flowClust: Keep the Outlier Rule in the Gating Record

What happened to the events outside the model?

A clustering summary can be clear about the retained populations while leaving excluded events almost invisible. In a methods review, ask for the disposition of those events before accepting a final population count. The practical aim is a record that lets another analyst reconstruct what entered and left the analysis.

The original software demonstration

Lo, Hahne, Brinkman and Gottardo introduced flowClust in BMC Bioinformatics in 2009. The package uses multivariate t-mixture models with a Box-Cox transformation, including options for transformation and degrees-of-freedom parameters. Bayesian Information Criterion helps compare candidate cluster counts.

In their graft-versus-host disease dataset demonstration, the authors first cluster scatter parameters and then use fluorescence parameters within the selected population. Outlier rules are configurable, and subsetting removes outliers by default. The paper notes that a relatively flat model-selection curve may leave several plausible solutions requiring expert judgement. A software demonstration of this type does not establish clinical diagnostic performance for an unrelated assay.

Write down the exclusions before naming the clusters

A local review can use three short records: the inputs to the fit, the rule governing exclusions, and the evidence supporting each biological label. Keep them separate so that changing a label does not silently change the event-selection history. If the team chooses among several plausible fits, document why the chosen solution serves the intended analysis.

As a handover exercise, ask another analyst to trace one excluded event category through the workflow. Is the rule stored with the result? Can they tell whether exclusion happened during preprocessing, fitting or extraction? If those stages are indistinguishable in the record, improve the audit trail before comparing outputs across analysts.

A historical method is not a current implementation claim

This review reads the original paper for its analytical lessons. It does not assert that a current software release has identical defaults, dependencies or behaviour. Verify the actual installed version and its documentation when implementing a workflow. Clinical adoption also requires validation suited to the local use case; a familiar cluster shape is insufficient evidence.

Daily Learning Pill

A population label does not explain an exclusion. Keep the event-selection rule accessible beside the model result.

Bring this audit question to your next methods meeting and explore the CFCM cytometry resources. Original ChatGPT/Codex educational analysis, not independent scientific review or clinical validation.