Effective Feature Extraction of ECG for Biometric Application

Available Available online online at at www.sciencedirect.com www.sciencedirect.com

ScienceDirect ScienceDirect Procedia Computer Science 00 (2017) 000–000 ScienceDirect Procedia Computer Science 00 (2017) 000–000 ScienceDirect

Available online at www.sciencedirect.com

Available online online at Available atwww.sciencedirect.com www.sciencedirect.com

ScienceDirect

Procedia Computer Science 00 (2017) 000–000 Procedia Computer Science 00 (2017) 000–000

Procedia Computer Science 115 (2017) 296–306

www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia www.elsevier.com/locate/procedia

7th 7th International International Conference Conference on on Advances Advances in in Computing Computing & & Communications, Communications, ICACC-2017, ICACC-2017, 22-24 22-24 August 2017, Cochin, India August 2017, Cochin, India 7th International Conference on Advances in Computing & Communications, ICACC-2017, 22-24 7th International Conference on Advances in Computing Communications, ICACC-2017, 22-24 August 2017, Cochin, & India Effective of August 2017, Cochin,for IndiaBiometric Effective Feature Feature Extraction Extraction of ECG ECG for Biometric Application Application a* b a*, P.Rajesh b Kiran Kumar Kumar Effective Extraction of for Biometric Application Kumar Patro Patro , P.Rajesh Effective Feature FeatureKiran Extraction of a*ECG ECG for Kumar Biometric Application b a aDepartment bDepartment bDepartment

of ECE, AUCE, Andhra University, Visakhapatnam-530003, India of ECE, AUCE, Andhra University, Visakhapatnam-530003, a* b India of ECE, AUCE, Andhra University, Visakhapatnam-530003, India Department of ECE, AUCE, Andhra University, Visakhapatnam-530003, India a [email protected], [email protected] Department of ECE, AUCE, Andhra University, Visakhapatnam-530003, India [email protected], [email protected] a bDepartment of ECE, AUCE, Andhra University, Visakhapatnam-530003, India Department of ECE, AUCE, Andhra University, Visakhapatnam-530003, India b Department of ECE, AUCE, Andhra University, Visakhapatnam-530003, India [email protected], [email protected] [email protected], [email protected]

Kiran Kumar Patro , P.Rajesh Kumar Kiran Kumar Patro , P.Rajesh Kumar

Abstract Abstract Biometric systems performing identity recognition based upon extracted informative data from an individual are vital for security Biometric Abstract systems performing identity recognition based upon extracted informative data from an individual are vital for security applications. The Abstract applications. The vital vital characteristics characteristics of of an an ECG ECG signal signal depend depend upon upon its its Characteristic Characteristic points’ points’ P, P, Q, Q, R, R, S S and and T. T. In In this this paper, paper, an an Biometric systems performing identity recognition basedforupon extracted informative data from anfragments individualareare vital foraccording security effective feature extraction method is proposed, in which each record of ECG, the best 6-PQRST extracted effective feature extraction method is proposed, in which for each record of ECG, the best 6-PQRST fragments are extracted according Biometric systems performing identityofrecognition baseddepend upon extracted informative data from P,anQ,individual forpaper, security applications. Theand vital characteristics an ECG signal upon itsfeatures Characteristic points’ R, S andare T. vital In this an to their positions normalized. A of finally the of set to priority priority basis basis their positions are are of normalized. A total totaldepend of 72 72 different different features are are calculated, calculated, finally theSperformance performance of feature feature applications. Theand vital characteristics an ECG signal upon its points’ P, fragments Q, R, and In this paper, set an effective feature extraction method is proposed, in which for eachis record of Characteristic ECG, the best 6-PQRST are T. extracted according is examined and compared using ANN. The proposed algorithm tested for MIT-BIH ECG ID database signals. is examined and compared using ANN. The proposed algorithm is tested for MIT-BIH ECG ID database signals. effective feature extraction method is proposed, in which for each record of ECG, the best 6-PQRST fragments are extracted according to priority basis and their positions are normalized. A total of 72 different features are calculated, finally the performance of feature set to priority basis and their positions are normalized. A total of 72 isdifferent features are calculated, finally the performance of feature set © 2017 Authors. Published Elsevier is examined compared usingby ANN. The B.V. proposed algorithm tested for MIT-BIH ECG ID database signals. © 2017 The The and Authors. Published Elsevier is examined and compared usingby ANN. The B.V. proposed algorithm is 7th tested for MIT-BIH ECG ID database signals. Peer-review under responsibility of the scientific committee of the International Conference on Advances Peer-review under responsibility of the scientific committee of the 7th International Conference on Advances in in Computing Computing & & Communications. © 2017 The Authors. Published by Elsevier B.V. Communications. © 2017 The under Authors. Published by B.V. committee of the 7th International Conference on Advances in Computing & Peer-review responsibility of Elsevier the scientific Keywords: Electrocardiogram (ECG); Time-Frequency approach; Biometric; PQRST fragments; Pattern Recognition (ANN); MIT-BIH ECG ID; Keywords: Electrocardiogram (ECG); approach; of Biometric; PQRST fragments; Patternon Recognition MIT-BIH Peer-review under responsibility of Time-Frequency the scientific committee the 7th International Conference Advances (ANN); in Computing & ECG ID; Communications. MATLAB MATLAB Communications. Keywords: Electrocardiogram (ECG); Time-Frequency approach; Biometric; PQRST fragments; Pattern Recognition (ANN); MIT-BIH ECG ID; Keywords: MATLAB Electrocardiogram (ECG); Time-Frequency approach; Biometric; PQRST fragments; Pattern Recognition (ANN); MIT-BIH ECG ID; MATLAB 1. Introduction

1. Introduction ECG signal ECG signal is is aa universal universal characteristic. characteristic. It It has has been been used used for for several several decades decades as as an an efficient efficient and and reliable reliable diagnostic diagnostic tool tool in in 1. Introduction medical applications. Recently, the possibility of using this ECG signal as a biometric tool has been suggested. 1. Introduction medical applications. Recently, the possibility of using this ECG signal as a biometric tool has been suggested. Its Its ECG signal is a supported universal characteristic. It has been used for several decades as an differences efficient andofreliable diagnostic tool in validity is by the and the under validity is well well supported by the the fact fact that thatIt both both the physiological physiological and geometrical geometrical differences ofreliable the heart heart under different different ECG signal is acertain universal characteristic. has been used for several as aanbiometric efficient and diagnostic tool Its in medical applications. Recently, the possibility of using this ECG decades signal signal as tool hasindividual been suggested. subjects reveal uniqueness in the signal characteristics [1][2] .ECG of each and every contains subjects reveal certain Recently, uniquenessthe in the signal characteristics [1][2] .ECG signal of each and every individual containsItsaa medical applications. possibility of using this ECG signal as a biometric tool has been suggested. validity is welldue supported by the fact that both the physiological and geometrical differences of the heart under different unique to differences in among The Electrocardiogram (ECG) indicates the unique pattern pattern due to existing existing differences in morphology morphology among individuals. individuals. The differences Electrocardiogram (ECG) indicates the validity is well supported by the fact thatsignal both the physiological and .ECG geometrical the heart under different subjects reveal certain uniqueness in the characteristics [1][2] signalofof each and of every individual contains a recording of electrical activity of human heart over aa period of time. The shape the waveform reveals the current state recording of electrical activity of human heart over period of time. The shape of the waveform reveals the current state subjects reveal certain uniqueness in the signal characteristics [1][2] .ECG signal of each and every individual contains a unique pattern due to existing differences in morphology among individuals. TheofElectrocardiogram (ECG) indicates the of the heart and it offers helpful information regarding the rhythm and function the heart [3] .One cardiac cycle in an of the heart anddue it offers helpful information regarding theamong rhythm and function ofElectrocardiogram the heart [3] .One(ECG) cardiacindicates cycle inthe an unique pattern to existing differences in morphology individuals. The recording ofconsists electrical activity of human heart over waves a period .. of time. The shape of the waveform reveals the current state ECG of the Characteristic ECG signal signal ofactivity the P-QRS-T P-QRS-T Characteristic recording ofconsists electrical ofinformation human heartregarding over waves a period of time.and Thefunction shape of reveals the current state of the heart and it offers helpful the rhythm ofthe thewaveform heart [3] .One cardiac cycle in an of the heart and it offers helpful information regarding the rhythm and function of the heart [3] .One cardiac cycle in an ECG signal consists of the P-QRS-T Characteristic waves. ECG signal consists of the P-QRS-T Characteristic waves.

Fig. 1. ECG signal Pulse [2] Fig. 1. ECG signal Pulse [2]

The depends upon the The biometric biometric systems systems which which are are based based on on ECE ECE generally generally depends upon[2] the nature nature of of the the features features such such as as Fiducial Fiducial or or Fig. 1. ECG signal Pulse Non-Fiducial [2][4][5]. For Fiducial based method, the detection of characteristic waves i.e. P-QRS–T. Fig. 1. ECG signal Pulse [2] Non-Fiducial [2][4][5]. For Fiducial based method, the detection of characteristic waves i.e. P-QRS–T. The biometric systems which characteristic are based on wave ECE generally depends upon the nature of the features suchfinding as Fiducial or Generally in the is Detection of is crucial task. R-peak Generally in ECG, ECG, the major major characteristic wave is R-peak. R-peak.depends Detection of R-peak R-peak is aaof crucial task. After After finding R-peak The biometric systems which areS,based on ECE generally upon theasnature the features such as fro Fiducial or Non-Fiducial [2][4][5]. For P, Fiducial based method, thetaking detection of characteristic waves i.e. P-QRS–T. location, other components Q, T are detected by R-peak location reference and tracing to and from Rlocation, other components P, Q, S, T are detected by taking R-peak location as reference and tracing to and fro from RNon-Fiducial [2][4][5]. For Fiducial based method, the detection of characteristic waves i.e. P-QRS–T. Generally in ECG, the major characteristicnormal wave is R-peak. Detection of R-peak is a crucialbased task. After finding R-peak peak relative position. The ECG with waves. represent the peak relative position. The Fig. Fig. 11 shows shows normal ECG with all all P-QRS-T P-QRS-T waves. isFiducial Fiducial based features features represent the Generally in ECG, the major wavemiscellaneous is of R-peak a crucial After finding R-peak location, other components P, characteristic Q,angle S, T are detected byR-peak. taking Detection R-peak location as reference and task. tracing to and fro from Rtime, amplitude, distance, slope, and some features [2][5][6]. time, amplitude, distance, slope, angle anddetected some miscellaneous features [2][5][6]. location, other components P, Q, S, T are by taking R-peak location as reference and tracing to and fro from Rpeak relative position. The Fig. 1 shows normal ECG with all P-QRS-T waves. Fiducial based features represent the *Corresponding author. Tel.:+91 9441029815. peak relative position. The Fig. 1 shows normal ECG with all P-QRS-T waves. Fiducial based features represent the *Corresponding author. Tel.:+91slope, 9441029815. time, amplitude, distance, angle and some miscellaneous features [2][5][6]. E-mail address:[email protected]. time, amplitude, distance, slope, angle and some miscellaneous features [2][5][6]. E-mail address:[email protected].

1877-0509 © 2017author. The Authors. Published by Elsevier B.V. *Corresponding Tel.:+91 9441029815. *Corresponding author. Tel.:+91Published 9441029815. E-mail address:[email protected]. Peer-review under responsibility of the by scientific committee of the 7th International Conference on Advances in Computing & 1877-0509 © 2017The Authors Elsevier B.V. 1877-0509 © 2017The Authors Published by Elsevier B.V. E-mail address:[email protected]. Communications Peer-review under responsibility of the scientific Peer-review under responsibility of the scientific committee committee of of the the 7th 7th International International Conference Conference on on Advances Advances in in Computing Computing & & 10.1016/j.procs.2017.09.138 Communications. 1877-0509 © 2017The Authors Published by Elsevier B.V. Communications. 1877-0509 © 2017The Authors Published Elsevier B.V. Peer-review under responsibility of thebyscientific committee of the 7th International Conference on Advances in Computing & Peer-review under responsibility of the scientific committee of the 7th International Conference on Advances in Computing & Communications. Communications.

Kiran Kumar Patro, P Rajesh Kumar/ Procedia Computer Science 00 (2017) 000–000

Kiran Kumar Patro et al. / Procedia Computer Science 115 (2017) 296–306

2. Methodology of Effective Feature Extraction In this paper, a new effective approach to the feature extraction is developed for biometric application which include four major steps; Data Pre-processing, Feature Extraction and Effective feature extraction finally Classification.

Fig. 2. Block Diagram for Effective Feature Extraction

2.1 Pre-Processing: The initial step of the Feature Extraction process is the removal of noise from the signal. The raw ECG is rather noisy and contains distortions of various origins. The different frequency components used while acquisition of the ECG signal may produce interference in the signal recording process. Such interference can add noise to the ECG signal. This unwanted modification in the signal may tamper the original information in the ECG signal and lead to false ECG data. For each possible case corruption in real ECG signal made by the introduction of unwanted noise, filtering methods has to be used. The appropriate filter has to be selected on the basis of possible type of noise present in the signal. In this paper, a cascaded digital filters configuration is used for removal of three major noises of Baseline Drift, Power Line Interference and EMG noise [7]. After noise filtering, Smoothing filter is applied to the de-noised ECG signal. Smoothing of signal creates an approximating function that attempts to capture important patterns in the data, while leaving out noise or other finescale structures/rapid phenomena. Thus, the roughness in the signal is removed by smoothing filter. The smoothed signal is good for accurate detection of peaks in ECG signal. A (weighted) moving average filter (Direct Form II Transposed) was used for the signal smoothing [8].

Fig. 3. Cascaded FIR Filter configuration [7]

297


298


2.2 Feature Extraction : In the feature extraction stage, all the Characteristic waves are identified and their intervals are calculated. Identification of R-peak is more important in ECG diagnosis. By taking R-peak as reference, the other characteristic wave was identified using Frequency-Time based methodology and Pan Tompkins algorithm [14]. In this, frequency domain analysis (wavelet transform) is used for detection of R-peak and Time domain analysis is used for detection of other characteristic waves [9]. 2.2.1 Wavelet Transform Approach: The Wavelet Transform is a time-scale analysis and used on wide range of applications, in particular signal compression. Nowadays, wavelet transform is used to solve problems in Electro cardiology, including compression of data, ventricular late potentials analysis, and finally ECG characteristic waves detection [9][10]. Wavelet transform decomposes the given signal into a number of levels related to signal frequency components and analyses each level with particular resolution [11]. In this, the ECG signal is decomposed into 4 levels using Symlet wavelet (sym4) transform for finding R-peak. In wavelet decomposition, it down samples the original ECG signal, as a result the samples are reduced and QRS complex is retained. Symlet wavelet transform is the modified version of Daubechies wavelet with increased symmetry [9]. Decomposed signals are noise free signals, and by making a threshold of 60% of maximum value. The values which are above threshold are invariably R-peaks [12]. The decomposed signal can be reconstructed into actual signal by first multiplying the down sampled signal into 4, so that R-peaks are detected in actual signal.

2.2.2 Time domain Analysis: Frequency domain approach is used for the detection of R-Peak only and after time domain approach is used for other characteristic wave’s detection [13]. R-R interval can be calculated by �𝑅𝑅𝑅𝑅𝑅 �� =

𝑅𝑅��𝑅𝑅𝑅�� 𝑓𝑓𝑠𝑠

�𝑠𝑠 = �� = ��

��

For identifying P-wave, a window in time domain is created with time gap limits from 65% of R-R interval to 95% R-R interval which is added to same R-peak location. In that window, the maximum value will represent P-wave. For P- wave � = �� 0.� � �𝑠𝑠 � � = �� 0.0� � �𝑠𝑠 � Where a & b are time domain window ranges. The Q-wave is identified by choosing minimum value in Time based window starting from 20ms before corresponding R-peak. For Q- wave � = �� 0.0�� 𝑠𝑠 � � = �� Where a & b are time domain window ranges. Similarly S-peak is detected by selecting least value in time based window after R-peak location. For S- wave � = �� = �� 0.0� � �𝑠𝑠 � For identifying T-wave, a window in time domain is created with time gap limits from 15% of R-R interval to 55% R-R interval which is added to same R-peak location. In that window, the maximum value will represent T-wave. For T- wave � = �� 0.0� � �𝑠𝑠 � � = �� 0.� � �𝑠𝑠 �


299

Kiran Kumar Patro, P Rajesh Kumar/ Procedia Computer Science 00 (2017) 000–000 Filtered Ecg Signal with Peaks Detected 1

ECG Signal R(avoided) R P Q S T R (best 6) P (best 6) Q (best 6) S (best 6) T (best 6)

0.8

electrical activity

0.6

0.4

0.2

0

-0.2

-0.4

0

1000

2000

3000 ECG sample

4000

5000

6000

Fig. 4. ECG signal with Peaks detected

The time domain windows are adaptive because they depend on R-R interval values. The detected peaks are shown in Fig.4. The PQRST fragment marked with red are 6 best fragments among all available in the given signal. The best fragments were selected by their least difference from the mean of the Euclidian distance of the peaks. Only these 6 best fragments are used for the data extraction. 2.3. Effective Feature Extraction From the sets of time and amplitude relation between peaks of ECG signal, several features of the given ECG signal can be extracted. For each ECG record, the signal was broken into each independent pulse ranging from the P-peak to T-peak of the pulse (PQRST-fragment) since the information from PQRST-fragment samples are useful to define the state of the heart. The process of Effective feature extraction explained by below steps as follows [15]. 1.

From the set of PQRST-fragments, the "mean" PQRST- fragment was estimated using Euclidean distance. Only the 6 closest PQRST-fragments were selected for further analysis. Best 6 ECG Pulses 1

0.8

electrical activity

0.6

0.4

0.2

0

-0.2

-0.4

0

20

40

60

80

100 ECG sample

120

140

160

180

200

Fig. 5. PQRST-fragments

2. To get comparative data from the best 6 PQRST fragments, their position has to be normalized at any standard point. Thus R-peak of all such fragments were assumed to be at origin and the position of their respective P, Q, S and T peaks were translated to new co-ordinates keeping their original aspect ratio of position unchanged. Best 6 ECG Pulses with R peak position normalized 0

-0.2

electrical activity

-0.4

-0.6

-0.8

-1

-1.2

-1.4 -100

-50

0

50 ECG sample

Fig. 6.PQRST-fragments with position normalized w.r.t. ‘R’

100

150



300

Thus in the initial feature space (dimension N=72) the ECG appears as a set of 6 PQRST- fragments with each seen as a separate pattern at subsequent system stages, to be interpreted and classified independently. All the features were calculated with respect to the position of R being origin. The features extracted from PQRST fragment were related to amplitude (y), time duration (x), difference between amplitude(amp), distance (dis), difference between times (time),slope, angle (angl), area and ratio of some features [16]. Table I. The set of features (Time, Amplitude) extracted with Respect to the position of ‘R’ as origin Features

I. Time Features

II.Amplitude Features

III. Distance Features

IV.Slope Features

V. Angle Features

VI. Miscellaneous Features

Feature Description

1.Px 2.Qx 3.Sx

4.Tx 5.PQ 6.PT

7.QS 8.QT 9.ST

10.PS 11.PT/QS 12.QT/QS

13.Py

19.RS

25.ST/QS

31.PQ/RS

14.Qy

20.ST

26.RS/QR

32.RS/QS

15.Sy

21.QS

27.PQ/QS

33.RS/QT

16.Ty

22.PS

28.PQ/QT

34.ST/PQ

17.PQ

23.PT

29.PQ/PS

35.ST/QT

18.QR

24.QT

30.PQ/QR

36.PQ

38.RS

40.QS

42. ST/QS

37.QR

39.ST

41.PR

43.RS/QR

44.PQ

47.ST

50.PS

45.QR

48.QS

51.QT

46.RS

49.PT

52.PR

53.PQR

55.RST

57.RSQ

54.QRS

56.RQS

58.RTS

59.QRS area

63.S angl/QTtime

67.(Q/T) angl

71.QRSxcentroid

60.QRSarea/RSˆ2

64.S angl/PQ dis

68.QRSarea/QRamp

72.QRSycentroid

61.(R/S)angl

65.(R/Q) angl

69.QRS perimeter

62.R angl/QStime

66.(R/T) angl

70.QRS in radius

2.4 Classification The biometric identification of the ECG signal can be done by feeding the data of ECG features to an Artificial Neural Network (ANN). For the purpose of generating Neural Network for Pattern Recognition a two layered feed forward network was used. A feed forward neural network is an artificial neural network where connections between the units do not form a directed cycle. A two layered feed forward network, with sigmoid hidden and softmax output neurons, can classify vectors arbitrarily well, given enough neurons in its hidden layers. All other layers are called hidden layers. A layer that produces the network output is called an output layer [17]. At the primary stage the neural network has to be trained with the ECG data of different persons. Then afterwards the neural network generated from the training is used for biometric identification of the persons.



301

3. Experimentation and Results In this Paper for numerical experimentation, MIT-BIH ECG ID data base with 12 subjects was used for training and testing purposes [19]. The database includes five records for each individual subject in those two signals for training, three signals for testing. MATLAB Version 8.3.0.532 (R2014a) was used for all data processing, Data Analysis and application of Effective feature extraction algorithm [20]. Raw Ecg Signal 150

electrical activity

100

50

0

-50

0

500

1000

1500

2000

2500

3000

3500

4000

4500

5000

ECG Sample

Fig. 7. ECG signal before processing

Fig. 7 represents the raw ECG signal obtained on examining a person. At this stage, the signal contains some artifacts which are to be removed before extracting data from it. Noise filtered ECG Signal 150

electrical activity

100

50

0

-50

0

1000

2000

3000 ECG sample

4000

5000

6000

Fig. 8.Noise filtered ECG signal Smoothed ECG Signal 1 0.8

electrical activity

0.6 0.4 0.2 0 -0.2 -0.4

0

1000

2000

3000 ECG sample

4000

5000

6000

Fig. 9.SmoothedECGsignal

Fig. 8 represents Noise filtered signal and Fig. 9 represents Smoothed ECG signal. In this, a cascaded digital filters configuration is used for removal of three major noises of Baseline Drift, Power Line Interference and EMG noise. Smoothing filter is applied for accurate detection of peaks .

Kiran Patro et al. / Procedia Computer Science 115000–000 (2017) 296–306 Kiran Kumar Patro,Kumar P Rajesh Kumar/ Procedia Computer Science 00 (2017)

302

ECG signal with all Peaks 1 0.8

electrical activity

0.6 0.4 0.2 0 -0.2 -0.4 -0.6 -0.8

400

500

600

700 ECG Sample

800

900

1000

Fig. 10. ECG signal with Characteristic Waves

Fig. 10 represents ECG signal with all characteristic waves P-QRS-T and their ON-OFF intervals. ECG Signal with Characteristic Points Detected 1

ECG Signal R(avoided) R P Q S T R (best 6) P (best 6) Q (best 6) S (best 6) T (best 6)

0.8

electrical activity

0.6 0.4 0.2 0 -0.2 -0.4

0

1000

2000

3000 ECG sample

4000

5000

6000

Fig. 11. ECG signal with peaks detected

The detected peaks are shown in Fig.11. The PQRST fragments marked with red are 6 best fragments among all available in the given signal. Best 6 ECG Pulses with R-Peak Position normalized 0

0.8

-0.2

0.6

-0.4 electrical activity

electrical activity

Best 6 ECG Pulses 1

0.4 0.2

-0.6 -0.8

0

-1

-0.2

-1.2

-0.4

0

20

40

60

80

100 ECG sample

120

140

Fig. 12.ECG pulse of best six PQRST fragment

160

180

200

-1.4 -80

-60

-40

-20

0

20 ECG sample

40

60

80

100

120

Fig. 13. ECG pulse of best six fragments with position Normalized with respect to R.

Fig. 12 shows the best 6 PQRST fragments before their position was normalized. The best fragments were selected by their least difference from the mean of the Euclidian distance of the peaks. Only these 6 best fragments are used for the data extraction. But the broken pulse is still not appropriate enough to compare their data. Before extracting data from such pulses, their position has to be normalized taking a standard origin point common for all of them. The nonuniformity in the pulse can be seen in Fig. 12. Fig. 13 shows the ECG pulse of best 6 PQRST fragments plotted at normalized position considering R as at origin. All the features were calculated with respect to the position of R being the origin.



303

Peak points of Best 6 ECG Pulses with their combined mean 0 -0.2

electrical activity

-0.4 -0.6 -0.8 -1 -1.2 -1.4 -80

-60

-40

-20

0

20

40

60

80

100

120

ECG sample

Fig. 14. ECG pulse of best six PQRST fragment with their combined mean

Fig. 14 shows the ECG pulse of best 6 PQRST fragments with their combined mean. Mean is the statistical feature which is helpful to reduce feature size. The ECG signal Features are extracted with a sampling frequency �𝑠𝑠 = �� and reference R peak at origin (�� ) = (0,0) Table 3 describes Time, Amplitude, Distance, Slope, Angle & Miscellaneous features of best 6 P-QRS-T fragment pulses in addition to their mean value. The mean of each feature was recorded for all ECG records in training and test sets The ECG signal was taken from MIT-BIH ECG ID Database/Person. A total of 72 features were extracted . The obtained features were subjected to a two layered feed forward network, with Sigmoid hidden and softmax output neurons, which can classify vectors arbitrarily well, given enough neurons in its hidden layers. The network were trained with scaled conjugate gradient back propagation and tested over various ECG sets of different individuals. The 72 extracted features are applied to input and 22 hidden layers were taken for the 12 output target (Person) to be tested

Fig. 15. Neural Network used for Pattern Recognition

.

MIT-BIH ECG ID Database was used for training and testing purposes. This database was chosen because it includes more than one record for some of its subjects. Each individual person having 5 records out of 5 records 2 records for train set and 3 records for testing. The Biometric system is analyzed for 12 subjects of each 5 signals with duration of 6 months has been taken, so a total of 24 signals for training and 36 signals for testing . Table II. Train and Test signals of ECG Train Set Includes record-1 and record-2 of each Subject P3, P10, P24, P25, P30,P32, P34, P36, P52, P53, P59, P72

Test Set Includes record-3, record-4, record-5 of each subject P3, P10, P24, P25, P30,P32, P34, P36, P52, P53, P59, P72

The biometric system performance from ANN is obtained by examining the Confusion Matrix; generally Confusion Matrix is a specific table layout that allows visualization of the performance of an algorithm. Each column of the matrix represents the instances in a predicted class, while each row represents the instances in an actual class.

Patro et al. / Procedia Computer 115000–000 (2017) 296–306 Kiran KumarKiran Patro,Kumar P Rajesh Kumar/ Procedia Computer ScienceScience 00 (2017)

304

TRUE CLASSIFICATION OF PULSES ANN Confusion Mat 18 0 6 0 0 4 0 0 0 0 0 0 64.3% 7.1% 0.0% 2.4% 0.0% 0.0% 1.6% 0.0%0.0% 0.0% 0.0% 0.0% 0.0%35.7% 0 21 0 0 0 0 0 3 0 0 7 16 44.7% P10 0.0% 8.3% 0.0% 0.0% 0.0% 0.0% 0.0%1.2% 0.0% 0.0% 2.8% 6.3%55.3% 3 0 15 0 0 0 3 0 0 0 0 0 71.4% P24 1.2% 0.0% 6.0% 0.0% 0.0% 0.0% 1.2%0.0% 0.0% 0.0% 0.0% 0.0%28.6% 0 0 0 20 0 1 0 0 0 0 0 0 95.2% P25 0.0% 0.0% 0.0% 7.9% 0.0% 0.4% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 4.8% 0 0 0 1 21 0 0 0 0 0 0 0 95.5% P30 0.0% 0.0% 0.0% 0.4% 8.3% 0.0% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 4.5% 0 0 0 0 0 16 0 0 0 0 0 0 100% P32 0.0% 0.0% 0.0% 0.0% 0.0% 6.3% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 18 0 0 0 0 0 100% P34 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 7.1%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 14 0 0 0 0 100% P36 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%5.6% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 0 21 0 0 0 100% P52 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 8.3% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 0 0 21 0 0 100% P53 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 8.3% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 4 0 0 13 0 76.5% P59 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%1.6% 0.0% 0.0% 5.2% 0.0%23.5% 0 0 0 0 0 0 0 0 0 0 1 5 83.3% P72 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 0.4% 2.0%16.7% 85.7%100%71.4% 95.2%100%76.2% 85.7% 66.7%100%100%61.9% 23.8% 80.6% 14.3%0.0%28.6%4.8% 0.0%23.8% 14.3% 33.3%0.0% 0.0%38.1% 76.2% 19.4% P3 P10 P24 P25 P30 P32 P34 P36 P52 P53 P59 P72

Output Class

P3

Target Class Fig. 16. Confusion Matrix for True Classification of ECG Pulse data of Person

The Fig. 16 shows the confusion matrix for the true classification of ECG pulse data of Person used for testing the pattern recognition neural network. The confusion matrix shows the number of pulses used for the test with their corresponding True Positive Rate (TPR). According to the given confusion matrix, the TPR of overall true classification of pulses in testing session was 80.60%. TRUE IDENTIFICATION OF PULSES ANN Confusion Matrix 21 0 0 0 0 7 0 0 0 0 0 0 75.0% 8.3% 0.0% 0.0% 0.0% 0.0% 2.8% 0.0%0.0% 0.0% 0.0% 0.0% 0.0%25.0% 0 21 0 0 0 0 0 0 0 0 7 14 50.0% P10 0.0% 8.3% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 2.8% 5.6%50.0% 0 0 21 0 0 0 0 0 0 0 0 0 100% P24 0.0% 0.0% 8.3% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 21 0 0 0 0 0 0 0 0 100% P25 0.0% 0.0% 0.0% 8.3% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 21 0 0 0 0 0 0 0 100% P30 0.0% 0.0% 0.0% 0.0% 8.3% 0.0% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 14 0 0 0 0 0 0 100% P32 0.0% 0.0% 0.0% 0.0% 0.0% 5.6% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 21 0 0 0 0 0 100% P34 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 8.3%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 21 0 0 0 0 100% P36 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%8.3% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 0 21 0 0 0 100% P52 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 8.3% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 0 0 21 0 0 100% P53 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 8.3% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 0 0 0 14 0 100% P59 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 5.6% 0.0% 0.0% 0 0 0 0 0 0 0 0 0 0 0 7 100% P72 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 0.0% 2.8% 0.0% 100%100%100%100%100%66.7%100%100%100%100%66.7% 33.3%88.9% 0.0% 0.0% 0.0% 0.0% 0.0%33.3%0.0%0.0% 0.0% 0.0%33.3% 66.7%11.1% P3 P10 P24 P25 P30 P32 P34 P36 P52 P53 P59 P72

Output Class

P3

Target Class Fig. 17. Confusion Matrix for True Classification of ECG Pulse data of Person

The Fig. 17 shows the confusion matrix for the true identification of ECG pulse data of Person used for testing the pattern recognition neural network. From the confusion matrix, the TPR of overall true identification of pulses in testing session was 88.90% which is 8.30% increment compared to true classification of pulses True Identification is, in each set of pulse data contains 7 pulses. But while applying neural network some of the pulses of a particular set may be misclassified under two or more different person. These misclassified pulses are then categorized for



true person for whom the true classified number of pulses out of 7 pulses in the pulse set is maximum. TRUE IDENTIFICATION OF PERSON ANN Confusion Matrix 3 0 0 0 0 1 0 0 0 0 0 0 75.0% 8.3% 0.0% 0.0% 0.0% 0.0% 2.8% 0.0%0.0% 0.0% 0.0% 0.0% 0.0%25.0% 0 3 0 0 0 0 0 0 0 0 1 2 50.0% P10 0.0% 8.3% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 2.8% 5.6%50.0% 0 0 3 0 0 0 0 0 0 0 0 0 100% P24 0.0% 0.0% 8.3% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 3 0 0 0 0 0 0 0 0 100% P25 0.0% 0.0% 0.0% 8.3% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 3 0 0 0 0 0 0 0 100% P30 0.0% 0.0% 0.0% 0.0% 8.3% 0.0% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 2 0 0 0 0 0 0 100% P32 0.0% 0.0% 0.0% 0.0% 0.0% 5.6% 0.0%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 3 0 0 0 0 0 100% P34 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 8.3%0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 3 0 0 0 0 100% P36 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%8.3% 0.0% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 0 3 0 0 0 100% P52 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 8.3% 0.0% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 0 0 3 0 0 100% P53 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 8.3% 0.0% 0.0% 0.0% 0 0 0 0 0 0 0 0 0 0 2 0 100% P59 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 5.6% 0.0% 0.0% 0 0 0 0 0 0 0 0 0 0 0 1 100% P72 0.0% 0.0% 0.0% 0.0% 0.0% 0.0% 0.0%0.0% 0.0% 0.0% 0.0% 2.8% 0.0% 100%100%100%100%100%66.7%100%100%100%100%66.7% 33.3%88.9% 0.0% 0.0% 0.0% 0.0% 0.0%33.3%0.0%0.0% 0.0% 0.0%33.3% 66.7%11.1% P3 P10 P24 P25 P30 P32 P34 P36 P52 P53 P59 P72

Output Class

P3

Target Class Fig. 18. Confusion Matrix for True Identification of Person

Each set of pulse data contains 7 pulses, thus the 7 pulse represents one individual. Therefore the number of persons involved in the test is 7 times less than the number of total pulses in each category. The Fig. 18 shows the confusion matrix for the true identification of Person used for testing the pattern recognition neural network. According to the confusion matrix, the TPR of overall true identification of person involved in testing session was 88.90%. The performance of the classifier can be estimated by examining the performance parameters of Classification Accuracy, Sensitivity, Specificity, Precision and F-score. All the performance parameters are calculated using Confusion Matrix for true identification of Person. Finally the proposed system for biometrics, with effective feature extraction method gives best performance compared to other approaches [18]. Table III. Performance Parameters of the system S.No

Performance Parameter

Value

1

True_Positive Rate (Classification)

80.5556%

2

True_Positive Rate (Identification)

88.889%

3

Over all_Accuracy

98.1481%

4

Over all_Recall

88.8889%

5

Over all_Specificity

98.9899%

6

Over all_Precision

93.750%

7

Over all_F Score

88.5317%

4.Conclusion This paper provides an overview of major steps in ECG signal analysis of De-noising ECG, Characteristic Points identification, feature extraction and effective feature extraction finally classification. The feature extraction methodology extracts the features of each heartbeat after automatic detection of R-peak. The proposed work concentrates

305

306



mainly on Effective feature extraction algorithm that contains both morphological and temporal features of each heart beat in ECG signal. For each ECG signal, the best 6 P-QRS-T fragments were selected by their least difference from the mean of the Euclidian distance of the peaks and their positions were normalized according to the position of R-peak. Only these 6 best fragments were considered for data extraction. A good feature extraction methodology can accurately work for Biometric applications. A total of 72 features were calculated. Performance of the feature set is examined by ANN for Pattern Recognition classifier. The results obtained are very promising to support ECG as one of the Biometric feature with accurate feature extraction scheme. For analysis the data from 12 subjects are taken from MIT-BIH ECGID database with 6 months of duration, a total of 24 signals for testing and 36 signals for training. Finally the proposed system shows 98.1481% over all accuracy for Biometric recognition. References [1] K. Revett, Behavioral Biometrics: A Remote access Approach, John Wiley & Sons, ISBN: 978-0-470-51883-0, 2008; 145-152. [2] Y. Wang, F. Agrafioti, D. Hatzinakos, K. Plataniotis, “Analysis of Human Electrocardiogram for Biometric Recognition”, EURASIP Journal of advances in signal process, 2008; 2 : 1-11. [3] A. C. Guyton, J. E. Hall, Textbook of Medical Physiology (11th ed.). Philadelphia: Elsevier Saunder. 2006. [4] M.Tantavi, K. Revett “ On the use of the electrocardiogram for biometric authentication”, IEEE International Conference on Informatics and Systems (INFOS), 2012; 2 (1):14-16. [5] M.Tantawi, A. Salem and M.F. Tolba “ ECG signal analysis for Biometric Recognition” , I IEEE International Conference on Hybrid Intelligent Systems (HIS), 2014;1(2) : 169-175 [6] R. Murugavel, Heart-Rate and EKG Monitor Using the MSP430FG439, Application Report SLAA280A, 2005; 5:1-12. [7] Kiran Kumar Patro, Dr. P Rajesh Kumar “De-Noising of ECG raw Signal by Cascaded Window based Digital filters Configuration”, IEEE Conference IEEE PCITC ,2015 ; 120-124. [8] D. Jeyarani, T. Jaya Singh, “Analysis of Noise Reduction Techniques on QRS ECG Waveform - by Applying Different Filters”, IEEE conference on Recent Advances in Space Technology Services and Climate Change (RSTSCC), Chennai, 2010; 149-152. [9] L. Williams and Wilkins, ECG Interpretation made incredibly easy. 5th edition, Kluwer Wolters, 2011. [10] Saminu, S.; Ozkurt, N.; Karaye, I.A. “Wavelet feature extraction for ECG beat classification” IEEE 6th International Conference on Adaptive Science & Technology (ICAST), 2014; 1-6. [11] P.S. Addison, “Wavelet transforms and the ECG: a re-view,” Physiological measurement, 2005; 26 (5): 155-159. [12] B.U. Kohler, C. Hennig, and R. Orglmeister, “The principles of software QRS detection,” IEEE Eng Bio, 2002; 21: 42–57. [13] Kiran Kumar Patro, Prof. P Rajesh Kumar “A Novel Frequency-Time based approach for the Detection of Characteristic Waves in Electrocardiogram Signal” Springer LNEE, 2016; 372: 57-67. [14] J. Pan, W.J. Tompkins, A real-time QRS detection algorithm. IEEE Transactions on Biomedical Engineering, 1985; 32: 230-236. [15] LugovayaT.S.,“Biometric human identification based on electrocardiogram”, [Master's thesis] Faculty of Computing Technologies and Informatics, Electrotechnical University "LETI", Saint-Petersburg, Russian Federation; 2005. [16] Singh, Y.N.; Gupta, P. “ECG to Individual Identification” 2nd IEEE International Conference on Biometrics Theory, Applications and Systems, 2008. BTAS 2008; 2 (1): 1-8. [17] Ashutosh Gupta, Betsy Thomas, “Neural Network based indicative ECG Classification” 5th IEEE international Conference (Confluence),2014; 5 (2): 277-279. [18] Takoua Hamdi, Anis Ben Slimane, Anouar Ben Khalifa, “A Novel Feature Extraction Method in ECG Biometrics”,IEEE International Image Processing Applications and Systems Conference 2014; 1-5. [19] http://www.physionet.org/cgi-bin/atm/ATM – “MIT-BIH ECG ID Database. [20] Matlab help, MATLAB MATHWORKS.