2124 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO.

5, MAY 2022

Analysis of ECG Signals by Dynamic

Mode Decomposition
Honorine Niyigena Ingabire , Kangjia Wu, Joan Toluwani Amos, Sixuan He, Xiaohang Peng ,
Wenan Wang , Min Li, Jinying Chen, Yukun Feng, Nini Rao , and Peng Ren

Abstract—Objective: Based on cybernetics, a large electrical changes that occur as a result of depolarization and
system can be divided into subsystems, and the stability repolarization of the cardiac muscle over each heartbeat (cardiac
of each can determine the overall properties of the system. cycle) [1]. These changes reflect a series of waves, P, Q, R, S,
However, this stability analysis perspective has not yet
been employed in electrocardiogram (ECG) signals. This is and T waves. The electrical potential difference between a pair of
the first study to attempt to evaluate whether the stability electrodes during each cardiac cycle is graphically represented
of decomposed ECG subsystems can be analyzed in order as a lead. The standard ECG is comprised of 12-leads including
to effectively investigate the overall performance of ECG bipolar limb leads (I, II, and III) and unipolar leads (augmented
signals, and aid in disease diagnosis. Methods: We used
limb leads i.e., aVR, aVL, and aVF, and precordial chest leads
seven different cardiac pathologies (myocardial infarction,
cardiomyopathy, bundle branch block, dysrhythmia, hyper- i.e., V1, V2, V3, V4, V5, and V6). The 12-lead ECG reflects the
trophy, myocarditis, and valvular heart disease) to illustrate 3-dimensional electrical activity of the heart captured from 12
our method. Dynamic mode decomposition (DMD) was different viewpoints (or leads). These leads reflect the electrical
first used to decompose ECG signals into dynamic modes activity of different anatomic areas of the heart.
(DMs) which can be regarded as ECG subsystems. Then, Cardiac diseases, such as myocardial infarction (MI), car-
the features related to the DMs stabilities were extracted,
and nine common classifiers were implemented for diomyopathy (CM), bundle branch block (BBB), hypertrophy
classification of these pathologies. Results: Most features (HT), and valvular heart disease (VHD), have been widely diag-
were significant for differentiating the above-mentioned nosed by monitoring 12-lead ECG changes [2]–[4]. Analysis of
groups (p value<0.05 after Bonferroni correction). In a single-lead ECG signal has low computational cost and is more
addition, our method outperformed all existing methods easily interpretable than 12-lead ECG signals. However, 12-lead
for cardiac pathology classification. Conclusion: We
have provided a new spatial and temporal decomposition ECG has many advantages: (1) It can fully capture signs of
method, namely DMD, to study ECG signals. Significance: cardiac abnormalities located in any anatomic area of the heart,
Our method can reveal new cardiac mechanisms, which contributing to cardiac disease diagnosis. (2) It can fully reflect
can contribute to the comprehensive understanding of its the underlying dynamics of the heart by providing comprehen-
underlying mechanisms and disease diagnosis, and thus, sive information about its activities. This paper regards the heart
can be widely used for ECG signal analysis in the future.
as a complex large-scale system, and its dynamic behaviors are
Index Terms—Cardiac pathologies, dynamic mode captured from 12-lead ECG signals in macroscopic quantities.
decomposition (DMD), electrocardiogram (ECG), multi-lead Since single-lead ECG only captures limited information on
ECG, stability, subsystems.
heart mechanisms, therefore, 12-lead ECG signals were used
I. INTRODUCTION for further analysis.
Stability analysis is an important issue in cybernetics (or
A. Stability Analysis of ECG Signals
system science) with many applications across fields, such as
Electrocardiogram (ECG) is an important tool to capture the power systems, product design, and industrial manufacturing
electrical activity of the heart over time. Conductive electrodes [5]–[8]. Recently, it has received increased attention in the life
are selectively placed on the body’s surface to capture the small science. Previous studies have also revealed that the health status
of individuals can be reflected in instability of their physiological
Manuscript received November 4, 2020; revised October 30, 2021;
signals, including blood oxygen saturation, respiration, and elec-
accepted November 15, 2021. Date of publication November 24, 2021; troencephalography signals. For example, analysis of the stabil-
date of current version May 5, 2022. This work was supported in part ity of physiological signals, such as the heart rate, respiration,
by the National Natural Science Foundation of China under Grants
62171109 and 61872405, in part by the Key R&D Project of Sichuan
and blood oxygen saturation of a patient with sleep apnea led
Province under Grants 2020YFS0094 and 2020YFS0243, and in part to the successful extraction of abnormal breathing patterns [9].
by University of Electronic Science and Technology of China. (Corre- Similarly, Hocepied et al. found that electroencephalography
sponding authors: Nini Rao; Peng Ren.)
The authors are with the Department of Biomedical Engineering,
signals of epileptic patients are less stable compared with those
University of Electronic Science and Technology of China, Chengdu of healthy individuals [10]. According to these findings, Glass et
610054, China (e-mail: raonn@uestc.edu.cn; pren28@uestc.edu.cn). al. concluded that most pathological conditions can be reflected
This article has supplementary downloadable material available at
https://doi.org/10.1109/JBHI.2021.3130275, provided by the authors.
in the instability of physiological signals [11]. Apart from this,
Digital Object Identifier 10.1109/JBHI.2021.3130275 it should also be noted that according to cybernetics, the overall

This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/
NIYIGENA INGABIRE et al.: ANALYSIS OF ECG SIGNALS BY DYNAMIC MODE DECOMPOSITION 2125

performance of a large-scale system can be determined by the not only aid in revealing new cardiac dynamic mechanisms, but
stability of many subsystems decomposed from it [12]. As also contribute to research on heart-related diseases.
mentioned above, in this study, the heart is regarded as a complex
large system with dynamic behaviors which can be reflected
and measured by multi-lead ECG signals. Thus, it is essential C. Cardiac Abnormalities
to evaluate how the stability of decomposed ECG subsystems We evaluated our approach using the ECG signals of patients
affects the overall properties of multi-lead ECG signals from with various cardiac abnormalities, such as MI, CM, BBB, HT,
subjects suffering from diseases. This perspective has not yet VHD, and myocarditis (MCD). MI, also known as heart attack,
been evaluated by previous studies. is the most severe cardiovascular disease and one of the top
Finally, it must be noted that some previous studies evaluated causes of mortality in the world. Every year, more than 8 million
the stability of ECG signals based on the variability of different people die globally from MI [20]. It occurs as a result of necrosis
ECG waves, such as the T-wave and QT-interval [13]. However, of heart cells and permanent damage of the heart muscle due
this study refers to the stability of ECG signals from system to prolonged insufficient oxygen supply (ischemia), which is
science theory, which has an entirely different physiological caused by narrowed coronary arteries. This can lead to both
meaning from the previous studies (see the Methods section acute infarction and sudden death [4]. MI characteristics are
for more details). We attempt to use a novel approach, namely captured by monitoring ECG changes, including ST-segment
dynamic mode decomposition (DMD), to demonstrate the ef- elevation and depression, P-wave, T-wave, and QRS-complex
fectiveness of our perspective for analyzing ECG signals. abnormalities [4]. Dysrhythmia (DT) occurs as the result of
problems in the electrical conduction system. Cardiac muscle
diseases including CM and HT are characterized by the presence
of T-wave inversion, ST-segment depression, and deep T-wave
B. Dynamic Mode Decomposition
inversion in leads I, II, aVL, aVF, V4, V5, and V6. In VHD, the
DMD is a powerful new decomposition method capable of heart valves do not close or open properly. MCD occurs when
capturing coherent spatial-temporal patterns from complex data there is inflammation of the myocardium (heart muscle), which
by assuming that the spatial pattern of the observed data at each is as a result of the immune response to infections. BBB occurs
time point is given by a linear combination of spatial patterns due to delays in the heart conduction process. The characteristics
at previous time points [14]. This method was initially imple- of right BBB include the presence of wide R-wave and S-wave
mented for the analysis of fluid flows [15]. It is a data-driven in the leads V1 and V6 [2]. For left BBB, the amplitude of the
approach, which does not require any governing equation or R-wave and the duration of the QRS complex are high, and the
prior assumptions of underlying system dynamics. In addition, it T-wave inversion is present in the precordial leads [2].
encompasses the ability of both singular value decomposition (to The aforementioned pathologies have been detected us-
extract inherent modes from high-dimensional data) and spectral ing various approaches, including time-domain, WT, discrete
analysis (to assess oscillatory frequencies correspond to those Fourier transform (DFT), discrete cosine transform, PCA, EMD,
inherent modes). For these reasons, DMD has recently been used and neural network methods [13], [21]–[32]. For example, Sad-
to analyze high-dimensional and dynamic physiological signals hukhan et al. extracted the phases after implementing DFT
and public health data, such as epidemiological data, brain- on ECG signals for MI detection [22]. Acharya et al. used
related signals, and others. For instance, DMD was successfully three types of coefficients based on discrete wavelet transform,
employed to interpret the spread of three infectious diseases, discrete cosine transform, and EMD for the detection of MI [23].
using flu activity data provided by Google’s Flu Trends tool, pre- In addition, Reasat and Shahnaz, and Liu et al., implemented
vaccination measles from the U.K., and type-1 paralytic polio different types of neural networks for MI detection [33], [34].
cases in Nigeria [16]. Furthermore, Brunton et al. used DMD Tripathy et al. implemented principal component multivariate
to analyze large-scale sleeping electrocorticography data, and multiscale sample entropy for detection and classification of MI,
were successful in detecting spindle networks during sleep [17]. CT, DT, and HT [35]. In [36], BBB was detected based on a
Similarly, Solaija et al. used DMD to accurately detect epileptic complex wavelet sub-band dual-spectrum. Jain and Bhaumik
seizures of electroencephalography signals, and Casorso et al. applied a specific signal processing technique for integrated
also used this method to analyze resting-state and motor-task circuits on ECG signals for HT, DT, and BBB detection [37].
function magnetic resonance imaging data to model the brain’s Meanwhile, different approaches have been proposed for mul-
spatial-temporal organization [18], [19]. However, it has not yet ticlass ECG classification [35], [38]–[40]. For example, Deng et
been used to analyze multi-lead ECG data, which is also dynamic al. and Dey et al., respectively used convolution neural network
and high-dimensional. In addition, in contrast to other tradi- (CNN) for eight classes of ECG (HC, MI, BBB, CT, DT, HT,
tional data decomposition methods, such as principal component MC, and VHD) and three classes of ECG (HC, MI, and non-MI)
analysis (PCA), empirical mode decomposition (EMD), and classification [39], [40]. This study attempts to implement DMD
wavelet transform (WT), only DMD can decompose complex to analyze ECG signals in order to improve classification of
(high-dimensional and dynamic) data into subsystems, namely binary-class and multi-class ECG signals. Specifically, our pro-
dynamic modes (DMs), with degree of stability. Therefore, this posed method was evaluated on different types of binary-class
study attempts to analyze ECG signals based on stable and ECG signals (MI versus normal ECG signals, BBB versus
unstable DMs from multi-lead ECG signals. This approach may normal ECG signals, CT versus normal ECG signals, DT versus
2126 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 5, MAY 2022

In the section below, we demonstrate the proposed method

to analyze multi-lead ECG signals to detect of cardiovascular
pathologies. The detection block (see Fig. 1) consists of prepro-
cessing, feature extraction based on DMD, and classification.

A. Preprocessing
The preprocessing includes filtering and beat segmentation.
For filtering, artifacts such as noise and baseline wander were
filtered out by the Daubechies wavelet basis function, and pow-
erline interference was removed by a second-order notch filter,
then a fourth-order butterworth low pass filter with a cutoff
frequency of 100 Hz was used to reduce very high-frequency
content [42]. These preprocessing techniques have been applied
by different studies for ECG analysis and are effective in filtering
different artifacts, including motion and baseline wander.
For beat segmentation, the R-wave is considered as a distinc-
tive point due to its high amplitude and clearly visible peak.
The Pan-Tompkins algorithm has been widely used for R-peak
detection [43]. We used this algorithm similarly to identify
the R-peaks. Once the R-peaks were identified, a multi-lead
segment of 651 time points, including 250 points before the
R-peak, and 400 points after, was considered as the ECG beat.
Fig. 1. The flow chart of this study. Note: classification performance This beat length was chosen because it has been employed in
includes accuracy, precision, recall, and area under the curve (AUC)
values. previous studies which used the PTB database [23], [31]. These
studies have demonstrated that this length generally contains
a large proportion of information on the heartbeat, and can be
normal ECG signals, and HT versus normal ECG signals) and successfully used to detect ECG beats for different individuals.
classification of eight-class ECG signals, including HC, MI, It should be noted that to avoid inconsistency in beat length, the
BBB, CT, DT, HT, MCD, and VHD. first and last beats of each subject were not considered.
The flowchart of this study is shown in Fig. 1. First, the multi- The aforementioned pathologies, such as DT (ventricular
lead ECG data was preprocessed in order to remove artifacts, and arrhythmia, supraventricular arrhythmia, premature ventricular
further segmented into ECG beats and frames. Each individual ectopic beats, and sinus arrhythmia), HT, and BBB, are mainly
beat and frame signal were then decomposed into DMs using diagnosed using frame-based processing; in other words, beat
the DMD algorithm. Then, eigenvalues were used to determine segmentation is not required during preprocessing [44]. In this
whether their corresponding DMs were either stable or unstable. way, during preprocessing, multi-lead ECG signals were seg-
After, the features related to the stability of these DMs were mented into non-overlapping frames of 4096 time points (4.096
extracted, and the statistical tests were performed on each feature seconds). This frame size was chosen by following the previous
to assess the differences between the stability changes of DMs studies which used the PTB database [30], [35], [36], [38]. Thus,
of ECG beats and frames of patients with different cardiac in order to easily compare our proposed method with those
abnormalities and HCs. Last, classification performance was of existing studies that used either beat-based or frame-based
evaluated via nine common classifiers. processing, both processing methods were considered. After pre-
processing, the multi-lead ECG signal features were extracted
using DMD.
II. METHODS
This study used Physikalisch-Technische Bundesanstalt
(PTB), a well-known public diagnostic ECG database provided
by the national metrology institute of the Federal Republic of B. Dynamic Mode Decomposition
Germany, for analysis [41]. The PTB ECG database contains DMD is a powerful decomposition approach that is used
digitized ECG recordings from 148 MI patients, 15 BBB pa- to evaluate dynamic systems with high-dimensional data. The
tients, 18 CT patients, 14 DT patients, 7 HT patients, 4 MC DMD algorithm has the benefit of decomposing time series data
patients, 6 VHD patients, and 52 HCs with an average age of 57.2 into a set of DMs that contain spatial and temporal patterns.
years. Each ECG time series was recorded by the 12 standard Tu et al. has refined the definition of DMD algorithm, as shown
leads (I, II, III, aVR, aVL, aVF, V1, V2, V3, V4, V5, V6). below [45]. Let us consider two n × (k − 1) raw data matrices X
The duration of each recording was approximately 2 minutes, and Xs , whose rows and columns respectively denote the leads
and each ECG signal was sampled at 1000 Hz with a 16-bit and sampling points of ECG beats. In the case of 12-lead ECG
resolution over a range of ±16.384 mV. signals, n is equal to 12. These matrices can be constructed by
NIYIGENA INGABIRE et al.: ANALYSIS OF ECG SIGNALS BY DYNAMIC MODE DECOMPOSITION 2127

arranging measurements from k sampling time points as follows:

⎛ ⎞
x11 · · · x1(k−1)
⎜ ⎟ 
X = ⎝ ... . . . ... ⎠ = x1 · · · xk−1 (1)
xn1 · · · xn(k−1)
⎛ ⎞
x12 · · · x1k
⎜ ⎟ 
Xs = ⎝ ... . . . ... ⎠ = x2 · · · xk (2)
Fig. 2. The procedures to compute the stability of DMs and their
xn2 · · · xnk corresponding frequencies.

where columns (or snapshots) of Xs are obtained by shifting

those in X by one time point, therefore the data in these matrices
largely overlap.
DMD assumes that the temporal progression from X to Xs is
governed by a linear operator A, which satisfies the following
relationship:
Xs = AX (3)
Linear regression of the non-linear dynamics between these
consecutive data matrices, i.e., X and Xs , can be determined by
estimating the eigendecomposition of this operator A based on
one possible approach, which is to calculate the pseudoinverse
of X using its SVD.
It is important to note that the snapshots in the data matrix X
are assumed to be high dimensional, i.e., if the measurements
n of a snapshot are larger than the total number of snapshots
k − 1, then the transition matrix A (with the size n × n) may be Fig. 3. The magnitudes and oscillatory frequencies of eigenvalues of
high dimensional. As a result, it may not be straightforward DMs for one ECG beat. (a) Healthy control (HC); and (b) Myocardial
to compute the eigendecomposition of A. Thus, instead of infarction (MI) patient.
calculating A directly, the DMD algorithm uses a reduced matrix
à given by projecting Aonto the leading singular vectors of X
to compute the eigendecomposition of A using the following Step 5: Compute the DMs Φ of X using the eigenvectors W
procedure: and time-shifted snapshot matrix Xs :
Step 1: Determine the SVD of the first data matrix: Φ = Xs VΣ−1 W (9)
∗
X ≈ UΣV (4) It should be noted that these DMs Φ denote the eigenvectors of
Then, substitute (4) into (3) to get the SVD of X : s the high-dimensional operator A, and each DM φi corresponds
to an eigenvalue λi given by Λ as shown in Tu et al. [45].
Xs = AUΣV∗ (5) Finally, the observed data can be approximately constructed


where U, Σ, and V represent left singular vectors, singular as the simple dynamic model X(t):
values, and right singular vectors, respectively. 
X(t) ≈ X(t) = Φ exp(Ωt)b (10)
Step 2: Compute the pseudo-inverse of X to get the matrix A:
where Ω = log(Λ)/Δt is a diagonal matrix containing eigen-
A = Xs X−1 = Xs VΣ−1 U∗ (6) values in continuous time, t is time, Δt is the time difference
Step 3: Project A onto the proper orthogonal decomposition between two consecutive points, and b is a vector containing a
modes of U to get à : set of weights to match the initial time point measured, such that
b = Φ−1 x1 .
à = U∗ AU = U∗ Xs VΣ−1 (7) The essence of the above algorithm is to decompose data
arranged, as in (1) into a set of coupled spatial-temporal patterns.
Step 4: The eigendecomposition of à is calculated as follows: Note that both Φ and Λ are complex values. Fig. 4 shows an
ÃW = WΛ (8) example of one stable and unstable DMs for a single ECG beat
(each DM contains real and imaginary parts). An eigenvalue
where the columns of W are the eigenvectors of Ã, and the can be expressed as λi = ri jωi , where ri denotes the damp-
elements λi of the diagonal matrix Λ are the eigenvalues of the ing ratio and ωi (element of Ω) denotes the frequency of φi .
full matrix A, which are also the DMD eigenvalues of data X. The oscillatory frequency Fi of each DM can be determined by
2128 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 5, MAY 2022

DMs. Each element of unstable DM φ ud , namely φ ud (l),

corresponds to each lead where l = 1, 2, . . . , 12. Φs is the
matrix containing stable DMs, and Λs is the diagonal matrix
containing their corresponding eigenvalues, as follows:

Φs = [φ s1 , . . . , φ sc , . . . , φ sC ] (15)
⎡ s ⎤
λ1 0 . . . . . . 0
⎢ . ⎥
⎢ 0 . . . 0 0 .. ⎥
⎢ ⎥
⎢ ⎥
Λs = ⎢ ... 0 λs 0 ... ⎥ (16)
⎢ c ⎥
⎢ . ⎥
⎣ .. 0 0 . . . 0 ⎦
0 . . . . . . 0 λsC
where φ sc is a vector which represents stable DM and λ sc is
the element which denotes its corresponding eigenvalue, where
c = 1, 2, . . . , C, and C denotes the total number of stable DMs.
Each element of stable DM φ sc , namely φ sc (l) , corresponds
to each lead. Since the number of ECG leads (12 in this study)
is always less than the number of snapshots (or time points of
each beat), i.e., 12 < k, the observed non-zero singular values
given by SVD X is smaller than both the number of leads
Fig. 4. The values of real and imaginary parts of single stable and
(12) and the total number of snapshots (k − 1). Consequently,
unstable DMs for one ECG beat. The arrows indicate the position of the the maximum number of DMs is limited to 12, and these few
leads. The color bar indicates the values of real and imaginary parts of DMs are insufficient to fully capture the dynamic activity of the
each element of these DMs. Each of these elements correspond to in-
dividual lead. Note: These DMs were randomly selected from individual
cardiac system. To solve this problem, we followed the method
beat’s 12-lead ECG signals of one subject for illustration. suggested by Brunton et al., which is to increase rows of the data
matrix X to at least twice the number of columns by stacking
the imaginary part of ωi as follows: time-shifted versions of the original signal to get an augmented
data matrix Xaug . It should be noted that in practical application,
imag(ωi ) DMD is implemented on the augmented matrix Xaug , rather
Fi = (11)
2π than X. The detailed steps for data augmentation are shown
According to the theory of discrete-time linear systems, the in the Supplementary Material. After applying DMD on the
magnitude of eigenvalue |λi | indicates the stability of the system, augmented data Xaug , we restacked the extracted DMs from this
where the system is considered as asymptotically stable when augmented data matrix to get DMs whose elements correspond
|λi | < 1. However, if |λi | > 1, the system is considered unstable to the 12 leads as shown in the Supplementary Material.
[46]. Therefore, the stability and frequency of each DM can be
indicated by its corresponding eigenvalue (see Figs. 2 and 3) C. Feature Extraction
[47]. The stable and unstable DMs of a 12-lead ECG signal and
their corresponding eigenvalues are shown as follows: After preprocessing, we decomposed the signal contents of

u  each beat and frame into a set of stable and unstable DMs in
Λ 0 each time interval, then we derived three types of features: the
Φ = [Φu , Φs ] ⇔ Λ = (12)
0 Λs feature reflecting the ratio of number of unstable DMs to total
DMs, the features derived from eigenvalues, and the features
where Φu is the matrix containing unstable DMs, and Λu is the
derived from DMs (or eigenvectors).
diagonal matrix containing their corresponding eigenvalues, as
The first type of feature was derived in order to evaluate the
follows:
ratio of unstable to total DMs as follows:
Φu = [φ u1 , . . . , φ ud , . . . , φ uD ] (13) D
⎡ u ⎤ RN = (17)
λ1 0 . . . . . . 0 (D + C)
⎢ . ⎥
⎢ 0 . . . 0 0 .. ⎥ where RN represents the ratio of unstable DMs to total DMs.
⎢ ⎥
⎢ ⎥ Then, we derived the parameters based on eigenvalues, be-
Λu = ⎢ ... 0 λu 0 ... ⎥ (14)
⎢ d ⎥ cause they can generally reflect information inherent to their cor-
⎢ ⎥
⎣ . . . 0 0 ... 0 ⎦ responding DMs. These parameters were extracted as follows.
0 . . . . . . 0 λuD We examined the relationship between eigenvalues of stable and
unstable DMs as shown below:
where φ ud is a vector which represents unstable DM, and λ ud is D
d=1 |λ d |
u
an element which denotes its corresponding eigenvalue, where Rλ =  D C (18)
d = 1, 2, . . . , D , and D denotes the total number of unstable d=1 |λ d | +
u
c=1 |λ c |
s
NIYIGENA INGABIRE et al.: ANALYSIS OF ECG SIGNALS BY DYNAMIC MODE DECOMPOSITION 2129

where Rλ represents the ratio of eigenvalue magnitudes of Since the raw ECG signal data are strictly real values, the
unstable DMs to those of total DMs. Furthermore, we examined decomposition generates complex conjugate pairs of DMs and
the eigenvalue magnitudes of the most stable and unstable DMs, eigenvalues. Thus, the phases of each pair of modes are opposite
respectively demonstrating the fast convergent and divergent to each other, and the observed results in (24) and (26) will be
DMs, as follows: equal to 0. In order to solve this problem, for each DM, the
phase of lead I was subtracted from the phase of each lead, then
λmin = min (|Λs |) (19)
the relative phases of all DMs were averaged. It should be noted
λmax = max (|Λu |) (20) that these newly developed parameters were extracted from ECG
signals in each beat as mentioned above.
where λmin and λmax respectively indicate the eigenvalues of
the most and least stable DM.
Finally, we derived the features based on DMs (or eigenvec- D. Statistical Tests
tors). For an individual beat and frame in 12-lead ECG signals, For two-class ECG analysis, a total of 54 features extracted
each element of DM φ ud or φ sc (with the size 12 × 1), namely from the stable and unstable DMs of each ECG beat and frame
φ ud (l) or φ sc (l), contains two important pieces of information: for the HC group and patients with each of the mentioned
the magnitude of the element (absolute value), providing a pathologies (MI, CM, DT, BBB and HT) were utilized for
measure of the lead’s participation within DM; and the phase of analysis. The Jarque-Bera test was applied on each feature
the element (angle between the real and imaginary components). for normality evaluation at a significance level of 0.05. This
Therefore, this study attempts to use the magnitude as well as test demonstrated that most of the extracted features were not
phase of the stable and unstable DMs as useful information for normally distributed. Therefore, the Wilcoxon rank-sum test
the 12-lead ECG signals. We quantified the relationship between was implemented to investigate the DMs’ stability differences
the magnitudes of stable and unstable DMD modes as follows: between the two groups (i.e., HC group and each of the afore-
12 D mentioned pathologies group). The p value was considered
d=1 |φ d (l)|
u
RM = 12 D l=1
 12 C (21) significant if it was less than 0.05 after Bonferroni correction.
d=1 |φ d (l)| + c=1 |φ c (l)|
u s
l=1 l=1 For eight-class ECG analysis, 54 features were also extracted
where RM represents the overall ratio of the magnitudes of the from the stable and unstable DMs of each ECG beat and frame.
12 leads of unstable DMD modes to total DMD modes. We also Then, an analysis of variance test was conducted on each ex-
quantified how the phases of the stable DMs are related to those tracted feature to assess the significant differences between the
of unstable DMs as follows: DMs’ stability of ECG signals (ECG beats and frames) for the
normal, MI, CM, BBB, DT, HT, VHD, and MCD subjects.
RP
12 D E. Classification
d=1 angle (φ d (l))
u
= 12 D l=1
 12  C
l=1 d=1 angle (φ u (l))+
d l=1 c=1 angle (φ c
s (l)) After conducting the statistical tests, to prevent overfitting,
(22) only significant features were used to assess the performance of
our proposed approach for two-class (i.e., HC group and each
where RP denotes the overall ratio of the oscillatory phases of of the aforementioned pathologies group) and eight-class clas-
the 12 leads of unstable DMs to total DMs. Furthermore, we sification of ECG signals. This performance was assessed using
computed the average magnitude and phase of stable DMs of nine commonly used classifiers: k-nearest neighbor (KNN), J48
each lead as shown below: decision tree, random forest, random tree, AdaBoost, Bayes net,
C
|φ s | vote, support vector machine (SVM), and multilayer perceptron
Mlead = c=1 c
s
(23)
C (MLP) [48]–[56]. The open-source software WEKA was used to
C build these classifiers [57], and 10-fold stratified cross-validation
angle (φ sc )
Pslead = c=1 (24) (CV) was chosen to evaluate the performance of these models.
C Finally, the performance parameters of the mentioned classi-
where both Mslead and Pslead are 12 × 1 vectors whose elements fiers, including accuracy, precision (or positive predictive value
represent the average magnitudes and phases respectively of (PPV)), recall (or sensitivity), and area under the curve (AUC)
stable DMs of each lead. Finally, we computed the average were considered to evaluate the ability of our proposed method
magnitudes and phases of unstable DMs from each lead as shown for ECG signal analysis.
below:
D
|φ u | III. RESULTS
Mulead = d=1 d (25)
D A. Statistical Analysis
D
angle (φ ud ) For two-class ECG analysis, we found that out of the total
Plead = d=1
u
(26)
D of 54 features, 39 features were significant for the detection of
where Mulead and Pulead are also 12 × 1 vectors whose elements MI, 28 for CM, 32 for DT, 26 for BBB, and 15 for HT versus
represent average magnitudes and phases respectively of unsta- the HC, using either beat-based or frame-based processing (after
ble DMs from each lead. Bonferroni correction). Notably, the results obtained in assessing
2130 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 5, MAY 2022

for multiclass (eight-class) ECG signals, is consistent between

beat-based and frame-based processing. Specifically, the highest
accuracy values achieved when classifying beats and frames
were 99.88% and 99.95%, respectively.
Finally, we compared our proposed approach with those of
previous studies for classification of two-class and eight-class
ECG signals. The highest performance achieved by these studies
are summarized in Tables I, II, and III. It is clear that our
proposed method outperforms all of the existing methods.

IV. DISCUSSION
In this study, we consider patients with different cardiac
pathologies (MI, CM, DT, BBB, HT, MC, and VHD) and HCs as
an example, and applied a new decomposition method, namely
Fig. 5. The means and standard deviations of our features for MI
DMD, to decompose multi-lead ECG signals into stable and
patients and HCs. (a) RN denotes the ratio of number of unstable DMs unstable modes (or subsystems). Then, we extracted the features
to total DMs; (b) Rλ represents the ratio of the eigenvalue magnitudes of to quantify the relationships between these two types of modes.
unstable DMs to those of all DMs; (c) λmin denotes the eigenvalue of the
most stable DM; (d) RP represents the overall ratio of oscillatory phases
The results verified our hypothesis that the stability of ECG
of the 12 leads of unstable DMs to total DMs; and (e) RM denotes the modes (or subsystems) reflect the health status of individuals,
overall ratio of the magnitudes of 12 leads of unstable DMs to total DMs; which could be a new perspective for ECG signal analysis.
and (f) λmax represents the eigenvalue of the least stable DM (after
Bonferroni correction, ∗p < 0.05, ∗∗p < 0.0001).
As mentioned in the introduction, this study has hypothesized
that the stability changes of DMs in ECG signals are influenced
when individuals suffer from cardiac diseases, and the results
the significance of the differences between the beats and frames
conform to the hypothesis. We extracted 54 features which
of the patients with each of the aforementioned pathologies and
reflect the relationships between stable and unstable DMs. These
those from HCs exhibited a similar trend. Thus, we only show
features can be divided into three parts: (1) Feature which reflects
the results for MI detection. As illustrated in Fig. 5, MI patients
the proportion of number of unstable DMs to stable DMs RN . (2)
generally exhibit more instability in the DMs of ECG signals
Features extracted according to eigenvalues, including Rλ , λmin ,
compared to those from HCs. Furthermore, as shown in Fig. 6,
and λmax . (3) Features extracted from magnitudes and phases
for both MI patients and HCs, the average magnitude and phase
of eigenvectors, including RM , RP , Mslead , Pslead , Mulead , and
in each lead in both stable and unstable DMs are different from
Pulead . Interestingly, we found that 39, 28, 32, 26, and 15 features
those of other leads. Additionally, as shown in Fig. 6(a), most
from total of 54 were significant for the detection of MI, CM,
leads’ stable DMs in HCs (except leads II, III, and aVF) were
DT, BBB, and HT, respectively, using either beat-based or frame-
greater in magnitude than those from MI patients. Similarly, as
based processing. This demonstrates that our proposed method
shown in Fig. 6(d), except for lead II, and aVF, the other leads’
can be generally effective for two-class ECG classification.
oscillatory phases in unstable DMs from HCs were higher than
Also, 45 and 39 features out of the 54 were significant for the
those of MI patients.
classification of beats and frames respectively of seven cardiac
For eight-class ECG signal analysis, from the total of 54
pathologies. As mentioned previously, the statistical results
features, 45 were significant for distinguishing the beats of those
obtained for the detection of the aforementioned pathologies
eight classes, and 39 for the frames (after Bonferroni correction).
(MI, CM, DT, BBB, and HT) were similar. Therefore, we only
showed the results for MI detection. As illustrated in Fig. 5,
B. Classification greater instability in the DMs of MI patients was observed
As mentioned previously, 10-fold CV was used and per- compared to those in HCs. This may be due to the presence
formance parameters of the mentioned classifiers, including of ECG changes (such as variation of P-wave, QRS-complex,
accuracy, PPV, sensitivity, and AUC, were considered to assess T-wave, and ST-segment) during MI. All of these observed
the ability of our proposed method for ECG signal analysis. results conform to findings from previous studies, which state
For two-class ECG signal analysis, as shown in Figs. 7 and 8, that changes in the stability of physiological signals reflect
respectively, all the classifiers accurately identified the normal individuals’ pathological conditions. In addition, according to
beats and frames from MI, CM, BBB, DT, and HT beats and the theory of cybernetics, stable subsystems can generally assist
frames. Specifically, the highest accuracy values achieved in the whole system to work more efficiently, orderly, and adaptive,
detecting MI, CM, BBB, DT, and HT beats were 99.97%, 99.5%, while unstable subsystems are almost the opposite. This study
100%, 99.85%, and 99.61%, respectively (see Fig. 7). Accuracy has found a similar phenomenon in the cardiac system - that
values of 100%, 100%, 100%, 100%, and 99.8% were obtained is, the DMs of HCs are less stable compared to those of MI
for detection of MI, CM, BBB, DT, and HT frames, respectively patients, which indicates that their cardiac systems function in
(see Fig. 8). a more efficient, orderly, and adaptive manner than those of MI
As shown in Fig. 9, the classification performance achieved patients. However, we also found some interesting phenomena:
in the case of seven different abnormal and normal ECG signals (1) As shown in Fig. 6, for both stable and unstable DMs,
NIYIGENA INGABIRE et al.: ANALYSIS OF ECG SIGNALS BY DYNAMIC MODE DECOMPOSITION 2131

Fig. 6. The means and standard deviations of averaged magnitudes and phases of DMs of individual leads of MI patients and HCs. (a)
Mslead represents the averaged magnitudes of stable DMs; (b) Mu lead denotes the averaged magnitudes of unstable DMs; (c) Plead denotes
s
the averaged phases of stable DMs; and (d) Pu
lead represents the averaged phases of unstable DMs (after the Bonferroni correction, ∗p < 0.05,
∗∗p < 0.0001).

Fig. 7. Classification performance (accuracy, precision, recall, and AUC) for our features when identifying MI, CM, DT, BBB and HT from normal
beats.

Fig. 8. Classification performance (accuracy, precision, recall, and AUC) for our features for identifying MI, CM, DT, BBB and HT frames from
normal frames.
2132 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 5, MAY 2022

TABLE I
PERFORMANCE COMPARISON OF THE PROPOSED METHOD WITH EXISTING APPROACHES USED THE PTB ECG DATABASE FOR DETECTION
OF ECG PATHOLOGIES

Note: #: number; &: and; Acc.: accuracy; PPV: positive predictive value; Sens.: sensitivity; N/R: not reported; WT: wavelet transform;
DWT: discrete wavelet transform; DCT: discrete cosine transform; EMD: empirical mode decomposition; DFT: discrete Fourier transform;
PCA: principal component analysis; CNN: convolution neural network, and RNN: recurrent neural network. It should be also noted that
confidence intervals obtained by these methods were not reported.

TABLE II
PERFORMANCE COMPARISON OF THE PROPOSED METHOD WITH EXISTING APPROACHES USED THE PTB ECG DATABASE FOR MULTICLASS
ECG SIGNAL CLASSIFICATION

Note: #: number; Acc.: accuracy; Sens.: sensitivity; N/R: not reported; It is also worth noting that the precision and AUC values achieved by these studies were not reported.
NIYIGENA INGABIRE et al.: ANALYSIS OF ECG SIGNALS BY DYNAMIC MODE DECOMPOSITION 2133

Fig. 9. The classification performance (accuracy, precision, recall, and AUC) for our developed features on multiclass (eight classes) ECG
(MI, CM, DT, BBB HT, VHD, MCD and HC) beats and frames.

TABLE III
PERFORMANCE COMPARISON OF THE PROPOSED METHOD WITH THE EXISTING APPROACHES USED PTB ECG DATABASE FOR MULTICLASS
ECG SIGNAL CLASSIFICATION

Note: #: number; Acc.: accuracy; PPV: positive predictive value; Sens.: sensitivity; N/R: not reported; It is also worth noting that PPV and AUC
value achieved by these studies were not reported.

the magnitudes and phases of different leads vary from each classifiers were implemented, because each of them utilizes dis-
other. This demonstrates that different leads can convey different tinct techniques for data classification. Encouragingly, as shown
useful information for disease diagnosis, hence it is meaningful in Fig. 7, each of the classifiers was able to accurately identify
to analyze these leads separately. (2) Furthermore, our results MI, CM, BBB, DT, and HT beats from those of HCs, with
ascertain that most leads of HCs in stable DMs (except lead II, respective highest accuracy values of 99.97%, 99.5%, 100%,
III, and aVF) exhibit larger magnitudes compared with those of 99.85%, and 99.61%. Also, 100% accuracy was obtained for
MI patients (see Fig. 6(a)). This demonstrates that the leads in the detection of MI, CM, BBB, and DT frames (see Fig. 8).
stable DMs of HCs are more active compared to those of MI Similarly, the highest accuracy values of 99.88% and 99.95%
patients, which can reflect the normal and sufficient adaptability were respectively achieved when classifying beats and frames
of their cardiac systems. of multiple heart pathologies (MI, CM, BBB, DT, MC, VHD,
(3) Similarly, as illustrated in Fig. 6(d), most leads of HCs in and HT) (see Fig. 9). Furthermore, we also summarized the
unstable DMs (except lead II and aVF) exhibit higher oscillatory results from previous studies for the classification of two-class
phases compared to those of MI patients. This may be because and multiclass ECG signals, for comparison with our observed
the cardiac systems of healthy individuals are capable of detect- results (see Tables I, II, and III). Interestingly, our proposed
ing and responding quickly to any changes subjected to them. (4) approach outperformed all of the existing approaches; in other
We also found that the average phases for some leads in stable words, this demonstrates the potential of our proposed approach
DMs, such as II, III, aVF, and V4, for the MI patient group for ECG signal analysis, helping cardiologists to enhance the au-
differed significantly from those in the HC group. However, tomated diagnosis of different cardiac pathologies. Additionally,
the average magnitudes of those leads were not significant (see it should be noted that our method was compared with a previous
Fig. 6(a) and (c)). Similarly, the average phases for some leads in study which used univariate ECG signal analysis (single-lead
unstable DMs, such as I, aVR, aVF, V3, V4, V5, and V6, differed ECG) in terms of storage and complexity [31]. We used a
significantly between the two groups. However, the average computer with Intel CORE 2.4 GHz (i5–4210U) processor and
magnitudes of those leads were not significant (see Fig. 6(b) 4 GB RAM. The execution times used for feature extraction
and (d)). This demonstrates that the phase and magnitude of and detection of MI (KNN classifier, training and testing using
individual leads contain different information, thus they all need 10-fold CV as in [31]) were 14.14 seconds and 30.06 seconds,
to be considered when analyzing stable and unstable DMs. respectively. The execution time of [31] was considerably less,
In order to further investigate the effectiveness of our pro- and their proposed method is good, but our proposed method is
posed approach in differentiating two and eight groups, nine still the best. Also, the performance of our proposed method was
2134 IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 26, NO. 5, MAY 2022

evaluated for detection of other four cardiac pathologies. The in [23] and [31], we followed their strategies, using 40182
classification of these pathologies achieved promising results. MI and 10546 HC beats, achieving the highest accuracy, PPV,
However, the method of Sharma et al. has not yet been tested sensitivity, and AUC values of 99.95%, 99.93%, 99.95%, and
for detection of other heart pathologies, we are not certain if 99.94%, respectively. These observed results are consistent with
their method is also suitable for detecting other heart diseases, those for 35010 MI and 10140 HC beats.
or for multiclass ECG signal classification [31]. Finally, the major contributions of our proposed approach can
It should also be noted that in order to conveniently and be summarized as follows. (1) We are the first to use DMD to
objectively compare our proposed approach with the existing evaluate how the stability of ECG modes (or subsystems) affects
approaches, we repeated the strategies of most previous studies the macrodynamic patterns of multilead ECG signals. No one
which used the PTB dataset. These strategies involve preprocess- used this stability concept to analyze physiological signals. The
ing and use of 12-lead ECG signals in the analysis. However, to results show that this stability is an inherent phenomenon for all
further demonstrate the improvement of our proposed approach, mentioned types of cardiac pathologies. Therefore, our proposed
we used MI as an example and included two supplementary method can reveal new cardiac mechanisms, which cannot be
evaluations: (1) we evaluated our proposed method on 9-lead achieved by other existing methods and can effectively help in
ECG signals after removing some leads (augmented leads aVR, disease diagnosis. Thus, it might be used in clinical studies. (2)
aVL, and aVF) to reduce the effect of the data redundancy In addition, in contrast to other traditional data decomposition
in leads I, II, III, aVR, aVL, and aVF, which may affect our methods, such as PCA, EMD, and WT, only DMD can de-
conclusion. The observed results were consistent with those compose high-dimensional and dynamic data into subsystems,
obtained on 12-lead ECG signals. We have reported the results of namely DMs, with stability. Therefore, this study is the first
12-lead ECG signals for comparison; but in the future, we might attempt to analyze ECG signals based on stable and unstable
also use 9-lead ECG signals to reduce computation time. (2) DMs of multi-lead ECG signals, which may not only help to
To assess our denoising performance, we evaluated the perfor- reveal new cardiac dynamic mechanisms, but also contribute
mance of our proposed method on noisy ECG signals, obtaining to research on heart-related diseases. (3) DMD was initially
respective highest accuracy, precision, recall, and AUC values of implemented in the analysis of fluid flows and has recently been
99.94%, 99.92%, 99.94%, and 99.93%. This demonstrates the used in analysis of high-dimensional and dynamic physiological
robustness of our proposed method in the case of noisy signals, signals and public health data, such as epidemiological data,
and can be useful in clinical data, which often suffer from noise brain-related signals, and others [15]–[19]. However, DMD has
and artifacts. However, compared with noisy ECG signals, we not yet been used to analyze multi-lead ECG signals, which are
achieved the highest classification performance on preprocessed also dynamic and high-dimensional.
ECG signals. This indicates good denoising performance on
ECG signals. V. CONCLUSION
Several further points also need to be mentioned. (1) As
mentioned before, the Pan-Tompkins algorithm is commonly In this study, we proposed a novel perspective for the analysis
used for R-peak detection [43]. Although it is not optimal, it of ECG signals, which relies on the stability analysis of ECG
can achieve the promising accuracy of 99.32% and is currently subsystems by decomposing ECG signals into stable and un-
regarded as one of the best approaches for detection of R-peaks stable modes using DMD. Furthermore, this study also demon-
[40]. Some R-peaks cannot be accurately detected by the algo- strates that the stability analysis of ECG modes (or subsystems)
rithm, but our proposed approach can still achieve the highest can reveal the underlying spatiotemporal dynamics of the cardiac
performance for the classification of cardiac pathologies. This system, and reflects the heart condition status of an individual.
demonstrates the robustness of our proposed method for cardiac Finally, our proposed method exhibits great potential as well as
disease diagnosis. (2) Regarding fast Fourier transform (FFT), high accuracy for diagnosis of heart disease, and might be widely
the power spectrum of each frequency point contains magnitude applied in clinical studies as well as various engineering-related
and phase information. Similarly, DMD can be regarded as a applications. In future, our proposed approach should be ap-
new way to calculate the spectra of multivariate signals, but it plied to other physiological signals, such as brain-related and
is not the same as FFT (which is used for univariate signals). In electromyography signals, to evaluate the stability of their DMs
this study, we calculated the phase, which has a totally different (or subsystems). This can help clinicians to comprehensively
physiological meaning from that of leads in raw ECG signals. understand their underlying mechanisms and diagnose their
Specifically, we computed the phase of DMs (which can be related diseases.
regarded as subsystems) of ECG signals. (3) The DMD al-
gorithm revealed the amplitude variation of two neighbor REFERENCES
points, then accurately uncovered the shape of each ECG wave, [1] G. Guven et al., “Biometric identification using fingertip electrocardio-
revealing the relationship among ST-segment, P-wave, T-wave, gram signals,” Signal, Image Video Process., vol. 12, no. 5, pp. 1–8,
Jan. 2018.
and QRS complex. Therefore, DMD can be considered as [2] A. L. Goldberger, Ed., Clinical Electrocardiography: A Simplified Ap-
another technique for ECG morphological analysis. (4) In the proach. 7th ed., Philadelphia, PA, USA: Mosby, 2006, pp. 329–337.
PTB database, apart from MI patients, most normal, CM, BBB, [3] J. A. Drezner et al., “Abnormal electrocardiographic findings in athletes:
Recognising changes suggestive of cardiomyopathy,” Brit. J. Sports Med.,
DT, HT, MCD, and VHD subjects have one ECG recording. vol. 47, no. 3, pp. 137–152, Feb. 2013.
Thus, we randomly selected one ECG recording for each subject. [4] K. Thygesen et al., “Third universal definition of myocardial infarction,”
In order to compare our method with the MI detection methods Circulation, vol. 126, no. 16, pp. 2020–2035, Jun. 2012.
NIYIGENA INGABIRE et al.: ANALYSIS OF ECG SIGNALS BY DYNAMIC MODE DECOMPOSITION 2135

[5] R. C. Dorf and R. C. Bishop, Modern control systems, 13th ed., New Jersey, [31] M. Sharma et al., “A novel automated diagnostic system for classification
NJ, USA: Pearson, 2016. of myocardial infarction ECG signals using an optimal biorthogonal filter
[6] P. Kundur et al., Power System Stability and Control. New York, NY, USA: bank,” Comput. Biol. Med., vol. 100, pp. 100–113, Nov. 2018.
McGraw-Hill, 1994. [32] Z. Lin et al., “Automated detection of myocardial infarction using robust
[7] M. V. Cook, Flight Dynamics Principles: A Linear Systems Approach features extracted from 12-lead ECG,” Signal, Image Video Process.,
to Aircraft Stability and Control. Oxford, U.K.: Butterworth-Heinemann, vol. 14, pp. 857–865, Jul. 2020.
2012. [33] T. Reasat and C. Shahnaz, “Detection of inferior myocardial infarction
[8] F. Tahami, R. Kazemi, and S. Farhanghi, “A novel driver assist stability using shallow convolutional neural networks,” in Proc. 5th IEEE R10-
system for all-wheel-drive electric vehicles,” IEEE Trans. Veh. Technol., HTC, 2017, pp. 718–721.
vol. 52, no. 3, pp. 683–692, May 2003. [34] W. Liu et al., “Real-time multilead convolutional neural network for my-
[9] L. A. Aguirre and Á. V. P. Souza, “Stability analysis of sleep apnea time ocardial infarction detection,” IEEE J. Biomed. Health Informat., vol. 22,
series using identified models: A case study,” Comput. Biol. Med., vol. 34, no. 5, pp. 1434–1444, Sep. 2018.
no. 3, pp. 241–257, Apr. 2004. [35] R. K. Tripathy et al., “A new way of quantifying diagnostic informa-
[10] G. Hocepied et al., “Stability analysis of epileptic EEG signals,” in Proc. tion from multilead electrocardiogram for cardiac disease classification,”
8th IEEE Int. Conf. BIBE, 2008, pp. 1–5. Healthcare Technol. Lett., vol. 1, no. 4, pp. 98–103, Oct. 2014.
[11] L. Glass and M. C. Mackey, “Pathological conditions resulting from [36] R. K. Tripathy and S. Dandapat, “Automated detection of heart ailments
instabilities in physiological control systems,” Ann. New York Acad. Sci., from 12-lead ECG using complex wavelet sub-band Bi-spectrum features,”
vol. 316, no. 1, pp. 214–235, Feb. 1979. Healthcare Technol. Lett., vol. 4, no. 2, pp. 57–63, Apr. 2017.
[12] J. Lunze, “Stability analysis of large-scale systems composed of strongly [37] S. K. Jain and B. Bhaumik, “An energy efficient ECG signal processor
coupled similar subsystems,” Automatica, vol. 25, no. 4, pp. 561–570, detecting cardiovascular diseases on smartphone,” IEEE Trans. Biomed.
Jul. 1989. Circuits Syst., vol. 11, no. 2, pp. 314–323, Apr. 2017.
[13] M. A. Hasan et al., “Increased beat-to-beat T-wave variability in my- [38] R. K. Tripathy and S. Dandapat, “Detection of cardiac abnormalities from
ocardial infarction patients,” Biomed. Eng., vol. 63, no. 2, pp. 123–130, multilead ECG using multiscale phase alternation features,” J. Med. Syst.,
Nov. 2018. vol. 40, no. 6, Jun. 2016, Art. no. 143.
[14] K. K. Chen et al., “Variants of dynamic mode decomposition: Boundary [39] M. Deng et al., “Extracting cardiac dynamics within ECG signal for human
condition, Koopman, and Fourier analyses,” J. Nonlinear Sci., vol. 22, identification and cardiovascular diseases classification,” Neural Netw.,
no. 6, pp. 887–915, Dec. 2012. vol. 100, pp. 70–83, 2018.
[15] C. W. Rowley et al., “Spectral analysis of nonlinear flows,” J. Fluid Mech., [40] M. Dey, N. Omar, and M. A. Ullah, “Temporal feature-based blassification
vol. 641, pp. 115–127, Sep. 2009. into myocardial infarction and other CVDs merging CNN and Bi-LSTM
[16] J. L. Proctor and P. A. Eckhoff, “Discovering dynamic patterns from from ECG signal,” IEEE Sensors J., vol. 21, no. 19, pp. 21688–21695,
infectious disease data using dynamic mode decomposition,” Int. Health, Oct. 2021.
vol. 7, no. 2, pp. 139–145, Nov. 2015. [41] A. L. Goldberger et al., “PhysioBank, Physiotoolkit, and Physionet: Com-
[17] B. W. Brunton et al., “Extracting spatial-temporal coherent patterns in ponents of a new research resource for complex physiologic signals,”
large-scale neural recordings using dynamic mode decomposition,” J. Circulation, vol. 101, pp. 215–220, Jun. 2000.
Neurosci. Methods, vol. 258, pp. 1–15, Jan. 2016. [42] J. P. Martínez, R. Almeida, S. Olmos, A. P. Rocha, and P. Laguna, “A
[18] M. S. J. Solaija et al., “Dynamic mode decomposition based epileptic wavelet-based ECG delineator evaluation on standard databases,” IEEE
seizure detection from scalp EEG,” IEEE Access, vol. 6, pp. 38683–38692, Trans. Biomed. Eng., vol. 51, no. 4, pp. 570–581, Apr. 2004.
Jul. 2018. [43] J. Pan and W. J. Tompkins, “A real-time QRS detection algorithm,” IEEE
[19] J. Casorso et al., “Dynamic mode decomposition of resting-state and task Trans. Biomed. Eng., vol. 32, no. 3, pp. 230–236, Mar. 1985.
fMRI,” NeuroImage, vol. 194, pp. 42–54, Mar. 2019. [44] E. A. Whitsel et al., “RR interval variation, the QT interval index and risk
[20] World Health Organization, “Global health estimates 2015: Deaths by of primary cardiac arrest among patients without clinically recognized
cause, age, sex, by country and by region, 2000-2015,” WHO, Geneva, heart disease,” Eur. Heart J, vol. 22, no. 2, pp. 165–173, Jan. 2001.
2016. [45] J. H. Tu et al., “On dynamic mode decomposition: Theory and applica-
[21] M. Arif et al., “Detection and localization of myocardial infarction using tions,” ACM J. Comput. Dyn., vol. 1, no. 2, pp. 391–421, Jun. 2014.
K-nearest neighbor classifier,” J. Med. Syst., vol. 36, no. 1, pp. 279–289, [46] M. H. Hayes, Schaum Outline of Digital Signal Processing. New York,
Mar. 2012. NY, USA: McGraw Hill Education, 1998.
[22] D. Sadhukhan et al., “Automated identification of myocardial infarction [47] K. Lehnertz et al., “State-of-the-art of seizure prediction,” J. Clin. Neuro-
using harmonic phase distribution pattern of ECG data,” IEEE Trans. physiol., vol. 24, no. 2, pp. 147–153, Apr. 2007.
Instrum. Meas., vol. 67, no. 10, pp. 2303–2313, Oct. 2018. [48] T. M. Cover and P. E. Hart, “Nearest neighbor pattern classification,” IEEE
[23] U. R. Acharya et al., “Automated characterization and classification Trans. Inf. Theory, vol. 13, no. 1, pp. 21–27, Jan. 1967.
of coronary artery disease and myocardial infarction by decomposition [49] U. Bashir and M. Chachoo, “Performance evaluation of J48 and Bayes
of ECG signals: A comparative study,” Inf. Sci., vol. 377, pp. 17–29, algorithms for intrusion detection system,” Int. J. Netw. Secur. Appl., vol. 9,
Jan. 2017. no. 4, pp. 1–11, Jul. 2017.
[24] A. K. Dohare et al., “Detection of myocardial infarction in 12 lead ECG [50] A. Liaw and M. Wiener, “Classification and regression by random forest,”
using support vector machine,” Appl. Soft Comput., vol. 64, pp. 138–147, R News, vol. 2, no. 3, pp. 18–22, Dec. 2002.
Mar. 2018. [51] A. Cutler and G. Zhao, “Pert-perfect random tree ensembles,” Comput.
[25] B. Halder et al., “Classification of complete myocardial infarction using Sci. Statist., vol. 33, pp. 490–497, Jan. 2001.
rule-based rough set method and rough set explorer system,” IETE J. Res., [52] W. Hu et al., “Adaboost-based algorithm for network intrusion detection,”
Apr. 2019. IEEE Trans. Syst., Man, Cybern. B. Cybern., vol. 38, no. 2, pp. 577–583,
[26] S. G. Al-Kindi et al., “Towards real-time detection of myocardial infarction Apr. 2008.
by digital analysis of electrocardiograms,” in Proc. 1st MECBME, 2011, [53] N. Friedman et al., “Bayesian network classifiers,” Mach. Learn., vol. 29,
pp. 454–457. no. 2/3, pp. 131–163, Nov. 1997.
[27] L. Sun et al., “ECG analysis using multiple instance learning for my- [54] R. E. Schapire et al., “Boosting the margin: A new explanation for the ef-
ocardial infarction detection,” IEEE Trans. Biomed. Eng., vol. 59, no. 12, fectiveness of voting methods,” Ann. Statist., vol. 26, no. 5, pp. 1651–1686,
pp. 3348–3356, Aug. 2012. 1998.
[28] E. S. Jayachandran et al., “Analysis of myocardial infarction using discrete [55] J. A. K. Suykens and J. Vandewalle, “Least squares support vector machine
wavelet transform,” J. Med. Syst., vol. 34, no. 6, pp. 985–992, May. 2010. classifiers,” Neural Process. Lett., vol. 9, no. 3, pp. 293–300, Jan. 1999.
[29] B. Liu et al., “A novel electrocardiogram parameterization algorithm and [56] S. K. Pal and S. Mitra, “Multilayer perceptron, fuzzy sets, and classifica-
its application in myocardial infarction detection,” Comput. Biol. Med., tion,” IEEE Trans. Neural Netw., vol. 3, no. 5, pp. 683–697, Sep. 1992.
vol. 61, pp. 178–184, Jun. 2015. [57] E. Frank et al., “Weka-a machine learning workbench for data mining,” in
[30] L. N. Sharma, R. K. Tripathy, and S. Dandapat, “Multiscale energy and Proc. Data Min. Knowl. Discov., 2009, pp. 1269–1277.
eigenspace approach to detection and localization of myocardial infarc-
tion,” IEEE Trans. Biomed. Eng., vol. 62, no. 7, pp. 1827–1837, Jul. 2015.