DeepECG: Image-based electrocardiogram interpretation with deep convolutional neural networks

Abstract

Electrocardiogram (ECG) plays a critical role in the diagnosis of cardiovascular disease (CVDs). In this paper, we develop DeepECG, a system that diagnoses 7 kinds of arrhythmia from 51,579 ECGs. DeepECG takes ECG images as inputs, and performs arrhythmia classification using deep convolutional neural network models (DCNN) and transfer learning. We conduct a comprehensive study of different neural network architectures, where the best model Inception-V3 achieves mean balanced accuracy of 98.46 %, recall of 95.43 %, and specificity of 96.75 %. The experimental results have successfully validated that our system can achieve excellent multi-label classification based on image formats, making it possible for cardiologists to use image-based ECG interpretation with DCNN to aid diagnosis and reduce misdiagnosis rates.