Digital Signal Processing and MATLAB application

© All Rights Reserved

0 views

Digital Signal Processing and MATLAB application

© All Rights Reserved

- a-new-ecg-sign
- Ecg Cheat Sheet 35
- unit plan
- P80 Six
- SalilPrabhakar_FpMatchingFilterbank_PhD01
- Comparative Analysis of PCA and 2DPCA in Face Recognition
- Engineering Mechanics Meriam Kraige 6th Edition
- rr311105-bio-medical-equipment
- MYOB Biometric Reader Fact Sheet
- A Survey on Human Face Recognition Invariant to Illumination
- ECG_FINAL
- 219
- eclipse-phase-cheat-sheet.pdf
- Bio Potentials
- SIRILANKA1.ppt
- Biometric Verification System
- Framework for biometric match score fusion using statistical and belief models
- Basic ECG Lecture - New
- 12-Lead-Ischemia-Injury-and-Infarct-1 (1).pdf
- Atria 3000 Service Manual.pdf

You are on page 1of 4

University of Toronto,

10 King’s College Road, Toronto, ON, Canada, M5S 3G4

{shari,foteini,dimitris}@comm.utoronto.ca

ABSTRACT 1.5

A) ECG heart-beats

R R'

Voltage (mV)

feature extraction and matching are performed in Euclidean 0.5

T

P'

T'

P

spaces. However, there are many scenarios (e.g., biometric 0

Q S

Q'

S'

Time (samples)

200 250 300

B) Autocorrelation segment

Normalized Power

0.8

P, QRS duration

0 50 100 150

Time (samples)

200 250 300

closed-set matching and 14.82% for open set matching.

Fig. 1. (a) Basic components of ECG heart-beats

Index Terms— Autocorelation, electrocardiogram, SPEC- (b)Autocorrelation (zoomed-in) of an ECG signal

Hashing

The next generation of biometric systems and modalities

1. INTRODUCTION are being developed in light of the threats listed above. Medi-

cal biometrics, a new but promising category of biometric fea-

1.1. Medical Biometrics tures, is one such example. Medical biometrics utilize physi-

Biometric recognition systems are rapidly replacing tradi- ological signals of the human body (vital signals) that contain

tional identification systems based on PIN numbers, tokens subject-specific characteristics. Since these vital signals are

or passwords. Fingerprint, face and iris biometric systems are internal to the human body, medical biometrics offer an in-

increasingly being used to secure passports and grant access herent robustness to circumvention, replay and obfuscation

to high security environments. attacks. Furthermore, biometric liveness is inherently guar-

However, as the technology for biometric recognition ad- anteed.

vances and its use becomes more widespread, so too do the

methods and technology for biometric falsification. Commer- 1.2. ECG Biometrics

cial fingerprint sensors are defeated with off-the-shelf compo- The focus of this paper is on ECG-based biometric recog-

nents that can be used to create identical copies of a person’s nition. The ECG is a vital signal of cardiovascular origin

fingerprint features. Replay attacks are already a credible which reflects the cardiac electrical potential of the heart. It is

threat to voice and fingerprint biometrics. Biometric obfusca- formed during the depolarization and repolarization of spe-

tion, i.e. removal of biometric features to avoid establishment cific parts of the myocardium and can be measured at the

of one’s true identity, is another prominent challenge (for ex- surface of the body using electrodes which can be placed in

ample, asylum-seekers in Europe intentionally damaged their various configurations. From an engineering point of view,

fingerprints). With the wide deployment of biometrics, these the ECG has a non-periodic but highly repetitive pattern. Ev-

attacks are becoming more frequent and concerns are being ery quasi-period of the ECG corresponds to a pulse (or heart-

raised about the kind of security that biometric technologies beat) and it signifies a full cycle of the cardiac function. There

can offer. are particular points of interest on a heart-beat called fiducial

This work was supported by the Natural Sciences and Engineering Re- points, such as the ones shown in Figure 1. These are primar-

search Council of Canada (NSERC). ily related to the main ECG waves, namely the P wave, the

QRS complex and the T wave. the Euclidean space: y = WφT . Details pertaining to the

There are two main approaches for feature extraction from AC/LDA can be found in [6] and [8].

ECG signals, namely fiducial points-dependent or indepen- 2.2. Binarization

dent. Fiducial-dependent approaches extract features based In [7], a quantization approach was suggested (using a Max-

on the local characteristics of heart-beats [1, 2]. For exam- Lloyd quantizer), together with a Gray-code mapping. While

ple, temporal and amplitude distances between consecutive Gray-mapping is effective when the number of bits is small

fiducial points on the ECG were proposed in [1, 3]. On the (< 5bits), this approach does not scale well. In addition, bio-

other hand, non-fiducial approaches explore the ECG wave- metric feature vectors have unique properties, such as small

form as a whole thereby eliminating the need for heart-beat intra-class variability and high inter-class variability, which

segmentation and fiducial-point detection [4, 5, 6]. In the are not taken into account in ML quantization.

above works, the resulting feature vectors lie in Euclidean This motivates the need for class-specific or similarity-

spaces. To the best of our knowledge, there hasn’t been sig- preserving binarization techniques, which is the focus of this

nificant prior work that addresses the problem of binarization work.

(i.e. transformation of feature vectors from Euclidean into There has been significant amount of work on discovering

binary Hamming spaces) for ECG signals. In [7], a quantiza- good projections from arbitrary feature spaces (metric spaces)

tion approach was suggested (using a Max-Lloyd quantizer), into binary vector-spaces. This interest has been motivated by

together with a Gray-code mapping. However, quantization- the need for fast nearest neighbor search in face-recognition

based approaches do not necessarily preserve distance prop- and image-retrieval applications. Some methods include the

erties, i.e two feature vectors that are separated by large Eu- traditional KD-tree, and the more recent Locality Sensitive

clidean distances are not guaranteed to be mapped to two Hashing, Parameter Sensitive Hashing, RBM-based learning,

binary vectors that have a large Hamming-distance separa- Spectral Hashing, SPEC-Hashing, random quantization using

tion.This observation has motivated the subsequent analysis Fourier features etc. Some qualitative discussion and relevant

for ECG biometric processing in the Hamming space. references can be found in [9].

2. OVERVIEW OF THE PROPOSED SYSTEM In this paper, the possibility of adapting the SPEC-

The proposed method for ECG analysis in the hamming space Hashing algorithm to the needs of biometric systems is evalu-

is based on the Autocorrelation/ Linear Discriminant Analy- ated. The purpose is to map (component wise) feature vectors

sis (AC/LDA) [6] in conjunction with SPEC-Hashing. We from the space containing the image of the LDA-based trans-

use the following notation : we denote vectors using bold form (i.e., Im(W)) into binary vectors in such a way that

lower-case x. xm denotes the m-th component of the vec- two feature vectors which have a small Euclidean distance

tor x, whereas x(i) denotes the i-th vector in a set. We use between them are mapped to two binary vectors that are sep-

bold upper-case X for matrices. f (·) denotes that the output arated by small Hamming distance. With this manner, the

of the function is a vector. intra-class and inter-class variabilities will be carried to the

new space. SPEC-Hashing is a similarity preserving algo-

2.1. Feature extraction rithm for entropy-based coding that was developed by Lin

The AC/LDA is a two step procedure in which the autocor- et al.[9] for fast nearest neighbor search in high-dimensional

relations of 5-sec ECG windows are first estimated, and then feature spaces, primarily in the context of image retrieval and

subjected to dimensionality reduction using the LDA. celebrity face recognition applications. The nearest neigh-

Figure 1 shows an example of AC for an ECG reading. bors are defined according to the semantic similarity between

The autocorrelation is computed as: objects in the feature space.

N −|m|−1

In this work, the SPEC-Hashing approach is evaluated for

X closed-set and open-set biometric matching. In the first case,

bee [m] =

R e[i]e[i + m] (1)

a fraction of each user’s enrollment data is used for training

i=0

of the SPEC-Hashing algorithm. For the open-set scenario, a

where e[i], i = 0, 1...(N − 1), are samples of the ECG signal, generic database of non-enrollees is input for SPEC-Hashing

bee

m is the time lag, and N is the length of the signal. Out of R training while the output is applied to a new set of users i.e.,

only a segment φ[m], m = 0, 1...M , starting from the zero the enrollees of the system under evaluation.

lag instance and extending to approximately the length of P The input to the SPEC-Hashing block is the set of LDA-

and QRS waves (as shown in Figure 1) is input to the LDA feature vectors {x(i) } ⊂ RM corresponding to the training

training block. These waves are the least affected by heart data, along with a similarity matrix S. Sij , the (i, j)-th en-

rate changes, and consequently, utilizing only this segment try of the similarity matrix, denotes the semantic similarity

for discriminant analysis makes the system robust to heart rate between the i-th and j-th feature vectors in the training set.

variability. The SPEC-Hashing algorithm is run separately for each

The output of the LDA-training is a transformation ma- of the M components of the LDA- feature vectors. The out-

trix W. The AC feature vectors are projected with W into put is a list of M threshold-vectors {λ1 , λ2 , · · · , λM }. Each

3010

of these threshold-vectors corresponds to a component of the Algorithm 1 SPEC-Hashing

LDA- feature vectors, and they can have varying lengths (their Inputs : LDA-feature vectors (for training set), Similarity

length is optimally determined by the algorithm). matrix

Once the list of thresholds is obtained from the SPEC- Outputs : an ordered list of thresholds

Hashing algorithm during the training phase, the binarization

of the ECG feature vectors is done as follows : 1: Set L = ∅

Let us first look at the m-th component. Suppose λm = 2: for each component m = 1, 2, · · · , M do

[λm,1 , λm,2 , · · · , λm,n ]. Then, the m-th component xm of 3: Set flag = 0.

a feature vector x is mapped to an n-bit vector bm (xm ) = 4: while flag = 0 do

[bm,1 (xm )bm,2 (xm ) · · · bm,n (xm )] where 5: Find λopt that minimizes ∆(λ).

6: if ∆(λ) < 0 then

0 xm ≤ λm,i Update L = [L, λ]

bm,i (xm ) = (2) 7:

1 xm > λm,i 8: else

Once we have the binary vectors for each component xm , 9: Set flag = 1.

we simply form the binary representation of x by concatena- 10: end if

tion, i.e., b(x) = [b1 (x1 )b2 (x2 ) · · · bM (xM )]. 11: end while

In the enrollment phase, for each feature vector x(i) we 12: end for

compute the corresponding binary representation b x(i) .

Here on, we will denote this as b(i) for simplicity, and refer

collection sessions took place, scheduled a few weeks apart in

to these as the binary feature vectors.

order to evaluate the stability of the signal with time. Every

During the authentication phase, the feature vector y ex-

recording is 3 minutes long and the lead orientation matches

tracted from the user’s ECG reading is mapped to the binary

that of Lead II of the standard 12-lead ECG system. 16 out

vector b = b(y) using the same binary decision stumps b(·).

of 52 volunteers were recorded in both sessions and 36 in just

We now comment briefly on the operation of the SPEC-

one session. The sampling frequency is 200Hz.

Hashing algorithm (a detailed treatment can be found in [9]).

Open-set recognition setup. To simulate an open-set bio-

Suppose we have a list of thresholds L, and the corresponding

metric system the dataset was split between a training and an

binary mappings obtained using Eq.(3). Let dH (i, j) denote

evaluation set. The training set included the 36 subjects for

the Hamming distance between the binary vectors b(i) and

who just one ECG reading is available.The proposed algo-

b(j) (corresponding to the feature vectors x(i) and x(j) ). We

rithm was trained on signals from this set. The evaluation set

can define the matrix T (corresponding to list L )as follows

included signals from the 16 volunteers for which two record-

1 −dH (i,j) ings are available. The earlier recording was used for enroll-

Tij = e (3) ment and the later for testing.

Z

P Closed-set recognition setup. To simulate the closed-set

where Z = i,j e−dH (i,j) . Also, let us denote the normal- biometric system the proposed algorithm was trained on ECG

ized similarity matrix by S̃. We can then view the operation signals from all users i.e., the enrollees of the system. The

of the SPEC-Hashing algorithm as finding an optimum list enrollment data includes the first ECG readings from the 16

of thresholds that minimizes the Kullback-Liebler divergence volunteers and the first half of the ECG reading from the 36

between the distribution T and the target distribution S̃. The volunteers.

closeness of the two distributions represents how closely the The resulting binary vectors, obtained using the list of bi-

similarities are preserved by the mapping from the Euclidean nary decision-stumps learnt by the SPEC-Hashing algorithm

space to the Hamming feature space. were 345 bits long for the open-set setup and 468 bits long for

The SPEC-Hashing algorithm builds the list L in an in- the closed-set setup (we set the maximum number of bits per

cremental fashion. Let L∗ = [L, λ]. Define dimension to be 25 bits). The difference of the bit length be-

tween the two setups is due to the dimensionality after LDA

∆(λ) = KL(S̃||T∗ ) − KL(S̃||T) (4) projection.

Figure 2 shows the histograms for the intra-subject and

If ∆(λ) is negative, then adding λ to the list reduces the di- inter-subject Euclidean and Hamming distances for the

vergence from the target distribution, and the list is updated closed-set recognition setup. For the Euclidean space, the

to include λ. average inter-subject distance is 0.2656 and the average intra-

3. EXPERIMENTAL RESULTS

subject distance is 0.0757. For the Hamming space, the

The proposed system was evaluated over ECG signals from average inter-subject distance is 0.2496 (116 bits) and the

52 different volunteers that were collected at the Biometrics intra-subject 0.0719 (33 bits).

Security Laboratory of the University of Toronto. Two signal Figure 3 shows the performance of three open-set sys-

3011

A) AC/LDA feature space B) SPEC-Hash feature space

50 40

Percentage of Pairs

Percentage of Pairs

40

30

30

20

20

10

10

0 0

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7

Euclidean Distance Hamming Distance

System Setup EER 4. CONCLUSION

AC/LDA Euclidean Classifer Open-set 18.6% As ECG biometric systems begin to scale, it is important to

Hamming Classifier (SPEC-Hashing) Open-set 14.82% study binarization of the ECG feature vectors and design of

AC/LDA Euclidean Classifer Closed-set 11.4% Hamming-distance classifiers, in order to enable important

Hamming Classifier (SPEC-Hashing) Closed-set 5.5% applications such as template protection using biometric en-

cryption, efficient low-complexity matching for identification

mode of operation etc. A method for binarization of the

Table 1. System equal-error-rate (EER) ECG feature vectors based on SPEC-Hashing was proposed

Open-set Biometric System Performance and evaluated over ECG signals from 52 subjects. With the

proposed treatment there is virtually no loss in performance

60

AC/LDA Euclidean Classifier

SPEC-Hashing

compared to nearest neighbor classification in the Euclidean

ML Quantization LDA-projection space.

50

False Acceptance Rate (%)

EER 5. REFERENCES

40

[1] S. A. Israel, J. M. Irvine, A. Cheng, M. D. Wiederhold, and B. K.

Wiederhold, “ECG to identify individuals,” Pattern Recognition, vol.

30 38, no. 1, pp. 133–142, 2005.

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

approach in human identification,” IEEE Trans. on Instrumentation and

20

Measurement, vol. 50, no. 3, pp. 808–812, 2001.

[3] S.I. Safie, J.J. Soraghan, and L. Petropoulakis, “Electrocardiogram ECG

10 biometric authentication using pulse active ratio PAR,” Information

10 15 20 25 30 35 40 45 50

Forensics and Security, IEEE Transactions on, vol. 6, no. 4, pp. 1315–

False Rejection Rate (%) 1322, dec. 2011.

[4] G. G. Molina, F. Bruekers, C. Presura, M. Damstra, and M. van der

Fig. 3. ROC: FAR versus FRR for the Euclidean and Ham- Veen, “Morphological synthesis of ECG signals for person authentica-

tion,” in Proceedings of 15th European Signal Proc. Conf., Poland, Sept.

ming (SPEC-Hash and ML) spaces 2-7 2007.

[5] I. Odinaka, Po-Hsiang Lai, A.D. Kaplan, J.A. O’Sullivan, E.J. Sirevaag,

S.D. Kristjansson, A.K. Sheffield, and J.W. Rohrbaugh, “ECG biomet-

tems: the AC/LDA Euclidean-distance classifier (which is rics: A robust short-time frequency analysis,” in Proceedings of IEEE In-

the baseline system), the Max-Lloyd quantization approach ternational Workshop on Information Forensics and Security, Dec. 2010,

pp. 1–6.

(for the sake of completeness), and the proposed Hamming-

[6] F. Agrafioti and D. Hatzinakos, “Fusion of ECG sources for human

distance classifier with binary feature vectors designed using identification,” in Proceedings of 3rd Int. Symp. on Communications

the SPEC-Hashing algorithm. The Max-Lloyd quantization Control and Signal Processing, Malta, March 2008, pp. 1542–1547.

approach does not perform well, because binarization does [7] F. Agrafioti, F. M. Bui, and D. Hatzinakos, “Medical biometrics in mo-

not take into consideration class information. The perfor- bile health monitoring,” Security and Communication Networks, vol. 4,

no. 5, pp. 525–539, July 2010, .

mance of the proposed Hamming-distance classifier using bi-

[8] Foteini Agrafioti and Dimitrios Hatzinakos, “ECG biometric analysis

narization based on SPEC-Hashing outperforms all cases. As

in cardiac irregularity conditions,” Signal, Image and Video Processing,

stated in Table 1, the Equal Error Rate (EER) for the AC/LDA pp. 1863–1703, 2008.

based Euclidean-distance classifier was found to be 18.6% [9] Ruei-Sung Lin, D.A. Ross, and J. Yagnik, “Spec hashing: Similarity

(Open-set), and the EER of the proposed Hamming-distance preserving algorithm for entropy-based coding,” in Computer Vision

classifier is 14.82% (Open-set). The EER in the closed set and Pattern Recognition (CVPR), 2010 IEEE Conference on, june 2010,

pp. 848–854.

setup of the proposed system is 5.5%.

3012

- a-new-ecg-signUploaded byDrscottmccall
- Ecg Cheat Sheet 35Uploaded byjessjaylee
- unit planUploaded byapi-336004351
- P80 SixUploaded byPablo Ceretani
- SalilPrabhakar_FpMatchingFilterbank_PhD01Uploaded byVincent Geoca
- Comparative Analysis of PCA and 2DPCA in Face RecognitionUploaded byDinh Khai Lai
- Engineering Mechanics Meriam Kraige 6th EditionUploaded byahmad atta
- rr311105-bio-medical-equipmentUploaded bySRINIVASA RAO GANTA
- MYOB Biometric Reader Fact SheetUploaded byrjmv busisolutions
- A Survey on Human Face Recognition Invariant to IlluminationUploaded byIAEME Publication
- ECG_FINALUploaded byJoseph Mathis
- 219Uploaded byArunKumar Mani
- eclipse-phase-cheat-sheet.pdfUploaded byLisa Solis
- Bio PotentialsUploaded bysonu24me
- SIRILANKA1.pptUploaded byk000
- Biometric Verification SystemUploaded byBob
- Framework for biometric match score fusion using statistical and belief modelsUploaded byMia Amalia
- Basic ECG Lecture - NewUploaded byAradhanaRamchandani
- 12-Lead-Ischemia-Injury-and-Infarct-1 (1).pdfUploaded byErica Reimer
- Atria 3000 Service Manual.pdfUploaded byPamela Zavala
- ch 02Uploaded byAgus Kusnayat
- 1Uploaded byAman Aryan
- Texture TMIUploaded byAnonymous 6WeX00mx0
- lung cancer detection using computer vision by mutiullah et alUploaded byMutiullahJamil
- Pelayanan Medik Dan Keperawatan_9Uploaded byArih Ginting
- MEC291 Lab Manual Sheet Dynamics DecApr2016Uploaded byHaFiy HaZim
- ECG BasicsUploaded bysuzane christopher
- ECG Basics PptUploaded bysuzane christopher
- Design_and_implementation_of_student_att.docUploaded byFrederick Lawrence
- EKGUploaded byJay Jiang

- Network ScurityUploaded bySarvjeet Singh Sohal
- A Chaos-based Visual Encryption Mechanism InUploaded bySarvjeet Singh Sohal
- Data Encryption Using ECG Signals (4)Uploaded bySarvjeet Singh Sohal
- Calculation of Lyapunov Exponent by time series analysisUploaded bySarvjeet Singh Sohal
- Linear Integrated Circuits Course Outline.pdfUploaded bySarvjeet Singh Sohal
- Project ReportBIIRECTIONAL VISITOR COUNTERUploaded bySarvjeet Singh Sohal
- THE HEARTUploaded bySarvjeet Singh Sohal
- TOC JAVAUploaded bySarvjeet Singh Sohal
- Inter networking and Internet ProtocolsUploaded bySarvjeet Singh Sohal
- Assign. (Autosaved).docxUploaded byVikas Singh Kohaal
- Strategic Management and Business PolicyUploaded bySarvjeet Singh Sohal
- Bridge Chapter 01Uploaded bySiti Hajar
- MAT LAB By Sarvjeet Singh Sohal Free- Lance Computer TechnicianUploaded bySarvjeet Singh Sohal
- The Compact Hydraulic Excavator Can Be a Tracked or Wheeled Vehicle With an Approximate Operating Weight of 13Uploaded bySarvjeet Singh Sohal
- 1169-2156-1-PBUploaded bySarvjeet Singh Sohal
- Conclusion.docxUploaded bySarvjeet Singh Sohal
- 2218.docxUploaded bySarvjeet Singh Sohal
- PLASMA ANTENEUploaded bySarvjeet Singh Sohal

- CPUploaded byStephanie Joy Escala
- Invited Review Dynamic Exercise Performance in Masters AthletesUploaded bySoulyoga
- Chordates.pdfUploaded byNil Banerjee
- Enzymatic Therapy CanadaUploaded bycureself.com
- 11 GeobiologyUploaded byLuis Castillo
- ARTRITIS REUMATOIDEUploaded byiris alicia salas magos
- Assessment of Intestinal MalabsorptionUploaded bysdasdsd
- BAB 7Uploaded byDenny Ramdhan
- Neonatal Neurology 4th ed - G. Fenichel (Churchill Livingstone, 2007) WW.pdfUploaded byCampan Alina
- Mekanisme Transpor Dian[1]Uploaded bylaksmiyuni
- F&E Memorization ChartUploaded byKrCMommy522
- acute limb ischemiaUploaded byMohammad Husni Banisalman
- Fundamentals of IV TherapyUploaded byMilo Almirol
- careplanUploaded byapi-253285688
- IB Multiple Choice QuestionsUploaded byJohn Osborne
- Vit B1 and Cancer Cell MetabolismUploaded byLateecka R Kulkarni
- MolecularUploaded byErik Joyce
- PREMATURE EJACULATION.pdfUploaded byxristosxristos
- Stroke Volume 49 Issue 2 2018 [Doi 10.1161%2FSTROKEAHA.117.020453] -- Stroke - Highlights of Selected Articles(1)Uploaded byUmam Fauzi
- NHS - Buidling Self-Esteem.pdfUploaded byVer Chew
- BelonocystisUploaded byWillbour
- ALCIAN Blue Staining IMPUploaded byPallaval Veerabramhachari
- Drug StudyUploaded byMax Loco
- 495.fullHigh-Rising Epiglottis in Children: Should It CauseConcernUploaded byAndri wijaya
- The Cosmetic Dye Quinoline Yellow Causes DNA Dmage in VITROUploaded byCarlos Alberto Bayona López
- 12 - Drug Design - Optimising Access to the TargetUploaded byDuy Tran
- Oral Cavity Considerations for the Patient with Renal DiseaseUploaded byEditor IJTSRD
- fluid volume excessUploaded byTamil Villardo
- SAPS Sheet 3 - Teachers Notes - Leaf DiscsUploaded bymizo1994
- Adrenal Glands From Pathophysiology to Clinical Evidence PDFUploaded byDr-Sumanta Banerjee

## Much more than documents.

Discover everything Scribd has to offer, including books and audiobooks from major publishers.

Cancel anytime.