Multiple-feature-branch convolutional neural network for myocardial infarction diagnosis using electrocardiogram

Generally, 12-lead electrocardiogram (ECG) is widely used in MI diagnosis. It has two unique attributes namely integrity and diversity. But most of the previous studies on automated MI diagnosis algorithm didn't utilize these two attributes simultaneously. In this paper, a novel Multiple-Feature-Branch Convolutional Neural Network (MFB-CNN) is proposed for automated MI detection and localization using ECG. Each independent feature branch of the MFB-CNN corresponds to a certain lead. Individual features of a lead can be learned by a feature branch, exploiting the diversity among the 12 leads. Global fully-connected softmax layer can exploit the integrity, summarizing all the feature branches. Based on deep learning framework, no hand-designed features are required for analysis. Furthermore, patient-specific paradigm is adopted to manage the inter-patient variability, which is a significant challenge for automated diagnosis. Also, class-based experiment (regardless of the inter-patient variability) is performed. The proposed algorithm is evaluated using the ECG data from PTB diagnostic database. It can achieve a good performance in MI diagnosis. For class-based MI detection and localization, the average accuracies are up to 99.95% and 99.81%, respectively; for patient-specific experiment, the average accuracies of MI detection and localization are 98.79% and 94.82%, respectively. Considering its excellent performance, the MFB-CNN can be applied to computer-aided diagnosis platform to assist the real-world MI detection and localization.