Comparative Study of Medical-grade and Off-the-Person ECG Systems

Carlos Carreiras

, Andr

e Lourenc¸o

1,2

, Hugo Pl

acido da Silva

and Ana Fred

Instituto de Telecomunicac¸

oes, Av. Rovisco Pais 1, Lisboa, Portugal

Instituto Superior de Engenharia de Lisboa, R. Cons. Em

ıdio Navarro 1, Lisboa, Portugal

Keywords:

Pervasive Electrocardiology, Off-the-Person ECG Sensor, Morphological Analysis.

Abstract:

Simpliﬁed, off-the-person electrocardiogram (ECG) sensor designs enable the introduction into everyday life

of continuous, and pervasive ECG acquisition paradigms. This signiﬁcantly augments the potential of ap-

plications such as remote health monitoring, emotion assessment, and ECG biometrics, among others. We

accomplish this by reducing the number of contact points to just two, making use of a groundless setting. Ad-

ditionally, acquisition is made at the hand palms or ﬁngers through dry electrodes. In this paper, we describe

such a system, comparing it against a standard 12-lead, medical-grade ECG system. Our analysis is based

on the morphological similarity between individual heartbeat waveforms, as well as the general similarity be-

tween the synchronized time series. We show that the ECG signal acquired at the hands with our sensor is

highly correlated with lead I from the standard system.

1 INTRODUCTION

The measurement and recording of the electrical ac-

tivity of the heart, commonly known as the electro-

cardiogram (ECG), has come a long way since its

introduction at the end of the 19th century. The

ﬁrst systematic study of the electrical activity of the

human heart was performed by Augustus Waller in

1887 (Besterman and Creese, 1979). He used a Lipp-

mann capillary electrometer to measure the electri-

cal potential between two surface electrodes placed

on the front and back of the chest. This device con-

sisted of a capillary tube partially ﬁlled with mercury

and a solution of sulfuric acid. The electrodes were

connected to either side of the tube; when an electric

pulse arrives, the surface tension of the mercury is al-

tered, allowing it to climb up a small distance in the

capillary tube. Weller combined the use of the Lipp-

mann electrometer with a light source, allowing him

to project the oscillations of the mercury in the capil-

lary onto a photographic plate; the plate moved with

the help of a toy train, producing a real-time recording

of the heartbeat.

A few years later, Willem Einthoven used a string

galvanometer to signiﬁcantly advance the knowledge

about the electrical activity of the heart, naming

the various deﬂections of the ECG (P, Q, R, S, T,

and later U), standardizing the use of three limb

leads (Einthoven’s Triangle), and correlating numer-

ous electrocardiographic features with cardiovascular

disorders (Barold, 2003). The string galvanometer

consisted of a long conductive ﬁlament; when sub-

jected to a strong magnetic ﬁeld, the ﬁalment oscil-

lates in proportion to the current traveling through it.

This movement is then ampliﬁed and projected onto a

moving photographic plate.

Fast forward to the present, the ECG has become a

perfectly established and mainstream technique, pro-

viding vital information for the diagnosis and preven-

tion of a wide array of cardiovascular disorders (Drew

et al., 2004; Chung, 1996). Clinical practice relies

mainly on the widespread short-term (< 1 minute)

12-lead ECG for diagnosis, and, in selected cases,

on Holter monitors (∼ 24 hour assessment). Nev-

ertheless, the outreach of ECG data acquisition and

processing can still be signiﬁcantly improved upon,

in the context of a pervasive healthcare framework,

with off-the-person ECG sensor designs. The goal

of off-the-person approaches is not to replace exist-

ing data acquisition procedures, but to enhance and

complement current practices with a simpliﬁed sen-

sor setup that can be introduced, transparently to the

subject, in multiple aspects of his everyday life. This

enables a more comprehensive assessment of cardio-

vascular function, contributing to the development

of preventive behaviors and methodologies. Also, it

opens the door to many potential applications, such as

continuous monitoring, non-intrusive emotion assess-

115

Carreiras C., Lourenço A., Plácido da Silva H. and Fred A..

Comparative Study of Medical-grade and Off-the-Person ECG Systems.

DOI: 10.5220/0004675501150120

In Proceedings of the International Congress on Cardiovascular Technologies (IWoPE-2013), pages 115-120

ISBN: 978-989-8565-78-5

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

ment (Haag et al., 2004; Medina, 2009) and ECG bio-

metrics (Lourenc¸o et al., 2011; Odinaka et al., 2012),

among others.

In previous work by our team, we have proposed

a simpliﬁed, off-the-person ECG sensor design tar-

geted at data acquisition in a pervasive framework

(Silva et al., 2011). In this paper, we compare the

performance of this sensor against a medical-grade

12-lead standard ECG device, assessing the morpho-

logical and structural differences between both sig-

nals. The remainder of this paper is organized as fol-

lows: Section 2 provides an overview of our off-the-

person sensor design, highlighting the main technical

options, and details the acquisition setup; Section 3

describes the signal processing techniques employed,

and the similarity measures used to compare the sig-

nals under study; ﬁnally, Section 4 outlines the main

conclusions.

2 EXPERIMENTAL SETUP

With the goal of designing a non-intrusive, pervasive

ECG sensor, we have been focusing our work on min-

imizing the number of electrical contact points with

the subject’s body, and eliminating the need for any

gel or conductive paste in the electrode-skin inter-

face. In this section, we present the general design

guidelines adopted for the development of our off-the-

person sensor, as well as the description of the signal

acquisition setup used for the comparison of our sen-

sor with a medical-grade ECG system.

2.1 Off-the-Person Sensor

Targeting the context of pervasive ECG acquisition,

our sensor design uses only two dry electrodes, be-

ing based on the classical voltage potential differen-

tial principle. In particular, the sensor does not re-

quire the placement of a ground electrode, which is

replaced by a reference voltage produced by the cir-

cuit (commonly known as a virtual ground).

According to the literature (Webster, 2009), typi-

cal physiological ECG signals range from 0.5 to 4 mV

in amplitude, with a spectral content between 0.01

and 250 Hz. These small amplitudes require the use

of high-quality, low-noise instrumentation and oper-

ational ampliﬁers, with high common-mode rejection

at the input. The global gain of our system is set at

2000. However, given the use of dry electrodes and a

groundless setting, our system is more prone to high-

frequency noise. Therefore, the sensor incorporates a

band-pass ﬁlter with a passing band between 0.5 and

40Hz, simultaneously reducing the effect of the low-

frequency respiratory modulation, as well as the pow-

erline and high-frequency noise.

Figure 1: Schematic representation of the classical 12-

lead ECG placement system, along with the placement of

our off-the-person sensor approach, with the dry electrodes

placed at the hands (drawings not to scale).

Figure 1 illustrates the typical 12-lead ECG sys-

tem, as well as the placement of our off-the-person

sensor, using two dry electrodes placed at the hands.

Our ECG sensor is integrated into a broader signal ac-

quisition hardware framework, the BITalino platform

(Guerreiro et al., 2013), which wirelessly transmits

the acquired ECG signal to a computing device. Note

that, although our sensor was speciﬁcally designed

for 1-lead measurements at the hand palms or ﬁngers

with virtual ground, it can also be used with a ground

electrode, being capable of emulating any of the bipo-

lar ECG leads (the Einthoven leads: I, II and III).

2.2 Acquisition and Preprocessing

We conducted experimental tests involving 11 vol-

unteers (5 females and 6 males) in which simul-

taneous recordings were performed using a Philips

PageWriter Trim series ECG device (hereinafter re-

ferred to as Philips), and our off-the-person sensor

design (hereinafter referred to as BITalino) with data

acquisition performed using the SignalBIT platform

(Alves et al., 2013). The acquisitions were made on

two sessions, separated by ∼ 1.5 months, but three of

CARDIOTECHNIX2013-InternationalCongressonCardiovascularTechnologies

116

the subjects only partook in one of the sessions (sub-

jects S1, S3 and S6).

The Philips equipment was used in the standard

12-lead conﬁguration (I-III, V1-V6, aVF, aVL, aVR),

as shown in Figure 1, with conductive gel applied to

each of the electrodes, which were placed on the sub-

jects by an experienced technician. It acquires sig-

nal data at 500 Hz, with 16 bit resolution, and storage

is made in a proprietary ﬁle format. Our sensor was

used in the virtual ground setting with dry Ag/AgCl

electrodes, and the subjects were told to hold the elec-

trodes at the hands. The BITalino device acquires sig-

nal data at 1000 Hz, with 10 bit resolution, and stor-

age is made in the open StorageBIT ﬁle format (Car-

reiras et al., 2013a). Given the fact that the Philips

system records only 10 seconds of data for each run,

we acquired three sequential runs, thus amounting to

30 seconds of data per subject per session.

The main application of our sensor has been to

ECG biometrics (Lourenc¸o et al., 2011; Silva et al.,

2013; Carreiras et al., 2013b), reason for which both

signals were ﬁltered with a Finite Impulse Response

(FIR) bandpass ﬁlter with a Hamming window of

300ms, and cutoff frequencies of 5 − 20Hz. These

parameters were determined to be appropriate for this

speciﬁc application (Lourenc¸o et al., 2012).

Additionally, for the comparison measures we em-

ploy in this paper, it is necessary to have both signals

with the same sample rate. Therefore, and given the

characteristics of the ﬁlter just described, we down-

sampled the BITalino signal by a factor of 2, making

it match the 500 Hz sampling rate of the Philips de-

vice.

3 COMPARATIVE STUDY

We base our comparative study of the BITalino sen-

sor and the Philips system on two major approaches:

morphological and synchronized analyses. While the

ﬁrst approach focuses on the shape similarity between

individual heartbeat waveforms, the second focuses

on analyzing the correlation between the signals after

a synchronization step, i.e. the correction of the time

delay between the two independent data acquisition

systems.

3.1 Morphological Analysis

In this paper, morphological analysis is understood to

be the comparison of the shape similarity between in-

dividual heartbeat waveforms, as measured by a spec-

iﬁed metric. This requires the use of a QRS detection

algorithm, which enables the segmentation and ex-

traction of the individual heartbeat waveforms from

the ECG time series. For that purpose, we used the

modiﬁed Engzee algorithm described in (Lourenc¸o

et al., 2012), which identiﬁes the positions of the R

peaks in the ECG. We then extract the signal seg-

ment corresponding to 200ms before and 400 ms after

each R peak, amounting to a total of 600ms, which

corresponds to the RR interval with a heart rate of

100bpm (see Figure 2). We applied this segmenta-

tion algorithm to the signals from both devices, col-

lecting, from each subject, all the identiﬁed heartbeat

waveforms across both sessions. Note that in the case

of the Philips device, segmentation was performed us-

ing the signal from lead I, using it as reference for the

other leads, i.e. the positions of the R peaks found

using lead I were used to extract the individual heart-

beats from the other leads.

10 .0 10 .5 11 .0

11 .5

12 .0 12 .5 13 .0

Tim e (s)

15 0

10 0

15 0

20 0

25 0

Am p lit u d e

200 ms 400 ms

Figure 2: Example of the segmentation and heartbeat wave-

form extraction procedure; the positions of the R peaks are

plotted in red vertical lines; the limits of one of the heart-

beats are shown in green.

Figure 3 shows the heartbeat waveforms obtained

for one of the subjects, comparing the BITalino heart-

beats with each of the leads from the Philips system.

We can observe that the most visually similar Philips

lead to the BITalino heartbeats is lead I, as intended

by the design of the sensor described in the previous

section, with an almost complete overlap between the

two. It is also possible to distinguish two groups of

heartbeats in the Figure, identiﬁable, for example, by

looking at the T waves. This is due to the fact that

the Figure shows the heartbeats from both acquisition

sessions, for which the subject happened to have dif-

ferent resting heart rates. As it is known (Simoons

and Hugenholtz, 1975), the heart rate inﬂuences the

relative position of the ECG waves, in particular the

T wave.

For the quantitative analysis, given the use of dif-

ferent electrode types and gains between the Philips

and BITalino devices, we adopted the cosine distance

as similarity metric (Equation 1), which is bounded

between 0 and 1, respectively representing complete

similarity (vectors with same orientation) and com-

plete dissimilarity (vectors at π/2).

ComparativeStudyofMedical-gradeandOff-the-PersonECGSystems

117

Figure 3: Segmented heartbeat waveforms obtained for sub-

ject S8 with both devices; the BITalino heartbeats are repro-

duced in each plot, comparing them with each of the leads

from the Philips system.

cos

) = 1 −

∑

k=1

[k]x

[k]

∑

k=1

[k]

∑

k=1

[k]

(1)

We use this metric to compare, for each subject,

the distance of the segmented heartbeat waveforms:

a) within the BITalino sensor (BIT vs BIT); b) within

lead I of the Philips system (I vs I); and c) of

the BITalino device against the segmented heartbeat

waveforms of each of the Philips leads (BIT vs X).

Figure 4 shows the boxplots describing the general

distribution characteristics, across all subjects, of the

cosine distances obtained for each of the previously

mentioned items. We can see that the most simi-

lar Philips lead to the BITalino sensor is lead I, as it

presents the lowest median and narrowest Interquar-

tile Range (IQR) of the compared leads. Additionally,

these values are the most similar to the ones obtained

in the BIT vs BIT and the I vs I cases. This reinforces,

with quantitative information, the conclusion drawn

earlier when comparing the heartbeat waveforms of

one of the subjects. Also of note is the fact that the

BIT vs BIT distribution appears to have a wider IQR

when compared to that of the I vs I case, as well as

two distinct groups of outliers. This suggests that the

BITalino sensor is somewhat more sensitive to noise.

To better analyze this question, we also show in Fig-

ure 5 the histograms of these two distributions, for

all subjects. Indeed, the BITalino sensor produces

distributions with higher means, with a slower decay

as the distance increases. In two speciﬁc cases (sub-

jects S6 and S9), the BITalino system originates some

heartbeats with high dissimilarity to the others, aris-

ing from the existence of spurious, noisy heartbeats

for these subjects (e.g. resulting from motion arti-

facts). Again, note that the two distribution modes

observable in subject S8 result from different resting

heart rates between the acquisition sessions.

Figure 4: Boxplots, across all subjects, of the cosine dis-

tance of the BITalino heartbeats against the heartbeats of

each of the Philips leads; the whiskers extend to the lowest

and highest data points still within 1.5 times the Interquar-

tile Range; crosses represent outliers.

Figure 5: Histograms, for all subjects, of the cosine dis-

tance between the BITalino heartbeats (BIT), and between

the heartbeats from lead I of the Philips system (I).

3.2 Synchronized Analysis

The synchronized analysis corresponds to the com-

parison of the two ECG sensor systems after the sig-

nals have been aligned in the time domain. We ac-

complish this by ﬁrst normalizing the DC offset and

amplitude of the signals, as shown in Equation 2,

where the original signal x[k] is subtracted of its mean

µ, and divided by the total amplitude variation.

norm

[k] =

x[k] − µ

max

{

x[k]

}

− min

{

x[k]

}

(2)

We then aligned the signals between corresponding

BITalino and Philips runs by computing the time

delay for which the cross-correlation between them

(Equation 3) is maximum. Note that we use the lead I

from the Philips system to determine the delay, given

that we demonstrated in the previous section that this

derivation is the same as the one used by the BITalino

sensor.

∗

= argmax

d∈Z

+∞

∑

k=−∞

x[k] y[k + d] (3)

Figure 6 shows an example of the synchronized

signals for one of the acquisition runs, where we can

observe that there is an almost exact match between

CARDIOTECHNIX2013-InternationalCongressonCardiovascularTechnologies

118

the two signals. This is further validated by comput-

ing the Root Mean Square Error (RMSE), as deﬁned

in Equation 4.

Figure 6: Synchronized BITalino and Philips lead I signals

(subject S8); normalization using Equation 2.

RMSE(x,y) =

∑

k=1

(x[k] − y[k])

(4)

Figure 7 shows, for each subject, the averaged

(across all runs) RMSE between the synchronized

BITalino signal and each of the Philips leads. We can

see that the error to lead I is indeed small, although we

lack a reference with which to formulate stronger con-

clusions. It is also interesting to note that subject S10

exhibits a different pattern to that of the other subjects

of the RMSE across the various leads. This arises due

to the subject having a rotation of the heart axis, re-

sulting in different projections onto the typical ECG

derivations.

Figure 7: Root Mean Square Error between the BITalino

and Philips lead I synchronized signals, for all subjects; val-

ues are the averages across all runs.

Finally, we use the cosine distance to compare

the individual heartbeats of the synchronized signals,

computing the distance between each BITalino heart-

beat waveform and the corresponding aligned Philips

heartbeat waveform. We show the results in Figure 8,

where we can observe that the comparison against

lead I exhibits a very small variability, and a median

close to zero. This allows us to conclude that the

BITalino and the Philips lead I signals are highly cor-

related.

Figure 8: Boxplots, across all subjects, of the cosine dis-

tance of the synchronized BITalino heartbeats against the

heartbeats of each of the Philips leads; the whiskers extend

to the lowest and highest data points still within 1.5 times

the Interquartile Range; crosses represent outliers.

4 CONCLUSIONS

In this paper we described a simpliﬁed ECG sensor

design targeted at pervasive, off-the-person acquisi-

tion, comparing it with a standard 12-lead, medical-

grade ECG system. Our analysis was based on the

morphological similarities between individual heart-

beat waveforms, and also on the general similarity be-

tween the synchronized time series.

The main conclusion of this paper is that the ECG

signal acquired at the hands with our sensor is most

similar to lead I from the standard system, showing a

high degree of correlation between them. This result

strongly encourages us to further investigate the off-

the-person acquisition paradigm, especially for ECG

biometrics. In this particular case, it is important,

in order to develop a practical system, to have ECG

acquisition setups that are non-intrusive and seam-

lessly integrate into everyday objects. In the future,

we would like to compare the recognition accuracy

of an ECG biometric system using our off-the-person

sensor design, against the standard 12-lead ECG sys-

tem.

ACKNOWLEDGEMENTS

This work was partially funded by Fundac¸

ao para a

encia e Tecnologia (FCT) under grants PTDC/EEI-

SII/2312/2012, SFRH/BD/65248/2009 and SFRH/PR

OTEC/49512/2009, whose support the authors grate-

fully acknowledge.

REFERENCES

Alves, A. P., Silva, H., Lourenc¸o, A., and Fred, A. (2013).

SignalBIT A web-based platform for real-time biosig-

ComparativeStudyofMedical-gradeandOff-the-PersonECGSystems

119

nal visualization and recording. In Proc. of the 10th

SIGMAP.

Barold, S. S. (2003). Willem Einthoven and the birth

of clinical electrocardiography a hundred years ago.

Cardiac electrophysiology review, 7(1):99–104.

Besterman, E. and Creese, R. (1979). Waller–pioneer of

electrocardiography. Brit. Heart J., 42(1):61.

Carreiras, C., da Silva, H. P., Lourenc¸o, A., and Fred, A.

L. N. (2013a). StorageBIT: A Metadata-aware, Exten-

sible, Semantic, and Hierarchical Database for Biosig-

nals. In Proc. of the 6th HEALTHINF.

Carreiras, C., Lourenc¸o, A., Silva, H., and Fred, A. L. N.

(2013b). A Unifying Approach to ECG Biometric

Recognition Using the Wavelet Transform. In Proc.

of the 10th ICIAR.

Chung, E. K. (1996). Pocket guide to ECG diagnosis.

Blackwell Science.

Drew, B. J., Califf, R. M., Funk, M., Kaufman, E. S., Kru-

coff, M. W., Laks, M. M., Macfarlane, P. W., Sommar-

gren, C., Swiryn, S., and Van Hare, G. F. (2004). Prac-

tice standards for electrocardiographic monitoring in

hospital settings. Circulation, 110(17):2721–2746.

Guerreiro, J., Martins, R., Silva, H., Lourenc¸o, A., and Fred,

A. L. N. (2013). BITalino: A Multimodal Platform

for Physiological Computing. In Proc. of the 10th

ICINCO.

Haag, A., Goronzy, S., Schaich, P., and Williams, J. (2004).

Emotion recognition using bio-sensors: First steps to-

wards an automatic system. In Aff. Dial. Syst., pages

36–48. Springer.

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

the biometric potential of Finger-Based ECG signals.

Computational Intelligence and Neuroscience, 2011.

Lourenc¸o, A., Silva, H., Leite, P., Lourenc¸o, R., and Fred,

A. (2012). Real time electrocardiogram segmentation

for ﬁnger based ECG biometric. In Proc. of the 5th

BIOSIGNALS, pages 49–54.

Lourenc¸o, A., Silva, H., and Fred, A. L. N. (2012). ECG-

based biometrics: A real time classiﬁcation approach.

In Proc. of the IEEE MLSP Workshop.

Medina, L. (2009). Identiﬁcation of stress states from ECG

signals using unsupervised learning methods. Mas-

ter’s thesis, Universidade T

ecnica de Lisboa, Instituto

Superior T

ecnico.

Odinaka, I., Lai, P.-H., Kaplan, A., O’Sullivan, J., Sirevaag,

E., and Rohrbaugh, J. (2012). ECG biometric recogni-

tion: A comparative analysis. IEEE Trans. Inf. Foren-

sics Security, 7(6):1812 –1824.

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 Conf.

Silva, H., Lourenc¸o, A., Canento, F., and Fred, A. (2013).

ECG Biometrics: Principles and Applications. In

Proc. of the 6th BIOSIGNALS.

Simoons, M. and Hugenholtz, P. (1975). Gradual changes

of ECG waveform during and after exercise in normal

subjects. Circulation, 52(4):570–577.

Webster, J. G. (2009). Medical Instrumentation Application

and Design. Wiley, 4th edition.

CARDIOTECHNIX2013-InternationalCongressonCardiovascularTechnologies

120