Flow Cytometry in AML: What Ten Years of EuroFlow Data Teach About AI and Quality

What this multicentre study actually establishes

Bonte and colleagues analysed ten years of standardised EuroFlow diagnostic acute myeloid leukaemia (AML) flow cytometry data from eight centres. They began with 5,366 FCS files from 885 patients. After compensation quality control and annotation selection, 4,078 files representing 683 patients were retained, and 433 patients had complete tube sets for key downstream clustering and machine-learning analyses. The distinction matters: performance estimates do not describe every patient originally collected. A FlowSOM representation enabled a comparison across centres without additional batch correction in this particular dataset. The genetic-class random forest achieved mean cross-validated balanced accuracy of 0.68 and balanced accuracy of 0.69 in a separate 73-patient validation cohort. These are exploratory study-specific results, not an authorised molecular diagnosis or a validated clinical decision system. Source: Bonte et al., Cytometry Part B (2026).

Quality control precedes artificial intelligence

A computational classifier cannot recover information that was never recorded or that was removed during preprocessing. Preserve laboratory, instrument, fluorochrome configuration, tube version, specimen type, compensation matrix, sample age, acquisition settings and software version as explicit metadata. Track each exclusion at file and patient levels. Compensation failures can have clinically non-random distributions; always examine the retained cohort before discussing performance. In this study, stringent compensation QC removed files, so the final analysis cohort was smaller than the initial collection. The authors note that the QC procedure did not distinguish overcompensation from fluorescence spreading.

Standardisation is valuable but not magical. An absent detectable batch effect under a selected test does not guarantee platform portability, analytical equivalence or freedom from future drift. Investigate site-specific marker distributions, reagent changes and patient case-mix before choosing to correct or combine datasets.

Check the biological denominator

FlowSOM metacluster percentages derived from non-lymphocyte events describe the mixture selected by the analysis, not necessarily the fraction of residual leukaemic blasts. Different marrow recovery states and different blast burdens may change the apparent patient signature. For every predictive feature, record its parent population and denominator. A reproducible model may be predicting a mixture of neoplastic and non-neoplastic differentiation rather than a single molecular lesion.

Why label verification belongs inside model governance

Historic WHO classifications are not interchangeable with more recent classifications. A disagreement between a flow-based hypothesis and a stored molecular diagnosis can reveal wrong metadata, missing molecular analysis, overlapping abnormalities or a real model error. Preserve the original label, send the case for blinded adjudication and record the eventual reason for a change. Report model performance before and after label reconciliation; otherwise a cleaned retrospective dataset may make a system look more accurate than it would be in live practice.

External validation must challenge intended use

Cross-validation tests generalisation within the source data system. A clinical-facing model also needs independent hospitals or temporal holdouts, documented software and panel changes, calibration, class-specific confidence intervals and predefined rules for unknown or insufficient-quality cases. Minority classes require deliberate assessment, not just an attractive overall score. A flow-based class hypothesis could help triage confirmatory testing only after its use case is prospectively evaluated; the paper does not demonstrate a patient outcome benefit or support selecting therapy without genetic confirmation.

Daily Learning Pill — batch correction can erase biology

Do not correct batch effects automatically. First show that a site-dependent difference exceeds expected clinical and biological variation. Preserve raw data and test whether correction reduces true within-patient differences. Uniform plots may be less truthful than carefully qualified heterogeneity.

Related CFCM learning and next step

Explore the CFCM clinical flow cytometry library for panel design, quality control and validation; compare your own pipeline's source-to-feature chain with the four checks above before proposing an AI assay.

Source and editorial scope

Technical Review, based on the publisher's Introduction, Methods 2.1–2.3, Results 3.1–3.6 and Discussion. Supplementary tables and raw FCS files were not independently reanalysed. This original CFCM educational analysis does not claim independent review, RUO clearance or clinical validation. DOI 10.1002/cyto.b.70077.