Analysis
AI & Data
AI speeds B-ALL MRD interpretation—but what happens below 0.01%?
Across 260 clinical files from 153 patients, AI-assisted review retained 139/139 MRD-positive files at or above 0.01%, versus 28/32 below that threshold. Expert review and denominator correction remain essential.
Clinical question. A marrow file contains 1.2 million events, but only a tiny fraction might represent B-ALL measurable residual disease (MRD). Can a computer make the search faster without losing the rare cells that matter? Seheult and colleagues tested an AI-assisted transformation of routine flow cytometry files that changes how an analyst reaches a decision. The result is promising—but the boundary between a faster review and an independently validated diagnostic decision remains important.
The study, first published on 22 September 2025 and appearing in Blood Advances volume 10 (2026), tested 260 files from 153 patients: 171 reported MRD-positive and 89 MRD-negative. The original clinical assay was a ten-color B-ALL MRD workflow; the AI system was trained on 31 expert-gated negative controls. This was not a prospective multicentre comparison of AI-only reporting with expert clinical diagnosis.

Compression without discarding the diagnostic target
Cluster-informed downsampling reduced the average number of events per file from approximately 1.2 million to 155,884—an 87% reduction. Crucially, the authors designed the reduction to spare small phenotypically distinctive populations. All 139 conventionally positive cases with MRD at or above 0.01% retained detectable MRD in AI-enhanced exports, whereas the 89 conventional negatives remained negative. These are within-study agreements against conventional interpretation, not proof of perfect sensitivity on new instruments, antibody combinations or patient populations.
The distinction becomes visible below that threshold: of 32 samples called positive below 0.01% conventionally, 28 remained positive after AI-assisted processing (88% agreement). A reassuring headline of 100% positive agreement applies only to the prespecified at-or-above-0.01% subgroup and should never be extended to all rare events. Deliberate downsampling also inflates the apparent percentage of rare cells in the exported file, so the pipeline supplies a correction factor to estimate the original denominator. Analysts must not report an uncorrected exported percentage as the clinical MRD value.
What the network actually learned
A deep neural network classified normal hematopoietic categories, with macro-averaged F1 of 0.86. Performance was uneven: a rare CD24-negative hematogone subset had an F1 of 0.25. The pipeline added a quantitative aberrancy scale and a multidimensional projection to highlight events that differed from the negative controls. Nearly all positive cases could be located rapidly, but some still required extra cluster-based gating. The AI system therefore assisted expert scrutiny rather than eliminating it.
The authors reported an aberrancy AUC of 0.98 against benign mononuclear cells and 0.94 against benign B-lymphoid cells. Those discrimination measures describe their study datasets and must not be translated into patient-level clinical accuracy without independent testing. An unusual hematogone population, treatment effect or altered staining panel can create an out-of-reference phenotype without necessarily indicating leukemia.
Time saved, and the design of the comparison
Analysis of AI-enhanced files took an average of 1.01 minutes per case in one 60-case assessment. A separate expert-comparison cohort found an average of 0.86 minutes for enhanced files versus 4.85 minutes for conventional review. A historical laboratory comparison of 14.16 minutes for the existing workflow was measured under a different design and should not be mistaken for a randomized head-to-head outcome. The time-saving signal is strong within these settings, but total laboratory turnaround also includes specimen preparation, quality control, review, information systems and sign-out.
What a clinical laboratory would have to validate
Before deploying similar software, lock the intended assay, staining reagents, instrument configuration, normalization and reference-population library. Test negative recovery marrows with hematogones, post-therapy immunophenotypic shifts, antigen-loss variants, hemodilution and low-event files. Verify LoB, LoD, LoQ where applicable, precision of reported MRD percentages after downsampling, operator override, audit trails and false-negative rates in the low-MRD range. Run an independent local challenge set with blinded expert comparisons and explicit referral rules for difficult cases.
A software model that makes a plot easier to interpret could improve access to high-quality cytometry. It has not automatically acquired the clinical authority to release a result. The appropriate trajectory is decision support with traceable evidence and validated human oversight—not a claim that AI is more accurate than every specialist.
Daily Learning Pill: When assessing AI for rare events, inspect which events survive preprocessing and the denominator used for the final report, not only AUC and minutes per case.
Primary reference: Seheult JN et al., Blood Advances (2026) 10:58–69, DOI 10.1182/bloodadvances.2025016126. Peer-reviewed original research, published online 22 September 2025. Original CFCM explanatory figure; no reproduced source figure. The source article is published CC BY-NC-ND 4.0, which does not authorize commercial adaptation of its figures.