Automating detection and localization of myocardial infarction using shallow and end-to-end deep neural networks

Abstract Myocardial infarction (MI), also known as a heart attack, is one of the common cardiac disorders caused by prolonged myocardial ischemia. For MI patients, specifying the exact location of a heart muscle suffering from blood shortage or stoppage is of crucial importance. Automatic localization systems can support physicians for better decisions in emergency situations. Using 12-lead electrocardiogram, in this paper, two MI detection and localization methods are proposed with classic and end-to-end deep machine learning techniques. For the feature extraction phase, the classic approach performs a Discrete Wavelet Transform (DWT) and Principal Component Analysis (PCA) on the pre-processed signals followed by a shallow neural network (NN) for the classification phase. However, in the end-to-end residual deep learning technique, a Convolutional Neural Network (CNN) is directly employed on the pre-processed input signals. For specifying the infarcted region of myocardium, 6 classes of subdiagnosis are considered. Proposed models are verified with the Physikalisch-Technische Bundesanstalt (PTB) dataset, where the data of each patient is first grouped and then carefully partitioned to training, validation, and test datasets. The results of K-fold cross-validation indicate that the general model achieves over 98% accuracy for both MI detection and localization with fewer number of feature sets compared to previous studies. Moreover, the end-to-end CNN model shows superior performance by achieving perfect results. Thus, with the larger size of CNN models, one may choose a perfect system that requires larger memory compared to another system that requires less computational power and accepts nearly 2% of false positives and/or false negatives.