Artificial intelligence-enhancement of flow cytometry data accelerates the identification of measurable residual chronic lymphocytic leukemia

Leukemia ARTICLE www.nature.com/leu OPEN Artificial intelligence-enhancement of flow cytometry data accelerates the identification of measurable residual chronic lymphocytic leukemia ✉ April Chiu 1,3 , Jansen N. Seheult1,3, Min Shi 1, Dragan Jevremovic 1, Clarissa E. Jordan1, Mathew J. Weybright Michael M. Timm1, Gregory E. Otteson1, Horatiu Olteanu 1, Sameer A. Parikh 2 and Pedro Horna 1 1 , 1234567890();,: © The Author(s) 2026 Flow cytometry (FC) is essential for detecting measurable residual disease (MRD) in chronic lymphocytic leukemia (CLL), but its use is limited by the expertise and time required for manual analysis. We developed an artificial intelligence (AI) pipeline, Clustering/ Classification of All Events, Dimensionality reduction, Downsampling, and Aberrancy Scaling (CCADDAS), to automatically enhance raw FC files, streamlining CLL MRD detection using a single-tube 10-color panel. FC files from 166 MRD-positive and 61 MRDnegative cases were processed in a cloud environment. Automated steps included error correction (FlowCut), clustering (PARC), dimensionality reduction (UMAP), anomaly detection against negative controls, and cluster-informed downsampling that preserved rare MRD events. A deep neural network trained on expert-defined normal subsets enabled automated gating. AI-enhanced files were analyzed in standard FC software, yielding results highly concordant with conventional expert review (R² = 0.98). Downsampling reduced cellularity by 85% and file size by 78%, while retaining low-level MRD events. An AI-generated aberrancy scale distinguished CLL MRD from background B cells with excellent performance (AUC = 0.98). Manual analysis time decreased from 9.0 to 0.9 min per case (90% reduction). CCADDAS provides a largely unsupervised, software-agnostic method that accelerates and simplifies CLL MRD detection without compromising test performance compared to conventional analysis, enabling broader adoption of FC-based MRD testing. Leukemia; https://doi.org/10.1038/s41375-026-02986-3 INTRODUCTION Chronic lymphocytic leukemia (CLL) is a lymphoid neoplasm characterized by accumulation of clonal mature B-cells in the peripheral blood, bone marrow, lymph nodes, and/or spleen [1]. CLL is currently the most common leukemia in the Western Hemisphere, with approximately 24,000 new cases diagnosed in the United States in 2025 [2]. Although patients with CLL typically present with indolent disease and can be managed by observation alone, most eventually require therapeutic intervention due to progressive or symptomatic disease. CLL-directed therapeutic options have dramatically expanded over the last decade and continue to do so, which consist of chemoimmunotherapy (including CD20 targeting monoclonal antibodies such as rituximab and obinutuzumab), agents targeting B-cell receptor signaling (e.g., ibrutinib), B-cell lymphoma 2 inhibitors (BCL2i, such as venetoclax), and chimeric antigen receptor T-cell (CART) therapy [3]. While many patients do achieve complete remission (CR) according to the 2018 International Workshop on Chronic Lymphocytic Leukemia (iwCLL) criteria [4], residual disease may still be detectable, termed measurable (or minimal) residual disease (MRD), by using highly sensitive assays. MRD detection in patients with CLL following therapy is of great significance as this has been established as an independent prognostic factor for progression free survival and overall survival [5–7]. A negative MRD status in CLL patients is increasingly used as a surrogate endpoint to establish efficacy and consideration for cessation of therapy [8–12]. CLL MRD can be assessed by a variety of methodologies including flow cytometry (FC), allele-specific oligonucleotide quantitative polymerase chain reaction (ASO-qPCR), next generation sequencing (NGS), droplet digital PCR (ddPCR), and cell-free DNA analysis [13]. Of these, FC is the most assessable, fastest, and likely most cost effective; and therefore most widely used. Furthermore, FC does not require prior knowledge of diseasespecific immunoglobulin sequences or immunophenotypic abnormalities at diagnosis. However, while modern FC platforms are capable of achieving the recommended assay sensitivity of at least 0.01% (10−4), which is recommended by both 2018 iwCLL [4] and 2026 NCCN guidelines [14], molecular methods have the advantage of detecting even lower MRD levels (up to 10−6) which may have prognostic significance in some settings. In addition, conventional FC MRD testing is typically performed by manual gating and analysis of cell populations on two-dimensional dot plots, and heavily relies upon technical expertise of clinical laboratory staff. Contributing to this highly manual and timeconsuming process is the large amount of data analyzed, requiring 1 Department of Laboratory Medicine and Pathology, Division of Hematopathology, Mayo Clinic, Rochester, MN, USA. 2Department of Medicine, Division of Hematology, Mayo Clinic, Rochester, MN, USA. 3These authors contributed equally: April Chiu, Jansen N. Seheult. ✉email: Received: 29 December 2025 Revised: 3 April 2026 Accepted: 15 May 2026 A. Chiu et al. 2 robust computational processing power and advanced software analytical capabilities compared to non-MRD-based FC analysis. Therefore, the availability of MRD testing by FC is generally confined to large academic centers and reference laboratories. To overcome these limitations, we implemented a cloud-based computational artificial intelligence (AI) pipeline recently developed by our group [15], which simplifies and accelerates the identification and quantification of MRD without compromising test performance compared to conventional analysis. This pipeline, CCADDAS (Clustering and Classification of All Events, Dimensionality reduction, Downsampling, and Aberrancy Scaling), transforms raw data files produced by the FC analyzer into markedly smaller, AI-annotated, and software-agnostic files, and provides a measure of aberrancy in comparison to negative controls. We applied this pipeline on a large cohort of FC data files produced from our 10-color CLL MRD clinical assay, and compared the pipeline’s performance against that of conventional FC CLL MRD analysis. METHODS Patient and sample selection For training of the pipeline, 29 negative control samples (15 bone marrow aspirates and 14 peripheral bloods) were selected; including 24 CLL MRD negative cases analyzed at Mayo Clinic, Rochester, MN, between January 2021 and July 2023; and 5 normal marrow specimens procured from hip replacements and analyzed between January and May 2021. The selection of controls was based on adequate representation of benign normal subsets as assessed by expert analysts. A separate research cohort included 227 specimens (74 bone marrow aspirate and 153 peripheral blood) received for CLL MRD flow cytometric analysis at Mayo Clinic in Rochester, MN, between April 2023 and February 2024; including 166 specimens from 122 patients reported as MRD- (...truncated)