Clinically validated machine learning algorithm for detecting residual diseases with multicolor flow cytometry analysis in acute myeloid leukemia and myelodysplastic syndrome

• Published in: EBioMedicine Vol. 37; pp. 91 - 100
• Main Authors: Ko, Bor-Sheng, Wang, Yu-Fen, Li, Jeng-Lin, Li, Chi-Cheng, Weng, Pei-Fang, Hsu, Szu-Chun, Hou, Hsin-An, Huang, Huai-Hsuan, Yao, Ming, Lin, Chien-Ting, Liu, Jia-Hau, Tsai, Cheng-Hong, Huang, Tai-Chung, Wu, Shang-Ju, Huang, Shang-Yi, Chou, Wen-Chien, Tien, Hwei-Fang, Lee, Chi-Chun, Tang, Jih-Luh
• Format: Journal Article
• Language: English
• Published: Netherlands Elsevier B.V 01.11.2018; Elsevier
• Subjects: Acute myeloid leukemia; Advanced Basic Science; Artificial intelligence; Disease-Free Survival; Female; Flow Cytometry; Humans; Internal Medicine; Leukemia, Myeloid, Acute - blood; Leukemia, Myeloid, Acute - mortality; Leukemia, Myeloid, Acute - therapy; Machine Learning; Male; Minimal residual disease; Multicolor flow cytometry; Myelodysplastic syndrome; Myelodysplastic Syndromes - blood; Myelodysplastic Syndromes - mortality; Myelodysplastic Syndromes - therapy; Neoplasm, Residual; Predictive Value of Tests; Research paper; Survival Rate; Minimal residual disease; Multicolor flow cytometry; Acute myeloid leukemia; Artificial intelligence; Myelodysplastic syndrome
• ISSN: 2352-3964; 2352-3964
• DOI: 10.1016/j.ebiom.2018.10.042

Multicolor flow cytometry (MFC) analysis is widely used to identify minimal residual disease (MRD) after treatment for acute myeloid leukemia (AML) and myelodysplastic syndrome (MDS). However, current manual interpretation suffers from drawbacks of time consuming and interpreter idiosyncrasy. Artificial intelligence (AI), with the expertise in assisting repetitive or complex analysis, represents a potential solution for these drawbacks. From 2009 to 2016, 5333 MFC data from 1742 AML or MDS patients were collected. The 287 MFC data at post-induction were selected as the outcome set for clinical outcome validation. The rest were 4:1 randomized into the training set (n = 4039) and the validation set (n = 1007). AI algorithm learned a multi-dimensional MFC phenotype from the training set and input it to support vector machine (SVM) classifier after Gaussian mixture model (GMM) modeling, and the performance was evaluated in The validation set. Promising accuracies (84·6% to 92·4%) and AUCs (0·921–0·950) were achieved by the developed algorithms. Interestingly, the algorithm from even one testing tube achieved similar performance. The clinical significance was validated in the outcome set, and normal MFC interpreted by the AI predicted better progression-free survival (10·9 vs 4·9 months, p < 0·0001) and overall survival (13·6 vs 6·5 months, p < 0·0001) for AML. Through large-scaled clinical validation, we showed that AI algorithms can produce efficient and clinically-relevant MFC analysis. This approach also possesses a great advantage of the ability to integrate other clinical tests. This work was supported by the Ministry of Science and Technology (107-2634-F-007-006 and 103–2314-B-002-185-MY2) of Taiwan.