Accurate classification of plasma cell dyscrasias is achieved by combining artificial intelligence and flow cytometry

Abstract

Summary Monoclonal gammopathy of unknown significance (MGUS), smouldering multiple myeloma (SMM), and multiple myeloma (MM) are very common neoplasms. However, it is often difficult to distinguish between these entities. In the present study, we aimed to classify the most powerful markers that could improve diagnosis by multiparametric flow cytometry (MFC). The present study included 348 patients based on two independent cohorts. We first assessed how representative the data were in the discovery cohort (123 MM, 97 MGUS) and then analysed their respective plasma cell (PC) phenotype in order to obtain a set of correlations with a hypersphere visualisation. Cluster of differentiation (CD)27 and CD38 were differentially expressed in MGUS and MM (P < 0·001). We found by a gradient boosting machine method that the percentage of abnormal PCs and the ratio PC/CD117 positive precursors were the most influential parameters at diagnosis to distinguish MGUS and MM. Finally, we designed a decisional algorithm allowing a predictive classification ≥95% when PC dyscrasias were suspected, without any misclassification between MGUS and SMM. We validated this algorithm in an independent cohort of PC dyscrasias (n = 87 MM, n = 41 MGUS). This artificial intelligence model is freely available online as a diagnostic tool application website for all MFC centers worldwide (https://aihematology.shinyapps.io/PCdyscrasiasToolDg/).

Artificial intelligence: Training the trainer

“…Although I have to admit that ANN are increasingly designed as self‐learning networks, they still require human validation, which just shifts the task a little bit. To address the established haematologists that are in leading positions, I hope the current discussion in the British Journal of Haematology 1–3 will contribute.…”

Multiparameter flow cytometry in plasma cell disorders: when in doubt, go with the flow

“…1,3 Although MFC is called to play a broader role in the post-treatment monitoring setting (Paiva et al 4 ; Goicoechea et al 5 ), the extent of its role at the diagnostic stage remains to be ascertained. In this issue, Clichet et al 6 highlight a potential new role for upfront MFC assessment. In this work, using Gradient-Boosting Machine-learning, they demonstrate and independently validate that a simple routine MFC panel with a handful of markers could be used to both optimise diagnosis and successfully discriminate myeloma from precursor conditions such as monoclonal gammopathy of undetermined significance (MGUS).…”

“…Clichet et al 6 . circumvent the reproducibility issue by using a common, well‐ established panel 17 .…”

“…This explains why, so far, the only firm criteria to differentiate MGUS, SMM, and MM is the percentage of PCs in the marrow. 1 Clichet et al 6 circumvent the reproducibility issue by using a common, well-established panel. 17 By taking an original statistical approach, they show that the percentage of abnormal PCs, CD27 expression, the PC/CD117 + precursors ratio and the total PCs/haematogones ratio were the most influential in predicting disease subtype.…”

“…), the extent of its role at the diagnostic stage remains to be ascertained. In this issue, Clichet et al 6 . highlight a potential new role for upfront MFC assessment.…”