Background: The low rate of detection of abnormalities has been a major problem with current artificial intelligence-based electrocardiogram diagnostic algorithms, particularly when applied under real-world clinical scenarios.
Methods: We proposed an aggregation attention multilabel electrocardiogram classification model (AA-ECG) that can be applied directly to raw images to identify cardiac abnormalities using image-level annotation only. To develop and validate the model, we conducted a prospective two-site study to build two large-scale real-world datasets of 12-lead electrocardiogram images, annotated by clinical experts in a multilabeled manner. We compared the proposed model with seven state-of-the-art classifiers on both datasets in 27 main categories.
Results: In total, 47,733 electrocardiogram images from 37,442 consecutive patients were included in the development set, while 18,581 from 18,345 in the external set. The proposed model achieved better overall performance than the other seven models.The visualization of the attention maps provided an approach to build medical interpretability for machine intelligence.
Conclusions: The proposed model had high diagnostic accuracy in identifying cardiac abnormalities on two real-world datasets. It has the potential to help clinicians provide more efficient cardiac care with fewer medical resources.
Keywords: Artificial intelligence; Deep learning; Electrocardiogram diagnosis.
Copyright © 2023. Published by Elsevier Ltd.