ECG Quality Assessment Framework by Using Attentional Convolution Neural Network

Abstract

ECG is an important means of diagnosis of arrhythmia. In daily health monitoring, serious noise pollution, reverse leads connection, and so on make cannot meet the requirements of subsequent automatic diagnosis. Thus, it is of great significance to further evaluate the ECG quality and screen out the ECG that meet the requirements of subsequent diagnosis. However, complex interference factors affect the quality of the signal and has brought the huge challenge to quality assessment. Additionally, the current algorithms depend on the wave detection, which also brings additional cumulative error. Meanwhile, the current algorithms cannot intuitively present the attention degree to ECG signals during the assessment process. This paper proposes a novel method (ACNN) for evaluating the ECG quality. ACNN directly targets the whole ECG signal and does not detect the waveform of the ECG signal. Then, ACNN uses convolutional blocks to extract the deep features and designs a novel attention layer to enhance the beneficial features of the results. Finally, the fully connected layer is employed for obtaining the final quality evaluation. Compared with existing methods, ACNN obtains better performance, with 100.0% sensitivity, 83.33% specificity and 98.0% accuracy, which shows ACNN can be applied in clinical scenarios.