CLL MRD Flow Cytometry AI: Interpreting 0.927 Balanced Accuracy

Weakly supervised UMAP and random forest analysis classified 287 CLL specimens—but excluded below-LoQ suspicious results. Clinical validation demands more than cross-validation.

Published study and clinical question

Dinalankara and colleagues published an original, peer-reviewed paper on 17 February 2026 in Cytometry Part B: Clinical Cytometry (DOI 10.1002/cytob.70008). It examines weakly supervised learning for CLL measurable residual disease using multiparameter flow cytometry. The model learns from specimen-level reported disease status and burden, not manually drawn labels for every event. Agreement with a historical report is not independent proof of biological cell identity or external clinical validity.

The population behind the performance figure

The authors selected 310 cases sharing one antibody panel and excluded the suspicious category below the laboratory's 0.01% limit of quantification. That left 287 specimens: 192 MRD positive and 95 MRD negative. Of these, 285 were peripheral blood, one lymph-node FNA and one bronchial lavage. A BD FACSCanto II acquired at least a million events per specimen between 2020 and 2021. The analysis used 14 scatter and fluorescence features after omitting time, with CD19, CD5, CD20, CD22/CD79b, CD43, CD81, CD45 and light chains among the informative measurements. No multi-instrument or prospective clinical deployment cohort was evaluated.

Algorithm: from events to patient-level cluster proportions

For each training fold, one million sampled cells were embedded into two dimensions with UMAP. K-means divided the embedding into 1,000 clusters. Each specimen was projected onto the embedding and reduced to a vector giving the fraction of cells within every cluster. A random forest used those vectors to estimate an MRD positive/negative label; related regression estimated neoplastic percentage. A cluster's statistical importance is not itself proof that every event within it is neoplastic or that an emerging minor population has been clinically verified.

Threefold stratified cross-validation yielded aggregate balanced accuracy of 0.900 with all 1,000 clusters and 0.927 after feature-importance-driven cluster reduction. These estimates came from folds of the same retrospective case set. They should not be described as external validation, analytical sensitivity at a defined event threshold, a regulatory diagnostic performance claim or an independent cell-level adjudication.

CFCM original diagram illustrating CLL MRD AI case-level training, UMAP clusters, 0.927 internal balanced accuracy, and external validation requirements. — Original CFCM conceptual diagram. Educational synthesis; not clinical validation.

The missing low-frequency challenge set

Excluding the suspicious category means that events below the established 0.01% local quantification threshold were outside the headline performance analysis. These borderline specimens are precisely where an MRD assay needs defensible event statistics, thresholding, reproducibility, flow phenotype reconciliation and reporting rules. The study does not establish LoB, LoD or LoQ for the machine-learning method, nor prove safe classification at 10^-4 or deeper disease levels.

Cluster maps can also reflect batch effects, viability and instrument differences. The authors identify clusters overlapping populations compatible with normal B cells and discuss stage-3 hematogones, which can express CD5, as a possible confusion after anti-CD20 therapy. Their panel did not include CD10. A local transfer study must challenge such biology rather than relying only on cross-validation folds.

Regression, computation and deployment limits

The paper reports regression correlations with clinical percentages, but correlations across high and low burdens do not establish quantitative agreement in the near-negative range. Filtering low-CD19 cells changed the regression result; a feature that improves binary classification may discard useful burden information. The authors also describe substantial computational cost for constructing and projecting training cohorts, although prediction once a model is configured may be faster. Hardware, preprocessing and versioning belong in a laboratory verification plan.

Before considering clinical deployment, define the intended specimen, minimum events, antibody panel mapping and unmixing or compensation conventions. Freeze training and validation cohorts by patient and time, use genuine external prospective cases, audit near-LoQ false negatives, compare with qualified review and define manual escalation for inconclusive results. Training labels derived from historical laboratory reports do not erase the need for independent confirmation.

Daily Learning Pill: internal accuracy is not clinical detectability

Three questions matter more than 0.927: Which ambiguous cases were excluded? Which assay and laboratory supplied the training labels? What prospective external challenge demonstrated safe decisions around the reporting boundary? Until those questions are answered, this is promising AI research rather than a clinically validated automated MRD interpretation.

Sources and CFCM learning

This original CFCM Technical Review covers the publisher's Introduction, Methods, Results, Tables 1-4 and Discussion. We did not reanalyse raw FCS files or claim independent human review, IVDR approval or local clinical validation. Read the full primary paper and explore CFCM clinical flow cytometry resources for rare-event analysis, validation and troubleshooting. Subscribe through the site for further evidence-led reviews.