Artificial Intelligence-Enabled 8-Channel ECG Diagnosing of Abnormalities with Wide QRS Complexes

TL;DR

This study shows that an AI model using 8-channel ECG data can accurately diagnose heart rhythm issues with wide QRS complexes, outperforming both 4- and 12-channel formats and human experts.

Contribution

The novel contribution is the development of an AI model using 8-channel ECG data to accurately classify abnormalities with wide QRS complexes, outperforming both fewer channels and human experts.

Findings

The AI model using 8-channel ECG data achieved 95.0% accuracy, outperforming both 4- and 12-channel formats.

The model's performance on the JX-Test set showed a mean F1 score of 0.917 and a mean AUROC score of 0.994.

The model outperformed cardiologists in accuracy, F1 score, and AUROC on the same datasets.

Abstract

Background: There is a substantial number of research exploring the application of artificial intelligence (AI) in identifying electrocardiogram (ECG) abnormalities related to heart rhythm or conduction with the 12-channel format. However, there is a scarcity of studies focusing on refined differentiation of serials of ECG abnormalities with wide QRS complexes in a simplified channel format. Methods: We constructed an ECG dataset (standard 10-s, 12-channel format) from adult patients from Tongji Hospital of Huazhong University of Science and Technology, Wuhan, China. This dataset was consisted of 5 kinds of ECG abnormalities with wide QRS complexes in the normal heartbeat (60 to 100 beats per minute) and the normal ECGs. Convolutional neural network was developed to classify these abnormalities. Four-channel (I, II, V1, and V5) and 8-channel (I, II, and V1 to V6) formats, compared to the standard 12-channel format (I, II, III, aVR, aVL, aVF, and V1 to V6), were chosen as the input channel format of the model. Other unreplicated ECGs from Tongji Hospital (TJ-Test set), annotated by a committee of board-certified cardiologists, served as the test dataset. The F1 score, area under the receiver operating characteristic curve (AUROC), and accuracy were calculated to assess the performance of the model, which were further compared with diagnoses of 6 ECG cardiologists who were informed that the final objective was classifying among 6 classes with the 12-channel format. In addition, a dataset of 291 ECGs from The First People’s Hospital of Jiangxia District (JX-Test set) and a public dataset of 64 ECGs were used to assess model generalizability Results: The dataset consisted of 11,808 ECGs from 8,542 patients from 2012 January 1 to 2020 November 30 and divided into training and validation datasets in the ratio of 9:1. The test dataset included unreplicated 480 ECGs from 480 new adult patients recorded from 2014 January 1 to 2017 November 30. The model shows a superior performance in the 8-channel format compared to that of 4- and 12-channel formats. As for the 8-channel format, the model obtained an accuracy of 95.0%, a mean F1 score of 0.969 (0.943 to 0.997), and a mean AUROC score of 0.997 (0.975 to 1.00) compared to an accuracy of 89.9%, an F1 score of 0.898 (0.863 to 0.932), and an AUROC score of 0.941 (0.918 to 0.963) of physicians assessing the same datasets. The model exhibited a mean F1 score of 0.917 (0.943 to 0.997) and a mean AUROC score of 0.994 (0.975 to 1.00) on the JX-Test set, and mean F1 scores of 0.708 for the left bundle branch block and 0.828 for the right bundle branch block for the external published validation data both with the 8-channel format. Conclusion: Our model distinguishes a range of distinct abnormalities focusing on abnormal morphology on QRS complexes in normal heartbeats with high accuracy, providing a foundation for AI-aided clinical decision-support systems in ECG differential diagnosis.

Figures