Seheult, Jansen N.

Abstract Background Extramedullary disease (EMD) in multiple myeloma (MM) represents an aggressive phenotype with poor outcomes. We performed single-cell RNA sequencing (scRNA-seq) on clonal plasma cells (CPCs) from EMD tissue, solitary plasmacytomas (SPM), and bone marrow aspirates (BMAs) from patients with MM to identify transcriptional states unique to EMD. Methods We processed FFPE tumor samples from EMD and SPMs and cryopreserved CD138+ BMAs from MM, using the 10x Genomics FLEX platform. Cells/nuclei were fixed, hybridized with Human WTA probes, pooled, and sequenced (10,000–40,000 cells/library) on NovaSeq X. After removing low quality cells, data was normalized and integrated using RPCA in Seurat. Plasma cell identity was assigned via Cellarium and canonical markers. Clonality was inferred by light chain restrictions and refined via random forest modeling. Cell cycle effects were regressed out to reduce proliferation bias. Results We analyzed 128, 908 CPCs from 16 samples [8 EMD (62, 669 CPCs), 4 SPM (20, 607 CPCs), 4 BMA (45, 632 CPCs)] and identified eight transcriptionally distinct cell clusters. Differential expression and pathway comparisons were performed between EMD/SPM/BMA within each of these clusters. Integration of gene expression and pathway analyses revealed consistent upregulation of cell cycle and E2F pathways in EMD. We identified two highly proliferative cell cycle clusters (Cluster 2 with S-phase genes and Cluster 4 with a G2/M transcriptional program) that were expanded in EMD compared to BMA and SPM. Cluster 0 comprised a quiescent CPC population enriched for RNA processing pathways alongside downregulation of ER stress and autophagy. EMD cells demonstrated suppressed interleukin signaling [fold change (FC) –33.8), and class I MHC antigen processing (FC –23.5) relative to BMA, suggestive of an immune suppressive transcriptional program. Angiogenesis-related genes [VEGFA (log2FC 2.4), ADM2 (log2FC 2.2)] and EZH2 [log2FC 2.1; q<0.0001)] were upregulated in EMD. Compared to SPM, EMD plasma cells in this cluster demonstrated marked downregulation of CD74 (log2FC –2.6; q<0.0001), a key modulator of MHC class II antigen presentation. Cluster 1 comprised of metabolically stressed secretory CPCs and displayed downregulation of early growth response pathways in EMD compared to BMA, concurrent with upregulation of E2F signaling. Suppression of IL-2/STAT5, IL-6/JAK/STAT3, and TGF-β pathways in EMD suggests reduced reliance on bone marrow–derived cytokines and impaired immunoregulatory signaling. CD44 was upregulated in EMD versus BMA (log2FC 3.2), consistent with its role in EMD spread and marrow independence. Cluster 3 was enriched for extracellular matrix–interacting plasma cells (COL1A1, MMP2) with upregulated epithelial–mesenchymal transition (EMT) pathways in EMD compared to both BMA (FC 7.5) and SPM (FC 7.07). Cluster 6 was defined by a migratory, motility-primed plasma cell population (CD44, AHNAK, LMNA and S100A10). Notably, IL32 was overexpressed in EMD compared to BMA (log2FC 5.1; adj. p = 2.02×10⁻¹⁴), a finding linked to hypoxia-driven proliferation. Compared to SPM, EMD cells also had elevated MYC/E2F activity and DNA repair, hallmarks of a proliferative phenotype. Cluster 5 displayed a strong interferon-stimulated gene signature (ISG15, IFIT1, TNFSF10), with enrichment of type I IFN, JAK–STAT, and TNF pathways. Within this inflammatory cluster, EMD exhibited marked suppression of immune signaling, metabolism (TCA, OXPHOS), and TP53 programs, consistent with an immune-evasive, metabolically suppressed, and deregulated state. Finally, within Cluster 7, a smaller subset defined by NFκB–driven anti-apoptotic transcriptional activity, EMD exhibited significant downregulation of KLF2 (log2FC –3.7; adj. p = 1.77×10⁻²⁷), likely reflecting disruption of the KDM3A–KLF2–IRF4 axis, a transcriptional network critical for bone marrow homing and plasma cell survival. We consistently observed downregulation of the MAPK pathway in EMD compared to BMA, aligning with prior bulk RNA-seq data and contrasting with the near-universal presence of MAPK mutations, highlighting a disconnect between mutational status and pathway activation. Conclusion In the largest yet reported single-cell transcriptional atlas of EMD and SPM, we demonstrate EMD to be highly proliferative with metabolic reprogramming, EMT, and downregulation of key immunologic signaling pathways, indicating immune escape and microenvironmental independence.

Abstract Background: Recent advances in computer vision have given rise to image foundation models (FMs), large-scale neural networks pre-trained on millions of unlabeled histopathology patches, which can learn generalizable representations of tissue morphology. In many cancers, these FMs have predicted genetic mutations and prognoses directly from histology, suggesting that morphology encodes “hidden” molecular information. In lymphoma, this field is nascent but growing.General-purpose histopathology FMs have limited representation of lymphoid morphology. We developed and evaluated LymphoVision, a lymphoma-specialized image foundation model pretrained with self-supervised learning, to extract clinically meaningful features from routine hematoxylin and eosin (H&E) histology. Methods: LymphoVision was pre-trained using knowledge distillation and consistency (DINOv2) on over 37 million multi-resolution patches (5x, 10x, 20x, 40x) sampled using a density-aware strategy from 31,211 H&E whole slide images (WSIs) from 9,155 archival lymphoma and reactive lymphoproliferative cases. To ensure diverse, representative samples, our patch sampling method uses an efficient graph-based clustering algorithm and prioritizes morphologically informative regions by clustering patch embeddings and performing density sampling within each cluster. The pretrained encoder was evaluated using a clustering-constrained attention multiple instance learning (MIL) framework on three diagnostic tasks: (1) cell-of-origin (COO) classification in diffuse large B-cell lymphoma (DLBCL), distinguishing germinal center B-cell (GCB) from non-GCB subtypes based on the Hans immunohistochemistry algorithm; (2) follicular lymphoma (FL) grading, differentiating low-grade (grades 1–2) from high-grade (grade 3); and (3) multiclass lymphoma subtype classification across six categories: DLBCL, FL, marginal zone lymphoma (MZL), mantle cell lymphoma (MCL), classical Hodgkin lymphoma (CHL), and benign/reactive lymphoid conditions. WSI-level labels were derived from pathology reports, and none of the evaluation cases were included in pretraining. Models used a ViT-Giant LymphoVision backbone to generate 20x patch embeddings, followed by MIL using 50% training data and 2-fold cross-validation (25% validation, 25% test). Performance was assessed using area under the receiver operating characteristic curve (AUC), accuracy, recall (sensitivity) and F1 score. Results: For DLBCL COO classification (total n=358 cases), LymphoVision achieved a mean test AUC of 0.93 (range 0.92–0.94), accuracy of 84.5% and F1 of 0.85 across two independent test folds (n=91 per fold, 50% GCB). These results outperformed general-purpose FMs on the same task and cases, including Virchow-v2 (AUC 0.85) and UNI-v2 (AUC 0.88), despite those models being pretrained on 10–100 times more images. For FL grading (total n=660 cases), the test set in each fold included 166 cases (18% grade 3), yielding a mean test AUC of 0.88 (range 0.87–0.90), accuracy of 89.2% and F1 of 0.80. In multiclass lymphoma subtype classification (total n=1,176 cases), the test set included 296 cases per fold. LymphoVision achieved a mean test AUC of 0.98 (range 0.97-0.99), mean test accuracy of 87.1% and mean F1 of 0.84. Class-wise recall was 92.0% for DLBCL (n=112), 88.2% for FL (n=72), 77.2% for MZL (n=29), 70.8% for MCL (n=35), 87.5% for CHL (n=24), and 95.8% for benign/reactive lymphoid conditions (n=24). Conclusion: LymphoVision is a lymphoma-specialized FM trained using self-supervision that achieves high diagnostic accuracy across multiple clinically meaningful tasks using weak supervision and standard H&E images, suggesting the encoder has learned fine-grained lymphoma morphology. Based on historical benchmarks from the 1997 International Lymphoma Study Group classification project, expert hematopathologists achieved diagnostic agreement rates of 55% to 84% using H&E morphology alone for the non-Hodgkin lymphoma (NHL) entities included in our multiclass task. LymphoVision approached or exceeded these benchmarks. These findings underscore the value of our approach using disease-specialized pretraining and intelligent patch sampling for data-efficient histopathology learning. LymphoVision has the potential to augment diagnostics, enable scalable AI-powered classification in real-world workflows, and serve as a foundation for prognostic and theragnostic applications. External validation and multimodal integration are ongoing.

Abstract Introduction: Only 30% of patients (pts) presenting with abnormal bleeding symptoms receive a clear diagnosis, with many classified as having bleeding disorder of unknown cause, or BDUC (Baker et al. 2021). Currently, the recommended diagnostic workup for BDUC by the International Society on Thrombosis and Haemostasis (ISTH) includes plasmatic assays for coagulation factors, von Willebrand factor (vWF) antigen/activity, and platelet function testing (PFT), specifically platelet function analyzer-100 (PFA-100) and light transmission aggregometry (LTA). However, PFTs are not widely available and cannot be mailed to reference laboratories. We studied the diagnostic yield of an expanded testing panel which includes quantitative platelet glycoprotein flow cytometry (GpFL) and platelet electron microscopy (PTEM), which we have optimized for specimen mailing. Methods: We conducted a minimal risk, retrospective study by reviewing the electronic health records of pts (18 years and older) who were seen in the Coagulation Clinic of a tertiary academic medical center for evaluation of a bleeding disorder between October 2017 and June 2025. Testing included plasma-based assays including PT, APTT, vWF assay, FVIII, FIX, FXIII screen, and thrombin time (and additional reflex assays as needed). At the discretion of the treating clinician, PFTs (PFA-100 and LTA), GpFL, PTEM, and fibrinolysis assays (plasminogen activator inhibitor-1 antigen [PAI-1 Ag] and alpha-2 plasmin inhibitor) were included. We abstracted clinical data on pt demographics, ISTH bleeding scores, and the results of coagulation testing. Pts with an established diagnosis of a bleeding disorder, underlying malignancy, and those who did not have plasma-based coagulation testing were excluded from this study. Results: 109 pts met study criteria, and 82.6% were females (n = 90) with a median age of 45 years (range = 18-88 years). The median ISTH-BAT score was 7 (range: 2-21). About half of the pts, 49.5% (n = 54), had normal coagulation profiles and PFTs. All 54 had PTEM performed, of which 14.8% (n = 8) were abnormal (alpha and/or dense granule deficiencies). 85.2% (n = 46) had GpFL testing performed, with one showing reduced Gp2b/CD42 and Gp3a/CD61 expression, four showing reduced Gp1a/CD49b expression, and three showing reduced GpVI expression. Half of the pts, 50.5% (n = 55) had an abnormal coagulation profile from plasma and/or PFTs, of which 20% (n = 11) had a clinically significant coagulation factor deficiency (FV, FVII, FX). Roughly 9.1% (n = 5) had type 1 vWD of which 1 also had PAI-1 deficiency. One pt had isolated PAI-1 deficiency. Abnormal coagulation profiles were noted in 53 pts who had PFA-100 and LTA done, and both tests were normal in 22.6% (n = 12). Within the abnormal coagulation profile group, 10.9% (n = 6) had ultrastructural abnormalities on PTEM (dense and alpha granule abnormalities). Majority of the pts (n = 46) had GpFL done, and one had abnormal reductions in Gp1b-aCD42b, three in Gp1aCD49b, and seven in GPVI. Comparison of the ISTH BAT score between the normal and abnormal coagulation profiles/PFTs groups by non-parametric (Mann-Whitney U test) and parametric (Welch's t-test) showed no significant difference (p = 0.56 and p = 0.87) respectively. Conclusion: In this cohort of pts with abnormal bleeding, standard recommended testing resulted in a diagnosis in 66.4% pts. The remainder would have been classified as BDUC. The addition of GpFL and PTEM resulted in establishing a diagnosis that explains the bleeding phenotype in 14.8% of pts with an otherwise normal coagulation profile and PFT. Although abnormal PFTs indicate a platelet function defect, GpFL and PTEM may further define the abnormalities particularly differentiating syndromic from non-syndromic inherited platelet disorders. Additional studies are needed to confirm our findings and explore the possibility of addition of GpFL and PTEM for patients with BDUC.

Abstract Measurable residual disease (MRD) assessment by flow cytometry (FC) plays an essential role in prognosis and therapy escalation of B-cell acute lymphoblastic leukemia (B-ALL). However, the high degree of expertise and manual analysis time required limits the availability of this assay. To overcome this limitation, we developed a data-enhancing artificial intelligence (AI) pipeline that accelerates and simplifies MRD analysis. Unaltered FC files from 171 B-ALL MRD-positive and 89 MRD-negative cases were processed through an AI pipeline trained with 31 expert-gated negative controls. Cluster-informed downsampling reduced FC files from 1.2 million to 155 884 cells per case, on average, (87% cellularity reduction), whereas preserving small MRD populations (median, 100% retention for MRD of <1%) and allowing for true percentage MRD estimates using a correction factor. A deep neural network cell classifier automatically identified normal hematopoietic subsets (macro-averaged F1 score of 0.86); and an AI measure of anomaly discriminated B-ALL from benign mononuclear (area under the curve [AUC] of 0.98) or B-lymphoid cells (AUC of 0.94). Manual analysis of AI-enhanced files was completed in only 1.01 minutes per case, on average (standard deviation of ±0.57); with 100% positive agreement with conventional analysis (for MRD of ≥0.01%), 100% negative agreement, and excellent quantitative correlation (R2 = 0.92). Our cloud-based AI-enhancement solution accelerates B-ALL MRD identification without compromising test performance and has the potential of facilitating B-ALL MRD analysis by more clinical laboratories.

Background: Flow cytometric detection of minimal residual disease (MRD) in chronic lymphocytic leukemia (CLL) is complex, time-consuming, and subject to inter-operator variability. Deep neural networks (DNNs) offer potential for standardization and efficiency improvement, but require rigorous validation and monitoring for safe clinical implementation. Methods: We evaluated a DNN-assisted human-in-the-loop approach for CLL MRD detection. Initial validation included method comparison against manual analysis (n = 240), precision studies, and analytical sensitivity verification. Post-implementation monitoring comprised four components: daily electronic quality control, input data drift detection, error analysis, and attribute acceptance sampling. Laboratory efficiency was assessed through a timing study of 161 cases analyzed by five technologists. Results: Method comparison demonstrated 97.5% concordance with manual analysis for qualitative classification (sensitivity 100%, specificity 95%) and excellent correlation for quantitative assessment (r = 0.99, Deming slope = 0.99). Precision studies confirmed high repeatability and within-laboratory precision across multiple operators. Analytical sensitivity was verified at 0.002% MRD. Post-implementation monitoring identified 2.97% of cases (26/874) with input data drift, primarily high-burden CLL and non-CLL neoplasms. Error analysis showed the DNN alone achieved 97% sensitivity compared to human-in-the-loop-reviewed results, with 13 missed cases (1.5%) showing atypical immunophenotypes. Attribute acceptance sampling confirmed 98.8% of reported negative cases were true negatives. The DNN-assisted workflow reduced average analysis time by 60.3% compared to manual analysis (4.2 ± 2.3 vs. 10.5 ± 5.8 min). Conclusions: The implementation of a DNN-assisted approach for CLL MRD detection in a clinical laboratory provides diagnostic performance equivalent to expert manual analysis while substantially reducing analysis time. Comprehensive performance monitoring ensures ongoing safety and effectiveness in routine clinical practice. This approach provides a model for responsible AI integration in clinical laboratories, balancing automation benefits with expert oversight.

Abstract Background The lupus anticoagulant (LAC) is an important laboratory criterion in the diagnosis of antiphospholipid antibody syndrome. LAC testing at the Mayo Clinic Special Coagulation Laboratory includes up to 13 tests, which are interpreted by trained physicians to identify the presence of LAC and rule out anticoagulant interferences. This feasibility study explored the use of two deep neural network (DNN) architectures for multilabel classification of LAC profiles as a first step toward automating interpretation. Methods Seven thousand two hundred and two retrospective cases were randomly split (64:16:20) for training, validation, and test, respectively. LAC positivity by dilute Russell's viper venom time (LAC-DRVVT) and activated partial thromboplastin time (LAC-APTT) and the presence of warfarin (WAR) and heparin (HEP) were adjudicated by one expert. DNN architectures included: single-output DNNs using domain-knowledge input feature selection and a single-column multioutput DNN using all 13 inputs. Results Domain-knowledge-naïve multioutput DNN achieved similar or improved performance for all 4 label prediction tasks compared with domain knowledge optimized DNNs: F1 scores of 0.977 vs 0.968 for LAC-DRVVT, 0.954 vs 0.945 for LAC-APTT, 0.961 vs 0.957 for HEP, and 0.995 vs 0.977 for WAR, respectively. Conclusions The comparable performance of the 4 domain knowledge optimized DNNs and the multioutput DNN using all 13 input features suggests that the DNN may learn feature importance or mapping to a task without explicit input selection. Given its relative simplicity and versatility, the multioutput DNN is the preferred choice for implementation in a clinical laboratory to standardize LAC diagnosis.

ABSTRACTIntroductionVon Willebrand factor (VWF) multimer analysis is essential for diagnosing and classifying von Willebrand disease (VWD) but requires expert interpretation and is subject to inter‐rater variability. We developed an automated image analysis pipeline using deep learning to improve the reproducibility and efficiency of VWF multimer pattern classification.MethodsWe trained a YOLOv8 deep learning model on 514 gel images (6168 labeled instances) to classify VWF multimer patterns into 12 classes. The model was validated on 192 images (2304 instances) and tested on an independent set of 94 images (1128 instances). Images underwent preprocessing, including histogram equalization, contrast enhancement, and gamma correction. Two expert raters provided ground truth classifications.ResultsThe model achieved 91% accuracy compared to Expert 1 (macro‐averaged precision = 0.851, recall = 0.757, F1‐score = 0.786) and 87% accuracy compared to Expert 2 (macro‐averaged precision = 0.653, recall = 0.653, F1‐score = 0.641). Inter‐rater agreement was very high between experts (κ = 0.883), with strong agreement between the model and Expert 1 (κ = 0.845) and good agreement with Expert 2 (κ = 0.773). The model performed exceptionally well on common patterns (F1 > 0.93) but showed lower performance on rare subtypes.ConclusionAutomated VWF multimer analysis using deep learning demonstrates high accuracy in pattern classification and could standardize the interpretation of VWF multimer patterns. While not replacing expert analysis, this approach could improve the efficiency of expert human review, potentially streamlining laboratory workflow and expanding access to VWF multimer testing.

AbstractBackgroundAs low titer group O whole blood (LTOWB) increases in popularity, blood centers are finding innovative ways of maintaining the supply. One potential way is collecting LTOWB from parous female donors without testing for HLA antibodies. This in silico simulation predicted the risk of an LTOWB unit containing an HLA antibody and the subsequent risk for an HLA‐incompatible transfusion.MethodsAn LTOWB blood bank with 1 million units was simulated consisting of male, nulliparous, and parous female donors. The proportion of each donor type was modeled after the sex distribution at US blood centers. The parity of female donors was calculated based on the average number of live births per female depending on her age. HLA‐alloimmunization risk was determined by her parity status. The HLA haplotypes of the simulated recipients were derived from the 100 most common HLA haplotypes in the US National Marrow Registry Program database. The proportion of different race/ethnic groups in the US was used to simulate 100,000 LTOWB recipients to whom between 1 and 10 units were administered.ResultsOverall, the risk of an LTOWB unit containing at least one HLA antibody was 12.2% and the rate of receiving an HLA‐incompatible unit was 21.3%. The risk of receiving an HLA‐incompatible unit rose from 4.8% after receipt of one unit to 36.5% after 10 units.ConclusionBlood collectors and hospitals should evaluate the potential TRALI risk against the benefit of a potentially expanded inventory of LTOWB before collecting plasma‐containing products from non‐HLA‐tested parous female donors.

Abstract Background and Objectives Artificial intelligence (AI) has been gaining increasing interest in healthcare. During the 2024 International Society of Blood Transfusion (ISBT) Congress, the Clinical Transfusion Working Party (CTWP) conducted a session to explore the exciting intersection of AI in transfusion medicine (TM) practice, education and research. We report here the potential applications and the session outcome. Materials and Methods A pre‐workshop survey explored the participants' demographics and areas of use of AI and whether they have had any AI‐specific training or education. The workshop included presentations on the regulatory aspects of AI use and its application in TM practice, education and research. These were followed by round‐table discussions to explore participants' experience and concerns. Results The workshop had 72 attendees, with 38% falling in the 36–45‐year age group. A total of 70% indicated the use of AI, but only 12% reported having specific training or education. Participants expressed interest in different potential applications but also shared concerns on over‐reliance, potential loss of skills, the accuracy of provided information and content plagiarism. Conclusion The findings of the workshop highlight the need for training, educational resources, standards and regulatory frameworks to guide the use of AI tools in the field of TM.

Abstract Objectives Fluorescence resonance energy transfer (FRET)–based ADAMTS13 activity assays are critical for the diagnosis of thrombotic thrombocytopenic purpura. However, these assays are susceptible to interference. As iodide has been suggested to interfere in laboratory testing via fluorophore quenching or promotion, we aimed to determine whether iodinated contrast (Omnipaque) interferes with the ATS-13 ADAMTS13 Activity Assay 2.0. Methods We evaluated the excitation, emission, and absorbance spectrum of Omnipaque alone and spiked in patient plasma with known ADAMTS13 activity and ADAMTS13 activity on Omnipaque alone, an abnormal control of patient plasma previously observed to display elevated baseline relative fluorescent units, and variable concentrations of patient plasma with known ADAMTS13 activity spiked with Omnipaque. Results No atypical fluorescent peaks were observed on any sample (Omnipaque alone or spiked in plasma) between 250 and 700 nm. There was no difference in the mean ADAMTS13 activity among the various concentrations of plasma spiked with Omnipaque or plasma spiked with saline. Conclusions Iodinated contrast does not appear to interfere—either via spectral overlap of the fluorophore or through physiologic inhibition of the ADAMTS13 enzyme—with ADAMTS13 activity FRET-based assays based on the findings from this in vitro analysis. Delaying sample collection for ADAMTS13 activity testing from suspected patients with thrombotic thrombocytopenic purpura following administration of iodinated radiocontrast agents is not necessary, and recent contrast administration should not yield erroneous ADAMTS13 activity results.