Модель обработки и классификации сигналов ЭКГ на основе интеграции CNN, LSTM и метода двойного внимания

Abstract

Listen

Translate article

Heart rhythm abnormalities diagnosis or classification via electrocardiogram (ECG) in the field of signal processing is regarded as a challenging problem. Classifying ECG signals under the rules of the "Association for the Advancement of Medical Instrumentation (AAMI)" and the inter-patient scheme increases the difficulty of the challenge. AAMI classifies the ECG signal into the following categories: N, S, V, F, and Q. The S category is difficult to detect since its shape pattern is like the N category. To overcome these challenges, this study proposes a novel, robust, end-to-end, and efficient model for ECG signal classification by integrating the "Convolutional Neural Network (CNN)", "Long Short-Term Memory (LSTM)" and dual attention techniques. The architecture of the model includes an encoder-decoder unit built from LSTMs and a dual attention unit with two levels: the first is based on Squeeze-and-Excitation (SE) networks that are incorporated into CNN sections, and they are responsible for weighting the Local Heartbeat Shape Pattern (LHSP) features caught by CNN sections, while the second is based on the Bahdanau scheme, which is incorporated into the encoder-decoder unit, and it is responsible for weighting the Global Heartbeat (GH) features caught by the LSTMs. The suggested model can extract the distinctive features that balance local and global information for optimizing the ECG signal diagnosis. The presented approach is superior to previous methods in the literature and comparable to a cardiologist's performance. The introduced model attained the following F1-Scores: 99.12%, 99.89%, 96.48%, and 99.89% for classifying F, V, S, and N beats, respectively.