Automated Detection of Sudden Cardiac Death by Discrete Wavelet Transform of Electrocardiogram Signal

Abstract

Sudden cardiac death (SCD) results in millions of deaths annually; as it is a fatal heart abnormality, early prediction of SCD could save peoples’ lives to the greatest extent. Symmetry and asymmetry play an important role in many fields. Electrocardiograms (ECG) as a noninvasive process for acquiring the electrical activity of the heart, has both asymmetric and non-stationary characteristics; it is frequently employed to diagnose and evaluate the heart’s condition. In this work, we have detected SCD 14 min (separately for each one-minute interval) prior to its occurrence by analyzing ECG signals using discrete wavelet transform (DWT) and locality preserving projection (LPP). In the experiment, we have performed DWT on ECG signals to obtain coefficients, then LPP as a reduction methodology was used to cut down these obtained coefficients. Then, the acquired LPP features were ranked using various methods, including the T-test, Bhattacharyya, Wilcoxon, and entropy. At last, the highly ranked LPP features were subjected to decision tree, k-nearest neighbor (KNN), and support vector machine classifiers for distinguishing normal from SCD ECG signals. Our proposed technique has achieved a highest accuracy of 97.6% for the detection of SCD 14 min prior using the KNN classifier, compared to the existing works. Our proposed method is capable of predicting the people at risk of developing SCD 14 min before its onset, and, hence, clinicians would have enough time to provide treatment in intensive care units (ICU) for a subject at risk of SCD. Thus, this proposed technique as a useful tool can increase the survival rate of many cardiac patients.

1. Introduction

Sudden cardiac death (SCD) describes an unanticipated sudden death of a person who had or not had primary cardiovascular abnormalities, and known or even unknown history of cardiac diseases [1]. A number of 250,000 to 300,000 US lives per year have died from SCD [2]. For the past two decades, despite declines in death caused by cardiac disease [3], about 20% of all unexpected and sudden deaths are still reported to happen from SCD, most commonly resulted from asystole or ventricular fibrillation (VF) [4]; VF can result in the heart’s failure to effectively pump blood and hence can lead to death within a few minutes [5]. There are only 1–2% of out-of-hospital heart disease patients that survive when SCD happens, as most of subjects at the risk of SCD cannot receive timely care [6]. Therefore, early identification of an unanticipated SCD risk in a subject experiencing VF is very important for timely treatment and reducing mortality.

Ordinarily, symmetry represents a correspondence or balance between different parts of an object [3]. Symmetry and asymmetry are the fundamental aspects of life [7]. Electrocardiogram (ECG) signals demonstrate the cardiac electrical activity, and the morphology of the ECG signal varies with respect to time [8], it contains substantial vital information about the human body. The asymmetry of ECG signals is extremely significant to capture alarming features for the early identification of SCD. The morphological variations of the ECG signal are found when the heart undergoes episode of ischemia or arrhythmias [9]. Researchers worldwide have captured the subtle variation residing in the ECG signal to discern the unique characteristics which can early predict the risk of SCD occurrence. QT interval and heart rate variability (HRV) extracted from the ECG signal are proven to be an independent risk indicator to identify SCD caused by cardiac arrest [10]. However, it is time consuming and tedious to manually examine and identify these minute morphological variations in long continuous ECG signals, and this process is easy to result in errors because of fatigue. Therefore, an automated diagnostic technique for predicting SCD early is indispensable; it can overcome these shortcomings of manual estimation of ECG signals.

Three major methodologies, including classical linear, time-frequency (TF), and nonlinear methodologies, have been frequently used to analyze ECG or HRV signals. Mean of standard deviation (SD) and SD of the mean sinus R-R intervals as a time domain feature of the classical linear features, computed from HRV signals in normal groups, was higher than those in people at risk of SCD [11]. Various frequency domain features, such as low frequency, high frequency, and very low frequency, obtained from the analysis of HRV signals using the fast Fourier transformation method, were used to identify SCD [6]. Smoothed pseudo Wigner-Ville distribution and Wigner-Ville transform were subjected to the HRV signals to acquire the time-frequency features for SCD detection [2,12]. Additionally, the nonlinear methods such as symbolic dynamic [13], Fractal dimension, correlation dimension, approximate entropy, sample entropy, and so on, were employed to quantify complexities residing in ECG or HRV signals due to the nonlinear, non-stationary, and asymmetrical characteristics of these signals [14].

Automated diagnostic techniques for the early prediction of SCD using ECG or HRV signals have been put forward and are summarized in

Table 1. Overview of study using ECG/HRV signals for early SCD detection.

Ref.	Data	Method	Classification	Cross Validation	Accuracy
Early SCD Detection Using HRV Signals
[6]	No. of data: 23 SCD, 20 normal Source: the MIT/BIH SCD ans Normal Sinus Rhythm (NSR) databases Prediction time resolution: 2 min interval	Wavelet analysis and TF domain method	Artificial neural networks (ANN); Back propagation (BP)	Not mentioned	87.5% (2 min before)
[12]	No. of data: 35 SCD, 35 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	Linear (time, frequency domain), TF domain, and nonlinear methods (poincaré plot (PP), Detrended fluctuation analysis (DFA))	KNN; Multilayer perceptron (MLP)	Not mentioned	91.23% (2 min before)
[2]	No. of data: 35 SCD, 35 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	Linear (time, frequency domain), TF domain, and nonlinear methods (PP, DFA)	KNN; Multilayer Perceptron Neural Network (PNN)	Leave one out cross-validation	83.93% (4 min before)
[16]	No. of data: 23 SCD, 18 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	Linear (time, frequency domain), TF domain, and nonlinear methods (PP, DFA)	SVM, PNN	10-fold cross-validation	96.36% (2 min before)
[14]	No. of data: 40 SCD, 36 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	Recurrence Quantification Analysis (RQA), and Kolmogorov, complexity parameters	SVM, KNN, PNN, DT	10-fold cross-validation	86.8% (4 min before)
[17]	No. of data: 19 SCD and18 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	Linear (time, frequency domain), TF domain, and nonlinear methods (PP)	SVM	10-fold cross-validation	83.24% (1 min before)
[1]	No. of data:20 SCD and18 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	Wavelet transform, nonlinear methods (Renyi entropy (REn), fuzzy entropy (FuEn), Hjorth’s parameters, Tsallis entropy)	DT, KNN, SVM	10-fold cross-validation	94.7% (4 min before)
[18]	No. of data: 35 SCD, 35 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	Linear (time, frequency domain), TF domain and, nonlinear methods (PP, DFA)	MLP	Leave one out cross-validation	95.23% (4 min before) 83.88% (12 min before)
[19]	No. of data: 35 SCD, 35 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	Linear (time, frequency domain), TF domain, and nonlinear methods (PP, DFA)	MLP, SVM, KNN	Not mentioned	95.24% (4 min before) 84.28% (13 min before)
[20]	No. of data: 40 SCD, 36 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	EEMD, linear (time, frequency domain), TF domain, and nonlinear methods (REn, FuEn, dispersion entropy, Renyi distribution entropy (RdisEn), and improved multiscale permutation entropy	KNN	10-fold cross-validation	94.7% (4 min before) 96.1% (14 min before)
Early SCD detection using ECG signals
[21]	No. of data: 40 SCD, 36 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	DWT, Nonlinear methods (Fractal Dimension (FD), DFA Hurst’s exponent (H), Sample Entropy, Approximate Entropy, and Correlation Dimension (CD))	KNN, SVM, DT	10-fold cross-validation	92.11% (4 min before)
Current study	No. of data: 40 SCD, 36 normal Source: the MIT/BIH SCD ans NSR databases Prediction time resolution: 1 min interval	DWT and LPP	KNN, SVM, DT	10-fold cross-validation	98.1% (4 min before) 97.6% (14 min before)

Note: Sen = sensitivity, Spe = specificity, Acc = accuracy, No. of data = number of data.

; it can be seen from the literature review (Table 1) that most researchers employed the HRV signals with a 1 min interval of a prediction time resolution, and linear (time, frequency domain), TF domain, and nonlinear methods including poincaré plot, detrended fluctuation analysis, sample entropy (SamEn), approximate entropy (ApEn), etc., were used for the analysis of HRV, but the recommended duration of HRV for the frequency domain and nonlinear (SamEn, ApEn, etc.) methods is 5 min; employing more short-term intervals (1 min interval) can cause unreliability of the frequency domain method and the undefined values of the classical entropy algorithm (SamEn, ApEn, etc.) [15], thereby resulting in the instability of these automated diagnostic techniques.

Therefore, we proposed an automated diagnostic technique using ECG beats with 651 data points for automated identification of subjects at risk of SCD one to fourteen minutes before its occurrence. Figure 1 shows the block diagram of the proposed method. Initially, discrete wavelet transform (DWT) is performed on normal and SCD ECG beats to obtain coefficients. The locality preserving projection (LPP) reduction method is then applied to the coefficients obtained from DWT, to extract the LPP features with significant information. Later, the reduced LPP features are ranked using various methods, including the T-test, Bhattacharyya, Wilcoxon, and entropy. Finally, the ranked LPP features are subjected to decision tree (DT), k-nearest neighbor (KNN), and support vector machine (SVM) classifiers for classification of SCD risk and normal groups. The proposed system yielded a highest accuracy of 97.6% for predicting if the person is susceptible to SCD or not 14 min earlier, using only five LPP features.

2. Materials and Methods

2.1. Data Acquisition

The SCD and normal ECG signal data used for our experiment were downloaded from two open access databases, namely the MIT/BIH SCD and PhysioBank MIT-BIH Normal Sinus Rhythm (NSR) databases [22]. The SCD database contains 23 patients, including 8 females, 13 males and 2 of unknown sex (age: 18–89). Among the 23 SCD patients, 18 patients had underlying sinus rhythm, 1 was continuously paced, and 4 had atrial fibrillation. Most of these patients had an actual cardiac arrest, and all patients had a sustained ventricular tachyarrhythmia. Moreover, only 20 subjects were employed for further analysis in the current study, due to the ECG signals of the other 3 SCD patients did not exhibit any VF episodes. The normal database contains 18 subjects, including 13 females (age: 20–50) and 5 males (age: 26–45). Each SCD record contains leads Ⅱ and Ⅲ, and each normal record contains leads Ⅰ and Ⅱ ECG signals. Therefore, a total of 36 ECG records obtained from the MIT-BIH NSR database and 40 SCD ECG records obtained from the MIT/BIH SCD database were used.

2.2. Pre-Processing and Beats Segmentation

For 24 h of ECG records of the SCD subjects, only the ECG records 14 min prior to VF onset were utilized, for simulating 14 min before SCD occurrence. As for the normal subjects, 14 min of the ECG records were selected arbitrarily. The ECG records from the two databases were sampled at 250 Hz for SCD and 128 Hz for normal classes [22]. The ECG records of the two databases were sampled up to 1000 Hz using the cubic interpolation method, for the purpose of maintaining uniformity between the two databases. Because there are various noises, such as power line interference (>50 Hz) and baseline wander (<0.5 HZ) in the ECG records, DWT with Daubechies wavelet 6 (db6) was performed on the ECG records with the ECG records decomposed to 11 levels, in order to denoise the noises [9]. The highest-level approximation coefficients and the first three detail coefficients were set to zero, and the inverse wavelet transform was calculated from the fourth to the eleventh level detail sub-bands to acquire the denoised ECG records.

These 14 min ECG segments were uniformly divided into five 1 min intervals (i.e., the 1st 1 min, 2nd 1 min, 3rd 1 min, 4th 1 min, etc.); the 1st 1 min prior to SCD onset was referred to as the first 1 min interval, the 2nd 1-min prior to SCD onset was referred to as the second 1 min interval, and so on. Figure 2 displays a typical SCD ECG signal 1 to 4 min intervals before its onset.

Additionally, the Pan-Tompkins method was performed on the preprocessed 1 min ECG intervals in order to detect R-peak for the ECG beat segmentation [23]. The R-peak was selected as the unusual point due to its high amplitude. In this experiment, all the 1 min ECG intervals were segmented into the ECG beats, using the selected R-peaks, without including the first and last beats. Numbers of 250 and 400 samples before and after the R-peak were taken to perform the segmentation of ECG beats. The overview of the number of ECG beats used in each group and case is demonstrated in

Table 2. Summary of the ECG beats used in fifth cases.

Case	Number of Beats
	SCD	Normal
1st 1-min	2975	2934
2nd 1-min	2849	2938
3rd 1-min	2925	3036
4th 1-min	3112	3021
5th 1-min	3013	3100
6th 1-min	2923	3052
7th 1-min	3001	3043
8th 1-min	3038	2961
9th 1-min	2908	3032
10th 1-min	2904	3135
11th 1-min	2907	3175
12th 1-min	2758	3147
13th 1-min	2865	3062
14th 1-min	2861	3026

. Figure 3 shows the SCD and normal beats segmented from the continuous ECG intervals.

2.3. Feature Extraction-Discrete Wavelet Transformation

Wavelet transformation is a signal processing technique, employing mathematical models to uncover the characteristic of the analyzed signal [24]. The assumption of this technique is that the analyzed signal consists of a set of analogously small functions, namely ‘mother wavelet’, composed of shifting and scaling. Detailed explanation of the discrete wavelet transformation method is stated below.

$${\psi }_{a,b}(t)=\frac{1}{\sqrt{a}}\psi \left(\frac{t-b}{a}\right)$$

(1)

where $\psi (t)$ is function of mother wavelet, t is the time, $a$ and $b$ are the scale and shifting parameters, respectively. The discrete wavelet transforms (DWT) are obtained by the scale and shifting position of mother wavelet and can be computed using

$$DWT(m,n)=\frac{1}{\sqrt{a_0^m}}{\displaystyle \sum _{k}f(k)}\psi \left[\frac{n-k{b}_0{a}_0^m}{a_0^m}\right]$$

(2)

where m and k are the coefficients of the scaling and shifting, n is number of points [24]. The Mallat algorithm was proposed in 1988 and uses a filter to perform DWT; it is an effective method. The low-pass and high-pass filters adopted in the Mallat algorithm filter out the high-frequency and low-frequency components of a signal analyzed, respectively, in order to acquire the detail (high-frequency) and the approximation (low-frequency) coefficients; the detail and approximation coefficients obtained from the wavelet decomposition procedure are denoted by $a_i$ and $d_i$, where i represents the level of the decomposition. Figure 4 shows a five-level decomposition of DWT for the ECG beats, extracted from normal and patients at the risk of SCD, respectively. Therefore, each SCD ECG and normal beat obtained from the four cases were subjected to five levels of DWT with Daubechies 6 (db6) mother wavelet to obtain their coefficients, respectively. A total of 31 detail coefficients at the fifth level from each ECG beat were extracted for feature extraction [25]. The acquired subband coefficients were connected with the kind of wavelet basis, so DWT with Daubechies (orders 1–5), Coiflets (orders 1–3), and Harr wavelet functions with 5-level decomposition were also were explored for a better classification accuracy on SCD prediction.

2.4. Features Reduction—Locality Preserving Projections (LPP)

LPP is a dimensionality reduction methodology. It constructs an adjacency map using data maps, the data points analyzed are mapped to the subspace in order to acquire the transformation matrix, employing the Laplace transform of the graph, which attains the goal of reducing dimensions [26]. The data set $X=\left\{x_1,x_2,\cdots ,x_n\right\}$ with high-dimensional space $R^n$ is mapped to the data set $Y=\left\{y_1,y_2,\cdots ,y_n\right\}$ with low-dimensional space $R^m\hspace{0.17em}(m<<n)$, after the transformation matrix W

$$y_i=W^T{x}_i, i=1,2,\cdots ,k$$

(3)

as $x_i$ is the k nearest neighbor of $x_j$, similarity matrix $S_{ij}$ is computed by $S_{ij}=e^{-\frac{{\Vert x_i-x_j\Vert }^2}{t}}$, otherwise $S_{ij}$ = 0. The LPP can be obtained by working out the minimum problem:

$${\partial }_{opt}=\arg \min {\alpha }^{T}XL{X}^T\alpha$$

(4)

where ${\alpha }^{T}XD{X}^T\alpha =1$, $D_{ii}$ is the diagonal weight matrix and its element is calculated by the summation of the row element of $S_{ij}$, the Laplacian matrix $L$ is computed by $D_{ii}-S_{ij}$. Transformation matrix W is the following formula:

$$\min ({{\displaystyle \sum _{i,j}(W^T{x}_i-W^T{x}_j)}}^2{S}_{ij})$$

(5)

Furthermore, it can be used to solve the eigenvalue problem

$$XL{X}^{T}W=\lambda XD{X}^{T}W$$

(6)

where $\lambda$ is the eigenvalue [26].

The feature extraction procedure generally causes high-dimensional features, and thus results in many difficulties to machine learning tasks; LPP as the dimensionality reduction methodology was used in this paper, aiming to find few low-dimensional features that are representative for the high-dimensional features. It can effectively improve learning speed and save memory space [27].

2.5. Feature Ranking

Not all LPP features obtained from the feature reduction procedure have important information to effectively discriminate the patients with SCD from unaffected people. In current work, we used the T-test to rank the LPP features obtained, in order to reserve features with important clinical information. The T-test algorithm can determine whether the LPP features obtained have statistical significance, and the features are ranked according to their p-values, generated by the T-test algorithm; the p-value less than 0.05 is considered as statistically significant and lower p-values demonstrate better ranking. Additionally, the other three ranking methods such as Bhattacharyya, Wilcoxon, and entropy were explored for a better separability between the two groups [21].

2.6. Classification

The ranked LPP features were input to the classifiers one by one to acquire the classification accuracy with minimum number of features. In this work, KNN, SVM with Gaussian kernel function, and DT were used for the sake of exploring the best classifier to differentiate unaffected from SCD-affected ECG beats; two kinds of KNN (k = 1.10 denoted by 1-NN and 10-NN) were employed in current work. We adopted accuracy, specificity, and sensitivity for purpose of assessment of the classifiers. Furthermore, 10-fold cross-validation methodology was employed to assess the performance of the classifiers. These normal from SCD affected ECG beats in each case, were divided into ten parts, and each part had the same samples from the two groups. One part was employed for testing the classifier and the rest of the 9 parts were employed for training the classifier. The performance indicators such as accuracy, sensitivity, and specificity, based on the results of testing data, were then computed [21]. This process was repeated 10 times, utilizing a different testing data each time, and the average of performance indicators from all the 10 test results were computed

3. Results

In current work, a total of 83,701 ECG beats including 41,039 SCD and 42,662 normal, were segmented from normal and SCD groups. All ECG segments were obtained from 14 cases, such as (1) 2934 normal and 2975 SCD segments from normal and SCD 1 min before its onset, (2) 2938 normal and 2849 SCD segments from normal and SCD 2 min before its onset, (3) 3036 normal and 2925 SCD segments from normal and SCD 3 min before its onset, (4) 3021 normal and 3112 SCD segments from normal and SCD 4 min before its onset, (5) 3100 normal and 3013 SCD segments from normal and SCD 5 min before its onset, (6) 3052 normal and 2923 SCD segments from normal and SCD 6 min before its onset, (7) 3043 normal and 3001 SCD segments from normal and SCD 7 min before its onset, (8) 2961 normal and 3038 SCD segments from normal and SCD 8 min before its onset, (9) 3032 normal and 2908 SCD segments from normal and SCD 9 min before its onset, (10) 3135 normal and 2904 SCD segments from normal and SCD 10 min before its onset, (11) 3175 normal and 2907 SCD segments from normal and SCD 11 min before its onset, (12) 3147 normal and 2758 SCD segments from normal and SCD 12 min before its onset, (13) 3062 normal and 2865 SCD segments from normal and SCD 13 min before its onset, (14) 3026 normal and 2861 SCD segments from normal and SCD 14 min before its onset (Table 2).

Take the fourth case as an example; the five levels DWT with db6 wavelet basis was performed on each ECG beat to acquire 31 detailed coefficients on the fifth. Amounts of 31 × 3021 and 31 × 3112 DWT coefficients were obtained from the two groups. Furthermore, the LPP methodology was used to reduce the obtained DWT coefficients. Subsequently these LPP features were ranked by their p-values, acquired from the T-test, in descending order, and there exists significant differences for these ranked LPP features between the two classes, as shown in Figure 5.

Table 3. The mean ± SD of LPP features from normal and SCD ECG signals 4th 1 min before SCD onset.

Feature	Normal		SCD		p-Value
	Mean	SD	Mean	SD
LPP2	−0.032	0.475	0.749	1.307	5.52 × 10−199
LPP1	0.443	0.843	1.567	2.457	1.05 × 10−122
LPP3	−0.118	0.294	0.104	0.539	4.04 × 10−87
LPP4	−0.072	0.444	0.082	0.257	3.33 × 10−60
LPP5	−0.070	0.330	−0.164	0.321	3.57 × 10−29
LPP6	−0.010	0.179	0.043	0.259	2.00 × 10−20
LPP11	−0.060	0.172	−0.085	0.127	4.88 × 10−10
LPP7	0.011	0.379	−0.020	0.146	2.32 × 10−5
LPP10	−0.061	0.216	−0.071	0.119	0.032
LPP8	0.019	0.350	0.012	0.115	0.293
LPP9	−0.011	0.236	−0.011	0.149	0.934

tabulates the mean and standard deviation (SD) of LPP features computed from normal and SCD ECG beats 4 min before the SCD occurrence. Ranked LPP features were subjected to various classifiers, such as KNN(1-NN, 10-NN), SVM, and DT, one by one, to achieve the highest classification accuracy, and the classification results are shown in

Table 4. Performance measures for normal and SCD HRV signals four minutes before SCD onset using various classifiers.

Classifier	Features	TP	TN	FP	FN	Accuracy	Sensitivity	Specificity
KNN (1)	5	3054	2962	58	59	98.1%	98.1%	98.0%
KNN (10)	5	3050	2959	62	62	98.0%	98.0%	97.9%
SVM(Gaussian)	5	3043	2960	69	61	97.9%	97.8%	98.0%
DT	8	3012	2904	100	117	96.5%	96.8%	96.1%

Note: TP = True Positive; TN = True Negative; FP = False Positive; FN = False Negative.

. Figure 6 illustrates the plot of accuracy (percentage) versus the number of LPP features employed for the classification of normal and SCD ECG beats. It is evident from Figure 6 that the 1-NN classifier achieves the highest accuracy of 98.5% using only seven LPP features, and outperforms other classifiers. In order to explore the influence of different wavelet basis, such as Daubechies (orders 1–6), Coiflets (orders 1–3), and Harr, various ranking methods were used, such as the T-test, Bhattacharyya, Wilcoxon, and entropy, for the best classification results.

Table 5. Performance measures for normal and SCD HRV signals four minutes before SCD onset using various ranking methods and the 1-NN classifier.

Ranking Method	Features	TP	TN	FP	FN	Accuracy	Sensitivity	Specificity
t-test	5	3054	2962	58	59	98.1%	98.1%	98.0%
Bhattacharyya	5	3042	2838	70	83	97.5%	97.8%	97.2%
Wilcoxon	8	3059	2945	53	76	97.9%	98.3%	97.5%
entropy	6	3050	2941	62	80	97.7%	98.0%	97.4%

Note: TP = True Positive; TN = True Negative; FP = False Positive; FN = False Negative.

and

Table 6. Performance measures for normal and SCD HRV signals four minutes before SCD onset using DWT with various wavelet basis and the 1-NN classifier.

Wavelet Basis	Features	TP	TN	FP	FN	Accuracy	Sensitivity	Specificity
db1	7	3062	2944	50	77	97.9%	98.4%	97.5%
db2	12	3067	2942	45	79	98.0%	98.6%	97.4%
db3	12	3058	2954	54	69	98.0%	98.3%	97.7%
db4	9	3020	2911	92	110	96.7%	97.0%	96.4%
db5	7	3053	2940	59	81	97.7%	98.1%	97.3%
db6	5	3054	2962	58	59	98.1%	98.1%	98.0%
haar	8	2976	2862	136	159	95.2%	95.6%	94.7%
coif1	7	3058	2954	54	69	98.0%	98.3%	97.7%
coif2	5	3046	2955	66	66	97.8%	97.9%	97.8%
coif3	10	3041	2920	71	101	97.2%	97.7%	96.7%

Note: TP = True Positive; TN = True Negative; FP = False Positive; FN = False Negative.

summarize the performance of different wavelet basis and ranking methods, using the 1-NN classifier for the categorization of normal and 4 min before SCD ECG beats. Obviously, the highest classification accuracy was achieved using db6 wavelet basis and the T-test ranking method. Therefore, we adopted the db6 wavelet basis and T-test ranking method for the rest of the thirteen cases for SCD identification.

Table 7. The maximum accuracy obtained for all 14 cases.

Case	Features	TP	TN	FP	FN	Accuracy	Sensitivity	Specificity
The 1st 1-min	9	2875	2816	100	118	96.3%	96.6%	96.0%
The 2nd 1-min	7	2771	2814	78	124	96.5%	97.3%	95.8%
The 3rd 1-min	7	2848	2915	77	121	96.7%	97.4%	96.0%
The 4th 1-min	5	3054	2962	58	59	98.1%	98.1%	98.0%
The 5th 1-min	6	2957	3009	56	91	97.6%	98.1%	97.1%
The 6th 1-min	4	2797	2876	126	176	94.9%	95.7%	94.2%
The 7th 1-min	4	2914	2886	87	157	96.0%	97.1%	94.8%
The 8th 1-min	6	2973	2869	65	92	97.4%	97.9%	96.9%
The 9th 1-min	6	2842	2918	66	114	97.0%	97.7%	96.2%
The 10th 1-min	6	2844	3033	60	102	97.3%	97.9%	96.7%
The 11th 1-min	4	2828	3056	79	119	96.7%	97.3%	96.3%
The 12th 1-min	6	2688	3072	70	75	97.5%	97.5%	97.6%
The 13th 1-min	7	2811	2997	54	65	98.0%	98.1%	97.9%
The 14th 1-min	5	2800	2944	61	82	97.6%	97.9%	97.3%

Note: TP = True Positive; TN = True Negative; FP = False Positive; FN = False Negative.

tabulates the confusion matrix for all 14 cases using the 1-NN classifier, with different numbers of the LPP features ranked by the T-test method. The technique proposed in this paper achieved an accuracy of 97.6% for 14 min prior to SCD occurrence. Only 61 SCD ECG beats were wrongly categorized into normal in the fourteenth case. This obviously demonstrates that number of false positive is very low and our technique is robust.

4. Discussion

Symmetry and asymmetry coexist as fundamental forms among mathematical structures; asymmetry drives evolution and the generation of complexity [28]. ECG as an asymmetrical signal is most frequently used for reflecting the cardiac electrical activity, it provides significant information to detect heart’s condition for clinicians. DWT is a multiresolution methodology, and has capability of breaking down the ECG signal being analyzed into several frequency sub-bands at various resolutions. This decomposition technique generates fewer coefficients without losing the information contained in the signal during both wavelet and inverse transformation [29]. In this work, A novel method for the early detection of SCD using DWT of the ECG signal is proposed. However, the performance of the DWT methodology is closely correlated with selection of the basis function and its similarity to the signal analyzed. Therefore, we explored the performance of several of the wavelet basis functions (Daubechies (orders 1–6), Coiflets (orders 1–3), and Harr) on prediction of developing SCD. Numerical experiments support the proposed technique; using DWT with db6 can distinguish normal from SCD ECGs 4 min before SCD onset with an accuracy, sensitivity, and specificity of 98.1, 98.1, and 98.0%, respectively, and perform better than the others’ basis function. The results of the studies on the early SCD identification using ECG/HRV analysis is provided in Table 1.

Shen et al. used four HRV indices (heart rate variability mean, very high frequency, low frequency/high frequency, and standard deviation) for analysis of two-minute HRV signal beats and achieved an accuracy of 87.5% [6]. Acharya et al. proposed a technique for SCD detection based on DWT and non-linear analysis, such as Fractal dimension, Hurst’s exponent, detrended fluctuation analysis (DFA), approximate entropy, sample entropy, and correlation dimension of ECG signals with a 1 min interval; they obtained an accuracy of 92.11, 98.68, 93.42, and 92.11% for one to four minutes before SCD occurrence, respectively [21]. They also used recurrence quantification analysis and Kolmogorov complexity features acquired from HRV with 1 min intervals to automatically detect SCD four minutes prior, with an average accuracy of 86.8% employing the KNN classifier [14]. Additionally, Elias et al. proposed several SCD detection methods [2] using time domain features, such as the mean of all NN intervals (MNN), root mean square of successive differences, frequency domain features, such as high frequency, low frequency, and ratio of LF and HF bands’ power, as well as TF features obtained from the Wigner-Ville transformation of the HRV signal and nonlinear features extracted from HRV signals with 1 min interval; their methods can predict SCD four minutes before SCD onset [2].

It can be seen from Table 1 that, 1 min and 2 min intervals of the analyzed signals (ECG or HRV) have been frequently used as a prediction resolution for early SCD identification. It is obvious that the bigger interval, such as over 2 min intervals, will increase the prediction accuracy due to having more samples. However, it will correspondingly result in a decrease of the prediction resolution. In addition to that, it becomes apparent that almost all studies tabulated in Table 1 used the analysis of HRV signals with a 1 min interval, but for time and frequency metrics of HRV analysis, it requires about 5 min of heart beat measurements to get reasonable and reliable results [30]. The number of HRV signals with 1–2 min is no more than 200 data points, but for some nonlinear methods used in SCD detection mentioned above, such as approximate entropy, sample entropy, and correlation dimension, too short of a time series results in a highly unstable results due to their great sensitivity to the length of the signal analyzed, which that discounts the clinical value of these methods [31]. Therefore, in this ongoing study, we developed an automatic SCD identification scheme combining DWT and LPP using ECG beats with 651 points in each interval, instead of HRV signals with 1 min interval, that is to say, this proposed scheme in this paper can be utilized to detect patients at risk of SCD using a single ECG beat. When compared with the prediction time of most research being 4 min at the most, the proposed technique can efficiently predict SCD occurrence risk 14 min earlier and achieved significantly good results, with the highest accuracy of 97.6% when compared with the state-of the-art studies. Furthermore, the 10-fold cross-validation methodology made our developed technique more robust and reliable. The proposed technique is completely automated, economical, and non-invasive, it can be used in the real-time monitoring of heart conditions with the development of cloud computing, smart phones, and wearable sensors [32], and can increase the survival rate of patients at risk of SCD.

5. Conclusions

Providing accurate early warning information for impending likely SCD, can trigger preemptive action for patients at risk of SCD, which would be extremely important and of great clinical meaning. In this study, an efficient technique was developed for an automated prediction of SCD 14 min before its onset using DWT of ECG signals. Our study result showed 98.1% accuracy, 98.1% sensitivity, and 98.0% specificity, as well as 97.6% accuracy, 97.9% sensitivity, and 97.3% specificity by employing the LPP features extracted from the asymmetric ECG signal in the SCD prediction 4 and 14 min earlier, respectively, which outperformed all the reported previous studies. The proposed technique can be tested using huge diverse database in the future.