Impact of sampling rate reduction on automatic ECG delineation.
ABSTRACT Electrogram (EGM) delineation is an increasingly important task to be performed in implantable cardiac devices such as pacemakers and defibrillators. Reliable detection and classification of EGM components might help to minimize the risk of false detections. Efforts are therefore undertaken to examine whether existing ECG delineators can be adapted for the delineation of EGMs. One issue to be solved is the low sampling rate at which EGMs are acquired. In this study we investigate performance degradation of an existing wavelet-based ECG delineator by a stepwise reduction of the sampling rate. It is shown that for signals sampled at 1 kHz, no significant performance degradation occurs in P or T wave delineation. The performance of QRS delineation is affected only at the lowest sampling rate of 62.5 Hz. For signals originally sampled at 250 Hz, no degradation in delineation performance is observed. It is concluded that the automatic delineation of ECGs can be performed at sampling rates as low as 62.5 Hz and that the low sampling rate does not significantly degrade the reliability of automatic delineation.
- [Show abstract] [Hide abstract]
ABSTRACT: Arrhythmia is one of the preventive cardiac problems frequently occurs all over the globe. In order to screen such disease at early stage, this work attempts to develop a system approach based on registration, feature extraction using discrete wavelet transform (DWT), feature validation and classification of electrocardiogram (ECG). This diagnostic issue is set as a two-class pattern classification problem (normal sinus rhythm versus arrhythmia) where MIT-BIH database is considered for training, testing and clinical validation. Here DWT is applied to extract multi-resolution coefficients followed by registration using Pan Tompkins algorithm based R point detection. Moreover, feature space is compressed using sub-band principal component analysis (PCA) and statistically validated using independent sample t-test. Thereafter, the machine learning algorithms viz., Gaussian mixture model (GMM), error back propagation neural network (EBPNN) and support vector machine (SVM) are employed for pattern classification. Results are studied and compared. It is observed that both supervised classifiers EBPNN and SVM lead to higher (93.41% and 95.60% respectively) accuracy in comparison with GMM (87.36%) for arrhythmia screening.Journal of Medical Systems 06/2010; 36(2):677-88. · 1.37 Impact Factor
- [Show abstract] [Hide abstract]
ABSTRACT: Low-power design has become a key technology for battery-power biomedical devices in Wireless Body Area Network. In order to meet the requirement of low-power dissipation for electrocardiogram related applications, a down-sampling QRS complex detection algorithm is proposed. Based on Wavelet Transform (WT), this letter characterizes the energy distribution of QRS complex corresponding to the frequency band of WT. Then this letter details for the first time the process of down-sampled filter design, and presents the time and frequency response of the filter. The algorithm is evaluated in fixed point on MIT-BIH and QT database. Compared with other existing results, our work reduces the power dissipation by 23%, 61%, and 72% for 1 ×, 2 ×, and 3 × down-sampling rate, respectively, while maintaining almost constant detection performance.IEEE Signal Processing Letters 05/2013; 20(5):515-518. · 1.64 Impact Factor
Impact of Sampling Rate Reduction on Automatic ECG Delineation
Fernando Simon∗, Juan Pablo Martinez,
Communication Technologies Group, I3A,
Mar´ ıa de Luna 1, 50018 Zaragoza Spain
Bart van Grinsven∗, Cyril Rutten∗,
Medtronic Bakken Research Center B.V.
Endepolsdomein 5, 6229 GW Maastricht, The
Abstract—Electrogram (EGM) delineation is an increasingly
important task to be performed in implantable cardiac devices
such as pacemakers and defibrillators. Reliable detection and
classification of EGM components might help to minimize the
risk of false detections. Efforts are therefore undertaken to
examine whether existing ECG delineators can be adapted
for the delineation of EGMs. One issue to be solved is the
low sampling rate at which EGMs are acquired. In this study
we investigate performance degradation of an existing wavelet-
based ECG delineator by a stepwise reduction of the sampling
rate. It is shown that for signals sampled at 1 kHz, no significant
performance degradation occurs in P or T wave delineation.
The performance of QRS delineation is affected only at the
lowest sampling rate of 62.5 Hz. For signals originally sampled
at 250 Hz, no degradation in delineation performance is
It is concluded that the automatic delineation of ECGs can
be performed at sampling rates as low as 62.5 Hz and that the
low sampling rate does not significantly degrade the reliability
of automatic delineation.
Most of today’s implantable cardiac devices such as pace-
makers and defibrillators are equipped with algorithms that
make decisions based on sensed cardiac events. The intro-
duction of real-time digital signal processing technology has
added a new dimension to cardiac signal analysis. Besides
conventional event-based decisions, digital signal processing
enables morphological analysis of the electrogram (EGM),
including characterization of the individual components. This
development not only enables new sensing applications but
may also reduce false detections.
Since electrograms are in nature similar to surface ECGs,
existing ECG delineation techniques are good candidates
to analyze EGMs. However, to optimize delineation per-
formance there are several issues that need specific atten-
tion. First, the ECG and EGM have some morphological
differences, i.e. characteristic components of the ECG are
not necessarily present in the EGM. Moreover, the spectral
content of the EGM differs from that of the ECG. Since this
affects any filtering-based approach including wavelet based
delineation, spectral analysis is required to reveal the relevant
differences. Finally, there are several technical constraints
that lead to marked differences between electrograms and
∗This work was supported by Medtronic Bakken Research Center
B.V. and by projects TEC2004-05263-C02-02/TCM from MCYT/FEDER,
CB06/01/0062 from ISCIII, and Grupos Consolidados GTC T30 from DGA
the surface ECG. Most critical is the low sampling rate
at which EGMs are typically acquired due to the lim-
ited power budget and storage capacity. The present work
aims to investigate this issue by studying the degradation
in delineation performance when applying a wavelet-based
ECG delineator recently described by Martinez et al.  on
datasets with stepwise reduced sampling rates. In addition,
initial morphological and spectral differences between the
ECG and EGM are studied.
II. MATERIALS AND METHODS
A. Signal Databases
Automatic delineation was performed on two publicly
available signal databases.
1) PTB Diagnostic ECG Database (PTBDB): contains
549 high-resolution 15-lead ECG recordings (12 standard
leads together with Frank XYZ leads) sampled at 1 kHz,
having a resolution of 0.5 µV and being of variable duration
, . The recordings include 54 healthy controls and 240
patients with different cardiopathies. The delineation was
performed on beats from the lead II signal, manually selected
and annotated for QRS onset and T wave end by 5 experts
individually as described by Christov et al. . The median
of the 5 expert annotations was considered the gold standard.
2) QT Database (QTDB): contains 105 two-lead (mostly
MLII and V5) ECG recordings of 15 minutes each acquired
at a sampling rate of 250 Hz . In each recording, a
minimum of 30 consecutive beats were annotated by an
expert cardiologist for the onset, peak, and end of the P, QRS,
T, and (where present) U wave. P waves were annotated
on 3335 beats, QRS delineation on 3623 beats, and T
wave delineation on 3542 beats. Eleven recordings were also
annotated by a second cardiologist.
B. QRS detection and wave delineation
Automatic waveform delineation of the ECG was per-
formed by a multiscale wavelet-based ECG delineator pre-
viously described and validated . Detection of all fiducial
points (onset, peak and end) of the ECG components was
based on the quadratic spline wavelet transform producing
smoothed ECG derivatives at four dyadic scales.
Proceedings of the 29th Annual International
Conference of the IEEE EMBS
Cité Internationale, Lyon, France
August 23-26, 2007.
1-4244-0788-5/07/$20.00 ©2007 IEEE 2587
C. Signal downsampling
To evaluate the effect of sampling frequency reduction
on delineation performance, the signals in the PTBDB were
consecutively downsampled by a factor 2, 4, 8 and 16, thus
obtaining ECG signals sampled at 500, 250, 125 and 62.5
Hz, respectively. The QTDB signals, originally sampled at
250 Hz, were downsampled only by a factor 2 and 4.
D. Evaluation of the delineation performance
Delineation performance was evaluated as the difference
(mean ± SD) between automatic detection and median expert
annotation. The inter-expert variability was calculated as the
mean ± SD between the individual expert annotation and
the median (PTBDB) or mean (QTDB) expert annotation.
The delineation tolerances reported in  are also used for
In the PTBDB, the delineator was applied to lead II. In the
QTDB, the delineation was run on leads MLII and V5. For
each point, the annotation with lowest error was selected.
In automatic ECG delineation, large errors are usually
associated with missed waves, whereas small errors relate
to the delineation accuracy in waves with morphologies
correctly identified. Because this study will primarily focus
on reduction of delineation accuracy at lower sampling rates,
performance was also estimated exclusively on correctly
delineated beats at the highest sampling rate. Therefore beats
with a QRS onset error >10 ms and a T wave end error
exceeding 30 ms were excluded.
E. Spectral analysis
Power spectra were obtained by computing the average
fast-Fourier transform on a selected number of beats.
1) ECG: 300 normal sinus rhythm beats were randomly
selected from the MIT-BIH Arrhythmia Database, consisting
of 48 fully annotated recordings sampled at 360 Hz with an
11-bit resolution over a 10 mV amplitude range .
2) EGM: 18 normal sinus rhythm beats were selected
from an electrogram database containing EGM from im-
plantable cardioverter defibrillators and comprised recordings
of heart beat rhythm prior and after therapy delivery. These
signals were recorded unipolarly (between the device and the
ring of the ventricular lead) at 64 Hz with a 12-bit resolution
over a 8 mV range.
Automatic delineation was successful in 471 out of 549
beats. We excluded 37 records in which the annotated beat
was the first one, 35 beats due to false-positive detections
and 6 because of false-negative detections of QRS fiducial
points. In Table I (first and second column), the delineation
errors at the various sampling rates are shown for QRS onset
and T wave end. The inter-expert variability is included for
The third and fourth column summarize the results after
omitting measurements with large delineation error at 1 kHz.
In this situation, QRS onset was performed on 375 records
DELINEATION ERRORS IN RECORDS OF THE PTBDB (IN MS)
With error thresholding
Q R S onset T end Q R S onsetT end
−4. 5± 11. 02. 9± 33. 1
−2. 6± 4. 51. 5± 12. 3
−4. 4± 10. 72. 9± 33. 1
−2. 6± 4. 5 1. 5± 12. 3
−4. 4± 10. 72. 9± 33. 1
−2. 6± 4. 51. 5± 12. 3
−4. 6± 11. 52. 8± 33. 3
−3. 0± 7. 01. 3± 12. 6
−8. 1± 14. 1 3. 2± 33. 2
−7. 9± 13. 21. 6± 13. 6
0. 0± 3. 20. 0± 8. 0
−0. 5± 3. 1 0. 1± 7. 5
− ± 6. 5
− ± 30. 6
− ± 6. 5
− ± 30. 6
(a) fs = 500 Hz. (b) fs = 250 Hz.
(c) fs = 125 Hz.(d) fs = 62.5 Hz.
Fig. 1. Bland-Altman plots for error measurement at the PTBDB.
and T wave detection on 390 records correctly delineated at
To evaluate delineation performance on other ECG compo-
nents, we compensated for the absence of expert annotations
by taking the automatic detections at 1 kHz as a reference.
This enabled the additional delineation of P wave onset, P
wave end, QRS end and T wave onset (Table II).
The Bland-Altman plots in Fig. 1 illustrate the agreement
between QT interval estimation for downsampled signals and
In this database, delineation of the P wave was successful
in 94.96% (n=3335), QRS delineation in 99.99% of the cases
(3622 out of 3623 beats) and T wave delineation in 99.78%
(3534 out of 3542 beats). These numbers did not change as
sampling rates were reduced.
Table III summarizes the results of the automatic delin-
eation at all sampling rates compared to expert annotations.
The agreement between QT measurements at original and
downsampled signals is shown in Fig. 2 by means of Bland-
Fig. 2.Bland-Altman plots for error measurement at the QTDB.
(a) ECG morphology.(b) ECG Power spectral density.
(c) EGM morphology.(d) EGM Power spectral density.
morphology of a single beat of normal sinus rhythm ECG. Visible are the
P wave, QRS complex and T wave; (b) Power spectral density function
(PSD) computed on the average of 300 beats of normal sinus rhythm ECG.
(c) Morphology of a single beat of normal sinus rhythm EGM recorded
unipolarly from a ventricular lead. Visible components include the QRS
complex and T wave. (d) PSD computed on the average of 18 beats of
normal sinus rhythm EGM acquired from a unipolar ventricular lead.
Time and frequency plots for the ECG and EGM. (a) Typical
C. Morphological and spectral analysis
Time and frequency plots for ECG and EGM are shown in
Fig 3. The time series in the panel a) and c) show the typical
morphology of a single beat of ECG and EGM recorded
during normal sinus rhythm. Panels b) and d) contain the
power spectral density functions (PSD) for the ECG and
EGM, and their main components.
IV. DISCUSSION AND CONCLUSION
The main objective of the present work was to quantify the
impact of reduced sampling rates, as used for EGM recording
in implantable devices, on the performance of a wavelet-
based ECG delineator.
Table I shows that the error in QRS onset detection
remains essentially constant when downsampling from the
original frequency (1000 Hz) down to 125 Hz, but it
markedly increases at 62.5 Hz (the QRS onset detection is
located on average 4 ms before those at higher sampling rates
and the standard deviation increases from 11 ms to 14 ms).
When focusing on beats well delineated at 1000 Hz, a
degradation in QRS onset delineation is already observed
at 125 Hz. On the other hand, the T wave end delineation
error is similar at all sampling frequencies. The error is
substantially higher in the case of T wave end than in QRS
onset, as could be expected.
The degradation in delineation performance of the QRS
onset can be attributed to the loss of high-frequency com-
ponents of the QRS complex at lower sampling rates. The
T wave end error does not increase significantly since the
low-frequency content of the T wave is preserved even at a
sampling rate of 62.5 Hz.
For ECG fiducial points other than the QRS onset and T
wave end, similar conclusions can be derived from Table II.
QRS delineation is most notably affected by the sampling
The results of the QTDB delineation show that the error
remains fairly constant for all fiducial points. In contrast to
the PTBDB, the SD of the QRS onset delineation does not
increase at 62.5 Hz, as shown in Table III. As can be seen in
Fig. 2, QT interval estimations at 62.5 Hz are less stable than
at 125 Hz. However, this increased variability is negligible
compared to the SD at the original sampling frequency.
The differences in the results obtained in both databases
could be explained by the different equivalent recording
cutoff frequencies that were used (500 Hz in the PTBDB
vs 100 Hz in the QTDB).
Acceptable limits for automatic delineation errors are
usually based on comparison with inter-expert variability.
In the case of the PTBDB the variability between the five
expert annotators is much lower than the error obtained
by automatic delineation, as well as the inter-expert error
found in other studies  or the tolerances proposed in .
This may be explained as being the result of the three-
round feedback procedure that was used to reduce the largest
discrepancies between expert annotators . In the PTBDB,
T wave end delineation errors are in the order of the CSE-
proposed tolerances, while QRS onset delineation errors are
larger even at the original sampling frequency, and even more
pronounced at 62.5 Hz. Most large QRS onset errors are
due to misinterpretation of small Q waves. However, in the
QTDB, the differences of automatic delineation, even at 62.5
Hz, are below the inter-cardiologist differences at the same
From the results obtained in the PTBDB and QTDB ECG
databases, it can be concluded that the delineation remains
essentially unaffected for sampling frequencies higher than
125 Hz, while a slight but not negligible loss in delineation
accuracy of the waves with higher frequency content, i.e. the
QRS complex is observed when sampling at 62.5 Hz.
Before adapting the wavelet-based ECG delineator to
EGMs, it is important to identify morphological and spectral
differences between both signals. However, the shape of
the EGM signal is strongly affected by the lead configura-
tion. Unipolar EGMs differ from bipolar EGMs and signals
DELINEATION ERRORS WITH RESPECT TO THE AUTOMATIC DETECTIONS AT 1000 HZ OF THE PTBDB (IN MS)
P onsetP endQ R S onset Q R S endT onsetT end
5000. 0± 0. 10. 0± 0. 10. 1± 2. 40. 0± 1. 0 0. 0± 0. 10. 0± 0. 1
−0. 1± 2. 90. 0± 0. 40. 1± 2. 50. 1± 1. 00. 0± 0. 30. 0± 0. 3
1250. 0± 5. 40. 2± 4. 0
−0. 1± 6. 70. 1± 7. 70. 0± 4. 70. 0± 3. 2
−0. 7± 14. 4 0. 1± 11. 5
−3. 6± 16. 33. 4± 17. 5
−0. 4± 8. 9 0. 4± 7. 4
DELINEATION ERRORS WITH EXPERT ANNOTATIONS AS REFERENCE OF THE QTDB (IN MS).
P onsetP end Q R S onsetQ R S endT peakT end
2502. 0± 15. 01. 9± 12. 94. 6± 7. 70. 8± 8. 70. 2± 13. 9
−1. 6± 18. 1
1252. 2± 14. 7 1. 9± 12. 74. 7± 7. 70. 9± 8. 50. 3± 13. 9
−1. 5± 18. 5
62. 52. 5± 14. 52. 4± 13. 64. 0± 7. 61. 0± 8. 40. 1± 13. 8
−1. 5± 17. 8
5. 3± 11. 1
−0. 1± 12. 25. 1± 15. 92. 1± 22. 4
− ± 10. 2
− ± 12. 7
− ± 6. 5
− ± 11. 6
− ± 30. 6
recorded from a ventricular lead are very distinct from those
recorded atrially. We therefore consider only one type of
configuration here: the unipolar EGM of a ventricular lead.
The time domain plots in Fig. 3 show that morphological
differences can be found in the lack or presence of a P wave
and the individual deflections of the QRS complex. In the
spectral domain, the PSD plots show that the peak signal
power is similarly distributed for the entire beat as well as
for QRS complex and T wave components. This suggests
that a wavelet-based approach for EGM delineation, similar
to that described in , may be effective for EGM signals.
 J.P. Mart´ ınez, R. Almeida, S. Olmos, A.P. Rocha and P. Laguna, “A
Wavelet-Based ECG Delineator: Evaluation on Standard Database.”
IEEE Trans. Biomed. Eng.,, vol. 51(4), pp. 570–581, 2004.
 A.L. Goldberger, LAN Amaral, L. Glass, J.M. Hausdorff, P.Ch Ivanov,
R.G. Mark, J.E. Mietus, G.B. Moody, C.K. Peng, H.E. Stanley.
“PhysioBank, PhysioToolkit, and PhysioNet: Components of a New
Research Resource for Complex Physiologic Signals”. Circulation
101(23):e215-e220; 2000 (June 13).
 R. Bousseljot, D. Kreiseler, A. Schnabel, “Nutzung der EKG-
Signaldatenbank CARDIODAT der PTB ¨ uber das Internet”. Diomedi-
zinische Technik, Band 40, Erg¨ anzungsband 1 (1995) S 317.
 D. Kreiseler, R. Bousseljot, “Automatisierte EKG-Auswertung mit
Hilfe der EKG-Signaldatenbank CARDIODAT der PTB”. Biomedi-
zinische Technik, Band 40, Erg¨ anzungsband 1 (1995) S 319.
 I. Christov, I. Dotsinsky, I. Simova, R. Prokopova, E. Trendafilova
and S.Naydenov. “Dataset of manually measured QT intervals in the
electrocardiogram”, in Biomedical Engineering Online 2006, 5:31.
 P. Laguna, R.G. Mark, A. Goldberger and G.B. Moody. “A Database
for Evaluation of Algorithms for Measurement of QT and Other
Waveform Intervals in the ECG”. In Proc. Computers in Cardiology,
vol. 24. pp. 673-676, 1997.
 “Recommendations for measurement standards in quantitative electro-
cardiography”, in Eur. Heart J.: The CSE Working Party, 1985, vol.
6, pp. 815–825.