چكيده لاتين
Nowadays Immediate diagnosis of fatal ventricular arrhythmias is one of the most challenging medical issues. A large number of yearly deaths are related to heart disease. During the research of experts, heart failure is the highest-ranked among other heart diseases and it is one of the highest-ranked in the overall global per-capita death. One way to diagnose cardiovascular failure is to diagnose and analyze heart failure from ECG signals. Because of the low possibility of survival after malignant arrhythmias, quick detection and immediate action are so important. An efficient portable device is highly needed for fast and precise detection. A device that can constantly monitor talented people, must use a high-performance algorithm and good construction. So far, lots of diagnostic methods have been proposed by scientists. Some methods pay attention to the time features of the ECG, some rely-on frequency content, some methods concentrate on extraction and finding features and parameters of ECG, and some others are relyed on statistical methods. This study handles on a way to have both time and frequency in hand. In this method, non-overlapping three-second intervals from raw ECG signal are converted into colored images using CWT and then provide a two-dimensional time-frequency mapping of those segments. Then these images are fed input to a convolutional neural network for training. Finally, after a sufficient and appropriate amount of training, the coefficients matrix is used to recognize the test specimens on an ARM V3s microcontroller. The dataset used for this study is downloaded from commonly access MIT-BIH and Creighton university databases. This CNN obtained 94.44% of accuracy, 91.71% of sensitivity, and 97.14% of specificity on this dataset, and the response time for the microcontroller is as fast as about 0.7 seconds on the classification of each test set image.
Compared to similar studies, this study uses the machine learning method, but methods presented in previous studies which used a convolutional neural network used a 1-dimensional ECG signal only. This method has also been proposed in another study to classify atrial fibrillation arrhythmia, which provided good results. In a similar study using a deep CNN, accuracy, sensitivity, and specificity were 94.70, 95.93, and 94.38% for the 10-second window, respectively, and for a 2-second window obtained 93.18, 95.32, and 91.04, respectively, and for The 3-second interval, it was reported at 98.4, 92.05 and 99.1, respectively. This study has good accuracy and relatively less detection time for the real-time system compared to the same study. The network of this study is simpler and has fewer layers due to the real-time approach of the system and the ability to run on the microcontroller. Therefore, a relative decrease is felt in some statistics reports.
