Abstract 13745: Exploring the Segment-Wise Explainability of Electrocardiogram on Deep Learning-Based Models for Acute Myocardial Infarction

• Published in: Circulation (New York, N.Y.) Vol. 148; no. Suppl_1
• Main Authors: Ji, Younghoon, Cho, Kyung Hoon, Chang, Mineok, Joo, Sunghoon, Park, Minje, Kim, Kyung Geun, Lee, Yeha, Kim, Min Chul, Sim, Doo Sun, Hong, Young Joon, kim, Juhan, Ahn, Young Keun, Jeong, Myung Ho, Shin, Dong-Geum, Cho, Jung Rae, Kang, Min-Kyung, Seonghoon, Choi, Lee, Namho, Han, Donghoon
• Format: Journal Article
• Language: English
• Published: 07.11.2023
• ISSN: 0009-7322; 1524-4539
• DOI: 10.1161/circ.148.suppl_1.13745

Abstract only Background: While DL models for AMI detection using ECGs have advanced, their 'black box' nature hinders seamless integration into clinical practice, as it limits our understanding of their decision-making process. Research Questions: Which specific ECG segments (P, QRS, T) are considered important by the DL model for the detecting AMI, and does it fit into the ECG criteria of AMI? Aims: We aimed to examine the importance of each ECG segment in the detection of DL model for AMI. Methods: We examined the impact of various ECG segments by replacing their values with Gaussian noise and verify the decrease in AUC of a DL model for detecting AMI. The importance of each ECG segment in diagnosing AMI using a DL model was assessed by observing the performance drop in AUC of the model. DL-based model detecting AMI trained with 17,907 ECGs from 5,872 coronary angiography-validated AMI patients and 705,482 ECGs from 294,755 non-AMI patients at Chonnam National University Hospital (May 2013 - December 2020). External validation was performed at Hallym University Kangnam Sacred Heart Hospital using 357,072 ECGs from 163,870 patients (January 2006 - December 2020). For ECG segmentation, a U-Net architecture trained on LUDB was employed. Results: The test set was divided into STEMI and NSTEMI groups within internal and external sets. The occurrence rates of events in ECGs were 0.498% (1895/380149) for STEMI and 0.67% (2550/380804) for NSTEMI. In the STEMI group, the AUC values were as follows: 12-lead raw ECG - 0.988, P-wave masking ECG - 0.986, QRS-complex masking ECG - 0.943, and T-wave masking ECG - 0.939. In the NSTEMI group, the AUC values were 0.943, 0.936, 0.835, and 0.867, respectively. Conclusion: The DL model assigns varying degrees of importance to individual ECG segments when determining AMI. The results show that the AMI model effectively utilizes ECG characteristics such as QRS and T waves, which are also highly valued in clinical practice, to determine the presence of AMI.