FlowSOM for Plasma Cells: Why an Excellent B-Cell AUC Can Mislead

When a very good AUC hides the harder cell type

Automated gating is increasingly discussed for plasma-cell flow cytometry, yet identifying B cells is a different task from identifying scarce, aberrant plasma cells. Gantana and colleagues combined FlowSOM unsupervised clustering with a transparent rule-based Excel classifier using median fluorescence intensities from a commercial plasma-cell screening panel. Their open-access peer-reviewed study is Cytometry Part B (2026), DOI 10.1002/cyto.b.70058. The question for laboratories is not whether an algorithm can produce a polished ROC curve; it is which cell class and clinical use that curve actually evaluates.

Understand the source cohort before the model

The retrospective dataset comprised 101 diagnostic bone marrow FCS cases investigated for plasma-cell disorders from 2023 to 2025. Nine were excluded for predefined disease, specimen or method reasons, leaving 92 eligible cases. Investigators developed the rule set on 40 cases (30 clonal and ten polyclonal), then evaluated it on 52 different cases. The validation group had 51 clonal and only one polyclonal case. That imbalance is not a footnote: it constrains what predictive values and generalisation claims can mean outside the selected referral population.

Data came from a BD OneFlow plasma-cell screening tube and BD FACSLyric workflow. Approximately 30,000 events were acquired per case at the median. FlowSOM used CD38, CD56, CD19, CD45, CD138 and beta-2 microglobulin, while kappa and lambda channels were left out of the clustering feature space and reserved for downstream interpretation. Eight clusters were evaluated per case, producing 416 validation clusters. Clinical performance at a very different event depth, instrument or panel composition was not demonstrated.

Report B-cell and plasma-cell performance independently

For B-cell clusters, the reported validation sensitivity was 0.902, specificity 0.984, precision 0.885 and AUC 0.988. For plasma-cell clusters, sensitivity fell to 0.651, specificity to 0.828 and precision to 0.458, with an AUC of 0.753. These are cluster-level model metrics against the study reference, not specimen-level clinical sensitivity for measurable residual disease. A laboratory that quotes only the larger B-cell AUC would obscure the more clinically challenging classification problem.

A plasma-cell positive predictive value below one-half in the evaluated clusters also means expert confirmation remains important. Predictive value depends on prevalence and selection: because the validation cohort was strongly enriched for clonal cases, its positive and negative predictive values cannot be transported unadjusted to population screening or remission monitoring.

Why the workflow is interpretable but still fragile

The design deliberately uses a B-cell internal reference to normalise several fluorescence measures after FlowSOM assignment. That makes the scoring pathway visible and potentially easier to audit than a black-box classifier. It also makes the B-cell anchor a dependency. If the expected reference population is missing, poorly gated or biologically altered, a ratio may become unstable even though the spreadsheet continues to output a score. Record such failure modes as an explicit no-call rather than forcing a class.

Plasma-cell phenotype can vary with differentiation, treatment, preanalytics and sample handling. CD138 and CD38 measurement is not invariant across delayed specimens, cell lysis conditions, antibody clones and optical settings. A model tuned on diagnostic marrow cannot silently become a therapy-era MRD method. Also consider doublet exclusion, which can remove unusual larger events, and the difference between cluster medians and the distribution of individual rare cells.

Expert agreement is a separate performance ceiling

Three haematopathologists provided the study reference interpretations. They showed higher complete agreement for B-cell than plasma-cell clusters: 94.7% versus 71.6% across 416 observations. The authors used multiple interobserver coefficients because simple agreement and kappa behave differently under class imbalance. An automation result should be compared with the reference ambiguity, not judged against a fiction that every manual plasma-cell gate has an unambiguous label.

Even excellent agreement with reviewers would not establish patient-level clinical benefit without a predefined intended use. Define acceptable false-negative risk, abnormal population size, event count and when a human must review a discordant cluster before moving toward an LDT or regulated clinical method.

What an MRD-ready validation would require

Design a separate verification set enriched for low-count and post-treatment specimens, abnormal light-chain ratios, antigen-shifted plasma cells and difficult normal regeneration. Stratify by acquisition volume and viable events, assess imprecision between operators and instruments, and report failure-to-classify frequency. Lock the model and thresholds before external validation; a spreadsheet does not remove the need for change control. Retain raw FCS, versioned metadata and a route for expert adjudication.

Daily Learning Pill: specificity to the wrong question

A B-cell AUC of 0.988 says nothing by itself about accurately detecting residual plasma-cell clones. Read the positive class, unit of analysis, denominator and intended-use population before quoting an automation performance metric.

Next step for the laboratory

For any automated plasma-cell gating proposal, request cluster-level and patient-level confusion matrices separately, plus the exact rules for anchor failure and rare-event no-calls. Read more method-validation and flow-cytometry analysis material at CFCM. This Technical Review is an original educational appraisal, not independent peer review or clinical validation of the reported Excel/FlowSOM tool.