4
978-1-4799-1467-8/13/$31.00 ©2013 IEEE ECG Analysis for Person Identification Somsanouk Pathoumvanh, Surapan Airphaiboon Electronics Department, Faculty of Engineering King Mongkut’s Institute of Technology Ladkrabang Ladkrabang, Bangkok 10520 Thailand [email protected] Benjawan Prapochanung, Thurdsak Leauhatong Electronics Department, Faculty of Engineering King Mongkut’s Institute of Technology Ladkrabang Ladkrabang, Bangkok 10520 Thailand [email protected] AbstractElectrocardiogram (ECG) has been actively proposed as aliveness biometric. In this paper, the study which concern to a realistic application is proposed. Firstly, a single lead normal ECG signal is acquired from individuals of 10 subjects. Then, each single beat ECG is segmented and analyzed in Continuous Wavelet Transform (CWT) domain. Total energy of wavelet coefficients for each P, QRS, and T segment is calculated. Next, the Fisher Linear Discriminant Analysis (FLDA) is applied. Finally, normalized Euclidean distance is implemented as a classifier. In experimental results, 97% of classification accuracy is achieved in case of a normal ECG (with non-variation of heart rate). KeywordsECG Biometrics; Single Beat ECG features extraction; ECG Identifications I. INTRODUCTION The authentication system using biometric is newly accepted as another standard for person identification. Because the biometric provides higher security and the biometric features are usually based on the anatomical, physiological and/or behavioral characteristics which are significantly unique to individual. ECG features has more advantages than other biometrics, such as: difficult to disguise, regenerate, and can be acquired with a simple low-cost technology, (electrodes placed on the appropriate location on a human body). Typically ECG can be acquired less than 500ms and is painless. However, the disadvantage of ECG signal contains the heart rate variability (HRV) among the heartbeats. However, the disadvantage of ECG signal contains the heart rate variability (HRV) among the heartbeats. However, there are many researchers have been investigated and validated the characteristics of ECG signal as a new biometric such as: L. Biel et al. (1) who are the first to introduce the application of ECG as biometric, S. A. Israel et al. (2) are employed the features name: fiducial points detection by measure the temporal on the ECG features, and as well as the group of Foteini Agrafioti et al. (3)-(5) are implemented both fiducial points (temporal and amplitude) and non-fiducial features. From above literature reviews, two problems are needed to be improved. (i) The features extraction method by using fiducial points detection are very difficult to obtain the onset position of P, QRS, and T wave as reported by Foteini et al. (6) . Therefore, a combination of autocorrelation and discrete cosine transform (AC/DCT) is proposed. However, a few seconds of delay time from the ECG features acquiring process is consumed. (ii) In training process, most of research works with a huge numbers of ECG signal from each subjects. This shown that the learning ability of the system is very slow. Thus, the main contributions in this research work are the proposals of the reasonable solution for mention problems. (i) Proposed a robust features extraction using coefficients for each P, QRS, and T segment of ECG in CWT domain. (ii) Study the desire “single beat ECGas biometric features. In practical application, the learning ability of the system must be as fast as possible while a classification accuracy of the system still remains. II. PROPOSED METHODS A. Training and Testing Process The training process is mainly consisting of, FIR filter and DC baseline drift correction, Single Beat ECG feature extraction, and continuous wavelet transform. The overall block diagram is shown in Fig. 1. Fig. 1. Block diagram of proposed system B. Single Beat Length Selection After filtering, each individual single beat ECG signal is extracted. Normally, a single beat length of each dataset is non- linearity and has a small variation R-R of length. Then, each window of individual beat is considered by the novel beat length selection process (7) . The only requirement is the window of length N should not longer than the average heart The 2013 Biomedical Engineering International Conference (BMEiCON-2013)

[IEEE 2013 6th Biomedical Engineering International Conference (BMEiCON) - Amphur Muang, Krabi, Thailand (2013.10.23-2013.10.25)] The 6th 2013 Biomedical Engineering International

Embed Size (px)

Citation preview

978-1-4799-1467-8/13/$31.00 ©2013 IEEE

ECG Analysis for Person Identification

Somsanouk Pathoumvanh, Surapan Airphaiboon Electronics Department, Faculty of Engineering

King Mongkut’s Institute of Technology Ladkrabang Ladkrabang, Bangkok 10520 Thailand

[email protected]

Benjawan Prapochanung, Thurdsak Leauhatong Electronics Department, Faculty of Engineering

King Mongkut’s Institute of Technology Ladkrabang Ladkrabang, Bangkok 10520 Thailand

[email protected]

Abstract—Electrocardiogram (ECG) has been actively proposed as aliveness biometric. In this paper, the study which concern to a realistic application is proposed. Firstly, a single lead normal ECG signal is acquired from individuals of 10 subjects. Then, each single beat ECG is segmented and analyzed in Continuous Wavelet Transform (CWT) domain. Total energy of wavelet coefficients for each P, QRS, and T segment is calculated. Next, the Fisher Linear Discriminant Analysis (FLDA) is applied. Finally, normalized Euclidean distance is implemented as a classifier. In experimental results, 97% of classification accuracy is achieved in case of a normal ECG (with non-variation of heart rate).

Keywords—ECG Biometrics; Single Beat ECG features extraction; ECG Identifications

I. INTRODUCTION The authentication system using biometric is newly

accepted as another standard for person identification. Because the biometric provides higher security and the biometric features are usually based on the anatomical, physiological and/or behavioral characteristics which are significantly unique to individual. ECG features has more advantages than other biometrics, such as: difficult to disguise, regenerate, and can be acquired with a simple low-cost technology, (electrodes placed on the appropriate location on a human body). Typically ECG can be acquired less than 500ms and is painless. However, the disadvantage of ECG signal contains the heart rate variability (HRV) among the heartbeats. However, the disadvantage of ECG signal contains the heart rate variability (HRV) among the heartbeats. However, there are many researchers have been investigated and validated the characteristics of ECG signal as a new biometric such as: L. Biel et al. (1) who are the first to introduce the application of ECG as biometric, S. A. Israel et al. (2) are employed the features name: fiducial points detection by measure the temporal on the ECG features, and as well as the group of Foteini Agrafioti et al. (3)-(5) are implemented both fiducial points (temporal and amplitude) and non-fiducial features. From above literature reviews, two problems are needed to be improved. (i) The features extraction method by using fiducial points detection are very difficult to obtain the onset position of P, QRS, and T wave as reported by Foteini et al. (6). Therefore, a combination of autocorrelation and discrete cosine transform (AC/DCT) is proposed. However, a few seconds of delay time from the ECG features acquiring process is consumed. (ii) In training process, most of research works

with a huge numbers of ECG signal from each subjects. This shown that the learning ability of the system is very slow.

Thus, the main contributions in this research work are the proposals of the reasonable solution for mention problems. (i) Proposed a robust features extraction using coefficients for each P, QRS, and T segment of ECG in CWT domain. (ii) Study the desire “single beat ECG” as biometric features. In practical application, the learning ability of the system must be as fast as possible while a classification accuracy of the system still remains.

II. PROPOSED METHODS

A. Training and Testing Process The training process is mainly consisting of, FIR filter and

DC baseline drift correction, Single Beat ECG feature extraction, and continuous wavelet transform. The overall block diagram is shown in Fig. 1.

Fig. 1. Block diagram of proposed system

B. Single Beat Length Selection After filtering, each individual single beat ECG signal is

extracted. Normally, a single beat length of each dataset is non-linearity and has a small variation R-R of length. Then, each window of individual beat is considered by the novel beat length selection process (7). The only requirement is the window of length N should not longer than the average heart

The 2013 Biomedical Engineering International Conference (BMEiCON-2013)

978-1-4799-1467-8/13/$31.00 ©2013 IEEE

beat rate, and it is not contain multiple R peaks in one window, but it is necessary to contain all P, QRS, and T wave in one complete heartbeat periods. The algorithm is shown in Fig. 2.

Fig. 2. Single beat length estimation.

Where BR and AR are the window’s lenght before and after considering R peak respectively. R-R is a complete beat length.

C. Continuous Wavelet Transform The CWT is decomposing operation of ECG signal with different window width, called the scale or resolution of observation. The wavelet transform of a continuous time signal, S(t), is defined in equation (1).

(1)

Where, a is a dilation parameter and b is a location (shifting) parameter of the wavelet respectively.

D. Fisher Linear Discriminant Analysis Fishers Linear Discriminant (FLD) (8) is a widely used method for dimension reduction. Fishers idea utilizes a ratio of between-class and within-class scatters of a maximized training sample set.

All partition of ECG signal is projected onto FLD space. The trained features vector can be represented by all biometrics features in matrix Y, shown in equation (2).

(2)

Where Fmyn represent the feature of m beats ECG random input that is used for training and their projected FLD component into n classes.

E. Square Norm Euclidean Distance The number of the new bases is obtained from the number of the vectors that give the best discriminated class when measuring with the normalized Euclidean distance, and the nearest neighbor is the result of the classifier. The normalized Euclidean distance function D(y,μ) measures the distance between FLD feature vectors y and the center of each class μ in FLD feature space as defined in equation (3).

(3)

Where, v is the number of dimension in the feature vectors, y is the location of the Testing feature and μ is the center of the each Training feature.

After obtaining the distance between the unknown subject and all trained vectors, the classifier applies to the majority of the shortest distances between the trained vectors of the cluster.

F. Evaluation Methods To evaluate the performance of the proposed method, the accuracy parameter is calculated as shown in equation (4).

(4)

Where, TC is true classification. FC is false classification. AC is the accuracy of the total correctly classify

This equation is the ratio between the True classifications over the True and False classifications or total subjects uses in this experiment.

III. EXPERIMENTALS SETUP AND RESULT A normal ECG signal is captured by using Bio-Pac system

(9) as shown in Fig.3. The ECG measurements are the standard modified limb lead II (MLII), measure with Ag/AgCl surface electrode, with sampling rate at 500 samples per second.

Fig. 3. Bio-Pac system.

There are 10 volunteers are involved in the ECG signal recording to be representing for 10 IDs. There are 8 male and 2 female. In each record, a random selected of 10 sets of a single heartbeat for training and another 50 sets of heartbeats for testing are performed.

Before start recording, the volunteers have to relax, listen to the instruction of recording and answer some questions as shown in Table 1.

978-1-4799-1467-8/13/$31.00 ©2013 IEEE

TABLE I. analysis results from the questionnaire

To demonstrate the features extraction as referred to the experimental setup. After the novel beat length is selected, the single beat ECG feature extraction is analyzed with the continuous wavelet transform and then the RMS value for each particular segment is calculated as shown in Fig.4.

Fig. 4. Demonstration for CWT coefficients features vector calculation.

Fig. 5 show a results of the ECG features extraction for NORM-ECG

Fig. 5. NORM-ECG Features.

An expectations for the continuous wavelet transform and the RMS spectrum values on each P, QRS and T segment is extracted, and the amplitude of each feature vector are scaling into a unit. The comparative result of experiments on the features extraction of this proposed method (CWT) and AC/DCT method are considered.

(a) Single Beat ECG.

(b) AC/DCT Feature.

(c) CWT Feature.

Fig. 6. Example of NORM-ECG Features for 2 subjects.

The results show in Fig. 6 (a), (b), and (c) are depicted a single beast NORM-ECG signal and extracted features for AC/DCT and CWT of 2 individual subjects (red and blue), respectively. Finally, all-classes scattered-plots in the FLDA feature vector space are calculated as shown in Fig. 7 (a) and (b). It is clear that all 10 different classes of each subject are well clustering by using this proposed method.

978-1-4799-1467-8/13/$31.00 ©2013 IEEE

(a) AC/DCT scattered-plots in the FLDA feature vector space.

(b) CWT scattered-plots in the FLDA feature vector space.

Fig. 7. All-classes scattered-plots in the FLDA feature vector space.

The classification results are depicted in Table II. This proposed method (CWT) is given better recognition rate than the AC/DCT algorithms which are evaluated based on the similar database and classification tools.

IV. RECOGNITION RESULT The accuracy evaluation is focused only on the NORM-

ECG trained data set. The recognition results and percent of the accuracy are shown in Table II. The overall accuracy is around 97%. There are only 3 data sets obtained the accuracy around 90-92% at recognition rate at 45/50, 46/50 and 48/50 of the unknown heartbeat. The rest of data sets are achieved the recognition rate 50/50 or 100% of the accuracy.

The ID number N001, N005, N009, and N010 are slightly decreased around 10% of recognition rate. By evaluation, an ocurring recognition error may be effect from ECG data capturing procedure, including many preprocessing algorithms.

TABLE II. analysis results from the questionnaire

V. CONCLUSION This paper is interested in study the robustness of ECG

biometric identification in normal ECG data. All subjects assume to be no heart diseases. The most significant condition for recognition accuracy control in this paper is a Heart Rate Variability (HRV), which is caused from physical activities. HRV effect will be left for a future work.

ACKNOWLEDGMENT The authors appreciate valuable comments from Dr.

Yuttana Kitjaidure and the Bio-SIS Laboratory at Electronics Department, KMITL for supporting the BioPAC Measurement System.

REFERENCES [1] L. Biel, O. Pettersson, L. Philipson and P. Wide, ECG analysis: a new

approach in human dentification, IEEE Transactions on Instrumentation and Measurement, vol. 50, no. 3, pp. 808-812, 2001.

[2] S. A. Israel, J. M. Irvine, A. Cheng, M.D, Wiederhold and B.K. Wiederhold, ECG to identify individuals, Pattern Recognition, vol. 38, pp. 133-142, 2005.

[3] R. Matta, J.K. H. Lau, F. Agrafioti, D. Hatzinakos, Real-time Continuous Identification System using ECG Signals , in IEEE 24th Canadian Conf. on Electrical and Computer Engineering (CCECE2011), 8-11 May, Niagara Falls, Canada.

[4] Z. S. Fatemian, F. Agrafioti, D. Hatzinakos, HeartID: Cardiac Biometric Recognition , IEEE Fourth International Conference on Biometrics: Theory, Applications and Systems (BTAS2010), Sept. 27-29, 2010, Washington DC, USA.

[5] D. Hatzinakos, F. Agrafioti, Signal Validation for Cardiac Biometrics, IEEE 35th International Conference on Acoustics, Speech, and Signal Processing (ICASSP 2010), March 14-19, 2010, Dallas, Texas, USA.

[6] F. Agrafioti, D. Hatzinakos, ECG Biometric Analysis in Cardiac Irregularity Conditions , Signal, Image and Video Processing, Springer, pp 1863-1703, September 2008.

[7] S. Pathoumvanh, K. Hamamoto, and S. Airphaiboon, A NOVEL BEAT LENGTH SELECTION FOR ECG FEATURES EXTRACTION AND CLASSICATION, BMEiCON 2010, August 27-28, 2010, Kyoto, Japan.

[8] S.Y. Khung, M. W. Mak, S. H. Lin, Biometric Authentication A machine learning Approach, Prentice Hall, 2005.

[9] Bio-Pac System, world class data acquisition systems and data loggers for scientific, life sciences research, data analysis and research purposes. http://www.biopac.com/