Statistical analysis of ST segments in ECG signals for detection of ischaemic episodes

Abstract

This paper highlights a new method for the detection of ischaemic episodes using statistical features derived from ST segment deviations in electrocardiogram (ECG) signal. Firstly, ECG records are pre-processed for the removal of artifacts followed by the delineation process. Then region of interest (ROI) is defined for ST segment and isoelectric reference to compute the ST segment deviation. The mean thresholds for ST segment deviations are used to differentiate the ischaemic beats from normal beats in two stages. The window characterization algorithm is developed for filtration of spurious beats in ischaemic episodes. The ischaemic episode detection is made through the coefficient of variation (COV), kurtosis and form factor. A bell-shaped normal distribution graph is generated for normal and ischaemic ST segments. The results show average sensitivity (S_e) 97.71% and positive predictivity (+P) 96.89% for 90 records of the annotated European ST-T database (EDB) after validation. These results are significantly better than those of the available methods reported in the literature. The simplicity and automatic discarding of irrelevant beats makes this method feasible for use in clinical systems.

Keywords

ECG myocardial ischaemia QRS complex detection ST segment statistical analysis wavelet transform

Introduction

Cardiovascular deaths (CVDs) are projected the leading cause of deaths globally. It is likely that 17.3 million people died because of CVD in 2013, representing 31.5% of total global deaths. The expected deaths due to CVD will increase to 23.3 million by 2030. With the increasing sedentary lifestyle, more people are becoming prone to CVD. Important modulators of CVD include unhealthy diet, lack of exercise, consumption of snuff, alcohol, diabetes, raised lipids and stress (WHO, 2014). As myocardial ischaemia is the forerunner of CVD, so improvement of CVD management is a primary task that requires advances in its diagnosis, treatment and prevention. Ischaemia is the result of plaque building and accumulation of cholesterol in the inner walls of the arteries of the heart. As it progresses, a smaller amount blood flows through the arteries; the heart muscle cannot receive sufficient blood and oxygen, which usually lead to chest pain (angina), myocardial ischaemia (MI) or sometimes infarction (Brownfield and Herbert, 2008; Dash, 2002; Kusumoto, 2004). The most common method for investigating ischaemia is through ECG. The ECG is a process of interpretation of the electrical activity of the heart over a period of time. ECG presents the depolarization and repolarization of the cardiac muscle (Singh, 2010). The ECG is comprised of P, QRS complex and T waves. The most important ECG change caused by myocardial ischaemia includes ST deviation (elevation or depression); an example is shown in Figure 1. Elevation and depression of the ST segment with respect to isoelectric reference in ECG signal is the investigation of myocardial ischaemia around the applied lead. Elevated ST segments (or episodes) appear usually in patients with transmural (sub-epicardial) ischaemia, whereas depressed ST segments (or episodes) appear in patients with sub-endocardial ischaemia (Goldberger, 1981; Park et al., 2012). During recording of ECG, the artifacts, i.e. anything other than muscular activity of heart, are contaminated on the ECG. Commonly noticed artifacts include baseline wanders and muscle tremors. The reasons may be a loose electrode connection, patient movements, etc. These artifacts lead to misdiagnosis of ST segments. Removal of these artifacts is primarily required to facilitate easy, accurate and automatic detection of ST segment deviations in pathological cases (Gacek and Pedrycz, 2012). A number of methods have been developed, which may have linear or non-linear structures for the removal of the artifacts and detection of characteristic points of ECG signals. These methods are broadly divided into four categories – time-based, frequency-based, time–frequency-based and non-linear methods (Li and Wang, 2013; Li et al., 2014, 2015; Martis et al., 2014). Similarly, advanced methods like KICA non-linear, wavelet packet decomposition (WPD) and approximate entropy (ApEn) have been implemented for extraction of ECG features (Li et al., 2016a, 2016b). Time-frequency based, wavelet transform, can provide discrimination between two different signals with the same spectrum magnitude. Hence, wavelet transform is found more appropriate for analysis of non-stationary ST segments in ECG signal via wavelet coefficients thresholding (Beena and Prabavathy, 2014; Bianchi et al., 2000). A number of methods have been reported in literature for the detection of ischaemic beats and episodes, including support vector machine (SVM), kernel density estimation (KDE) (Gacek and Pedrycz, 2012), neural network (Baxt, 1991; Maglaveras et al., 1998; Papaloukas et al., 2002; Silipo and Marchesi, 1998; Silipo et al., 1994; Stamkopoulos et al., 1992, 1998), the self organizing map (SOM) model (Papadimitriou et al., 2001), KLT (Afsar et al., 2008; Jager et al., 1992), genetic algorithms (Papaloukas et al., 2004), hidden Markov models (HMM) (Andreao and Dorizzi, 2004), fuzzy models (Exarchos et al., 2007; Tsipouras and Costas, 2007; Vila et al., 1997; Zahan, 2001), RMS series (Garcia et al., 2000), wavelets (Gramatikov et al., 2000; Lemire et al., 2000; Martis et al., 2013; Ranjith et al., 2003; Taddei et al., 1995), the ant-miner algorithm (Murugan and Radhakrishnan, 2010), isoelectric energy (Kumar and Singh 2016a), statistical analysis (Kumar and Singh, 2016b; Kumar and Sonam, 2016), hermite functions (Rocha et al., 2010), and fractal and statistical features (Don et al., 2013). These methods have their own advantages and disadvantages. Similarly, statistical features using wavelet power spectra for normal and abnormal beats have been identified by Hegde et al. (2013). Real-time detection of myocardial infarction in digital ECG is presented by Kindi and Reza (2011). A survey of ischaemia detection methods has been performed by Manocha and Singh (2011). The presented paper contributes a simplified method for the detection of ischaemic episodes based on statistical analysis, which does not involve any complicated calculation and training. Based on statistical features, additional morphological feature of ST segments may be analysed. The paper is organized in nine sections; the first section introduces the ECG and literature survey for ischaemia, the second discusses deals with resources and methods, the third covers methodology, the fourth covers classification of beats, the fifth divulges episode recognition, the sixth presents statistical analysis, the seventh discusses the results, the eighth makes comparison with existing methods and the last section covers conclusions and future work.

Resources and methods

European ST-T database

The European ST-T Database (EDB) records are projected to be used for performance evaluation of the developed method for validation of the results. This database contains 90 ECG recordings; beat-by-beat annotation is mentioned for 79 subjects. The database contains records of two leads acquired from V₁–V₅, MLI, MLIII and D leads. The database has total 367 ischaemic episodes, corresponding to ST segment change in the durations between 30 s and several minutes. Each record contains 2-h recording of two channels with 12-bit resolution over a nominal 20 mV input range and a sampling frequency of 250 samples/s. The sample values were re-scaled after digitization, in order to obtain a consistent scale of 200 ADC units/mV for all ECG signals. Three independent cardiologists were appointed to annotate beat by beat in each record for changes in ST segment. The 90 records altogether contain 802,866 annotations (Taddei et al., 1992).

Wavelet transform

The wavelet transform is a representation of a wavelet function (Graps, 1995). Through wavelet transform, a given continuous time signal is divided into different scale components, in which a range of frequency is assigned to each scale component. The mother wavelet function $ψ (t)$ $\in L^{2} (R)$ , limited in time domain, i.e. $ψ$ (t) has values in a certain range and zeros elsewhere (Liu, 2010). The wavelet transform has properties of zero-mean and normalization (Daubechies, 1988; Strang and Nguyen, 1996) as presented in Equations (1) and (2), respectively.

\int_{- \infty}^{\infty} ψ (t) dt = 0

(1)

\int_{- \infty}^{\infty} ψ (t) ψ^{*} (t) dt = 1 = {‖ ψ (t) ‖}^{2}

(2)

The dilation and translation properties (Equation 3) states that the mother wavelet can form a basis set denoted by

ψ_{s, u} (t) = \frac{1}{\sqrt{s}} ψ (\frac{t - u}{s}) | u, s \in R

(3)

where u is the translating and s a scaling parameter holding property greater than zero, because defining negative scaling is not possible. For s=2^J, the wavelet transform is called a dyadic digital wavelet transform (DWT). The mallet algorithm (Mallat, 1999) is used to compute the DWT of a digital signal f(n); it is a digitized ECG signal in the form of samples, as expressed in Equations (4) and (5).

S_{2^{j}} f (n) = \sum h_{k} S_{2^{j - 1}} f (n - 2^{J - 1} k)

(4)

W_{2^{j}} f (n) = \sum g_{k} W_{2^{j - 1}} f (n - 2^{J - 1} k)

(5)

where $S_{2^{0}} f (n) = d (n), S_{2^{j}}$ is a smoothing operator and $W_{2^{j}} f (n)$ is the wavelet transform of the digital signal f(n) to be analysed. Low pass filter H(w) and high pass filter G(w) are further represented by coefficients $\sum^{g_{k}}$ and $\sum^{h_{k}}$ , respectively, as

H (w) = \sum h_{k} e^{- jkw} and G (w) = \sum g_{k} e^{- jkw}

Pre-processing

The main goal of artifact removal or pre-processing is to formalize the accurate and efficient delineation of characteristic points in ECG signals, further used for detection of the region of interest (ROI) for ST segment and isoelectric reference. It is significant to recognize the artifact frequencies and to discriminate these artifact changes from genuine changes to prevent misdiagnosis. The pre-processing (i.e. discarding of wavelet coefficients corresponding to artifact frequencies) process should not have any destructive effects on the morphologies of the original ECG signal to lose the diagnostic information. The choice of wavelet function depends upon the type of signal being analysed. The wavelet with similarity to the processed signal is usually selected.

Removal of baseline wanders

For correction of normalization (DC component), the mean of 1-min ECG record samples have been subtracted from the original ECG signal. Baseline wandering is a low-frequency noise in the range 0.5–1 Hz; hence the ECG signal (with a sampling frequency of 250 Hz) is decomposed to the eighth level using Daubechies (db4) wavelet transform for removal of baseline wander. The db4 wavelet wavelet gives details more correctly than the other wavelets for this noise, because db4 has similarity to the QRS complex and the energy spectrum is concentrated around low frequencies. Thresholding is applied after that, which involves the discarding of approximation coefficients to zero at the eighth decomposition level followed by reconstruction (Beena et al., 2014; Kumar and Singh, 2015). The results for removal of baseline wanders for e0103m record of EDB are shown in Figure 2.

Removal of muscle tremors

Muscle tremors are low amplitude noises usually lying in the frequency range 100–110 Hz. The Coiflet (Coif4) wavelet function for removal of these noises to the first level of wavelet decomposition is applied. First, the noise is estimated in the approximate coefficients of the first level of the ECG signal, named sigma, and then the threshold is deliberated by using this sigma value. The soft thresholding technique (Beena et al., 2014; Kumar and Singh, 2015) is applied for the discarding of detail coefficients, while maintaining the approximation coefficients at unity. The results of removing this noise for the e0103m record of EDB are shown in Figure 3.

Delineation of characteristic points

Detection of R peaks

While looking for peak locations in the pre-processed ECG signal, we have appended the 100 sample points to the left- and right-hand sides of the pre-processed ECG samples in order to overcome the scope of window crossing the signal boundaries. Firstly, the modulus maximum (maxh) and threshold ε=(max (ECG)−mean (ECG))/2 for searching the highest bumps in the entire record have been computed. Then indices into the boundaries for these peaks for a value of adaptive thresholding {maxh*ε} are applied (Banerjee et al., 2012; Kumar and Singh, 2016c). The time position and amplitude of each R peak is put in the R_INDEX. The results for detection of R peaks for the e0103m record of the EDB are shown in Figure 4.

Detection of P, Q, S and T peaks

For detection of P, Q, T, S peaks, a traversing search for maxima and minima is carried out in the implicit window lengths, primarily decided as the normal range of medical societies. As for the Q peak, a search for a local minimum in R_Index-200ms: R_INDEX-40ms is performed, and in the same manner, for R_INDEX- 400ms: R_INDEX-200ms for the local maximum, for the P peak. With similar logic, a search for a local minimum in the window of R_INDEX+20ms: R_INDEX+100ms for S peak and for a local maximum in window of R_INDEX+200ms: R_INDEX+400ms for T peak is performed. Depending upon the polarity of the maximum, any of the six possible T wave morphologies – positive (+), negative (−), biphasic (−/+ or +/−), upwards and downwards – have been carefully identified. Customarily, the Q and S peaks are high-frequency and low-amplitude waves in the ECG signal. The detected sample locations and corresponding amplitudes of the characteristic waves are stored in P_INDEX, Q_INDEX, S_INDEX and T_INDEX. In Figure 4, blue stars symbolize the P, QRS and T peaks consequent to their amplitudes and sample values for the e0103m record of the EDB (Banerjee et al., 2012; Ghaffaria et al., 2009; Kumar and Singh, 2016c; Li et al., 1995).

Detection of onset and offset point

The detection of onset and offset point for P, QRS and T peak is carried out for a corresponding INDEX. The windows of INDEX-80ms: INDEX for onset and INDEX: INDEX+80ms for offset point are fixed. The slope sign inversion in the defined window is computed using Newton’s difference formula (Ghaffaria et al., 2009; Lin et al., 2014; Kumar and Singh, 2016c):

\frac{f (x + h) - f (x)}{h}

(11)

Two-point estimation is used to calculate the slope of a nearby secant line through the points (x, f(x)) and (x+h, f(x+h)), where x represents the sample value. We choose a small number h, where h is a small change in x, and it may be either positive or negative, depending on onset and offset sample value. Detected onset and offset sample points are represented by red and green stars, respectively, for the e0103m record in Figure 4.

Detection of J point

The J point is lies after the QRS complex within the range 80–120 ms. The J point is detected as the first inflection point after the S peak. As shown in Figure 4, magenta-coloured stars represent the J points.

Detection of iso-electric reference

The iso-electric reference (IR) is detected in the TP segment of the ECG signal, i.e. the region between the T offset of current beat and P onset of the next beat.

Heart rate calculation

The heart rate is calculated as the distance between the two corresponding R–R intervals and is expressed in beats per minute (bpm). It is the average time of R–R interval.

ROI and ST Segment deviation measurement

We have defined ROI for accurate detection of ST segment and isoelectric reference. The advantage of defining ROIs includes, the algorithm will involve only peak samples instead of small voltage samples across the boundaries in a segment. It has been assumed that the ST segment lies in the Jpoint to Tonset, depending on the heart rate, i.e.

ST Segment

J+80 ms for heart rate ≤120 bpm

J+60 ms for heart rate >120 bpm

An algorithm for ROI corresponding to the ST segment is

ROI_ST

ST segment=length (Jpoint: Tonset)

M_ST=20% of ST Segment

ROI_ST=Jpoint+M_ST : Tonset−M_ST

Similarly, the algorithm for the ROI corresponding to IR is

ROI_IR

IR=length (Toffset: Ponset)

M_IR=20% of IR

ROI_IR=Toffset+M_IR : Ponset−M_IR

The ST segment deviation is computed as the potential difference between the ROI_ST and the ROI_IR, as ST Deviation=ROI_ST−ROI_IR.

Classification of beats

After measurement of the ST segment deviation for each beat in a particular record, we define a threshold to classify the beats as either normal or ischaemic. The value of ST segment deviation lies in the range 54–74 µV for elevated ST segments and −52 to −73 µV for depressed ST segments; hence we have defined the mean value of ±65 µV threshold. Keeping this viewpoint, a normal beat always lies between these values. The elevated or depressed beats have a value greater than this threshold (Kumar and Singh, 2016b). The normal beats are assigned to 1 and ischaemic (elevated or depressed) to 0, called the preliminary stage.

If ST segment deviation is ≥±65 µV

The beat is ischaemic

Else The beat is normal

An ischaemic beat characterization algorithm has been developed to filter out the spurious ischaemic or normal beats in a record, called the final stage. This algorithm helps to filter out the superior beats in a record, in which the ischaemic episode starts to follow the normal episode (or vice versa). This algorithm decides the final beat annotation, when transitions occur between the same record, and also helps make the proposed method more sensitive and positively predictive. The steps involved in proposed algorithm are:

Divide the entire ECG record of 120 min into 120 sections of 1 min each.

Firstly, each section is de-noised, delineated and each ST segment examined.

Classify all beats as being ischaemic or normal.

Each beat is annotated as 1 if normal and as 0 if ischaemic.

Recombine all section beats of 120-min ECG records with annotations, having an array of L entries, where L is the number of beats in the 120-min ECG record.

To implement an algorithm for the nth beat where 1≤n≤L, the record is appended with four extra entries, i.e. two at the beginning and two at the end. The record now starts at −1 and ends at L+2. The extended array will have [−1, 0, 1, 2, …L, L+1, L+2].

The extended array is now filtered from n=1 to n=L for spurious entries using following algorithm:

Let

S_{n} = \sum_{k = 0}^{k = 2} x (n + k)

Where x(n)=1 for the nth beat being normal

and x(n)=0 for nth beat being ischaemic

S_{n}

=3, then x(n)=1

Else if

S_{n}

=0, then x(n)=0

Else let

P_{n} = \sum_{k = - 2}^{k = 0} x (n + k)

P_{n}

=3, then x(n)=1

Else if

P_{n}

=0, then x(n)=0

Else let

T_{n} = \sum (x (n - 2), x (n - 1), x (n + 1), x (n + 2)

If T_{n} =

4, then x(n)=1

Else if

T_{n} =

0, then x(n)=0

Else x(n)=0.5, beat is unclassified

Ischaemia episode detection

As per the advice of the European Society of Cardiology, the ischaemic episode detection process should consider a minimum 30-s duration of ECG signals for investigation. Accordingly, an ischaemic window is defined, if it contains at least 80% ischaemic beats (Kumar and Singh, 2016b). A definition of ischaemic episode detection has been presented in Figure 5.

If [(No. of ischaemic beats)/(All beats)]≥80%

Then ischaemic window is detected

Else normal window is detected

Statistical analysis of ST measurement

From all individual ST segment values of the enduring beats, we then perform a statistical analysis on the measured ST segment deviations, which may also be used for the validation of results. These involve mean, standard deviation (SD), coefficient of variation (COV), kurtosis, skewness and form factor. Here, x_i=ith beat and n=number of beats.

Mean = \frac{\sum x_{i}}{n}

(12)

Standard Deviation (SD) = \sqrt{\frac{\sum {(x_{i} - mean)}^{2}}{n}}

(13)

Coefficient of Variation (COV) = \frac{SD}{Mean} * 100

(14)

The SD measures the absolute dispersion in a record. It facilitates the evaluation of the whole dataset with its mean. COV is the ratio of standard deviation and mean. The COV represents the consistency. The mean, SD and COV definitions have been mentioned in Equations (12), (13) and (14), respectively. Skewness (S) is a measure of the symmetry of the data set. The skewness of any perfectly symmetric (normal) distribution is exactly zero, defined as:

S = E [x - μ) / σ]^{3}

(15)

The kurtosis (K) measures whether the data is flat or peaked in a normal distribution, defined as:

K = E [x - μ) / σ]^{4}

(16)

where μ is the mean of x, σ is the standard deviation of x and E represents the expected value. The histogram is a graphical method to represent skweness and kurtosis. The form factor is the ratio of standard deviation of second-order derivative and original ST segment deviation:

F F = \frac{σ_{x^{″}} / σ_{x^{'}}}{σ_{x^{'}} / σ_{x}}

(17)

Results and discussion

We have measured ST segment deviations, and classify the beat as ischaemic or normal in two stages and correspondingly normal or ischaemic episodes in ECG records. The pre-processing, delineation and ROI’s algorithms have been implemented in MATLAB 2012a and for statistical analysis, IBM SPSS 22 has been used. A general flow chart for the presented method for ischaemia detection is presented in Figure 6. We have validated the proposed algorithm for the entire 90 records of European ST-T database. Statistical analysis includes mean, COV, kurtosis and form factor. These features help to differentiate a record as normal, depressed or elevated. The mean value for normal, depressed and elevated record is 0.035, −0.105 and 0.104; hence the ±65 µV threshold is set to differentiate normal beats from ischaemic beats. The computed values of COV, kurtosis and form factor are 28.57, −0.879 and 3.43 for a normal ST segment. The values of COV, kurtosis and form factor for the depressed ST segment are 7.14, 0.29 and 2.10; similarly, for the elevated ST segment, the values measured are 7.69, −0.377 and 2.58. These features are sufficient to recognize a record. A bell-shaped normal distribution graph is generated for the measured ST segment deviations with its histogram. The normal distribution graphs of the selected 1-min duration ST segment for normal, depressed and elevated ECG records have been shown in Figures 7(a), (b) and (c), respectively. The corresponding values of standard deviation and skewness for these records are displayed in Table 1. The presented method is based on thresholding, and thus does not require any type of testing and training. The performance evaluation for the detected ischaemic episodes is made through the sensitivity (S_e) and positive predictivity (+P), defined as

Sensitivity (S_{e}) = TP / (TP + FN) * 100

(18)

Positive predictivity (+ P) = TP / (TP + FP) * 100

(19)

where TP denotes true positive, FP denotes false positive and FN denotes false negative detection for the detection of an ischaemic episode window. Table 2 shows 97.71% S_E and 96.89% +P for 90 ECG records of the annotated European ST-T Database.

Table 1.

Statistical features for normal, depressed and elevated ST segments.

Statistical feature	Normal ST segment	Depressed ST segment	Elevated ST segment
Mean	0.035	−0.105	0.104
Standard deviation	0.01	0.0075	0.0081
Coefficient of variation	28.57	7.14	7.69
Skewness	0.071	0.079	−0.506
Kurtosis	−0.879	0.289	−0.377
Form factor	3.43	2.10	2.58

Table 2.

Performance evaluation for proposed method: application to records of European ST-T database.

Sr. No.	Record No.	S_E	+P	Sr. No.	Record No.	S_E	+P
1	e0103m	98.25	96.56	46	e0211m	97.45	96.41
2	e0104m	98.32	97.67	47	e0212m	97.82	96.46
3	e0105m	97.62	97.76	48	e0213m	97.36	96.56
4	e0106m	98.46	97.37	49	e0302m	96.21	96.38
5	e0107m	98.51	97.12	50	e0303m	97.67	95.29
6	e0108m	97.82	97.46	51	e0304m	98.43	95.37
7	e0110m	97.64	97.39	52	e0305m	97.84	96.46
8	e0111m	96.47	97.36	53	e0306m	97.87	96.35
9	e0112m	97.93	97.09	54	e0403m	97.78	95.56
10	e0113m	97.36	97.2	55	e0404m	98.51	97.12
11	e0114m	97.83	96.54	56	e0405m	97.62	96.71
12	e0115m	98.54	96.32	57	e0406m	97.46	95.43
13	e0116m	98.26	97.49	58	e0408m	97.87	95.76
14	e0118m	96.12	97.23	59	e0409m	98.73	95.66
15	e0119m	97.38	97.25	60	e0410m	97.69	96.32
16	e0121m	98.21	97.87	61	e0411m	96.56	96.87
17	e0122m	97.95	97.23	62	e0413m	96.83	95.86
18	e0123m	97.89	96.29	63	e0415m	97.12	95.61
19	e0124m	97.54	96.11	64	e0417m	96.34	98.43
20	e0125m	97.89	96.48	65	e0418m	97.78	98.46
21	e0126m	97.81	97.98	66	e0501m	97.89	98.32
22	e0127m	97.68	97.68	67	e0509m	97.45	98.56
23	e0129m	96.42	97.87	68	e0515m	97.25	98.52
24	e0133m	97.57	97.86	69	e0601m	97.76	96.58
25	e0136m	97.89	96.36	70	e0602m	97.61	96.26
26	e0139m	96.91	96.31	71	e0603m	97.43	97.89
27	e0147m	97.68	96.92	72	e0604m	98.76	97.76
28	e0148m	97.36	97.37	73	e0605m	97.37	97.59
29	e0151m	97.62	97.86	74	e0606m	97.32	95.36
30	e0154m	97.91	97.35	75	e0607m	98.76	96.49
31	e0155m	97.62	96.24	76	e0609m	98.69	97.98
32	e0159m	97.58	96.19	77	e0610m	97.26	95.47
33	e0161m	97.07	95.57	78	e0611m	97.34	96.25
34	e0162m	98.22	96.43	79	e0612m	97.41	95.22
35	e0163m	98.74	97.15	80	e0613m	98.58	95.28
36	e0166m	97.65	98.97	81	e0614m	98.87	96.47
37	e0170m	98.42	98.73	82	e0615m	98.56	98.20
38	e0202m	97.84	98.46	83	e0704m	97.98	97.31
39	e0203m	97.39	98.41	84	e0801m	97.37	97.43
40	e0204m	97.65	97.42	85	e0808m	97.46	96.25
41	e0205m	98.16	97.88	86	e0817m	97.77	96.42
42	e0206m	97.83	96.32	87	e0818m	97.52	96.64
43	e0207m	97.92	96.38	88	e1301m	97.78	96.18
44	E0208m	97.52	96.37	89	e1302m	98.68	97.67
45	E0210m	97.87	95.43	90	e1304m	96.24	96.13
				Average		97.71	96.89

S_e, sensitivity; +P, positive predictivity.

Comparison with existing methods

This paper presents a novel method for detection of ischaemic episodes through statistical analysis of ST segments in ECG recordings of the European ST-T database. The presented method is very simple and does not involve any multifaceted calculations for investigation of ischaemia. The performance of the method has been validated on the entire 90 records of the European ST-T database. This database is commonly used as a standard reference by medical researchers internationally for validation of their algorithms or methods. Exclusively, it was considered that each annotated ischaemic episode in the database contains only ischaemic beats. The performance of our method cannot be judged against other methods (Afsar et al., 2008; Baxt, 1991; Correa et al., 2014; Don et al., 2010; Exarchos et al., 2007; Garcia et al., 2000; Gramatikov et al., 2000; Kumar and Singh, 2016a, 2016b; Lemire et al., 2000; Maglaveras, 1998; Murugan and Radhakrishnan, 2010; Papadimitriou et al., 2001; Papaloukas et al., 2004; Park et al., 2012; Ranjith et al., 2003; Rocha et al., 2010; Stamkopoulos et al., 1998; Tsipouras et al., 2007; Zahan, 2001), as they employed different performance measures or different databases or different datasets. The obtained results are better than those of other approaches (Andreao et al., 2004; Jager et al., 1992; Papaloukas et al., 2001, 2002; Rocha et al., 2010; Silipo and Marchesi, 1998; Silipo et al., 1994; Stamkopoulos et al., 1992; Taddei et al., 1995; Vila et al., 1997) in terms of both S_E and +P. Other methods have certain advantages and disadvantages, e.g. signal analysis methods (Gramatikov et al., 2000; Lemire et al., 2000; Martis et al., 2013; Ranjith et al., 2003; Taddei et al., 1995) are easy to implement in real time. Similarly, rule-based algorithms (Papaloukas et al., 2001) provide a good decision similar to that of the experts. A limitation of this method is that it requires a representative training set of rules for validation purposes. However, fuzzy logic methods (Exarchos et al., 2007; Tsipouras et al., 2007; Vila et al., 1997; Zahan, 2001) have enormous significance in the analysis of ischaemia for the validation of rules, but it required investigations to improve the analysis performance. HMM (Andreao et al., 2004) is not suitable for detection of non-ischaemic episodes. The ANNs-based approaches (Maglaveras et al., 1998; Papaloukas et al., 2002) cannot be achieved easily, because of ‘black box’ and the required tedious processing to interpret. In the same way, genetic algorithms (Papaloukas et al., 2004), SVM and KDE (Park et al., 2012) based algorithms take more time for decisions, as they involve many complex calculations for the optimization process. The presence of artifacts usually affects the detection of the characteristic points, ST segment and ischaemic beats in the ECG recordings. The wavelet transform-based filtration and delineation makes the proposed method perform better than the modern ECG recorders used currently. The proposed method has two major advantages. The method can provide an interpretation of results. This will help the cardiologist reach a diagnosis faster. Second, it involves direct analysis based on statistical features without involvement of any complex calculations and training. Table 3 shows a comparison of results for the presented method with existing methods in the literature in terms of S_E and +P for the same (entire) records of the European ST-T database. The results validate the efficacy of presented method by showing 97.71% Se and 96.89% +P. The above-mentioned methods demonstrate an Se that ranged from 77% to 96.7% and a +P that ranged from 79% to 96.2%. The presented method performance exhibits the highest Se and +P. The simplicity, robustness and automatic discarding of spurious beats in ECG records make the method feasible for use in clinical systems. A disadvantage of the presented method is that it is not applicable to analysis of ischaemia through T waves.

Table 3.

Performance evaluation, comparison of presented method with existing methods for European ST-T records.

Methods	S_E (%)	+P (%)
Maglaveras et al. (1998)	88.62	78.38
Jager et al. (1992)	85	86
Papaloukas et al. (2001)	94	79
Silipo and Marchesi (1998)	77	86
Silipo et al. (1994)	85	88
Silipo et al. (1994)	78	90
Stamkopoulos et al. (1992)	84	79
Taddei et al. (1995)	84	85
Jager et al. (1998)	83.8	87.1
Vila et al. (1997)	83	75
Papaloukas et al. (2002)	90	89
Rocha et al. (2010)	96.7	96.2
Andreao et al. (2004)	83	85
Presented method	97.71	96.89

S_e, sensitivity; +P, positive predictivity.

Conclusion and future scope

We have developed a simple method for the detection of ischaemic episodes in ECG records through statistical analysis. Although this research is effective in detecting myocardial ischaemia and achieved the highest performance, there are still many challenges that need to be addressed. The research has been developed using the datasets of the European ST-T database. In order for the developed method to be used in clinical practice, the database records may be comprehensive and additional types of ischaemic and normal ST segments may be incorporated. The developed method performance may be validated for additional functionality, e.g. for multiple leads and with other promising classifiers. Future work must include the detection of ST segment axis shift and 24-h ambulatory recordings in real-time systems.

Footnotes

Declaration of Conflicting Interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

References

Afsar

Arif

Yang

(2008) Detection of ST segment deviation episodes in ECG using KLT with an ensemble neural classifier. Physiological Measurement 29(7): 747–760.

Andreao

Dorizzi

Boudy

. (2004) ST segment analysis using hidden Markov model beat segmentation: application to ischemia detection. In: IEEE Conference on Computers in Cardiology, pp. 381–384.

Banerjee

Gupta

Mitra

(2012) Delineation of ECG characteristic features using multi-resolution wavelet analysis method. Measurement 45(3): 474–487.

Baxt

(1991) Use of an artificial neural network for the diagnosis of myocardial infarction. Annals of Internal Medicine 115(11): 843–848.

Beena

Prabavathy

Priestly

. (2014) Robust rule based de-noising scheme using wavelets. Journal of Medical Imaging and Health Informatics 4(4): 457–468.

Bianchi

Mainardi

Cerutti

(2000) Time–frequency analysis of biomedical signals. Transactions of the Institute of Measurement and Control 22(3): 215–230.

Brownfield

Herbert

(2008) EKG criteria for fibrinolysis: what’s up with the J point?Western Journal of Emergency Medicine 9(1): 40–42.

Correa

Arini

Correa

. (2014) Novel technique for ST-T interval characterization in patients with acute myocardial ischemia. Computers in Biology and Medicine 50: 49–55.

Dash

(2002) Electrocardiogram monitoring. Indian Journal of Anaesthesia 46(4): 251–260.

10.

Daubechies

(1988) Ortho-normal bases of compactly supported wavelets. Communication in Pure Applied Mathematics 41(7): 909–966.

11.

Don

Chung

Min

(2013) Analysis of electrocardiogram signals of arrhythmia and ischemia using fractal and statistical features. Journal of Mechanics in Medicine and Biology 13: 1350008-8.

12.

Exarchos

Tsipouras

Exarchos

. (2007) A methodology for the automated creation of fuzzy expert systems for ischemic and arrhythmic beat classification based on a set of rules obtained by a decision tree. Artificial Intelligence in Medicine 40(3): 187–200.

13.

Gacek

Pedrycz

(2012) ECG Signal Processing, Classification and Interpretation. London: Springer-Verlag.

14.

Garcia

Sörnmo

Olmos

. (2000) Automatic detection of ST-T complex changes on the ECG using filtered RMS difference series: application to ambulatory ischemia monitoring. IEEE Transactions on Biomedical Engineering 47(9): 1195–1201.

15.

Ghaffaria

MRH

Akraminiaa

Ataroda

. (2009) A robust wavelet-based multi-lead electrocardiogram delineation algorithm. Medical Engineering & Physics 31(10): 1219–1227.

16.

Goldberger

(1981) Electrocardiogram in ischemic heart disease. The Western Journal of Medicine 135(2): 104–109.

17.

Gramatikov

Brinker

Sun

. (2000) Wavelet analysis and time–frequency distributions of the body surface ECG before and after angioplasty. Computer Methods and Program Biomedicine 62(2): 87–98.

18.

Graps

(1995) An introduction to wavelets. IEEE Transaction on Computing Science & Engineering 2(2): 50–61.

19.

Hegde

Deekshit

Satyanarayana

(2013) Identification of electrocardiogram morphologies using continuous wavelet transform, Journal of Medical Imaging and Health Informatics 3(3): 417–423.

20.

Jager

Mark

Moody

. (1992) Analysis of transient ST segment changes during ambulatory monitoring using the Karhunen–Loève transform. In: Proceedings of the IEEE Conference on Computers in Cardiology, pp. 691–694.

21.

Jager

Moody

Mark

(1998) Detection of transient ST segment episodes during ambulatory ECG monitoring, Computers and Biomedical Research 31(5): 305–322.

22.

Kindi

ASG

Reza

(2011) Real-time detection of myocardial infarction by evaluation of ST-segment in digital ECG. Journal of Medical Imaging and Health Informatics 1(1): 225–230.

23.

Kumar

Singh

(2015) Optimal selection of wavelet function and decomposition level for removal of ECG signal artifacts. Journal of Medical Imaging and Health Informatics 5(1): 138–146.

24.

Kumar

Singh

(2016a) Ischemia Detection using isoelectric energy function. computers in biology & medicine. Computers in Biology and Medicine 68: 76–88.

25.

Kumar

Singh

(2016b) Statistical analysis of ST segments for ischemia detection in ECG signals. Journal of Medical Imaging and Health Informatics 6(2): 431–440.

26.

Kumar

Singh

(2016c) Robust multi-resolution wavelet analysis and window search based approach for electrocardiogram features delineation. Journal of Medical Imaging and Health Informatics 6(1): 146–156.

27.

Kumar

Sonam

(2016) Statistical analysis of QRS-complex to evaluate the QR versus RS interval alteration during ischemia. Journal of Medical Imaging and Health Informatics 6(1): 210–214.

28.

Kusumoto

(2004) Cardiovascular Patho-physiology. Raleigh, NC: Hayes Barton Press.

29.

Lemire

Pharand

Rajaonah

. (2000) Wavelet time entropy, T wave morphology and myocardial ischemia. IEEE Transaction on Biomedical Engineering 47(7): 967–970.

30.

Zheng

Tai

(1995) Detection of ECG characteristic points using wavelet transforms. IEEE Transaction on Biomedical Engineering 42(1): 21–28.

31.

Wang

(2013) Detection of electrocardiogram characteristic points using lifting wavelet transform and Hilbert transform. Transactions of the Institute of Measurement and Control 35(5): 574–582.

32.

Feng

Cao

. (2015) Heartbeat classification using different classifiers with non-linear feature extraction. Transactions of the Institute of Measurement and Control, doi: 10.1177/0142331215620697.

33.

Feng

Cao

. (2016b) A new ECG signal classification based on WPD and ApEn feature extraction. Circuits, Systems, and Signal Processing 35(1): 339–352.

34.

Liang

Miao

. (2016a) Novel ECG signal classification based on KICA nonlinear feature extraction. Circuits, Systems, and Signal Processing 35(4): 1187–1197.

35.

Wang

Chen

. (2014) Denoising and R-peak detection of electrocardiogram signal based on EMD and improved approximate envelope. Circuits, Systems, and Signal Processing 33(4): 1261–1276.

36.

Lin

Liang

. (2014) Discrete-wavelet-transform-based noise removal and feature extraction for ECG signals. IRBM 35(6): 351–361.

37.

Liu

(2010) A Tutorial of the Wavelet Transform. Available online: http://disp.ee.ntu.edu.tw/tutorial/WaveletTutorial.pdf

38.

Maglaveras

Stamkopoulos

Pappas

. (1998) An adaptive back propagation neural network for real-time ischemia episodes detection: development and performance analysis using the European ST-T database. IEEE Transactions on Biomedical Engineering 45(7): 805–813.

39.

Mallat

(1999) A Wavelet Tour of Signal Processing. Burlington, MA: Academic Press.

40.

Manocha

Singh

(2011) An overview of ischemia detection techniques. International Journal of Science and Engineering Research 2(11): 1–6.

41.

Martis

Acharya

Adeli

(2014) Current methods in electrocardiogram characterization. Computers in Biology & Medicine 48: 133–149.

42.

Martis

Acharya

Min

(2013) ECG beat classification using PCA, LDA, ICA and discrete wavelet transform. Biomedical Signal Processing and Control 8(5): 437–448.

43.

Murugan

Radhakrishnan

(2010) Rule based classification of ischemic ECG beats using ant-miner. International Journal of Engineering Science and Technology 2(8): 3929–3935.

44.

Papadimitriou

Mavroudi

Viadutu

. (2001) Ischemia detection with a self organizing map supplemented by supervised learning. IEEE Transaction on Neural Networks 12(3): 503–515.

45.

Papaloukas

Fotiadis

Liavas

. (2001) A knowledge-based technique for automated detection of ischemic episodes in long duration electrocardiograms. Medical Biology Engineering Computer 39(1): 105–112.

46.

Papaloukas

Fotiadis

Likas

. (2002) An ischemia detection method based on artificial neural networks. Artificial Intelligence in Medicine 24(2): 167–178.

47.

Papaloukas

Goletsis

Fotiadis

. (2004) Automated ischemic beat classification using genetic algorithm and multi-criteria decision analysis. IEEE Transaction on Biomedical Engineering 51(10): 1717–1725.

48.

Park

Pedrycz

Jeon

(2012) Ischemia episode detection in ECG using kernel density estimation, support vector machine and feature selection. Biomedical Engineering On-line, 11: 1–22, doi: 10.1186/1475-925X-11-30.

49.

Ranjith

Baby

Joseph

(2003) ECG analysis using wavelet transform: application to myocardial ischemia detection. ITBM-RBM 24(1): 44–47.

50.

Rocha

TS.

Carvalho

Henriques

. (2010) A lead dependent ischemic episodes detection strategy using Hermite functions. Biomedical Signal Processing and Control 5(4): 271–281.

51.

Silipo

Marchesi

(1998) Artificial neural networks for automatic ECG analysis. IEEE Transaction Signal Processing 46(5): 1417–1425.

52.

Silipo

Taddei

Marchesi

(1994) Continuous monitoring and detection of ST-T changes in ischemic patients. In: Proceedings of the IEEE Conference on Computers in Cardiology, pp. 225–228.

53.

Singh

(2010) Introduction to Biomedical Instrumentation. New Delhi: Prentice Hall of India.

54.

Stamkopoulos

Diamantaras

Maglaveras

. (1998) ECG analysis using nonlinear PCA neural networks for ischemia detection. IEEE Transactions on Signal Processing 46(11): 3058–3067.

55.

Stamkopoulos

Strintzis

Pappas

. (1992) One-lead ischemia detection using a new backpropagation algorithm and the European ST-T database. In: Proceedings of the IEEE Conference on Computers in Cardiology, pp. 663–666.

56.

Strang

Nguyen

(1996) Wavelets and Filter Banks. Wellesley, MA: Wellesley-Cambridge Press.

57.

Taddei

Costantino

Silipo

. (1995) A system for the detection of ischemic episodes in ambulatory ECG. In: Proceedings of the IEEE Conference on Computers in Cardiology, pp. 705–708.

58.

Taddei

Distante

Emdin

. (1992) The European ST-T database: standard for evaluating systems for the analysis of ST-T changes in ambulatory electrocardiography. European Heart Journal 13(9): 1164–1172.

59.

Tsipouras

Costas

Fotiadis

(2007) A framework for fuzzy expert system creation – application to cardiovascular disease. IEEE Transactions on Biomedical Engineering 54: 2089–2105.

60.

Vila

Presedo

Delgado

. (1997) SUTIL: intelligent ischemia monitoring system. International Journal of Medical Informatics 47(3): 193–214.

61.

WHO (2014) Global status report on noncommunicable diseases 2014, ISBN 9789241564854, available at http://www.who.int/mediacentre/factsheets/fs317/en/.

62.

Zahan

(2001) A fuzzy approach to computer-assisted myocardial ischemia diagnosis. Artificial Intelligence in Medicine 21(1–3): 271–275.