Effective tree-based classification for automated flow cytometry data analysis on samples with suspected haematological malignancy

Flow cytometry is a commonly used diagnostic technique for haematological malignancies. The gold standard method for analysis of flow cytometry data is manual gating, which is time consuming and requires a highly skilled operator, generating a bottleneck in the workflow and potentially increasing time to diagnose malignancy. For nearly 20 years attempts have been made at replacing manual analysis with automated algorithms, however these are not deemed accurate enough for clinical practice. Clustering methods have been the focus of previous automated attempts, though supervised methods have been shown to be more accurate and require less manual intervention. Tree-based classification algorithms make decisions using an analogous process to manual gating. One hundred and fifty-two flow cytometry files were generated from peripheral blood samples of patients with suspected haematological malignancies. A trained operator labelled events in these files as one of nine cell types. CART, Random Forest and XGBoost were trained on the labelled dataset and the performance was evaluated against previously published clustering methods. Classification algorithms showed higher mean F1 scores than clustering methods. There was no significant difference between CART, Random Forest and XGBoost mean F1 scores, and all three algorithms showed mean prediction times per sample of less than 25 seconds. Tree-based methods struggled to differentiate B cell subtypes, which show similar phenotypic signatures and present an area for future improvement. This work demonstrates the effectiveness of tree-based classification algorithms for flow cytometry analysis. Overall, CART may offer a solution to automated flow cytometry analysis for the purpose of haematological malignancies due to showing high agreement with manual analysis, and short prediction and training times.