ECG-BASED AUTHENTICATION

Bayesian vs. Nearest Neighbour Classifiers

Carla Oliveira and Ana Fred

Instituto de Telecomunicações, Instituto Superior Técnico, Tech. University of Lisbon, Lisbon, Portugal

Keywords: Bayesian, Biometric authentication, ECG, MAP, One-Class, 1-NN.

Abstract: This paper presents an approach for human authentication based on electrocardiogram (ECG) waveforms.

ECG data was collected from 24 individuals during the realization of cognitive tests, where subjects held a

surface mount triode placed on the V2 pre cordial derivation. Authentication is based on MAP, One-Class

and 1-NN classifiers. Results show that ECG-based authentication may be a feasible tool for biometric

systems. The One-Class classifier with class normalization has presented enhanced performance, with an

equal error rate of 3.5%.

1 INTRODUCTION

Biometric authentication is a promising tool for

security applications, attesting that the user of a

system is who he claims to be through the use of

some of its physical or behaviour characteristics

(e.g., a fingerprint). Recent work, (Biel et al., 2001)

and (Israel et al., 2005), suggests that the human

heartbeat is a characteristic that can be used in

biometric authentication schemes, as it exhibits

features that are unique to an individual.

Electrocardiogram (ECG) is the typical method to

measure heartbeat, being extensively used in

medicine. Figure 1 illustrates a typical ECG trace.

Figure 1: A typical heartbeat waveform (adapted from

(Wikipedia, 2008)). The R R interval indicates the

duration of a heartbeat. P, QRS, and T indicate the major

ECG complexes comprising one heartbeat.

Some feasibility studies on the potential of ECG

for biometrical applications are found in the

literature. For example, in the identification scheme

presented in (Wübbeler et al., 2007), authors use 234

ECG recordings of 10 s length, obtained during

several months up to several years. Records were

taken from 74 subjects in a supine position in a

resting state, from the three Einthoven leads.

Classification is based on the heart vector and a

simple distance measure, standard nearest

neighbour, and threshold schemes being used. For

verification, an error rate of 2.8% was achieved;

while a rate of 98.1% was obtained for

identification. Other study is presented in (Chan et

al., 2008), where ECG data was collected from 50

subjects during 3 sessions on different days, from

two electrodes on the pads of their thumbs using

their thumb and index fingers. Classification was

performed using percent residual difference,

correlation coefficient, and a novel distance measure

based on wavelet transform. The wavelet distance

measure has a classification accuracy of 89%,

outperforming the other methods by nearly 10%.

In this work we have addressed the problem of

user authentication from ECG records using a single

lead montage, while the subjects were performing

cognitive tests on a computer. Classification is based

on two Bayesian classifiers, the maximum a

posteriori (MAP) (Duda et al., 2001) and the

One-Class (Tax, 2001) classifiers, and also on a

distance based method, the 1-Nearest Neighbour

(1-NN) (Duda et al., 2001) classifier. The MAP

163

Oliveira C. and Fred A. (2009).

ECG-BASED AUTHENTICATION - Bayesian vs. Nearest Neighbour Classiﬁers .

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing, pages 163-168

DOI: 10.5220/0001549901630168

 SciTePress

classifier assigns an object x to the class k with the

largest a posteriori probability p(ω

|x). In One-Class

classification only p(x|ω

), the probability density of

the target class, ω

, is known. Estimating the

probability density from the training data and given

a threshold, the classifier accepts or rejects the test

samples. The 1-NN classifier assigns an object x to

its nearest class, with closeness measured by the

Euclidean distance between the vectors of inputs.

This paper is composed of 4 sections, besides the

current one. The next section presents the data

acquisition system from which ECG records were

obtained. Section 3 describes the authentication

system, detailing the implementation of the

classifiers. An overview and discussion of results is

provided in Section 4. Section 5 finalises the paper,

drawing the main conclusions.

2 DATA ACQUISITION AND

PROCESSING

The ECG data analysed in this work was acquired

within the scope of the HiMotion Project (HiMotion,

2008). The HiMotion Project consisted on the

design, implementation and administration of a set

of computer based experiments with cognitive tests

related to memory, concentration, association,

intelligence and insight (discovery). The underlining

idea is that these activities produce noticeable

changes in the physiological characteristics of

subjects, which, on one hand, are task dependent,

and therefore global task-related dynamics/features

can be recognized, and, on the other hand, individual

behavioural traits may be present in the acquired

data, and thus contribute for human authentication.

A set of physiologic signals was continuously

acquired during the realization of the tests:

electrodermal activity, blood volume pressure,

electroencephalography and ECG. A population of

24 male and female volunteers, with a mean age of

23.4±2.5 years, was asked to complete the series of

tests in individual sessions, designed to take, in

average, 30 minutes.

ECG measurements were taken using a surface

mount triode placed on the V2 pre-cordial

derivation. Each heartbeat waveform was

sequentially segmented from the full recording, and

then all individual waveforms were aligned by their

R peaks in segments of equal temporal length. The

mean wave for groups of 10 heartbeat waveforms

(without overlapping), was computed to minimize

the effect of outliers. A labelled database composed

by 137 samples was compiled, in which each pattern

corresponds to a mean wave. For each mean

waveform (Figure 1), the latency and amplitude for

each of the P QRS T peaks were extracted, along

with a sub sampling of the waveform itself,

providing a feature representation space of 53

features. In this work, only the latencies and

amplitudes of P, Q, S and T complexes were used,

resulting in 8 features, Table 1.

Table 1: Description of features.

Feature Description

1 Latency of P complex

2 Latency of Q complex

3 Latency of S complex

4 Latency of T complex

5 Amplitude of P complex

6 Amplitude of Q complex

7 Amplitude of S complex

8 Amplitude of T complex

Concluding, the available ECG data comprises

24 classes (each corresponding to each one of the

subjects under test) and 8 features.

3 AUTHENTICATION SYSTEM

The purpose of ECG-based authentication systems is

to attest that the user of a system is who he claims to

be, through the monitoring of its ECG records. In

this work, three classifiers were implemented using

Matlab (Matlab, 2007): MAP classifier, One-Class

classifier and 1-NN classifier.

The MAP classifier algorithm was constructed as

follows. Two mutually exclusive sub-sets from the

137 sample set are created, with 1 pattern for test

and the remaining 136 for training (“leave-one-out”

method). Then, density of the training data, p(x|ω

is estimated according to a maximum likelihood

technique later explained. p(ω

|x) is subsequently

computed for each test sample according to (1) and

the classifier decides on accepting test samples if (2)

is verified. This process is repeated for all the 137

samples. It is important to state that a Naive Bayes

model is considered for used features, thus assuming

statistical independence between them, (3). Also,

classes are assumed to be equiprobable, (4).

(| )( ) (| )( )

(|)

()

(| )( )

kk kk

px p px p

px p

ωωω

∑

(1)

(|)

(2)

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

164

(| ) ( | ) ( , )

kkkk

px px N

ωμσ

∏∏

(3)

() ( )

ωω

(4)

For the One-Class classifier, a similar algorithm

was adopted. The algorithm starts to estimate the

distribution of training data. Then, the probability

density of the target class, p(x|ω

), is estimated and

normalized within a factor F

class

, which is the

maximum value within each class, (5). Afterwards,

given a threshold λ, the classifier accepts the test

samples included in the acceptance region defined

by threshold according to (6). This process is

repeated for all the 137 samples. For comparison

purposes, another version of the One-Class classifier

was implemented, using a different normalization

factor, F

all

, which is the maximum p(x|ω

) value

found over all classes, (7).

max( ( | ))

class k

Fpx

(5)

(| )

(6)

max ( ( | ))

all k k

Fpx

(7)

Note that, regarding these two Bayesian

classifiers, the density model of the training data was

estimated based on its histogram plots, Figure 2. A

Gaussian distribution, with mean µ and variance σ

was assumed for each feature. It is important to state

that this is a simplistic approach (e.g., feature 2 in

Figure 2 is a Dirac function), with implicit

drawbacks on the performance of the authentication

system. A mixture of Gaussians will probably

provide refined results, but has the additional

complexity drawback.

Figure 2: Histogram plots for subject id 10.

The 1-NN classifier algorithm starts to compute

and store the Euclidean distances between all data

samples, and then normalizes the computed values

to F

dis

, which is the maximum distance found

between two samples x

and x

, (8). After the

creation of training and test sets, the minimum

distances for the training set within each class are

found and classification is based on that. Test

samples are accepted if (9) is verified. This process

is repeated 137 times, one per sample.

max( ( , ))

dis a b

Fdxx

(8)

min

(| )

(9)

4 RESULTS AND DISCUSSION

In what concerns the MAP classifier, the confusion

matrix (average values obtained for the 137 runs) is

presented in Figure 3. It is clear from this figure that

test samples from different individuals are extremely

uncorrelated, thus being correctly classified. About

60% of the test samples achieve p(ω

|x) > 0.85. In

Figure 4 and Figure 5 one may observe the average

Receiving Operating Characteristic (ROC) and the

False Acceptance Rate (FAR) / False Rejection Rate

(FRR) curves, respectively, for the 137 runs. It is

observed that FAR and FRR are dependant on the

adjustable chosen threshold. If the threshold value is

increased, FAR decreases, while FRR increases.

When the value of threshold is decreased, the

proportion FRR will decrease, while FAR increases.

FAR lies between 5% and 17%, while FRR achieves

values between 4% and 9%. The equal error rate

(EER) occurs for λ ~ 2.6E-04, corresponding to

FAR=FRR=7%. For this value of the threshold,

FAR and FRR values were analysed within each

class (see Figure 6). It is observed that classes 1, 18

and 22 present the worst results. In order to

determine EER within each class, the respective

FAR and FRR values were computed. Table 2

presents the λ

values corresponding to the EER for

each class k. An average EER of 5.9% was

estimated, being lower to the one obtained for the

MAP classifier with a global threshold. Thus, one

concludes that specific thresholds per class will

enhance the performance of the classifier.

ECG-BASED AUTHENTICATION - Bayesian vs. Nearest Neighbour Classifiers

165

Figure 3: Confusion matrix for the MAP classifier

(average values).

Figure 4: ROC curve for the MAP classifier.

Figure 5: FAR/FRR(λ) curves for the MAP classifier.

Figure 6: FAR/FRR curves for each class, with

λ ~ 2.6E-04 (MAP classifier).

Table 2: ERR within each class (MAP classifier).

Class, k λ

ERR

1 1.51E-04 27%

2 4.66E-04 7%

3 2.00E-09 0%

4 6.40E-03 6%

5 6.45E-06 4%

6 9.90E-01 5%

7 1.91E-06 11%

8 3.00E-06 1%

9 2.00E-03 1%

10 1.56E-04 8%

11 2.06E-08 2%

12 2.10E-05 7%

13 2.80E-10 2%

14 1.60E-01 3%

15 3.86E-04 4%

16 3.30E-03 2%

17 3.48E-06 9%

18 4.70E-03 10%

19 2.00E-08 0%

20 1.56E-04 3%

21 1.71E-04 5%

22 3.11E-04 15%

23 2.66E-04 7%

24 2.03E-03 4%

average 6%

Regarding the One-Class classifier, the confusion

matrix (average values) obtained for the 137 runs, is

presented in Figure 7. Again, it is observed that

samples are extremely uncorrelated and almost all

are correctly classified. 75% of the test samples

achieve p(x|ω

) > 0.85. ROC and FAR/FRR(λ)

curves for the One-Class classifier (with

normalization within each class) are represented in

Figures 8 and 9, respectively. One observes better

results for this classifier when compared to MAP,

with FAR ranging from 3% and 13%, while FRR

lies between 7% and 13%. EER happens for

λ ~ 1.4E-03, corresponding to FAR=FRR=3.5%.

Worse results were obtained for the version of

One-Class classifier with normalization to the

maximum p(x|ω

). In this case, FAR lies between

5% and 15%, while FRR ranges from 8% to 16%.

EER occurs for λ ~ 2.5E-08, corresponding to

FAR=FRR=10%.

Regarding the 1-NN classifier, Figure 10 and

Figure 11 represent the ROC and FAR/FRR(λ)

curves for this classifier. For 1-NN, a symmetric

trend is verified when compared to the Bayesian

classifiers. With higher threshold levels, FAR will

increase, while FRR decreases. This is intuitive, as

increasing the distance threshold will lead the

system to accept more users, thus increasing FAR

and decreasing FRR. Poor performances were

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

166

obtained for 1-NN, with EER occurring for

λ ~ 2.9E-02, corresponding to FAR=FRR=8%.

Figure 7: Confusion matrix for the One-Class classifier

with class normalization (average values).

Figure 8: ROC curve for the One-Class classifier.

Figure 9: FAR/FRR(λ) curves for the One-Class classifier.

Figure 10: ROC curve for the 1-NN classifier.

Figure 11: FAR/FRR(λ) curves for the 1-NN classifier.

An overview of the obtained results is presented

in Table 3, from which one concludes that all the

implemented classifiers are promising for an

ECG-based authentication scheme.

Table 3: Overview of classifiers.

Classifier EER

MAP 7%

One-Class

(class normalization)

3.5%

One-Class

(maximum normalization)

10%

1-NN 8%

5 CONCLUSIONS

This paper exploits possible approaches for

ECG-based authentication schemes, using real data

obtained by the HiMotion Project. ECG records

from 24 individuals were gathered during realization

of cognitive tests, where subjects held a surface

mount triode placed on the V2 pre cordial

derivation. Two Bayesian classifiers, MAP and

One-Class, and a standard 1-NN were implemented

using Matlab. Results show that the three schemes

achieve feasible performances for an authentication

system, with statistical classifiers presenting better

results.

Regarding the Bayesian classifiers, Gaussian

distributions were assumed to estimate p(x|ω

). In

the MAP classifier, decision is based on posterior

probabilities, given a global threshold for the 24

classes. This assumption results in an EER of 7%. It

was shown that enhanced performance could be

obtained, if one considers specific thresholds per

class. The same conclusion is valid for the One

Class classifier, which, in a first approach, considers

class normalization factors, leading to an error rate

of 3.5%. Without normalization to the maximum

ECG-BASED AUTHENTICATION - Bayesian vs. Nearest Neighbour Classifiers

167

p(x|ω

) within each class, performance degrades to

error rates of 10%. Regarding the 1-NN classifier,

which is based on distance measure, a slight worse

performance was achieved with EER of 8%.

It is concluded that the ECG biometric does

provide a simple method for human authentication,

which may be appropriate in some applications (e.g.,

sensor authentication in body area networks).

Moreover, ECG may be a good source of additional

information in a multi-biometrics approach, as well

as integrated in health surveillance systems.

REFERENCES

Biel,L., Pettersson,O., Philipson,L. & Wide,P. (2001).

ECG Analysis: A New Approach in Human

Identification. In IEEE Transactions on

Instrumentation and Measurement, Vol. 50, No. 3,

808-812.

Israel,S., Irvine,J., Cheng,A., Wiederhold,M. &

Wiederhold,B. (2005). ECG to identify individuals. In

Pattern Recognition, Vol. 38, No. 1, 133-142.

Sinus Rhythm. (2008). Retrieved 2008, from

http://en.wikipedia.org/wiki/sinus_rhythm.

Wübbeler,G., Stavridis,M., Kreiseler,D., Bousseljot,R. &

Elster,C. (2007). Verification of Humans Using the

Electrocardiogram. Pattern Recognition Letters,

Vol. 28, No. 10, 1172-1175.

Chan,A., Hamdy,M., Badre,A. & Badee,V. (2008).

Wavelet Distance Measure for Person Identification

Using Electrocardiograms. In IEEE Transactions on

Instrumentation and Measurement, Vol. 57, No. 2,

248-253.

Duda,R., Hart,P. & Stork,D. (2001). In Pattern

Classification. John Willey & Sons Inc.. New York,

2nd Edition.

Tax,D. (2001). One-Class classification: Concept-learning

in the absence of counter-examples. Ph.D. Thesis,

Delft University of Technology, Delft.

HiMotion Project. (2008). Retrieved 2008, from

http://www.it.pt/auto_temp_web_page_preview.asp?id

=305.

MATLAB Release 14 (2007). Retrieved 2008, from

http://www.mathworks.com/products/matlab/.

BIOSIGNALS 2009 - International Conference on Bio-inspired Systems and Signal Processing

168