Automated quantification of measurable residual disease in chronic lymphocytic leukemia using an artificial intelligence-assisted workflow

Abstract

Detection of measurable residual disease (MRD) in chronic lymphocytic leukemia (CLL) is an important prognostic marker. The most common CLL MRD method in current use is multiparameter flow cytometry, but availability is limited by the need for expert manual analysis. Automated analysis has the potential to expand access to CLL MRD testing. We evaluated the performance of an artificial intelligence (AI)-assisted multiparameter flow cytometry (MFC) workflow for CLL MRD. We randomly selected 113 CLL MRD FCS files and divided them into training and validation sets. The training set (n = 41) was gated by expert manual analysis and used to train the AI model. We then compared the validation set (n = 72) MRD results obtained by the AI-assisted analysis versus those by expert manual analysis using the Pearson correlation coefficient and Bland–Altman plot method. In the validation set, the AI-assisted analysis correctly categorized cases as MRD-negative versus MRD-positive in 96% of cases. When comparing the AI-assisted analysis versus the expert manual analysis, the Pearson r was 0.8650, mean bias was 0.2237 log₁₀ units, and the 95% limit of agreement (LOA) was ±1.0282 log₁₀ units. The AI-assisted analysis performed sub-optimally in atypical immunophenotype CLL and in cases lacking residual normal B cells. When excluding these outlier cases, the mean bias improved to 0.0680 log₁₀ units and the 95% LOA to ±0.2926 log₁₀ units. An automated AI-assisted workflow allows for the quantification of MRD in CLL with typical immunophenotype. Further work is required to improve performance in atypical immunophenotype CLL.