Evaluation Performance of ECG Leads in Myocardial Infarction Classification Using Deep Learning

Cardiovascular disease has become the leading cause of death globally. In Indonesia, the prevalence of heart disease is about 1.5%. An electrocardiogram (ECG) can be used to diagnose cardiac disorders. The main advantage of ECG is widely available in healthcare services and the result can be obtained faster than other tools. However, it needs expert physicians to interpret the diagnosis. This study tries to improve the performance of AI-based Cardiovascular Disease (CVD) classification using paper-based ECG. In this study, the ECGs are divided into four classes, i.e. ECG images of MI patients, patients that have abnormal heart rate, patients that have a history of myocardial infarction and normal person. The CVD classification using V3- V 4 ECG leads (anterior region) and ResNet152V2 give the area under the curve (AUC) of 0.999, F1-score of 0.973, recall of 0.965, precision of 0.981 and accuracy of 0.982. Moreover, single-lead of V2 using ResNet50 gives an AUC of 0.999, F1-score of 0.978, recall of 0.971, precision of 0.985 and accuracy of 0.982. The best performance is achieved by ResNet50 architecture without data augmentation using all 12-leads clean (the grid line background is removed) ECG. It gives an AUC of 1.000, F1-score of 0.989, recall of 0.986, precision of 0.992 and accuracy of 0.989. The implementation of this system can be used to improve the healthcare services in rural areas that have a limited number of experts to interpret ECG.