Automated prediction of sudden cardiac death using statistically extracted features from electrocardiogram signals

International Journal of Electrical and Computer Engineering (IJECE)
Vol. 12, No. 5, October 2022, pp. 4960~4969
ISSN: 2088-8708, DOI: 10.11591/ijece.v12i5.pp4960-4969  4960
Journal homepage: http://ijece.iaescore.com
Automated prediction of sudden cardiac death using
statistically extracted features from electrocardiogram signals
Karna Viswavardhan Reddy1
, Navin Kumar2
1
Department of Electronics Engineering, Jain University, Bengaluru, India
2
Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham,
Bengaluru, India
Article Info ABSTRACT
Article history:
Received Jul 8, 2021
Revised May 26, 2022
Accepted Jun 14, 2022
Sudden cardiac death (SCD) is becoming a severe problem despite
significant advancements in the usage of the information and communication
technology (ICT) in the health industry. Predicting an unexpected SCD of a
person is of high importance. It might increase the survival rate. In this
work, we have developed an automated method for predicting SCD utilizing
statistical measures. We extracted the intrinsic attributes of the
electrocardiogram (ECG) signals using Hilbert-Huang and wavelet
transforms. Then utilizing machine learning (ML) classifier, we are using
these traits to automatically classify regular and SCD existing risks. Support
vector machine (SVM), decision tree (DT), naive Bayes (NB), discriminate
k-nearest neighbors (KNN), analysis (Disc.), as well as an ensemble of
classifiers also utilized (Ens.). The efficiency and practicality of the
proposed methods are evaluated using a standard database and measured
ECG data obtained from 18 ECG records of SCD cases and 18 ECG records
of normal cases. For the automated scheme, the set of features can predict
SCD very fast that is, half an hour before the occurrence of SCD with an
average accuracy of 100.0% (KNN), 99.9% (SVM), 98.5% (NB), 99.4%
(DT), 99.5% (Disc.), and 100.0% (Ens.)
Keywords:
Electrocardiogram
Machine learning
Sudden cardiac arrest
Sudden cardiac death
Ventricular fibrillation
This is an open access article under the CC BY-SA license.
Corresponding Author:
Karna Viswavardhan Reddy
Department of Electronics Engineering, Jain University
#44/4, District Fund Road, Jayanagar 9th Block, Bengaluru, Karnataka 560069, India
Email: viswavardhanreddy@gmail.com
1. INTRODUCTION
Sudden cardiac death (SCD) arises when a patient’s heart starts to pump in an unstable or abnormal
rhythm (arrhythmia) and afterwards stops beating altogether. If the person survives, the condition is also
known as sudden cardiac arrest (SCA) [1]. SCD is induced by cardiovascular disease patients who have or
have not previously had a cardiac problem. The time and manner of death in such circumstances are
unanticipated [2], [3]. From the beginning of an unanticipated variations in health condition and
unconsciousness, up to a minute is typically assumed to be controlled [3]. SCD is generally the result of a
deadly cardiac activity disorder such ventricular fibrillation (VF) or ventricular tachycardia (VT) [4], or a
severe bradyarrhythmia [5]. Such arrhythmias typically result in SCA that decreases cardiac function and
found it challenging for the heart to effectively push blood out [6]. Whenever SCA issue is left neglected for
an extended period of time, it sends a message to SCD. Cardiomyopathy, coronary heart disease, valve
illnesses, and hereditary abnormalities are by far the most common causes of malignant ventricular
arrhythmias. Instantaneous death may occur if not detected accurately and treated quickly [2], [7]. According
to research findings, the rise in actual targeted treatment involvements such as, implantable cardioverter

Int J Elec & Comp Eng ISSN: 2088-8708 
Automated prediction of sudden cardiac death using statistically extracted … (Karna Viswavardhan Reddy)
4961
defibrillators (ICD) lowers SCD death [8], [9]. However, they are still expensive and only a small percent of
participants continues receiving appropriate ICD shock. The majority of unintentional deaths occur who have
not had qualified candidates [10]. As a result, the public viewing method via an automatic electrical
defibrillator (AED) has recently attracted a lot of attention as a technique for preserving patients without an
ICD against mortality after cardiogenic shock [11]. However, even now in countries when public AEDs are
commonly accessible and has strengthened recovery methods as well as when the first reaction controls are in
place, the median rate of survival with SCA continues to decrease [12]. Basic electrophysiological substrate,
showing dispersion of refractoriness, and sympathetic stimulation activity in the chambers of the heart was
all demonstrated to be harmful for cardiac arrhythmia SCD [13], [14]. As a result, early diagnosis of an
unexpected SCD in a patient having aggressive ventricular arrhythmias is critical to enhancing overall
survival.
Lately, research has focused on building effective models for calculating the risk of SCD utilizing
invasive and non-invasive methodologies including such electrophysiological scanning [11], left ventricular
ejection fraction (LVEF) [15], invasive hemodynamic assessment [16], and non-invasive electrocardiography
(ECG) [17], [18], among many others. These are still not cost-effective. In comparing to the intrusive or
imaging techniques mentioned above, ECG is much less expensive and much more generally sold. The
electrical action of the heartbeat generates an ECG, that is, an electronic signal. An organized meta-analysis
recently confirmed that certain ECG signal parameters could provide important information on the
underlying cardiac substrate abnormality that can contribute to ventricular arrhythmias including SCD. Few
of these metrics are pathophysiological control systems controlled by cardiac autonomic processes. Heart rate
variability (HRV) or heart rate turbulence (HRT) [19]–[21], echocardiography transfer procedures, and the
repolarization delay [22] are also significant. Other metrics are QRS (Q, R, and S waves in an
electrophysiological) time [23], QT (Q and T waves) gap and dispersion [24], and T-wave alternative [25].
HRV and HRT is a derivation of an ECG signal which is described as the measurement of the R-R interval’s
beat-to-beat variation. It has been seriously evaluated for SCD prediction and diagnosis. To reveal the
intrinsic features of an HRV signals for monitoring and detection of SCD, frequency domain [20], temporal
domain [26], and nonlinear techniques [27] have been proposed. This is achieved mostly through the use of
classifications to organize topics and the number of features in various processing domains. HRV or HRT
first provided impressive outcomes, but instead were eventually discovered to be unpredictive of arrhythmic
mortality [28], [29]. HRV measurement’s output in the first several days after a myocardial disease has
already been challenged and, its prognostic value has also been shown to really be low [28]. The efficacy of
HRV-based SCD risk stratification in cardiovascular events is unknown [30]. It cannot be evaluated in people
who have atrial fibrillation or to have a lot of arrhythmias [29]. As a consequence, these Electrocardiography
markers are convincing in identification of patients at higher risk of having malignancy ventricular
arrhythmia. The majority of the effort has really been centered on clinical studies. Using complex wavelet
transforms, statistical calculations, and/or electrophysiological indicators, a few research have achieved
automatic extraction of ECG characteristics immediately from ECG data [31], [32]. While, in some works, an
SCD index (SCDI) introduced technique of combining some of the features in such a way that they could
predict the SCD [33]–[35].
In this paper, our focus is to predict SCD automatically. Towards this, the contribution in this work
is at multiple levels: i) an algorithm is developed for constructing a labelled database by segmenting the
datasets; ii) data cleaning algorithm is developed to deal with missing values and removal of noise [36];
iii) feature extraction is incorporated using Hilbert-Huang transforms, or empirical mode decomposition
(EMD)-intrinsic mode functions (IMF1, IMF2, and IMF3) and wavelet transform, or multilevel 1-D wavelet
decomposition (DWT)-approximation coefficients vectors (cAs) and detail coefficients vector (cDs) [37],
[38]; iv) method developed for ranking of extracted features and selection using various statistical methods
such as analysis of variance (ANOVA), correlation analysis (dCor), and ReliefF to find the features with the
most deviations; and v) finally, we utilized artificial classifier such as k-nearest neighbor (KNN),
discriminant analysis (Disc), naive bayes (NB), support vector machine (SVM), decision tree (DT), and
ensembles of classifiers to identify regular and SCD risk areas and use these characteristics (Ens).
Contents of rest of the paper are: section 2 discusses about the datasets used in the work obtained
from various databases of both normal and SCD patients. Section 3 describes the schematic diagram of the
proposed methodology for SCD prediction along with algorithms developed. While section 4 discusses the
important performance measures, results, and comparison with state of the art. Finally, the conclusion and
future works are mentioned in section 5.
2. INPUT DATASET
For the SCD prediction process, at first, the ECG data are collected and pre-processed. The
Massachusetts Institute of Technology-Beth Israel Hospital (MIT-BIH) is the largest publicly available

 ISSN: 2088-8708
Int J Elec & Comp Eng, Vol. 12, No. 5, October 2022: 4960-4969
4962
database that provides the ECG signals for this work. Here, the two databases are considered from the
MIT-BIH namely, sudden cardiac death Holter (SCDH) and normal sinus rhythm (NSR) [39]. For the people
who are at the SCD risk stage, the SCDH database is examined and for the normal people, the NSR database
is examined. 18 recordings are involved in the SCD groups which are collected from the SCDH database.
The collection of onset ventricular fibrillation (VF) is done before 30 mins of the first lead signal of ECG.
This collection is taken at every recording of ECG. In this work, the VF is absent so that the SCDH having
the No. 40, 42, and 49 recordings were not used. Because the amplitude R wave is low, the No. 38 and 41
recordings were not incorporated due to unknown lead. Normal sinus rhythm database (NSRDB) database
built the recordings with the normal group of 18 half-hour. From Table 1, it is observed that the age of
61.1±18.7 years belongs to the SCD group as they range from 30-89 years and the age of 34.3±8.4 years
belongs to the normal group and this group ranges from 20-50 years. There are 9 males and 8 females from
the SCD group and 5 males and 13 females from the normal group but 1 SCD patient is omitted who is
considered as an unknown gender. In Table 2, the descriptions are elaborately provided with record number,
name, arrhythmia, and heart diseases group, and finally before the length of the VF. From the SCD group, it
is observed that the majority of the patients are affected by malignant ventricular arrhythmias those who are
undergoing heart disease and the abnormality of the cardiac substrate. The ECG signals with two examples
are shown in Figure 1. Figure 1(a) belongs to the SCD group that shows around the onset of VF and the
Figure 1(b) belongs to the normal group and this show around the signal’s middle part.
Table 1. The SCD and regular groups’ genders and ages
Group Total Gender Age
Male Female Unknown Range Mean±SD
SCD 18 9 8 1 30-89 61.1±18.7
Normal 18 5 13 0 20-50 34.3±8.4
Table 2. ECG data of SCD patients out from MIT-BIH dataset
Database Heart diseases No of
records
Record name Length before
VF onset
Arrhythmia
categories
NSRDB Not available (NA) 18 Whole database NA unknown
Sudden
Cardiac
Death
Holter
Database
(SDDB)
Cardiac surgery 4 32, 35, 36, 50 30 min Ventricular
tachycardia,
Ventricular
fibrillation,
Ventricular flutter
Coronary artery 1 43
Unknown 9 30, 33, 34, 37, 44, 46, 47, 48, 51
Heart failure 2 31, 52
Ventricular ectopy 1 45
Acute myelogenous leukemia 1 39
(a)
(b)
Figure 1. Two examples of ECG signals (a) 10 sec. of record 30 of SCD group around VF onset and
(b) 10 sec. of record 16265 of normal group around the middle of the signal

4963
3. ALGORITHM DEVELOPMENT
We developed an algorithm to predict SCD at early stage. Also, we have developed models using
ECG signals in which different combinations of the main components of models based on machine learning
are used. The overall stages of the proposed work are shown in Figure 2. In the following sections we briefly
describe the aspects of the components of the models.
Figure 2. Proposed methodology for SCD prediction
3.1. Pre-processing of datasets
We have applied data pre-processing mechanisms mainly for two reasons: i) enable the algorithms
to work on the datasets and ii) improve the quality of the datasets. Data pre-processing includes data
cleaning, data reduction, feature selection, and data transformation. Data cleaning on the datasets deals with
the missing values and/or reduce the noise.
In order to increase the precision and efficiency of the prediction algorithms, we designed a filter to
reduce the noise of the ECG signals. For this, a Butterworth filter of order 6 is designed to pass signals with
frequency higher than 0.5 Hz to remove the baseline wander and consider signals with frequency of higher
than 30 Hz as noise and filter them out. Normal and SCD are transformed from categories to numerical
indices. Without this transformation, the algorithms may not be able to work.
3.2. Construct a labeled database
The dataset’s (ECG data) records are collected and pre-processed initially. Continuous one-minute
signals image is segmented on each recording, and 30 one-minute ECG segments are generated, either
classified as SCD or regular. For SCD group, the 30-minutes data are selected before VF onset and for
normal group the 30-minutes data are selected from the middle of the records. At the end of this stage, we
have a labeled database consisted of 30 × (18 + 18) = 1080 one-minute fragments which are saved and
will be used for further study, i.e., to train and test our automated strategy of SCD prediction. This process is
illustrated in Figure 3.
3.3. Feature extraction
In this study, two types of transforms are used to find the intrinsic attribute curves of the ECG
signals: i) Hilbert-Huang transforms, or empirical mode decomposition (EMD) to find the intrinsic mode
functions (IMFs); ii) wavelet transform, or multilevel 1-D wavelet decomposition (DWT) to find the
approximate and detail coefficients (cAs and cDs). Figures 4 and 5 show some intrinsic curves found by
aforesaid transforms for a small duration.
Then four statistical measures (mean, variance, skewness, kurtosis) are applied on the intrinsic
curves found using Hilbert-Huang and wavelet transforms to specify a numerical value to each of these
curves, which we use them as features. Since we initially do not know how many of features distinct enough
are required to cover the whole aspects of ECG signal, we extracted quite a big number of them, 32 features.
However, it is not needed to use all features.

 ISSN: 2088-8708
4964
Segmentation of the datasets
Input: ECG signals {(𝐭𝑖,𝐫𝑖,𝑙𝑖,𝑡𝑖
𝑉𝐹𝑜𝑛𝑠𝑒𝑡
};𝑖 = 1,… , 𝑁, 𝑙𝑖 ∈ {𝑁𝑜𝑟𝑚𝑎𝑙, 𝑆𝐶𝐷}
Output: Segmented database {(𝐬𝑖,𝑦𝑖}; 𝑖 = 1, … ,30𝑁, 𝑦𝑖 ∈ {𝑁𝑜𝑟𝑚𝑎𝑙, 𝑆𝐶𝐷}
𝑐 ← 0
for i=1 to N do
if 𝑙𝑖 = 𝑁𝑜𝑟𝑚𝑎𝑙
for k=1 to 30
𝑐 ← 𝑐 + 1
𝑦𝑐 ← 𝑙𝑖
𝐣 ← 0.5𝑡𝑖(𝑒𝑛𝑑) + (−15 + 𝑘 − 1) × 60,… , 0.5𝑡𝑖(𝑒𝑛𝑑) + (−15 + 𝑘) × 60
𝐬𝑐 ← 𝐫𝑖(𝐣)
else
for k=1 to 30
𝑐 ← 𝑐 + 1
𝑦𝑐 ← 𝑙𝑖
𝐣 ← 𝑡𝑖
+ (−30 + 𝑘 − 1) × 60,… , 𝑡𝑖
+ (−30 + 𝑘) × 60
𝐬𝑐 ← 𝐫𝑖(𝐣)
Figure 3. The process of segmentation of ECG signals to prepare the database
Figure 4. Some IMFs of an ECG signal (record 16265 of NSRDB), extracted using EMD
Figure 5. Some cA and cDs of an ECG signal (record 16265 of NSRDB), extracted using DWT

4965
3.4. Feature selection
After feature extraction, we have applied three methods to rank the features: i) ANOVA (analysis of
variance) and ii) ReliefF algorithm (rank importance of predictors), and iii) dCor (correlation analysis). The
ANOVA using F-test technique is used to rank characteristics in terms of how significant they are in the
categorization [40]. The strengths of predictors are determined using ReliefF. Predictors that offer various
scores to neighbors in the very same class are penalized, while forecasters that provide different preferences
to neighbors in separate classes are awarded. To rank the characteristics, dCor utilizes a measurement of
interaction between matched vectors. Tables 3 and 4 demonstrate several of the features that were used in this
investigation.
Table 3. Sample features extracted using EMD and ranked with correlation analysis
Feature, X Label
Feature No. 2 6 1 4 8
Feature formula variance(lmf1) variance(lmfz) mean(lmf1) kurtosis(lmf1) kurtosis(lmfz) row of x record name DB
0.094411319 0.129388168 0.025532239 6.33964188 6.559645558 Normal 1 16265 NSRDB
0.181140341 0.14554032 0.089286928 6.347494434 3.986966356 Normal 2 16265 NSRDB
… … … … … … … … …
0.01723577 0.017637224 0.000622635 9.354896841 5.765803261 Normal 540 19830 NSRDB
0.00820838 0.024405427 0.005172282 8.178362444 7.110102235 SCD 541 30 SDDB
… … … … … … … … …
0.005173318 0.004320744 0.005415132 13.69867346 11.14347478 SCD 1080 52 SDDB
Table 4. Sample features extracted using DWT and ranked with ANOVA
Feature, X Label
Feature No. 2 6 1 4 8
Feature formula kurtosis(cD7) kurtosis(cD6) kurtosis(cD5) variance(cD1) variance(cA7) row of x record name DB
16.43466415 14.68811494 7.884046875 0.345274906 0.345274906 Normal 1 16265 NSRDB
20.41299529 18.90730632 9.593829463 0.306361446 0.306361446 Normal 2 16265 NSRDB
… … … … … … … … …
25.92484865 14.38789522 5.56937062 0.077736297 0.077736297 Normal 540 19830 NSRDB
31.7408592 15.51188951 5.659636272 0.134887891 0.134887891 SCD 541 30 SDDB
… … … … … … … … …
38.5807801 25.3291934 22.97580617 1.289219856 1.289219856 SCD 1080 52 SDDB
3.5. Classification
To classify ECG data into the regular and SCD groups, we utilized 6 major classifiers. The best
classification with the maximum accuracy was then chosen. These classifiers are: KNN, SVM, NB, DT,
Disc., and Ens. One such algorithm that is used in DT is shown in Figure 6.
Algorithm AdaBoostM1
Input: Dataset 𝑆 = {𝐱𝑖,𝑦𝑖}; 𝑖 = 1, …, 𝑁; 𝑦𝑖 ∈ {−1, +1},
T: Number of learners,
W: Algorithm of weak learner
Output: Boosted classifier 𝑓(𝐱) = ∑ 𝑎𝑡ℎ𝑡(𝐱)
𝑇
𝑡=1
𝑑𝑖
(1)
←
1
𝑁
; 𝑖 = 1,… , 𝑁
for t=1 to T do
ℎ𝑡 ← 𝑊(𝑆, 𝑑(𝑡)
)
𝑘𝑖
(𝑡)
← 0; 𝑖 = 1, … , 𝑁
if 𝑦𝑖 ≠ ℎ𝑡(𝐱𝑖);𝑖 = 1, … , 𝑁
𝑘𝑖
(𝑡)
← 1
𝑒𝑡 = ∑ 𝑑𝑖
(𝑡)
𝑘𝑖
(𝑡)
𝑁
𝑖=1
if 𝑒𝑡 > 0.5 then
𝑇 ← 𝑡 − 1
return
𝑎𝑡 ← 0.5ln(𝑒𝑡
−1
− 1)
𝑑𝑖
(𝑡+1)
← 𝑑𝑖
(𝑡)
exp(−𝑎𝑡ℎ𝑡(𝐱𝑖)𝑦𝑖); 𝑖 = 1, … , 𝑁
𝑑𝑖
(𝑡+1)
←
𝑑𝑖
(𝑡+1)
∑ 𝑑𝑖
(𝑡+1)
𝑁
𝑖=1
; 𝑖 = 1, …, 𝑁
Figure 6. Algorithm of AdaBoostM1

 ISSN: 2088-8708
4966
3.6. 5-fold cross-validation
The data are partitioned for classification. In other words, we have employed 5-fold cross-
validation. It means, we partition the instances to 5 parts and use 4 parts for training and 1 part for testing.
Therefore, 80% of the instances are used for training and 20% of them are used for testing. This scheme is
repeated for all parts, i.e., 5 times.
4. RESULTS AND DISCUSSION
4.1. Performance measures
As mentioned before, we employed 5-fold cross validation method to build and evaluate the
performance of the classifiers used. After finding the model some metrics are used to evaluate the
performance of models found using the classifiers. For this purpose, widely used classification performance
measures, sensitivity (or recall), specificity, accuracy, precision, and F1-score are computed by (1)-(5). They
are used to evaluate the performance of the proposed methods for prediction of SCD.
𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦 =
𝑇𝑃
𝑇𝑃+𝐹𝑁
(1)
𝑆𝑝𝑒𝑐𝑖𝑓𝑖𝑐𝑖𝑡𝑦 =
𝑇𝑁
𝑇𝑁+𝐹𝑃
(2)
𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =
𝑇𝑁+𝑇𝑃
𝑇𝑁+𝑇𝑃+𝐹𝑃+𝐹𝑁
(3)
𝑃𝑟 𝑒 𝑐𝑖𝑠𝑖𝑜𝑛 =
𝑇𝑃
𝑇𝑃+𝐹𝑃
(4)
𝐹1 − 𝑠𝑐𝑜𝑟𝑒 =
2.𝑃𝑟 𝑒𝑐𝑖𝑠𝑖𝑜𝑛.𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦
𝑃𝑟 𝑒𝑐𝑖𝑠𝑖𝑜𝑛+𝑆𝑒𝑛𝑠𝑖𝑡𝑖𝑣𝑖𝑡𝑦
(5)
In which the true positive (TP), true negative (TN), false positive (FP), and false negative (FN) are the
components of the confusion matrix and their meanings are summarized in Table 5.
Table 5. Meanings of the components of the confusion matrix
Term Meaning Meaning in this work
TP True Positive The number of SCD which are recognized as SCD
TN True Negative The number of not SCD which are recognized as not SCD
FP False Positive The number of not SCD which are recognized as SCD
FN False Negative The number of SCD which are recognized as not SCD
All the 1080 one-minute fragments are used to train and test our automated strategy of the SCD
prediction. These fragments are passed through a Butterworth filter to reduce the noise. Then they are
supplied to the transforms and found the intrinsic curves. The intrinsic curves are supplied to 4 statistical
measures and the results are used as extracted features. The extracted features are ranked and selected for
classification. Then the classifiers are employed to find and test the models. The details of these stages are
explained in the previous sections. Table 6 summarize the average performances of the various models we
used by different combinations of the feature extractions, feature selections, and classifiers. As it can be seen,
the best performance we got is 100% accuracy.
Detailed statistics of performance measures are presented in Table 6. With DWT as feature
extraction technique, ANOVA as feature ranking method and KNN classifier shows the best performance
measures (100%) for 6 features. Whereas, for 5 features-(DWT, dCor, Ens.), (DWT ReliefF, KNN) and
(DWT, ReliefF, Ens) show best performance measures of almost 100% and at the same time, other classifiers
were exhibiting poor performance. It is seen that DWT as a feature extraction technique has performed the
best when compared with EMD. KNN and ensemble of classifiers have shown the best performance among
other classifiers, where as in feature ranking, all the three have performed well when ranked with a smaller
number of features i.e., 5 and 6.
4.2. Comparative study
Table 7 (in appendix) compares the performance of numerous studies that was using ECG or HRV
data to identify SCD. We found 7 papers from 2015 to 2019 that can provide specific information on the
research in three areas: data material, technique, and results. We specified the sorts of signals used the

4967
databases from which the data was acquired, and the length of signals (in minutes) used in each study in
terms of materials. In terms of methodologies, we mention the techniques applied by each work separately,
and the number of markers and the classification. We offer performance data in terms of accuracy,
specificity, and sensitivity.
Table 6. The statistical measures and the average performance of methods we used
Method
Feature
ranking
No. of
features
TP TN FP FN Sensitivity Specificity Accuracy Precision F1-score
EMD+KNN ANOVA 10 444 499 41 96 82.2 92.4 87.3 91.5 86.6
EMD+SVM ANOVA 10 442 471 69 98 81.9 87.2 84.5 86.5 84.1
EMD+Ensemble ANOVA 10 534 533 7 6 98.9 98.7 98.8 98.7 98.8
EMD+NB ANOVA 10 447 466 74 93 82.8 86.3 84.5 85.8 84.3
EMD+DT ANOVA 10 433 490 50 107 80.2 90.7 85.5 89.6 84.7
EMD+Discrement ANOVA 10 357 496 44 183 66.1 91.9 79.0 89.0 75.9
DWT+KNN ANOVA 6 540 540 0 0 100.0 100.0 100.0 100.0 100.0
DWT+SVM ANOVA 6 501 498 42 39 92.8 92.2 92.5 92.3 92.5
DWT+Ensemble ANOVA 6 534 531 9 6 98.9 98.3 98.6 98.3 98.6
DWT+NB ANOVA 6 511 524 16 29 94.6 97.0 95.8 97.0 95.8
DWT+DT ANOVA 6 533 522 18 7 98.7 96.7 97.7 96.7 97.7
DWT+Discrement ANOVA 6 455 530 10 85 84.3 98.1 91.2 97.8 90.5
EMD+KNN
Correlation
analysis 5 526 518 22 14 97.4 95.9 96.7 96.0 96.7
EMD+SVM
Correlation
analysis 5 496 505 35 44 91.9 93.5 92.7 93.4 92.6
EMD+Ensemble
Correlation
analysis 5 536 538 2 4 99.3 99.6 99.4 99.6 99.4
EMD+NB
Correlation
analysis 5 523 447 93 17 96.9 82.8 89.8 84.9 90.5
EMD+DT
Correlation
analysis 5 510 523 17 30 94.4 96.9 95.6 96.8 95.6
EMD+Discrement
Correlation
analysis 5 511 461 79 29 94.6 85.4 90.0 86.6 90.4
DWT+KNN
Correlation
analysis 5 534 537 3 6 98.9 99.4 99.2 99.4 99.2
DWT+SVM
Correlation
analysis 5 535 537 3 5 99.1 99.4 99.3 99.4 99.3
DWT+Ensemble
Correlation
analysis 5 540 540 0 0 100.0 100.0 100.0 100.0 100.0
DWT+NB
Correlation
analysis 5 522 540 0 18 96.7 100.0 98.3 100.0 98.3
DWT+DT
Correlation
analysis 5 537 536 4 3 99.4 99.3 99.4 99.3 99.4
DWT+Discrement
Correlation
analysis 5 521 536 4 19 96.5 99.3 97.9 99.2 97.8
EMD+KNN ReliefF 5 510 518 22 30 94.4 95.9 95.2 95.9 95.1
EMD+SVM ReliefF 5 506 515 25 34 93.7 95.4 94.5 95.3 94.5
EMD+Ensemble ReliefF 5 519 517 23 21 96.1 95.7 95.9 95.8 95.9
EMD+NB ReliefF 5 382 524 16 158 70.7 97.0 83.9 96.0 81.4
EMD+DT ReliefF 5 507 523 17 33 93.9 96.9 95.4 96.8 95.3
EMD+Discrement ReliefF 5 499 480 60 41 92.4 88.9 90.6 89.3 90.8
DWT+KNN ReliefF 5 540 540 0 0 100.0 100.0 100.0 100.0 100.0
DWT+SVM ReliefF 5 539 540 0 1 99.8 100.0 99.9 100.0 99.9
DWT+Ensemble ReliefF 5 540 540 0 0 100.0 100.0 100.0 100.0 100.0
DWT+NB ReliefF 5 525 539 1 15 97.2 99.8 98.5 99.8 98.5
DWT+DT ReliefF 5 533 537 3 7 98.7 99.4 99.1 99.4 99.1
DWT+Discrement ReliefF 5 537 538 2 3 99.4 99.6 99.5 99.6 99.5
5. CONCLUSION AND FUTURE WORK
We have proposed an automated method for predicting SCD utilizing statistical measures in this
paper. Different classifiers are used in this work. With an accuracy rate of 100.0% (KNN), 98.5% (NB),
100.0% (Disc.), 99.9% (SVM), 99.4% (DT), 98.5% (NB), and the system can predict SCD very quickly, that
is 30 minutes before it occurs (Ens.). From the studies, it is observed that prediction of SCD is always a
challenging task. The current work is carried based on the ECG databases available publicly, i.e., NSRDB
and are of relatively small databases. More research is required to acquire a large amount of clinical data to
educate the suggested classifications and to assess their validity. Additionally, it should be noted that the
methodology used in this study must be tested in clinical conditions. Furthermore, such projections might be
more realistic and useful if they have been made in real time on mobile devices in hospitals or at home.

 ISSN: 2088-8708
4968
APPENDIX
Table 7. A comparison of our research with some other previous ECG or HRV signal-based approaches
Author (year) Material Methodology Best performance
Data
type
Dataset
used
Length of
signal(min)
Feature extraction
(No of features)
Classifier Acc (%) Sen (%) Spe (%)
Acharya
(2015) [31]
ECG SDDB
NSRDB
4 minutes
before SCD
Nonlinear features (18) and
SCDI
DT, SVM 92.11% 92.50% 91.67%
Fujita
(2016) [27]
HRV SDDB
NSRDB
4 minutes
before SCD
Nonlinear features (4), nonlinear
heart rate variability analysis
SVM, KNN 94.70% 95.00% 94.40%
Sanchez
(2018) [27]
ECG SDDB
NSRDB
20 minutes
before SCD
Nonlinear methods HI, Wave
packet transform
EPNN 95.80% unknown unknown
Khazaei
(2018) [21]
HRV SDDB
NSRDB
6 minutes
before SCD
Wave packet transform RQA
(13) and increment entropy
(2 out of 14) Nonlinear method
DT, KNN,
SVM, NB
95.00% 95.00% 95.00%
Ebrahimzadeh
(2018) [19]
HRV SDDB
NSRDB
12 minutes
before SCD
HRV features (23) Time local
subset feature selection
MLP 88.29% unknown unknown
Ebrahimzadeh
(2019) [20]
HRV SDDB
NSRDB
13 minutes
before SCD
HRV features (23) time local
subset feature selection
MLP 90.18% unknown unknown
Lai (2019)
[35]
ECG SDDB
NSRDB
AHADB
30 minutes
before SCD
Arrhythmias risk markers (5)
and SCDI
DT, KNN, SVM,
NB, RF
99.49% 99.75% 99.04%
Present work ECG SDDB
NSRDB
30 minutes
before SCD
Nonlinear (5) (EMD and
DWT)
KNN, SVM, NB,
DT, Dis, Ens
100% 100% 100%
REFERENCES
[1] A. G. Yow, V. Rajasurya, and S. Sharma, “Sudden cardiac death,” StatPearls. 2021, Accessed: Mar. 10, 2022. [Online].
Available: https://www.ncbi.nlm.nih.gov/books/NBK507854/
[2] D. P. Zipes and H. J. J. Wellens, “Sudden cardiac death,” Circulation, vol. 98, no. 21, pp. 2334–2351, Nov. 1998, doi:
10.1161/01.CIR.98.21.2334.
[3] R. J. Myerburg and A. Castellanos, “Cardiac arrest and sudden cardiac death,” in Braunwald’s Heart Disease: A Textbook of
Cardiovascular Medicine, Elsevier, 2012, pp. 845–884.
[4] R. M. John et al., “Ventricular arrhythmias and sudden cardiac death,” The Lancet, vol. 380, no. 9852, pp. 1520–1529, Oct. 2012,
doi: 10.1016/S0140-6736(12)61413-5.
[5] T.-W. Shen, H.-P. Shen, C.-H. Lin, and Y.-L. Ou, “Detection and prediction of sudden cardiac death (SCD) for personal
healthcare,” in 2007 29th Annual International Conference of the IEEE Engineering in Medicine and Biology Society, Aug. 2007,
pp. 2575–2578, doi: 10.1109/IEMBS.2007.4352855.
[6] N. J. Pagidipati and T. A. Gaziano, “Estimating deaths from cardiovascular disease: a review of global methodologies of mortality
measurement,” Circulation, vol. 127, no. 6, pp. 749–756, Feb. 2013, doi: 10.1161/CIRCULATIONAHA.112.128413.
[7] G. Finocchiaro, M. Papadakis, S. Sharma, and M. Sheppard, “Sudden cardiac death,” European Heart Journal, vol. 38, no. 17,
pp. 1280–1282, May 2017, doi: 10.1093/eurheartj/ehx194.
[8] R. Passman, “Prevention of sudden cardiac death in dialysis patients: drugs, defibrillators or what else,” Blood Purification,
vol. 35, no. 1–3, pp. 49–54, 2013, doi: 10.1159/000345178.
[9] A. Cheng et al., “Prospective observational study of implantable cardioverter‐defibrillators in primary prevention of sudden
cardiac death: study design and cohort description,” Journal of the American Heart Association, vol. 2, no. 1, Jan. 2013, doi:
10.1161/JAHA.112.000083.
[10] S. S. Chugh, “Sudden cardiac death with apparently normal heart: Clinical implications of progress in pathophysiology,” Cardiac
Electrophysiology Review, vol. 5, no. 4, p. 394, 2001.
[11] G. I. Fishman et al., “Sudden cardiac death prediction and prevention,” Circulation, vol. 122, no. 22, pp. 2335–2348, Nov. 2010,
doi: 10.1161/CIRCULATIONAHA.110.976092.
[12] Fei Zhang, Pengye Li, Fan Jiang, and Dakun Lai, “A shockable rhythm detection algorithm for automatic external defibrillators
by combining a slope variability analyzer with a band-pass digital filter,” in 2014 IEEE Workshop on Electronics, Computer and
Applications, May 2014, pp. 828–831, doi: 10.1109/IWECA.2014.6845749.
[13] R. Passman and J. J. Goldberger, “Predicting the future,” Circulation, vol. 125, no. 24, pp. 3031–3037, Jun. 2012, doi:
10.1161/CIRCULATIONAHA.111.023879.
[14] E. F. Aziz, F. Javed, B. Pratap, and E. Herzog, “Strategies for the prevention and treatment of sudden cardiac death,” Open Access
Emergency Medicine, Dec. 2010, doi: 10.2147/OAEM.S6869.
[15] S. D. Solomon et al., “Sudden death in patients with myocardial infarction and left ventricular dysfunction, heart failure, or both,”
New England Journal of Medicine, vol. 352, no. 25, pp. 2581–2588, Jun. 2005, doi: 10.1056/NEJMoa043938.
[16] R. Pacher et al., “Prognostic impact of big endothelin-1 plasma concentrations compared with invasive hemodynamic evaluation
in severe heart failure,” Journal of the American College of Cardiology, vol. 27, no. 3, pp. 633–641, Mar. 1996, doi:
10.1016/0735-1097(95)00520-X.
[17] S. Mandala and T. C. Di, “ECG parameters for malignant ventricular arrhythmias: a comprehensive review,” Journal of Medical
and Biological Engineering, vol. 37, no. 4, pp. 441–453, Aug. 2017, doi: 10.1007/s40846-017-0281-x.
[18] Y. Castro-Torres, “Ventricular repolarization markers for predicting malignant arrhythmias in clinical practice,” World Journal of
Clinical Cases, vol. 3, no. 8, p. 705, 2015, doi: 10.12998/wjcc.v3.i8.705.
[19] E. Ebrahimzadeh, M. S. Manuchehri, S. Amoozegar, B. N. Araabi, and H. Soltanian-Zadeh, “A time local subset feature selection
for prediction of sudden cardiac death from ECG signal,” Medical and Biological Engineering and Computing, vol. 56, no. 7,
pp. 1253–1270, Jul. 2018, doi: 10.1007/s11517-017-1764-1.
[20] E. Ebrahimzadeh et al., “An optimal strategy for prediction of sudden cardiac death through a pioneering feature-selection
approach from HRV signal,” Computer Methods and Programs in Biomedicine, vol. 169, pp. 19–36, Feb. 2019, doi:

4969
10.1016/j.cmpb.2018.12.001.
[21] M. Khazaei, K. Raeisi, A. Goshvarpour, and M. Ahmadzadeh, “Early detection of sudden cardiac death using nonlinear analysis
of heart rate variability,” Biocybernetics and Biomedical Engineering, vol. 38, no. 4, pp. 931–940, 2018, doi:
10.1016/j.bbe.2018.06.003.
[22] G. Tse and B. P. Yan, “Traditional and novel electrocardiographic conduction and repolarization markers of sudden cardiac
death,” EP Europace, vol. 19, no. 5, pp. 712–721, May 2017, doi: 10.1093/europace/euw280.
[23] D. P. Morin et al., “QRS duration predicts sudden cardiac death in hypertensive patients undergoing intensive medical therapy:
the LIFE study,” European Heart Journal, vol. 30, no. 23, pp. 2908–2914, Dec. 2009, doi: 10.1093/eurheartj/ehp321.
[24] K. S. Spargias et al., “QT dispersion as a predictor of long-term mortality in patients with acute myocardial infarction and clinical
evidence of heart failure,” European Heart Journal, vol. 20, no. 16, pp. 1158–1165, Aug. 1999, doi: 10.1053/euhj.1998.1445.
[25] R. L. Verrier and T. Ikeda, “Ambulatory ECG-based T-wave alternans monitoring for risk assessment and guiding medical
therapy: mechanisms and clinical applications,” Progress in Cardiovascular Diseases, vol. 56, no. 2, pp. 172–185, Sep. 2013, doi:
10.1016/j.pcad.2013.07.002.
[26] I. Cygankiewicz et al., “Heart rate turbulence predicts all-cause mortality and sudden death in congestive heart failure patients,”
Heart Rhythm, vol. 5, no. 8, pp. 1095–1102, Aug. 2008, doi: 10.1016/j.hrthm.2008.04.017.
[27] H. Fujita et al., “Sudden cardiac death (SCD) prediction based on nonlinear heart rate variability features and SCD index,”
Applied Soft Computing, vol. 43, pp. 510–519, Jun. 2016, doi: 10.1016/j.asoc.2016.02.049.
[28] M. T. La Rovere, J. T. Bigger, F. I. Marcus, A. Mortara, and P. J. Schwartz, “Baroreflex sensitivity and heart-rate variability in
prediction of total cardiac mortality after myocardial infarction,” The Lancet, vol. 351, no. 9101, pp. 478–484, Feb. 1998, doi:
10.1016/S0140-6736(97)11144-8.
[29] R. Liew, “Electrocardiogram-based predictors of sudden cardiac death in patients with coronary artery disease,” Clinical
Cardiology, vol. 34, no. 8, pp. 466–473, Aug. 2011, doi: 10.1002/clc.20924.
[30] H. Evrengul et al., “The relationship between heart rate recovery and heart rate variability in coronary artery disease,” Annals of
Noninvasive Electrocardiology, vol. 11, no. 2, pp. 154–162, Apr. 2006, doi: 10.1111/j.1542-474X.2006.00097.x.
[31] U. R. Acharya et al., “An integrated index for detection of sudden cardiac death using discrete wavelet transform and nonlinear
features,” Knowledge-Based Systems, vol. 83, pp. 149–158, Jul. 2015, doi: 10.1016/j.knosys.2015.03.015.
[32] J. P. Amezquita-Sanchez, M. Valtierra-Rodriguez, H. Adeli, and C. A. Perez-Ramirez, “A novel wavelet transform-homogeneity
model for sudden cardiac death prediction using ECG signals,” Journal of Medical Systems, vol. 42, no. 10, Oct. 2018, doi:
10.1007/s10916-018-1031-5.
[33] D. Ghista, Applied biomedical engineering mechanics. CRC Press, 2008.
[34] U. R. Acharya et al., “An integrated diabetic index using heart rate variability signal features for diagnosis of diabetes,” Computer
Methods in Biomechanics and Biomedical Engineering, vol. 16, no. 2, pp. 222–234, 2013, doi: 10.1080/10255842.2011.616945.
[35] D. Lai, Y. Zhang, X. Zhang, Y. Su, and M. B. Bin Heyat, “An automated strategy for early risk identification of sudden cardiac
death by using machine learning approach on measurable arrhythmic risk markers,” IEEE Access, vol. 7, pp. 94701–94716, 2019,
doi: 10.1109/ACCESS.2019.2925847.
[36] J. R. Mohammed, “Low complexity adaptive noise canceller for mobile phones based remote health monitoring,” International
Journal of Electrical and Computer Engineering (IJECE), vol. 4, no. 3, pp. 422–432, Jun. 2014, doi: 10.11591/ijece.v4i3.5534.
[37] N. E. Huang and N. O. Attoh-Okine, Eds., The Hilbert-Huang transform in engineering. CRC Press, 2005.
[38] T. L. T. da Silveira, A. J. Kozakevicius, and C. R. Rodrigues, “Single-channel EEG sleep stage classification based on a
streamlined set of statistical features in wavelet domain,” Medical and Biological Engineering and Computing, vol. 55, no. 2,
pp. 343–352, Feb. 2017, doi: 10.1007/s11517-016-1519-4.
[39] A. L. Goldberger et al., “PhysioBank, PhysioToolkit, and PhysioNet,” Circulation, vol. 101, no. 23, Jun. 2000, doi:
10.1161/01.CIR.101.23.e215.
[40] A. Grünauer and M. Vincze, “Using dimension reduction to improve the classification of high-dimensional data,”
arxiv.org/abs/1505.06907, May 2015.
BIOGRAPHIES OF AUTHORS
Karna Viswavardhan Reddy received the Bachelor of Technology, degree in
Electronics and Communication Engineering from JNTU-Hyderabad, A. P, India, in 2008, the
M. S. degree in Telecommunication Systems from Blekinge Institute of Technology
Karlskrona, Sweden, in 2011. Currently, he is a research assistant at Department of Electronics
Engineering, Jain University, Bengaluru, India. His research interests include machine
learning, artificial intelligence, wireless body area networks, wireless sensor networks, and
software defined networks. He can be contacted at email: viswavardhanreddy@gmail.com,
viswavardhank@rvce.edu.in.
Navin Kumar graduated from a joint PhD program from the university of Minho,
Aveiro and Porto (MAP Tele) of Portugal in Europe from 2007-2011 and worked as research
collaborator at Institution of telecommunication, Aveiro, Portugal until 2012. Currently, serves
as head of the department at the Department of Electronics and Communication, School of
Engineering, Amrita Vishwa Vidyapeetham, Bengaluru, India. He is Sr Member of IEEE,
AIENG (HK), a Life Member of IETE and Fellow IE (India). His research areas include 5G
(mmWave and massive MIMO), visible light communication, optical wireless communication,
IoT health care and smart city and intelligent transportation systems. He is also the part of
IEEE 5G and future network initiative. He can be contacted at email: navinkumar@ieee.org,
navin_kum3@yahoo.com.

Automated prediction of sudden cardiac death using statistically extracted features from electrocardiogram signals

More Related Content

Similar to Automated prediction of sudden cardiac death using statistically extracted features from electrocardiogram signals (20)

More from IJECEIAES (20)

Recently uploaded (20)

Automated prediction of sudden cardiac death using statistically extracted features from electrocardiogram signals