Detection of First-Degree Atrioventricular Block on Variable-Length Electrocardiogram via a Multimodal Deep Learning Method

Automatic detection of first-degree atrioventricular block (I-AVB) from electrocardiogram (ECG) is of great importance in prevention of more severe cardiac diseases. I-AVB is characterized by a prolonged PR interval. However, due to various artifacts and diversity of ECG morphology, existing ECG delineation algorithms is unable to provide robust measurement of the PR interval. Deep neural network is good at extracting high-level feature from ECG waveform, but merely using waveform as input of neural network may aggravate overfitting when lack of I-AVB records. In this paper, we propose a multimodal-input deep learning method to effectively detect I-AVB from 12-lead ECG records. We utilize ECG waveform and delineation result as the multimodal input of neural network. Our neural network, mainly composed of convolutional neural network and Long Short-term Memory, is well designed to adapt to variable-length ECG. Our method is evaluated on dataset of CPSC2018, and outperforms the baseline methods in F1 score.