| Abstract [eng] |
Myocardial infarction is the leading cause of death in the world. With the rapid development of smart, worn electrocardiogram recorders, there are more and more opportunities to recognize myocardial infarction at an early stage. Record a 12-standard lead electrocardiogram (ECG) using a wrist-worn electrocardiogram recorder, unfortunately, is impossible. In this study, an algorithm for the detection of myocardial infarction in ambulatory electrocardiogram signals with reduced number of leads was developed and investigated. The proposed algorithm was tested using various single leads and combinations of leads. To detect myocardial infarction the algorithm used a hybrid neural network consisting of a convolutional neural network (CNN) and a recursive long-term memory neural network (LSTM). PTB Diagnostic ECG Database from the Physionet was used for the study. ECG recordings from 347 cases of acute myocardial infarction and 80 ECG recordings from healthy volunteers were used. The database is divided into training, validation and testing databases by patient. 10s ECG segments were used to train and test the algorithm. The best results were obtained using three-lead (I, II, III) ECG signals with 92.3 % accuracy, 93 % sensitivity, and 91.7 % specificity. Given that the quality of ECG signals, when recording ECG signals with wrist-worn electrocardiogram recorders, depends on electrode attachment site, elektrode attachment site impact on algorithm accuracy was also investigated in this work. This study shows that while recording chest leads the accuracy of the algorithm decreases by 27.28 % when electrode attachment site deviating from the standard location by approximately 10 cm. When electrode attachment site deviating from the standard location by approximately 6.2 cm, the accuracy of the algorithm decreases by 13.3%. The influence of noise on the performance of the algorithm was also investigated, taking into account noises that may occur when using wrist-worn electrocardiogram recorder. The MIT-BIH Noise Stress Test Database was used for this study. Baseline variation noises, electrode movement artifacts, and muscle contraction artifacts were added to the ECG signals. This study shows that the algorithm is most affected by the artifacts of muscle contractions and electrode movement. Elliptic high-frequency IIR filters were not effective in attempting to filter out these artifacts when their amplitude was increased. In this regard, when recording ECG signals, the patient should be at rest to minimize the artifacts of electrode movement and muscle contractions. |
