Automated <scp>B</scp> ‐cell and plasma cell identification using unsupervised clustering by <scp>FlowSOM</scp> and an excel‐based classification model

Abstract Manual gating for plasma cell (PC) identification in multiparametric flow cytometry (MFC) is time‐consuming and operator‐dependent, especially when PCs are scarce. Artificial intelligence approaches such as unsupervised clustering (e.g., FlowSOM) map high‐dimensional data that still require expert interpretation. The primary aim of this work was to develop a lightweight, transparent, spreadsheet‐based algorithm for a classification model that integrates with automated clustering outputs for standardized B‐cell and PC identification in research flow cytometry datasets. Bone marrow aspirates were stained with a standard BD OneFlow™ PC screening tube (CD38, CD56, β2‐microglobulin, CD19, cyIgκ, cyIgλ, CD45, CD138) and acquired on a BD FACSLyric™ flow cytometer. FCS files were exported to CellEngine cytometry software and singlet nucleated events underwent FlowSOM clustering (8 clusters/sample). Cluster‐level median fluorescence intensities (MFIs) were exported to an Excel “PC Trainer Classifier” that (i) normalizes markers to an in‐sample B‐cell anchor, (ii) computes a PC score with CD138 as a hard gate and CD38 as a soft gate, plus secondary features (CD19↓, CD45↓, CD56↑), (iii) applies a forced core fallback (highest CD38/CD138 core score) when strict criteria yield no PCs, and (iv) derives a NEO score (CD56↑, CD19↓, CD45↓) for neoplastic phenotype. The rule‐based classifier was trained on expert assigned PC and B‐cell clusters from 40 samples (30 clonal and 10 polyclonal). Validation was done on a new set of 52 samples, independent of the model. Elements of the Excel formula design and error‐proofing were co‐developed with ChatGPT (OpenAI); all outputs were verified by the authors. Across 52 validation cases (8 clusters/case; 416 clusters total), B‐cell detection achieved: Sensitivity 0.902 (0.79–0.96), Specificity 0.984 (0.96–0.99), Precision 0.885 (0.77–0.94), Accuracy 0.973 (0.95–0.99) and F1 0.893. PC identification achieved: Sensitivity 0.651 (0.54–0.75), Specificity 0.828 (0.79–0.86), Precision 0.458 (0.37–0.55), Accuracy 0.796 (0.76–0.83) and F1 0.537. A transparent Excel‐based classifier integrated with FlowSOM clustering enables highly reproducible B‐cell identification and provides a structured approach to PC classification in research flow cytometry datasets. While the B‐cell classifier demonstrated excellent discriminatory performance, PC identification yielded moderate sensitivity and precision, likely reflecting underlying biological and phenotypic heterogeneity. Consequently, the PC classification component is best interpreted as a triage or augmented‐intelligence tool intended to support, rather than replace, expert assessment. This approach provides a structured and auditable framework that reduces operator dependency and improves inter‐case harmonization. Its interpretability, low cost and portability make it particularly suited to research laboratories operating in resource‐variable settings. Further optimisation and prospective validation may refine PC classification performance.