Abstract:
Myocardial Infarction (MI) is a common Cardiovascular Disease (CVD) caused when blood
flow through the coronary artery to a segment of the myocardium is disrupted and causes the
perfusion of imbalance between demand and supply. Electrocardiogram (ECG) is the most
commonly used for the detection and localization of MI. MI may change the ST−segment,
T−waves, and Q−waves of ECG signal. A manual interpretation of the12-lead ECG signal
to identify MI is time consuming, tedious, subjective, and error prone. To overcome these
problems several computer aided diagnoses based on conventional machine learning and deep
learning approach were developed in the past decade. Those methods work well for binary
classification (MI and Healthy Control (HC)) and localization of blockage of the coronary
artery. However, there are other CVDs that deviate ST− segment of the ECG signal and
mimic the ECG signal of MI. These diseases are the primary causes of the difficulties in
detecting MI from an ECG signal, and previously developed methods did not take into ac count these diseases. This research implements a hybrid of Residual Network (ResNet)-Long
Short Term Memory (LSTM) for the classification of CVDs that mimic MI detection and MI
localization from the 12-lead ECG signal. A hybrid deep learning model produces better re sults than a pure Convolutional Neural Network (CNN) or sequential model. The developed
model was trained with the online database PTB-XL and the results show that the achievable
accuracy, precision, recall, and f1-score using this technique are 89%, 85%, 90%, and 87%,
respectively, for CVDs classification. For MI localization, it achieves 93% accuracy, 86%
precision, 91% recall, and 89% f1-score. This indicates that even in the presence of other
CVDs, which are the main cause of pitfalls in MI detection from a 12-lead ECG signal, the
developed model can automatically diagnose MI and localize MI from a 12-lead ECG signal.
