D2AFNet: A dual-domain attention cascade network for accurate and interpretable atrial fibrillation detection

Highlights

• The D2AFNet can exploit the channel-spatial and time series features to mine discriminative atrial fibrillation patterns.

• The D2AFNet can profoundly explore the different contributions of spatial and temporal segments for excellent interpretation.

• The D2AFNet method is tenfold cross-validated on the CPSC 2018 dataset and independently tested on the MIT-BIH dataset.

• The D2AFNet method achieves the accuracies of 99.49% and 99.28% in the two-class and three-class AF detection tasks.

Abstract

Atrial fibrillation is one of the common and potentially dangerous persistent cardiac arrhythmias that are generally associated with the risk of stroke and heart failure. Manual electrocardiography diagnosis is the gold standard for the clinical detection of atrial fibrillation, but it has some drawbacks, such as being time-consuming and prone to misclassification due to inter-patient variability. Due to the powerful ability of deep learning to learn and extract rich features from huge datasets, end-to-end deep learning models are generally designed to detect abnormal atrial fibrillation signals automatically. However, these approaches usually ignore the key factors that feature maps from different channels and sequences may contribute differently to atrial fibrillation detection, making it challenging to implement accurate and interpretable models with better generalization performance. To tackle this challenge, we develop a dual-domain attention cascade D2AFNet for accurate and interpretable atrial fibrillation detection by cascading attention-based bidirectional gated recurrent units and densely connected networks embedded with channel-spatial information fusion modules. The D2AFNet can take full advantage of channel-spatial features to enhance the feature representation in the spatial domain, and then combine with the time series features in the temporal domain to form spatial–temporal fusion attention mechanisms to mine discriminative atrial fibrillation patterns. Besides, the D2AFNet can profoundly explore the different contributions of different spatial and temporal segments of feature maps for excellent interpretation. The proposed D2AFNet method is performed ten-fold cross-validation on the publicly available CPSC 2018 dataset, and achieves the accuracies of 99.49% and 99.28% in the two-class and three-class classification tasks, outperforming cutting-edge atrial fibrillation detection methods. In addition, the powerful generalization performance and inference efficiency of the D2AFNet method are also proved on another publicly available MIT-BIH dataset. The advantages of high performance and interpretability indicate that the D2AFNet method has huge potential in the computer-aided diagnosis of atrial fibrillation.

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.