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

Paper Details

Abstracts, authors, affiliations, and disclosures.

Sunday, December 13, 2026
06:00 p.m. - 08:00 p.m.
MCCNO - Hall F
Technology and Procedures, Assays, Imaging, Molecular Testing, Omics Technologies, Pathology, Bioinformatics, Clinical Procedures, Computational Biology, Emerging Technologies, Gene Editing, Profiling, Treatment Considerations, Drug Development, Therapy Sequence, Biological Therapies, Transplantation, (CAR)-T Cell Therapies, Workforce, Adverse Events, Artificial Intelligence (AI), Computer Vision, Deep Learning, Generative AI, Machine Learning, Natural Language Processing, Biological Processes, Pathogenesis, CHIP, Genomics, Hematopoiesis, Immune Mechanism, Metabolism, Microbiome, Molecular Biology, Multi-Systemic Interactions, Clinical Practice (Health Services and Quality), Quality Improvement, Systems-Based Hematology, Diseases, Bone Marrow Failure Syndromes, Lymphoid Malignancies, Lymphoid Leukemias, Lymphomas, Plasma Cell Disorders, Myeloid Malignancies, Acute Myeloid Malignancies, Chronic Myeloid Malignancies, Metabolic Disorders, Measurable Residual Disease, ALL, CLL, LGL, Hodgkin Lymphoma, Research, Fundamental Science, AML, APL, Translational Research, MDS, MPN, Clinical Research, Clinical Trials, Epidemiology, Health Outcomes Research, Real-World Evidence, Registries, Survivorship, Study Population, Human, Adult, Elderly
PAPER

4343 Self-supervised learning to overcome batch variability and enable attention-based screening for T-LGLL–Associated abnormalities in clinical flow cytometry

Vahid Mehrtash, MD, Megan Nakashima, MD2, Kelly Bowers3
1 Cleveland Clinic, Cleveland, Ohio, United States
2 Cleveland Clinic, Cleveland, Ohio, United States
3 Cleveland Clinic, Cleveland, Ohio, United States

Introduction: T-cell large granular lymphocytic leukemia (T-LGLL) can be challenging to recognize by flow cytometry, particularly when the abnormal T-cell population is small or demonstrates only subtle abnormalities. Automated screening could support detection in high-volume clinical laboratories; however, developing such tools from longitudinal clinical data is complicated by variation in instrument settings, reagent lots, and compensation, especially when standardized internal controls are unavailable. We evaluated a clinically oriented artificial-intelligence workflow that uses self-supervised learning to reduce sensitivity to historical technical variation and attention-based modeling to identify cases with abnormal T-cell flow cytometric findings.

Methods: We analyzed 189 T-cell flow cytometry cases performed from August 2022 through December 2025, including 134 abnormal and 55 normal cases. Antibodies against CD2, CD3, CD4, CD5, CD7, CD8, CD13, CD16, CD19, CD45, CD56, CD57, and TRBC1 were used. Flow cytometry data files underwent compensation, acquisition-quality assessment, singlet selection, and lymphocyte gating pre-processing. A flow cytometry–specific self-supervised learning (SSL) framework with a Barlow Twins-based redundancy-reduction objective was used to encode each cell as a 32-dimensional latent feature representation, which allowed for learning of meaningful cellular features without individual cell annotation. To address the absence of uniform internal controls or batch metadata, the model was trained using case-level nuisance augmentations that simulated plausible run-to-run variation, including fluorescence-intensity shifts, nonlinear transformation drift, gain and reagent-lot scaling, compensation perturbations, and detector noise, while preserving biologically relevant features. A gated-attention multiple-instance learning (MIL) classifier then evaluated all cells within each case, allowing a small abnormal population to have greater influence on the final case-level prediction. Performance was assessed using three repeats of five-fold cross-validation with each case evaluated by models not trained on that case (fully out-of-fold [OOF] predictions). Attention-weighted biaxial plots were generated to identify the cells most responsible for each prediction and to determine whether they formed a coherent abnormal population.

Results: The model achieved an OOF area under the receiver operating characteristic curve of 0.857 and an average precision of 0.927, compared with an abnormal case prevalence of 0.709. The mean fold-level area under the curve was 0.855±0.059. An exploratory Youden-optimized threshold of 0.510 yielded 80.6% sensitivity and 81.8% specificity, with 108 true-positive, 10 false-positive, 26 false-negative, and 45 true-negative classifications. The attention mechanism localized the cell populations contributing to high-probability predictions, allowing targeted review of clinically relevant features such as TRBC1 restriction, CD4/CD8 distribution, loss or dim expression of pan–T-cell antigens, and aberrant expression levels. These findings demonstrate that useful case-level discrimination can be obtained from heterogeneous longitudinal clinical flow cytometry data without standardized historical controls.

Conclusions: A self-supervised encoder combined with attention-based MIL offers a potential strategy for managing batch-related variability in routine flow cytometry. After training, a frozen encoder and MIL classifier could function as a rapid screening or triage tool in a busy laboratory, flagging high-probability cases for focused technologist and pathologist review. This method has potential applications in other flow cytometric diagnostics, including minimal residual disease, as well as across different flow cytometry platforms/analyzers. Future work will focus on characterizing the attention-weighted cell populations with biologically and clinically relevant features with the goal of extending the model beyond screening and into a discovery platform for identifying clinically relevant immunophenotypic populations associated with diagnosis, disease behavior, treatment opportunities, and other clinical outcomes. External and prospective validation, including across different patient populations, time periods, institutions, and flow cytometry platforms, and comparison to definitive clinical T-LGLL diagnoses and longitudinal outcome are required before clinical implementation.
Vahid Mehrtash, MD: No relevant conflicts of interest to declare.