Robust QRS Detection Using High-Resolution Wavelet Packet Decomposition and Time-Attention Convolutional Neural Network

Abstract

QRS detection is a crucial step in analyzing the electrocardiogram (ECG). For ECG collected by wearable devices, a robust QRS detection algorithm that yields high accuracy in spite of abnormal QRS morphologies and severe noise is needed. In this paper, we propose a QRS detection method based on high-resolution wavelet packet decomposition (HR-WPD) and convolutional neural network (CNN). Firstly, we design the HR-WPD that decomposes the ECG into multiple signals with different frequency bands to provide detailed QRS features. Secondly, all the decomposed signals are forwarded to a CNN for comprehensive morphology analysis and QRS prediction. To further improve the robustness, a time-attention module acting on the input signals is added to the CNN. Finally, a variable threshold is imposed to locate the QRS. The proposed method is validated by using two noisy databases (i.e., Telehealth Database (TELEDB) and MIT-BIH Noise Stress Test Database (NSTDB)) and one database with multiple ECG morphologies (i.e., MIT-BIH Arrhythmia Database (ARRDB)). The experiment results show that the proposed method achieves a comparable or even better performance compared with state-of-art methods on the TELEDB (SE 98.99%, P+ 95.57%, ER 5.61%, F1 97.25%), NSTDB (SE 99.25%, P+ 96.31%, ER 4.55%, F1 97.76%) and ARRDB (SE 99.89%, P+ 99.90%, ER 0.21%, F1 99.89%), suggesting that it is highly applicable to the QRS detection for ECG collected by wearable devices. INDEX TERMS Electrocardiogram, convolutional neural network, wavelet packet decomposition, QRS detect.

DeepCEDNet: An Efficient Deep Convolutional Encoder-Decoder Networks for ECG Signal Enhancement

“…Besides, when the training reaches a certain epoch, the loss difference between the training set and the validation set is relatively small, which means that the proposed DeepCED-Net has almost no overfitting, in other words, the DeepCED-Net has strong learning ability. CNN and FCN are two classic neural networks, and have been widely applied in ECG signal analysis [27], thus, they This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/ This article has been accepted for publication in a future issue of this journal, but has not been fully edited.…”

“…Recently, deep learning techniques have been gaining attention due to their powerful capability in learning the characteristics of ECG signal [23]- [27]. The noise suppression techniques based on denoising autoencoder (DAE) have shown the excellent performance than conventional denoising methods.…”

Applications of Machine Learning

“…Wavedet [48] and WQRS [85] are examples of ECG frequency filters. Other examples of frequency filters include [49,50,51,52,53,54,55,86,87,88].…”

R-R Interval Estimation for Wearable Electrocardiogram Based on Single Complex Wavelet Filtering and Morphology-Based Peak Selection

“…However, this magnitude thresholding is not necessarily effective when it comes to selecting applicable QRS complexes: fixed thresholding may cause overlook or overdetection depending on the setting value, whereas variable thresholding may sometimes fail to follow the change and result in overlook due to the changed threshold [22]. Although the recently proposed time-attention method [23] predicts the probabilities of the regions of QRS complexes by using a convolutional neural network (CNN), this approach may not be free from the aforementioned thresholding issue because its final decision is based on a variable threshold, and it also requires a large amount of computation.…”