• No results found

Automatic Detection of Atrial Fibrillation Using Empirical Mode Decomposition and Statistical Approach

N/A
N/A
Protected

Academic year: 2021

Share "Automatic Detection of Atrial Fibrillation Using Empirical Mode Decomposition and Statistical Approach"

Copied!
8
0
0

Loading.... (view fulltext now)

Full text

(1)

Procedia Technology 10 ( 2013 ) 45 – 52

2212-0173 © 2013 The Authors. Published by Elsevier Ltd. Open access under CC BY-NC-ND license.

Selection and peer-review under responsibility of the University of Kalyani, Department of Computer Science & Engineering doi: 10.1016/j.protcy.2013.12.335

ScienceDirect

International Conference on Computational Intelligence: Modeling Techniques and Applications

(CIMTA) 2013

Automatic Detection of Atrial Fibrillation using Empirical Mode

Decomposition and Statistical Approach

U Maji

a

**, M Mitra

b

S Pal

b

aDepartment of Applied Electronics and Instrumentation Engineering, Haldia Institute of Technology, Haldia, West Bengal, India bDepartment of Applied Physics, University of Calcutta, Kolkata, West Bengal, India

Abstract

Automatic detection of different cardiac abnormalities is an emerging field of study in assistive diagnosis technology for cardiac diseases. Atrial fibrillation (AF) is a kind of arrhythmia which increases risk of heart attack especially to the older people. Detection of AF at the early stage may cause prevention of serious stoke. This paper presents a method of automatic detection of AF by using higher order statistical moments of ECG signal in Empirical Mode Decomposition (EMD) domain. The proposed technique operates in two stages. First stage consists of decomposition of denoised ECG into intrinsic mode functions (IMF) and derives the statistical parameters like variance and standard deviation for classification from each IMF. In the second stage these features are used by a supervised classifier to distinguish between normal and AF ECG rhythms. The performance of this method is tested with the MIT-BIH arrhythmia data base. Possibility to use different IMFs is also tested and 96% sensitivity is achieved in IMF 4.

© 2013 The Authors. Published by Elsevier Ltd.

Selection and peer-review under responsibility of the University of Kalyani, Department of Computer Science & Engineering.

"Keywords: Atrial Fibrillation, EMD, IMF, Variance, Standard deviation;"

1.Introduction

Atrial fibrillation (AF) is the most common arrhythmia caused by the irregular conduction of impulses to the ventricles. It may sustain for a minute or a day long. More risk may be caused by the AF are cardiovascular and

* Corresponding author. Tel.: 9143098701

E-mail address: udaymaji@gmail.com.

© 2013 The Authors. Published by Elsevier Ltd. Open access under CC BY-NC-ND license.

(2)

metabolic comorbidities like hypertension, coronary artery disease, congestive heart failure, and myocardial infarction [1, 2].

In comparison to normal sinus rhythm, AF rhythm is mainly characterized by the absence of P wave where the conventional P waves are replaced by rapid oscillation of fibrillatory waves in the ratio 2:1, 4:1 [3] that are varying in amplitude, shape and time of occurrence. Another important characteristic of the AF wave is the irregular RR intervals (figure 1) which are caused due to uncoordinated electrical impulses that disrupt the steady atrial activation in the heart.

Various works have been done to detect the AF beats from the ECG signal by using characteristic features along with a specialized classification algorithm. XiuhuaRuan et al. [4] and K Tateno et.al. [5] have developed a method based on irregularity of RR interval. But there is a possibility of misinterpretation of AF if only RR interval feature is considered as the same feature alters in other cardiac abnormalities like Premature Ventricular Construction (PVC). A model based analysis of features is also proposed in [6], [7] with a sequential analysis of atrial activity in ECG signals. By coupling genetic programming with orthogonal least squares (GP/OLS) and simulated annealing (GP/SA) based optimization is also applied to the detection of atrial fibrillation episodes in [8]. This method provides good detection accuracy at the cost of more mathematical complexity with higher computational cost.

Fig. 1. RR interval and P peaks of atrial fibrillation rhythm.

In this work a new technique is proposed to detect the AF by Empirical Mode Decomposition (EMD) based analysis. EMD is a fully data driven signal decomposition technique mostly applied for analysis of morphologically complicated non-stationary signals like ECG. The detail of the process is explained in section 2.2. Apart from RR intervals, P wave morphology is abruptly changed in AF. Hence P wave related features are used in the proposed algorithm. Statistical feature[9] has been used to which can discriminate the AF episode from the normal sinus rhythm. Variance of the selected Intrinsic Mode Function differs significantly for either rhythm. Supervised learning based classification technique is used for discrimination of Atrial Fibrillation from normal sinus rhythms.

2.Methods

The proposed method of detection involves two steps. In the first step a filtering of ECG signal and feature extraction technique has been described. Second section mainly comprised of AF classification based on the characteristic of P-wave.

2.1.ECG signal processing and SQ interval selection

It is required to filter the ECG signal before extracting the features as it contains various kinds of noises such as electromyographic noise, power frequency noise, baseline wondering etc. The signal is filtered with a standard Butterworth band pass filter of pass band frequency 0.5Hz to 45Hz[10] to remove above-mentioned high and low

1 2 3 4 5 6 7 950 1000 1050 1100 1150 1200 1250 Time Am plit ud e Irregular RR interval Absence of P-wave

(3)

frequency noises. In this proposed work, section of the ECG signal between two R peaks is selected to obtain the percentage of distortion of the P wave of a cycle.

2.2.Feature extraction

The ECG data used for this study contain various sets of normal rhythm and AF rhythm of different person. Among this data 102 different cycles of normal rhythm and 63 different cycles for AF rhythms are arbitrarily chosen from the different data sets. Each cycle contain duration between two R-peaks, from S peak of one rhythm to Q peak of the next rhythm. This particular cycle is decomposed into various IMF as describe below.

EMD is a procedure of obtaining modulated pattern from a time domain data. This pattern is called IMFs [11]. This method has been used in other field like Oceanography, Geography, Medicine, fault diagnosis etc. S pal et al [17] has been used EMD technique for removable of noise from ECG signal. In this work EMD is first time used for the feature extraction of AF rhythm. The steps of Empirical Mode decomposition of any signal

x

(

t

)

are

1. At first all the local maxima and minima of a given signal are identified.

2. Cubic spline interpolation is used to connect all the local maxima and thus upper envelope of mother signal is constructed.

3. This procedure is repeated for the local minima to produce the lower envelope.

4. The mean

m

1 of upper and lower envelope is calculated and the difference

d

1 between the signal

x

(

t

)

and

1

m

is computed as

d

1

(

t

)

, i.e.

)

(

)

(

t

m

1

d

1

t

x

5. If

d

1satisfies the condition of IMF, then

d

1 is the first frequency and amplitude modulated oscillatory mode of

x

(

t

)

.

6. If

d

1is not an IMF, then the shifting process described in step (1), (2), (3) are repeated on

d

1. Thus

d

11is calculated as,

)

(

11 11 1

m

d

t

d

(1)

In which

m

11is the mean of upper and low envelope value of

d

1

7. Let after

k

cycles of operation,

d

1kbecomes an IMF, that is

)

(

1 1 ) 1 ( 1

m

d

t

d

k

k k (2)

8. Then, it is designated as

c

1

d

1k, the first IMF component from the origin data. 9. Subtracting

c

1from

x

(

t

)

,

r

1 is calculated as,

1

1 x(t) c

r (3)

Which is treated as the original data for the next cycle.

10. Repeating the above process for n times n no. of IMFs are obtained along with the final residue

r

n. The decomposition process can be stopped when

r

n becomes a monotonic function, from which no more IMF can be extracted. A popular stopping criteria is to have the value of normalized standard difference(NSD) within a predefined threshold where

¦

r k k k k t d t d t d NSD 1 2 2 1 ) ( ) ( ) ( (4)

By summing finally we obtain

¦

N n n n t r t c t x 1 ) ( ) ( ) ( (5)

Residue

r

n is the mean trend of

x

(

t

)

. IMFs

c

1,

c

2

,...

c

nrepresent the finally obtained amplitude and frequency

modulated output set. Their frequency gradually decreases as the order of the IMFs increases. It is already mentioned that in normal sinus rhythm, all the characteristic waves of ECG signal are properly visible whereas, in AF rhythms the P wave is distorted. Hence in the IMFs, these changes are to be reflected.

Based on the above algorithm the derived IMFs of one normal sinus rhythm and AF affected rhythm is shown in fig.2

(4)

Fig. 2. IMFs of Normal and AF rhythm

It is observed from the figure 2 that, there is a variation of peaks of the IMF for normal and AF rhythm as expected. In case of normal rhythm the oscillations due to T and P peaks are prominent in IMFs leading to large positive peaks of the IMFs at the beginning and end compare to middle. But for AF rhythm due to absence of P peaks, the amplitudes of the IMFs are decreasing in nature towards the end. Thus if the peaks are connected by straight lines, the slope of the line will be gradually increasing in case of normal beats and decreasing for AF rhythms (fig3).

Fig. 3. Plot of the slope of the IMF peaks

There are 6 to 8 IMFs obtain from each cycle (RR interval) of the selected ECG signal depending upon the complexity of the signal. Among these IMFs, no. 3, 4, 5 are identified for the present study as the morphology of P wave is prominent in these IMFs. This is also supported by visual inspection. The positive halves of IMFs are considered to measure the changes between the normal and AF rhythm (fig4).

2.3.P- wave detection

The absence of the P- wave in the ECG rhythm is one of the important characteristics of AF. Instead a relatively small amplitude fibbrillatory wave is present. Because of this fact IMFs of AF and normal rhythm are showing a

(5)

distinct feature. It is evident from the figure 3 that the oscillation of the signal towards the end of the proposed test section of ECG is more in normal beats in comparison to the AF beats.

Fig. 4. Positive halves of the selected IMFs

In this work, higher order statistical parameters such as variance and standard deviation are used to classify the ECG signal in EMD domain. It is used to measure the variation of each cycle of the selected and rectified IMFs around the mean value. It is evident from figure 4 that, towards the end amplitude of IMFs of normal and AF rhythms are different. Variance and standard deviation measure the percentage of changes in IMF amplitude. It measured as follows

For an N point data standard deviation

V

is given as

¦

N i i x N 1 2 ) ( 1 P ) V ; (6)

¦

N i i x N 1 1 P (7)

Where

x

iand μ is

i

thvalue and mean value of the set respectively. Whereas variance is the square of the standard deviation i.e.

V

2. Variance is calculated and normalized for all the selected IMFs (fig.5). Variance profile indicates that, for normal rhythm it is minimum at the middle of the graph and increases at the end from 50 to 100%. Whereas percentage of variance at the end is remain within 0 to 40%. It reflects at the end variance is increasing for NSR and decreasing in nature for AF rhythm. Main reason of this characteristic lies in the fact that presence of a prominent P-peak for NSR which is absent in case of AF wave and replace byfibrillatory waves.yy

3. Classification

The irregularity of P- wave has been used to classify the fibrillatory wave from the normal wave. It is very difficult to detect the presence of P- wave, as its amplitude is very small. Empirical mode decomposition and higher order statistics like standard deviation and variance are used to classify the signals in [12]. In this study this feature being used and empirical mode decomposition technique is used to extract the feature. Statistical parameter like standard deviation and variance are used to classification of the signals.

It shown in the fig.4 that, ends of the selected IMFs and variance of the last cycle (fig.5) are increasing in nature for natural rhythm and decreasing for AF. This feature is motivated to select the variance of last two cycle is selected and average of them is plotted against the corresponding standard deviation. For the scatter three different IMF as 3rd, 4th and 5th are selected among the IMFs obtained by EMD.

(6)

1. The mean value is calculated for the standard deviation and variance from the normal and AF training data set. The mean value for the training data set are (80.94 70.18) and (35.69 17.01) for normal and AF group respectively for IMF 3.

Fig. 5. Plot of Variance of normal and AF rhythm.

2. The equation of the straight line joining the two coordinate is computed. The normal bisector of this line is the optimal linear decision function.

3. The equation of the linear decision function is given by

0 8 . 635 9 . 11 Y X (8)

The decision rule is as follows:

8 . 635 9 . 11 Y

X If > 0 beat under test is Normal And if ≤ 0 beat under test is AF

Similarly decision function is calculated for IMF4 and IMF5. Training is done with 102 normal and 63 AF rhythms as shown in fig 6.

Fig. 6. Scatter plot of stander deviation Vs. Variance (a) IMF3; (b) IMF4; (c) IMF5

0 20 40 60 80 100 0 10 20 30 40 50 60 70 80 90 100 5Th IMF Standard Deviation (%) (c) V a ri a n c e (% ) 0 20 40 60 80 100 0 10 20 30 40 50 60 70 80 90 100 Standard Deviation (%) (a) V a ri a n c e (% ) 3rd IMF 0 20 40 60 80 100 0 10 20 30 40 50 60 70 80 90 100 Standard Deviation (%) (b) 4Th IMF V a ri a n c e (% ) Bisecter AF Normal

(7)

4.Result

To evaluate the performance of the proposed algorithm, its detection accuracy is analysed and compare to other state-of-the-art algorithms. For the measurement the AF detection accuracy of the proposed algorithm, it is used the MIT-BIH AF database of 30 minute duration. This proposed algorithm is tested with total 110 cycles of normal rhythm and 68 cycles of AF rhythm. In order to measure the performance of this work sensitivity (SN) and specificity (SP) are calculated for all the selected IMFs. It is calculated as follows.

FN TP TP SE FP TN TN SP

Where true positive (TP): AF is classified as AF; true negative (TN): non-AF is classified as non-AF; false negative (FN): AF is classified as non-AF; false positive (FP): non-AF is classified as AF. Based on this result shows in table1. This shows sensitivity of 4th IMF is maximum, whereas specificity is maximum for 3rd. But result

based on the 5th IMF is very poor. A comparison with other few previously published algorithms onthe MIT-BIH

AF database is listed in Table 2. The proposedalgorithm shows a moderate result in compare to other published result.

Table 1.Performance result of the scatter plot

IMF No. Sensitivity (SE) Specificity (SP)

3 94.72 94.87

4 96.14 93.51

5 87.81 83.07

Table 2. Comparison of algorithm performance

Method Sensitivity (SE) Specificity (SP)

XiuhuaRuan [4] 100 100 J. Lee et al [5] 97.41 97.54 Eric Helfenbein et al [6] 76 97 S Dash et al [14] 94 95 Tran Thong [15] 89 91 Francisco Rinc´on [16] 96 93 Richard P. M [19] 96 90 Proposed algorithm 96 93 5.Conclusion

This paper presents a method to detect atrial fibrillation rhythm by using the irregularity of P-wave feature. The proposed method uses the segment of the ECG signal between two R peaks. Though the variation of R to R interval is an important indication of AF but is not unique. Hence P wave texture based feature is considered here. EMD based signal decomposition technique is a versatile tool to analyze a signal presently being applied in many areas including biomedical signal processing. In this research this method is used to extract the lack of oscillatory nature of P wave section in AF waves. Higher order statistical parameters like variance and standard deviation are used to quantify the difference of oscillations present in normal sinus rhythm and AF rhythm. These statistical parameters show good difference in magnitude leading to better classification using conventional training-testing based supervised classification technique. Hence the proposed algorithm can be applied to identify the instant of change in rhythm from normal to AF of a cardiac patient under observation.

(8)

This proposed technique is more advantageous with respect to other because EMD technique which is used for feature extraction also can use for noise elimination [17]. Hence single tool perform multiple roll in the process, where as other technique may show little higher sensitivity but uses different method for signal processing and feature extraction.

The proposed method can be used to detect other cardiac rhythms like atrial bradycardia, ventricular tachycardia etc. and the study can be extended to identify different rhythmic abnormalities using a single classifier in intensive cardiac care units and in holter monitoring.

References

[1] Benjamin EJ, Wolf PA, D'Agostino RB, Silbershatz H, Kannel WB, Levy D., “ Impact of atrial fibrillation on the risk of death: the

Framingham Heart Study”. Circulation, 1998; 98(10):946-52.

[2] N. Hannon, O. Sheehan, L. Kelly, M. Marnane, A. Merwick, A. Moore, L. Kyne, J. Duggan, J. Moroney, P. M. E. McCormack, L. Daly, N. Fitz-Simon, D. Harris, G. Horgan, E. B. Williams, K. L. Furie, and P. J. Kelly, “Stroke associated with atrial fibrillation – Incidence and early

outcomes in the North Dublin population stroke study,” Cerebrovasc. Diseases, vol. 29, no. 1, pp. 43–49, 2010.3. Y. Dimitriev and E. Kashchieva, J.Mater. Sci. 10 (1975) 1419.

[3] ECG Made Easy, A. Luthra, J.B. Medical Publishers Pvt. Ltd. ISBN. 81-7179-538-2, pp. 179

[4] Xiuhua Ruan, Changchun Liu, Chengyu Liu, Xinpei Wang, Peng Li, “Automatic Detection of Atrial Fibrillation Using R-R Interval

Signal”,in: International Conference on Biomedical Engineering and Informatics, pp. 644-647, 2011

[5] J. Lee, D. McManus and K. Chon, “Atrial Fibrillation Detection using Time-Varying Coherence Function and Shannon Entropy”, in: 33rd

Annual International Conference of the IEEE EMBS Boston, Massachusetts USA, pp- 4685-4688, August 2011

[6]. Eric Helfenbein, Richard Gregg, James Lindauer, Sophia Zhou, “An Automated Algorithm for the Detection of Atrial Fibrillationin the

Presence of Paced Rhythm”, Computing in Cardiology 2010; vol 37: pp.113−116.

[7] I Christov, G Bortolan, I Daskalov, “Sequential Analysis for Automatic Detection of Atrial Fibrillation and Flutter”, in: Computers in Cardiology 2001,pp-293-296, IEEE.

[8] F. Yaghouby, et al., Towards automatic detection of atrial fibrillation: A hybrid computational approach, Comput. Biol. Med. (2010), doi:10.1016/j.compbiomed.2010.10.004

[9] S. M. Shafiul Alam and M. I. H. Bhuiyan, “Detection of Seizure and Epilepsy Using Higher Order Statistics in the EMD Domain”, IEEE

journal of biomedical and health imformatics, vol. 17, No. 2, March 2013

[10] J. J. Bailey, A. S. Berson, A. Garson, L. G, Horan, P. W. Macfarlanc, and C. Zywictz. Recomendation for the standarddization and specification in automated electrocardiography: bandwidth and signal processing. Circulation, vol. 81, pp. 730-739, 1990.

[11] N.E. Huang, Z. Shen, S.R. Long, M.C.Wu, H. H.Shih, Q. Zheng, N.-C.Yen, C. C. Tung, and H. H. Liu, “The Emperical Mode

Decomposition and the Hilbert spectrum for nonlinear and nonstationary time series analysis,” proc. Roy.Lond, vol A 454, pp. 903-995,1998.

[12]. Jinseok Lee,, David D. McManus and Ki H. Cho, “Automatic Motion and Noise Artifact Detection in Holter ECG Data Using Empirical

Mode Decomposition and Statistical Approaches”, IEEE transactio on biomedical engineering , VOL. 59, NO. 6, pp- 1499-1506, JUNE 2012 [13]. Biomedical Signal Analysis, R.M. Rangayyan, Wiley Interscience,ISBN 9814-12-611-X, pp-247

[14] ] S Dash, E Raeder, S Merchant, K Chon, “A Statistical Approach for Accurate Detection of Atrial Fibrillation and Flutter”, Computers in Cardiology-2009, vol. 36, pp-137−140.

[15] Tran Thong, James McNames, Mateo Aboy, and Brahm Goldstein, “Prediction of Paroxysmal Atrial Fibrillation by Analysis of Atrial

Premature Complexes”, ieee tranction on biomedical engineering, Vol. 51, no. 4, pp-561-569, 2004

[16] Francisco Rinc´on, Paolo Roberto Grassi, Nadia Khaled, David Atienza and Donatella Sciuto, “Automated Real-Time Atrial Fibrillation

Detection on a Wearable Wireless Sensor Platform”, 34th Annual International Conference of the IEEE EMBS-2012, pp-2472-2475.

[17] S Pal, M Mitra, “ Empercal mode decomposition based ECG enhansment and QRS detection’, Computer in biology and medicine-2011, pp 84-92

[18] Sanchez, C., Millet, J., Rieta, J. J., Rodenas, J., Castells, F., Ruiz, R., and Ruiz, V., "Packet Wavelet Decomposition: An Approach For Atrial Activity Extraction," IEEE Computers in Cardiology, vol. 29 pp. 33-36. Memphis (TN), Sept.2002

[19] Richard P. M. Houben, Senior Member, IEEE, Natasja M. S. de Groot, and Maurits A. Allessie, “Analysis of Fractionated Atrial Fibrillation

Figure

Fig. 1. RR interval and P peaks of atrial fibrillation rhythm.
Fig. 2. IMFs of Normal and AF rhythm
Fig. 4. Positive halves of the selected IMFs
Fig. 6. Scatter plot of stander deviation Vs. Variance (a) IMF3; (b) IMF4; (c) IMF5
+2

References

Related documents

Unlike normal sinus rhythm, atrial fibrillation is characterized by disorganized, nearly continuous, and widespread electrical activity throughout the upper chambers of the

Antiarrhythmic Drug Therapy to Maintain Sinus Rhythm in Patients With Recurrent Paroxysmal or Persistent Atrial Fibrillation (Updated).. Maintenance of

Here, we provide deeper insights into the posttranscriptional mechanisms that regulate Tfh cell differ- entiation, function, and plasticity through the actions of RNA-binding

Since we observed a negative effect of homophily on polarization in the empirical analysis of climate change discussions on the Twitter platform, we conclude that tweets

At the dawn of 2020, though, when the threat from Covid-19 arose, the international community was more fragmented than at any time since 1945 because of a populist wave that has

There are 3 main types of financial reports generated from the recording model, namely Profit / (Loss) Report, Balance Sheet Report, and Cash Flow Statement. The Balance

The purpose of science education is no longer simply to train that tiny fraction of the population 

With the intermediate population projection, the OASDI payroll tax rate is 19.3 [22.2] percent higher and the benefit replacement rates are 13.9 [15.4] percent lower in 2078 (after