Automated Detection and Localization of Myocardial Infarction with Staked Sparse Autoencoder and TreeBagger

Novel techniques in deep learning networks are proposed for the staked sparse autoencoder (SAE) and the bagged decision tree (TreeBagger), achieving significant improvement in detection and localization of myocardial infarction (MI) from single-lead electrocardiograph (ECG) signals. With our layer-wise training strategies, the SAE-based diagnostic feature extraction network can automatically and steadily extract the deep distinguishing diagnostic features of the single-lead ECG signals and avoid the vanishing gradient problem. This feature extraction network is formed by stacking shallow SAEs. In addition, to automatically learn the stable distinctive feature expression of the label-less input ECG signals, this feature extraction network adopts unsupervised learning. Moreover, TreeBagger classifier can optimize the results of multiple decision trees to more accurately detect and localize MI. The experiment and verification datasets include healthy controls, various types of MI with anterior, anterior lateral, anterior septal, anterior septal lateral, inferior, inferior lateral, inferior posterior, inferior posterior lateral, lateral, posterior, and posterior lateral, from PTB diagnostic ECG database. The evaluation results show that the new techniques can effectively and accurately detect and localize the MI pathologies. For MI detection, the accuracy, the sensitivity, and the specificity rates achieve as high as 99.90%, 99.98%, and 99.52%, respectively. For MI localization, we obtain consistent results with the accuracy of 98.88%, sensitivity 99.95%, and specificity 99.87%. The comparative studies are conducted with the state-of-the-art techniques, and significant improvements by our methods are presented in the context. Success in the development of the accurate and comprehensive tool greatly helps the cardiologists in detection and localization of the single-lead ECG signals of MI.