Journal of Ambient Intelligence and Humanized Computing (2021) 12:3227–3244

https://doi.org/10.1007/s12652-020-02536-4

ORIGINAL RESEARCH

Accurate detection of myocardial infarction using non linear features

with ECG signals
Chaitra Sridhar1 · Oh Shu Lih2 · V. Jahmunah2 · Joel E. W. Koh2 · Edward J. Ciaccio3 · Tan Ru San4 · N. Arunkumar5 ·
Seifedine Kadry6 · U. Rajendra Acharya2,7,8

Received: 11 March 2020 / Accepted: 5 September 2020 / Published online: 22 October 2020
© Springer-Verlag GmbH Germany, part of Springer Nature 2020

Abstract
Interrupted blood flow to regions of the heart causes damage to heart muscles, resulting in myocardial infarction (MI). MI
is a major source of death worldwide. Accurate and timely detection of MI facilitates initiation of emergency revasculariza-
tion in acute MI and early secondary prevention therapy in established MI. In both acute and ambulatory settings, the elec-
trocardiogram (ECG) is a standard data type for diagnosis. ECG abnormalities associated with MI can be subtle, and may
escape detection upon clinical reading. Experience and training are required to visually extract salient information present
in the ECG signals. This process of characterization is manually intensive, and prone to intra-and inter-observer-variability.
The clinical problem can be posed as one of diagnostic classification of MI versus no MI on the ECG, which is amenable
to computational solutions. Computer Aided Diagnosis (CAD) systems are designed to be automated, rapid, efficient, and
ultimately cost-effective systems that can be employed to detect ECG abnormalities associated with MI. In this work, ECGs
from 200 subjects were analyzed (52 normal and 148 MI). The proposed methodology involves pre-processing of signals
and subsequent detection of R peaks using the Pan-Tompkins algorithm. Nonlinear features were extracted. The extracted
features were ranked based on Student’s t-test and input to k-Nearest Neighbor (KNN), Support Vector Machine (SVM),
Probabilistic Neural Network (PNN), and Decision Tree (DT) classifiers for distinguishing normal versus MI classes. This
method yielded the highest accuracy 97.96%, sensitivity 98.89%, and specificity 93.80% using the SVM classifier.

Keywords Myocardial infarction · Computer aided diagnostic system · Electrocardiogram · Pan Thompkins algorithm ·
Classifiers

1 Introduction
* U. Rajendra Acharya
[email protected] The heart is an essential organ composed primarily of mus-
1 cle tissue. The function of the coronary arteries is to sup-
Schiller Healthcare India Private Limited, Bangalore, India
ply oxygenated blood to the heart muscle. Thus, coronary
2
School of Engineering, Ngee Ann Polytechnic, artery disease obstructs the arteries, and reduces the blood
Singapore 599489, Singapore
supply to downstream muscle, potentially causing damage
3
Division of Cardiology, Department of Medicine, Columbia to the heart muscle and the possibility of myocardial infarc-
University, New York, USA
tion (MI) (Thomas 2015). If the extent of the MI is large,
4
National Heart Centre, Singapore, Singapore the affected poorly contracting wall segments can introduce
5
Biomedical Engineering Department, Rathinam Technical increased mechanical stress to the heart, resulting in mor-
Campus, Coimbatore, India phological and conformational alterations in response, i.e.,
6
Department of Mathematics and Computer Science, Beirut left ventricular remodeling, which results in inefficient pump
Arab University, Beirut 115020, Lebanon functioning and contributes to heart failure. The MI seg-
7
Department of Bioinformatics and Medical Engineering, ments can undergo substrate alterations, e.g. fibrosis, that
Asia University, Taichung, Taiwan renders the tissue arrhythmogenic, which can be lethal.
8
International Research Organization for Advanced Science Symptoms of acute MI involve breaking into cold sweats,
and Technology (IROAST), Kumamoto University, chest pain, shortness of breath, feeling faint, nausea, as well
Kumamoto, Japan

13
Vol.:(0123456789)
3228 C. Sridhar et al.

as discomfort that may radiate to the arm, shoulder, and analysis speed. Student’s t-test was utilized to obtain sig-
neck (Setiawan et al. 2014). Major risk factors associated nificant features, which were ranked based on their t-values
with MI are high cholesterol, diabetes, high blood pressure, and then input to the classifier. The classifier was designed
physical inactivity, obesity, and unhealthy diets. These risk to distinguish between normal and MI signals.
factors promote atherosclerotic plaque formation in the coro-
nary artery wall that results in arterial narrowing. As fat
content in the plaque increases, it becomes vulnerable and 3 Computer aided diagnostic systems (CAD)
susceptible to surface rupture. The latter triggers a second-
ary pathophysiological phenomenon that activates blood clot ACAD system comprises four main units: preprocessing of
formation within the coronary artery lumen to interrupt the signals, nonlinear feature extraction, selection of the most
distal blood supply suddenly and completely, resulting in significant features, and classification of these selected fea-
MI (Roger 2007). tures into normal versus MI. Figure 1 illustrates a proposed
In 2017, 800,000 Americans died from MI (WHO fact CAD system.
2012; Ley 2015). Of these, 280,000 had prior MI and the
remainder were first-time presentations (Ley 2015) under- 3.1 Raw ECG signals
scoring the need for accurate diagnostics. There are sev-
eral diagnostic instruments utilized to characterize MI. ECG signals were obtained from PhysioBank, the Physi-
The diagnostic tests include the exercise stress test (EST), kalisch-Techische Bundesanstalt (PTB) Diagnostic ECG
cardiac catheterization, and electrocardiogram (ECG). Car- database (Goldberger et al. 2000). This database contains
diac catheterization is an invasive procedure that requires publicly available digitized ECG data from patients with
expertise and training. In addition, patients are exposed to different heart diseases, for training of models. The signals
procedural risk, albeit small, as well as radiation and poten- were recorded using a PTB sample recorder. The database
tially nephrotoxic contrast. With EST, ECGs are recorded contained 52 healthy and 148 subjects with acute and/or
during treadmill exercise, which can be associated with a chronic MI and the entire dataset was used in this study. The
very small risk of cardiac arrest (Robert et al. 2013). Hence, normal and MI signals were sampled at a rate of 1000 Hz. In
not all MI patients can undergo this test due to the associ- this work, only Lead II ECG signals were considered.
ated risk. Thus, the standard ECG remains the most com-
mon diagnostic tool for detection of MI, especially in the 3.2 Pre‑processing
acute setting. As the ECG abnormalities associated with
MI, both acute and chronic, can be subtle, experts’ read- At this phase, the discrete wavelet transform (DWT) was
ings are crucial to ensure the accuracy of interpretation. utilized for noise removal. DWT was employed to decom-
To overcome these drawbacks, computer aided diagnostic pose the signals, using the Daubechies wavelet 6 (db6). The
systems (CAD) should be designed to extract the pertinent approximate and detail coefficients with high and low pass
parameters. These parameters can then be input to classifiers filters were obtained to eliminate noise including baseline
for the categorization of normal versus MI patients (Robert wander (0–0.5 Hz) and power-line interference (50–150 Hz)
et al. 2013; Liu et al. 2014). (Acharya et al. 2016a, b).
CAD tools are developed to detect a disorder and to mini- After DWT, empirical mode decomposition (EMD) (Pal
mize intra- and inter-observer variability (Jahmunah et al. and Mitra 2012) was then employed. Various intrinsic mode
2019a, b). In general, CAD systems can be classified into functions (IMFs) were obtained by decomposing the EMD
online- and offline-based systems. Table 1 shows the sum- signals, and low frequency components were removed using
marized works for MI detection using a CAD system. a notch filter. The denoised signals were reconstructed once
the low frequency components were removed.
DWT is dependent on a prior choice of wavelet basis
2 Literature review and taps on experts’ experience in determining the level of
decomposition to extract the signals (Labate et al. 2013).
Herein, a novel algorithm involving an offline processing EMD on the other hand, does not require the level of decom-
system for diagnosing MI with the ECG is described and position to be set beforehand and hence, overcomes the
discussed. Acquired ECG signals were subjected to CAD limitations of DWT (Labate et al. 2013). After the baseline
after pre-processing to remove noise. A nonlinear feature wander and power-line interference were removed, the Pan-
extraction (Paul et al. 2019) operation was implemented to Tompkins algorithm was employed for R-peak detection
extract suitable features from the signals. These features (Pan and Tompkins 1985). Once the R point was detected, a
were then subjected to a feature selection process, wherein two-second signal segmentation was done, considering 1749
the system selected the best features which would improve samples to the left of R peak and 250 samples to the right of

13
Accurate detection of myocardial infarction using non linear features with ECG signals 3229

Table 1  A summary of studies using CAD systems for the classification of MI using ECG signals
Author (year published) Methods used Participant information Classification results

Bhaskar (2015) Artificial neural network N0: 52 (82 records) SVM classifier
SVM classifier N1: 148 (367 records) Acy: 91.1%
Pan Tompkins algorithm Neural network
Wavelet transformation Acy: 82.1%
Principal component analysis
Sharma et al. (2015) Multiscale energy and eigen space N0: 52 SVM classifier with RBF kernel
Wavelet decomposition N1: 148 Acy: 96%,
SVM with radial basis function Spe: 99%,
Sen: 93%
Acharya et al. (2016a, b) Discrete wavelet transform N0: 52 (125 652 beats) Acy: 98.80%,
12 nonlinear features N1: 148 (485 753 beats) Spe: 96.27%,
T-test Sen: 99.45%
k-NN classifier
Ten-fold cross-validation
Seenivasagam and Chitra (2016) Feed forward neural network N0 + N1: 770 records Acy: 89.61%,
Particle swarm optimised neural net- Spe: 88.98%,
work Sen: 90.13%
Classifiers
Dohare et al. (2018) Clinical features N0: 60 Acy: 96.66%,
peak to peak amplitude, area, standard N1: 60 Spe: 96.66%,
deviation, skewness, kurtosis and Sen: 96.66%
mean, are determined
SVM classifier
PCA
Kora (2017) Hybrid firefly N0: 18 (1500 normal) Acy: 96.7%,
PSO N1: 26 (1306 beats) Spe: 95.89%,
ANN structure Sen: 94.45%
KNN, SVM, LMNN classifiers
Padhy and Dandapat (2017) 2-D multi lead ECG data matrix N0: 52 Acy: 95.3%,
3rd order tensor structure N1: 148 Spe: 96.0%,
Intra-beat, inter-beat, inter-lead relation- Sen: 94.6%
ship of wavelet transformed MECG
tensor
SVM classifier
Kumar et al. (2017) Flexible analytic wavelet transform N0: 52 (10 546 beats) Least-squares support vector machine:
Sample entropy N1: 148 (40 182 beats) classifier
least-squares support vector machine Acy: 99.31%
classifier
Sadhukhan et al. (2018) Harmonic phase distribution pattern N0: 52 (79 records) Acy: 95.6%,
Threshold-based, logistic regression N1: 148 (368 records) Spe: 92.7%,
classifiers Sen: 96.5%
2 discriminative features
Sharma et al. (2018) Biorthogonal filter bank N0: 52 Noisy dataset:
Decomposition of signals N1: 148 Acy: 99.62%
Fuzzy entropy, signal-fractal dimen- Clean dataset:
sions, Rényi entropy Acy: 99.74%
k-NN classifier
Ten-fold validation
Bharadwaj et al. (2018) ECG sensor – Acy: 91.89%
Wavelet decomposition
ST segment detection
Han and Shi (2019) Energy entropy + morphological features N0: 52 SVM with RBF kernel:
Maximal overlap discrete wavelet packet N1: 148 Acy: 93.5%,
transform Spe: 92.8%,
Area, kurtosis coefficient, skewness Sen: 93.7%
coefficient, standard deviation
SVM with radial basis kernel
Ten-fold validation

13
3230 C. Sridhar et al.

Table 1  (continued)
Author (year published) Methods used Participant information Classification results

Zhang et al. (2019) Gramian angular field N0: 52 Acy: 99.49%,

Principal component analysis N1: 148
fivefold cross validation
Lin et al. (2020) Maximal overlap discrete wavelet – Acy: 99.57%,
transform Spe: 98.79%,
Variance, interquartile range, Pearson Sen: 99.82%
correlation coefficient, Hoeffding’s
D correlation coefficient, Shannon
entropy
k-nearest neighbour
Type of technique: deep learning (normal vs MI)
Acharya et al (2017a, b, c, d, e) Convolutional neural network N0: 52 (10 546 beats) With noise:
Ten-fold validation N1: 148 (40,182 beats) Acy: 93.5%,
Training, testing, validation phases Spe: 92.8%,
Sen: 93.7%
Without noise:
Acy: 95.2%,
Spe: 94.2%,
Sen: 95.5%
Reasat and Shahnaz (2018) Shallow convolutional neural network N0: 52 Acy: 84.54%,
coupled with inception network N1: 148 Spe: 84.09%,
Geometric separability index Sen: 85.33%
Compared against benchmark database
Liu et al. (2018a, b) Multiple-feature-branch Convolutional N0: 52 Class-based detection:
Neural Network N1: 148 Acy: 99.95%
Patient-specific paradigm Patient-specific experiment:
Class-based, patient-specific experiment Acy: 98.79%
Liu et al. (2018a, b) Multi-lead Convolutional Neural N0: 52 (80 records) Acy: 96%,
Network N1: 148 (368 records) Spe: 97.37%,
Five-fold validation Sen: 95.40%
Lui and Chow (2018) Convolutional Neural Network N0: 52 (80 records) Spe: 97.7%,
Recurrent neural network N1: 148 (368 records) Sen: 92.4%,
Multi-layer perceptron classifier F1 score: 94.6%
Ten-fold validation
Zhang et al. (2019) Autoencoder N0: 52(80 records) Acy: 99.9%,
Tree bag classifier N1: 148 (368 records) Spe: 99.52%,
Ten-fold validation Sen: 99.98%
Tripathy et al. (2019) Deep neural network N0: 52 (72 records) Acy: 99.74%%,
Fourier–Bessel series expansion-based N1: 100 (100 records) Spe: 99.60%,
empirical wavelet transform Sen: 99.87%
Five-fold validation
Statistical features
Alghamdi et al. (2020) Convolutional Neural Network N0: 143 VGG-M12:
VGG-M11, VGG-M12 networks N1: 147 Acy: 99.22%,
Transfer learning method Spe: 99.49%,
Sen: 99.15%
Baloglu et al. (2019) Deep convolutional neural network N0: 52 Acy: 99.78%
10 layers N1: 148
Lodhi et al. (2018) Convolutional Neural Network N0: 52 (31 722 beats) Acy: > 90%
Ten-fold validation N1: 100 (49 930 beats) Sen: > 94%
Spe: > 86%
Strodthoff and Strodthoff (2019) Convolutional Neural Network N0: 52 Spe: 89.7%,
Ten-fold validation N1: 127 Sen: 93.3%
Comparison with long short-term
memory model
Feng et al. (2019) Convolutional Neural Network N0: 52 Acy: 95.4%,
Long short-term memory network N1: 148 Spe: 86.5%,
Ten-fold validation Sen: 98.2%

13
Accurate detection of myocardial infarction using non linear features with ECG signals 3231

Table 1  (continued)
Author (year published) Methods used Participant information Classification results

Venu et al. (2019) Convolutional Neural Network N0: 450 signals Acy: 87%
5 layers N1: 720 signals
Haddadi et al. (2019) Discrete wavelet transform N0: 50 Acy: 94.83%,
Convolutional neural network N1: 50 Spe: 94.93%,
Sen: 94.75%
Han and Shi (2020) Multi-lead residual neural network N0: 52 Intra-patient scheme:
Intra-patient, inter-patient strategies N1: 148 Acy: 99.92%
Five-fold validation
Liu et al (2020) Multiple-feature-branch convolutional N0: 52 Class-based method:
bi-directional recurrent neural network N1: 148 Acy: 99.90%
Lead random mask optimisation Subject-based method:
Class-based, subject-based five-fold Acy: 93.08%
validation
Fu et al. (2020) Multi-lead mechanism coupled with N0: 52 (127 188 beats) Intra-patient performance:
convolutional neural network N1: 148 (632 940 beats) Acy: 99.93%,
and bidirectional gated recurrent unit Spe: 99.63%,
Temporal features, spatial features Sen: 99.99%
Inter-patient performance:
Acy: 96.50%,
Spe: 93.34%,
Sen: 97.10%
Ribeiro et al. (2020) Deep neural network N0 + N1 = 2,322,513 F1 score: > 80%
12 leads ECG recordings
Ramesh et al. (2020) Morphological, temporal, statistical 47 subjects Acy: 98%(arrhythmia classification)
features
Convolutional Neural Network
Huang et al. (2020) Fast compression residual convolutional 47 subjects Acy: 98.79%(arrhythmia classification)
neural network
Maximal overlap wavelet packet trans-
form
Type of classification: normal vs MI/CAD/CHF
Acharya et al. (2017a, b, c, d, e) Discrete cosine transform N0: 52 Discrete wavelet transform with k-NN
(3-Class) Discrete wavelet transform N1: 148 classifier:
Empirical mode decomposition, IMFs N2: 7 Acy: 98.5%,
Locality preserving projection Spe: 98.5%,
k-nearest neighbour classifier Sen: 99.7%
Acharya et al. (2017a, b, c, d, e) Contourlet, shearlet transformations N0: 52 Contourlet transform:
(4-class) Continuous wavelet transform N1: 148 Acy: 99.5%,
Entropies, 1st and 2nd order statistical N2: 7 Spe: 99.2%,
features N3: 15 Sen: 99.9%
Improved binary particle swarm optimi-
sation for feature selection
Analysis of variance, relief methods
Decision tree, k-nearest neighbour clas-
sifiers
Ten-fold validation
Present study Pan Tompkins algorithm N0: 52 SVM classifier
Nonlinear features N1: 148 Acy: 97.96%,
T-test Spe: 93.80%,
SVM classifier Sen: 98.89%

Type of technique: machine learning (normal vs MI)

N0 normal subjects, N1 MI patients, N2 coronary artery disease (CAD) patients, N3 congestive heart failure (CHF) patients

R peak at a 1 kilo Hertz sampling rate. Thus, a total of 2000 Two-second segments were extracted, compris-
sample points were selected for each ECG beat analysis. ing 16826 normal 3796 MI segments, respectively. Fig-
ure 2 shows the 2-s ECG signals in (a) normal and (b) MI.

13
3232 C. Sridhar et al.

Fig. 1  Proposed CAD system

Fig. 2  Two seconds duration of ECG Signals: a normal and b MI

3.3 Feature extraction 3.3.1 Bispectrum

Naturally, ECG signals are disorderly and vary over time. The Higher order spectrum (HOS) is a used to evaluate
These characteristics cause the extraction of nonlinear non-stationary and non-Gaussian signals (Acharya et al.
features to be used preferably in the development of CAD 2015, 2017a; b, c, d, e; Pham et al. 2020a, b). It detects
systems for MI diagnosis. Time, frequency, and time–fre- the diversion from phase correlations and Gaussian level
quency domain analyses are unable to detect the inher- among frequency components of the signal (Jahmunah
ent variations in the ECG, but nonlinear techniques can et al. 2019a, b; Acharya et al. 2015). HOS retains phase
be used to extract characteristic features. Many studies information as it is more immune to noise.
have reported using various nonlinear methods (Jahmu-
nah et al. 2019a). The computational demands and time
increase when nonlinear methods are utilized to obtain 3.3.2 Recurrence qualitative analysis (RQA)
features from DWT coefficients, but it facilitates the
extraction of large numbers of suitable features (Acha- The RQA computes the total number of recurrences in
rya et al. 2017a; b, c, d, e). In our study, we extracted order to calculate ECG signal complexity (Webber and
the following nonlinear features: Bispectrum, Recur- Zbilut 1994). The RQA includes parameters of transitiv-
rence Qualitative Analysis (RQA), Approximate Entropy, ity, determinism, laminarity, mean diagonal line length,
Permutation Entropy, Detrended Fluctuation Analysis, the entropy of diagonal length, trapping time, recurrence
Fractal Dimension, Largest Lyapunov Exponent, Sample time, recurrence rate, and recurrence time entropy (Zbilut
Entropy, Rényi Entropy, Hurst Exponent, Tsallis Entropy, and Webber 1992; Zbilut et al. 2002).
Kolmogorov Sinai Entropy, Fuzzy Entropy, Modified
Multiscale Entropy (MMSE), Wavelet Entropy, and cor-
relation Dimension.

13
Accurate detection of myocardial infarction using non linear features with ECG signals 3233

3.3.3 Approximate entropy (ApEn) signal is less predictable; if SampEn is of low value, then it
implies that the signal is more predictable.
This entropy helps to characterize the instability or irregu-
larity (Pincus 1991) in the ECG signal. The approximate
3.3.10 Tsallis entropy (TEnt)
entropy is higher for more irregularity.
It is used to calculate the differences and memory con-
3.3.4 Fractal dimension (FD) sequences in the signal (Acharya et al. 2018a, b). Tsallis
coefficients help in characterizing ECG bursts, spikes, and
It is a powerful tool used to measure the complexity of continuous rhythms.
fractals by changing the measurement scale. The fractal
patterns are characterized by quantifying their complexity 3.3.11 Fuzzy entropy (FEnt)
as the ratio of change in the detail to the change in scale
(Acharya et al. 2011). Fuzzy entropy is used to estimate ECG signal unpredict-
ability (Acharya et al. 2018a, b). This entropy indicates
the degree of randomness, and is calculated as the entropy
3.3.5 Permutation entropy (PEnt) of a fuzzy set whose elements have varying degrees of
membership.
Permutation entropy is a measure of random time series
data depending upon the examination of the permutation
pattern. It measures the complexity in ECG signals by cal- 3.3.12 Kolmogorov‑Sinai entropy (K‑SEnt)
culating the coupling between time series data (Bandt and
Pompe 2002). Kolmogorov Sinai entropy is used to calculate the uncer-
tainty in the ECG over time (Acharya et al. 2019; Ziv and
Lempel 1977).
3.3.6 Detrended fluctuation analysis (DFA)
3.3.13 Multivariate multi‑scale entropy (MMSEn)
To obtain the self-similarity characteristics of ECG signals,
DFA is used (Peng et al. 1996). The variants such as multi In MMSEn, nonstationary ECG signals are considered to
fractal DFA and root mean square (RMS) are implemented calculate the intrinsic correlation and express the degree of
along with DFA features (Jahmunah et al. 2019a, b). correlation in the time series (Acharya et al. 2018b, a; Hu
and Liang 2012).
3.3.7 Hurst exponent (HE)
3.3.14 Rényi entropy (RE)
The Hurst Exponent (Hurst 1956) is an evaluation of self-
similarity and predictability in the ECG. If the magnitude Rényi entropy is commonly known as the generalized type
of the Hurst exponent is high, it indicates a smoother and of Shannon entropy. The sudden variations in the time series
simpler signal (Acharya et al. 2019). data can be elucidated through RE (Renyi 1961; Shannon
1948).

3.3.8 Largest Lyapunov exponent (LLE)

3.3.15 Wavelet entropy (WE)
The LLE is evaluated to assess the degree of chaos existing
Wavelet entropy is used to measure the level of disorder as
in the signals (Rosenstein et al. 1993). If the value of LLE
well as to compute the energy levels in the various frequency
is higher, then the signals are more complex.
bands of the ECG signals (Rosso et al. 2001).

3.3.9 Sample entropy (SampEn) 3.3.16 Correlation dimension (D2)

This technique calculates the complexity and regularity of Correlation dimension calculates self-similarity in the ECG
a physiological time series signal, and is not constrained by signals (Renyi 1961). The correlation integral C(r) is calcu-
the pattern length (Richman and Randall 2000; Song and Liò lated first, and then the gap between N pairs of data points
2010). If the value of SampEn is high, then the time series are measured (Jahmunah et al. 2019a, b).

13
3234 C. Sridhar et al.

3.4 Feature selection Figure 3 presents the 45 nonlinear parameters that were

extracted from the ECG signals of MI patients and normal
Feature selection determines the best features and inputs subjects using lead II.
them to the classification algorithm, to increase the speed
of the CAD system. In the current work, the t test is used
to select highly significant features. The t-distribution is 3.5 Classification
defined as a continuous probability distribution for calcu-
lating population mean when sample size is small and the Classification of data is broadly grouped into two types:
variance of the population is unknown. The t distribution parametric and nonparametric classifiers. The parametric
is similar to the normal distribution. It is bell-shaped and type classifies the data based upon the statistical distri-
symmetrical about the mean (Student Biometrika 1908). bution of the class, whereas nonparametric classifiers
Features with a p-value lower than threshold are consid- rely upon conditions of the unknown density function.
ered to be significant (Acharya et al. 2013). Here, the In this work, K-Nearest Neighbor and Decision Tree are
significant level corresponds to the probability of reject- implemented, and they are categorized as nonparametric
ing the null hypothesis, given that the condition is true. classifiers.

Fig. 3  a and b indicate 45 non-linear parameters extracted from MI and normal ECG signals using lead II

13
Accurate detection of myocardial infarction using non linear features with ECG signals 3235

3.5.1 K‑Nearest Neighbor (KNN) seconds. A feature is statistically significant if p < 0.05; the
t value indicates ranking of features (Hagiwara et al. 2018;
KNN is a nonparametric classification method. KNN clas- Bishop 2006) If the t-value is higher, it represents features
sifies the test sample by estimating the distance between that are more significant. Hence, there is an inverse relation-
the training and testing set. The class of the test sample is ship between t-value and p-value.
decided based on the nearest samples present in the train- Table 3 shows the best performance of KNN, DT, SVM,
ing set. The distance is estimated using Euclidean distance and PNN classifiers used in our study. The highest accu-
(Acharya et al. 2013). racy, sensitivity, and specificity values of 97.96%, 98.89%,
93.81% were achieved respectively, with the SVM classifier.
3.5.2 Decision Tree (DT)

DT is a graph in which a branching technique is used to 5 Discussion

decide possible outcomes. The decision tree splits the com-
plex decisions into simpler decisions in order to extract the In this work, a nonlinear feature extraction method was
rules for the recognition of the class of data (Peng et al. investigated for the detection of normal versus MI from lead
1996). The output of the decision tree is a binary structure. II ECG signals. This technique was used because it detects
Here the input and outputs are deliberated as the root and intrinsic complexity existing in the time series (Hurst 1956).
terminal/leaf nodes respectively. Table 1 is a comparison of the results obtained from different
CAD systems using a similar database.
3.5.3 Support vector machine (SVM) From Table 1, it is apparent that various methods were
proposed to develop CAD systems. The nonlinear analysis
This is a supervised learning tool which simultaneously min- and time–frequency domain approaches were commonly
imizes the empirical classification error and maximizes the employed (Hariharan et al. 2012). Hence, CAD systems
geometric margin. SVM classifiers have two types of data: along with a feature extraction technique yielded enhanced
separable and non-separable data. For non-separable data, diagnostic accuracy as compared to single-type feature
kernels such as radial basis aid in the classification process extraction techniques (Hagiwara et al. 2018). Bhaskar
(Liu et al. 2013). (2015), Dohare et al. (2018), Kora (2017), Padhy and Dan-
dapat (2017), Sadhukhan et al. (2018), Bharadwaj et al.
3.5.4 Probabilistic Neural Network (PNN) (2018), Han and Shi (2019), Seenivasagam and Chitra
(2016) achieved lower classification accuracies as com-
PNN is a supervised learning algorithm. It comprises an pared to our study. Although Acharya et al. (2016a, b),
input layer, representing the feature vector. This input layer Kumar et al. (2017), Sharma et al. (2018), Lin et al. (2020)
is totally interconnected to the concealed layer, wherein achieved higher classification accuracies, the methods they
these layers connect to the output representing the choices in had proposed are computationally intensive as compared
which the input data can be classified (Acharya et al. 2019). to the Pan-Thompkins algorithm, which is computationally
The three classifiers are trained and tested using ten-fold more efficient. Although Liu et al. (2018a, b), Zhang et al.
cross-validation strategy. (2019), Tripathy et al. (2019), Alghamdi et al. (2020), Han
and Shi (2020), Fu et al. (2020), Ramesh et al. (2020),
Huang et al. (2020) achieved relatively higher classifi-
4 Results cation accuracies than our study, these authors explored
deep learning techniques, which require a lot of time to
20,622 beats were segmented in total, from 52 normal and train the model as compared to our technique. Also, these
148 MI subjects. The two second segmentation was done in authors have used small data sizes to validate their models;
order to find the R-peak. Performance of the classifier was hence robustness is not guaranteed. Deep learning models
estimated based on the ranked features and was assessed by should be validated with a larger data. Ribeiro et al. (2020)
calculating the average accuracy, Positive Predictive Value had employed a deep neural network with larger data, but
(PPV), sensitivity, and specificity (Acharya et al. 2011). still obtained a lower accuracy than our study. Acharya
Table 2 provides a summary of the diagnostic performance et al. (2017a, b, c, d, e) also achieved higher classification
for detection of MI versus normal using recurrence time, accuracies, but the authors investigated a different clas-
HOS, entropy, bispectrum, DFA, LLE, Hurst exponent, frac- sification from our study. They did a 3-class classifica-
tal dimension, permutation entropy, MMSE, Tsallis entropy, tion and 4-class classification study, respectively. Venu
wavelet entropy, sample entropy, and approximate entropy et al. (2019), Haddadi et al. (2019) had employed deep
features extracted from ECG signals with a duration of two learning methods but still achieved lower accuracies than

13
3236 C. Sridhar et al.

Table 2  Range (Mean ± standard deviation) of features ranked using the t test
Parameters Normal MI

Mean Standard deviation Mean Standard deviation p Value t Value

T2 107.243 28.450 66.425 23.744 0 92.052
<L> 13.517 5.151771 7.298457 3.397544 0 91.510
TT 18.783 7.33713 9.897124 4.942379 0 90.529
ENTR 0.367 0.129463 0.199021 0.105068 0 85.052
Vmax 205.136 99.09857 93.54796 70.68333 0 80.960
Fuzzy Entropy 0.523 0.175765 0.785387 0.206109 0 72.608
Fractal Dimension 1.080 0.042248 1.145388 0.056665 0 66.738
Trans 0.789 0.023603 0.770115 0.01662 0 60.347
Lmax 806.736 231.6037 525.355 267.7788 0 59.884
MMSE2 0.133 0.105154 0.276806 0.160778 0 52.509
MMSE3 0.180 0.140709 0.366986 0.212359 0 51.679
MMSE4 0.216 0.162438 0.431577 0.246438 0 51.275
MMSE5 0.246 0.173875 0.475803 0.265502 0 50.867
MMSE6 0.271 0.179371 0.502349 0.271802 0 50.119
Approximate Entropy 0.455 0.167186 0.667381 0.255679 0 48.886
MMSE1 0.076 0.060224 0.153577 0.096167 0 47.805
Permutation Entropy 1.137 0.050133 1.182912 0.055447 0 47.311
MAmp 8.57E+09 1.49E+10 2.62E+09 4.72E+09 0 43.017
H5 7.22E+12 1.24E+13 2.29E+12 4.16E+12 0 42.147
Wavelet Entropy − 2.5E+09 2.76E+09 − 1.2E+09 1.71E+09 1.1E–291 37.098
H1 1719.991 145.041 1635.583 131.6278 7.8E–261 35.005
Sample Entropy 0.381 0.082902 0.444815 0.110283 3.2E–241 33.615
Tsallis Entropy − 2.1E+08 1.96E+08 − 1.2E+08 1.39E+08 8.2E–223 32.261
H2 118.689 9.47417 113.7473 8.412088 3.2E–218 31.915
H4 1.14E+08 26,625 1.01E+08 25,219 3.8E–192 29.884
H3 979.602 86.752 939.748 79.654 3.7E–162 27.380
Awcoby 2.095 0.052 2.149 0.117 3.2E–159 27.124
T1 7.304 2.029 8.381 2.523 2.9E–131 24.555
K-S Entropy 0.414 0.108 0.363 0.121 2.6E–125 23.976
D2 0.657 0.236 0.791 0.325 1E–124 23.916
DET 0.974 0.032 0.948 0.066 4E–123 23.759
RR 0.147 0.038 0.131 0.042 6.59E–94 20.663
LAM 0.987 0.015 0.975 0.038 3.4E–78 18.800
Largest Lyapunov Exponent 10.805 1.311 11.228 1.347 2.1E–68 17.543
RTE 0.164 0.035 0.148 0.053 1.64E–67 17.424
Ent2 0.036 0.023 0.047 0.041 2.79E–59 16.288
Ent3 0.009 0.011 0.016 0.027 2.25E–54 15.573
Ent1 0.244 0.056 0.265 0.079 6.78E–52 15.199
DFA 0.598 0.155 0.642 0.186 6.91E–43 13.759
Hurst’s Exponent 0.773 0.063 0.788 0.067 3.96E–38 12.936
Awcobx 3.028 0.299 3.136 0.557 1.31E–30 11.519
RényiEntropy − 20.322 0.842 − 20.179 0.901 5.14E–19 8.918
EntPh 2.669 0.360 2.647 0.439 0.004 2.901
Wcoby 2.051 0.054 2.067 0.415 0.022 2.298
Wcobx 2.919 0.335 2.965 2.346 0.234 1.189

13
Accurate detection of myocardial infarction using non linear features with ECG signals 3237

Table 3  Best results of KNN, Classifier N TrPo TrNe FaPo FaNe Acy Sen Spe Popv
DT, SVM, PNN classifiers
kNN 42 16,554 3522 274 272 0.973523 0.983835 0.927819 0.983718
DT 39 16,381 3316 480 445 0.955145 0.973553 0.873551 0.971532
SVM 42 16,640 3561 235 186 0.979585 0.988946 0.938093 0.986074
PNN 43 16,413 2237 1559 413 0.904374 0.975455 0.589305 0.913254

The best results are shown in bold

N number of features, TrPo true positive, TrNe true negative, FaPo false positive, FaNe false negative,
Acy accuracy, Popv positive predictive value, Sen sensitivity, Spe specificity

Fig. 4  Recurrence plots for a normal and b MI

our study. Hence, our method is technically sound as we 5.1 Benefits

have achieved a relatively high classification accuracy with
the usage of a less computationally intensive algorithm. 1. The recommended system enables rapid diagnosis of
Therefore, a faster diagnosis of MI is possible. MI.
It should be noted that recurrence time yielded more 2. The system has been validated by ten-fold cross-valida-
statistically significant features. The means for normal tion; hence it is reliable.
class features are greater than for the MI class. Figures 4, 3. The system could be used to find anomalies associated
5, 6 indicate recurrence, bispectrum and cumulant plots, with MI in real-time analysis.
respectively. These plots are different for normal versus 4. The top few features exhibit a high degree of separation
MI classes. Figure 4 shows recurrence plots for normal and hence, can be used to develop the index (Pham et al.
and MI. In this, the recurrence parameter indicates the 2020a, b)
degree of feature repetition in the ECG signals. In the plot,
blue dots indicate recurrence at a specific time interval.
Figure 5 represents bispectrum and contour plots. The 5.2 Limitations
cumulant and contour plots of normal and MI signals are
presented in Fig. 6. It is evident that the recurrence plots 1. Only a small data size can be used for this technique.
are distinct in the normal and MI classes. Also, the pat- 2. The manual extraction and selection of features can be
terns in the bispectrum and cumulant plots are unique and a laborious task.
show good separation between the two classes. Hence,
the non-linear features used in our study are highly dis- A CAD system is advantageous due to its versatility,
criminatory. There are some benefits and limitations of rapidness, reliability, and the fact that it can be used to
our study, following. find anomalies associated with MI in real-time analysis.

13
3238 C. Sridhar et al.

Fig. 5  Bispectrum plots for a

normal and b MI

13
Accurate detection of myocardial infarction using non linear features with ECG signals 3239

Fig. 6  Cumulant plots for a

normal and b MI

13
3240 C. Sridhar et al.

5.3 Challenges in myocardial infarction prediction as they can be implemented even under noisy conditions
(Ansari et al. 2017; Acharya et al. 2017a, b, c, d, e). Addi-
The features of ECG, including its peak deflections, time tionally, more data should be added to test our recommended
domain amplitude, and duration, provide information regard- model to improve performance.
ing myocardial characteristics. The time domain features are
not able to provide exact discrimination between normal
and MI beats (Martis et al. 2014). The wavelet domain can 6.1 ECG in mobile healthcare and telecare
be used to distinguish two dissimilar ECG signals with the
same magnitude due to an increase in time resolution and Apart from modifications done for ECG algorithmic analy-
a compromise in the frequency resolution. The challenge lies sis, there is additional scope for mobile healthcare technol-
in selecting an optimal wavelet basis function. ogy, wherein ECG signals are recorded with mobile devices
Assuming that noise and ECG signals are seen in separate (Constantinescu and Jinman 2012). Patients are connected
frequency components, linear models can be used. However to Holter-like devices for ECG signal recording. Any emer-
when ECG signals and noise overlap the same frequency gency in patient health can be readily reported to the critical
spectrum, random processes cannot be described using lin- care unit. Adeli and Sankari (2011) invented a cost-effective
ear models (Chua et al. 2010). Nonlinear methods have some mobile medical device for the purpose of real-time monitor-
limitations including: (i) these techniques do not follow both ing of cardiac health. A future challenge is to develop an
principles of superposition and homogeneity, (ii) they are efficient remote monitoring device to rapidly provide infor-
computationally rigorous, (iii) linear shift invariance is not mation about patient health and suitable treatments, using
valid for a nonlinear system, and (iv) reflection and sym- a decision support system (DSS) (Tamura 2012). There is
metry properties are not followed. a requirement for DSS in cardiac telecare, and a necessity
for an ECG monitoring unit via wearable electrodes. This
can help to capture data via a web-based unit, and can be
6 Future work employed for instantaneous screening of ECG signals by
doctors. The system should be cost-effective, low in power
The amplitude peaks provide information concerning car- consumption, and should be affordable to the patient (Ven-
diac characteristics. However, minute alterations in peak katesan et al. 2018; Pandey et al. 2012) Fig. 7 presents the
deflection morphology and position cannot be identified ECG recognition system in a clinical setting.
clearly by visual inspection. In signal processing terms, the In this system, mobile devices and sensors are connected
time domain features are not able to provide discriminative wirelessly through cloud computing devices. Without pro-
information for normal versus abnormal beats (Lih et al. cessing ECG data, signals are sent to mobile devices from
2020; Clifford et al. 2006). Herein, we have listed differ- different sensing devices. As the data is not processed and
ent non-linear features used by different authors for testing is generated from different devices with continuous moni-
(Kannathal et al. 2006). The linear method provides satisfac- toring, it will require more space. To store large volumes
tory classification accuracy. Hence, for future work, testing of data, cloud storage space is utilized. The unknown data
should be done on noise-free data using nonlinear methods, is sent to the developed model at the cloud for diagnosis.

Fig. 7  Proposed ECG recognition system in a clinical setting

13
Accurate detection of myocardial infarction using non linear features with ECG signals 3241

The diagnostic results can be sent to healthcare centers and is a common type of RNN. The LSTM consists of three
physicians (Acharya et al. 2018a, b). gates to incorporate memory block; they are: input, output,
The advent of cloud computing has enabled host software and forget gate. These gates assist in adding and removal of
packs to analyze ECG data, including Software-as-a-Service information from the network, based on cell state (Hopfield
(SaaS), Platform-as-a-Service (PaaS), and Infrastructure- 1987).
as-a-Service (IaaS). The SaaS layer is used for organizing For future work, we intend to employ a deep learning
custom-designed analysis of current and historic ECG data. model and use a larger dataset to train our model. We hope
The PaaS layers supervise the execution of software into to integrate this developed model to the cloud system such
three major units: (i) Workflow Engine, (ii) Container scal- that information about patient health from the acquired ECG
ing manager, and (iii) Aneka. The IaaS layers are self-ser- signals could be sent to the clinicians to aid them in their
vice models for monitoring, accessing, and managing remote diagnostic and treatment decisions.
datacenter infrastructures, such as compute, storage, and net-
working services (LeCun et al. 2015).
Deep learning is the part of machine learning methods in
which hidden layers of neurons are used to construct inher- 7 Conclusion
ent features. For big datasets, deep learning methods can
be used as it performs better than machine classification MI is an irreversible damage to the myocardium due to coro-
and classic analysis methods. The word “deep” is obtained nary artery blockage. The expansion of MI can be rapid if
from various hidden layers in an artificial neural network it is not treated expediently. If left untreated, further dam-
(ANN) (Fukushima and Miyake 1982). The ANN structure age can occur to myocardial function and the structure of
comprises the input, hidden, and output layers. Connec- left ventricle. The recommended approach to distinguish
tion link is used to connect every nerve cell of one layer between normal and MI signals involves a non-invasive
to every nerve cell of another layer. These computational tool. In this work, nonlinear parameters were extracted from
nerve cells are composed of dendrites (input), axon (output), lead II ECG signals. The performance of each classifier was
nucleus (activation function), node (soma), and synapses measured using accuracy, sensitivity, and specificity. The
(weights) (Goodfellow et al. 2016). Here, the input signals highest accuracy of 97.96% was achieved with the SVM
are dendrites, weight models are synapses, and the activa- classifier using nonlinear features with a ten-fold cross-
tion function is the nucleus in the biological neuron. The validation strategy. We have also proposed unique recur-
ANN structure adversely affects the performance of clas- rence and HOS plots for normal and MI ECG classes to
sification due to susceptibility to translation and shift devia- identify the two classes qualitatively. The proposed system
tion (LeCun and Bengio 1995). To overcome this drawback, outperforms other recently developed techniques, and has
a Convolutional Neural Network (CNN) can be used. The the potential to be readily deployed in hospitals and clinics,
CNN ensures shift invariance and translation. CNN is a feed- even in remote areas. CAD systems provide rapid and accu-
forward network which consists of pooling, convolution, and rate results, and can be implemented for real-time diagnosis
fully connected layers (Hinton and Salakhutdinov 2006; Lih of MI.
et al. 2020; Murat et al. 2020).
Other deep learning techniques include the autoencoder,
Funding Not applicable.
Recurrent Neural Network (RNN), and deep generative
models, which are used to evaluate physiological signals. Availability of data and material The data used in this study is available
For the autoencoder, the input dimension is the same as in the publicly available PTB ECG database.
the output, and it is an unsupervised neural network. The
autoencoder has three layers: input, output, and hidden. For Compliance with ethical standards
the autoencoder, encoding and decoding are two important
steps (Hinton et al. 2006; Hopfield 1987). Conflict of interest The authors declare that they have no conflicts of
The Deep Generative model comprises the Restricted interest.
Boltzmann Machine (RBM) and Deep Belief Network
(DBN). The RBM consists of two layers: (i) visible and
(ii) hidden layers. The DBN is the probabilistic model with
References
various hidden layers. In RBM, the first layer is trained to
reconstruct the input (Hinton et al. 2006). Acharya UR, Chua ECP, Faust O, Lim TC, Lim LFB (2011) Auto-
The RNN is a recurrent network which follows a recursive mated detection of sleep apnea from electrocardiogram sig-
approach. The Long Short-Term Memory (LSTM) network, nals using nonlinear parameters. Physiol Meas. https​: //doi.
org/10.1088/0967-3334/32/3/002
which has the capability to learn long term dependencies,

13
3242 C. Sridhar et al.

Acharya UR, Faust O, Sree SV, Ghista DN, Dua S, Joseph P et al Bharadwaj AV, Upadhyaya SM, Sharath L, Srinivasan R (2018)
(2013) An integrated diabetic index using heart rate variability Early diagnosis and automated analysis of myocardial infarc-
signal features for diagnosis of diabetes. Comput Methods Bio- tion (STEMI) by detection of ST segment elevation using
mech Biomed Eng 16:222–234. https​://doi.org/10.1080/10255​ wavelet transform and feature extraction. Int Conf Des Innov
842.2011.61694​5 3Cs Comput Commun Control ICDI3C, pp 24–28. https​://doi.
Acharya UR, Sudarshan VK, Adeli H, Santhosh J, Koh JEW, Adeli A org/10.1109/ICDI3​C.2018.00014​
(2015) Computer-aided diagnosis of depression using EEG sig- Bhaskar NA (2015) Performance analysis of support vector machine
nals. Eur Neurol. https​://doi.org/10.1159/00038​1950 and neural networks in detection of myocardial infarction. Pro-
Acharya UR et al (2016a) Automated detection and localization of cedia Comput Sci 46:20–30
myocardial infarction using electrocardiogram: a comparative Bishop CM (2006) Pattern recognition and machine learning.
study of different leads. Knowledge Based Syst 99(2016):146– Springer-Verlag, Berlin, Heidelberg
156. https​://doi.org/10.1016/j.procs​.2015.01.043 Chua KC, Chandran V, Acharya UR, Lim CM (2010) Application of
Acharya UR, Fujita H, Sudarshan VK, Lih OS, Adam M (2016b) Auto- higher order statistics/spectra in biomedical signals—a review.
mated detection and localization of myocardial infarction using Med Eng Phys 32:679–689. https​: //doi.org/10.1016/j.meden​
electrocardiogram: a comparative study of different leads. Knowl gphy.2010.04.009
Based Syst 9(8):1–39. https​://doi.org/10.1016/j.knosy​s.2016.01.040 Clifford GD, Azuaje F, McSharry P (2006) Advanced methods and
Acharya UR et al (2017a) Application of higher-order spectra for the tools for ECG data analysis. Artech House, Norwood
characterization of Coronary artery disease using electrocar- Constantinescu L, Jinman K (2012) Feng DD (2012) Spark Medi-
diogram signals. Biomed Signal Process Control. https​://doi. cal: a framework for dynamic integration of multimedia
org/10.1016/j.artme​d.2019.07.006 medical data in to distributed m-health systems. IEEE Transl
Acharya UR, Fujita H, Oh SL, Hagiwara Y, Tan JH, Adam M (2017b) Inf Technol Biomed 16:40–52. https ​ : //doi.org/10.1109/
Application of deep convolutional neural network for automated TITB.2011.21740​64
detection of myocardial infarction using ECG signals. Inf Sci Dohare AK, Kumar V, Kumar R (2018) Detection of myocardial infarc-
415–416:190–198. https​://doi.org/10.1016/j.ins.2017.06.027 tion in 12 lead ECG using support vector machine. Appl Soft
Acharya UR, Fujita H, Adam M, Lih OS, Sudarshan VK et al (2017c) Comput 64:138–147. https​://doi.org/10.1016/j.asoc.2017.12.001
Automated characterization and classification of coronary artery Feng K, Pi X, Liu H, Sun K (2019) Myocardial infarction classifica-
disease and myocardial infarction by decomposition of ECG tion based on convolutional neural network and recurrent neural
signals: a comparative study. Inf Sci 377:17–29. https​://doi. network. Appl Sci 9(9):1–12. https:​ //doi.org/10.3390/app909​ 1879​
org/10.1016/j.ins.2016.10.013 Fu L, Lu B, Nie B, Peng Z, Liu H, Xi P (2020) Hybrid network with
Acharya UR et al (2017d) Automated characterization and classifi- attention mechanism for detection and location of myocardial
cation of coronary artery disease and myocardial infarction by infarction based on 12-lead electrocardiogram signals. Sensors
decomposition of ECG signals: a comparative study. Inf Sci (Basel). https​://doi.org/10.3390/s2004​1020
377:17–29. https​://doi.org/10.1016/j.ins.2016.10.013 Fukushima K, Miyake S (1982) Neocognitron: a self-organizing neural
Acharya UR et al (2017e) Automated characterization of coronary network model for a mechanism of visual pattern recognition, In:
artery disease, myocardial infarction, and congestive heart fail- Competition and cooperation in neural nets. Springer, pp 267–
ure using contourlet and shearlet transforms of electrocardiogram 285. https​://doi.org/10.1007/BF003​44251​
signal. Knowl Based Syst 132:156–166. https:​ //doi.org/10.1016/j. Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC (2000)
knosy​s.2017.06.026 PhysioBank, Physiotool Kit, and Physionet components of a new
Acharya UR et al (2018a) Entropies for automated detection of coro- research resource for complex physiologic signals. Circulation
nary artery disease using ECG signals: a review. Biocybern 101(4):215–220
Biomed Eng. https​://doi.org/10.1016/j.bbe.2018.03.001 Goodfellow I, Bengio Y, Courville A (2016) Deep learning. MIT Press,
Acharya UR et al (2018b) Automated identification of shockable and Cambridge
non-shockable life-threatening ventricular arrhythmias using con- Haddadi R, Abdelmounim E, El Hanine M, Belaguid A (2019) A wave-
volutional neural network. Futur Gener Comput Syst. https​://doi. let-based ECG delineation and automated diagnosis of myocardial
org/10.1016/j.futur​e.2017.08.039 infarction in PTB database, 2284216
Acharya UR et al (2019) Application of nonlinear methods to discrimi- Hagiwara Y, Fujita H, Oh SL, Acharya UR et al (2018) Computer-aided
nate fractionated electrograms in paroxysmal versus persistent diagnosis of atrial fibrillation based on ECG signals: a review. Inf
atrial fibrillation. Comput Methods Programs Biomed. https:​ //doi. Sci 32(12):2–38. https​://doi.org/10.1016/j.ins.2018.07.063
org/10.1016/j.cmpb.2019.04.018 Han C, Shi L (2019) Automated interpretable detection of myocar-
Alghamdi A et al (2020) Detection of myocardial infarction based on dial infarction fusing energy entropy and morphological fea-
novel deep transfer learning methods for urban healthcare in smart tures. Comput Methods Programs Biomed 175:9–23. https​://doi.
cities. Multimed Tools Appl. https​://doi.org/10.1007/s1104​2-020- org/10.1016/j.cmpb.2019.03.012
08769​-x Han C, Shi L (2020) ML–ResNet: a novel network to detect and locate
Ansari S, Farzaneh N, Duda M, Horan K, Andersson HB, Gold- myocardial infarction using 12 leads ECG. Comput Methods Pro-
berger ZD, Nallamothu BK, Najarian K (2017) A review of grams Biomed. https​://doi.org/10.1016/j.cmpb.2019.10513​8
automated methods for detection of myocardial ischemia and Hariharan M, Vijean V, Yaacob (2012) Objective analysis of vision
infarction using electrocardiogram and electronic health records. impairments using single trial VEPs. Int Conf Biomed Eng
IEEE Rev Biomed Eng 10:264–298. https​://doi.org/10.1109/ (ICoBE). https​://doi.org/10.1109/CCDC.2013.65617​36
RBME.2017.27579​53 Hinton GE, Salakhutdinov RR (2006) Reducing the dimensionality of
Baloglu UB, Talo M, Yildirim O, Tan RS, Acharya UR (2019) Clas- data with neural networks. Science 313(5786):504–507. https​://
sification of myocardial infarction with multi-lead ECG signals doi.org/10.1126/scien​ce.11276​47
and deep CNN. Pattern Recognit Lett 122:23–30. https​://doi. Hinton GE, Osindero S, The YW (2006) A fast learning algorithm for
org/10.1016/j.patre​c.2019.02.016 deep belief nets. Neural Comput 18(7):1527–1554. https​://doi.
Bandt C, Pompe B (2002) Permutation entropy: a natural complex- org/10.1162/neco.2006.18.7.1527
ity measure for time series. J Physiol 88(16):3–17. https​://doi. Hopfield JJ (1987) Neural networks and physical systems with emer-
org/10.1103/PhysR​evLet​t.88.17410​2 gent collective computational abilities, in: spin glass theory and

13
Accurate detection of myocardial infarction using non linear features with ECG signals 3243

beyond: an introduction to the replica method and its applications. Liu W, Wang F, Huang Q, Chang S, Wang H, He J (2020) MFB-
World Sci. https​://doi.org/10.1073/pnas.79.8.2554 CBRNN: a hybrid network for MI detection using 12-Lead
Hu M, Liang H (2012) Adaptive multiscale entropy analysis of mul- ECGs. IEEE J Biomed Health Inf 24(2):503–514. https​://doi.
tivariate neural data. IEEE Trans Biomed Eng 59:12–15. https​ org/10.1109/JBHI.2019.29100​82
://doi.org/10.1109/TBME.2011.21625​11 Lodhi AM et al (2018) A novel approach using voting from ECG
Huang JS, Chen BQ, Zeng NY, Cao XC, Li Y (2020) Accurate clas- leads to detect myocardial infarction. IntelliSys. https​: //doi.
sification of ECG arrhythmia using MOWPT enhanced fast org/10.1007/978-3-030-01057​-7_27
compression deep learning networks. J Ambient Intell Humaniz Lui HW, Chow KL (2018) Multiclass classification of myocardial
Comput. https​://doi.org/10.1007/s1265​2-020-02110​-y infarction with convolutional and recurrent neural networks for
Hurst HE (1956) Methods of using long-term storage in reservoirs. portable ECG devices. Inf Med Unlocked 13:26–33. https​://doi.
Proc Inst Civ Eng 5(5):519–543. https​://doi.org/10.1680/iicep​ org/10.1016/j.imu.2018.08.002
.1956.11503​ Martis JR, Acharya UR, Adeli H (2014) Current methods in elec-
Jahmunah V et al (2019a) Automated detection of schizophrenia trocardiogram characterization. Comput Biol Med 48:133–149.
using nonlinear signal processing method. Artif Intell Med. https​://doi.org/10.1016/j.compb​iomed​.2014.02.012
https​://doi.org/10.1016/j.artme​d.2019.07.006 Murat F, Yildirim O, Talo M, Baloglu UB, Demir Y, Acharya UR
Jahmunah V et al (2019b) Computer-aided diagnosis of congestive (2020) Application of deep learning techniques for heartbeats
heart failure using ECG signals—a review. Phys Med. https​:// detection using ECG signals-analysis and review. Comput Biol
doi.org/10.1016/j.ejmp.2019.05.004 Med 120:103726
Kannathal N, Lim CM, Rajendra Acharya U, Sadasivan PK (2006) Padhy S, Dandapat S (2017) Third-order tensor based analysis
Cardiac state diagnosis using adaptive neuro-fuzzy technique. of multilead ECG for classification of myocardial infarc-
Med Eng Phys. https​://doi.org/10.1109/IEMBS​.2005.16153​04 tion. Biomed Signal Process Control 31:71–78. https​: //doi.
Kora P (2017) ECG based myocardial infarction detection using org/10.1016/j.bspc.2016.07.007
hybrid firefly algorithm. Comput Methods Programs Biomed Pal S, Mitra M (2012) Empirical mode decomposition based ECG
152:141–148 enhancement and QRS detection. Comput Biol Med 42(1):83–
Kumar N, Pachori RB, Acharya UR (2017) Automated diagnosis 92. https​://doi.org/10.1016/j.compb​iomed​.2011.10.012
of myocardial infarction ECG signals using sample entropy in Pan J, Tompkins WJ (1985) A real time QRS detection algo-
flexible analytic wavelet transform framework. Entropy. https​:// rithm. IEEE Trans Biomed Eng 32(3):230–236. https​: //doi.
doi.org/10.3390/e1909​0488 org/10.1109/TBME.1985.32553​2
Labate D, Foresta FL, Occhiuto G, Morabito FC, Lay-Ekuakille A, Pandey S, Voorsluys W, Niu S, Khandoker A, Buyya R (2012) An
Vergallo P (2013) Empirical mode decomposition vs. wavelet autonomic cloud environment for hosting ECG data analysis
decomposition for the extraction of respiratory signal from sin- services. Future Gener Comput Syst 28:147–154. https​://doi.
gle-channel ECG: a comparison. IEEE Sens J 13(7):2666–2674. org/10.1016/j.futur​e.2011.04.022
https​://doi.org/10.1109/JSEN.2013.22577​42 Paul JK, Iype T, Hagiwara DRY, Koh JEW, Acharya UR (2019)
Lecun Y, Bengio Y (1995) Convolutional networks for images, Characterization of fibromyalgia using sleep EEG signals with
speech, and time series. In: Arbib MA (ed) The hand book on nonlinear dynamical features. Comput Biol Med. https​://doi.
brain theory neural networks. MIT Press org/10.1016/j.compb​iomed​.2019.10333​1
Lecun Y, Bengio Y, Hinton G (2015) Deep learning. Nature Peng CK, Havlin S, Hausdorf JM, Mietus JE et al (1996) Frac-
521(7553):436–444. https​://doi.org/10.1038/natur​e1453​9 tal mechanisms and heart rate dynamics. J Electro Cardiol
Ley K (2015) Atherosclerosis, thrombosis and vascular biology. 8(15):59–64. https​://doi.org/10.1016/S0022​-0736(95)80017​-4
J Am Heart Assoc 15(9):436–448. https​: //doi.org/10.1161/ Pham TH, Vicnesh J, Wei JKE, Oh SL, Arunkumar N, Abdulhay
ATVBA​HA.117.30981​3 EW (2020a) Autism spectrum disorder diagnostic system using
Lih OS, Jahmunah V, San TR, Ciaccio EJ, Yamakawa T, Tanabe HOS bispectrum with EEG signals. Int J Environ Res Public
M, Kobayashi M, Faust O (2020) Comprehensive electrocar- Health 17(3):971
diographic diagnosis based on deep learning. Artif Intell Med Pham TH, Raghavendra U, Koh JEW, Gudigar A, Chan WY, Hamid
103:101789 MTR, Rahmat K, Fadzil F, Ng KH, Ooi CP, Ciaccio EJ, Fujita
Lin Z, Gao Y, Chen Y, Ge Q, Mahara G, Zhang J (2020) Auto- H, Acharya UR (2020b) Development of breast papillary index
mated detection of myocardial infarction using robust features for differentiation of benign and malignant lesions using ultra-
extracted from 12-lead ECG. Signal Image Video Process. https​ sound images. J Ambient Intell Humaniz Comput. https​://doi.
://doi.org/10.1007/s1176​0-019-01617​-y org/10.1007/s1265​2-020-02310​-6
Liu C, Wang H, Lu Z (2013) EEG classification for multiclass Pincus SM (1991) Approximate entropy as a measure of system
motor imagery BCI. Control and decision conference (CCDC) complexity. Proc Natl Acad Sci 88(12):2297–2301. https​://doi.
25th Chinese, pp 4450–4453. https ​ : //doi.org/10.1109/ org/10.1073/pnas.88.6.2297
CCDC.2013.65617​36 Ramesh G, Satyanarayana D, Sailaja M (2020) Composite feature
Liu B, Liu J, Wang G, Huang K, Li F et al (2014) A novel electrocar- vector based cardiac arrhythmia classification using convolu-
diogram parameterization algorithm and its application in myo- tional neural networks. J Ambient Intell Humaniz Comput. https​
cardial infarction detection. Jpn J Clin Med 27(12):2892–2900. ://doi.org/10.1007/s1265​2-020-02259​-6
https​://doi.org/10.1016/j.compb​iomed​.2014.08.010 Reasat T, Shahnaz C (2018) Detection of inferior myocardial infarc-
Liu W, Huang Q, Chang S, Wang H, He J (2018a) Multiple-feature- tion using shallow convolutional neural networks. In: 5th IEEE
branch convolutional neural network for myocardial infarction Reg 10 Humanit Technol Conf 2017, R10-HTC 2017, pp 718–
diagnosis using electrocardiogram. Biomed Signal Process 721. https​://doi.org/10.1109/R10-HTC.2017.82890​58
Control 45:22–32. https​://doi.org/10.1016/j.bspc.2018.05.013 Renyi A (1961) On measures of entropy and information. Math Stat
Liu W et al (2018b) Real-time multilead convolutional neural net- Probab 1(4):547–561
work for myocardial infarction detection. IEEE J Biomed Heal Ribeiro AH, Ribeiro MH, Paixão GMM, Oliveira DM, Gomes PR,
Inf 22(5):1434–1444. https​://doi.org/10.1109/JBHI.2017.27717​ Canazart JA (2020) Automatic diagnosis of the 12-lead ECG
68 using a deep neural network. Nature Commun 11(1):1–9

13
3244 C. Sridhar et al.

Richman JS, Randall MJ (2000) Physiological time-series analysis Student Biometrika (1908) The probable error of a mean. Biometrika
using approximate entropy and sample entropy. Am J Physiol 6(1):1–25. https​://www.jstor​.org/stabl​e/23315​54
Heart Circ Physiol 278:H2039–H2049 Tamura T (2012) Home geriatric physiological measurements. Physiol
Robert H, Mangner N, Schuler G, Erbs S (2013) Physical exercise Meas 33(10):R47. https:​ //doi.org/10.1088/0967-3334/33/10/R47
training and coronary artery disease. Rev Health Care 4(3):1–17 Thomas FL (2015) Myocardial infarction: mechanisms, diagnosis,
Roger VL (2007) Epidemiology of myocardial infarction. Med and complications. Eur Heart J 36(16):947–949. https​://doi.
Clin North Am 91(7):537–552. https ​ : //doi.org/10.1016/j. org/10.1093/eurhe​artj/ehv07​1
mcna.2007.03.007 Tripathy RK, Bhattacharyya A, Pachori RB (2019) A novel approach
Rosenstein M, Colins JJ, Luca CJ (1993) A practical method for calcu- for detection of myocardial infarction from ECG signals of mul-
lating largest Lyapunov exponents from small data sets. Physica tiple electrodes. IEEE Sens J 19(12):4509–4517
D 65:117–134. https​://doi.org/10.1016/0167-2789(93)90009​-P Venkatesan C, Karthigaikumar P, Satheeskumaran S (2018) Mobile
Rosso OA, Blanco S, Yordanova J, Kolev V et al (2001) Wavelet cloud computing for ECG telemonitoring and real-time coro-
entropy: a new tool for analysis of short duration electrical sig- nary heart disease risk detection. Biomed Signal Process Control
nals. J Neurosci Methods 105:65–75. https​://doi.org/10.1016/ 44:138–145. https​://doi.org/10.1016/j.bspc.2018.04.013
S0165​-0270(00)00356​-3 Venu K, Natesan P, Krishnakumar B, Sasipriya N (2019) Classification
Sadhukhan D, Pal S, Mitra M (2018) Automated identification of myo- of myocardial infarction using convolution neural network. Int J
cardial infarction using harmonic phase distribution pattern of Recent Technol Eng 8(4):12763–12768
ECG Data. IEEE Trans Instrum Meas 67(10):2303–2313. https​ Webber CL, Zbilut JP (1994) Dynamical assessment of physiological
://doi.org/10.1109/TIM.2018.28164​58 systems and states using recurrence plot strategies. J Appl Physiol
Sankari Z, Adeli H (2011) Heart Saver: a mobile cardiac monitoring 76(2):965–973. https​://doi.org/10.1152/jappl​.1994.76.2.965
system for auto-detection of arial fibrillation, myocardial infarc- WHO fact sheet (2012) https​://www.who.int/media​centr​e/facts​heets​/
tion, and atrio-ventricular block. Comput Biol Med 41:211–220. fs310​/en/index​.html. Accessed 19 Jan 2016
https​://doi.org/10.1016/j.compb​iomed​.2011.02.002 Zbilut JP, Webber CL (1992) Embeddings and delays as derived from
Seenivasagam V, Chitra R (2016) Myocardial infarction detection using quantification of recurrence plots. Phys Lett 171(4):199–203. https​
intelligent algorithms. Neural Netw World 26(1):91–110 ://doi.org/10.1016/0375-9601(92)90426​-M
Setiawan NA, Prabowo DW, Nugroho HA (2014) Benchmarking of Zbilut JP, Thomasson N, Webber CL (2002) Recurrence quantification
feature selection techniques for coronary artery disease diagno- analysis as a tool for nonlinear exploration of nonstationary car-
sis. In: 6th International Conference on Information Technology diac signals. Med Eng Phys 24(4):53–60. https​://doi.org/10.1016/
and Electrical Engineering, Yogyakarta, Indonesia, vol 3(1), pp S1350​-4533(01)00112​-6
123–131. https​://doi.org/10.1109/ICITE​ED.2014.70078​98 Zhang J et al (2019) Automated detection and localization of myo-
Shannon C (1948) A mathematical theory of communication. Bell Syst cardial infarction with staked sparse autoencoder and treebagger.
Tech J 27:379–423. https​://doi.org/10.1002/j.1538-7305.1948. IEEE Access 7:70634–70642. https​://doi.org/10.1109/ACCES​
tb013​38.x S.2019.29190​68
Sharma LN, Tripathy RK, Dandapat S (2015) Multiscale energy and Ziv J, Lempel A (1977) A universal algorithm for sequential data
eigenspace approach to detection and localization of myocardial compression. IEEE Trans Inf Theory 23:337–343. https​://doi.
infarction. IEEE Trans Biomed Eng 62(7):1827–1837 org/10.1109/TIT.1977.10557​14
Sharma M, Tan RS, Acharya UR (2018) A novel automated diagnostic
system for classification of myocardial infarction ECG signals Publisher’s Note Springer Nature remains neutral with regard to
using an optimal biorthogonal filter bank. Comput Biol Med jurisdictional claims in published maps and institutional affiliations.
102:341–356. https:​ //doi.org/10.1016/j.compbi​ omed.​ 2018.07.005
Song Y, Liò P (2010) A new approach for epileptic seizure detec-
tion: sample entropy based feature extraction and extreme
learning machine. J Biomed Sci Eng 3(6):556–567. https​://doi.
org/10.1152/ajphe​art.2000.278.6.H2039​
Strodthoff N, Strodthoff C (2019) Detecting and interpreting myocar-
dial infarction using fully convolutional neural networks. Physiol
Meas 40(1):1–11. https​://doi.org/10.1088/1361-6579/aaf34​d

13