Background and Objective
Measurable residual disease (MRD) status is strongly associated with clinical outcomes in haematological malignancies and can be monitored by multi-colour flow cytometry. Conventional MRD analysis by manual gating is time-consuming, operator-dependent and requires significant expertise. Challenges unique to MRD analysis include small size of the target population, and unpredictable target population immunophenotype due to disease heterogeneity and immunophenotypic shifts during treatment. FlowSOM is a clustering tool that employs unsupervised machine learning for analysis of flow cytometric data. This study aimed to explore the performance of FlowSOM in analysis of T-lymphoblastic leukaemia (T-ALL) MRD flow cytometric data.

Methods
Flow cytometric data files from samples sent to Queen Mary Hospital for T-ALL MRD testing from January 2021 to March 2022 were retrieved. FlowSOM was applied retrospectively for automatic clustering of events and the antigen expression profile for each cluster was reviewed by qualified pathologist/flow cytometrist to identify cluster(s) representing residual disease. Results were compared with those obtained using manual gating.

Results
A total of 142 tubes from 36 samples were analysed. These included different leukaemia-associated immunophenotypes (LAIPs), including early T-cell precursor (ETP-ALL) or near-ETP phenotypes, and samples with antigenic shifts during monitoring. Sampled timepoints included post-induction chemotherapy, post-consolidation, maintenance, pre-haemopoietic stem cell transplant (HSCT) and post-HSCT.

Residual disease was positive in 14 and negative in 22 samples by conventional analysis. Analysis by FlowSOM achieved a 100% concordance rate in terms of positivity / negativity call for each sample.

Amongst 142 tubes analysed, residual disease was quantifiable in 62 tubes by conventional analysis. MRD level ranged from 0.038% to 35.146%. Residual disease was quantifiable by FlowSOM in 60 tubes. Residual disease identified on conventional analysis was not identifiable by FlowSOM in 3 tubes. On the other hand, residual disease was quantifiable by FlowSOM in 2 tubes where conventional analysis failed to isolate the leukaemic population, and in total, there were 23 tubes where FlowSOM analysis outperformed human analysis in terms of quantification accuracy.

MRD levels derived from conventional analysis and FlowSOM analysis showed clinically acceptable agreement and high degree of correlation (R2 = 0.997).

Conclusion
This study serves as proof of concept for the use of machine in clinical flow cytometric data analysis. FlowSOM facilitates quick, accurate and reproducible analysis of flow cytometric data in T-ALL MRD monitoring and can serve as an alternative or supplementary method to conventional analysis in the clinical laboratory.