Automated detection of atrial fibrillation and atrial flutter in ECG signals based on convolutional and improved Elman neural network

Abstract

Atrial fibrillation (AF) and atrial flutter (AFL) are two common life-threatening arrhythmias. Both are not only easily transformed into each other, but also often cause misdiagnosis due to the similar clinical symptoms. The early efficient and accurate detection of AF and AFL is helpful to reduce the pain and injury of patients suffering from these diseases, and the traditional detection is often inefficient and laborious. Therefore, we propose a new model mechanism using an 11-layers network architecture to automatically classify AF and AFL signals. It is mainly constructed with the convolutional neural network (CNN) and the improved Elman neural network (IENN). Besides, we specifically design two relative models as control subjects to validate the classification performance of the proposed model. 10-fold cross-validation is also implemented on the MIT-BIH AF database (AFDB) and the MIT-BIH arrhythmia database (MITDB), respectively. The obtained results show that the model achieved the accuracy, specificity, and sensitivity of 98.8%, 98.6%, and 98.9% on the AFDB database and 99.4%, 99.1%, and 99.6% on the MITDB database, respectively. The model mechanism has been demonstrated to have more superior performance than two relative models and some advanced algorithms, which can be considered as a reliable and efficient identification system to aid physicians and save lives.

Introduction

According to the World Health Organization (WHO), arrhythmia is a common life-threatening heart disease, affecting more than 12 million people in the United States and over 350 million people worldwide [1], [2]. Among the diverse arrhythmia conditions, atrial fibrillation (AF) and atrial flutter (AFL) have higher morbidity and mortality. AF is an irregular abnormal rhythm along with the absence of P-wave and AFL is a relatively regular atrial rhythm due to the macrorentrant circuit phenomenon [3]. Notably, both are not only easily transformed into each other, but also often easily confused by cardiologists due to their similar clinical symptoms [2]. This undoubtedly increases the difficulty of diagnosis and even causes irreparable medical accidents if misdiagnosis occurs. Therefore, it is indispensable to combat this situation for the sake of public health.

However, the traditional AF and AFL detection are mainly dependent on visual observation of electrocardiogram (ECG) data by professional doctors, their personal experience often determines the final diagnosis. Thus, this visual analysis is relatively laborious and subjective. Also, a great quantity of ECG data is putting a heavy burden on physicians [2], [4]. These limitations have motivated efforts to develop a computer-automated diagnosis system to discriminate AF and AFL signals from mass ECG data more efficiently and accurately.

The ECG is a primary non-invasive diagnostic tool that has been widely used in clinical practice due to its low-cost and simplicity. It is installed on the human body by electrodes for a period, which provides a large amount of key information about the electrical activity of the heart [1], [5]. At the same time, the time-domain, frequency-domain and other features of ECG signals can be well revealed [6], [7]. In recent years, many algorithms based on the ECG features have seen great success in boosting the performance of ECG signals analysis.

Currently, the majority of them mainly depend on machine learning algorithms [8], [9], [10]. The algorithm structure mainly includes feature extraction, feature selection, and classification. For instance, Henzel et al. [11] combined four statistical characteristics extracted from RR intervals (adjacent heartbeat intervals) with the generalized linear classifier for the discrimination of AF signals. Kumar et al. [12] performed the decision mechanism using the ECG features from log-energy entropy (LEE) and the random forest (RF) classifier to distinguish AF segments. Safarbali et al. [13] presented a nonlinear method based on fractal dimension and persistent homology of heart rate variation (HRV) phase space. A 3-layers artificial neural network (ANN) was performed to discriminate normal sinus rhythm (NSR) and AF episodes. Kong et al. [14] proposed the decision approach using the integrated radial basis function (IRBF) and relevance vector machine (RVM). The related features of RR intervals were fed into the proposed classifier to distinguish ECG episodes. Kalidas et al. [15] developed an automated technique based on discrete-state Markov models (DMM) and RF for the detection of AF episodes. These algorithms are helpful to identify ECG data, but they still have some disadvantages. Firstly, how to formulate appropriate feature extraction and feature selection strategies remains a challenge. At present, however, there are few foolproof means to confirm which manual design theory can match the given problem; secondly, the traditional hand-designed method has limited ability to extract non-linear features. Especially when faced with different types of data sets, the robustness of the method cannot be well ensured, and thus it also suffers from the bottleneck of low generalization ability.

Therefore, to tackle the above problems, deep learning (DL) technology has been widely developed and achieved huge success in many fields [16], [17], [18]. In contrast with traditional learning methods, the end-to-end learning paradigm of DL greatly alleviates the predicament caused by artificial design, which integrates feature extraction, feature selection, and classification into a model. The trained deep model is capable to mine various data sources while making the computer self-learn essential features for any given problems. The network architecture primarily includes stacked autoencoder (SAE), convolutional neural network (CNN), recurrent neural network (RNN), deep belief network (DBN) and so on [19]. At present. with the advancement of computer processing power and data-driven application, DL models are becoming more promising in several fields such as image processing and behavior classification, especially in ECG signals analysis. Recently, Acharya et al. [20] constructed a deep network model to discriminate four kinds of heartbeats. Faust et al. [21] presented the novel network structure using RNN and the long short-term memory (LSTM) structure to perform AF segments classification. Fujita et al. [22] proposed an 8-layers CNN model for the detection of AF and AFL fragments. Andersen et al. [23] presented the identification system based on CNN and RNN. The high-level features of RR intervals were used as model input to classify ECG episodes. Besides, Yao et al. [24] designed a novel attention-based time-incremental CNN (ATI-CNN) structure. The model was integrated with CNN, recurrent cells and attention module to discriminate AF signals.

Since some previous studies based on RNN and its variations have achieved well classification results [21], [23], [24], the structure also has attracted increasing attention and become the focus of research nowadays. Meanwhile, Elman neural network (ENN) is standing out from many classical models of RNN [25], [26]. In its network structure, due to the existence of the context units, the output of the last moment of hidden layer could be well preserved. Moreover, it can be updated over time to ensure the persistence of the spatio-temporal information [27]. Therefore, ENN often has a well dynamic information processing ability than conventional static networks such as multi-layer perceptron (MLP), and thus it is widely applied in various scenarios nowadays [25], [28], [29].

So far, however, there are few studies to implement ECG signals analysis using the ENN structure. Further, the design of the ENN model can be modified by diverse improvement approaches. It is promising to integrate CNN and the improved ENN (CNN-IENN) to discriminate AF and AFL signals from mass ECG data. Thus, in our research, a new network mechanism using an 11-layers CNN-IENN model is constructed. Additionally, we specifically design two relative models as control subjects to better validate the classification performance of the proposed model. 10-fold cross-validation is also performed on the MIT-BIH AF database (AFDB) and the MIT-BIH arrhythmia database (MITDB), respectively.