Proceedings of the 2005 IEEE

Engineering in Medicine and Biology 27th Annual Conference

Shanghai, China, September 1-4, 2005

Rank Based Selection of Electrode Positions for a

Multi-Lead ECG Electrode Array
Christoph Hintermüller, Gerald Fischer, Michael Seger,
Bernhard Pfeifer, Robert Modre, Bernhard Tilg

Institute for Biomedical Signal Processing and Imaging

University for Health Sciences, Medical Informatics and Technology, Innsbruck Austria

Abstract— Methods for noninvasive imaging of electrical func- The local linear dependency (LLA) can be used as a
tion of the heart seem to become a clinical standard procedure measure for the information content in the BSP. We used this
the next years. Thus, the overall procedure has to meet clinical measure in [5] to locate sensitive areas by calculating so called
requirements as easy and fast application.
In this study we propose a new electrode array meeting clinical LLA maps and constructed an initial regular spaced layout
requirements such as easy to apply and compatibility with routine consisting of two L-shaped parts. This array, in the following
leads. Within body surface regions of high sensitivity, identified in addressed as scheme SB , is compared to a layout SA initially
a prior, information content based study, the number of required developed at the University of Amsterdam [6]. The electrode
electrodes was optimized using effort-gain plots. These plots were spacing of 4 cm for this prototype was governed by various
generated by applying a so called type one detector criterion.
The optimal array was selected from a set of 12 electrode
limitations and the necessity to compare it to scheme SA .
arrays. Each of them consists of two L-shaped regular spaced One of these limitations is that the posterior electrode po-
parts. The optimal array was found by comparing several layouts sitions have to be extracted and identified form the individual
and electrode densities to the electrode array we use for clinical MRI scans. As this is a rather tedious and cumbersome process
studies. It consists of 125 electrodes with a regular spacing we were forced to modify SA keeping the number of dorsal
between 2 cm and 3 cm.
electrodes low. Thus, we compare SB to a modified version of
SA instead of the original scheme SA proposed by [6].
Keywords ecg mapping, electrode array, forward and inverse
problems of electrocardiography, effort-gain plot
Technical limitations are imposed by the process of manu-
facture and the physical characteristics of the electrodes, such
I. I NTRODUCTION as tolerances in positioning the electrodes and the minimal
electrode diameter. As a tradeoff between methodical and tech-
Methods to image electrical excitation of the human heart
nical limitations and the intent to maximize the information
in a noninvasive way, also addressed as inverse electrocardio-
content of the recorded BSP we set the minimal electrode
graphic problem, have been developed the past 15 years by
spacing to 2 cm .
several groups (see [1]–[4]). For all methods body surface
In the following we describe methods developed to select
potentials (BSP) are recorded using a multi-lead electrode
the array matching the clinical, methodical and technical
array (typically 32-250 electrodes). As everyday clinical use
requirements best.
is intended the electrode array not only has to meet technical
needs but also clinical requirements as: II. M ETHODS
• easy and fast application by clinicians and nurses To solve the inverse electrocardiographic problem a volume
• low cost for storage and logistics conductor described by a patient individual torso model and
Recording the body surface potential is one possibility to the conductivities of different compartments is used. The
noninvasively obtain information upon the electric sources of corresponding electrocardiographic forward problem can be
the heart. It is used in combination with an individual volume described as follows
conductor model to solve the underlying electrocardiographic
inverse problem. The corresponding solution is sensitive to Ve (t) = INϕm (t) = Lϕm (t) . (1)
the recording location of the BSP due to varying linear
The lead field matrix L describes the relation between the
dependencies of the equation system obtained for the inverse
transmembrane potential vector ϕm (t) on the heart surface
electrocardiographic problem. The volume conductor model
and the electrode potentials Ve (t). It consists of the matrix
required to solve the electrocardiographic inverse problem is
I interpolating the discrete BSP to gain Ve (t) and the node
based upon a patient individual torso model consisting of the
field matrix N describing the patients volume conductor.
body surface, the right and left lung, the atria and ventricles
We computed the singular values si for the leadfield matrix
and the associated bloodmasses. The torso model is generated
L of an electrode array along equation
using volume data acquired via magnetic resonance imaging
(MRI) recored prior to the electrophysiological intervention. Σ QT .
L = UΣ (2)

0-7803-8740-6/05/$20.00 ©2005 IEEE. 2379

 index
r
m
si
∆ϕm 10
-1
r
SA,59 1.9
SA,62

SB,158
SB,62 1.8

SB,125S
SB,81
SB,102 1.7
-2
10 SB,125L
SB,125S 1.6
SB,158
SB,177
1.5

SB,202
SB,102
SB,199

SB,177

SB,199

SB,225
SB,202

SB,272
1.4

SB,253
-3 SB,225
10
SB,253
SB,272 1.3
kσkmax

SB,125L
kσkmean 1.2

SB,62
kσkmin
-4
10 1.1

SA,59
1

SB,81
0.9

SA,62
0 20 40 60 80 100 1 1.5 2 2.5 3 3.5 4 4.5
index m

Fig. 1. Singular values for the ventricular leadfield matrix L. The horizontal Fig. 2. Effort gain plot for the the ventricles. The circle indicates the
lines indicate the maximum, the mean and the minimum of the detector noise array SA,59 measured during clinical intervention. Its idealized version SA,62 is
kσk observed when estimating the effective rank R based upon a type one indicated by the square and the array SB,125S considered to be optimal is is
detector criterion. marked by the diamond.

The rectangular matrix U and Q contain the eigenvectors and the norm of kVe ( j)k can in this special case be estimated
spanning the orthonormal basis of the electrode ui and the as follows
source space qi and the diagonal matrix Σ contains the singular kVe ( j)k = kLqi= j ∆ϕm k = si= j ∆ϕm . (6)
values si as depicted by fig. 1. The potentials ϕm (t) generated
by the source under investigation can in general be expressed A detector in general is characterized by its measurement
by the following linear combination of qi range
√ ∆Vl , the noise σl detectable per lead, the norm kσk =
Mσl of the noise vector σ and the number of channels M.
ϕm (t) = ∆ϕm ∑αi (t)qi . (3) The type one detector criterion uses these numbers to classify
i all Ve ( j) with respect to their information content. It separates
vectors Ve ( j) resulting from the mapping of ϕm ( j) to the target
with the action potential amplitude ∆ϕm . The time dependent
space U of L described by (1) from vectors Ve ( j) = σ. The
coefficients αi (t) map qi onto ϕm (t). They range from αi (t) =
numerical rank R of L can be now be estimated by comparing
0 corresponding to a total suppression of frequencies contained
all Ve ( j) to σ applying
in qi to αi (t) = 1 which reflects a full mapping of qi onto
ϕm (t). R = j ⇔ si= j+1 ∆ϕm ≤ kσk < si= j ∆ϕm . (7)
For estimating the effective rank R of L we applied a so To be able to compare two array layouts Si and S0 we
called type one detector or receiver criterion based upon the computed the corresponding measures for effort mi,0 and gain
following assumptions. ri,0 based upon their number of electrodes Mi , M0 and the
1) The source distribution of ϕm (t) is equal to ϕm (t) = effective rank Ri and R0 of the corresponding Li L0 as follows
ϕm = const for all time points t reducing (3) to Ri Mi
ri,0 = mi,0 = . (8)
R0 M0
ϕm = ∆ϕm ∑αi qi . (4)
i If effort mi,0 and gain ri,0 are computed for several Si in
relation to S0 then the entire comparison between the arrays
2) The source space Q is successively mapped onto ϕm can be done by an effort gain (EG) plot as depicted in fig.
with all αi = 0 except for one αi = α j = 1. The index j 2. Another possibility is to estimate the relative gain ∆r =
is used to select one specific qi to map onto ϕm ( j) = ϕm . ri+1 − ri and compare it to the relative effort ∆m = mi+1 − mi
It ranges from j = 1 and the entire number of singular necessary to convert the Si to Si+1 . We used ∆m and ∆r to
values si of L. find an electrode array exposing a large r at m being as low
Applying this assumptions to (1) the corresponding BSP as possible based upon the following relation
vectors Ve ( j) measured by a detector such as an electrode
∆r ≤ β ∆m . (9)
array coupled to a digital amplifier can be expressed as
The factor β ensures that the results gained by (9) are similar
Ve ( j) = L∆ϕm qi= j (5) to the ones deduced by looking at the EG plot (fig. 2).

2380
 15.4 16,6

2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0 model SA,62 we were able to select the array matching clinical,
H methodical and technical requirements best from a set of
14 arrays.
2,0

We computed the corresponding rank R for 11 patients with

2,0

different size and body shape assuming the action potential

2,0

amplitude ∆ϕm to be 100 mV. The measures m and r for effort

2,0

and gain ((8)) were computed with respect to SA,62 resulting in

16,6

m = 1 and r = 1 comparing SA,62 to itself as indicated by the

2,0

square shaped marker in fig. 2.

2,0

From m and r a patient individual EG plot was generated

2,0

and the array scheme yielding a high r at a low m was selected

2,0

kσkmax

according to (9). We found that selecting the factor β to be

kσkmean

kσkmin
si
∆ϕm

index
r
m
1/4 returned the best results for the optimal array. The optimal
2,0

SB,272

array scheme for recording BSP linked to atrial activity is

SB,253

SB,225
7.4

SB,202
L
2,0

SB,199

equal to the 62 leads L-shaped array SB,62 .

SB,177

SB,158

SB,125L

SB,125S
2,0

When recording ECG related to the whole heart and/or the

SB,102

SB,81

SB,62

SA,62

SA,59
F
ventricles the 125 lead array SB,125S proved to be optimal for
all patients as indicated by the gray lines in tab. I. Increasing
12,0 8,0

2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0 2,0
the number of electrodes to more than 125 electrodes did not
H
increase the rank R significantly if at all. For patients with a
2,0

large torso the optimal number of electrodes is even lower as

indicated by the grey figures.
2,0
8,0

The same findings could be made by looking at the effort

2,0

gain plot as depicted in fig. 2. The optimal array selected

2,0

is indicated by the diamond shaped marker while the circle

2,0

PSfrag replacements
R
indicates SA,59 measured during the clinical intervention the
2,0

patient had to undergo.

kσkmax

kσkmean

kσkmin
si
∆ϕm
2,0

index
r
m
L
12,0

IV. D ISCUSSION
2,0

SB,272

SB,253

SB,225

The results gained by the virtual array based linear de-

SB,202
2,0

SB,199

SB,177

SB,158

SB,125L

pendency analysis in [5] lead us to the assumption that

SB,125S
2,0

SB,102

SB,81

SB,62

it is possible to design an electrode array meeting clinical

SA,62

SA,59
F

Fig. 3. The initial 62 lead layout for the L-shaped scheme SB . The frontal requirements. Such an array basically may consist of two L-
part (top) consists of 41 electrodes and the dorsal part (bottom) of 21. The
horizontal and vertical spacing was selected to be 4 cm for both parts. shaped pieces. For the application of the frontal array we
suggest to use the six standard ECG leads V1 -V6 as landmarks.
The posterior array may be aligned along the spinal cord
III. R ESULTS touching the tip of the left shoulder blade with its horizontal
part.
We limited our search for the ideal number and density The EG analysis imposed to increase the number of elec-
of electrodes to arrays derived form the initial prototype of trodes to 92 frontal and 33 electrodes dorsal. It revealed
scheme SB as depicted in fig. 3. This decision is based upon that the highest gain r at a low effort m could be achieved
the findings given in [5]. In the following we will use the by increasing the electrode density within the most sensitive
number of electrodes to distinguish the different electrode body surface regions. Such regions are indicated by low LLA
arrays developed and tested. For example the initial prototype measures as stated by [5]. Within regions exposing medium
for scheme SB consists of 62 electrodes, thus, in the following LLA measures the electrode density may be lowered. The
it will be addressed as SB,62 . best results were obtained, when shrinking the horizontal
In a first step we generated four new arrays by reducing and vertical spacing between left frontal electrodes to 2 cm,
the rectangular electrode spacing to 2 cm by subdividing the while the spacing of the dorsal electrodes was set to ∼ 3 cm.
the spacing of 4 cm proposed by [5]. By omitting electrode These measures were mainly imposed by technical limitations
positions we gained four new electrode arrays with 102, 125, such as the non infinitesimal electrode diameter, the minimum
158 and 202 electrodes. Further we added an additional row distance necessary to avoid electrode to electrode interference
and column to SB,62 while keeping the spacing of 4 cm in and the manufacturing tolerances for the electrode size and
order to gain the array SB,81 based upon 81 electrodes. The position. Finally we could increase the number of frontal
reduction of spacing and omission of electrode positions as electrodes to 92 and the dorsal to 33 for the final 125 lead
done for SB,62 resulted in a set of another five arrays. Including array SB,125S in contrast to 41 and 21 for the initial 62 lead
SA,59 measured during clinical intervention and its idealized array SB,62 .

2381
 TABLE I
C OMPARISON BETWEEN 8 ARRAYS BASED UPON SCHEMES SA AND SB FOR ALL 11 PATIENTS USED TO RECORD ECG SIGNALS RELATED TO
VENTRICULAR ACTIVITY. I NCREASING THE NUMBER OF ELECTRODES M DOES NOT IMPLY AN INCREASE IN RANK R AS R DEPENDS UPON SIZE AND
WEIGHT OF THE ENTIRE PATIENT ( GREY NUMBERS ). T INY AND LIGHT PATIENTS EXPOSE A LARGER MAXIMAL R THAN TALL AND HEAVY PATIENTS . T HE
GRAYED ROW INDICATES THE ARRAY LAYOUT WE THINK IS OPTIMAL FOR OBSERVING AND RECORDING VENTRICULAR EVENTS . I T WAS SELECTED
BASED UPPON THE EFFORT AND GAIN MEASURES WE COMPUTED .

Patient 10

Patient 11
Patient 2

Patient 4

Patient 8
Patient 1

Patient 3

Patient 5

Patient 6

Patient 7

Patient 9
Array Scheme

R 25 17 20 18 22 25 18 26 27 19 17
SA,59 R/R0 1 0.944 0.952 0.9 0.917 1 1 0.963 1 1 1
kσk[µV] 153.6 145.6 151 140 134.2 151 149.7 147 148.3 157.5 145.6
R 25 18 21 20 24 25 18 27 27 19 17
SA,62 R/R0 1 1 1 1 1 1 1 1 1 1 1
kσk[µV] 157.5 157.5 157.5 157.5 157.5 157.5 157.5 157.5 157.5 157.5 157.5
R 28 20 24 22 28 30 20 30 28 23 19
SB,62 R/R0 1.12 1.11 1.14 1.1 1.17 1.2 1.11 1.11 1.04 1.21 1.12
kσk[µV] 157.5 157.5 157.5 157.5 157.5 157.5 157.5 157.5 157.5 157.5 157.5
R 28 20 24 22 28 30 20 30 28 23 19
SB,81 R/R0 1.12 1.11 1.14 1.1 1.17 1.2 1.11 1.11 1.04 1.21 1.12
kσk[µV] 180 180 180 180 180 180 180 180 180 180 180
R 39 20 27 23 46 33 20 31 37 23 19
SB,125S R/R0 1.56 1.11 1.29 1.15 1.92 1.32 1.11 1.15 1.37 1.21 1.12
kσk[µV] 223.6 223.6 223.6 223.6 223.6 223.6 223.6 223.6 223.6 223.6 223.6
R 40 20 27 23 46 34 21 32 37 23 19
SB,158 R/R0 1.6 1.11 1.29 1.15 1.92 1.36 1.17 1.19 1.37 1.21 1.12
kσk[µV] 251.4 251.4 251.4 251.4 251.4 251.4 251.4 251.4 251.4 251.4 251.4
R 39 20 27 22 44 34 21 31 37 23 19
SB,202 R/R0 1.56 1.11 1.29 1.1 1.83 1.36 1.17 1.15 1.37 1.21 1.12
kσk[µV] 284.3 284.3 284.3 284.3 284.3 284.3 284.3 284.3 284.3 284.3 284.3
R 39 20 25 23 42 33 21 31 36 23 18
SB,272 R/R0 1.56 1.11 1.19 1.15 1.75 1.32 1.17 1.15 1.33 1.21 1.06
kσk[µV] 329.8 329.8 329.8 329.8 329.8 329.8 329.8 329.8 329.8 329.8 329.8

Taking into account the three Wilson leads used to form ACKNOWLEDGMENT
the Wilson central terminal required by the solution to the This study was supported by the Austrian Science Fund
electrocardiographic inverse problem (see [7]), one has to (FWF) under grant START Y144-N04.
record the ECG signals for 128 leads. This results in the
additional advantage that the optimal sampled array layout R EFERENCES
SB,125S does not require amplifiers and/or digitizers providing [1] R. Modre, B. Tilg, G. Fischer, F. Hanser, B. Messnarz, M. Seger,
more than 128 channels for recording the ECG signals. This M. Schocke, T. Berger, F. Hintringer, and F. X. Roithinger, “Atrial
noninvasive activation mapping of paced rhythm data,” Journal of Car-
keeps the expenses for the recording equipment low and allows diovascular Electrophysiology, vol. 14, no. 7, pp. 712–719, July 2003.
to use standard 128 channel digitizing amplifier providing a [2] C. Ramanathan, R. N. Ghanem, P. Jia, K. Ryu, and Y. Rudy,
measurement range ∆V = ±32 mV and a noise level of σl = “Noninvasive electrocardiographic imaging for cardiac electrophysiology
and arrhythmia,” Nature Medicine Advanced Online Publication,
20 µV per lead such as the Mark-8 (Biosemi V.O.F.Amsterdam, vol. 10, no. 4, pp. 422–428, Apr. 2004. [Online]. Available:
The Netherlands). Until now we used the Mark-8 in connection http://www.nature.com/naturemedicine
with the electrode arrays SA,62 and SB,62 for recording 62 lead [3] F. Greensite, “Cardiac electromagnetic imaging as an inverse problem,”
Electromagnetics, vol. 21, pp. 559–557, 2001.
plus the three Wilson leads. [4] O. Dössel and F. R. Schneider, “Opportunities and limitations of non-
invasive bioelectric imaging of the heart,” in European Medical and
Biomedical Engineering Conference, 1999, pp. 1226–1227.
[5] C. Hintermüller, G. Fischer, M. Seger, B. Pfeifer, F. Hanser, R. Modre,
According to tab. I, array layouts based upon scheme SB and B. Tilg, “Multi-lead ecg electrode array for clinical application of
electrocardiographic inverse problem,” in Proceedings of the 26th Annual
seem to be preferable over scheme SA based ones. We found International Conference of the IEEE EMBS, Sept. 2004, pp. 1941–1944.
that increasing the numbers of electrodes within sensitive [6] A. SippensGroenewegen, H. A. P. Peeters, E. R. Jessurun, A. C. Linnen-
areas increases the effective rank R of the leadfield matrix bank, E. O. R. de Medina, M. D. Lesh, and N. M. van Hemel, “Body
surface mapping during pacing at multiple sites in the human atrium,”
L. Similar results were gained from tab. I. The large inter- Circulation, vol. 1997, no. 4, pp. 369–380, Feb. 1998.
patient variations of the effective rank R and the gain r are [7] G. Fischer, B. Tilg, R. Modre, F. Hanser, B. Messnarz, and P. Wach, “On
mainly caused by the different size of the patients and thus by modeling the wilson terminal in the boundary and finite element method,”
IEEE Transactions on Biomedical Engineering, vol. 49, no. 3, Mar. 2000.
differing distance between heart and body surface.

2382