MultiECGNet: Evaluation of a Deep-Learning Ensemble Method for Image-Based 12-Lead Electrocardiographic Detection of Atrial Fibrillation

Abstract

Background

Artificial intelligence holds significant promise for analysis and interpretation of electrocardiograms (ECGs) in cardiovascular disease. However, many ECGs are still stored as paper records or static images, necessitating deep-learning models that can operate on image-based ECG formats. This study aims to evaluate the performance of an ensemble classifier, MultiECGNet, using multi-format ECG images for the diagnosis of atrial fibrillation (AF).

Methods

An ensemble classifier was developed using 4 models derived by truncating the pretrained EfficientNet B3 model at different feature extraction layers. Transfer learning was employed to train the ensemble on the publicly available PTB-XL dataset for AF detection. External validation was performed using the Chinese Physiological Society Signal Challenge dataset. ECG samples were converted into 2 image formats (2 x 6 and 4 x 3), and performance was evaluated across same-format, cross-format, and mixed-format classification tasks.

Results

The classifier detected AF in the validation dataset with an accuracy of 0.95 with an F1-score of 0.87, comparable to a signal-based model (F1-score: 0.87 vs 0.83) and better than a single EfficientNet-B3 model (F1-score: 0.87 vs 0.71). Training on one ECG format and testing on a different format resulted in reduced performance (F1-score: 0.66-0.71). However, training on a dataset containing a mix of both formats closely matched the performance of the original same-format training (F1-score: 0.86).

Conclusions

The proposed image-based ensemble classifier demonstrated comparable performance to a strong signal-based model for AF detection. Although cross-format generalization posed a challenge, incorporating multiple ECG formats during training mitigated this limitation and improved model robustness.