On Real Time ECG Segmentation Algorithms for Biometric Applications

Filipe Canento

, Andr´e Lourenc¸o

1,2

, Hugo Silva

and Ana Fred

Instituto de Telecomunicac¸˜oes, Instituto Superior T´ecnico, Lisbon, Portugal

Instituto Superior de Engenharia de Lisboa, Lisbon, Portugal

Keywords:

QRS Detection, ECG Segmentation, Biometrics, Identity Recognition, Real-time Analysis.

Abstract:

Recognizing an individual’s identity through the use of characteristics intrinsic to that subject is a biometric

recognition problem with increasingly number of modalities and applications. Recently, the electrical activity

of the heart (the Electrocardiogram or ECG) has been explored as an additional modality to recognize indi-

viduals. The ECG signal contains several features, which are unique to each individual. The preprocessing

of the ECG signal and the feature extraction steps are crucial for biometric recognition to be successful. In

ﬁducial approaches, this last step is accomplished by correctly detecting the heart beats, and performing their

segmentation to extract the biometric templates afterwards. In this work, we present an overview of the differ-

ent steps of an ECG biometric system, focusing on the evaluation and comparison of multiple real-time heart

beat detection and ECG segmentation algorithms, and their application to biometric systems. An evaluation

and comparison of the algorithms with annotated datasets (MITDB, NSTDB) is presented, and methods to

combine them in order to improve performance are discussed.

1 INTRODUCTION

Biometric recognition, or biometrics, is a growing re-

search ﬁeld in which the purpose is to automatically

recognize a subject using his intrinsic characteristics

(Jain et al., 2004; Jain et al., 2007), such as: phys-

iological (e.g., Electrocardiogram), anatomical (e.g.,

ﬁngerprint, iris) or behavioral (e.g., keystroke, signa-

ture). The latest trends in the ﬁeld point towards the

multibiometrics approaches, which use one or more

of these inputs and they work by performing a pat-

tern recognition scheme, and assessing whether the

user being tested is genuine or not, i.e., the system

is able to recognize the user, or imposter (a user that

is not who he/she claims to be or is not registered in

the system’s database). Three concepts stand out in

any biometric application: i) enrollment: the initial

registration process, so that the system has informa-

tion to use later for identity recognition; ii) authenti-

cation: the recognition process in which the subject

ﬁrst provides a claimed identity, the biometric tem-

plates are acquired and compared to the templates

of the claimed user, and a decision is made (gen-

uine/imposter); iii) identiﬁcation: the recognition pro-

cess in which the subject does not provide any prior

information about his/her identity, hence the biomet-

ric system has to compare the acquired templates with

all the templates stored in the database.

Multiple scenarios and applications can beneﬁt

from biometrics, such as granting/denying access to

secure areas or resources (e.g., a computer), remote

identity veriﬁcation, and personalization of assets

(e.g., intelligent houses and cars), among many oth-

ers (Jain et al., 2011).

Our work is focused on Electrocardiogram (ECG)

based biometrics (Biel et al., 2001; Wang et al., 2008;

Shen, 2005; Israel et al., 2005; Lourenc¸o et al., 2011),

a recent trend in this ﬁeld. The ECG of each in-

dividual has been shown to possess unique features

among individuals (Israel et al., 2005); furthermore,

the ECG holds several other biometric-relevant prop-

erties, such as: (a) it is hard to synthesize; (b) it is

continuously available; (c) it may be acquired non-

intrusively; and (d) it has a straightforward liveliness

detection.

In developing a real-time ECG based biometric

system, different challenges emerge, such as: (i)

data acquisition; (ii) signal processing; (iii) informa-

tion extraction; and (iv) biometric pattern recognition.

The ﬁrst step deals with obtaining the raw ECG data

through a reliable, repeatable and non-intrusive pro-

cedure. This raw data may contain noise and other ar-

tifacts, and as such, signal processing techniques are

needed to enhance the signal quality. The next step is

to extract relevant subject-dependent information and

feed it to the biometric system so that recognition can

228

Canento F., Lourenço A., Silva H. and Fred A..

On Real Time ECG Segmentation Algorithms for Biometric Applications.

DOI: 10.5220/0004245902280235

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2013), pages 228-235

ISBN: 978-989-8565-36-5

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

be performed.

There are two types ECG based biometrics meth-

ods: ﬁducial (Biel et al., 2001; Wang et al., 2008;

Shen, 2005; Israel et al., 2005) and non-ﬁducial

(Coutinho et al., 2010; Chan et al., 2008). In the

ﬁducial case, the crucial step in extracting informa-

tion from the ECG is to correctly detect each of the

heart beats. The non-ﬁducial methods extract infor-

mation without the help of reference points.

We focus on the ﬁducial case and therefore, the

heart beat detection and ECG segmentation algorithm

is of utmost importance.

One of the major research lines in ECG biomet-

rics has been focusing on the development of non-

intrusive measurement methods (Silva et al., 2011b),

leading to increased measurement noise. Further-

more, so far, researchers have focused on ofﬂine pro-

cessing, which limits the deployment and applicabil-

ity of currently available ECG-biometrics techniques

in a real-world scenario. In this paper, we present an

overview and comparison of multiple real-time seg-

mentation algorithms, taking into account the inﬂu-

ence of measurement noise on the biometric perfor-

mance. To improve robustness and performance, we

propose and evaluate a set of different voting methods

to combine the output information of the segmenta-

tion algorithms.

The rest of the paper is organized as follows: Sec-

tion 2 provides background information about the

ECG and it’s application to biometric recognition;

Section 3 deals with the ECG based biometrics steps;

Section 4 presents the experimental protocol and dis-

cusses the results obtained; and Section 5 concludes

the paper and gives some future work ideas.

2 BACKGROUND

The heart has a set of specialized cells with self-

excitatory properties, which produce the electrical im-

pulses that trigger the mechanical action of the car-

diac muscle ﬁbers; the Electrocardiogram (ECG) is

the measurement of its electrical activity over time.

Typically, the ECG is acquired with a set of elec-

trodes on the thorax. Placing electrodes on the thorax

is not practical in a biometric point-of-view, and re-

cently other approaches have been presented, such as

the one in (Lourenc¸o et al., 2011; Silva et al., 2011b),

in which a 1-lead setup for ECG signal acquisition at

the ﬁngers using Ag/AgCl electrodes without gel is

proposed.

For a healthy human heart, the ECG waveform for

each heart beat resembles the one depicted in Fig. 1:

it is composed of the P, Q, R, S, and T waves and beat

Figure 1: ECG heart beat example. The corresponding ECG

signal was acquired at the ﬁngers using the setup proposed

in (Silva et al., 2011a).

detection is usually performed by searching for each

R-peak or QRS complex.

Before being able to extract any information, the

ECG signal needs preprocessing due to measurement

noise and other artifacts which may be contaminating

the signal (e.g., motion artifacts). This step is per-

formed by applying a speciﬁc ﬁlter, designed to en-

hance the signal quality, while avoiding distortion and

loss of relevant information.

3 ECG BASED BIOMETRICS

The architecture of an ECG biometric system follows

the block diagram of Fig. 2: the raw ECG is acquired,

preprocessed and convertedto a digital format; digital

signal ﬁltering techniques are applied and QRS detec-

tion is performed. With the QRS complexes detected,

the ECG signal is divided in segments corresponding

to individual heart beats. Different features may be

extracted from the heart beat waveform, such as: QRS

complex duration, P, Q, R, S, and T waves amplitudes

and onsets, (Biel et al., 2001; Chung, 2000). Features

are then used as input to the biometric system, where

pattern recognition is performed. In our case, the pat-

terns correspond to ECG signal windows of 600ms

around the QRS complexes: 200ms to the left and

400ms to the right of the R peak. These values were

selected according to the typical physiological dura-

tion of the P-Q and S-T complexes.

3.1 Preprocessing

To enhance the ECG signal quality and increase the

signal-to-noise ratio, we designed a 300 order band

pass Finite Impulse Response (FIR) ﬁlter with a Ham-

ming window, and cutoff frequencies of 5Hz to 20Hz.

These speciﬁcations, take into account the ECG infor-

OnRealTimeECGSegmentationAlgorithmsforBiometricApplications

229

Figure 2: Block diagram of an ECG based biometric sys-

tem.

mation bandwidth, and empirical considerations ex-

tracted from the heuristic analysis of the signal. Con-

cerning real time applications, the signal is acquired

and ﬁltered in frames, and thus a method to process

each frame and merge them into a meaningful sig-

nal is required. One way to solve this problem is to

use the overlap-add method (Oppenheim and Schafer,

1975). Fig. 3(a) shows an example of a raw ECG sig-

nal acquired at the ﬁngers using the protocol proposed

in (Silva et al., 2011a) and Fig. 3(b) is the correspond-

ing ﬁltered signal with the overlap-add FIR method.

3.2 QRS Detection

We evaluated ﬁve QRS detection-ECG segmentation

algorithms: I. Christov (CHRIS) (Christov, 2004),

Engelse and Zeelenberg(EZEE) (Engelse and Zeelen-

berg, 1979), P. Hamilton (HAM) (Hamilton, 2002),

H. Gamboa (GAMBOA) (Gamboa, 2008), and ECG

Slope Sum Function (ESSF) (Zong et al., 2003).

I. Christov proposes a QRS complex detection

algorithm which applies an adaptive threshold to a

constructed complex lead signal. This threshold is

a linear combination of the following components:

a steep-slope threshold (M); an integrating threshold

(F); and a beat expectation threshold (R), (Christov,

2004).

The algorithm by Engelse and Zeelenberg was

proposed in 1979. In this work, we used the mod-

iﬁed real time version of this algorithm proposed in

(Loureno et al., 2012) (EZEEMod). It applies a differ-

entiator, a low pass ﬁlter and scans the resulting sig-

nal, in a 160ms moving window, with adaptivethresh-

olds. These are updated each time and are a function

of the maximum signal amplitude. Furthermore, the

algorithm ignores peaks within a 200ms interval of

the previous detected R-peak.

Hamilton proposed a QRS complex detection al-

gorithm that works by ﬁrst preprocessing the ECG

signal and then scanning and evaluating it according

to a set of rules related to the interval between con-

(a) Raw ECG Signal.

(b) Overlap-add FIR ﬁltered ECG Signal.

Figure 3: Raw and ﬁltered versions of an ECG signal ac-

quired at the ﬁngers using a sampling frequency of 1000Hz

and a frame size of 100ms.

secutive R peaks, (Hamilton, 2002).

For QRS detection, H. Gamboa proposes in (Gam-

boa, 2008) an algorithm that involves signal normal-

ization via histogram computation and threshold set-

ting and R peak detection via ECG signal derivative

and threshold surpassing.

The Slope Sum Function (SSF) is a weighted

moving average function which enhances the upslope

of the ECG signal and thereby makes the R-peak de-

tection easier. It was applied to the Blood Volume

Pulse (BVP) signal in (Zong et al., 2003). In this

work, we adapted the SSF for the ECG signal.

3.2.1 QRS Complex Validation through Voting

We have also implemented a QRS detection algorithm

that combines and validates the information from the

above ﬁve algorithms. In summary, we ran each algo-

rithm separately and then decide if each detected QRS

complex is valid by voting. Two criteria were used: a)

one of the algorithms is taken as reference and each

QRS complex is considered valid only if at least two

of the other algorithms also detected the same QRS

complex (within a 50ms tolerance) - majority voting;

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

230

and b) from the set of all QRS complexes detected

(by the ﬁve algorithms), we validate only those that

are common to all ﬁve algorithms - unanimity voting.

3.3 Segmentation

After the ECG preprocessing, we perform ECG seg-

mentation by evaluating each QRS complex and clip-

ping the ECG signal in an interval 200ms to the left of

the R-peak to 400ms to the right (values based on the

typical duration of the P-Q and S-T complexes). In

this way, information from the QRS complex as well

as from the P and T waves is assured, and an ECG pat-

tern is generated for each heartbeat. Fig. 4 illustrates

the procedure.

Figure 4: ECG segmentation example: the arrows and

dashed lines indicate the length and limits of each segment.

On a side note, compression may occur in the

ECG signal when the heart rate increases signiﬁcantly

and the ﬁxed -200ms to +400ms interval would not be

the most suited one as it may contain information of

2 heart beats. This was not a problem in our exper-

imental protocol but future work should take this in

consideration.

3.4 Pattern Recognition

In this work, pattern recognition is performed by eval-

uating the Euclidean Distance (ED) between differ-

ent ECG patterns. In the enrollment phase, the sub-

ject interacts with the biometric system, ECG data is

acquired, ﬁltered, segmented and the ECG segments

(or combinations of them) are saved as ECG patterns.

These are called the train patterns. In the authentica-

tion or identiﬁcation phases, the subject goes through

the same process, but this time the ECG segments

saved as ECG patterns are considered test patterns.

These are compared with the train patterns available

and the user is recognized if the k-smallest EDs be-

tween the test patterns and the train patterns are below

a deﬁned threshold. In summary, we perform pattern

matching with a k-Nearest Neighbor (k-NN) classiﬁer

followed by a threshold test.

4 EXPERIMENTAL RESULTS &

DISCUSSION

4.1 QRS Detection Performance

We implemented each of the QRS detection algo-

rithms in Python and ran them with data from the

Physionet database (Goldberger et al., 2000): the

MIT-BIH Arrhythmia Database (MITDB) (Moody

and Mark, 2001) which contains 48 half-hour ECG

recordings from 47 subjects, and the MIT-BIH Noise

Stress Test Database (NSTDB) (Moody et al., 1984)

which consists of 12 half-hour ECG recordings.

Table 1 summarizes the QRS detection results

(mean and standard deviation) for each of the previ-

ously presented QRS detection algorithms and for the

two annotated databases used (column 1). The sec-

ond column presents the algorithm’s performance as

measured by the ratio between the number of correct

beats and the number of total detected beats (correct

plus incorrect).

Per formance =

correctbeats

incorrectbeats+ correctbeats

A detected beat is correct if it is within a 50ms toler-

ance interval from an annotated beat. Otherwise it is

considered incorrect. The third column presents the

correctly detected beats deviation error (DE), com-

puted as the absolute value of the difference between

the detected beat time instant and the annotated beat

time instant. The fourth column presents the false

positive results. It shows the mean and standard devi-

ation values of the percentage of false positives, i.e.,

the ratio between the number of incorrect beats and

the total number of detected beats. Below columns 2

to 4, there are 2 sub-columns indicating whether the

results correspond to the algorithm on its own (Solo)

or whether the results were evaluated using the major-

ity criterion.

The QRS detection performance is above 90% for

all ﬁve algorithms and the MITDB dataset. Also, all

of them have a worse performance in the NSTDB

dataset with the EZEEMod, HAM, and GAMBOA al-

gorithms dropping to a performance around 75%, and

the CHRIS and ESSF algorithms to around 85%. The

deviation error is the lowest in the ESSF case with a

mean of 3ms and 2ms for the MITDB and NSTDB

datasets. The number of incorrect beats is below 10%

in all cases and for the MITDB dataset and is higher

OnRealTimeECGSegmentationAlgorithmsforBiometricApplications

231

Table 1: QRS detection performance, deviation error and false positive results. Values are: mean±standard deviation.

Algorithm Performance (%) Deviation Error (ms) False Positives (%)

Database Solo Majority Solo Majority Solo Majority

EZEEMod

MITDB 90.5±17.3 92.0±15.3 5.2±5.6 5.1±5.3 9.5±17.3 8.0±15.3

NSTDB 75.6±12.7 85.0±9.1 6.3±2.8 5.9±2.4 24.4±12.7 15.0±9.1

CHRIS

MITDB 94.3±7.8 93.3±12.6 13.1±6.1 13.1±5.9 5.7±7.8 6.7±12.6

NSTDB 83.3±12.1 85.5±10.9 13.6±4.2 13.3±4.0 16.7±12.1 14.5±10.9

HAM

MITDB 90.9±13.6 92.5±12.8 9.8±6.2 9.8±6.2 9.1±13.6 7.5±12.8

NSTDB 73.2±15.3 84.4±10.4 11.7±7.6 11.1±7.4 26.8±15.3 15.6±10.4

GAMBOA

MITDB 92.3±13.2 92.6±13.1 14.3±17.0 14.3±17.0 7.7±13.2 7.4±13.1

NSTDB 74.1±30.2 74.9±30.5 18.3±17.2 18.3±17.2 25.9±30.2 25.1±30.5

ESSF

MITDB 91.6±15.5 92.0±15.1 3.4±3.9 3.4±3.9 8.4±15.5 8.0±15.1

NSTDB 88.8±7.2 89.8±7.3 2.3±1.1 2.2±1.2 11.2± 7.2 10.2±7.3

for the NSTDB dataset, ranging from 11% (ESSF) to

27% (CHRIS).

Overall, the CHRIS and ESSF algorithms cor-

rectly detects more beats, with the latter presenting

a better accuracy with respect to the beat position.

Also, the ESSF algorithm is the most robust one

as the results with the two different datasets match

more closely than in the other algorithms, and as it

presents a better performance in the presence of noise

(NSTDB dataset).

The application of the majority criterion reduces

the number of false positives detected, which boosts

performance (by increasing the ratio between cor-

rect beats and total detected beats) in all cases except

one. In the CHRIS-MITDB, the case with higher per-

formance in the solo mode, the majority voting de-

creased the number of true positives (correctly de-

tected beats), and consequently decreased the perfor-

mance, due to the fact that the majority of the other

algorithms were not able to detect some of the beats

detected with the CHRIS algorithm.

Table 2 complements Table 1 by presenting the re-

sults obtained using the unanimity criterion. In this

case, each QRS complex detected is validated only

if it was detected by all ﬁve algorithms. This re-

duces the number of false positives detected but also

removes beats that were correctly detected by some

algorithms but not all. The performance is similar to

the mean value of the performances of each algorithm

separately.

Table 3 also complements Table 1 by presenting

the mean and standard deviation of the number of de-

tected QRS complexes for each algorithm and dataset

(MITDB and NSTDB), as well as the annotated data

values (ANNOTATED).

Table 2: QRS detection results with the Unanimity crite-

rion. Values are: mean±standard deviation.

Unanimity MITDB NSTDB

Performance (%) 93.8±12.7 79.0±29.7

False Positives (%) 6.2±12.7 21.0±29.7

4.2 Biometric Recognition Evaluation

In this case, a dataset previously collected by our

group was used (Silva et al., 2011a). It is composed

of data from 62 subjects (47 males and 15 females)

with ages in the 31.1±9.46 range. The ECG data was

collected from the subject’s hands for approximately

2 minutes using Ag/AgCl electrodes. The ECG sen-

sor is characterized by a gain of 1000 and an analog

band pass ﬁltering (1Hz to 30 Hz).

We used a subset of 31 individuals and choose this

dataset as our interest is in hand ECG based biomet-

rics. This allows us to test the biometric performance

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

232

Table 3: Mean and standard deviation of the number of de-

tected QRS complexes.

Algorithm

QRS complexes

MITDB NSTDB

ANNOTATED 2079.4±468.4 1854.5±311.5

EZEEMod 2176.0±493.0 2160.3±155.7

CHRIS 2162.5±389.9 2008.8±252.4

HAM 2275.8±429.6 2441.3±319.7

GAMBOA 2136.4±516.9 1526.1±694.6

ESSF 2114.3±588.8 1759.7±309.7

of the algorithms with signals acquired using a practi-

cal and non-intrusiveprocedure, in a biometrics point-

of-view, and that have a lower signal-to-noise ratio

(when compared with ECG signals acquired with a

set of electrodes at the thorax). The QRS detection

performance was not evaluated with this dataset as

ground-truth annotations are not available.

We performed three experiences: one with ECG

segments, one with mean waves of 3 ECG segments,

and one with mean waves of 5 ECG segments. ECG

segments for each individual were generated from

the raw data, following the steps presented in Sec-

tion 3. For each QRS detection algorithm, we cre-

ated a dataset with these ECG segments and two other

datasets in which each entry corresponds to a mean

curve of (3 or 5) adjacent ECG segments. For each

dataset constructed, the biometric recognition evalua-

tion was performed10 times using 60 random patterns

for each individual (30 for training and 30 for test).

To measure the system’s effectiveness, the False

Acceptance Rate (FAR), False Rejection Rate (FRR),

Receiver Operating Characteristic (ROC), and the

Equal Error Rate (EER) measurements are commonly

used. The FAR measures the rate at which a non-

authorized individual is accepted; the FRR is the fre-

quency at which the system rejects an authorized indi-

vidual; the ROC plots the rate at which an authorized

person is granted access (True Positive Rate) versus

the FAR; and the EER is the rate at which the FRR

equals the FAR. The EER is inversely proportional to

the biometric system’s accuracy: the lower the EER,

the more accurate the system is.

4.2.1 Authentication

Table 4 presents the authentication results obtained.

The performance is similar in all ﬁve algorithms and

the EZEEMod and ESSF algorithms are the ones with

the best results overall. This is justiﬁed by the fact

that, as we saw in the previous section, these algo-

rithms present higher accuracy with respect to the R-

peak position detection, which is crucial for the seg-

mentation step and the comparison between patterns.

Furthermore, using mean waves, as opposed to single

segments, presents better results in most cases (EER

is lower in all cases except in the CHRIS case).

The recognition rates using the majority criterion

are similar to the ones using the algorithm separately.

This is justiﬁed with the fact that all algorithms have

similar QRS detection results with this dataset as we

observed posteriorly.

The unanimity criterion results compete with the

best results obtained in all other cases: 5.8±0.5

(Unanimity-Segments) versus 5.7±0.3 (EZEEMod

and SSF-Solo-Segments); 3.0±0.4 (Unanimity-Mean

Waves of 3) versus 3.2±0.5 (EZEEMod-Solo-Mean

Waves of 3); and 2.0±0.4 (Unanimity-Mean Waves

of 5) versus 2.4±0.6 (SSF-Majority-Mean Waves of

5). Fig. 5 presents the FAR and FRR curves for the

lowest EER obtained (Unanimity-Mean Waves of 5).

Table 4: Authentication results: mean and standard devia-

tion of the EER for the 10 runs of the different trials per-

formed.

Algorithm

EER (%)

Segments

Mean Waves of

3 5

EZEEMod

Solo 5.7±0.3 3.2±0.5 2.5±0.7

Majority 6.2±0.4 3.4±0.3 3.0±0.5

CHRIS

Solo 5.8±0.5 7.2±0.4 6.3±0.9

Majority 5.9±0.4 7.5±0.7 6.1±0.7

HAM

Solo 5.8±0.3 3.7±0.4 3.1±0.6

Majority 6.0±0.4 3.9±0.4 3.4±0.6

GAMBOA

Solo 5.7±0.4 5.6±0.6 4.7±1.1

Majority 5.7±0.5 5.2±0.6 4.3±1.0

ESSF

Solo 5.7±0.3 3.5±0.3 2.5±0.5

Majority 5.8±0.4 3.5±0.3 2.4±0.6

Unanimity

5.8±0.5 3.0±0.4 2.0±0.4

OnRealTimeECGSegmentationAlgorithmsforBiometricApplications

233

Figure 5: FAR and FRR curves obtained using the Unanim-

ity criterion.

4.2.2 Identiﬁcation

In the identiﬁcation scenario, we evaluate the biomet-

ric system’s performance by computing the identiﬁ-

cation error. Each time the system fails to correctly

recognize a registered user is an identiﬁcation error.

Identiﬁcation results are presented in Table 5 and

a similar analysis to the one performed before for the

authentication scenario is used. As before, the er-

ror decreases as we switch from Segments to Mean

Waves of 3 and to Mean Waves of 5, for the majority

of the cases. Also, the use of the majority criterion

yields results similar to the ones obtained by the algo-

rithm on its own (Solo). When Segments are used, the

results are around 5-6% and HAM (Solo) presents the

lowest Identiﬁcation Error: 5.0±1.1%. Using Mean

Waves, the results improve(lower identiﬁcation error)

for all cases except for the CHRIS algorithm. Here,

the algorithms EZEEMod (Majority) and Unanimity

tie in ﬁrst place: best result is 2.5±0.7%. The lowest

identication error overall is obtained by applying the

Unanimity criterion to Mean Waves of 5: 1.6±0.6%.

5 CONCLUSIONS AND FUTURE

WORK

We evaluated and compared ﬁve real-time segmenta-

tion algorithms with respect to their performance and

biometric application. In future work, the warping ef-

fects, which results in the compression (expansion) of

the ECG beat waveform due to higher (lower) heart

rate should be considered and handled using adaptive

segmentation limits, for example. We also plan to an-

notate the hand ECG dataset and re-run the beat de-

tection performance tests to clearly know how each

algorithm behaves with the hand ECG signals, which

present a low signal-to-noise ratio, and high sensitiv-

ity to motion artifacts. About the biometric system,

Table 5: Identiﬁcation results: mean and standard deviation

of the identiﬁcation error for the 10 runs of the different

trials performed.

Algorithm

Identiﬁcation Error (%)

Segments

Mean Waves of

3 5

EZEEMod

Solo 5.8±1.1 2.8±0.9 2.1±0.7

Majority 5.5±0.8 2.5±0.7 2.1±0.7

CHRIS

Solo 6.7±0.8 9.1±1.3 7.4±1.8

Majority 6.6±0.8 8.8±1.3 7.6±1.7

HAM

Solo 5.0±1.1 3.6±1.0 3.3±1.5

Majority 5.1±1.2 4.0±0.8 3.5±1.1

GAMBOA

Solo 6.2±0.8 4.7±1.1 3.1±1.3

Majority 5.1±1.1 3.8±0.8 3.0±1.0

ESSF

Solo 5.1±1.3 3.1±0.8 2.4±1.2

Majority 5.3±1.1 2.9±0.8 2.3±1.1

Unanimity

5.6±1.0 2.5±0.7 1.6±0.5

other pattern matching techniques should also be con-

sidered (e.g., using the cosine distance between pat-

terns), as well as different values for the following

parameters: number of mean waves, number of train-

ing patterns, and number of nearest neighbors used in

the k-NN classiﬁer.

All ﬁve evaluated algorithms run in real-time: the

one by Hamilton, Gamboa, and the ESSF are sim-

pler to implement; Christov’s algorithm and the ESSF

present higher robustness and beat detection perfor-

mance; the Engelse and Zeelenberg’s modiﬁed ver-

sion and the ESSF are the ones with higher beat po-

sition accuracy. Overall, we were able to boost the

QRS detection performance by combining the infor-

mation from all algorithms in a voting manner - either

by majority or unanimity.

For biometric recognition, combining the infor-

mation from all algorithms with the unanimity crite-

rion and using mean curves of adjacent 5 ECG seg-

ments results in the lowest error percentage: 2.0±0.4

and 1.6±0.5 for the authentication and identiﬁcation

scenarios, respectively.

BIOSIGNALS2013-InternationalConferenceonBio-inspiredSystemsandSignalProcessing

234

ACKNOWLEDGEMENTS

This work was partially funded by the IT - In-

stituto de Telecomunicac¸˜oes under the grant, “An-

droid Biometrics Platform”, and by the Fundac¸˜ao

para a Ciˆencia e Tecnologia (FCT) under grants

SFRH/BD/65248/2009 and SFRH/PROTEC/49512/

2009, whose support the authors gratefully acknowl-

edge.

REFERENCES

Biel, L., Petterson, O., Phillipson, L., and Wide, P. (2001).

ECG analysis: A new approach in human identiﬁca-

tion. IEEE Trans Instrumentation and Measurement,

50(3):808–812.

Chan, A. D. C., Hamdy, M. M., Badre, A., and Badee, V.

(2008). Wavelet distance measure for person identi-

ﬁcation using electrocardiograms. IEEE Trans on In-

strumentation and Measurement.

Christov, I. I. (2004). Real time electrocardiogram QRS

detection using combined adaptive threshold. Biomed

Eng Online, 3(1).

Chung, E. K. (2000). Pocketguide to ECG Diagnosis.

Blackwell Publishing Professional.

Coutinho, D. P., Fred, A. L. N., and Figueiredo, M. A. T.

(2010). String matching approach for ECG-based bio-

metrics. Pattern Recognition, 16th Portuguese Con-

ference on.

Engelse, W. A. H. and Zeelenberg, C. (1979). A single scan

algorithm for QRS-detection and feature extraction.

Computers in Cardiology, 6:37–42.

Gamboa, H. (2008). Multi-Modal Behavioural Biometrics

Based on HCI and Electrophysiology. PhD thesis,

Universidadte T´ecnica de Lisboa, Instituto Superior

T´ecnico.

Goldberger, A., Amaral, L., Glass, L., Hausdorff, J., Ivanov,

P., Mark, R., Mietus, J., Moody, G., Peng, C., and

Stanley, H. (2000). Physiobank, physiotoolkit, and

physionet: Components of a new research resource for

complex physiologic signals. Circulation 101.

Hamilton, P. (2002). Open source ecg analysis. Computers

in Cardiology.

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

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

Pattern Recognition, 38(1):133–142.

Jain, A., Flynn, P., and Ross, A. (2007). Handbook of Bio-

metrics. Springer.

Jain, A. K., Ross, A., and Prabhakar, S. (2004). An intro-

duction to biometric recognition. IEEE Trans.Circuits

Syst. Video Techn., 14(1):4–20.

Jain, A. K., Ross, A. A., and Nandakumar, K. (2011). In-

troduction to Biometrics. Springer, 2011 edition.

Lourenc¸o, A., Silva, H., and Fred, A. (2011). Unveiling

the biometric potential of Finger-Based ECG signals.

Computational Intelligence and Neuroscience, 2011.

Loureno, A., Silva, H., Leite, P., Loureno, R., and Fred, A.

L. N. (2012). Real time electrocardiogram segmenta-

tion for ﬁnger based ecg biometrics. In Huffel, S. V.,

Correia, C. M. B. A., Fred, A. L. N., and Gamboa, H.,

editors, BIOSIGNALS, pages 49–54. SciTePress.

Moody, G. and Mark, R. (2001). The impact of the mit-bih

arrhythmia database. IEEE Eng in Med and Biol.

Moody, G., Muldrow, W., and Mark, R. (1984). A noise

stress test for arrhythmia detectors. Computers in Car-

diology.

Oppenheim, A. V. and Schafer, R. W. (1975). Digital Signal

Processing. Prentice Hall.

Shen, T. (2005). Biometric Identity Veriﬁcation Based on

Electrocardiogram. PhD thesis, University of Wis-

consin.

Silva, H., Fred, A., and Lourenc¸o, A. (2011a). Check your

biosignals here: Experiments on affective computing

and behavioral biometrics. In Proc Portuguese Conf.

on Pattern Recognition - RecPad, Porto, Portugal.

Silva, H., Lourenc¸o, A., Lourenc¸o, R., Leite, P., Coutinho,

D., and Fred, A. (2011b). Study and evaluation of

a single differential sensor design based on electro-

textile electrodes for ECG biometrics applications. In

Proc IEEE Sensors, Limerick, Ireland.

Wang, Y., Agraﬁoti, F., Hatzinakos, D., and Plataniotis,

K. N. (2008). Analysis of human electrocardiogram

for biometric recognition. EURASIP J. Adv. Signal

Process, 2008:19.

Zong, W., Heldt, T., Moody, G., and Mark, R. (2003). An

open-source algorithm to detect onset of arterial blood

pressure pulses. Computers in Cardiology, 30:259–

262.

OnRealTimeECGSegmentationAlgorithmsforBiometricApplications

235