Detection and localization of myocardial infarction based on a convolutional autoencoder

• Published in: Knowledge-based systems Vol. 178; pp. 123 - 131
• Main Authors: Sugimoto, Kaiji, Kon, Yudai, Lee, Saerom, Okada, Yoshifumi
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier B.V 15.08.2019; Elsevier Science Ltd
• Subjects: Classification; Convolutional autoencoder; Echocardiography; Electrocardiogram; Electrocardiography; Feature extraction; Heart attacks; Heart diseases; k-nearest neighbor; Localization; Methods; Myocardial infarction; Signal processing; Visual observation; Visual signals; Waveforms; Myocardial infarction; k-nearest neighbor; Convolutional autoencoder; Electrocardiogram
• ISSN: 0950-7051; 1872-7409
• DOI: 10.1016/j.knosys.2019.04.023

Twelve-lead electrocardiograms (ECG) are widely used for the diagnosis of myocardial infarction (MI). For MI detection and localization, 12 ECG signals should be comprehensively checked through visual observation. This process is time-consuming, requires significant effort, and is prone to inducing errors. Hence, computer-aided automatic detection technology is required. Many existing methods perform MI detection and localization using features extracted from normal and abnormal ECG data. However, abnormal ECG signals show various waveforms for the same heart disease; therefore, it is difficult to extract the waveform features common to all the waveforms. In addition, ECG data is extremely imbalanced, and the minority class, including abnormal ECG data, may not be adequately learned. Because of the difficulty of feature extraction in the imbalanced data, in this study, we propose a new method for MI detection and localization that learns only normal ECG data in the public ECG database. This method is based on a convolutional autoencoder (CAE) model for normal ECG waveforms. The CAE model is constructed for each lead and outputs reconstructed input ECG data if normal ECG data is inputted. Otherwise, the waveform is distorted and outputted. MI detection and localization is performed by a k-nearest neighbor (k-NN) classifier using an error vector whose dimension corresponds to each lead and whose element is a degree of deviation between the normal ECG data and the output waveform. In the experiments, the classification performance of the proposed method was evaluated using 353640 beats obtained from the ECG data of MI patients (10 class infarct sites) and healthy subjects. Consequently, the proposed scheme demonstrated a classification performance higher than or comparable to that of existing methods, and the false positive and false negative rates could be reduced compared to existing methods.