Machine learning optimized multiparameter radar plots for B-cell acute lymphoblastic leukemia minimal residual...

Area of Science:

• Hematology
• Immunology
• Computational Biology

Background:

• Flow cytometry is crucial for B-cell acute lymphoblastic leukemia (B-ALL) minimal residual disease (MRD) analysis.
• Distinguishing B-lymphoblasts from regenerative B-cells poses a challenge in MRD detection.
• Radar plots offer a potential solution by visualizing multiple markers simultaneously.

Purpose of the Study:

• To develop and optimize radar plots for accurate B-ALL MRD analysis.
• To improve the discrimination between malignant B-lymphoblasts and non-malignant B-cells.

Main Methods:

• Utilized Children's Oncology Group (COG) flow cytometry data from B-ALL cases.
• Employed an automated approach to generate and rank hundreds of radar plots using support vector machine (SVM) models.
• Compared optimized radar plots with principal component analysis (PCA), t-distributed stochastic neighbor embedding (t-SNE), and uniform manifold approximation and projection (UMAP) plots.

Main Results:

• Optimized radar plots significantly improved SVM area under the ROC curve (AUC) for B-ALL MRD detection.
• The performance of optimized radar plots was superior to PCA and comparable to UMAP, with enhanced generalizability.
• Radar plots accurately distinguished B-lymphoblasts from other CD19-positive populations and correlated strongly with serial gating for MRD quantification.

Conclusions:

• Optimized radar plots are effective tools for discriminating between B-lymphoblasts and non-malignant CD19-positive populations in B-ALL MRD analysis.
• The developed radar plot optimization strategy is adaptable to other MRD panels and clinical applications.