Local Deep Field for Electrocardiogram Beat Classification

Abstract

To reduce the high mortality rate among heart patients, electrocardiogram (ECG) beat classification plays an important role in computer aided diagnosis system, but this issue is challenging because of the complex variations of data. Since ECG beat data lie on high-dimension manifold, we propose a novel method, named “local deep field”, in purpose of capturing the devil in the details of such data manifold. This method learns different deep models within the local manifold charts. Local regionalization can help models focus on the particularity of local variations, while deep architecture can disentangle the hidden class information within local distributions. The advantage of the proposed method has been experimentally demonstrated in terms of MIT-BIH Arrhythmia database.