REAL TIME ELECTROCARDIOGRAM SEGMENTATION FOR

FINGER BASED ECG BIOMETRICS

Andr´e Lourenc¸o

1,2

, Hugo Silva

2,3

, Paulo Leite

, Renato Lourenc¸o

and Ana Fred

DEETC, ISEL-IPL, Lisbon, Portugal

Instituto de Telecomunicac¸˜oes, IST-UTL, Lisbon, Portugal

PLUX - Wireless Biosignals, S.A., Lisbon, Portugal

Keywords:

Biometrics, Electrocardiography (ECG), Biosignal processing, QRS-complexes detection, Real-time segmen-

tation.

Abstract:

In biometric recognition based on Electrocardiographic (ECG) signals, there are two main approaches for

feature extraction: ﬁducial and non-ﬁducial. Fiducial methods use points of interest within single heartbeat

waveforms, obtained by segmenting the ECG signal using QRS complexes as a reference. In this paper we

study several QRS detection algorithms, with the purpose of determining what is the best algorithm in the

context of ﬁnger based ECG biometrics using ﬁducial approaches; our main focus is the real-time segmen-

tation of ECG signals resulting on a set of single heart beats. We propose a method combining the adaptive

characteristics of the algorithm by Christov, with the strategy of the widely adopted Engelse and Zeelenberg

algorithm. Experimental results obtained for real-world data show that online approaches are competitive with

ofﬂine versions, and represent a contribution for the realization of real-time biometric recognition.

1 INTRODUCTION

Electrocardiographic (ECG) signals are a recent trend

in biometric recognition; they exhibit very appealing

characteristics, such as intrinsic liveliness detection,

and the fact that they do not depend on external phys-

ical landmarks, therefore being difﬁcult to spoof.

There are two main approaches for ECG fea-

ture extraction: ﬁducial and non-ﬁducial. Fiducial

methods use points of interest within single heartbeat

waveforms (Biel et al., 1999; Shen et al., 2002; Is-

rael et al., 2005; Silva et al., 2007; Lourenc¸o et al.,

2011), while non-ﬁducial aim at extracting discrimi-

native information without localizing reference points

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

Fiducial approaches locate reference points based

on the detection of single heartbeat waveforms. These

are obtained by segmenting the ECG signal, and QRS

complexes are generally used as a reference due to

their singularity. Figure 1 shows the typical ECG sig-

nal, with QRS complexes identiﬁed. This complex

represents the depolarization and re-polarization phe-

nomenon of the ventricles.

As with other biometric traits, current research fo-

cuses on the usability and the design of more conve-

nient acquisition setups, that can be used for practical

m V

Figure 1: ECG acquired at the chest from one patient of the

PTB-BIH (Oeff et al., ) control subjects database. The ECG

waveform is labeled with the corresponding complexes.

The P wave corresponds to the sinoatrial node triggering

impulse, the QRS complex is associated with the depolar-

ization process, and ﬁnally the T wave reﬂects the repolar-

ization process.

and daily applications. The acceptance of ECG based

methods requires real-time or near real-time authen-

tication/identiﬁcation, which can be obtained dimin-

ishing the acquisition and processing time.

In this paper we review several QRS detection al-

gorithms, with the purpose of determining what is the

best algorithm for ﬁnger based ECG biometrics using

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

REAL TIME ELECTROCARDIOGRAM SEGMENTATION FOR FINGER BASED ECG BIOMETRICS.

DOI: 10.5220/0003777300490054

In Proceedings of the International Conference on Bio-inspired Systems and Signal Processing (BIOSIGNALS-2012), pages 49-54

ISBN: 978-989-8425-89-8

 2012 SCITEPRESS (Science and Technology Publications, Lda.)

ﬁducial approaches, focusing on real-time identiﬁca-

tion of single heart beats.

The remainder of the paper is organized as fol-

lows. In Section 2, an overview of QRS detection

algorithms is presented. In Sections 3 and 4 we pro-

vide a brief overview of representative algorithms of

ofﬂine and online QRS detection algorithms, and de-

scribe the proposed modiﬁcations for both ofﬂine and

online algorithms. Finally, in Sections 5, 6 and 7 we

outline the experimental setup, the main results and

conclusions, respectively.

2 RELATED WORK: QRS

DETECTION ALGORITHMS

The automation of electrocardiogram analysis pro-

cesses found in the Holter test and in real-time pa-

tient monitoring, led to the development of algorithms

for the detection of QRS complexes. Due to its mor-

phology (see Figure 1), it serves as basis for com-

puting the heart rate, as a reference point for cardiac

cycle classiﬁcation schemes, and ECG data compres-

sion algorithms (Kohler et al., 2002). There are ap-

proaches based on signal derivatives and digital ﬁl-

tering (Friesen et al., 1990), or more complex ap-

proaches based on artiﬁcial neural networks, genetic

algorithms, wavelet transform (for more details see

(Kohler et al., 2002)).

Using the ECG as biometric, the work in

(Lourenc¸o et al., 2011) proposes the acquisition of

this signal at ﬁnger level, through a minimally in-

trusive 1-lead ECG setup recurring to Ag/AgCl elec-

trodes without gel as interface with the skin. This type

of setup implies more noise than the traditional acqui-

sition at the chest, in particular when using multiple

leads. Figure 2 illustrates signals collected at the ﬁn-

gers using dry Ag/AgCl electrodes, showing common

artifacts. Figure 2(a) corresponds to the presence of

motion artifacts; in this case the amplitude of the ECG

signal is quite close to the noise amplitude; Figure

2(b) illustrates a case where signals are corrupted by

high frequency powerline noise and electromyogram

noise.

The need for robust algorithms is clear when we

compare the abovementioned signals, with traditional

ECG signals collected at the chest using conductive

gel, as depicted in Figure 1, where the noise is practi-

cally inexistent.

The development of algorithms for real time anal-

ysis of the ECG begun in the 80’s. In (Pan and Tomp-

kins, 1985), an online QRS detection algorithm was

implemented in assembly language. It consisted in

digital bandpass ﬁltering to remove the noise, differ-

(a) Motion artifact

(b) Electrical and electromyogram interference

Figure 2: Example of ﬁnger ECG readings from two dif-

ferent subjects, collected using dry Ag/AgCl electrodes. As

we can observe, the signal quality is signiﬁcantly worst than

the one obtained at the chest (see ﬁgure 1.).

entiation to obtain information about the slope of the

QRS, followed by squaring to intensify this slope and

ﬁnally a moving window integrator to produce a sig-

nal that included information about the slope and the

width of the QRS complex. This process is divided

in three phases: learning phase 1; learning phase 2;

detection. The learning phase 1 required 2s to ini-

tialize the detection thresholds. The learning phase

2 requires two heartbeats to initialize the RR-interval

average and RR-interval limit values. The detection

phase produces pulses for each QRS complex. The

thresholds and other parameters of the algorithm are

periodically adjusted, to adapt to changing character-

istics of the signal.

In (Christov, 2004), an algorithm is presented that

follows the same principles, proposing the detection

based on an adaptative threshold. over the so called

complex lead signal, y[n]. This signal is obtained

by averaging the absolute value of the differentiated

versions of all available leads. The process is initi-

ated by digitally ﬁltering the input signal, x[n], to re-

move power-line interference and electromyographic

noise. The detection is performed by verifying when

y[n] is higher than a threshold, obtained through the

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

window W

Figure 3: Adaptation of the Engelse and Zeelenberg algo-

rithm for QRS detection algorithms (Engelse and Zeelen-

berg, 1979). In blue the original signal, x[n], and in red

[n].

linear combination of three components: M (Steep-

slope threshold); F (Integrating threshold for high

frequency signal components); R (Beat expectation

threshold).

Before presenting our QRS-complex detection in

real time, we overview the state of the art in ofﬂine

QRS detection, since in previous works on ECG bio-

metric recognition, this type of algorithms was used.

3 OFFLINE QRS DETECTION

AND IMPROVEMENTS

In (Friesen et al., 1990), nine different ofﬂine QRS

detection algorithms are compared in the presence of

noise. We build on the work by (Engelse and Zee-

lenberg, 1979), since it is considered to be one of the

more robust.

The QRS detection method by Engelse and Zee-

lenberg consists of the following. A digitally ﬁltered

version of the ECG signal, x[n], is passed through a

differentiator (Eq. 1), and then by the low pass ﬁlter

(Eq.2):

[n] = x[n] − x[n − 4], (1)

[n] =

∑

i=0

[n− i], where c

= [1, 4,6, 4, 1] (2)

Figure 3 presents the input signal, x[n], and the

corresponding processed signal y

[n]. The R peak de-

tection is based on the analysis of the negative lobs of

7200 7400 7600 7800 8000 8200 8400 8600 8800 9000 9200

−0.1

−0.05

0.05

0.1

0.15

0.2

x[n]

DF1

7200 7400 7600 7800 8000 8200 8400 8600 8800 9000 9200

−4

−2

[n]

Figure 4: Missing R detection.

[n], and it is identiﬁed automatically, by scanning

[n] using two thresholding operations.

The detection begins by ﬁnding the interval n ∈

thi+

, n

th f+

}, indicated in Figure 3, whose ampli-

tudes verify the condition: y

[n] > Th, with Th =

0.6max(y

[n]). Then, we look for a 160ms long win-

dow, W, to the right of n

th f+

, where the condition

[n] < −Th holds for a speciﬁed number of consec-

utive points (experimentally we found this number to

be at least 10 points). In (Engelse and Zeelenberg,

1979) more thresholding operations are proposed, but

we considered the described conditions enough.

Upon ﬁnding a candidate R peak, the original sig-

nal, x[n] is scanned inside the obtained windows, W,

and the peak is determined as the time instant corre-

spondig to the highest amplitude signal.

Taking as reference the identiﬁed R peak, we con-

tinue the analysis of y

[n] within its left and right

neighborhoods, determining the time instants were

it starts to be positive and comes down to negative

again. Within these intervals, we take the minimum

values of x[n] as the beginning of the Q and S com-

plexes, respectively.Using as reference the starting of

the Q and S complexes, we analyze x[n] ﬁnding the

maximum on their neighborhood to ﬁnd the complex

Q and S.

This algorithm presents the problem of having

ﬁxed thresholds, and if the analyzed ECG signal has

amplitude variations, its robustness is affected. As an

example, let’s consider the case presented in Figure

4, where one of the single heartbeat waveforms is not

recognized. As we can see by analyzing y

[n] (in red),

the negative lob is much less pronounced than in the

case of the complexes on the left and right, correctly

identiﬁed.

In order to solve this situation we tried to lower

the threshold Th, but doing so, the number of false

REAL TIME ELECTROCARDIOGRAM SEGMENTATION FOR FINGER BASED ECG BIOMETRICS

positives increases. To compensate for this phenom-

ena we introduce a second threshold, which enables

the identiﬁcation of negative lobes comprising R-peak

time instants. Its value is obtained from the lowest

value of y

[n] found in the analysis, and the condition

deﬁned as:

[n] < ThNew

ThNew = 0.7min(y

[n])

(3)

4 PROPOSAL FOR ONLINE QRS

DETECTION

Motivated by the ofﬂine algorithm of Engelse and

Zeelenberg, where the R complex is identiﬁed thresh-

olding two lobes of a differentiated version of x[n],

we propose to combine that strategy with adaptative

thresholds estimated along the acquisition process.

To determine this adaptive threshold we build on

the work by (Christov, 2004), using a threshold esti-

mation scheme similar to the one in (Christov and T,

2002) for the threshold M. In that work, the thresh-

old M is calculated using a temporal sliding window

of 5s. They form a buffer MM = M

, M

consisting in the concatenation of 5 partial thresholds,

, calculated by Equation 4 in each 5s sliding win-

dow; M is obtained according to Eq. 5.

= 0.6max(y[n]), (4)

M =



during the initial 5s

∑

i=1

rest of acquisition

(5)

The process continuously updates the partial

thresholds, erasing the older threshold, left shifting

the intermediate thresholds and calculating a new M

This new partial threshold, M

new

, can become quite

high, due to premature ventricular contraction, so if

new

> 1.5M

, then M

new

= 1.1M

The algorithm doesn’t allow detections of QRS

peaks 200 ms after the last one, and the M threshold is

decreased during the interval 200 to 1200ms follow-

ing the last detection, at a low slope, until reaching

60% of its value at 1200ms. After 1200ms, M remains

unchanged until a new detection is obtained.

Notice that the R peak detection algorithm by En-

gelse and Zeelenberg can be performed in real-time,

since the computations of equations 1 and 2 only re-

quires a buffer of 4 samples. Moreover the scanning

of y

[n] can be performed based on a sliding window,

following the approach of Christov.

With this approach we expect to obtain better de-

cision rules for the detection of R peaks.

5 EXPERIMENTAL SETUP

To evaluate and compare the described algorithms,

we performed extensive acquisitions, collecting data

from 62 subjects (47 males and 15 females) with an

average age of 31.1±9.46 years. Subjects were only

asked to rest their left/right hands in a setup built for

this propose.

Two custom ECG sensors (Silva et al., 2011)

were used for signal acquisition, one connected to the

Ag/AgCl electrodes, and another connected to Elec-

trolycra strips placed at the index and middle ﬁnger

levels, as depicted in Figure 5. The ECG sensors add

a total gain of 1000 and analog band pass ﬁltering be-

tween the 1-30Hz range.

Figure 5: Experimental apparatus.

To avoid ground coupling between both sen-

sors, two independent biosignalacquisition units were

used, one per sensor. Data acquisition was performed

using the commercially available bioPLUX research

system (PLUX, ), which enables Bluetooth wireless

transmission of the collected signals to the base sta-

tion. We used a sampling frequency of 1000Hz, and

12−bit resolution.

Synchronization of the acquisition units was per-

formed optically using a syncPLUX kit and a light-

dependent resistor (LDR) (PLUX, ). To one of the

systems a triggering switch was connected, which si-

multaneously activated the digital input port of the

system and a LED. To the other system, a LDR was

connected to one of the analog input channels, and

placed in direct contact the LED of the ﬁrst system, in

such way that a synchronization signal was obtained

whenever the LED was lit.

This allowed us to have the data collected by each

system synchronized, without recurring to any electri-

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing

cal connection between them. Signals were acquired

during a period of approximately 2 minutes, in which

the supervisor in charge of the experimental proce-

dure would describe the experiment, goals and related

work.

6 RESULTS AND DISCUSSION

Table 1 summarizes the results of the several ap-

proaches, both for the signal acquired using the

Ag/AgCl electrodes and Electrolycras. The ofﬂine al-

gorithms of Engelse and Zeelenberg are denoted by

EG-Butter and EG-FIR; the improvement obtained

using a third threshold by EG-3; and for the online

algorithms, the Christov algorithm is denoted by Chr,

and the proposed combination by Chr+EG.

To quantify the performance of each algorithm,

we present the number of segmented waves, and the

mean and standard deviation of the percentage of the

ones considered valid, taking as population the full

data set. A segment is considered valid if the obtained

RR interval does not deviate from the mean RR inter-

val in more than 10%. The rational of this criterion

is that a deviation above this margin should be due

to segmentation errors, since acquisitions at rest and

during a short period of time have very stable RR in-

tervals.

The number of segmented heartbeat waveforms

obtained from the dry Ag/AgCl electrode signals is

higher than the ones obtained by the Electrolycra, re-

vealing that the later signals have more noise. Com-

paring the performance of the online vs ofﬂine algo-

rithms, as expressed in the table, one can observe that

ofﬂine algorithms obtain more segments than online

versions, but the percentages of valid segments are

similar.

The algorithm that exhibits better performance is

the Engelse and Zeelenberg, both using Butterworth

ﬁlters - EG-Butter, or using Fir ﬁlters - EG-FIR. Re-

garding the online algorithms, the original algorithm

of Christov - Chr, is the one that presents higher

number of segments. Our approach that combines

the adaptative threshold with the Engelse and Zeelen-

berg is slighty worst. Figure 6 presents an example

these segmentations,with the single heartbeats super-

imposed and aligned according to the R peak. One

can see than even highly noisy single heartbeats are

being correctly segmented.

7 CONCLUSIONS

In this work we studied the segmentation of ECG

(a) Segmentation 1

(b) Segmentation 2

Figure 6: Example of segmentation of ﬁnger ECG readings

from two different subjects.

signals acquired at the ﬁngers using QRS-complex

detection algorithms. This type of process is of

paramount importance in the ﬁducial based algo-

rithms, since they require reference points to be cor-

rectly located on single hearbeat waveforms.

We focused an algorithm for ofﬂine ECG segmen-

tation by adaptation of the Engelse and Zeelenberg

algorithm, and the Christov algorithm for online ECG

segmentation. The online approaches have shown to

be competitive with ofﬂine versions, but their perfor-

mance is slightly worst. On the ECG biometric point

of view, these algorithms represent a contribution for

performing real-time biometric recognition.

As future work we intent to use a benchmark an-

notated ECG dataset to corroborate these conclusions.

ACKNOWLEDGEMENTS

This work was partially funded by Fundac¸˜ao para a

Ciˆencia e Tecnologia (FCT) under grants SFRH/BD/

REAL TIME ELECTROCARDIOGRAM SEGMENTATION FOR FINGER BASED ECG BIOMETRICS

Table 1: Experimental results for ECG Segmentation.

Type of Electrodes Type of Processing Algorithm #segments %valid - mean %valid - std

Ag/AgCl

Ofﬂine

EG-Butter 7614 96.5 6.6

EG-FIR 7322 97.9 4.3

EG-Butter-3 7625 96.4 6.7

EG-FIR-3 7719 97.8 4.6

Online

Chr 6482 94.5 12.5

Chr+EG 5971 92.4 10.4

Electrolycra

Ofﬂine

EG-Butter 6692 94.0 11.4

EG-FIR 6244 93.4 11.6

EG-Butter-3 6712 93.9 11.4

EG-FIR-3 6545 94.2 10.5

Online

Chr 5550 90.8 15.7

Chr+EG 5044 84.5 18.5

65248/2009 and SFRH/PROTEC/49512/2009, and by

the Departamento de Engenharia de Electr´onica e

Telecomunicac¸˜oes e de Computadores - ISEL, whose

support the authors gratefully acknowledge.

REFERENCES

Biel, L., Pettersson, O., Philipson, L., and Wide, P. (1999).

Ecg analysis: a new approach in human identiﬁcation.

In Proceedings of the 16th IEEE Instrumentation and

Measurement Technology Conference, IMTC/99., vol-

ume 1, pages 557 –561 vol.1.

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

Wavelet distance measure for person identiﬁcation us-

ing electrocardiograms. IEEE Transactions on Instru-

mentation and Measurement, pages 248–253.

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

detection using combined adaptive threshold. Biomed

Eng Online, 3(1).

Christov, I. I. and T, S. (2002). Steep slope method for real

time qrs detection. Electrotechniques and Electronics,

pages 13 –17.

Coutinho, D. P., Fred, A., and Figueiredo, M. (2010). Per-

sonal identiﬁcation and authentication based on one-

lead ECG using ziv-merhav cross parsing. In Proc.

Int. Workshop PRIS’2010, pages 15–24.

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

rithm for QRS detection and feature extraction. IEEE

Comp in Card, pages 37–42.

Friesen, G., Jannett, T., Yates, M. J. S., Quint, S., and Nagle,

H. (1990). A comparison of the noise sensitivity of

nine QRS detection algorithms. IEEE Transactions

on Biomedical Engineering, 37(1):85 –98.

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

and Wiederhold, B. K. (2005). ECG to identify indi-

viduals. Pattern Recogn., 38:133–142.

Kohler, B. U., Hennig, C., and Orglmeister, R. (2002). The

principles of software QRS detection. Engineering in

Medicine and Biology Magazine, IEEE, 21(1):42–57.

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

the biometric potential of Finger-Based ECG signals.

Computational Intelligence and Neuroscience, 2011.

Oeff, M., Koch, H., Bousseljot, R., and Kreiseler, D. The

PTB diagnostic ECG database. National Metrology

Institute of Germany.

Pan, J. and Tompkins, W. J. (1985). A real-time qrs detec-

tion algorithm. Biomedical Engineering, IEEE Trans-

actions on, BME-32(3):230 –236.

PLUX. bioPLUX research. http://www.plux.info.

Shen, T., Tompkins, W., and Hu, Y. (2002). One-lead QRS

for identity veriﬁcation. In Proceedings of the Second

Joint EMBS/BMES Conference, pages 62–63.

Silva, H., Gamboa, H., and Fred, A. (2007). One lead

ECG based personal identiﬁcation with feature sub-

space ensembles. In 5th International Conference on

Machine Learning and Data Mining, pages 770–783.

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

D., and Fred, A. (2011). Study and evaluation of a sin-

gle differential sensor design based on electro-textile

electrodes for ECG biometrics applications. In Proc

IEEE Sensors, Limerick, Ireland.

BIOSIGNALS 2012 - International Conference on Bio-inspired Systems and Signal Processing