João Paulo do Vale Madeiro, ... Angelo Roncalli
Alencar Brayner, in Developments and Applications for ECG Signal Processing, 2019 The first-degree AV block is characterized by the difficulty of the atria impulses in reaching the ventricles, reflecting in the ECG a sustained delay prolongation between atrial and ventricular depolarization, a P–R interval higher than 0.2 s in all beats [Fig. 1.13]. It can be caused by structural defects of the conduction pathway as seen, for example, in chronic conduction degenerative diseases and myocardial
infarction [Lilly, 2012].Classical and Modern Features for Interpretation of ECG Signal
Atrioventricular Node
First-Degree AV
Block
Figure 1.13. An excerpt of ECG signal showing an episode of First-degree AV block.
Second-Degree AV Block
In the second-degree AV block, the atria impulses fail to reach the ventricles in all beats. It is due a conduction failure from the AV node, where atria impulses propagates intermittently. The ECG of this intrinsic behavior is characterized by a QRS complex not always preceded by a P wave [Lilly, 2012]. The second-degree AV block can have two different forms:
•Mobits Type I block is characterized by a gradual and progressive conduction defect between atria and ventricles until an impulse is completely blocked and ventricular stimulation lacks as a result of a single beat [Pastore et al., 2016; Lilly, 2012]. The ECG shows a progressive P–R interval increment and R–R interval reduction between beats, until the next QRS complex is absent, starting the cycle anew. During the pause, the R–R interval is twice shorter than the previous, see Fig. 1.14 [this figure was published in Textbook of Medical Physiology, Arthur C. Guyton and John E. Hall, Chapter 13: Cardiac Arrhythmias and Their Electrocardiographic Interpretation, Page 149, Copyright Elsevier Inc. [2006]] [Pastore et al., 2009];
Figure 1.14. Example of a second degree AV block, showing a dropped beat [failure of the ventricles to receive the excitatory signals] [lead V3].
•Mobits Type II block is characterized by a sudden conduction interruption, where the AV node ceases to conduct two or more beats, unexpectedly, without any ECG warning. As consequence, the ECG shows sequential P waves without a correspondent QRS complex [Fig. 1.15]. The His-Purkinje areas might play a role in this ceased behavior, resulting in abnormally wide QRS complexes and extensive infarction or chronic degenerative of the conduction pathway [Lilly, 2012].
Figure 1.15. An excerpt of ECG signal showing an episode of Mobits Type-II block: register 231 from MIT–BIH Arrhythmia Database.
Third-Degree AV Block
The third-degree AV block is characterized by a complete collapse of the impulses conduction between the atria and ventricles, dividing the heart into two unconnected zones, without P waves and QRS complexes sequential relationships [see Fig. 1.16; this figure was published in Textbook of Medical Physiology, Arthur C. Guyton and John E. Hall, Chapter 13: Cardiac Arrhythmias and Their Electrocardiographic Interpretation, Page 149, Copyright Elsevier Inc. [2006]]. In this scenario, the atria depolarize by the SA node at a higher rate [Pastore et al., 2009], and the ventricles by their own intrinsic escape rate, usually between 40 bpm to 55 bpm [Lilly, 2012].
Figure 1.16. Example of a complete AV block [lead II].
Read full chapter
URL: //www.sciencedirect.com/science/article/pii/B9780128140352000074
The Electrocardiogram—A Brief Background
Leif Sörnmo, Pablo Laguna, in Bioelectrical Signal Processing in Cardiac and Neurological Applications, 2005
6.2.3 ECG Waves and Time Intervals
We will now describe some important ECG wave characteristics, central to the development of signal processing algorithms, along with the wave-naming convention. Atrial depolarization is reflected by the P wave, and ventricular depolarization is reflected by the QRS complex, whereas the T wave reflects ventricular repolarization, see Figure 6.10. Atrial repolarization cannot usually be discerned from the ECG since it coincides with the much larger QRS complex. The amplitude of a wave is measured with reference to the ECG baseline level, commonly defined by the isoelectric line which immediately precedes the QRS complex. The duration of a wave is defined by the two time instants at which the wave either deviates significantly from the baseline or crosses it.
Figure 6.10. Wave definitions of the cardiac cycle and important wave durations and intervals. The J point defines the point in time when the QRS complex curves into the ST segment.
The P wave reflects the sequential depolarization of the right and left atria. In most leads, the P wave has positive polarity and a smooth, monophasic morphology. Its amplitude is normally less than 300 µv, and its duration is less than 120 ms. An absent P wave may, for example, suggest that the rhythm has its origin in the ventricles, i.e., a ventricular ectopic focus has taken precedence over the SA node causing atrial depolarization to coincide with ventricular depolarization.
The spectral characteristic of a normal P wave is usually considered to be low-frequency, below 10–15 Hz [Figure 6.11]. However, the application of ensemble averaging techniques to produce a noise-reduced ECG has helped demonstrate that much higher frequency components of the P wave exist; such components have been found useful for predicting the occurrence of certain arrhythmias of atrial origin.
Figure 6.11. Power spectrum of the P wave, QRS complex, and T wave. The diagram serves primarily as a rough guide to where the spectral components are located; large variations exist between beats of different lead, origin, and subjects.
It is sometimes problematic to determine the time instants that define the onset and end of a P wave because of a low amplitude and smooth morphology. As a result, the analysis of individual P waves is excluded from certain ECG applications where the presence of noise is considerable.
The QRS complex reflects depolarization of the right and left ventricles which in the normal heart lasts for about 70–110 ms. The first negative deflection of the QRS complex is denoted the Q wave, and the first positive is denoted the R wave, while the negative deflection subsequent to the R wave is denoted the S wave [Figure 6.10]. Although the QRS complex may be composed of less than three individual waves, it is nevertheless referred to as a QRS complex. The morphology of the QRS complex is highly variable and depends on the origin of the heartbeat: the QRS duration of an ectopic beat may extend up to 250 ms, and is sometimes composed of more than three waves.
Since the QRS complex has the largest amplitude of the ECG waveforms, sometimes reaching 2−3 mV, it is the waveform of the ECG which is first identified in any type of computer-based analysis. The algorithm that performs the search is termed the QRS detector and produces the “landmark information” required to further analyze the ECG characteristics, see Section 7.4.
Due to its steep slopes, the frequency content of the QRS complex is considerably higher than that of the other ECG waves and is mostly concentrated in the interval 10–50 Hz [Figure 6.11]. Similar to the P wave, ensemble averaging of the QRS complex has, in certain ECG recordings, uncovered high-frequency components which have been found to convey diagnostic information. In particular, the presence of late potentials in the terminal portion of the QRS complex has received considerable attention; see page 447 for further details.
The ST segment is not really a wave, but represents the interval during which the ventricles remain in an active, depolarized state. The ST segment begins at the end of the S wave [the J point] from where it proceeds nearly horizontally until it curves into the T wave [Figure 6.10]. Changes in the ST segment, which make it either more elevated, depressed, or more steeply sloped, often indicate various underlying cardiac conditions.
The T wave reflects ventricular repolarization and extends about 300 ms after the QRS complex. The position of the T wave is strongly dependent on heart rate, becoming narrower and closer to the QRS complex at rapid rates; this “contraction” property does not apply to the P wave or the QRS complex. The normal T wave has a smooth, rounded morphology which, in most leads, is associated with a single positive peak.
The T wave is sometimes followed by another slow wave [the U wave] whose origin is unclear but is probably ventricular after-repolarization. At rapid heart rates, the P wave merges with the T wave, causing the T wave end point to become fuzzy as well as the P wave onset. As a result, it becomes extremely difficult to determine the T wave end point because of the gradual transition from wave to baseline.
The RR interval represents the length of a ventricular cardiac cycle, measured between two successive R waves, and serves as an indicator of ventricular rate. The RR interval is the fundamental rhythm quantity in any type of ECG interpretation and is used to characterize different arrhythmias as well as to study heart rate variability.
The PQ interval is the time interval from the onset of atrial depolarization to the onset of ventricular depolarization. Accordingly, the PQ interval reflects the time required for the electrical impulse to propagate from the SA node to the ventricles. The length of the PQ interval is weakly dependent on heart rate.
The QT interval represents the time from the onset of ventricular depolarization to the completion of ventricular repolarization. This interval normally varies with heart rate and becomes shorter at more rapid rates. It is therefore customary to correct the QT interval for heart rate—using nonlinear [16] or, better, linear techniques [17]—so that the corrected QT interval allows an assessment that is roughly independent of heart rate. Prolongation of the QT interval has been observed in various cardiac disorders associated with increased risk of sudden death.
Read full chapter
URL: //www.sciencedirect.com/science/article/pii/B9780124375529500064
Overview of source separation applications
Y. Deville, ... R. Vigario, in Handbook of Blind Source Separation, 2010
16.7 Using physical mixture models or not
In addition to the dimension of the mixing model that we discussed in the previous section, a parameter of utmost importance when developing a BSS application is the nature [linear instantaneous, convolutive, nonlinear] of this model, because one should select accordingly the model used in the separating system. As already noted above, one should therefore derive a physical model of the relationship between the source signals and the observations, whenever this is possible. We have illustrated this approach in detail in this chapter, by means of two examples dealing with communications [see section 16.4.2.2] and astrophysics hyperspectral images [see section 16.5.1]. BSS investigations not based on such physical models have also been presented in the literature, however. Even for a single class of signals, different approaches may be used from that point of view, as shown below by considering two types of investigations that have been reported for electrocardiogram [ECG] analysis.
16.7.1 Mother vs fetus heartbeat separation from multi-channel ECG recordings
One of the most classical biomedical applications of BSS methods is related to ECG analysis. It concerns the extraction of fetal heartbeats from a set of ECG signals recorded by means of cutaneous leads placed on the mother’s skin. The signals thus obtained also include mother’s heartbeats and noise components. Moreover, physical considerations show that the recorded signals are linear instantaneous mixtures of fetal heartbeats, mother’s heartbeats and noise components [24].
This fetal heartbeat extraction problem was already tackled three decades ago by means of adaptive noise canceling techniques, which are now conventional [83]. The performance thus achievable was especially limited by the need for such techniques to use reference signals, i.e. unmixed sources. BSS algorithms now provide a more general framework which avoids this restriction and makes it possible to reconsider such applications while expecting higher performance. Results obtained in such applications are reported for example in [24,87,88]. A comparison of BSS methods and of more classical approaches based on Singular Value Decomposition [SVD] is also provided in [24].
A typical investigation of fetal ECG extraction, reported by De Lathauwer et al. [24], is illustrated in Figs 16.17 and 16.18. The measured signals are shown in Fig. 16.17 [such signals are available on the Internet: see section 16.8]. All recorded channels include an almost periodic component, corresponding to the mothers’s heartbeats. Each beat yields a major peak16 in these signals. Seven such peaks appear over the recorded period [i.e. 5 s]. In addition, the top three channels of Fig. 16.17 may be expected to contain a lower-magnitude, almost periodic, component. However, the associated waveform cannot be easily interpreted from these signals.
FIGURE 16.17. Eight-channel ECG recording [24] [courtesy of L. De Lathauwer].
FIGURE 16.18. Source estimates derived by BSS method from ECG recording [24] [courtesy of L. De Lathauwer].
Applying the linear instantaneous BSS method considered in [22] to these recordings yields the signals shown in Fig. 16.18 [two other methods led to similar results in [24]]. The mother’s 7-peak periodic signals clearly appear in the top three output channels [an interpretation of the dimension of the associated subspace is provided in [24]]. Moreover, another periodic signal is extracted, especially in the sixth channel. The considered recording period contains 12 peaks, i.e. 12 cycles of this signal. This periodic signal, which has a higher frequency than the mother’s heartbeats, corresponds to the fetal ECG.
16.7.2 Analysis of heart control from single-channel ECG
ECG signals have also been analyzed with other goals. In particular, Vetter et al. [81] presented an investigation where the BSS aspect of the problem appears in a much less natural way than above to non-specialists. This investigation concerns the analysis of the control of the heart by the autonomic nervous system, whose alterations have been shown to play an important role in many pathophysiological situations. This heart control system contains two antagonistic parts, corresponding to the cardiac sympathetic [CSNA] and parasympathetic [CPNA] nervous activities. Variations in these activities influence heart behavior and yield modifications in the ECG. The reported investigation aims at extracting the original CSNA and CPNA signals only from a single-channel observed ECG, or more precisely from two parameters derived from it. These parameters are the successive so-called RR and QT intervals, which respectively correspond to the time interval between adjacent heartbeats and to the duration of a specific portion of the ECG cycle.
This separation of CSNA and CPNA signals from their RR and QT mixtures is performed by means of a BSS method suited to linear instantaneous mixtures. Unlike in most applications, this linear instantaneous mixing structure is not selected as a result of detailed modeling of the considered physical system, which would show that the measured signals provide this type of mixture. Instead, the approach used here contains two aspects. Qualitative physiological knowledge corresponding to the above-mentioned heart control model is first used. It shows that each of the RR and QT parameters depends on the CSNA and CPNA signals, i.e. is a mixture [of unspecified type at this stage] of these signals [and possibly of others signals, which is confirmed below]. Then, a linear instantaneous mixture model is selected, based on two motivations. On the one hand, this investigation is focused on a “small-signal” approximation and only aims at extracting the most salient features of the variations of the CSNA and CPNA signals, which leads the authors to use linear modeling. On the other hand, they want to develop a simple analysis tool and they therefore only consider an instantaneous mixture model.
The other major issue of the BSS problem thus introduced concerns the independence of the sources to be restored [or at least their uncorrelation]. It has been shown that the CSNA and CPNA signals are not independent. However, previous work of the authors leads them to assume that two independent components, respectively sensitive to the CSNA and CPNA signals, may be derived from the observed ECG parameters.
The BSS method applied to the considered ECG parameters is a classical approach intended for temporally correlated sources, often referred to as SOBI [Second-Order Blind Identification] [7]. Moreover, it is preceded by a noise reduction stage based on PCA. This stage aims at reducing the influence of all the “noise” signals contained by the considered ECG parameters, in addition to the mixed contributions of CNSA and CPNA. These noise sources correspond to the influence of respiration and unknown stochastic phenomena on ECG parameters, together with measurement and quantization noise.
In biomedical applications, BSS methods often provide estimates of hidden variables, such as CSNA and CPNA here, which are not accessible in humans. These estimated variables cannot be compared to the lacking original sources and are therefore hardly interpreted. This makes it difficult to validate the operation of BSS methods in such applications. The authors here solve this problem by using specific experimental protocols. These protocols are selected because they elicit or inhibit sympathetic or parasympathetic responses and therefore make it possible to check if the proposed approach is able to highlight changes in the levels of CPNA and CNSA. This shows the effectiveness of linear instantaneous BSS methods in this application and the need to use a denoising stage. This approach outperforms the traditional indicator based on Fast Fourier Transform [FFT].17
It should be mentioned that the same authors previously reported a related approach [79,80]. However, the latter method requires non-causal convolutive BSS algorithms and simultaneous recordings of ECG and arterial blood pressure, which may be cumbersome in clinical applications. The above-described approach therefore has the advantage of requiring only instantaneous BSS methods and recording of one ECG channel.
16.7.3 Additional comments about performance evaluation
The application that we just described highlighted the problem of performance evaluation in applications where source signals are not accessible. A general solution to this problem however, exists in cases when, even if the source signals to be restored are not known, prior information about their properties is available. Indeed, one may indirectly verify that the considered BSS methods are successful, by checking to what extent the source estimates that they provide actually have the expected properties.
This approach is, for example, used in [16], still in the framework of cardiac signal analysis. This paper concerns the extraction of atrial activity during atrial fibrillation episodes. The authors take advantage of the fact that atrial activity has a narrowband spectrum, which contains a major peak at a frequency situated between 3 and 9 Hz, unlike the other signals involved in this application. The approach proposed for approximately evaluating the quality of the extraction of atrial activity by means of a BSS method then consists in measuring the spectral concentration of the estimated source. A BSS method is considered to yield good performance in this application if the spectrum of the signal that it provides is concentrated around its peak situated in the frequency band ranging from 3 to 9 Hz. On the contrary, this spectrum is more spread out if the restored signal contains undesired contributions resulting from other sources.
Read full chapter
URL: //www.sciencedirect.com/science/article/pii/B9780123747266000217
Biosignal monitoring using wearables: Observations and opportunities
Yashodhan Athavale, Sridhar Krishnan, in Biomedical Signal Processing and Control, 2017
2.3.1 About ECG Signals: properties and acquisition
Electrocardiography [ECG] involves recording of electrical impulses generated by heart muscles during its regular or irregular beating activity, by using electrodes placed over specific regions on the human torso [mostly around the chest region]. The intention is to capture minute heart beat signal changes which happens when the heart muscles depolarize during each beating cycle [33]. A typical ECG wave is characterized by three morphological patterns: a P-wave [atrial polarization wave], a QRS-complex wave [ventricular depolarization wave] and a T-wave [ventricular repolarization wave]. In terms of signal properties, a surface ECG signal has a frequency range of 0.05–150 Hz in diagnostic mode and 0.5–40 Hz in monitoring mode, with amplitude ranging from 0.1 to 5 mV [33].
Detecting abnormalities in heart rhythms usually involves analyzing for irregular patterns in either of the three aforementioned signal patterns. Most commonly occurring disorders include atrial and ventricular fibrillation, myocardial infarction and sudden cardiac death. ECG data when monitored with other body parameters such as blood pressure, glucose levels and pulse rate, could also serve as an indicator of diseases such as diabetes, high/low blood pressure and stress levels. In a clinical setting, ECG is typically monitored using a 12-electrode placement configuration [10] connected to a standard heart rhythm recording and monitoring system. This system allows for continuous signal acquisition, filtering and analysis which aids the doctor is making real-time decisions about a patient's cardiac health. But this system cannot be easily transferred to a home or remote setting due to its size and installation complexity.
Read full article
URL: //www.sciencedirect.com/science/article/pii/S1746809417300617
A survey on ECG analysis
Selcan Kaplan Berkaya, ... M. Bilginer Gulmezoglu, in Biomedical Signal Processing and Control, 2018
3.1 P-QRS-T complex features
The P-QRS-T complex features for an ECG signal basically correspond to the locations, durations, amplitudes, and shapes of particular waves or deflections inside the signal [106,107]. Typically, an ECG signal has a total of five major deflections, including P, Q, R, S, and T waves, plus a minor deflection, namely, the U wave, as shown in Fig. 2 [108]. The P wave is a small low-voltage deflection away from the baseline that is caused by the depolarization of the atria prior to atrial contraction as the activation [depolarization] wave-front propagates from the sinoatrial node through the atria. The Q wave is a downward deflection after the P wave. The R wave follows as an upward deflection, and the S wave is a downward deflection following the R wave. Q, R, and S waves together indicate a single event. Hence, they are usually considered to be the QRS complex. The features based on the QRS complex are among the most powerful features for ECG analysis. The QRS-complex is caused by currents that are generated when the ventricles depolarize prior to their contraction. Although atrial repolarization occurs before ventricular depolarization, the latter waveform [i.e., the QRS-complex] has a much greater amplitude, and atrial repolarization is, therefore, not seen on an ECG. The T wave, which follows the S wave, is ventricular repolarization, whereby the cardiac muscle is prepared for the next cycle of the ECG. Finally, the U wave is a small deflection that immediately follows the T wave. The U wave is usually in the same direction as the T wave.
Fig. 2. Standard fiducial points in the ECG [P, Q, R, S, T, and U] together with clinical features [listed in Table 2] [108].
Researchers utilize various attributes of the QRS complex as the features. Some of those attributes are the R wave duration, P+ amplitude, QRS p-p amplitude, R wave amplitude, ST amplitude, T+ amplitude, QRS wave area and ST slope. The R wave duration is the time that passes between the beginning and end of the R wave [109]. The P+ amplitude can be defined as the difference between the P point and the other subsequent points where the signal rises again. The QRS p-p amplitude is the difference between the R and Q points in the QRS complex in terms of the amplitude. The R wave amplitude can be defined as the height of the R wave from the baseline. The ST amplitude is the difference between the S and T points in terms of the amplitude values. The T+ amplitude can be defined as the difference between the T point and the subsequent point where the signal rises again. The QRS wave area is the area of the region when a rectangle is drawn on the QRS complex using the Q, R and S points [110]. The ST slope is the angle of the line, which can be drawn from the S point to the T point of the QRS complex. Some of the recent studies on ECG analysis that utilize QRS features are [35,40,67,70,76,80,88,89,91,106,107,111–117].
In addition, certain intervals within the ECG signal carry meaningful information and are employed as the features. For example, the PR interval is the duration between the beginning of the P wave and the QRS complex of an electrocardiogram. This duration contains signals between the onset of atrial depolarization and the onset of ventricular depolarization [118]. The QT-interval is the time between the onset of ventricular depolarization and the end of ventricular repolarization. The ST-interval is the time between the end of the S-wave and the beginning of the T-wave [119]. The RR interval [or heartbeat interval] is the time between the R peak of a heartbeat and the following heartbeat, which could be its predecessor or successor [9]. The RR-interval features are determined to realize the dynamic characteristics of the ECG signals [120]. Different RR interval features are used in [4,120,121].
Standard values for several QRS complex features of a normal ECG signal and healthy subjects with no cardiac abnormalities are listed in Table 2. The detailed information for these features is provided in [18] as well.
Table 2. Typical lead II ECG features and their normal values in the sinus rhythm at a heart rate of 60 bpm for a healthy male adult [18].
| 110 ms | ±20 ms |
| 160 ms | ±40 ms |
| 100 ms | ±20 ms |
| 400 ms | ±40 ms |
| 0.15 mV | ±0.05 mV |
| 1.5 mV | ±0.5 mV |
| 0 mV | ±0.1 mV |
| 0.3 mV | ±0.2 mV |
Since the extraction of the QRS features requires detection of the abovementioned fiducial points, various QRS complex detection algorithms are proposed in the literature. Although the most common one is the Pan-Tompkins algorithm [104], QRS detection algorithms can be categorized as derivative [122], digital filters [104,123,124], wavelet transform [125], neural networks [126] and phasor transform [127] based algorithms. In addition, high-order moments are used to detect QRS complexes in [111]. The difference operation method is used to find the fiducial points in [77]. There are also some studies that especially make use of R peak detection, such as [39,86,111,128–133]. In [105], a dynamic threshold based on a finite state machine is used to detect the R peaks. In [93], the techniques based on differentiation are used to detect the fiducial points. In [57,75,120], the Pan-Tompkins algorithm is used as well.
There also are several open-source QRS detectors that can be used by researchers. For example, EP-LIMITED, based on the Hamilton and Tompkins algorithm, and WQRS, based on a length transform, are used in [72]; the Augsburg Biosignal Toolbox is used in [81]; and ECGPUWAVE software is used in [134,135] to detect QRS and recognize the ECG wave boundary and extract the morphological features.
Read full article
URL: //www.sciencedirect.com/science/article/pii/S1746809418300636
Review of ECG detection and classification based on deep learning: Coherent taxonomy, motivation, open challenges and recommendations
Shan Wei Chen, ... Can Yang, in Biomedical Signal Processing and Control, 2022
1 Introduction
Cardiovascular diseases [CVDs] are listed as the leading single cause of death by the World Health Organization [1]. In recent years, the incidence and mortality of these diseases have continued to increase and have even exhibited a trend amongst the youth. At present, an estimated 2.3 million patients suffer from atrial fibrillation [AF] in the United States alone, whilst patients with AF in the European Union are estimated to be 4.5 million; the number of patients in the United States is expected to increase 2.5 times by 2050 [2]. CVDs primarily include coronary heart disease, hypertension, myocardial infarction [MI] and heart failure. Patients frequently suffer from heart arrhythmias before being diagnosed with CVD.
An electrocardiogram [ECG] is a non-invasive detection method that is widely used to detect potential heart conditions. The heart condition of a patient can be determined on the basis of ECG information. Early detection is beneficial for the diagnosis and treatment of various heart diseases, such as atrial premature beat, ventricular premature beat, left bundle branch block [LBBB] and right bundle branch block [RBBB], to ensure a high survival rate of patients.
In traditional medicine, the careful interpretation of ECGs by doctors in an artificial way is necessary for diagnosing life-threatening CVDs. The automatic classification of ECGs can not only provide an objective diagnosis but also save doctors’ time. Therefore, the detection and classification of ECGs exhibit high clinical significance, and they help promote clinical research on CVDs.
In the past few decades, researchers have developed many effective methods for automatically detecting ECG signals. For example, in the automatic detection of AF, some detectors can detect CVDs based on p-wave deletion or RR interval variation [R is a symbol for the beginning of ventricular depolarisation]. Dash et al. proposed an AF automatic detection algorithm based on the randomness, variability and complexity of heartbeat interval time series [3].
The morphological features of this method were composed of wavelet transform and independent component analysis extracted features. The wavelet features consisted of the fourth-order approximation coefficient and the third- and fourth-order detail coefficients [5]. The extracted features of the independent component analysis comprised a set of independent source signals recovered from the observation signals [6].
Although various feature classifiers have been developed, most of these methods involve manual extraction and traditional classification. To increase their accuracy, these methods must spend considerable time in determining the best combination of features. Moreover, the process requires extensive expertise in the field of communication. Therefore, feature extraction or selection poses a challenge to nonmedical researchers and even to those in the field.
In particular, the challenge originates from several aspects. Firstly, the ECG waveforms and time characteristics of different patients vary in basic ECG mode. Patients have different heartbeat waveforms, and even the same patient is likely to have varying heartbeat waveforms at different times. Secondly, the processing of heart rate variability also poses a problem to the classification of ECG signals because heart rate involves too many variables, including physiological and psychological conditions, such as stress, excitement, exercise and changes in ECG characteristics [e.g. RR and PR intervals]. Consequently, some researchers have begun using neural networks to extract features and heartbeat automatically.
Escalona-Moran et al. [7] presented a neural network based on the convolution of a 2D heartbeat as a classification method. In this method, the researcher converts a series of three adjacent beats into a 2D coupling matrix input, enabling the convolution filter to easily capture the continuous waveform of adjacent heartbeats and the correlation between beats. The method achieved a final sensitivity of 76.8%, with positive predictive values of 74.0% for SVEB and 93.8% for VEB.
Deep learning, which exhibits strong feature extraction capability, has achieved considerable success in computer vision in recent years [8]. In particular, the convolutional neural network [CNN] method is the most widely used deep learning model. Its applications to medical imaging, gene recognition, speech processing, sleep apnea detection and other aspects have demonstrated robustness.
Accordingly, researchers have exhibited an interest in using deep neural networks [DNNs] for the automatic detection of ECG signals. Salem al. [9] developed an automatic AF detection method based on CNN. AF features are automatically learned and applied to the classification module. This method simplifies the feature extraction process without requiring expert feature engineering to determine the suitability and criticality of features. Zhang et al. [10] proposed a multi-scale CNN [MCNN] that performs the timescale transformation of input signals and AF detection on the basis of the scale transformation input. The depth of MCNN in their model was strongly correlated with detection performance. Although the aforementioned methods are experimentally effective in addressing specific CVD detection problems, their good performance is typically based on carefully selected clean data or a small number of testers, and thus, their applicability is generally limited. Therefore, achieving the generalisation capability of models to detect CVDs reliably from limited single-lead ECG records remains a considerable challenge.
For medical and health-related applications, the authors of [11] mostly contributed to the integration of existing ECG detection classification and deep learning technologies through various components, such as algorithms, architecture, hardware design, software development, performance optimisation and CVD detection.
Through the review and analysis of relevant studies, the current work focuses on illustrating the effort of using deep learning technology in the diagnosis and detection of CVDs.
The current work also discusses the importance, motivation and challenges of deep learning in the application of ECG classification. It provides recommendations for future research and identifies trends from the perspective of how hospitals, healthcare professionals and patients can benefit from using ECG classification technology.
This paper is organised as follows. Section 1 presents basic information about deep learning and explains the necessity of using this method in ECG detection and classification. Section 2 discusses the research methodology used in the systematic review presented in this paper. Section 3 describes current research on using deep learning in ECG applications. Section 4 provides the results of the systematic review in terms of motivation, challenges and recommendations. Section 5 presents the conclusion of this work.
Read full article
URL: //www.sciencedirect.com/science/article/pii/S1746809422000155