Augmented Human Intelligence and Automated Diagnosis in Flow Cytometry for Hematologic Malignancies.

Abstract

Clinical flow cytometry is laborious, time-consuming, and expensive given the need for data review by highly trained personnel such as technologists and pathologists as well as the significant number of normal cases. Given these issues, automation in analysis and diagnosis holds the key to major efficiency gains. The objective was to design an automated pipeline for the diagnosis of B-cell malignancies in flow cytometry and evaluate its performance against our standard clinical diagnostic flow cytometry process. Using 3,417 cases of peripheral blood data over 6 months from our 10-color B-cell screening tube, we used a newly described method for feature extraction and dimensionality reduction called UMAP on the raw flow cytometry data followed by random forest classification to classify cases without gating on specific population. Our automated classifier was able to achieve greater than 95% accuracy in diagnosing all B-cell malignancies, and even better performance for specific malignancies for which the panel was designed, such as chronic lymphocytic leukemia. By adjusting classifier cutoffs, 100% sensitivity could be achieved with an albeit low 14% specificity. Hypothetically, this would allow 11% of the cases to be autoverified without human intervention. These results suggest that a clinical implementation of this pipeline can greatly assist in quality control, improve turnaround time, and decrease staff workloads.