EventHorizon: A Foundation Model for Clinical Flow Cytometry

Abstract

Clinical flow cytometry is central to diagnosing hematologic malignancies, but its reliance on manual expert interpretation and fixed panel designs limits scalability and adaptability. To address this, we developed EventHorizon, a self-supervised foundation model that combines a hierarchical, marker-aware transformer with self-distillation pre-training. EventHorizon integrates heterogeneous, multi-tube panels into a unified, specimen-level embedding. Pre-trained without labels on 100,937 routine clinical specimens, comprising over 50 billion cells, the model yields frozen embeddings that enable lightweight classifiers to accurately identify recurrent genetic abnormalities in acute myeloid leukemia (AML), including t(15;17) (AUROC 0.991) and an unexpected signal for DDX41 mutations (AUROC 0.970). EventHorizon maintained strong performance on a temporally separated 2026 cohort (macro AUROC 0.970 across 28 diagnoses) and demonstrated zero-shot transferability across a five-site B-cell lymphoma cohort (CLL vs. normal AUROC 0.988), the FlowCAP-II AML challenge (AUROC 0.973), and B-ALL measurable residual disease detection (AUROC 0.920 at [≥]1% disease burden). Although class ranking transferred reliably across sites, decision thresholds shifted; however, these were rapidly recalibrated using as few as four labeled AML cases. Sensitivity decreased at disease burdens below 0.1%, yet embeddings remained robust to simulated assay perturbations. Overall, EventHorizon offers a reusable, panel-agnostic representation for diagnostic flow cytometry.