An intelligent computer-aided approach for atrial fibrillation and atrial flutter signals classification using modified bidirectional LSTM network
Introduction
Arrhythmia is a serious and common pathophysiological mechanism that arises from abnormal electrical impulses [1]. In all types of arrhythmias, atrial fibrillation (AF) and atrial flutter (AFL) have a great effect on the health of patients due to their severe incidence rates and mortalities. Both arrhythmias are not easily perceived by human body and also characterized by similar clinical characteristics [2], [3], [5]. These problems have gradually become the main factors that make cardiologists easily confused or even misdiagnosed [3], [4]. Thus, establishing an early and accurate disease detection strategy to cure the patients is of considerable significance.
As a traditional AF and AFL detection method that mainly relies on visual inspection of electrocardiogram (ECG) signals, ECG is an effective and common approach to reveal the significant physiological information concerning the state of human heart [5], [6]. However, due to several emerging issues such as laborious visual inspection, inevitable human error and explosive ECG data growth, there is an urgent requirement to investigate the intelligent computer-aided diagnosis approach using ECG for AF and AFL signals identification with high efficiency and accuracy.
Nowadays, diverse algorithms have witnessed tremendous advances in the field of ECG signals identification and most of them are conventional machine learning algorithms. The robustness of these algorithms, despite some success with this kind of method, still needs to be strengthened especially with various data sources [5], [13]. Besides, it is indispensable to design the corresponding feature extraction and classification procedures manually. So far, however, there is no definite theoretical system to formulate the critical procedures accurately [14]. These issues are increasingly prominent and thus hampering the development of conventional strategies in several application scenarios.
Unlike traditional machine learning algorithms, deep learning models have exhibited superior performance in recent years [3], [15], [16]. They are designed to integrate corresponding algorithm steps into a model and have the capability to automatically learn the underlying data features. As several effective deep model frameworks, convolutional neural network (CNN) and recurrent neural network (RNN) that mainly includes long short-term memory (LSTM) and bidirectional LSTM (B-LSTM) networks have witnessed great success in ECG signals analysis [17], [18], [19], [20], [21], [22].
Taking these published researches into account, it is worth noting that RNN and its variations have become a promising recipe for obtaining state-of-the-art performances over the years [20], [21], [22]. In particular, LSTM is proposed to alleviate the vanishing gradient problem and difficulty in revealing long sequence dependence that occurs in standard RNN [20], [23]. However, due to the limitation of existing LSTM network that it only models the sequence along with one direction, B-LSTM is designed to simultaneously model the sequence along both forward and backward directions. Note that B-LSTM often exhibits better performance than LSTM in several scenarios, but several studies have shown that it suffers from the problem of information redundancy. For example, Schuster et al.[35] exhibit that B-LSTM works well in some tasks, but it is still blurry to effectively integrate feature representations of forward and backward directions, and this easily brings information redundancy to the network. Pascanu et al.[36] demonstrate that B-LSTM usually faces information redundancy and even the gradient exploding problem so that it is disadvantageous to extend the model to lightweight devices such as mobile phone. Besides, Zhang et al.[37] point out these issues by systematically exploring the connection architecture of several RNN’s variants.
In this work, to alleviate the above issues, we introduce a modified B-LSTM (MB-LSTM) network, a novel architecture module specially designed for AF and AFL detection. Inspired by Squeeze-and-Excitation network (SENet) [24], this work aims to redesign existing B-LSTM network by implementing a feature recalibration strategy. Specifically, squeeze operation and excitation operation are implemented on forward and backward directions in B-LSTM while generating the normalized importance of feature representations of both directions. Further, by implementing a scale operation on the obtained importance, the improved network is able to adaptively emphasize useful feature representations and suppress feature representations. Further, the proposed model will be embedded into a CNN architecture with attention mechanism and existing LSTM, B-LSTM networks are used as control groups to evaluate the effectiveness of MB-LSTM using a subject-independent validation strategy on the two public databases.
In summary, the innovations of this work can be summarized as follows:
- •
We put forward MB-LSTM network based on a feature recalibration strategy for AF and AFL signals classification while achieving higher model performance than several state-of-the-art methods on both public databases. The proposed model can be regarded as an improved version of existing LSTM and B-LSTM networks.
- •
To our knowledge, there is currently little research to alleviate the problem of information redundancy in B-LSTM. This work is the first attempt to redesign B-LSTM to compensate for this deficiency.
- •
In particular, a subject-independent validation strategy is used to ensure model robustness and detailed qualitative analysis is also provided for the mechanism of performance improvement to improve interpretability.
Section snippets
Related work
Extensive recent researches for ECG signals classification are mainly divided into two major categories: conventional machine learning based methods and deep learning based methods.
- •
: The traditional algorithm mainly consists of two main procedures: feature extraction and classification. Feature extraction is the core of algorithm system and designed by several hand-crafted strategies such as wavelet analysis [7], morphology [8], [46], radial basis function [9]
Database
The MIT-BIH AF database (AFDB) and arrhythmia database (MITDB) are used to assess the effectiveness of the proposed network, which are extensively used for several arrhythmia studies. The ECG records of both databases were labeled by professional physicians and the detailed information for these records can be summarized as follows:
- •
This database is primarily composed of 23 two-channel records from 25 subjects, every record is obtained by 250 Hz sample rate with a duration of 10 h [25].
- •
Results
In this work, the proposed models are evaluated using a subject-independent validation strategy with three standard evaluation measures of accuracy (ACC), sensitivity (SEN), and specificity (SPF) on two public databases. For a fair comparison, we employ existing LSTM and B-LSTM networks to devise C-LSTM and CB-LSTM as control groups, respectively. Both models are stacked by the same feature extraction and attention mechanism blocks as CMB-LSTM. The only difference among the three models is that
Qualitative analysis
To further elaborate on the effectiveness of the proposed MB-LSTM model, we provide the detailed theoretical analysis in the section. It is known that in an arbitrary network, the optimizer is aimed at the parameters of the network, so as to minimize the loss functionwhere is the sample number of training dataset. We use the mini-batch of size m to approximate the gradient of loss function with regard to each parameters as follows:
Besides, we
Conclusion
In this work, we introduce MB-LSTM network, an improved version of existing LSTM and B-LSTM networks for AF and AFL signals classification. With a subject-independent validation strategy, two control groups are constructed to evaluate the availability of MB-LSTM. The results exhibit more superior classification performance than several state-of-the-art methods across two public databases. The major contribution of the research is introducing a feature recalibration strategy based on B-LSTM
CRediT authorship contribution statement
Jibin Wang: Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Software, Supervision, Validation, Writing - original draft, Writing - review & editing.
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.
Acknowledgement
The author would like to express his sincere gratitude to the anonymous referees, the Editor and author’s parents for many valuable suggestions, comments and supports that helped to improve the paper.
References (52)
- et al.
Computer-aided diagnosis of atrial fibrillation based on ECG signals: a review
Inform. Sci.
(2018) Automated detection of atrial fibrillation and atrial flutter in ECG signals based on convolutional and improved Elman neural network
Knowl-Based. Syst.
(2020)- et al.
Automated detection of arrhythmias using different intervals of tachycardia ECG segments with convolutional neural network
Inform. Sci.
(2017) - et al.
Automated identification of shockable and non-shockable life-threatening ventricular arrhythmias using convolutional neural network
Future Gener. Comput. Syst.
(2018) - et al.
Application of deep convolutional neural network for automated detection of myocardial infarction using ECG signals
Inform. Sci.
(2017) - et al.
A comparative study of DWT, CWT and DCT transformations in ECG arrhythmias classification
Expert Syst. Appl.
(2010) - et al.
Automated detection of atrial fibrillation in ECG signals based on wavelet packet transform and correlation function of random process
Biomed. Signal Process.
(2020) - et al.
Detection of atrial fibrillation using discrete-state Markov models and random forests
Comput. Biol. Med.
(2019) Computer aided diagnosis for atrial fibrillation based on new artificial adaptive systems
Comput. Meth. Prog. Bio.
(2020)- et al.
Application of higher order statistics for atrial arrhythmia classification
Biomed. Signal Process.
(2013)
Deep learning approach for active classification of electrocardiogram signals
Inform. Sci.
Automated detection of atrial fibrillation using long short-term memory network with RR interval signals
Comput. Biol. Med.
A deep learning approach for real-time detection of atrial fibrillation
Expert Syst. Appl.
Automated arrhythmia classification based on a combination network of CNN and LSTM
Biomed. Signal Proces.
Multi-class arrhythmia detection from 12-lead varied-length ECG using attention-based time-incremental convolutional neural network
Inform. Fusion
Optimal selection of wavelet basis function applied to ECG signal denoising
Digital Signal Process.
Attention augmentation with multi-residual in bidirectional LSTM
Neurocomputing
Multi-domain modeling of atrial fibrillation detection with twin attentional convolutional long short-term memory neural networks
Knowl-Based. Syst.
Detecting atrial fibrillation by deep convolutional neural networks
Comput. Biol. Med.
Computer aided detection for fibrillations and flutters using deep convolutional neural network
Inform. Sci.
Automated diagnosis of atrial fibrillation ECG signals using entropy features extracted from flexible analytic wavelet transform
Biocybern. Biomed. Eng.
Atrial fibrillation detection using heart rate variability and atrial activity: a hybrid approach
Expert Syst. Appl.
Atrial fibrillation detection with and without atrial activity analysis using lead-I mobile ECG technology
Biomed. Signal Process.
Computer aided detection of breathing disorder from ballistocardiography signal using convolutional neural network
Inform. Sci.
A novel method for automated congestive heart failure and coronary artery disease recognition using THC-Net
Inform. Sci.
PlexNet: A fast and robust ECG biometric system for human recognition
Inform. Sci.
Cited by (45)
An atrial fibrillation classification method based on an outlier data filtering strategy and modified residual block of the feature pyramid network
2024, Biomedical Signal Processing and ControlEffective infant cry signal analysis and reasoning using IARO based leaky Bi-LSTM model
2024, Computer Speech and LanguageA generalizable and robust deep learning method for atrial fibrillation detection from long-term electrocardiogram
2024, Biomedical Signal Processing and ControlAn intelligent computer-aided diagnosis method for paroxysmal atrial fibrillation patients with nondiagnostic ECG signals
2024, Biomedical Signal Processing and ControlDiagnosis of atrial fibrillation based on lightweight detail-semantic network
2023, Biomedical Signal Processing and ControlDiagnosis of atrial fibrillation using self-complementary attentional convolutional neural network
2023, Computer Methods and Programs in Biomedicine