570 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 51, NO.

4, APRIL 2004

A Wavelet-Based ECG Delineator: Evaluation on

Standard Databases
Juan Pablo Martínez*, Rute Almeida, Salvador Olmos, Member, IEEE, Ana Paula Rocha, and
Pablo Laguna, Member, IEEE

Abstract—In this paper, we developed and evaluated a robust The topic of automatic delineation of ECG has been widely
single-lead electrocardiogram (ECG) delineation system based on studied. We can distinguish two main groups of algorithms:
the wavelet transform (WT). In a first step, QRS complexes are de- QRS detection algorithms and wave delineation algorithms.
tected. Then, each QRS is delineated by detecting and identifying
The QRS complex is the most characteristic waveform of the
the peaks of the individual waves, as well as the complex onset
and end. Finally, the determination of P and T wave peaks, on- ECG signal. Its high amplitude makes QRS detection easier
sets and ends is performed. We evaluated the algorithm on several than the other waves. Thus, it is generally used as a reference
manually annotated databases, such as MIT-BIH Arrhythmia, QT, within the cardiac cycle. A wide diversity of algorithms have

= 99 66%
European ST-T and CSE databases, developed for validation pur- been proposed in the literature for QRS detection. An exten-
+ = 99 56%
poses. The QRS detector obtained a sensitivity of sive review of the approaches proposed in the last decade can be
and a positive predictivity of over the first lead found in [1]. Older detectors are reviewed in [2]–[4]. A general-
+
of the validation databases (more than 980,000 beats), while for
the well-known MIT-BIH Arrhythmia Database, and over ized scheme [2] that matches most nonsyntactic QRS detectors
99.8% were attained. As for the delineation of the ECG waves, the presents a two-stage structure: a preprocessing stage, usually in-
mean and standard deviation of the differences between the auto- cluding linear filtering followed by a nonlinear transformation,
matic and manual annotations were computed. The mean error ob- and the decision rule(s).
tained with the WT approach was found not to exceed one sampling Concerning delineation (determination of peaks and limits of
interval, while the standard deviations were around the accepted the individual QRS waves, P and T waves), algorithms usually
tolerances between expert physicians, outperforming the results of
other well known algorithms, especially in determining the end of depart from a previous QRS location and define temporal search
T wave. windows before and after the QRS fiducial point to seek for the
other waves. Once the search window is defined, some tech-
Index Terms—ECG wave delineation, P wave, QRS detection, T nique is applied to enhance the characteristic features of each
wave, wavelets.
wave (e.g., its frequency band) in order to find the wave peaks.
The localization of wave onsets and ends is much more dif-
I. INTRODUCTION ficult, as the signal amplitude is low at the wave boundaries
and the noise level can be higher than the signal itself. It is
T HE analysis of the ECG is widely used for diagnosing
many cardiac diseases, which are the main cause of mor-
tality in developed countries. Since most of the clinically useful
also worthwhile to note that there is not any universally ac-
knowledged clear rule to locate the beginning and the end of
ECG waves, which complicates the systematization of onset and
information in the ECG is found in the intervals and amplitudes
end localization. One can find in the literature very different
defined by its significant points (characteristic wave peaks and
delineation approaches based on mathematical models [5]–[7],
boundaries), the development of accurate and robust methods
the signal envelope [8], matched filters [9], ECG slope criteria
for automatic ECG delineation is a subject of major importance,
[10]–[14], second-order derivatives [15], low-pass differentia-
especially for the analysis of long recordings. As a matter of
tion (LPD) [16]–[18], the wavelet transform (WT) [19]–[21],
fact, QRS detection is necessary to determine the heart rate, and
nonlinear time-scale decomposition [22], adaptive filtering [23],
as reference for beat alignment; ECG wave delineation provides
dynamic time warping [24], artificial neural networks [25], [26],
fundamental features (amplitudes and intervals) to be used in
or hidden Markov models [27]. Some of the algorithms pre-
subsequent automatic analysis.
sented in these works can only be used to obtain a subset of
the ECG characteristic points (e.g., QT interval, QRS limits, T
Manuscript received February 28, 2003; revised August 7, 2003. end, etc).
This work was supported in part by MCyT and FEDER under Project The validation of most recently published QRS detectors is
TIC2001-2167-CO2-02, in part by DGA under Project P075/2001, and in part based on standard databases. On the other hand, most of the
by the integrated action HP2001-0031/CRUP-E26/02. The work of R. Almeida
was supported by FCT and ESF (III CSF) under Grant SFRH/BD/5484/2001. works about ECG delineation do not use this approach, and that
Asterisk indicates corresponding author. makes the reproducibility and comparability of the performance
*J. P. Martínez is with the Communications Technology Group, Aragon Insti- results more difficult.
tute of Engineering Research, University of Zaragoza, María de Luna, 1, 50015
Zaragoza, Spain (e-mail: [email protected]).
The wavelet transform provides a description of the signal in
R. Almeida and A. P. Rocha are with the Departamento de Matemática Apli- the time-scale domain, allowing the representation of the tem-
cada, Faculdade de Ciências, Universidade do Porto, 4169 007 Porto, Portugal. poral features of a signal at different resolutions; therefore, it is
S. Olmos and P. Laguna are with the Communications Technology Group, a suitable tool to analyze the ECG signal, which is characterized
Aragon Institute of Engineering Research, University of Zaragoza, 50015
Zaragoza, Spain. by a cyclic occurrence of patterns with different frequency con-
Digital Object Identifier 10.1109/TBME.2003.821031 tent (QRS complexes, P and T waves). Moreover, the noise and
0018-9294/04$20.00 © 2004 IEEE

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
MARTÍNEZ et al.: A WAVELET-BASED ECG DELINEATOR. EVALUATION ON STANDARD DATABASES 571

artifacts affecting the ECG signal also appear at different fre-

quency bands, thus having different contribution at the various
scales.
In [19], a multiscale QRS detector including a method for de-
tecting monophasic P and T waves was proposed, although only
the QRS detector was validated. In this paper, we present a gen-
eralization of that method, including the determination of the
individual QRS waves, and a robust delineation of QRS, P, and
T waves for a wide range of morphologies. The performance
is assessed using standard manually annotated ECG databases,
where other algorithms have already been tested: MIT-BIH Ar-
rhythmia [28], QT [29], European ST-T [30], and CSE multi- Fig. 1. Two filter-bank implementations of DWT. (a) Mallat’s algorithm.
lead measurement [31] databases. Some of the evaluation results (b) Implementation without decimation (algorithme à trous).
obtained using a preliminary version of this delineator were al-
ready presented in [32]. The scale factor and/or the translation parameter can be
The paper is organized as follows: in Section II, we present discretized. The usual choice is to follow a dyadic grid on the
the detection and delineation algorithms and the validation time-scale plane: and . The transform is then
process. The results of the validation on several databases and called dyadic wavelet transform, with basis functions
their comparison to other algorithms are given in Section III and
discussed in Section IV. Finally, the conclusions are presented (3)
in Section V.
For discrete-time signals, the dyadic discrete wavelet trans-
form (DWT) is equivalent, according to Mallat’s algorithm, to
II. MATERIALS AND METHODS
an octave filter bank [35], and can be implemented as a cas-
A. Wavelet Transform cade of identical cells [low-pass and high-pass finite impulse
The wavelet transform is a decomposition of the signal as response (FIR) filters], as illustrated in Fig. 1(a). From the trans-
a combination of a set of basis functions, obtained by means formed coefficients and the low-pass residual, the
of dilation ( ) and translation ( ) of a single prototype wavelet original signal can be rebuilt using a reconstruction filter bank.
. Thus, the WT of a signal is defined as However, for this application we are only interested in the anal-
ysis filter bank.
The downsamplers after each filter in Fig. 1(a) remove the
(1) redundancy of the signal representation. As side effects, they
make the signal representation time-variant, and reduce the tem-
The greater the scale factor is, the wider is the basis function poral resolution of the wavelet coefficients for increasing scales.
and consequently, the corresponding coefficient gives informa- To keep the time-invariance and the temporal resolution at dif-
tion about lower frequency components of the signal, and vice ferent scales, we use the same sampling rate in all scales, what
versa. In this way, the temporal resolution is higher at high fre- is achieved by removing the decimation stages and interpolating
quencies than at low frequencies, achieving the property that the the filter impulse responses of the previous scale. This algo-
analysis window comprises the same number of periods for any rithm, called algorithme à trous [36], is shown in Fig. 1(b).
central frequency. Using this algorithm, the equivalent frequency response for the
If the prototype wavelet is the derivative of a smoothing th scale is
function , it can be shown [33], [34] that the wavelet trans-
form of a signal at scale is

(4)
(2)
B. Prototype Wavelet Used in This Paper
where is the scaled version of the In this paper, we used as prototype wavelet a quadratic
smoothing function. The wavelet transform at scale is pro- spline originally proposed in [33]. This wavelet was already ap-
portional to the derivative of the filtered version of the signal plied to ECG signals in [19] and [21], while in [20], the deriva-
with a smoothing impulse response at scale . Therefore, the tive of a Gaussian smoothing function was used. The quadratic
zero-crossings of the WT correspond to the local maxima spline Fourier transform is
or minima of the smoothed signal at different scales, and
the maximum absolute values of the wavelet transform are
associated with maximum slopes in the filtered signal. (5)
Regarding our application, we are interested in detecting
ECG waves, which are composed of slopes and local maxima From (5), the wavelet can be easily identified as the derivative of
(or minima) at different scales, occurring at different time the convolution of four rectangular pulses, i.e., the derivative of
instants within the cardiac cycle. Hence, the convenience of a low-pass function. Fig. 2 represents the wavelet and smoothing
using such a type of prototype wavelet. functions used in this paper.

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
572 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 51, NO. 4, APRIL 2004

some adaptation procedures are required for the system to be

able to handle equivalently ECG signals with different sampling
frequencies. Other previously published waveform delineators
using the WT [19]–[21] have not accounted for this fact, con-
sidering a unique sampling frequency ( Hz).
Resampling the signal is a time-demanding solution. A better
solution is to compute, for each , a new set of filters having
equivalent analogue frequency responses as close as possible to
the ones of Fig. 3. For this purpose, we resampled adequately
the equivalent filter impulse responses at 250 Hz
(where stands for the Fourier transform) to
other sampling rates. The equivalent frequency responses cor-
responding to the sampling rates of MIT-BIH Arrhythmia and
CSE databases (360 Hz and 500 Hz, respectively) are shown and
Fig. 2. Prototype wavelet (t) and smoothing function  (t).
compared with the ones at 250 Hz in Fig. 4. It can be observed
that the frequency responses of the adapted filter bank constitute
a good approximation of the original filters up to a frequency of
45–50 Hz.

D. Description of the Algorithms

The algorithms presented in this section apply directly over
the digitized ECG signal without any prefiltering. The ECG
signal can, in any case, be preprocessed as usual in order to re-
duce the noise level. Nevertheless, frequency domain filtering
is implicitly performed when computing the DWT, making the
system robust and allowing the direct application over the raw
ECG signal.
From the equivalent responses in Fig. 3 and according to the
spectrum of the ECG signal waves [38], it is clear that most of
the energy of the ECG signal lies within the scales to . For
Fig. 3. Equivalent frequency responses of the DWT at scales 2 , k = 1; . . . ; 5 scales higher than , the energy of the QRS is very low. The P
for 250-Hz sampling rate. and T waves have significant components at scale although
the influence of baseline wandering is important at this scale.
For the selected prototype wavelet, the filters and Fig. 5, inspired by [19, Fig. 1], shows several simulated waves
to implement the DWT as in Fig. 1 are [19], [37] similar to those in the ECG, together with the first five scales of
their DWT. As exemplified by (a), monophasic waves produce
a positive maximum-negative minimum pair along the scales,
with a zero crossing between them. Each sharp change in the
(6) signal is associated to a line of maxima or minima across the
scales. In wave (b), which simulates a QRS complex, it can be
which are FIR filters with impulse responses observed that the small Q and S wave peaks have zero crossings
associated in the WT, mainly at scales and . P or T-like
waves (c) have their major component at scales to , whereas
(7) artifacts like (d) produce isolated maximum or minimum lines
which can be easily discarded. If the signal is contaminated with
Using the algorithme à trous of Fig. 1(b) and (4) and (6), high-frequency noise (e), the most affected scales are and
the frequency responses of the first five scales are those repre- , being higher scales essentially immune to this sort of noise.
sented in Fig. 3, considering a sampling frequency of 250 Hz. It Baseline wander (f) affects only at scales higher than .
is noteworthy that the transfer functions show a low-pass differ- Using the information of local maxima, minima and zero
entiator characteristic. As the analysis filters have linear phase crossings at different scales, the algorithm identifies the
[19], the outputs of the filters can be realigned in order to present significant points in the following four steps: 1) detection of
the same delay with respect to the original ECG. Therefore, the QRS complexes; 2) detection and identification of the QRS
wavelet-based approach for ECG delineation can be considered individual waves (Q, R, S, R’), and determination of the QRS
as a differentiator filter-bank approach with the filter responses complex boundaries; 3) T wave detection and delineation; and
in (4). 4) P wave detection and delineation.
1) QRS Detection: First of all, QRS complexes are detected
C. Adaptation of the Filters to Other Sampling Frequencies using an algorithm based on the multiscale approach proposed
The result of using the same filters for a sampling rate other by Li and coworkers in [19]. This algorithm searches across
than 250 Hz is the frequency-scaling of the bands in Fig. 3. Thus, the scales for “maximum modulus lines” exceeding some

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
MARTÍNEZ et al.: A WAVELET-BASED ECG DELINEATOR. EVALUATION ON STANDARD DATABASES 573

Fig. 4. Equivalent frequency responses of the filters used for sampling rates of (a) 360 Hz and (b) 500 Hz (continuous lines). Dashed lines are the equivalent
frequency responses of the original filters for signals sampled at 250 Hz.

detection and identification of the QRS individual waves. The

algorithm departs from the position given by the detector,
, which must be flanked by a pair of maximum moduli
with opposite signs at scale , namely at and . The
delineator looks before and after for significant
maxima of accounting for other adjacent slopes
within the QRS complex. To consider a local maximum
modulus as significant, it must exceed the threshold,
or respectively for previous or subsequent waves. The
zero crossings between these significant slopes at scale are
assigned to wave peaks, and labeled depending on the sign and
the sequence of the maximum moduli. The algorithm considers
any possible QRS morphology with three or less waves (QRS,
RSR’, QR, RS, R, and QS complexes), and includes protection
measures, based on time interval and sign rules, to reject
notches in waves and anomalous deflections in the ECG signal.
See Fig. 7 for examples of these complexes.
The onset (end) of the QRS is before (after) the first (last) sig-
Fig. 5. WT at the first five scales of ECG-like simulated waves. (Inspired by nificant slope of the QRS, which is associated with a maximum
[19, Fig. 1].). of . So, we first identify the samples of the first and
last peaks associated with the QRS in , say and
thresholds at scales from to ; namely, , , . Then, candidates to onset and end are determined by ap-
, and (see Appendix for more details abouth plying two criteria: i) searching for the sample where
the thresholds). After rejecting all isolated and redundant is below a threshold ( or ) relative to the ampli-
maximum lines, the zero crossing of the WT at scale tude of the maximum modulus ( or );
between a positive maximum-negative minimum pair is marked ii) searching for a local minimum of before or
as a QRS. Other protection measures are taken, like a refractory after . Finally the QRS onset and end are selected as the
period or a search back with lowered thresholds if a significant candidates that supply the nearest sample to the QRS.
time has elapsed without detecting any QRS. Our implemen- 3) T Wave Detection and Delineation: The process for mul-
tation, though based on [19] is slightly different: the search tiscale T wave detection and delineation is as follows: first of
for the main wave of the QRS is not restricted to an R wave, all, we define a search window for each beat, relative to the QRS
allowing the detection of negative waves (negative minimum – position and depending on a recursively computed RR interval.
positive maximum pairs). Afterwards, the individual waves are Within this window, we look for local maxima of .
identified. Besides, our algorithm does not include regularity If at least two of them exceed the threshold , a T wave is
analysis, but only amplitude based criteria, and the thresholds considered to be present. In this case, the local maxima of WT
are not updated for each beat, but for each excerpt of with amplitude greater than are considered as significant
samples. slopes of the wave, and the zero crossings between them as the
In Fig. 6, some ECG excerpts from the MIT-BIH Arrhythmia wave peaks. Depending on the number and polarity of the found
database have been selected to illustrate the robustness of the maxima, we assign one out of six possible T wave morpholo-
detector dealing with motion artifacts, muscular noise, baseline gies: positive (+), negative (-), biphasic (+/- or -/+), only up-
wandering, and changes in the QRS morphology. wards, and only downwards (see Fig. 8). If the T wave is not
2) QRS Delineation (Onset, End and Individual found in scale we repeat the above process over .
Waves): One of the novelties with respect to [19] is the Attending to the loss of time resolution in the growing scales,

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
574 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 51, NO. 4, APRIL 2004

Fig. 6. Examples of the behavior of the QRS detector dealing with different kinds of noise and morphology changes. (a) Motion artifact, (b) muscular noise, (c)
baseline wandering, and (d) QRS morphology changes. The ECG signal, the WT’s first four scales and the detected QRS complexes (vertical lines) are shown in
each panel.

the peak(s) of the T wave correspond to the zero crossing(s) at marks (MITDB) or QRS onsets (EDB). The QTDB includes
scale , if they exist, or at the scale in which T wave was some records from EDB and MITDB and also from several other
found. To identify the wave limits, we used the same criteria as MIT-BIH databases (ST Change, Supraventricular Arrhythmia,
for QRS onset and end, with thresholds and applied Normal Sinus Rhythm, Sudden Death, and Long Term). This
to scale . database was developed for wave limits validation purposes and
4) P Wave Detection and Delineation: The P wave algo- it provides cardiologist annotations for at least 30 beats per
rithm is similar to the T wave algorithm, using an appropriate recording (ref1), with marks including QRS complexes, P and
RR-dependent search window and adequate thresholds ( , , T waves peaks, onsets and ends. That makes an amount of more
and ). For P wave only four different morphologies than 3600 annotated beats. The QTDB also includes, for 11
are admitted: positive (+), negative (-), and biphasic (+/-, -/+), out of its 105 records, an additional annotation performed by
as illustrated in Fig. 9. a second cardiologist (ref2), making a total of 404 beats with
a double reference annotation. In 79 of the 105 recordings ad-
E. Validation ditional annotation files are provided with the QRS positions
As there is no golden rule to determine the peak, onset and manually annotated or audited during the whole recording. The
end of the ECG waves, the validation of the delineator must be CSEDB signals include the 12 standard leads and the Frank
done using manually annotated databases. For these purposes, leads. This database supplies, in a limited number of beats (at
we used some easily-available or standard databases, namely most one beat per record), median referee annotations based
the MIT-BIH Arrhythmia database (MITDB), the QT database on five referee cardiologists. The cardiologists only analyzed
(QTDB), the European ST-T database (EDB) and the data set a subsample of the library (every fifth record) and, additionally
3 of the CSE multilead measurement database (CSEDB). The some waves for which a set of analysis programs differed signif-
main characteristics of these databases are compiled in Table I. icantly. Thus, the number of manually annotated beats is scarce
The MITDB includes specially selected Holter recordings (32 beats).
with anomalous but clinically important phenomena. The EDB For validation of QRS detection, we used the MITDB, the
files present ischemic episodes extracted from Holter record- EDB, and the QTDB (79 records) and to evaluate the delin-
ings. These databases include annotations of QRS positions: R eation performance, we used the whole QTDB and the CSEDB.

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
MARTÍNEZ et al.: A WAVELET-BASED ECG DELINEATOR. EVALUATION ON STANDARD DATABASES 575

Fig. 7. Examples of beats from the QTDB with different QRS complex morphologies with their WT at scales 2 and 2 and the peaks (short lines) and QRS
boundaries (long lines) determined by the algorithm. (a) QRS complex, (b) QRS complex (with notch), (c) R complex, (d) RSR’ complex (e), RS complex, and
(f) QS complex.

These are, to our knowledge, the only two standard databases To assess the QRS detector we calculated the sensi-
that have been used to test delineation techniques. Additionally, tivity TP ( TP FN) and positive predictivity
EDB database was also used for performance evaluation of the TP ( TP FP ), where TP is the number of true
QRS onset determination. positive detections, FN stands for the number of false negative
For QRS detection, only the first channel of each record detections, and FP stands for the number of false positive
was processed with the aim of comparing with other published misdetections. Regarding wave delineation, we calculated
works. The wavelet-based delineator works on a single-channel as the average of the errors, taken as the time differences
basis, while the manual annotation process was performed between automatic and cardiologist annotations. The average
having in sight all available leads. Therefore, to compare in standard deviation ( ) of the error was computed by averaging
a reasonable way the manual annotations on the QTDB and the intrarecording standard deviations. In the CSEDB, as
EDB with the two single-channel annotation sets produced by there is only one annotated beat per record, corresponds to
the delineator, we chose for each point the channel with less the standard deviation of the errors. For QTDB, was also
error. In the CSEDB we obtained with our system 15 different calculated, but given the format of this database, it was not
sets of annotations, one for each channel, and we needed a possible to quantify the , as it was already noted in [7]. As a
more complex rule for selecting a single location for each matter of fact, when there is not an annotation, we do not know
characteristic point. We used for that purpose a rule consisting either if the cardiologist considered that no wave was present
of ordering the 15 single-lead annotations and selecting as the or if he simply believed that he could not confidently annotate
onset (end) of a wave the first (last) annotation whose nearest the point (e.g., because of the noise). Anyway, for points other
neighbors lay within a ms interval. This rule had been already than the QRS complex, can be calculated considering each
used in [11] ( and and ms for QRS boundaries) and absent reference mark on an annotated beat as a not present
[16] ( , and ms for P wave limits and QRS onset, wave decision. The obtained values can be interpreted as a
ms for QRS end, and ms for T end). lower limit ( ) for the actual .

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
576 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 51, NO. 4, APRIL 2004

Fig. 8. Examples of different T wave morphologies, and their wavelet transform at scale 2 . Short vertical lines show peak annotations while long lines denote
wave boundaries. (a) Positive T wave, (b) negative T wave, (c) biphasic T wave, (d) ascending T wave, (e) descending T wave, and (f) T wave with low SNR.

Fig. 9. Examples of P waves with different polarities and in absence of P wave, their WT at scale 2 , and marks for peaks, onsets and ends. (a) Positive P wave,
(b) negative P wave, and (c) absent P wave.

III. RESULTS [39] (in single-channel mode), and other published detectors are
given in Table II. Most of the algorithms were tested on the
A. Results for QRS Detection MITDB. In some works [20], [40], the first 5 min of the MITDB
The detection performance on the MITDB, QTDB, and EDB were used as a learning period, and were not considered in the
obtained by our WT-based QRS detector, the software Aristotle validation. Since our algorithm does not need any learning pe-

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
MARTÍNEZ et al.: A WAVELET-BASED ECG DELINEATOR. EVALUATION ON STANDARD DATABASES 577

TABLE I IV. DISCUSSION

CHARACTERISTICS OF THE VALIDATION DATABASES

The proposed WT-based delineation system achieves very

good detection performance on the studied databases. The QRS
detector attains and (0.78% of
errors) on the first lead of the three databases: 983 423 beats
( hours of ECG). On the MITDB (widely used to eval-
uate QRS detectors), only three detectors based on WT ( [19],
riod, the entire recordings were considered. We excluded seg-
[21] and this paper) and a very recent one based on more clas-
ments with ventricular flutter in record 207 of MITDB (2 min
sical approaches [46], report and over 99.8%. However,
24 s) and those annotated as unreadable in the EDB (57 min 6
the extensive use of the MITDB as a testing database can hide
s in lead 1, and only 2 min 46 s in both leads simultaneously).
an overtunning of the detector parameters to fit this particular
Partial results are presented in [20] using the first channel of
database. Consequently, the validation of the same detector on
the eight first recordings of the MITDB (100 to 107) and leaving
a second data set without any later parameter tunning can help
out the first 5 min. That WT-based QRS detector achieve in
to obtain more reliable performance results. Actually, we fixed
those recordings 93 FP and 84 FN out of 14 481 analyzed beats
the thresholds using signals of the MITDB as training set, but as
( and ). Our WT-based detector
it can be observed in Table II, we obtained similar performance
presented in those excerpts 43 FP and 23 FN (
on the QTDB or the EDB, modifying only the resampling of the
and ).
filters as explained in Section II-C to cope with the differences in
B. Delineation Results the sampling frequency. It is important to remark that in EDB,
the first channel of record e0305 (which exhibits very narrow
The global validation results obtained with the WT-based de- and tall T waves) is responsible for 42% of the FP and 57% of
lineator (this paper) and a low-pass-differentiator-based method the FN. Excluding this record, the performance in the EDB is
(LPD) [16] on the QTDB are given in Table III. A previous ver- , and the total performance in the
sion of the LPD algorithm had already been validated on the three datasets increases to , over
QTDB in [47]. The results for T peak and T end of the recently 974 042 beats.
proposed T-U complex detector in [7] are shown as well in that
The WT-based detector, in contrast to most QRS detectors
table. In the last row, we include the accepted two-standard-de-
found in the literature, allows to take advantage of the same
viation tolerances given by the CSE working party from mea-
wavelet analysis stage for ECG wave delineation, due to the
surements made by different experts [48, Table 2].
particularly appropriate characteristics of time-scale analysis. In
In Table IV we compare, within the subset of the QTDB with
[19] and [21], the possibility of detecting monophasic P and T
double reference annotations, the delineation errors with respect
wave peaks was stated, but not evaluated. Only in [20], an al-
to both referees (ref1 and ref2) and the intercardiologist dif-
gorithm for detecting the peaks, onsets and ends of monophasic
ferences. Since the second cardiologist did not annotate any P
P and T waves was validated using the CSEDB. Our algorithm
wave, no results are given for this wave.
allows delineation of a wide variety of QRS complex, P wave
The special interest in T wave delineation and the high dif-
and T wave morphologies.
ficulty of the determination of its end justifies a more detailed
study. A record-by-record classification was performed as pro- The delineation performance on the QTDB (Table III) shows
posed in [47] and more recently in [7]. In order to facilitate the that our WT algorithm can detect with high sensitivity the P and
comparison with previous works we chose the same threshold T waves annotated by cardiologists in the ECG (
than in [7], i.e., 15 ms for the bias and 30.6 ms for the standard for the P waves and 99.77% for the T waves), and can delineate
deviation. Thus, the records are classified into four groups, ac- them with mean errors which are in all cases smaller than or
cording to: Group I: ms and ms; Group II: around one sample (4 ms). The standard deviations are around
ms and ms; Group III: ms and three samples for the P wave, two samples for QRS onsets and
ms and Group IV: ms and ms. The ends, and three to four samples for the T peak and the T end.
stratification results with the three algorithms for the T wave Despite is a conservative estimate of the actual , the
peak and T end are given in Table V. values found are also satisfactory ( for the P
In the CSEDB, we used the rule explained in Section II-E for waves and for the T waves).
selecting a single multilead position for each significant point. The comparison of our results with those of the LPD approach
The performance was found to be very sensitive to the chosen and the TU detector allows to observe that our algorithm outper-
and values. We obtained the best performance using forms the others clearly in the T wave delineation, especially
and ms for P limits, 10 ms for QRS end, and 12 ms for in the T wave end, where the standard deviation of the error is
QRS onset and T end. The multilead performance results, taking nearly reduced in one third. The results in the other points are
as reference the median cardiologist annotations are given in quite similar and the differences between them are small in com-
Table VI. parison to the sampling interval.
Finally, we also used the EDB for QRS onset validation, con- Some works considered the values given by the CSE Working
sidering that the QRS reference points annotated in this database Party in [48, Table 2] as a reference for delineation error toler-
are QRS onset marks. Selecting the channel with less error for ances. In the cited article, it was stated “the standard deviation of
each beat, we obtained a mean error of 0.1 ms and an averaged the differences [of a program results] from the reference should
standard deviation of 7.5 ms over 790 287 beats. not exceed certain limits as listed in Table II ”. However, the

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
578 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 51, NO. 4, APRIL 2004

TABLE II
QRS DETECTION PERFORMANCE COMPARISON ON SEVERAL DATABASES (FIRST CHANNEL). (N/R: NOT REPORTED)

TABLE III
DELINEATION PERFORMANCE COMPARISON IN THE QTDB (TWO-CHANNELS). (N/R: NOT REPORTED, N/A: NOT APPLICABLE)

TABLE IV
DELINEATION PERFORMANCE COMPARISON BETWEEN THE PRESENTED DELINEATOR AND BOTH REFEREES

limits given in that table, were obtained as two standard devia- algorithms, although the tolerance for T end would be nearly
tions of the differences (in millisseconds) between the median of accomplished by the WT approach.
the individual readers and the final referee estimates. As a conse- However, the interpretation of the CSE tolerances when as-
quence, some authors [7], [16], [20], [24] considered that an al- sessing the results on the QTDB is not so simple, since they
gorithm should accomplish (loose criterion), while were computed from a set of signals with different number of
for others [11], [22], a standard deviation should be channels, resolution, sampling frequency, quality, and rhythms.
attained (strict criterion). From the results in the QTDB, WT Given the 11 recordings annotated by a second cardiologist in
and LPD detectors would accomplish the “loose criteria” in P the QTDB, we could get an estimate of the intercardiologist dif-
end, QRS end, and T end, and nearly for QRS onset. The “strict ferences in this database. It can be seen from Table IV that the
criteria” would not be accomplished by any of the mentioned error between the WT-based delineation system and each of the

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
MARTÍNEZ et al.: A WAVELET-BASED ECG DELINEATOR. EVALUATION ON STANDARD DATABASES 579

TABLE V
QTDB RECORDING STRATIFICATION ACCORDING TO DELINEATION ACCURACY

TABLE VI
DELINEATION RESULTS COMPARISON IN THE CSEDB. (N/R: NOT REPORTED; N/A: NOT APPLICABLE; #: NUMBER OF ANNOTATIONS)

referees is lower or similar to the error between cardiologists in mance due to the lack of a conveniently annotated database for
all points, excepting an appreciable bias in the T end, which is that kind of validation.
insignificant when we take into account the whole database. To
have a more reliable validation, it would be beneficial to have
V. CONCLUSION
annotations of more than one cardiologist in the whole QTDB.
Regarding Table V, the stratification of the records according We have presented and validated in this paper a wavelet-based
to the error is quite similar for all three methods in the case of ECG delineation system which performs QRS detection and
the T peak, but the WT approach gives the largest percentage provides as well the locations of the peak(s) of P, Q, R, S, R’,
(around 77%) of well-detected recordings in the T wave end. and T waves, and the P, QRS, and T wave boundaries using a
In this group, the mean error was negligible while the standard single analysis stage: the dyadic wavelet transform of the ECG
deviation of the error in T end determination reduces to three signal. The algorithm has been validated using several standard
samples (12 ms). annotated databases, with different sampling rates and a wide
The CSEDB manual annotations were performed with infor- diversity of morphologies, making a total of more than 980 000
mation from all leads, while the presented WT-based delineator beats for QRS detection, more than 790 000 beats for QRS onset
works on a single-channel basis. We observed that the multi- and more than 3500 beats other significant points.
lead delineation performance was extraordinarily sensitive to The results have been compared with those of other pub-
the chosen parameters of the multichannel rule. In Table VI, lished approaches and have shown that the developed algorithm
we attained clearly worse results than those reported in [16] provides a reliable and accurate delineation of the ECG signal,
for QRS end and T end, with slight differences in the other outperforming other algorithms, and with errors well within
points. Sahambi et al. reported in [20] lower error standard de- the observed intercardiologist variations. The most significant
viation in the CSE, especially in the QRS limits, although we improvement was found in the T wave end, which shows,
have no information about what data set and what one-lead to in general, the greatest difficulty in its determination, which
multilead rule were used. As for the admitted tolerances, the is reflected in the largest intercardiologist differences. The
“loose criteria” were accomplished by our delineation system. clue for this performance improvement is, according to our
The “strict criteria” were only essentially accomplished in the P understanding, the multiscale approach, which permits to
wave limits. Anyway, the significance of these results is limited attenuate noise at rough scales, and then to refine the precision
by the reduced number of beats ( ), and by the great depen- of the positions with the help of finer scales.
dence on the multilead approach. While the assessment of QRS detectors can be reasonably
Finally, the results in the referee-reviewed automatic QRS and reliably done with the existing standard databases, we
onset annotations of the EDB allowed to validate the delineation found difficulties for the evaluation of one-lead delineation
in an extensive collection of beats ( ). Using such a algorithms. The number of annotated beats is still insufficient to
large data set, the bias was negligible, and the standard deviation have a good representation of the possible morphologies found
of the error was around two sampling intervals. in ECG signals. Additionally, there is not, to our knowledge,
In addition to the previously discussed features, the presented any available database with “single-lead annotations”, which
method can detect and delineate the individual waves of the would allow a more effective assessment and comparison of
QRS complex. However, we were not able to assess its perfor- single-lead delineation algorithms.

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
580 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 51, NO. 4, APRIL 2004

APPENDIX [6] I. Murthy and G. D. Prasad, “Analysis of ECG from pole-zero models,”
IEEE Trans. Biomed. Eng., vol. 39, pp. 741–751, July 1992.
The amplitude thresholds in the presented algorithms can be [7] J. Vila, Y. Gang, J. Presedo, M. Fernandez-Delgado, and M. Malik, “A
grouped into three types. new approach for TU complex characterization,” IEEE Trans. Biomed.
Eng., vol. 47, pp. 764–772, June 2000.
First, the thresholds used to decide if a pair of maximum [8] M. Nygards and L. Sörnmo, “Delineation of the QRS complex using the
moduli with opposite sign can account for a wave: , , envelope of the ECG,” Med. Biol. Eng. Comput., vol. 21, pp. 538–547,
, (for QRS detection), and (in T/P wave delin- Mar. 1983.
[9] A. S. M. Koeleman, H. H. Ros, and T. J. van den Akker, “Beat-to-beat in-
eation). These thresholds are proportional to the RMS value of terval measurement in the electrocardiogram,” Med. Biol. Eng. Comput.,
the WT at the corresponding scales. For QRS detection, in each vol. 23, pp. 213–219, 1985.
excerpt of samples [10] A. Algra and H. L. B. C. Zeelenberg, “An algorithm for computer mea-
surement of QT intervals in the 24 hour ECG,” in Computers in Car-
diology. Los Alamitos, CA: IEEE Computer Society Press, 1987, pp.
RMS (A.1) 117–119.
RMS (A.2) [11] P. de Chazal and B. Celler, “Automatic measurement of the QRS onset
and offset in individual ECG leads,” presented at the 18th Ann. Int. Conf.
IEEE Eng. Med. Biol. Soc., Amsterdam, The Netherlands, 1996.
For T and P waves [12] I. K. Daskalov, I. A. Dotsinsky, and I. I. Christov, “Developments
in ECG acquisition, preprocessing, parameter measurement and
RMS (A.3) recording,” IEEE Eng. Med. Biol. Mag., vol. 17, pp. 50–58, 1998.
[13] I. Daskalov and I. Christov, “Electrocardiogram signal preprocessing for
RMS (A.4) automatic detection of QRS boundaries,” Med. Eng. Phys., no. 21, pp.
37–44, 1999.
where the RMS is measured in each interval between two con- [14] , “Automatic detection of the electrocardiogram T-wave end,” Med.
secutive QRS. Biol. Eng. Comput., vol. 37, pp. 348–353, 1999.
The morphology of QRS complexes and the type of T/P [15] J. G. C. Kemmelings, A. C. Linnenbank, S. L. C. Muilwijk, A. Sippens-
Groenewegen, A. Peper, and C. A. Grimbergen, “Automatic QRS onset
waves depend of the number of significant maximum moduli. and offset detection for body surface QRS integral mapping of ventric-
The thresholds to determine if they are significant, , ular tachycardia,” IEEE Trans. Biomed. Eng., vol. 41, pp. 830–836, Sept.
, , and are related to the amplitude of the global 1994.
[16] P. Laguna, R. Jané, and P. Caminal, “Automatic detection of wave
maximum modulus within the corresponding search window boundaries in multilead ECG signals: Validation with the CSE data-
(sw) base,” Comput. Biomed. Res., vol. 27, no. 1, pp. 45–60, Feb. 1994.
[17] G. Speranza, G. Nollo, F. Ravelli, and R. Antolini, “Beat-to beat mea-
(A.5) surement and analysis of the R-T interval in 24 h ECG Holter record-
ings,” Med. Biol. Eng. Comput., vol. 31, no. 5, pp. 487–494, Sept. 1993.
(A.6) [18] S. Meij, P. Klootwijk, J. Arends, and J. Roelandt, “An algorithm for
automatic beat-to-beat measurement of the QT-interval,” in Computers
(A.7) in Cardiology: IEEE Computer Society Press, 1994, pp. 597–600.
(A.8) [19] C. Li, C. Zheng, and C. Tai, “Detection of ECG characteristic points
using wavelet transforms,” IEEE Trans. Biomed. Eng., vol. 42, pp.
21–28, Jan. 1995.
A third group of thresholds are used to determine the [20] J. S. Sahambi, S. Tandon, and R. K. P. Bhatt, “Using wavelet transform
onset/end of QRS complex, T and P waves. They are propor- for ECG characterization,” IEEE Eng. Med. Biol., vol. 16, no. 1, pp.
tional to the amplitude of the WT at the first/last maximum 77–83, 1997.
[21] M. Bahoura, M. Hassani, and M. Hubin, “DSP implementation of
modulus of the complex or wave wavelet transform for real time ECG wave forms detection and heart
rate analysis,” Comput. Meth. Programs Biomed., no. 52, pp. 35–44,
if 1997.
(A.9)
if [22] P. Strumillo, “Nested median filtering for detecting T-wave offset in
ECGs,” Electron. Lett., vol. 38, no. 14, pp. 682–683, July 2002.
if [23] E. Soria-Olivas, M. Martínez-Sober, J. Calpe-Maravilla, J. F. Guerrero-
(A.10)
if Martínez, J. Chorro-Gascó, and J. Espí-López, “Application of adap-
tive signal processing for determining the limits of P and T waves in an
(A.11) ECG,” IEEE Trans. Biomed. Eng., vol. 45, pp. 1077–1080, Aug. 1998.
(A.12) [24] H. Vullings, M. Verhaegen, and H. Verbruggen, “Automated ECG seg-
mentation with dynamic time warping,” in Proc. 20th Ann. Int. Conf.
(A.13) IEEE Engineering in Medicine and Biology Soc., Hong Kong, 1998, pp.
163–166.
(A.14) [25] Z. Dokur, T. Olmez, E. Yazgan, and O. Ersoy, “Detection of ECG wave-
forms by neural networks,” Med. Eng. Phys., vol. 19, no. 8, pp. 738–741,
1997.
REFERENCES [26] W. Bystricky and A. Safer, “Modelling T-end in Holter ECG’s by 2-layer
perceptrons,” in Proc. Computers in Cardiology. Los Alamitos, CA:
[1] B.-U. Köhler, C. Hennig, and R. Orglmeister, “The principles of soft- IEEE Computer Society Press, 2002, vol. 29, pp. 105–108.
ware QRS detection,” IEEE Eng. Med. Biol. Mag., vol. 21, pp. 42–57, [27] L. Clavier, J.-M. Boucher, R. Lepage, J.-J. Blanc, and J.-C. Cornily, “Au-
Jan./Feb. 2002. tomatic P-wave analysis of patients prone to atrial fibrillation,” Med.
[2] O. Pahlm and L. Sörnmo, “Software QRS detection in ambulatory mon- Biol. Eng. Comp., vol. 40, no. 1, pp. 63–71, Jan. 2002.
itoring – A review,” Med. Biol. Eng. Comp., vol. 22, pp. 289–297, 1984. [28] G. B. Moody and R. G. Mark, “The MIT-BIH arrhythmia database
[3] F. Gritzali, “Toward a generalized scheme for QRS detection in ECG on CD-ROM and software for use with it,” in Computers in Cardi-
waveforms,” Signal Processing, vol. 15, pp. 183–192, 1988. ology. Los Alamitos, CA: IEEE Computer Society Press, 1990, pp.
[4] G. M. Friesen et al., “A comparison of the noise sensitivity of nine QRS 185–188.
detection algorithms,” IEEE Trans. Biomed. Eng., vol. 37, pp. 85–98, [29] P. Laguna, R. Mark, A. Goldberger, and G. Moody, “A database for eval-
Jan. 1990. uation of algorithms for measurement of QT and other waveform inter-
[5] L. Sörnmo, “A model-based approach to QRS delineation,” Comput. vals in the ECG,” in Computers in Cardiology 1997. Los Alamitos,
Biomed. Res., vol. 20, pp. 526–542, 1987. CA: IEEE Computer Society Press, 1997, pp. 673–676.

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.
MARTÍNEZ et al.: A WAVELET-BASED ECG DELINEATOR. EVALUATION ON STANDARD DATABASES 581

[30] A. Taddei, G. Distante, M. Emdin, P. Pisani, G. B. Moody, C. Zeelen- Rute Almeida was born in Porto, Portugal, in 1979.
berg, and C. Marchesi, “The European ST-T database: Standards for She received a 4-year degree in mathematics applied
evaluating systems for the analysis of ST-T changes in ambulatory elec- to technology from the Faculty of Sciences, Univer-
trocardigraphy,” Eur. Heart J., vol. 13, pp. 1164–1172, 1992. sity of Porto (FCUP), in 2000. Since October 2001,
[31] J. L. Willems et al., “A reference data base for multilead electrocardio- she is working towards the Ph.D. degree in the Ap-
graphic computer measurement programs,” J. Amer. Coll. Cardiol., vol. plied Mathematics Department at FCUP, supported
10, no. 6, pp. 1313–1321, 1987. by a grant from Foundation for Science and Tech-
[32] J. P. Martínez, S. Olmos, and P. Laguna, “Evaluation of a wavelet-based nology (Portugal) and European Social Fund.
ecg waveform detector on the qt database,” in Computers in Cardi- She was with the Autonomic Function Study
ology. Los Alamitos, CA: IEEE Computer Society Press, 2000, pp. Center from Hospital S. João between March and
81–84. October 2000, working in methods for automatic
[33] S. Mallat and S. Zhong, “Characterization of signals from multiscale delineation of the ECG. Her main research interests are in time-scale methods
edge,” IEEE Trans. Pattern Anal. Machine Intell., vol. 14, pp. 710–732, and the automatic analysis of the ECG; namely, the study of ventricular
July 1992. repolarization.
[34] J. S. Sahambi, S. Tandon, and R. K. P. Bhatt, “Wavelet based ST-segment
analysis,” Med. Biol. Eng. Comput., vol. 36, no. 9, pp. 568–572, Sept.
1998.
[35] S. Mallat, “Multifrequency channel decompositions of images and
wavelet models,” IEEE Trans. Acoust. Signal Processing, vol. 37, pp.
2091–2110, Dec. 1989.
Salvador Olmos (A’01–M’03) was born in Va-
[36] A. Cohen and J. Kovačević, “Wavelets: The mathematical background,”
Proc. IEEE, vol. 84, pp. 514–522, Apr. 1996. lencia, Spain, in 1969. He received the Industrial
[37] M. Akay, Detecton and Estimation Methods for Biomedical Sig- Engineering degree and the Ph.D. degree from the
nals. San Diego, CA: Academic, 1996. Polytechnic University of Catalonia, Catalonia,
[38] N. V. Thakor, J. G. Webster, and W. J. Tompkins, “Estimation of QRS Spain, in 1993 and 1998, respectively.
He is currently an Associate Professor of Signal
complex power spectrum for design of a QRS filter,” IEEE Trans.
Biomed. Eng., vol. BME-31, pp. 702–706, Nov. 1984. Processing and Communications in the Department
of Electronics Engineering and Communications at
[39] G. B. Moody and R. G. Mark, “Development and evaluation of a 2-lead
ECG analysis program,” in Computers in Cardiology. Los Alamitos, Zaragoza University, Zaragoza, Spain. From August
2000 to August 2001, he was with the Department of
CA: IEEE Computer Society Press, 1982, pp. 39–44.
[40] V. X. Afonso, W. J. Tompkins, T. Q. Nguyen, and S. Luo, “ECG beat Electroscience, Lund University, Lund, Sweden, sup-
detection using filter banks,” IEEE Trans. Biomed. Eng., vol. 46, pp. ported by a post-doctoral research grant from Spanish Government. His profes-
sional research interests are in signal processing of biomedical signals.
192–201, Feb. 1999.
[41] J. Lee, K. Jeong, J. Yoon, and M. Lee, “A simple real-time QRS detection
algorithm,” in Proc. 18th Ann. Int. Conf. IEEE Engineering in Medicine
and Biology Soc., 1996, pp. 1396–1398.
[42] P. S. Hamilton and W. Tompkins, “Quantitative investigation of QRS
detection rules using the MIT/BIH arrhythmia database,” IEEE Trans
Biomed. Eng., vol. BME-33, pp. 1157–1165, Dec. 1986. Ana Paula Rocha was born in Coimbra, Portugal,
[43] J. Pan and W. J. Tompkins, “A real-time QRS detection algorithm,” in 1957. She received the Applied Mathematics de-
IEEE Trans. Biomed. Eng., vol. BME-32, pp. 230–236, Mar. 1985. gree and the Ph.D. degree in applied mathematics,
[44] R. Poli, S. Cagnoni, and G. Valli, “Genetic design of optimum linear systems theory and signal processing, from the Fac-
and nonlnear QRS detectors,” IEEE Trans. Biomed. Eng., vol. 42, pp. ulty of Sciences, University of Porto (FCUP), Porto,
1137–1141, Nov. 1995. Portugal, in 1980 and 1993, respectively.
[45] J. Moraes, M. Freitas, F. Vilani, and E. Costa, “A QRS complex detec- She is currently an Auxiliar Professor in the
tion algorithm using electrocardiogram leads,” in Computers in Cardi- Department of Applied Mathematics at FCUP. Her
ology Los Alamitos, CA, 2002, vol. 29, pp. 205–208. research interests are in biomedical signals and
[46] P. Hamilton, “Open source ECG analysis,” in Computers in Cardi- system analysis (EMG, cardiovascular systems
ology. Los Alamitos, CA: IEEE Computer Society Press, 2002, vol. analysis, and autonomic nervous system character-
29, pp. 101–104. ization), time-frequency/time-scale signal analysis, point processes spectral
[47] R. Jane, A. Blasi, J. García, and P. Laguna, “Evaluation of an automatic analysis, and data treatment and interpretation.
detector of waveforms limits in holter ECG with the QT database,” in
Computers in Cardiology. Los Alamitos, CA: IEEE Computer Society
Press, 1997, pp. 295–298.
[48] “Recomendations for measurement standards in quantitative electrocar-
diography,” in Eur. Heart J.: The CSE Working Party, 1985, vol. 6, pp.
815–825.
Pablo Laguna (M’02) was born in Jaca, Spain, in
1962. He received the M.S. degree in physics and the
Ph.D. degree from the University of Zaragoza (UZ),
Juan Pablo Martínez was born in Zaragoza, Zaragoza, Spain, in 1985 and 1990, respectively.
Aragon, Spain, in 1976. He received the M.Sc. The Ph.D. degree dissertation was developed at the
degree in telecommunication engineering (with Biomedical Engineering Division of the Institute
highest honors) from the University of Zaragoza of Cybernetics (IC), Polytechnic University of
(UZ), in 1999. Catalonia (UPC-CSIC), Barcelona, Spain.
From 1999 to 2000, he was with the Department He is currently an Associate Professor in the De-
of Electronic Engineering and Communications, UZ, partment of Electronic Engineering and Comunica-
as a Research Fellow. Since 2000, he is an Assis- tions, UZ, and researcher in the Aragon Institute of
tant Professor in the same department. He is also in- Engineering Research (I3A), UZ. From 1987 to 1992, he worked as Assistant
volved as Researcher with the Aragon Institute of En- Professor in the Department of Control Engineering at the Politecnic University
gineering Research (I3A), UZ. His professional re- of Catalonia, Barcelona, Spain and as a Researcher at the Biomedical Engi-
search activity lies in the field of biomedical signal processing, with main in- neering Division of the Institute of Cybernetics. His professional research inter-
terest in signals of cardiovascular origin. ests are in Signal Processing, in particular applied to biomedical applications.

Authorized licensed use limited to: Universidad de Zaragoza. Downloaded on February 22,2023 at 15:43:49 UTC from IEEE Xplore. Restrictions apply.