658 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO.

4, JULY 2012

Cross-Database Evaluation of a Multilead

Heartbeat Classifier
Mariano Llamedo, Antoun Khawaja, and Juan Pablo Martı́nez

Abstract—In this paper,we studied the improvement in heart- for the study of arrhythmias, and could represent a challenge for
beat classification achieved by including information from multi- cardiologists when the analysis is not restricted to short-term
lead ECG recordings in a previously developed and validated clas- recordings. Detection and diagnosis of infrequent or subtle ar-
sification model. This model includes features from the RR interval
series and morphology descriptors for each lead calculated from rhythmias require careful inspection of long-term recordings.
the wavelet transform. The experiments were carried out in the IN- Note that this type of arrhythmias may represent a long-term
CART database, available in Physionet, and the generalization was threat without a proper treatment. Therefore, the aid in the anal-
corroborated in private and public databases. In all databases, the ysis provided by automatic algorithms allows cardiologists to
AAMI recommendations for class labeling and results presentation improve their diagnostics.
were followed. Different strategies to integrate the additional infor-
mation available in the 12-leads were studied. The best performing Many algorithms for ECG heartbeat classification were de-
strategy consisted in performing principal component analysis to veloped in the last decades (see [3] and [4]) using the available
the wavelet transform of the available ECG leads. The performance two-lead databases. Few of these works used the joint infor-
indices obtained for normal beats were sensitivity (S) 98%, pos- mation present in both leads, as the vectocardiogram maximal
itive predictive value (P + ) 93%; for supraventricular beats, (S) vector (VCGM ) and angle (VCGφ ) used in [5]. Another multi-
86%, (P + ) 91%; and for ventricular beats (S) 90%, (P + ) 90%.
The generalization capability of the chosen strategy was confirmed lead strategy can be seen in [3], where a final decision is taken
by applying the classifier to other databases with different number from several posterior probabilities calculated from single-lead
of leads with comparable results. In conclusion, the performance features. This last approach is not practical for multilead clas-
of the reference two-lead classifier was improved by taking into sification because of the need of a different model designed for
account additional information from the 12-leads. each set of leads, and the consequent growth in features dimen-
Index Terms—Cross-database, heartbeat classification, linear sionality. Some methodological key points in the development
classifier, multilead, principal component analysis (PCA), wavelet of heartbeat classifiers allowed results comparison of several
transform (WT). works [3], [4], [6]. Probably, the most relevant aspects were the
I. INTRODUCTION use of public and standard databases, the fulfillment of Associ-
ation for the Advancement of Medical Instrumentation (AAMI)
ARDIOVASCULAR diseases are currently the biggest
C single cause of death in developed countries according
to their public health agencies [1]. The analysis of the ECG
recommendations [7], the patient-oriented data division [3], and
the generalization capability (at least to the available public
databases) of the classifier [4], [8]. The room for improvement
signal is a noninvasive, inexpensive, and well-established tech-
in the field of heartbeat classification, together with the avail-
nique to analyze the heart function. One important aspect that
ability of 3- and 12-lead Holter devices, makes necessary the
cardiologists must study in the ECG are arrhythmias, which
development of algorithms capable of exploiting the increase
are understood as any disturbance in the rate, regularity, site
of recorded information. Recently, moreover, the St. Petersburg
of origin, or conduction of the electrical impulses through the
Institute of Cardiological Technics 12-lead arrhythmia database
heart [2]. The classification of heartbeats is an important task
(INCART) has become freely available on Physionet [9], mak-
Manuscript received June 30, 2011; revised October 21, 2011; accepted ing possible the evaluation of multilead heartbeat classifiers in
March 29, 2012. Date of publication April 18, 2012; date of current version July a comparable way.
5, 2012. This work was supported by the Spanish Ministry of Science and In- The objective of this paper is to find an effective way of
novation under project TEC2010-21703-C03-02 and the Diputación General de
Aragón through Grupos Consolidados GTC T-30 (Spain). The Centro de Inves- accounting for morphologic information present in multilead
tigación Biomédica en Red of Bioengineering, Biomaterials and Nanomedicine ECG signals. For that purpose, we compare several multilead
is an initiative of Instituto de Salud Carlos III. classification strategies against the reference two-lead classifier
M. Llamedo is with the Electronic Department, National Technological Uni-
versity, Medrano 951, C1179AAQ, Buenos Aires, Argentina, and also with that we developed in [4]. We assess the improvement in clas-
the Communications Technology Group, Aragón Institute of Engineering Re- sification performance as well as the generalization capability
search, IIS Aragón, University of Zaragoza. Marı́a de Luna, 1, 50018 Zaragoza, to other databases not considered during the development. The
Spain, and also with the Centro de Investigación Biomédica en Red de Bioinge-
nierı́a, Biomateriales y Nanomedicina (CIBER-BBN), Zaragoza, Spain (e-mail: main novelty presented in this paper is the generalization of the
llamedom@electron.frba.utn.edu.ar). model developed in [4] to an arbitrary number of leads.
A Khawaja is with Biosigna GmbH, 80337 Munich, Germany (e-mail:
khawaja.antoun@biosigna.de).
J. P. Martı́nez is with the Communications Technology Group, Aragón In- II. METHODS
stitute of Engineering Research, IIS Aragón, University of Zaragoza. Marı́a
de Luna, 1, 50018 Zaragoza, Spain, and also with the Centro de Investigación A. ECG Databases
Biomédica en Red de Bioingenierı́a, Biomateriales y Nanomedicina (CIBER-
BBN), Zaragoza, Spain (e-mail: jpmart@unizar.es).
In this study, we used the well-known MITBIH arrhythmia
Digital Object Identifier 10.1109/TITB.2012.2193408 database (MITBIH-AR) [10] and other public and private
1089-7771/$31.00 © 2012 IEEE
LLAMEDO et al.: CROSS-DATABASE EVALUATION OF A MULTILEAD HEARTBEAT CLASSIFIER 659

TABLE I 4) Biosigna Database: This is a private database developed

DATABASES USED IN THIS STUDY
at Biosigna GmbH, Munich, Germany, which consists of 56
recordings containing a broad set of pathologies. Each recording
is 1-h length, sampled at 500 Hz with an amplitude resolution
of 410 increments per millivolt using a 12-bit ADC, allowing a
range of 10 mV approximately. The recordings were manually
annotated by experienced annotators. More detailed information
about this database can be found in [12].

B. Signal Processing
databases which are briefly described as follows. All pub-
The recordings of all databases were first resampled to 360
lic databases are available on Physionet [9]. For all databases,
Hz, after filtering with a tenth-order low-pass finite impulse re-
the AAMI recommendations for class labeling were adopted
sponse filter, without observing any notable distortion (resample
(see [7], Sec. 4.2]). The AAMI Q class (unclassified and paced
function, Signal Processing Toolbox of MATLAB, The Math-
heartbeats) was discarded since it is marginally represented. This
works Inc., Natick, MA). After this process, all databases were
limitation occurs to a lesser extent with the fusion (F) AAMI
filtered to remove artifacts as described in [3].
class, but instead of discarding the heartbeats of this class, a
1) Wavelet Transform (WT): The features that describes
class-labeling modification to the AAMI recommendation is
morphology (explained in following sections) are calculated
proposed here and was adopted. It consists in merging fusion
from the WT of the ECG signal. The WT is defined for a con-
(of normal and ventricular beats) and ventricular classes, as the
tinuous signal s(t) as
same ventricular class (V’). We will refer to this modification as  
 +∞
AAMI2 labeling. The class distribution for the databases used 1 t−b
Ws s(b) = √ s(t)ψ dt, s > 0. (1)
is shown in Table I. s −∞ s
1) MITBIH Arrhythmia Database (MITBIH-AR): The
This transformation maps the ECG signal into a time-scale
database consists of 48 two-lead recordings of approximately
plane (understanding scale as a surrogate of frequency). The
30 min and sampled at 360 Hz. The first 23 recordings were
responsible of the mapping is the prototype WT ψ(t), affected
extracted from routine ambulatory practice, while the remain-
by both scaling (s) and translation (b) parameters. The WT al-
ing 25 were selected because of the presence of less common
lows location of details or fast transitions when scale parameter
complex ventricular, junctional, and supraventricular arrhyth-
s is small, and coarser aspects or trends for higher values. The
mias. The two recorded leads are not the same in all recordings,
translation parameter b indicates the location of these finer or
depending on the arrhythmia and physical limitation of the sub-
coarser details, resulting in a time-scale plane with the same
ject’s body. The four recordings with paced beats were discarded
sampling rate at each scale (Algorithme à trous). We used the
in this study in accordance with AAMI [7].
derivative of a smoothing function (quadratic spline) as the pro-
2) MITBIH Supraventricular Arrhythmia Database
totype wavelet ψ(t), so the discrete wavelet transform (DWT)
(MITBIH-SUP): The database consists of 78 two-lead
components correspond to smoothed (or low pass) derivatives
recordings of approximately 30 min and sampled at 128 Hz.
of the ECG at different scales, implemented as a digital filter
The recordings were chosen to supplement the examples
bank. As a result, the DWT retains at certain scales the useful
of supraventricular arrhythmias in the MITBIH arrhythmia
information present in the ECG in form of absolute maxima and
database. The beat type annotations of the recordings were first
zero crossings. In this study, we only used scale 4 of the WT
automatically performed, by the Marquette Electronics 8000
since this scale (at a sampling frequency of 360 Hz) retains with
Holter scanner and later reviewed and corrected by a medical
good signal-to-noise ratio (SNR) the morphological features of
student [11]. The original labeling was also adapted to the
the QRS complex that we desire to model. For background and
AAMI recommendations and to the AAMI2 modification.
implementation details, the interested reader is referred to [13]
3) St. Petersburg Institute of Cardiological Technics (IN-
for a more detailed description of the WT and its implementation
CART) 12-Lead Arrhythmia Database: This database consists
for ECG delineation.
of 75 beat type annotated recordings extracted from 32 Holter
records. Each record is 30 min long and contains 12 standard
C. Heartbeats Classification: Classifier and Features
leads, each sampled at 257 Hz. The beat annotations were pro-
duced by an automatic algorithm and then corrected manually, We follow the results obtained in [4] where we developed
containing over 175 000 beat annotations in all. The original a heartbeat classifier with good generalization capability, us-
records were collected from patients undergoing tests for coro- ing rhythm and morphological features together with a linear
nary artery disease (17 men and 15 women, aged 18–80; mean classifier (compensated for the class imbalance).
age: 58). None of the patients had pacemakers; most had ven- 1) Classifier: Regarding to the classifier used, we found that
tricular ectopic beats. In selecting records to be included in the linear discriminant functions were suitable for the heartbeat
database, preference was given to subjects with ECG’s consis- classification task in terms of performance and generalization
tent with ischemia, coronary artery disease, conduction abnor- capability. Under the assumption of independent and normally
malities, and arrhythmias [9]. distributed data, the maximum a posteriori criterion leads to the
660 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 4, JULY 2012

TABLE II
FEATURES USED IN THE MODEL OBTAINED IN [4] ONLY FOR
TWO-LEAD RECORDINGS

linear classifier defined by the discriminant functions [14]

−1 1 T −1
gi (x) = μTi Σ x − μi Σ μi + log(P (ωi )) (2)
2
for the ith class, where x represents the feature vector describing
each heartbeat, μi is the mean vector, Σ is the covariance matrix,
and P (ωi ) is the prior probability.
The values of μi and Σ in (2) were computed from the training
data with the sample mean

1 i
M
μi = xm (3)
Mi m =1
and weighted covariance matrix expressions
C M i
m =1 (xm − μi ).(xm − μi )
T
i=1 wi
Σ= C
i=1 wi .Mi

while the values for the prior probabilities P (ωi ) were consid-
ered the same for all classes. For the classification of C classes,
where Mi is the number of examples (xm ) of the ith class, the
rule assigns an unlabeled observation x to the class i which
results in the maximum posterior probability gi (x). The class-
weighting possibility with wi is of much interest due to the heavy
class-size imbalance inherent to this application, where the nor-
mal class is in general one order of magnitude more represented
than other classes. The classification tasks were performed using
Fig. 1. Illustration of the features calculated from the wavelet correlation
the PRtools toolbox [15] for MATLAB (The Mathworks Inc.). signals. The autocorrelation sequence of the QRS complex at scale 4 is shown
2) Features: The features used in [4] are described in for both 12L-PCA and WT-PCA strategy. The calculated features, zero crossings
Table II. As the rhythm features of the model do not depend and peaks of the autocorrelation sequence, are indicated with an asterisk.
on the number of available leads, the first four features in
Table II remain the same. Therefore, we will focus the anal-
ysis on those features describing heartbeat morphology, which puted from the two most important components. 4) Finally, for
are the ones that can be improved by the addition of new leads. the fourth strategy, called WT-PCA, the PCA is applied not to
The morphology features calculated are the first zero crossing the ECG, but to the fourth scale of the WT (W4 s(k)), and the
(kZL ) and maximum position of the autocorrelation sequence of two morphological features kZL and kM L
were calculated from
L
the ECG WT at scale 4 (kM ) for each lead L (see [4] for more the principal components.
details). Both features were calculated in four sets of leads to The PCA is performed for each heartbeat in a 160-ms window
study the most suitable way of including the additional informa- centered at the QRS complex detection sample, or fiducial point
tion. 1) The first strategy consists of including kZL and kM
L
from (PCA window in Fig. 1). Then, the multilead signal is projected
all available ECG leads and is referred as 12L (or 3L when only into the PCA basis in a wider window starting 130 ms before
three leads are available), resulting in two morphology features and ending 200 ms after the QRS fiducial point. Then, one or
per lead. 2) The second strategy computes kZL and kM L
from the two principal components are selected to compute in them the
three vectocardiogram (VCG) leads X, Y, and Z, transformed morphological features, as it is known from previous works
from 12L by the Dower matrix. This strategy can only be per- that the first two PCA of the ECG retain most of the signal
formed in 12L recordings. 3) In the third strategy, referred as energy [16]. Note that the main difference between ECG-PCA
ECG-PCA, we apply principal component analysis (PCA) to the and WT-PCA strategies, as it can be seen in Fig. 1, is where the
available ECG leads; then, the morphology features are com- PCA is calculated: in the first case, PCA is applied to the ECG
LLAMEDO et al.: CROSS-DATABASE EVALUATION OF A MULTILEAD HEARTBEAT CLASSIFIER 661

signal, while in the second case, PCA is applied to the W4 s(k) The performance is calculated from this matrix, in terms of
signal. the class sensitivity (Si ), class positive predictive value (Pi+ ),
Then, for strategies 12/3L, VCG, and ECG-PCA, we calcu- global accuracy (A), global sensitivity (S), and global positive
lated W4 s(k) for each ECG lead. Remember that for WT-PCA, predictive value (P + ) as suggested in [7]. Then, Si and Pi+ for
the WT was calculated previous to the PCA. After that, the auto- the ith class are defined as
correlation sequence of W4 s(k) is calculated, obtaining rL (k), nTii
where the final step consists in detecting the first zero crossing Si = (4)
Ni
(kZL ) and the position of the first minimum (kM L
), as shown in
Fig. 1. nTii
Pi+ = (5)
Pi
D. Performance Evaluation and the global accuracy (A), sensitivity (S), and positive
predictive value (P + ) are calculated as
We use two approaches to evaluate the performance of a clas-
1  T 
sification experiment. One consists in estimating the parameters C C
Ni
of the classifier in a training dataset, for the subsequent eval- A= nii = Si (6)
NT i=1 i=1
N T
uation of its predictive performance in a disjoint test dataset.
The other approach used is known as cross validation, and is
1 
C
adopted when there are not many examples to build the train S= Si (7)
C i=1
and test datasets. It consists in dividing a dataset into k disjoint
folds (we use ten folds), and use them as test sets, obtaining,
1  +
C
therefore, k performance measures, one for each fold. The divi- P+ = P . (8)
sion is performed by patients to avoid the presence of heartbeats C i=1 i
of the same patient in both training and test datasets. Note that
each cross-validation fold implies training in k − 1/k of the From these equations, it is clear that any imbalance in the
database patients, and testing in the remaining 1/k. The result- class representation affects P + , Pi+ , and A calculation, but not
ing performance is the mean of the k performances. S and Si .
Now, we focus on how the performance is evaluated for both Although the AAMI recommendation does not suggest any
of the approaches described previously. In a multiclass clas- measure to deal with the strong class size imbalance (see Ta-
sification problem, the confusion matrix shows the outcome ble I), we considered weighting the classes previous to the cal-
achieved by a classifier and a detailed distribution of the mis- culation of Pi+ and A in order not to neglect the performance
classified events. For a C class problem, the confusion matrix is of the less represented classes. The balancing approach used
a square matrix of dimension C in this study consists in multiplying each row of the confusion
matrix by a constant such that the sum of each row Ni is equal
Estimated classes for all classes, or Ni = Nj , ∀i = j . This is equivalent to repeat
examples of the less represented classes, in order to balance the
1 ... i ... C
⎛ ⎞ class presence.
1 nT11 . . . n1iF
. . . nF1C N1
.. ⎜ .. .. .. ⎟ ..
. ⎜ . . . ⎟ . III. RESULTS
⎜ ⎟
True classes i ⎜ nFi1 . . . nTii . . . nFiC ⎟ Ni
⎜ ⎟ As a first experiment, we compared the classification perfor-
.. ⎜ .. .. .. ⎟ ..
. ⎝ . . . ⎠ . mance of the different multilead strategies (12L, VCG, ECG-
C nFC 1 ... nFC i T
. . . nC C NC PCA, and WT-PCA) in the INCART database. For each of
the strategies, we tested the performance in different subsets
P1 ... Pi . . . PC NT . of leads. Table III shows class and global classification per-
For the ith class, nTii is the number of correctly classified formances using a kfold (k =10) cross-validation approach.
examples and nFij is the number of examples of class i classified In Table IV, we show the results of a similar experiment in
as class j; Ni is the total number of examples for class i, Pi is the same database, but considering that only three pseudo-
the number of examples classified as class i, and NT is the total orthogonal leads were available (AvF, V2, and V5). In both
number of examples in the dataset cases, the best model resulted the one with features computed
from the first two leads of the WT-PCA set. Therefore, this

Ni = nTii + nFim multilead strategy is considered for the next experiment.
m = i
In the last experiment, we validated the generalization capa-
 bility of the classification model to other databases for different
Pi = nTii + nFm i number of leads. For all databases available for a given num-
m = i ber of leads, we assessed the performance using all possible
pairs of different databases as train and test sets. We also eval-

C 
C
NT = Ni = Pi . uated the cross validated performance within each database. To
i=1 i=1 have an upper bound reference, we additionally assessed the
662 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 4, JULY 2012

TABLE III
PERFORMANCE COMPARISON BETWEEN THE DIFFERENT STRATEGIES SEPARATING AAMI2 CLASSES (N, S, AND V’) IN INCART

TABLE IV
PERFORMANCE COMPARISON BETWEEN THE DIFFERENT STRATEGIES SEPARATING AAMI2 CLASSES (N, S, V’) IN THREE PSEUDO-ORTHOGONAL
LEADS FROM INCART

TABLE V
PERFORMANCE FOR ALL DATABASES WHERE THE GENERALIZATION OF THE WT-PCA STRATEGY WAS STUDIED
LLAMEDO et al.: CROSS-DATABASE EVALUATION OF A MULTILEAD HEARTBEAT CLASSIFIER 663

performance of the model when trained and tested in the same and achieve higher performances. Certainly, the biased perfor-
database. This optimistically biased performance serves as an mance can be thought as a metric of how difficult to classify is a
upper bound, and represents the performance of the model if the database by a given classification model, the closer to 100%, the
distributions of the examples in both training and test datasets easier. This last result reinforces the importance of evaluating
were identical. These results, grouped by test database, are pre- the performance of a classifier in several databases.
sented in Table V for databases with 12, 3, and 2 leads. Results One advantage of the proposed approach is that it can be used
show that the reference model extended with the selected WT- for an arbitrary number of leads, because after the PCA we only
PCA multilead strategy presents good generalization properties retain the two most important components for the morphologi-
for 3 and 12 leads, while a certain degradation is observed when cal feature calculation. These components are calculated specif-
using only two leads. ically for the QRS complex, and in the W4 s(k) signal (with
a band between 11.25 and 22.5 Hz), typically where the ECG
presents high SNR. However, in the case of a large-scale artifact
IV. DISCUSSION AND CONCLUSIONS during the QRS complex (as a lead disconnection), the PCA
In this paper, we have adapted and improved a two-lead heart- calculation would be corrupted, being this the main limitation
beat classifier by including the additional morphology infor- found for this approach.
mation present in multilead recordings, like those of INCART The performance improvement with respect to [4] is how-
database. We followed the concept of the morphology features ever moderate, probably because the automatic classification
assessed in [4], but calculating these features in sets of leads ob- approach is close to the performance limit achievable with the
tained by following different lead transformation strategies. The current classification model. The worst aspect of performance
simplest strategies consisted in computing the features kZL and remains classification of supraventricular ectopic beats, where
L
kM in all available leads (12L/3L), and in the derived orthogonal further study is needed. Regarding the ventricular class, tech-
leads (VCG). The other two strategies apply PCA to the ECG niques of patient adaptation as described in [17] are also under
or its WT previously to the morphology feature computation. development.
The results suggest that strategies using PCA performed better. This results represent an improvement in performance with
Moreover, the WT-PCA strategy obtains the best improvements respect to the previous two-lead classification model, conclud-
with respect to the two-lead classifier obtained in [4], either ing that the adequate addition of multilead information allows
in recordings of 12 or 3 leads. This can be explained because the performance improvement of a heartbeat classifier.
in WT-PCA, the PCA is calculated in W4 s(k) where most of
the noise and other components not related to the QRS have
been filtered out [13], and therefore, PCA provides a better rep- REFERENCES
resentation of the multilead evolution of the QRS complex. It
[1] World Health Organization. (2012). Cardiovascular diseases. [Online].
must be remembered that although both PCA and WT are linear Available: http://www.who.int/cardiovascular_diseases/en/
transformations, the eigenvector calculation is not linear, and [2] G. J. Taylor, 150 Practice ECGs: Interpretation and Review. Oxford,
therefore, sets ECG-PCA and WT-PCA differ in how the mul- U.K.: Blackwell, 2002.
[3] P. de Chazal, M. O’Dwyer, and R. B. Reilly, “Automatic classification of
tilead signal is projected into the principal components. Tables heartbeats using ECG morphology and heartbeat interval features,” IEEE
III and IV show that only the WT-PCA strategy showed the Trans. Biomed. Eng., vol. 51, no. 7, pp. 1196–1206, Jul. 2004.
largest observed performance improvement with respect to the [4] M. Llamedo and J. P. Martı́nez, “Heartbeat classification using feature
selection driven by database generalization criteria,” IEEE Trans. Biomed.
two-lead reference model developed in [4]. Eng., vol. 58, no. 3, pp. 616–625, Mar. 2011.
Results in Table V confirm the generalization capability of [5] I. Christov, G. Gómez-Herrero, V. Krasteva, I. Jekova, A. Gotchev, and
the model using the selected WT-PCA strategy to the rest of K. Egiazarian, “Comparative study of morphological and time-frequency
ECG descriptors for heartbeat classification,” Elsevier Med. Eng. Phys.,
the studied databases. For the case of 12L recordings, as those vol. 28, pp. 876–887, 2006.
included in the INCART and Biosigna databases, results show [6] K. Park, B. Cho, D. Lee, S. Song, and J. Lee, “Hierarchical support vector
very good generalization for both databases since the perfor- machine,” in Computers in Cardiology. vol. 35, Washington, DC: IEEE
Computer Society Press, 2008, pp. 229–232.
mance is slightly lower than the biased performance when train- [7] Testing and Reporting Performance Results of Cardiac Rhythm and
ing in the other database. In the same table, almost the same ST-Segment Measurement Algorithms, American National Standard
figures can be seen for the case of 3L. The last results presented ANSI/AAMI/ISO EC57, 1998.
[8] V. Chudácek, G. Georgoulas, L. Lhotská, C. Stylios, M. Petrı́k, and
in Table V are for databases of 2L. It can be noted that the in- M. Cepek. (2009). Examining cross-database global training to eval-
terdatabase dispersion in the performance increased, probably uate five different methods for ventricular beat classification. Phys-
because of the heterogeneity of the databases considered. iol. Meas. [Online]. 30(7), p. 661. Available: http://stacks.iop.org/0967-
3334/30/i=7/a=010.
The results suggest that databases INCART and MITBIH- [9] A. L. Goldberger, L. A. N. Amaral, L. Glass, J. M. Hausdorff, P. C. Ivanov,
AR share similar distributions in the feature space, since both R. G. Mark, J. E. Mietus, G. B. Moody, C.-K. Peng, and H. E. Stanley,
obtained the maximal reciprocal performance. Other interesting “PhysioBank, PhysioToolkit, and PhysioNet: Components of a new re-
search resource for complex physiologic signals,” Circulation, vol. 101,
aspect regarding the MITBIH databases is the lower perfor- no. 23, pp. e215–e220, 2000.
mance obtained even for the biased case. This fact evidences [10] G. Moody and R. Mark, “The impact of the MIT-BIH arrhythmia
the diversity of patients and ECG contamination (noise, lead database,” IEEE Eng. Med. Biol., vol. 20, no. 3, pp. 45–50, May/Jun.
2001.
disconnection, and misplacement) included in these databases, [11] S. Greenwald, “Improved detection and classification of arrhythmias in
and therefore the limitation of our classifier to model the data noise-corrupted electrocardiograms using contextual information,” Ph.D.
664 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 16, NO. 4, JULY 2012

dissertation, Harvard–MIT Division of Health Sciences and Technology, Antoun Khawaja was born in Damascus, Syria, in
Cambridge, 1990. 1977. He received the M.Sc. degree in biomedical
[12] R. Fischer, M. Sinner, R. Petrovic, E. Tarita, and S. Kääb, T. K. Zywi- engineering from Damascus University, Damascus,
etz, “Testing the quality of 12 lead Holter analysis algorithms,” in Proc. in 2001, and the Ph.D. degree in biomedical engi-
Comput. Cardiol., 2008, vol. 35, pp. 453–456. neering from the Karlsruhe Institute of Technology,
[13] J. P. Martı́nez, R. Almeida, S. Olmos, A. Rocha, and P. Laguna, “A Karlsruhe, Germany, in 2007.
wavelet-based ECG delineator: Evaluation on standard databases,” IEEE Since 2007, he has been a Researcher and Scien-
Trans. Biomed. Eng., vol. 51, no. 4, pp. 570–581, Apr. 2004. tific Developer in the industry. He is currently the
[14] F. van der Heijden, R. Duin, D. de Ridder, and D. Tax, Classification, Chief Scientist of Biosigna GmbH, Munich, Ger-
Parameter Estimation and State Estimation. New York: Wiley, 2005. many. His current research interests include the field
[15] R. Duin, P. Juszczak, P. Paclik, E. Pekalska, D. deRidder, D. Tax, and S. of biosignal processing and advanced automatic ECG
Verzakov. (2008). PR-tools, a MATLAB toolbox for pattern recognition. analysis.
[Online]. Available: http://www.prtools.org.
[16] B. Acar and H. Köymen, “SVD-based on-line exercise ECG signal or-
thogonalization,” IEEE Trans. Biomed. Eng., vol. 46, no. 3, pp. 311–321,
Mar. 1999.
[17] P. de Chazal and R. B. Reilly, “A patient-adapting heartbeat classifier
using ECG morphology and heartbeat interval features,” IEEE Trans.
Biomed. Eng., vol. 53, no. 12, pp. 2535–2543, Dec. 2006.

Juan Pablo Martı́nez was born in Zaragoza, Spain,

Mariano Llamedo was born in Buenos Aires,
in 1976. He received the M.S. degree in telecommu-
Argentina, in 1979. He received the M.Sc. degree
nication engineering and the Ph.D. degree in biomed-
in electronic engineering from the National Techno-
logical University of Buenos Aires (UTN- FRBA), ical engineering from the University of Zaragoza
(UZ), Zaragoza, in 1999 and 2005, respectively.
Buenos Aires, in 2005.
In 2000, he was an Assistant Professor at the
From 2005 to 2008, he was with the Department of
Aragon Institute of Engineering Research, UZ, where
Electronic Engineering, UTN-FRBA, as an Assistant
Professor and Research Fellow. Since 2008, he has since 2007, he has been an Associate Professor. He is
also with the Centro de Investigación Biomédica en
been a Researcher with the Aragon Institute of Engi-
Red en Bioingenierı́a, Biomateriales y Nanomedic-
neering Research, University of Zaragoza, Zaragoza,
ina, Zaragoza. His current research interests include
Spain. His professional research interests include the
field of biomedical signal processing, with main interest in signals of cardio- biomedical signal processing, with main interest in signals of cardiovascular
origin.
vascular origin.