Detection and Localization of Acute Myocardial Infarction Using Branch-Reduced Convolutional Neural Network

Myocardial infarction (MI) can be life-threatening and needs to be detected accurately and quickly. In this article, a branch-reduced network (BRN) is developed to detect and locate acute MI (AMI) using single-lead electrocardiogram (ECG) signals and 1-D convolutional neural network (CNN). A grouping method for single-lead ECG signals is introduced to reduce the neural network size. A periodic feature-embedded convolutional block attention module (PFECBAM) is proposed to capture periodic features in the input ECG signals and improve AMI detection performance. The proposed model has achieved AMI detection sensitivity, specificity, and accuracy of 95.82%, 93.38%, and 95.16%, respectively. The AMI localization accuracy is 85.31%. The proposed model is of 28 863 trainable parameters, which is two orders of magnitude smaller than the number of trainable parameters in the related works. The compact network structure makes it possible to implement the proposed model on low-power handheld devices for emergency rooms, remote areas, and developing countries.