Elsevier

Knowledge-Based Systems

Volume 258, 22 December 2022, 109926
Knowledge-Based Systems

A novel deep neural network for detection of Atrial Fibrillation using ECG signals

https://doi.org/10.1016/j.knosys.2022.109926Get rights and content

Abstract

In healthcare practice, one of the most predominantly occurring dysrhythmia is atrial fibrillation (AF). The anomalous heart rhythm and the deficiency of an evident P-wave signal are the consequences of AF, several cerebral apoplexies, thrombus, blood coagulation, cognitive impairments, and strokes. It is arduous to ascertain the symptoms of AF and clinically silent that might cause death. There are certain liabilities for diagnosis of AF in manual Electrocardiogram (ECG) since it demands high expertise; it is a time demanding and tedious process which is also accompanied by variations between intra- and inter-observer. Hence, to combat with this issue, a novel AF detection models has been proposed a conglomerate parallel structure of Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) framework which can deepen the understanding of features and classification. To test and legitimize the system, we utilize the data of the MIT-BIH Atrial Fibrillation Database. Parallel models for control subjects have been designed specifically to validate performance in terms of classification for more voracious categorization of 3 classes namely: Non-Atrial Fibrillation (N-AF), Atrial Fibrillation (AF) and Normal Sinus Rhythm (NSR). The model obtained an Accuracy, Sensitivity and Specificity of 99.6%, 98.64% and 99.01% respectively.

Introduction

The World Health Organization (WHO), states that over 6 million people were imperiled and around 90 million people worldwide have encountered AF in people aged 50 to 59 in the United States [1]. The mortality has outstretched to 5% and around 10% in senior citizens aged 80 to 89. This study has ameliorated prominently as age. Computer-Aided Design (CAD) system is used to monitor the human heart for longer durations. A device that displays the human heart condition and is extensively utilized in cardiac medical inspection is the Electrocardiogram [2]. One may record ECG using either non-invasive or invasive techniques. ECGs include a sequence of waveforms that replicate at regular intervals of P-waves, Q-waves, and R-waves (comprising the QRS complexes). In which, the T-waves and U-waves are not detectable. The essential physiological attributes that are extensively used to analyze the function of heart are denoted by the shifting of electrical activity in the heart over time using an ECG [3]. In the ECG complex, atrial depolarization is delineated by the P-wave, ventricular depolarization by QRS complex and ventricular repolarization by T-wave. Certain sections of the ECG signal are most significant for the diagnosis of diverse cardiac illnesses, particularly arrhythmia like RR interval, QT interval, PR interval, ST segment, PR segment, and QRS complex are shown in Fig. 1.

Arrhythmias might occur due to various physiological and pathological factors that affect the heartbeat leading to a sinus rhythm that is different from a normal sinus rhythm. The most frequently caused form of arrhythmia is AF, which causes non-uniform, anomalous rhythm accompanied by the lack of P-wave [3]. Prevention of complexities related to cardiogenic embolic illness, such as stroke, can be improved remarkably by an advanced and precise AF diagnosis. Almost 25% of all beings encountered with AF are asymptomatic [4]. Loss of sinus P-wave, serious non-uniformity of the QRS complex on the electrocardiogram and anomalous contraction of the upper atrium are the physiological symptoms of AF [5].

ECGs are the rudimentary and authentic diagnostic test of heart arrhythmias, so patients with miscellaneous and simultaneous arrhythmias are often diagnosed by ECG in clinical centers. Inappropriately interpreted ECGs would lead to inaccurate decisions made by the physicians, which results in unfavorable consequences [6]. In order to mitigate the inappropriate exposition, automated AF perception has been used. With the assistance of automated computer-aided design (CADx) machines, preliminary identification of heart-related illness can be made feasible to a certain extent which helps in decreasing the high mortality rate prevalent among patients encountering AF. CAD systems accentuate the diseased areas using image processing algorithms could be used to indicate the anomalies in the heart rhythm automatically [7].

The energy of signal spectra and versatility of time-frequency analysis of ECG can be improved by employing the Fractional Stockwell Transform (FrST) [8], which can also effectively ameliorate the time-frequency resolution capacity in comparison to Stockwell Transform (ST). Studies have shown that structural characteristics of time-frequency analysis of ECG signals are challenging to identify since they vary from patient to patient. Hence, Feature analysis is used for the extraction of information to quantify the structural properties of ECG signals. The AF signals can be effortlessly categorized from the statistical attributes (MVAR features) since AR coefficients are the most uncomplicated feature for AF classification [9]. The studies shown that the utilization of MVAR coefficient features produces preferable outcomes than that of original time series.

Parallel deep learning model is employed to classify the features. The Deep Learning (DL) significantly attenuates the difficulty originated by artificial design, which consolidates certain attributes like extraction, selection, and classification of features into a model. The extraordinary execution manifests that the posited model is an efficacious system to aid doctors and has good prospects of ameliorating early detection and mitigate mortality. However, DL methods have displayed masterly execution even with improper proportionality of data [10]. Convolutional Neural Networks (CNN) have been extensively used in the past for arrhythmia diagnosis and ECG signal classification. The prime disadvantage of ECG signal classification is a wide contrast in the quantity of information acquired between regular and irregular signals. CNN is a DL based algorithms which were the outcomes inspired by how the human brain and the biological neural systems work. Yet another kind of network structure especially created to incorporate sequential data or information is the Recurrent Neural Networks (RNN), which are frequently utilized in the classification of sequences. In contrast to a regular neural network, the RNN has the ability to learn temporal dependencies, due to which it has displayed precedence while contemplating time series data like the ECG signals [11], [12]. The combination of LSTM and RNN with 10-fold cross-evaluation to indicate rhythm of AF has attained an improved accuracy [13]. An automatic recognition technique called CNN-LSTM was put forward to spontaneously diagnose the AF heartbeats centered on deep learning [14]. In this work, we proposed an automatic AF detection method based on 2D CNN and LSTM architecture using ECG signals.

Section snippets

Overview

The workflow of proposed network architecture has been depicted in Fig. 2. An ECG is sectioned by a particular time to attain a good set of data segments whose length is T. Subsequently, every segment of data is sent across a band pass filter to eliminate noise. Later, Fractional Stockwell Transform is employed for each data segment to get spectra and versatility of ECG signals and to be administered to a CNN and RNN with Multivariate Auto-Regressive Algorithm. The FrsT scalogram image has a

Experimental Results

The proposed methodology was interpreted with the help of stratified 10-fold cross validation, where the sample set had been partitioned into 10 equal groups which have identical volume of information with a definite value. Of which, 9-fold were training data which is trained and evaluated on the remaining part. The training procedures were terminated in 100 spans. In addition to it, for each training event best weights were revived. To determine the performance of the model, we evaluated in

Conclusion

The model of 2D-CNN & RNN has been proposed for the Automation of Atrial Fibrillation detection on ECG Signals. Raw ECG signal need to be enhanced and its artifacts are removed using Fractional stockwell transform. Multivariate autoregressive model is adopted to extract features from denoised ECG signals. In this study, the MIT-BIH Atrial fibrillation database was employed to legitimize the structure, while yielding high performance extraction and classification. The performance of this model

CRediT authorship contribution statement

Lokesh Subramanyan: Conceptualization, Methodology, Writing – original draft, Software, Writing – review & editing. Udhayakumar Ganesan: Data curation, Investigation, Supervision, Writing – review & editing, Visualization.

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.

Acknowledgments

None. No funding to declare.

References (26)

  • ZhuH. et al.

    Automatic multilabel electrocardiogram diagnosis of heart rhythm or conduction abnormalities with deep learning: a cohort study

    The Lancet Digital Health

    (2020)
  • LiuJ. et al.

    A multiscale autoregressive model-based electrocardiogram identification method

    IEEE Access

    (2018)
  • ChenY.-H. et al.

    Atrial fibrillation risk prediction from electrocardiogram and related health data with deep neural network

  • Cited by (8)

    View all citing articles on Scopus
    View full text