A robust myocardial infarction localization system based on multi-branch residual shrinkage network and active learning with clustering

Generally, 12-lead electrocardiogram (ECG) is regarded as an effective noninvasive method for diagnosing myocardial infarction (MI). However, most previous studies required additional denoising operations and did not propose an effective method to overcome individual differences between patients. In this paper, we design a novel deep learning model named the multi-branch residual shrinkage network (MB-RSN) to locate MI via 12-lead ECG signals without denoising. It includes 12 branches that can automatically extract the heartbeat feature of the corresponding lead. Each branch is mainly composed of residual shrinkage blocks, and the shrinkage module eliminates unimportant features by a soft threshold function. Finally, all branch features are aggregated for MI localization. Also, to overcome individual differences and reduce the cost of manual labeling, we employ active learning (AL) to optimize the model. In particular, we proposed a novel query strategy called Best-versus-Second-Best with k-means (k-BvSB). k-BvSB can simultaneously consider the uncertainty and diversity of unlabeled samples to select the most valuable unlabeled samples. The proposed model and query strategy are evaluated under the intra-patient and patient-specific schemes using the PTB diagnostic database. The MB-RSN achieves accuracy and F1 of 99.89% and 99.88% under the intra-patient scheme. For the patient-specific scheme, the MB-RSN obtains accuracy and F1 of 98.35% and 98.19% based on k-BvSB. Compared with other studies on MI localization, our system achieves state-of-the-art performance. Therefore, it offers great potential for application in real-world MI diagnosis.