Abstract:

Objective To enhance the accuracy and reliability of 12-lead electrocardiogram (ECG) automatic diagnosis. Methods Herein we propose a 12-lead ECG automatic diagnosis model based on deep feature fusion (MRHL-ECGNet), which consists of a multi-scale feature extraction front-end, ResNet-34, a global feature mixing module, and a time-series analysis module. The Hyena Hierarchy Convolution Operator was applied to the 12-lead ECG automatic diagnosis task for more efficient capture of long-range dependencies while reducing computational complexity. Integrated Gradients (IG)-based interpretability analysis technology was used to achieve visualization of the decision-making basis of MRHL-ECGNet. The CPSC2018 dataset was used to train and test MRHL-ECGNet, and its performance was assessed using multiple quantitative evaluation indicators and evaluation experiments. Results In the 9-class ECG classification task on the test set, MRHL-ECGNet achieved an accuracy of 0.972, an AUC of 0.983, an F1 score of 0.864, a precision of 0.873, and a recall of 0.857, all surpassing other comparative models. This model only took 0.007 s to output a diagnosis for a single sample on a GPU and 0.156 s on a CPU, with a memory footprint of 67.196 MB. Conclusion The proposed MRHL-ECGNet model demonstrates excellent classification performance in 12-lead ECG automatic diagnosis with a lightweight design and good interpretability, and thus has great potential for clinical application in ECG-aided diagnosis.

Key words: electrocardiogram automatic diagnosis, deep learning, Hyena Hierarchy Convolution Operator, interpretability of model