D2AFNet: A dual-domain attention cascade network for accurate and interpretable atrial fibrillation detection
Introduction
Atrial fibrillation (AF) is the most frequent and potentially harmful persistent heart arrhythmia, characterized by the rapid, irregular, and uneven electrical activity of the heart caused by ineffective atrial contractions [1], [2], which can easily cause malignant diseases such as stroke, heart failure, and thromboembolism. At present, the prevalence of AF in China is 0.71 %, lower than the 1–4 % in western countries. However, with the increasing aging of the population worldwide, the prevalence of AF is showing a rapid year-on-year growth trend, and it is expected to reach 2–2.5 times by 2050 [3]. Therefore, developing appropriate AF detection devices or methods has recently received much attention.
Electrocardiogram (ECG) is the primary detection tool for AF, which determines the presence of AF by observing the changes in ECG waveform patterns and frequencies [4]. However, during manual diagnosis, some paroxysmal AF requires analyzing every heartbeat in each ECG, which is time-consuming and has the potential for misclassification, limiting the efficiency and accuracy of AF diagnosis. This limitation can be largely eliminated by using computational techniques for automated AF detection. Furthermore, due to the tremendous pressure of rapid data growth, more and more computational methods have begun to focus on intelligent computer-aided diagnosis strategies or algorithms to achieve automated detection of AF [5].
Over the past few decades, various computer-aided intelligent detection methods have been developed, including support vector machines (SVMs) [6], [7], [8], [9], [10], [11], k-nearest neighbor (KNN) [12], random forest (RF) [13], [14], [15], artificial neural networks (ANNs) [16], [17], [18], [19], assemble classifier [20], [21], etc. However, most of these approaches are machine learning strategies, and their algorithmic frameworks usually involve several steps, such as preprocessing, feature extraction, and classification. Moreover, the designs of these algorithms rely heavily on extensive and elaborate manual operations. Also, there is a risk of losing important information during the manual feature extraction stage, leading to misdiagnosis. Therefore, more advanced intelligent detection algorithms are urgently needed to achieve accurate and automated detection of AF.
Recently, deep learning (DL) technology, as a more effective and promising computer-aided intelligent detection algorithm, has been proposed to perform feature extraction and classification due to its unique self-learning and adaptive capabilities, and has been widely used for automated detection of AF [22], [23], [24], [25], [26], [27], [28], [29]. The commonly used deep learning models for AF detection are convolutional neural networks (CNNs). For a typical CNN structure, it sequentially places multiple convolutional and pooling layers to form a deep neural network capable of extracting the underlying features of input while reducing its dimensionality to achieve end-to-end AF detection studies. Besides CNNs, recurrent neural networks (RNNs) [30], [31], [32], [33], [34] are another typical DL method for time series analysis. The most representative one is long short-term memory (LSTM), which has the advantage of learning the temporal dynamic input data and selectively remembering or forgetting to check the information of the current memory state, and is becoming increasingly popular in AF classification.
Additionally, various combined models integrating CNN and RNN have gradually emerged and have been proven more effective than CNN or RNN alone in AF detection [35], [36], [37], [38], [39], [40]. The main idea behind these works is to use the CNN part as a feature extractor to provide the most discriminative input features for the RNN part. Although these methods can take full advantage of CNN and RNN to obtain better classification results, they ignore the key factors that feature maps extracted from different channels and sequences contribute differently to AF signal detection, resulting in low detection accuracy and lack of interpretability. Furthermore, simply combining CNN and RNN without considering the particularity of AF features in spatial and temporal dimensions may also produce less robust classification results, leading to poor generalization ability.
Therefore, in this study, we design a dual-domain attention cascade AF detection network, named D2AFNet, to achieve accurate AF detection with excellent interpretability. The proposed D2AFNet can extract the discriminative spatio-temporal fusion feature of AF by cascading attention-based bidirectional gated recurrent units and densely connected networks embedded with channel-spatial information fusion modules to achieve high detection accuracy. In addition, to improve the generalization capability of the proposed model, we provide an in-depth analysis of the operating principle of the model and obtain excellent interpretability by mining the relationship between the detection results and the input sequences. The contributions of our D2AFNet method are as follows:
- •
We propose a novel dual-domain attention mechanism whose weight parameters are calculated from channel-spatial features to enhance the feature representation in the spatial domain, which are then combined with the time-series features in the temporal domain to form dual-domain spatial–temporal fusion attention to mine discriminative AF patterns.
- •
Our proposed D2AFNet method can profoundly explore the specific contributions of different spatial and temporal segments of feature maps to obtain weighted parametric features for accurate AF detection with excellent interpretation.
- •
The proposed D2AFNet achieves the best classification performance in two-class and three-class AF detection tasks and outperforms the state-of-the-art methods.
The remainder of this study is arranged as follows. Section 2 describes the related works. Section 3 outlines the methodology of this study. Section 4 presents the experiments and results. Section 5 discusses the proposed D2AFNet method and compares the state-of-the-art methods. Section 6 concludes this study.
Section snippets
Related works
This section provides an overview of recent learning-based methods for AF classification. These studies mainly include traditional machine learning algorithms and deep learning algorithms. Deep learning methods mainly include CNN-based models, RNN-based models, and hybrid deep learning models combining CNNs and RNNs.
Problem formulation
The detection of AF could be viewed as a sequence classification task. This task requires designing a deep learning network model to capture helpful information from single-lead ECG recordings and determine the correct classification. The proposed model takes a dataset x={(x(1), y(1)), (x(2), y(2)), … ,(x(n), y(n))} consisting of input ECG recording x(i) and corresponding label y(i), where x(i) ∈ Rn and y(n) ∈ [1], [2]. The proposed D2AFNet takes x(n) as input and z(n) as output, given by (1):
Experimental environment
The proposed network was implemented using PyTorch on the Ubuntu 20.04 operating system powered by a 6-core Intel Core i7-10700K processor. The training and testing procedures were conducted on an NVIDIA GeForce GTX 3090 with 24 GB RAM.
Dataset
The China Physiological Signal Challenge 2018 (i.e., CPSC 2018) [42] was utilized to perform tenfold cross-validation experiments, and the specific details of the dataset are described in Table 1. Our goal was to distinguish between AF, normal, and other rhythms
Overview of the proposed method
In this work, we propose a powerful dual-domain attention cascade D2AFNet model that combines attention-based bidirectional gated recurrent units and densely connected networks embedded with channel-spatial information fusion modules to achieve accurate and interpretable AF detection. Firstly, the proposed D2AFNet can significantly reduce the reliance on manual feature extraction and accurately identify the salient features of different abnormal AF rhythms without computing feature parameters
Conclusion
This paper proposes a dual-domain attention cascade deep learning method for two-class and three-class AF detection tasks. The main finding of this work is that combining channel representation features in the spatial domain with time-series information in the temporal domain not only allows the D2AFNet to automatically extract discriminative AF features from ECGs in an interpretable way, but also indeed improves the overall classification performance. In addition, the proposed D2AFNet model
Funding
This work was supported by the National Natural Science Foundation of China (No. 62271023, 61871022), the Beijing Natural Science Foundation (No. 7202102), the 111 Project (No. B13003), the Fundamental Research Funds for Central Universities, and the Academic Excellence Foundation of BUAA for PhD Students.
CRediT authorship contribution statement
Peng Zhang: Conceptualization, Methodology, Investigation, Writing – original draft, Writing – review & editing, Funding acquisition. Chenbin Ma: Conceptualization, Methodology, Writing – review & editing. Fan Song: Methodology, Investigation. Yangyang Sun: Formal analysis, Visualization. Youdan Feng: Validation. Yufang He: Resources, Software. Tianyi Zhang: Data curation, Formal analysis. Guanglei Zhang: Supervision, Project administration, Writing – review & editing, Funding acquisition.
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (47)
- et al.
Atrial fibrillation in the ICU
Chest
(2018) - et al.
Explainable artificial intelligence to detect atrial fibrillation using electrocardiogram
Int. J. Cardiol.
(2021) - et al.
Automatic detection of atrial fibrillation using stationary wavelet transform and support vector machine
Comput. Biol. Med.
(2015) - et al.
Automated diagnosis of atrial fibrillation ECG signals using entropy features extracted from flexible analytic wavelet transform, Biocybernetics and Biomedical
Engineering
(2018) - et al.
Atrial fibrillation classification with artificial neural networks
Pattern Recogn.
(2007) - et al.
Automated detection of atrial fibrillation using R-R intervals and multivariate-based classification
J. Electrocardiol.
(2016) - et al.
Detecting atrial fibrillation by deep convolutional neural networks
Comput. Biol. Med.
(2018) - et al.
Accurate detection of atrial fibrillation from 12-lead ECG using deep neural network
Comput. Biol. Med.
(2020) - et al.
Global hybrid multi-scale convolutional network for accurate and robust detection of atrial fibrillation using single-lead ECG recordings
Comput. Biol. Med.
(2021) - et al.
An incremental learning system for atrial fibrillation detection based on transfer learning and active learning
Comput. Methods Programs Biomed.
(2020)
An intelligent computer-aided approach for atrial fibrillation and atrial flutter signals classification using modified bidirectional LSTM network
Inf. Sci.
Multi-domain modeling of atrial fibrillation detection with twin attentional convolutional long short-term memory neural networks
Knowl.-Based Syst.
Automated detection of atrial fibrillation and atrial flutter in ECG signals based on convolutional and improved Elman neural network
Knowl.-Based Syst.
Automated atrial fibrillation detection using a hybrid CNN-LSTM network on imbalanced ECG datasets
Biomed. Signal Process. Control
A novel multi-scale convolutional network with attention-based bidirectional gated recurrent unit for atrial fibrillation discrimination, Biocybernetics and Biomedical
Engineering
Research progress of LncRNAs in atrial fibrillation
Mol. Biotechnol.
Atrial fibrillation prevalence, awareness and management in a nationwide survey of adults in China
Heart
Electrocardiogram (ECG) for the prediction of incident atrial fibrillation: an overview
J. Atr. Fibrillation
A Support Vector Machine approach for reliable detection of atrial fibrillation events
Comput. Cardiol.
Detection of atrial fibrillation episodes in long-term heart rhythm signals using a support vector machine
Sensors
A support vector machine approach for AF classification from a short single-lead ECG recording
Physiol. Meas.
Cited by (6)
Multi-scale attention convolutional neural network for noncontact atrial fibrillation detection using BCG
2024, Biomedical Signal Processing and ControlA generalizable and robust deep learning method for atrial fibrillation detection from long-term electrocardiogram
2024, Biomedical Signal Processing and ControlAtrial Fibrillation Detection Method Based on Inception-CNN-BiLSTM Model
2023, Proceedings - 2023 China Automation Congress, CAC 2023
- 1
These authors contributed equally to this work.