF SVM algorithms can accurately reproduce expert manual flow cytometric gating across normal hematopoietic populations with high concordance in a well-controlled, stable flow cytometry assay. This approach demonstrates the potential for scalable, reproducible machine learning-assisted flow cytometry analysis, increasing efficiency while reducing subjective bias. This approach also provides a means for real-time quality control of flow cytometric data.

Parnaz Daneshpajouhnejad, MD: Hematologics Inc. - Consultancy (Includes expert testimony); Hematologics Inc. - Current Employment