do not necessarily reflect the views of UKDiss.com.
Early Prediction of Sudden Cardiac Death Using Nonlinear Analysis of Heart Rate Variability
Abstract ــ currently sudden cardiac death (SCD) is one of the most prevalent causes of death all around the world. With precise and early prediction of SCD, chance of survival can be improved through administering cardiopulmonary resuscitation (CPR) and defibrillation. Hence, there is a vital need for an automated SCD prediction system. In this work, a novel and efficient algorithm for automated prediction of SCD, five minutes before its onset proposed. This algorithm used nonlinear features of hearth rate variability (HRV) signals. In fact, after extraction of the HRV signals from electrocardiogram (ECG) recordings, increment entropy and recurrence quantification analysis based features extracted. Then, the one-way ANOVA applied to reduce the dimension of feature space, which resulted in lower computational cost. Finally, the best distinguishing features get fed to various classifiers such as Decision Tree, K-Nearest Neighbor, Naive Bayes, and Support Vector Machine. SCD prediction using the proposed algorithm with DT classifier was performed with accuracy, specificity and sensitivity of 95%. These experimental results demonstrated the superiority of the proposed algorithm to the existing ones in performance.
Sudden cardiac arrest (SCA) is known as a critical cardiovascular condition, in which sudden cardiac death (SCD) of an individual with or without preexisting cardiac disease is expected if the necessary proceedings for the heart revitalize does not take place within few minutes [1-5]. In the majority of times, cardiac arrhythmias such as ventricular fibrillation (VF), ventricular tachycardia (VT), and ventricular flutter (VFL) initiate the occurrence of SCD. A primary bradyarrhythmia is associated with a smaller number of SCD events . SCA immediately causes body wide circulatory failure, resulting in oxygen deprivation and subsequent loss of consciousness within one minute . When SCA occurs, electrical ventricular defibrillator is mainly utilized to stimulate the heart and reinstate its function. Although, during the last decades, the mortality rate caused by cardiac diseases has decreased, SCD is still reported as the pioneer of death causes in the USA as well as other countries . Due to the fact that the onset of symptoms appears approximately an hour before sudden death, early prediction of SCD is a critical issue for the clinicians, because the timely cardiopulmonary resuscitation (CPR) using ventricular defibrillator will lead to successful restoration of heart activity .
Many research works have developed invasive and non-invasive techniques for SCD prediction. For this aim, electrocardiogram (ECG) and heart rate variability (HRV) are non-invasively useful signals . Studies show that the most effective and accurate predictors for SCD are QT interval dispersion and HRV from the ECG recordings [11, 12]. There are several linear, nonlinear, time-domain, and frequency-domain approaches for analysis of the ECG and HRV signals [11-13].
Shen et al. , performed a study on ECG signals in which the results of performing modified zero-crossing approach and wavelet based methods for SCD prediction before its onset was reported. For this aim, 4-minute ECG signals, two minutes both before and after VF onset time of SCD occurrence was analyzed. The results indicated that the correct detection rate using wavelet and modified zero-crossing method were 92.31% and 98.48%, respectively.
In another attempt to predict SCD using ECG signals, Kora , applied fast conjugate symmetric sequency ordered complex Hadamard transform (CS-SCHT) coefficients to four minutes of ECG signals prior to SCD onset. The classification accuracy was 99.3% for the detection of SCD.
Acharya et al. , also proposed an ECG-based SCD prediction algorithm. They explored the feasibility of developing an integrated index for SCD prediction by ECG signal analysis. First, nonlinear features, including fractal dimension, Hurst’s exponent, detrended fluctuation analysis, approximate entropy, sample entropy, and correlation dimension from the second level discrete wavelet transform (DWT) of decomposed ECG signal were extracted. Then, SCD was predicted using the combination of DWT and nonlinear features. The accuracy of SCD detection was 92.11% for fourth minutes before its occurrence.
Commonly, nonlinear heart rate dynamics such as rapid spectral alterations and low frequency heart rate oscillations are appeared in patients at high risk of SCD . HRV signal analysis has demonstrated to be a strong and accurate mortality predictor after MI . Therefore, many researchers used features extracted from HRV signals to predict SCD. Ebrahimzadeh et al. , extracted time, frequency and time-frequency features of HRV signal and assessed them in order to predict SCD events. The accuracy of employed classifier was 99.16% by using ECG signals just prior to the SCD event.
This research group also presented the prediction of SCD using the combination of time-frequency and nonlinear features of HRV signals. The experimental results showed that the accuracy of their work for the first, second, and third one-minute intervals were 96.52%, 90.37% and 83.96%, respectively .
Murukesan et al. , showed how the combination of time, time-frequency, and nonlinear features from five-minute HRV signal (two minutes before SCD onset) provided satisfying performance with prediction rate of 96.36%.
As another example of usage of HRV as the SCD predictor we can point to study that has been performed by Acharya et al. . They predicted the occurrence of cardiac death, three minutes prior to its onset by means of recurrence quantification analysis (RQA) and Kolmogorov complexity parameters. The accuracy of their algorithm was 86.8%. In another work, they , also extracted various nonlinear features such as Hjorth’s parameters, fuzzy entropy, Tsallis entropy, Renyi entropy, and energy features of DWT coefficients for classification of HRV signals into normal and at the risk of SCD. They reached the accuracy of 94.7% in this recent research.
Murugappan et al. , predicted SCD events five minutes before its onset by exploiting time-domain features. The results proved that an accuracy of 93.71% was obtained using fuzzy classifies.
The main drawbacks of time domain and frequency domain techniques are inadequate representation of data and inaccurate time resolution, respectively [17, 22-24]. As ECG or HRV signals have nonlinear and non-stationary characteristics, linear algorithms will not provide enough information for accurate SCD prediction . Therefore, this work focuses on SCD detection using nonlinear methods to address these drawbacks and decipher the hidden information related to SCD in the HRV signals.
The block diagram of our proposed method is shown in Fig. 1. Nonlinear features are extracted from the one-minute interval of normal and SCD HRV signals. Then, the features with significant information (obtained by using one-way ANOVA method) are used for classification of normal and SCD risk groups by employing classifiers such as K-Nearest Neighbor (KNN), Support Vector Machine (SVM), Decision Tree (DT), and Naive Bayes.
Figure 1. Structure of the proposed method.
2. Material and Methods
In this study, the open source MIT-BIH SCD Holter database was used to obtain the SCD signals at a sampling frequency of 256 Hz and MIT/BIH Normal Sinus Rhythm database was utilized to acquire normal ECG signals at a sampling frequency of 128 Hz . Subsequently, in order to match normal sinus rhythm signals with SCD data, the ECG signals of SCD patients were downsampled to 128 Hz. Normal Sinus Rhythm database includes 18 ECG signals (ages ranging from 20 to 50 years) with no significant arrhythmias, while the MIT-BIH SCD Holter has ECG recordings acquired from 23 SCD patients (age: 18-89 years). Furthermore, only twenty SCD ECG signals were used for analysis, and the remaining three ECG signals were not used, as they did not have the VF episodes. It is reminded that in some cases that we have access to two leads of ECG recording, each lead was used as an input observation.
Fig. 2 illustrates the extraction procedure of ECG signals from SCD recordings in MIT/BIH database. The ECG signals were extracted at one-minute duration, which was six minutes earlier to the onset of SCA. Since, the ECG signals of normal subjects do not exhibit any pathological effects on ECG signals, one-minute episode ECG signals were captured randomly. Altogether, a total of 40 ECG recordings (for two leads) of SCD data and 36 ECG recordings of normal subjects (for two leads) was used to predict SCD. In this work, we used Pan-Tompkins algorithm to extract the HRV signals .
Figure 2. Extraction of ECG signals from SCD recording in MIT/BIH database.
In this section, nonlinear features which were used for SCD prediction is discussed. Since, the biological signals are naturally non-stationery and complex, nonlinear approaches are the most efficient tools for their analysis. Consequently, in this work, we investigated nonlinear algorithms, including RQA and increment entropy (IncEn) for feature extraction.
2.3.1 Recurrence Quantification Analysis
Recurrence plot (RP) is a non-linear dynamics measure, which provides functional information that is not achievable by others approaches . In order to visualize the time dependent behavior of the dynamics of systems, RP was introduced. In fact, the distance relationships between different points on a nonlinear and complex system could be revealed by RP plots. For one-dimensional time series
ui, we can reconstruct the phase space vectors,
xi∈R(i = 1, . . . , N), the th point in n-dimensional space, in this way :
In which the time delay
(τ)is estimated from mutual information approach and the dimension
(m )from the false nearest neighbors method [29, 30]. The RP is defined as an array of points in
N×Nsquare as following :
εiis the cut-off distance that defines a sphere centered at
|| . ||and
θ(x)are norm and the Heaviside function, respectively. In this case if
xj are adequately close, then
Zbilut et al. [32, 33], have quantified an RP using the RQA. They have used measures in the RP to define the recurrence point density, diagonal and vertical structures, recurrence rate, determinism, maximal length of diagonal structures, entropy and so on. Table.1 tabulates the measures that we used in this study.
Table 1. The features extracted from RQA.
Is the frequency distribution of the lengths
lof diagonal structures and
N lis the absolute number of diagonal lines.
Pv=vi;i…Nvdenotes the frequency distribution of the lengths
vof vertical structures.
Riare the recurrence points which belong to the state
P(t) is the recurrence period density function
twhere is the time difference between successive return.
Tmaxis the largest recurrence value.
2.3.2. Increment Entropy
The dynamic changes in complex signals can be characterized by IncEn. Hence, this measure quantifies the complexity of time series, effectively. Let
ai, i=1,…,Nbe a random vector and
λi,i=1,…,N-1be an increment vector from
λi=ai+1-a(i). Then we construct a vector
Sk+j=sgnλk+j, j=1,…, m-1as the sign and
pk+j , j=1,…,m-1, as the magnitude of each element in a vector
ϴ(k). Consequently, the IncEn can be defined as :
Where m (
m≥2) is the order,
Ris the resolution level and
zynis the relative frequency of any unique word
yk=⋃j=0N-1sk+jyk+j) as following:
Denotes the total amount of any unique word
2.4 Feature Ranking
In order to early predict of SCD occurrence, features with significant information are required. As all the features extracted from HRV signals do not contain these necessary details, feature ranking method is used for appropriate features selection. So, feature ranking technique ranks all features based on their significance. In this study, the one-way ANOVA is used as the ranking method to rank the extracted nonlinear features.
2.5. Methods for Classification
Classifiers of DT, KNN, Naive Bayes, and SVM are used to differentiate between the ECG of normal subjects and subjects at risk of SCD. In this work, 10-fold cross validation method is used to build classifiers and evaluate their performance. The features extracted from 76 ECG signals are divided into 10 parts, and each part consists of the equal number of signals of two classes. During the training phase, nine parts are used to train the classifier, and one part is used for testing the classifier. Then the performance measures are calculated according to test set results. This procedure is repeated ten times with different test set and the average performance measures such as accuracy, sensitivity and specificity are calculated based on all ten test results. The descriptions of different classifiers are given below.
2.5.1 Decision Tree
DT is a decision analysis tool that displays the possible consequences of a decision such as the outcome of the event, classes or class distributions and the resource costs . ID3, C4.5, and CART are the most popular types of DT algorithms. ID3 and C4.5 use information gain and gain ratio as the splitting criteria, respectively . Furthermore, for CART algorithm, Gini coefficient is used as the criteria of selecting the test attribute . Simple interpretation and understanding of DT is a benefit of using DT in data classification .
2.5.2 K-Nearest Neighbor
The KNN algorithm is an approach for data classification based on the votes of neighbors. The KNN is one of the simplest machine learning algorithms with low computation cost . This algorithm uses various types of distances, including Euclidean, the Chebyshev norm, the Manhattan distance, and the Mahalanobis. In this experiment, Euclidean and Minkowski distances were used.
2.5.3 Naive Bayes
Naive Bayesian classifiers are based on an assumption that there are not any dependency among features . This classifier has limited free parameters and uses an intuitive method which simplifies the design process tremendously. Accordingly, it is also called idiot Bayes, simple Bayes, or independent Bayes .
2.5.4 Support Vector Machine
This classifier constructs a separating hyper-plane in order to separate the training data into two groups . As nonlinear signals cannot be separated easily, kernel functions are used to convert them to a higher dimensional space. In this work, polynomial kernel functions of order 1, 2, 3 and 4 and radial basis function (RBF) kernels were used .
In this study, we totally extracted 14 features and ranked them from 6th one-minute interval of normal and SCD HRV signals. Table 2 presents the mean and standard deviation (SD) of these features. The value of F which demonstrates the classification ability, is very high in Laminarity and IncEn features. So, in this study these two features were selected for classification between SCD and normal subjects. Fig. 3 depicts the scattering of Laminarity and IncEn samples in two-dimensional feature space.
Table 2. Statistical significance of features using one-way ANOVA.
Figure 3. Scattering of samples in feature space.
The performance of the SCD classification is evaluated through the three following measures; accuracy, sensitivity, and specificity. Accuracy measures how many testing inputs are accurately classified into a respective group:
Sensitivity measures the proportion of true positive to the total positives:
Specificity measures the proportion of true negative to the total negatives:
Where TP: true positive; TN: true negative; FP: false positive; FN: false negative.
The performance of KNN, Naive Bayes, and SVM classifiers are shown in Tables 3, 4 and 5. We employed KNN classifier with various values for K and different distance metrics such as Euclidean and Minkowski. For Euclidean distance metric, we varied the value of K from 2 to 6 and achieved the best accuracy, sensitivity, and specificity for K=5. KNN classifier yielded the average classification accuracy of 93.571%, sensitivity of 92.5% and specificity of 95%. According to Table 3, different distance metrics did not change the performance of KNN classifier.
Table 3. Performance of KNN classifier.
In Naive Bayes classifier, the performance does not change by distribution alteration. This classifier achieved the highest average classification accuracy of 93.571%, sensitivity of 92.5% and specificity of 95%.
Table 4. Performance of Naive Bayes classifier.
Five kernel functions tested for SVM classifier, and the best performance obtained by polynomial kernel function with the degree of 2. In this case, the value for accuracy, sensitivity, and specificity were 93.571%, 92.5% and 95%, respectively. Finally, DT classifier yielded had the highest performance among all tested classifiers with classification accuracy of 95%, sensitivity of 95% and specificity of 95%.
Table 6 compares the result of our proposed SCD predication algorithm using HRV analysis with other recent researches. Based on this table, our approach has achieved the highest performance of SCD prediction, in comparison with previously reported works.
Table 5. Performance of SVM classifier.
Table 6. The comparison of our algorithm and other recent methods using HRV analysis.
SCD is still one of the prevalent and unresolved problems in clinical cardiology. In spite of the recent improvements in biological signal processing methods, the early prediction of SCD is still a challenging task. In fact, early SCD prediction can help in timely treatment and save many lives. In this work, an efficient algorithm is proposed for an automated prediction of SCD.
The unique feature of our work is the efficient prediction of SCD event, five minutes prior to its onset. We utilized IncEn and RQA-based features from the RPs to classify SCD subjects. For this purpose, first we obtained ECG signals (SCD and normal) from the MIT-BIH SCD Holter and the Normal Sinus Rhythm databases. Then, HRV signals are extracted from them and the IncEn and RQA-based features are obtained from HRV time series. After that, the one-way ANOVA is applied to the extracted features to rank them. This step results in reducing the dimension of feature space and identifying the features with significant information for accurate SCD prediction. Actually, through this procedure, we are ensured that only best features get fed to classifiers. It can be seen from Table 2 that the value of F is very high in Laminarity and IncEn features compared to other features. Therefore, we selected these two features for classification. Finally, healthy people and subjects at the risk of SCD, are classified using DT, KNN, Naive Bayes, and SVM. The best performance among these tested classifiers achieved by DT classifier. As a result, our algorithm can effectively predict the SCD occurrence, five minutes before its onset with an accuracy of 95%. Furthermore, in contrast with other works, our proposed method is required low computational power and time.
To best of our knowledge, this result is the best of its kind in an efficient prediction of SCD, five minutes prior to its onset. As a result, the clinicians have enough time to attend and treat the patients within these five minutes. As we access to a limited number of samples from the databases, this algorithm could not be generalized for larger population of samples. So, depending on the availability of new database, this algorithm can be implemented in real-time tele-monitoring of subjects at high risk inside or outside the hospital environment.