IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 14, NO.
3, MAY 2010 883
A Novel Neural-Network Model for Deriving
Standard 12-Lead ECGs From Serial Three-Lead
ECGs: Application to Self-Care
Hussein Atoui, Jocelyne Fayn, and Paul Rubel, Member, IEEE
Abstract—Synthesis of the 12-lead ECG has been investigated
in the past decade as a method to improve patient monitoring
in situations where the acquisition of the 12-lead ECG is cum-
bersome and time consuming. This paper presents and assesses a
novel approach for deriving 12-lead ECGs from a pseudoorthogo-
nal three-lead subset via generic and patient-specific nonlinear re-
construction methods based on the use of artificial neural-networks
(ANNs) committees. We train and test the ANN on a set of serial
ECGs from 120 cardiac inpatients from the intensive care unit
of the Cardiology Hospital of Lyon. We then assess the similar-
ity between the synthesized ECGs and the original ECGs at the
quantitative level in comparison with generic and patient-specific
multiple-regression-based methods. The ANN achieved accurate
reconstruction of the 12-lead ECGs of the study population using
both generic and patient-specific ANN transforms, showing sig-
nificant improvements over generic (p-value ≤ 0.05) and patient-
specific (p-value ≤ 0.01) multiple-linear-regression-based models.
Consequently, our neural-network-based approach has proven to
be sufficiently accurate to be deployed in home care as well as in
ambulatory situations to synthesize a standard 12-lead ECG from
a reduced lead-set ECG recording.
Index Terms—Acute ischemia, arrhythmia, ECG, eHealth,
embedded computing, neural networks, self-care.
I. INTRODUCTION
ECENT years have witnessed a growing interest for devel-
R oping personalized and nonhospital-based care systems to
improve the management of cardiac care, to reduce the time be-
fore treatment, and consequently, to reduce cardiac morbidity
and mortality [1]–[3]. The reason behind such interest is due
to the fact that in western countries, cardiovascular diseases are
a leading cause of early disability and premature death. Fur-
thermore, because the population is rapidly aging, the number
of cardiac patients is steadily increasing and almost 2/3 of the
cardiac patients die in the prehospital phase. Therefore, greater
deployment of resources for prehospital care is needed to reduce
the fatality rate.
The only easy to use diagnosis tool useful for assessing the
probability of cardiac events in home care, self-care, ambulatory,
or emergency recording conditions is the ECG. However, in such
Manuscript received November 3, 2006; revised March 15, 2007 and May 30,
2007. First published April 8, 2010; current version published June 3, 2010.
The authors are with the Department of Methodologies of Information Pro-
cessing in Cardiology, Institut National des Sciences Appliquées (INSA-Lyon),
Institut National de la Santé et de la Recherche Médicale (INSERM), Uni-
versité Lyon 1, Bron, F-69677, France (e-mail: hussein_atoui@hotmail.com;
jocelyne.fayn@insa-lyon.fr; paul.rubel@insa-lyon.fr).
Digital Object Identifier 10.1109/TITB.2010.2047754
1089-7771/$26.00 © 2010 IEEE
situations, recording a standard 12-lead ECG, the only ECG
representation that cardiologists have been trained to accurately
analyze and interpret, is often difficult and impractical. It would
thus be valuable to design a minimal, easy to use lead set from
which the 12-lead ECG can be accurately reconstructed [4].
The solution adopted by the enhanced personal, intelligent
and mobile system for early detection and interpretation of cardi-
ological syndromes (EPI-MEDICS) European project was to de-
sign an intelligent, portable Personal ECG Monitor (PEM), em-
bedding advanced decision making, which is capable to record
a simplified but professional quality three-lead ECG, to derive
a standard 12-lead ECG to detect arrhythmias and ischemia or
acute infarction, and to send an alarm message with a copy of
the electronic health record (EHR) and the concerned ECG to
the appropriate health care providers [3].
To achieve this goal, the EPI-MEDICS project determined an
easy-to-use pseudoorthogonal set of ECG leads, based on four
active electrodes located in the Mason Likar positions and in V 2,
respectively, which have proven to be adequate for home care
or ambulatory ECG recording [5]. The project also determined
a generic linear transformation matrix used to derive the five
missing leads (V 1, V 3, V 4, V 5, and V 6) of the standard 12-
lead ECG from the recorded three-lead PEM ECG. The overall
usability and diagnostic accuracy of the PEM device have been
judged very high. However, it is known that the accuracy of the
standard 12-lead ECG reconstruction can be further enhanced
by using patient-specific transforms that could be stored in the
smart media card embedded in the PEM device. The challenge is
thus to design a synthesis method that is able to derive a 12-lead
ECG from the pseudoorthogonal three-lead ECG that contains
as much diagnostic information as would have been contained
in the original 12-lead ECG if recorded simultaneously with the
three-lead subset.
Generally, methods used to derive the reconstruction trans-
forms are based on multiple-linear-regression techniques
[6]–[8]. However, most of the studies using linear patient-
specific transforms have only been assessed on subsets of the
standard 12-lead ECG recordings where the training and testing
phases were performed on the same data or, in the best cases,
the validation was conducted over ECGs recorded within laps
of time within 24 h with the same electrode positions as for the
learning phase. Also, it is essential to evaluate the accuracy of
lead reconstruction over ECGs acquired using personalized and
nonhospital-based care systems [9], such as the PEM system.
In addition, even though linear transforms yield high accu-
racy when electrode positions remain in the same place [6]–[8],
884 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 14, NO. 3, MAY 2010

such methods face serious challenges for serial ECGs recorded

at a few days or a few weeks interval by the patient himself
or by his/her relatives without any professional assistance. In
such conditions, the ECG recording can display noise or serial
changes caused by possible electrode misplacement that hinder
a high-quality synthesis.
Under such circumstances of noise and uncertainty, nonlinear
methods like artificial neural networks (ANNs) are believed to
further enhance the reconstruction accuracy and compensate the
conventional linear approaches weaknesses [10]. In fact, the sci-
entific literature is full of examples where the nonlinear methods
provide an efficient alternative to conventional approaches [11].
Also, it is not given, for certain, that the electrogenesis of the
cardiac signals is a purely linear process [12], [13].
The purpose of this study is to present novel, enhanced meth-
ods for deriving 12-lead ECGs from a pseudoorthogonal three-
lead subset. We investigate how well the 12-lead ECG is re-
constructed using both generic and patient-specific nonlinear
reconstruction methods that are based on the use of ANNs.

II. MATERIAL AND METHODS
A. Study Population
We used a database of 10 s ECGs that have been recorded on
cardiac inpatients from the intensive care unit of the Cardiology
Hospital of Lyon. Each patient had one or two pairs of ECGs,
each consisting of the following:
1) a standard 10 s 12-lead ECG, hereafter called reference
ECG, recorded by using a standard digital electrocardio-
graph (Cardiette);
2) a 10 s three-lead PEM ECG based on I, II, and V 2 (PECG).
The PEM ECGs were recorded in average at less than half an
hour interval of the recording of the reference ECGs.
An extensive work of quality assessment was performed on
all ECGs of the database, in collaboration with experienced
cardiologists of the hospital. The target was to remove ECGs
where situations of electrodes inversions, misplacements, or
diagnostic changes were noticed.
The study population thus consists of a series of 157 pairs of
digital ECGs (sampling rate: 500 samples/s; resolution: 5 µV)
collected on 120 patients (82 male, 38 female; mean age ± SD
= 61 ± 15 years). Acute coronary syndrome (ACS) was diag-
nosed in 46 (29.3%) of the 157 standard 12-lead ECG tracings.
The ECGs were processed by the Lyon program [14] that ex-
tracted a 1-s-duration representative cycle from every original
10 s ECG record and determined the corresponding P, QRS, and
T onsets and offsets. Finally, the ECG signal from the represen-
tative cycle was used in the interval [P-onset − 18 ms, T-offset
+ 38 ms] for computing the reconstruction transforms.
The study population was divided into two subsets DS1 and
DS2 of 83 and 74 pairs of ECGs, respectively. Dataset DS1
consists of the 83 patients (53 male and 30 female) of the
study population who had only one pair of ECGs (21 ACS
ECGs and 62 control ECGs). Dataset DS2 consists of the re-
maining 37 patients who had two pairs recorded in average at
three (range: 1–21) days interval (25 ACS ECGs and 49 control
ECGs).
Fig. 1. Architecture of the ANNs used to synthesize the five missing V 1, V 3,
V 4, V 5, and V 6 leads (output layer) of the 12-lead ECG using a three-lead
ECG (I, II, and V 2) as input layer, h (typically 1 or 2) hidden layers and n
neurons per hidden layer.
B. Neural-Network Architecture Design and Training
To synthesize the missing (V 1, V 3, V 4, V 5, and V 6) ECG
signals from the recorded (I, II, and V 2) 12-lead ECG sub-
set, we use an ensemble of N multilayer feedforward ANNs
trained by means of a supervised back-propagation algorithm.
Each individual ANN consists of one input layer with three
input neurons (one for each recorded signal), one output layer
with five output neurons (one for each derived signal), h = 1
hidden layer and n = 15 neurons per hidden layer, as shown
in Fig. 1. The relations between the input and output layers are
expressed through the weight and biases of the neurons of the
hidden layer. These weight and biases are chosen randomly at
the beginning of the training phase and are then iteratively opti-
mized during the back-propagation process. The hidden layers
number, the activation function type, and the numbers of neu-
rons per layer were chosen on the basis of previous experiments
and preliminary results obtained in synthesizing biomedical sig-
nals by nonlinear methods [15]. As usual in approximation
tasks, a linear activation function is used for the output neu-
rons. A sigmoid transfer function was chosen for the hidden
layer.
To overcome the performance limitations inherent to the train-
ing process, such as the random selection of the initial settings
of the ANN weights and biases of the individual networks, we
adopted a solution consisting of building up ANN committees
of N = 50 individual ANN of the Fig. 1 type. The five outputs
of the ANN committees are obtained by summing up and di-
viding by N , the outputs of the N individual ANNs. It is thus
expected that, as an application of the central limit theorem, the
performance of the ANN committee surpasses each individual
ANN performance [3], [11].
ATOUI et al.: NOVEL NEURAL-NETWORK MODEL FOR DERIVING STANDARD 12-LEAD ECGs
Fig. 2. Reconstruction of a standard 12-lead ECG out of the three-lead PEM ECG. The synthesis of the five missing V 1, V 3, V 4, V 5, and V 6 leads is obtained
by averaging the outputs of a committee of 50 neural networks.
1) Generic Neural-Network Training: To train each of the
N = 50 individual, generic ANN of the committee, we adopted
a cross-validation-like strategy [16] as follows.
The 83 standard 12-lead ECGs of DS1 were randomly divided
into m = 5 subsets. Each individual network of the committee
was trained m times, each time leaving out one of the train-
ing subsets. The ECGs of the m − 1 subsets were juxtaposed
to form one continuous ECG recording. Stopping the learning
process was performed by monitoring the reconstruction error
(rms) on the remaining subset, where the ECGs were also jux-
taposed to form one continuous ECG recording. This procedure
was repeated for each of the N different networks biases and
weights initializations and for each of the m subsets. A fixed
number of iterations K1 was then determined by recording, for
each training case, the number of iterations k and by averaging
these values over the N × m trainings. Finally, each individ-
ual network of the committee was trained over the 83 12-lead
ECGs of DS1 using the fixed number of iterations K1 as an
early stopping point.
2) Patient-Specific Neural-Network Training: Ideally, we
would need three serial 12-lead ECGs per patient for deriving
a patient-specific ANN ensemble: one for training the network
(ECG-L), one for stopping the learning process in order to avoid
network overdesign (ECG-S), and one for testing the general-
ization capacity of the trained network (ECG-T).
Unfortunately, our database comprised at maximum only two
serial standard 12-lead ECGs per patient (ECGs L and T). To
overcome this difficulty, instead of stopping the learning pro-
cess after a variable number of iterations that should ideally
885
be determined by ECG-S, we used a fixed number of itera-
tions K2 that was determined in a similar fashion, as shown in
Section II-B.1, but on a different, independent serial database
previously described in [15]. The learning process was then per-
formed by training each of the N individual ANN on the first
reference ECG (ECG-L) of each of the 37 patients of DS2.
3) Neural-Network Model Reutilization: Once trained, the
generic and patient-specific ANN transforms are tested on a
different dataset than the one used to train the generic and
patient-specific committees. This test is intended to assess the
ANN models performance by simulating their use in real-life
situations and measuring the reconstruction accuracy of both
transformation models. The simulation process however sub-
stantially differs from one reconstruction method to the other.
In the generic approach, we use the second standard 12-lead
and PEM ECG pairs of dataset DS2, i.e., ECG-T and the corre-
sponding PEM ECG from DS2, as a test set. The five “missing”
leads of every standard 12-lead and PEM ECG are synthesized,
as shown in Fig. 2, by the generic ANN committee designed in
Section II-B.1, and the reconstructed V 1, V 3 to V 6 leads are
then compared with the original leads of the corresponding stan-
dard 12-lead ECG of DS2. In the patient-specific approach, a
patient-specific committee is trained on the first standard 12-lead
ECG (ECG-L) of the first ECG pair of each of the 37 patients of
DS2, according to the method described in Section II-B.2. The
patient-specific committee is then tested on each second ECG
pair of the corresponding patient by deriving the five “missing”
leads for both the Cardiette (ECG-T) and the PEM ECGs. Fi-
nally, the reconstructed V 1, V 3 to V 6 leads are compared with
886 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 14, NO. 3, MAY 2010
the original leads of the corresponding Cardiette ECG (ECG-T).
This procedure is applied in turn for each case of the validation
set DS2.
C. Multiple-Regression-Synthesis-Coefficients Calculation
To calculate a generic multiple regression transform, we con-
sidered the 83 standard 12-lead ECGs of dataset DS1. As for
the generic ANN training, these 83 ECGs were juxtaposed into
one continuous ECG recording. We then computed a global
transformation matrix that synthesizes the missing V 1, V 3 to
V 6 leads of the PEM ECG, using a multiple-linear-regression
algorithm of the type
Vi = ai0 + ai1 I + ai2 II + ai3 V 2. (1)
Patient-specific multiple-regression-based transforms were
obtained by using the same algorithm to compute an individual
transformation matrix for every first 12-lead ECG (ECG-L) of
every patient of dataset DS2.
D. Validation Database (DS2) Three-Lead and 12-Lead ECG
Signals Alignment and Coherence Control
Since we were using two different recording systems (PEM
and Cardiette) with different electrodes positioning and because
the recording of the two ECGs was not performed at the same
time, it was essential to first proceed to the Cardiette and PEM
ECG signals alignment and to control the coherence between the
two serial recordings before measuring the goodness of fit of the
reconstructed leads of the dataset used for the validation phase:
DS2. This step is thus prefiguring a real situation of routine use
of the PEM.
1) PEM and Cardiette Signals Alignment: To establish the
time shift (delay) between the PEM and Cardiette ECGs, we
computed the cross correlation between the V 2 lead signals. In
order to be less sensitive to artifacts, we only considered the
signal in the following time interval [Q-onset − 18 ms, T-offset
+ 38 ms]. The algorithm consists of the following steps.
1) Calculation of the cross-correlation between the two
signals.
2) Estimation of the time shift between both signals by lo-
calization of the maximum value in the cross-correlation.
3) Alignment of the global QRS onset of the PEM ECG
with the global QRS onset of the standard 12-lead ECG
by subtraction of the time shift measured by the cross-
correlation technique.
2) PEM and Cardiette Recordings Coherence Control: Be-
cause the PEM and the standard 12-lead ECGs have not been
recorded synchronously and because of the fast changing car-
diac status usually found in an intensive care unit, there might
be large differences between the two ECGs. The objective of
this step is to evaluate if the difference between the recorded
PEM and Cardiette ECGs is not too large, and thus, could af-
fect the synthesis quality. The coherence control is performed
in two steps: for each of the 37 patients of DS2, we conducted
a correlation analysis between leads I, II, V 1, V 2 to V 6 of the
first and second Cardiette ECGs and between leads I, II, and V 2
of the first Cardiette ECG (ECG-L) and the second PEM ECG
TABLE I
RMS (IN MICROVOLTS) AND CORRELATION COEFFICIENTS (r) BETWEEN THE
ORIGINAL I, II, AND V 2 LEADS OF THE FIRST AND THE SECOND STANDARD
12-LEAD ECGS RECORDINGS FOR EVERY PATIENT OF DATASET DS2 (N = 35)
(PECG-T) associated with the second Cardiette ECG (ECG-T).
Then, we conducted another correlation analysis between leads
I, II, and V 2 of the second Cardiette ECG (ECG-T) and the
second PEM ECG (PECG-T) of dataset DS2.
E. Generalization Capability Verification and Validation
To assess the quality of the reconstruction of the generic and
patient-specific ANN-based transforms, we measured the lead-
by-lead signal differences between the derived and the original
ECGs. The overall waveform similarity between the original
and the reconstructed ECGs was assessed as follows.
1) Low-pass filtering of the original and reconstructed ECG
signals by means of a 21-term, 40 Hz cutpoint, moving
average filter.
2) Calculation of the rms and the correlation coefficients be-
tween the QT signal segments of the original and the
reconstructed ECG leads of DS2 for both the generic
and patient-specific ANN and multiple-linear-regression
models.
3) Exclusion for each of the four reconstruction methods,
and for every analyzed lead of 5% (two cases) of the rms
and correlation measurements, which have, respectively,
the highest rms or the lowest correlation coefficient be-
tween the original and the reconstructed ECGs. This step
is intended to overcome any bias generated by outliers in
the calculation of the mean and standard deviation [17].
Statistical analysis of the differences between the rms of the
ANN and linear-regression-based reconstruction methods was
performed by means of a paired t-test and a trimmed–paired
t-test. The trimmed t-test is used to overcome the t-test sensi-
tivity to outliers. Both tests are performed with the BioMeDical
statistical software package (BMDP).
III. RESULTS
A. Original ECG-Recordings Coherence Assessment
Table I displays the rms values and the correlation coefficients
between the first and second Cardiette ECGs of DS2, after re-
moval of the 5% extreme values (N = 35). The average rms
signal difference between the standard I, II, V 1, V 2, V 3, V 4,
V 5, and V 6 leads in the time interval [Q-onset − 18 ms, T-offset
+ 38 ms] is 107 µV. The average correlation coefficient is 0.92.
ATOUI et al.: NOVEL NEURAL-NETWORK MODEL FOR DERIVING STANDARD 12-LEAD ECGs 887

TABLE II
RMS (IN MICROVOLTS) AND CORRELATION COEFFICIENTS (r) BETWEEN THE
ORIGINAL I, II, AND V 2 LEADS OF THE FIRST CARDIETTE ECG AND THE
SECOND PEM ECG RECORDINGS FOR DATASET DS2 (N = 35)
TABLE IV
RMS (IN MICROVOLTS) AND CORRELATION COEFFICIENTS (r) BETWEEN THE
ORIGINAL AND THE RECONSTRUCTED V 1, V 3 TO V 6 LEADS FOR THE
GENERIC ANN COMMITTEE (ANN-G) AND REGRESSION-BASED
MODELS (REG-G) FOR DATASET DS2 (N = 35)

TABLE III
RMS (IN MICROVOLTS) AND CORRELATION COEFFICIENTS (r) BETWEEN THE
ORIGINAL I, II, AND V 2 LEADS OF THE SECOND CARDIETTE ECG AND THE
CORRESPONDING PEM ECG RECORDINGS FOR DATASET DS2 (N = 35)

TABLE V
RMS (IN MICROVOLTS) AND CORRELATION COEFFICIENTS (r) BETWEEN THE
ORIGINAL AND THE RECONSTRUCTED V 1, V 3 TO V 6 LEADS FOR THE
PATIENT-SPECIFIC ANN COMMITTEE (ANN-S) AND REGRESSION-BASED
MODELS (REG-S) FOR DATASET DS2 (N = 35)

Table II displays the rms values and the correlation coeffi-

cients between the first Cardiette ECG and the second PEM
ECG recordings of DS2. The average rms signal difference be-
tween the I, II, and V 2 leads is 107 µV. The average correlation
coefficient is 0.87.
Table III displays the rms values and the correlation coeffi-
cients between the second Cardiette ECG and the corresponding
PEM ECG for every patient of DS2. The average rms signal dif-
ference between the I, II, and V 2 leads is 80 µV. The average
correlation coefficient is 0.93.

B. Derived Versus Original ECG Signals

Differences Assessment
1) Precordial ECG Leads Derived From the (I, II, and V 2)
Subset of the Standard 12-Lead ECG: In this section, we present
the results obtained by deriving for all cases of DS2 the V 1, V 3
to V 6 leads from the I, II, and V 2 leads of the standard 12-lead
ECG-T by applying in turn the generic ANN-G and REG-G and
the patient-specific ANN-S and REG-S transforms constructed
according to the methods described in Section II.
Table IV displays the rms values and the correlation coeffi-
cients for the generic ANN committee and multiple-regression-
based models after removal of the two highest rms and the two
lowest correlation coefficients. It shows a slight superiority of
the ANN committee, with an average rms value of 122 µV
between the original and the reconstructed leads versus an aver-
age rms of 126 µV for the regression-based synthesis method.
However, the correlation coefficients between the original and
the reconstructed leads score almost the same values for the
ANN committee and the regression-based approaches, with an
average value of 0.93.
Table V displays the rms values and the correlation coef-
ficients for the patient-specific ANN committee and multiple-
regression-based models. In contrary to the generic approach, it
shows a more pronounced superiority of the ANN committee,
TABLE VI
P -VALUES OF THE PAIRWISE COMPARISONS OF THE RMS VALUES BETWEEN
THE ORIGINAL AND THE RECONSTRUCTED V 1, V 3 TO V 6 LEADS FOR THE
ANN AND THE REGRESSION-BASED MODELS (N = 35)
with an average QT interval rms of 94 µV between the original
and the reconstructed leads versus an average rms of 108 µV
for the regression-based synthesis method. The same yields for
the correlation coefficients between the original and the recon-
structed leads, which score 0.961 in average for the Q-onset to
T-offset segment for the patient-specific ANN committee versus
0.948 in average for the patient-specific regression model. The
median correlations are 0.975 and 0.967, respectively.
Table VI summarizes the results of the paired and trimmed–
paired t-tests. In the patient-specific approach, the difference
in accuracy, in terms of rms values, between the ANN and
the regression-based methods is statistically very significant
888 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 14, NO. 3, MAY 2010
TABLE VII
RMS (IN MICROVOLTS) AND CORRELATION COEFFICIENTS (r) BETWEEN THE
ORIGINAL CARDIETTE AND THE RECONSTRUCTED PEM V 1, V 3 TO V 6 LEADS
FOR THE GENERIC ANN COMMITTEE (ANN-G) AND REGRESSION-BASED
MODELS (REG-G) FOR DATASET DS2 (N = 35)
TABLE VIII
RMS (IN MICROVOLTS) AND CORRELATION COEFFICIENTS (r) BETWEEN THE
ORIGINAL CARDIETTE AND THE RECONSTRUCTED PEM V 1, V 3 TO V 6
LEADS FOR THE PATIENT-SPECIFIC ANN COMMITTEE (ANN-S) AND
REGRESSION-BASED MODELS (REG-S) FOR DATASET DS2 (N = 35)
(p-value ≤ 10−6 ). However, for the generic approach, the dif-
ferences are not significant (p-value = 0.066).
2) Precordial Leads Derived From the (I, II, and V 2) PEM
Leads: In this section, we present the same type of results, as
in Section II-B.1, but taking the PEM-T leads instead of the
(I, II, and V 2) subset of the ECG-T as input to the four
transforms.
Table VII displays the rms values and the correlation coeffi-
cients for the generic ANN committee and multiple-regression-
based models. It shows a slight superiority of the ANN commit-
tee only for the reconstruction of lead V 1 (p-value ≤ 0.05, paired
t-test). The average rms values between the original and the re-
constructed leads for the ANN versus the regression-based syn-
thesis method are 131 and 134 µV, respectively. The correlation
coefficients between the original and the reconstructed leads are
globally higher for the ANN committee than for the regression-
based approaches: average values are 0.917 and 0.901, respec-
tively. But, the median correlations are almost the same (0.945
versus 0.944).
Table VIII displays the rms values and the correlation coef-
ficients for the patient-specific ANN committee and multiple-
regression-based models. In contrary to the generic approach, it
shows a fair superiority of the ANN committee, with an average
TABLE IX
P -VALUES OF THE PAIRWISE COMPARISONS OF THE RMS VALUES BETWEEN
THE ORIGINAL CARDIETTE AND THE RECONSTRUCTED PEM V 1, V 3 TO V 6
LEADS FOR THE ANN AND THE REGRESSION-BASED MODELS (N = 35)
rms value of 118 µV between the original and the reconstructed
leads versus an average rms of 126 µV for the regression-based
synthesis method. The same yields for the correlation coeffi-
cients between the original and the reconstructed leads, which
score 0.932 in average for the patient-specific ANN commit-
tee, versus 0.920 for the patient-specific regression model. The
median correlations are 0.955 and 0.947, respectively.
Table IX summarizes the results of the paired and trimmed–
paired t-tests. In the patient-specific approach, the difference
in accuracy, in terms of rms values, between the ANN and the
regression-based methods is statistically significant (p-value ≤
0.05). However, in the generic approach, the difference is not
significant.
IV. DISCUSSION
In this paper, we designed and assessed the efficacy of
a nonlinear neural-network approach for missing ECG leads
synthesis.
A. Comparison to Existing Reconstruction Routines
The waveform similarity obtained in this study is better than in
prior publications using linear transforms [6]–[8]. The median
correlations for our ANN patient-specific approach and for a
patient-specific multiple-linear-regression model assessed over
ECGs recorded within 24 h are 0.975 and 0.953, respectively
[6]. In addition, when considering the fact that in these prior
investigations, training and testing were conducted over ECGs
recorded without changing the electrode positioning and without
introducing, as in our case, phase shifts due to the asynchronism
existing between the sampling clocks of the PEM and Cardiette
ECG recording devices, the difference becomes more obvious.
In our approach the standard and synthesized ECG leads are
similar enough, taking into consideration the small number of
leads that were used for the synthesis, and the fact that the
ECGs were recorded in challenging evolving clinical situations:
the standard 12-lead ECGs and the PEM three-lead ECGs were
not recorded simultaneously and were collected by two different
technicians in an intensive care unit of a cardiology hospital. In
addition, the PEM ECGs were recorded using a specific set of
leads based on the Mason Likar positions and V 2, whereas the
standard 12-lead ECG was recorded using the standard limb
leads positions. This has lead to a mean rms difference of 76 µV
between the original I and II leads of both ECGs and an average
correlation of 0.90 (see Table III), figures that are slightly higher
than the day to day variability of the standard 12-lead ECG (see
Table I).
ATOUI et al.: NOVEL NEURAL-NETWORK MODEL FOR DERIVING STANDARD 12-LEAD ECGs
B. Generic Versus Patient-Specific Reconstruction
Patient-specific transforms have proven to be more accurate.
Nevertheless, generic reconstructions remain particularly attrac-
tive in situations, where no standard 12-lead ECG could be
recorded and/or no patient-specific transform could be com-
puted and/or retrieved, prior to the recording of a reduced ECG
lead set. Also, generic reconstructions are less sensitive to arti-
facts and electrode displacements, since their coefficients were
computed from a very large set of ECGs of different origins.
For example, during the EPI-MEDICS project evaluation, we
found a few PEM ECG tracings in which lead V 2 could dis-
play up to 50% changes of the QRS peak-to-peak amplitude
between two recordings performed at a few days interval. In
these cases, the generic transform performed better because the
generic reconstruction of V 1, V 3, and V 4 was less influenced
by the amplitude of V 2 than was the patient-specific transform.
To optimize the ECG synthesis process, it would thus be valu-
able to implement an algorithm that first checks the quality of
the recorded ECG signals for artifacts or significant amplitude
changes before applying the patient-specific transform and, in
case of large differences, either request the user, whenever pos-
sible, to check the electrode positions, or switch to the less
sensitive generic 12-lead ECG synthesis method.
C. Study Limitations
In this study, we could only partly assess the performance of
the patient-specific ANN synthesis method, since we lack a third
serial ECG recording for stopping the learning process, and we
had to use a fixed number of iterations to avoid overfitting that
could limit the generalization capabilities.
V. CONCLUSION
This paper investigates the accuracy of the reconstruction of
the 12-lead ECG from a reduced lead set using generic and
patient-specific ANNs. The results suggest that ANN represent
a rather interesting and very promising approach to improve the
current 12-lead ECG synthesis method based on linear trans-
forms, especially for the patient-specific approach. The differ-
ences between the original and the reconstructed ECGs were
mainly ascribable to electrode displacements or to hour-to-hour
changes in the ECG and not to the synthesis method. ECG cases
in which the rms were high or the correlations were low were an-
alyzed by an experienced cardiologist. His conclusion was that
the differences between synthesized and original leads are diag-
nostically irrelevant in the majority of cases. The patient-specific
transform could be further improved if two serial standard 12-
lead ECGs are available for the learning process. We also believe
that the generic ANN synthesis method can be further optimized
by using a larger and more representative training database than
the one used in this study and should clearly supersede the linear
transform.
REFERENCES
[1] F. Axisa, P. M. Schmitt, C. Gehin, G. Delhomme, E. McAdams, and
A. Dittmar, “Flexible technologies and smart clothing for citizen medicine,
889
home healthcare, and disease prevention,” IEEE Trans. Inf. Technol.
Biomed., vol. 9, no. 3, pp. 325–336, Sep. 2005.
[2] P. Campbell, J. Patterson, D. Cromer, K. Wall, G. L. Adams, A. Albano,
C. Corey, P. Fox, J. Gardner, B. Hawthorne, J. Lipton, M. Sejersten,
A. Thompson, A. Thompson, S. Wilfong, C. Maynard, and G. Wagner,
“Prehospital triage of acute myocardial infarction: Wireless transmission
of electrocardiograms to the on-call cardiologist via a handheld computer,”
J. Electrocardiol., vol. 38, no. 4, pp. 300–309, 2005.
[3] P. Rubel, J. Fayn, L. Simon-Chautemps, H. Atoui, M. Ohlsson, D. Telisson,
S. Adami, S. Arod, M. C. Forlini, C. Malossi, J. Placide, G. L. Ziliani,
D. Assanelli, and P. Chevalier, “New paradigms in telemedicine: Ambient
intelligence, wearable, pervasive and personalized,” Stud. Health Technol.
Inf., vol. 108, pp. 123–132, 2004.
[4] B. J. Drew, M. M. Pelter, S. F. Wung, M. G. Adams, C. Taylor, G. T. Evans
Jr, and E. Foster, “Accuracy of the EASI 12-lead electrocardiogram com-
pared to the standard 12-lead electrocardiogram for diagnosing multiple
cardiac abnormalities,” J. Electrocardiol., vol. 32, pp. 38–47, 1999.
[5] P. Rubel, J. Fayn, G. Nollo, D. Assanelli, B. Li, L. Restier, S. Adami,
S. Arod, H. Atoui, M. Ohlsson, L. Simon-Chautemps, D. Télisson, C.
Malossi, G. L. Ziliani, A. Galassi, L. Edenbrandt, and P. Chevalier, “To-
ward personal eHealth in cardiology. Results from the EPI-MEDICS
telemedicine project,” J. Electrocardiol., vol. 38, no. 4 (suppl.), pp. 100–
106, 2005.
[6] S. P. Nelwan, “Evaluation of 12-lead ECG reconstruction methods for
patient monitoring,” Ph.D. dissertation, Erasmus Universiteit Rotterdam,
Rotterdam, The Netherlands, 2005.
[7] J. A. Scherer, J. M. Jenkins, and J. M. Nicklas, “Synthesis of the 12-Lead
electrocardiogram from a 3-lead subset using patient-specific transforma-
tion vectors. An algorithmic approach to computerized signal synthesis,”
J. Electrocardiol., vol. 22, pp. 128–136, 1989.
[8] S. P. Nelwan, J. A. Kors, S. H. Meij, J. H. van Bemmel, and M. L. Simoons,
“Reconstruction of the 12-lead ECG using reduced lead sets for patient
monitoring,” in Einthoven 2002: 100 years of Electrocardiography. M.
J. Schalij, M. J. Janse, A. van Oosterom, H. J. J. Wellens, E. E. van der
Wall, Eds. Leiden, the Netherlands: The Einthoven Foundation, 2002, pp.
545–551.
[9] L. Hadzievski, B. Bojović, V. Vukcević, P. Belicev, S. Pavlović, Z. Vasil-
jević-Pokrajcić, and M. Ostojić, “A novel mobile transtelephonic system
with synthesized 12-lead ECG,” IEEE Trans. Inf. Technol. Biomed., vol. 8,
no. 4, pp. 428–438, Dec. 2004.
[10] R. P. W. Duin, “Learned from neural networks,” in Proc. 6th Annu. Conf.
Adv. School Comput. Imag. (ASCI 2000), pp. 9–13.
[11] S. Mitra, S. K. Pal, and P. Mitra, “Data mining in soft computing frame-
work: A survey,” IEEE Trans. Neural Netw., vol. 13, no. 1, pp. 3–14, Jan.
2001.
[12] R. Modre, M. Seger, G. Fischer, C. Hintermuller, D. Hayn, B. Pfeifer,
F. Hanser, G. Schreier, and B. Tilg, “Cardiac anisotropy: Is it negligible
regarding noninvasive activation time imaging?” IEEE Trans. Biomed.
Eng., vol. 53, no. 4, pp. 569–580, Apr. 2006.
[13] R. Gulrajani, “The forward and inverse problems of electrocardiography,”
IEEE Eng. Med. Biol. Mag., vol. 17, no. 5, pp. 84–101, Sep./Oct. 1998.
[14] P. Arnaud, P. Rubel, D. Morlet, J. Fayn, and M. C. Forlini, “Methodology
of ECG interpretation in the lyon program,” Methods Inf. Med., vol. 29,
no. 4, pp. 393–402, 1990.
[15] H. Atoui, J. Fayn, and P. Rubel, “A neural network approach for patient-
specific 12-lead ECG synthesis in patient monitoring environments,” in
Computers in Cardiology, vol. 31, A. Murray, Ed., 2004, pp. 161–164.
[16] B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap. London,
U.K.: Chapman & Hall, 1993.
[17] “Recommendations for measurement standards in quantitative electrocar-
diography. The CSE Working Party,” Eur. Heart J., vol. 6, pp. 815–825,
1985.
Hussein Atoui received the B.S. degree in computer
science from Lebanese University, Beirut, Lebanon,
in 2001, and the M.S. and Ph.D. degrees in in-
formatics from the Institut National des Sciences
Appliquées de Lyon (INSA-Lyon), Lyon, France, in
2002 and 2006, respectively.
From 2002 to 2008, he has been with the
Méthodologies de Traitement de l’Information en
Cardiologie (MTIC) Research Laboratory of INSA-
Lyon and Université Lyon 1, Lyon. He is currently Re-
search Engineer in an IT solutions provider company.
890 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 14, NO. 3, MAY 2010
Jocelyne Fayn received the M.S. degree in mathe-
matics, specialty statistics and informatics, from the
Université Claude Bernard Lyon 1, Lyon, France,
and the Ph.D. degree in computer sciences from the
Institut National des Sciences Appliquées (INSA-
Lyon), Lyon. She is currently Research Engineer in
the Institut National de la Santé et de la Recherche
Médicale (INSERM) and Assistant-Director of the
Méthodologies de Traitement de l’Information en
Cardiologie (MTIC) Research Laboratory of INSA-
Lyon and Université Lyon 1, Lyon. She is also the
Coordinator of the study unit Medical Informatics and Communication Tech-
nologies of the Master Biomedical Research, University of Lyon. She was en-
gaged in three national French projects and eight UE-funded research projects.
She is an International Editor of the Journal of Electrocardiology and of the
International Journal of Telemedicine and Applications. She was a Reviewer of
a number of international journals. Her research interests include information
systems, pervasive computing, ambient intelligence, information and communi-
cation technologies, medical informatics, quantitative electrocardiology, serial
ECG analysis, and medical decision-making. She has authored or coauthored
more than 100 publications.
Dr. Fayn was the Chair of the 32nd IEEE Annual Conference on Com-
puters in Cardiology. She was an Invited Expert of European Committee for
Standardization (CEN)/Technical Committee 251/Working Group (WG) 4 on
Health Informatics Technology for interoperability and a member of the board
of the WG15 working group Computers in Cardiology of the European Society
of Cardiology for several years. She is a member of program committees of
several international conferences and workshops.
Paul Rubel (M’91) received the degree in electron-
ics engineering from the Institut National des Sci-
ences Appliquées (INSA), Lyon, France, in 1966,
and the Ph.D. degree in electronics from the Univer-
sité Claude Bernard Lyon 1, Lyon, in 1970.
He is currently a Professor of informatics in the
Department of Informatics, INSA de Lyon, and has
been an Advisor of circa 20 Ph.D. theses in medical
informatics at INSA and the University of Lyon 1.
From 1986 to 2000, he was the Head of the Research
Unit U121 on the “Electrical Activity of the Heart”
of the National Institute of Health and Medical Research (INSERM), and he is
currently the Head of research team MTIC on “Information Processing Method-
ologies in Cardiology” of INSA-Lyon and Université Lyon 1, Lyon. He has peer
reviewed several international journals and conferences. He has authored or
coauthored more than 220 publications.
Prof. Rubel was the Project Coordinator of the European Advanced Informat-
ics in Medicine and Information Society Technology projects: Open European
Data Interchange and Processing for Electrocardiography and Enhanced Per-
sonal, Intelligent, and Mobile system for Early Detection and Interpretation of
Cardiac Syndromes, and was the local Project Manager of AIM Project standard
communications protocol-computer-assisted electrocardiography (SCP-ECG),
and of five additional AIM and Health Telematics Applications Projects. He has
also been a member of the Steering Committee of the European concerted ac-
tion “Common Standards for quantitative Electrocardiography” (CSE) (DGXII,
1978–1990) for several years and is currently responsible for the world-wide
distribution of the CSE databases and for the quality assessment of ECG in-
terpretation programs since 1995. He has also been an active Expert of Euro-
pean Committee for Standardization (CEN)/Technical Committee 251/Working
Group 4 on Health Informatics Technology for interoperability, and was one of
the three Core-Team experts, which were responsible for the development of
the so called SCP-ECG CEN EN 1064 European Norm.