A machine learning framework for cross-institute standardized analysis of flow cytometry in differentiating acute myeloid leukemia from non-neoplastic conditions

This figure illustrates a cross-institute standardization framework for flow cytometry data, where heterogeneous single-cell measurements from different instruments are aligned via shared-marker intersection and transformed into a unified feature space for reliable acute myeloid leukemia (AML) diagnosis.

Sustainable Development Goals

Abstract/Objectives

Flow cytometry (FC) is a key diagnostic tool for acute myeloid leukemia (AML), but standardizing analysis across different laboratories is challenging. This study introduces a validated machine learning framework designed to improve cross-panel AML classification by utilizing shared parameters across various FC protocols. The model was trained on FC data from 215 samples (110 AML, 105 non-neoplastic) collected across five institutions. It relies on 16 common parameters like CD7, CD33, and HLA-DR, showing impressive accuracy with 98.15%, a 99.82% area under the curve (AUC), and high sensitivity and specificity. An independent validation of 196 additional samples confirmed these results, achieving 93.88% accuracy and 98.71% AUC. This research highlights the potential for standardized FC analysis across different panel configurations through machine learning, addressing current challenges in flow cytometry interpretation.

Results/Contributions

Flow cytometry (FC) remains a cornerstone diagnostic tool for acute myeloid leukemia (AML), yet standardizing panels across laboratories presents persistent challenges. Our study introduces a validated machine learning framework enabling cross-panel AML classification by leveraging common parameters shared across diverse FC protocols. We employed FC data from 215 samples (110 AML, 105 non-neoplastic) collected in five institutions using different panel configurations as model training set, and another 196 similarly collected samples (90 AML and 106 non-neoplastic) for independent validation set. The framework employs GMM-SVM classification based on 16 common parameters (FSC-A, FSC-H, SSC-A, CD7, CD11b, CD13, CD14, CD16, CD19, CD33, CD34, CD45, CD56, CD64, CD117, and HLA-DR) that are consistently present across various panel designs. The framework demonstrated robust performance with 98.15 % accuracy, 99.82 % area under curve (AUC), 97.30 % sensitivity, and 99.05 % specificity. Independent validation on 196 additional samples further confirmed the framework's effectiveness, maintaining high performance with 93.88 % accuracy and 98.71 % AUC. This research establishes the viability of standardized FC analysis across diverse panel configurations and instruments through machine learning implementation. The framework's robust performance suggests promising applications for harmonized multi-center FC analysis, potentially resolving current standardization challenges in flow cytometry interpretation. © 2025 The Authors

Keywords

flow cytometryacute myeloid leukemiamachine learningstandardizationcross-paneldetection panelGMM-SVM classificationaccuracyarea under the curvesensitivityspecificityindependent validationmulticenter

Contact Information

李祈均

cclee@ee.nthu.edu.tw