SoftwareX 13 (2021) 100639

Contents lists available at ScienceDirect

SoftwareX
journal homepage: www.elsevier.com/locate/softx

Original software publication

ECGdeli - An open source ECG delineation toolbox for MATLAB

∗
N. Pilia , C. Nagel, G. Lenis, S. Becker, O. Dössel, A. Loewe
Karlsruhe Institute of Technology (KIT), Institute of Biomedical Engineering, Fritz-Haber-Weg 1, 76131 Karlsruhe, Germany

article info a b s t r a c t

Article history: The electrocardiogram (ECG) is a standard cost-efficient and non-invasive tool for the early detection
Received 23 July 2020 of various cardiac diseases. Quantifying different timing and amplitude features of and in between
Received in revised form 1 November 2020 the single ECG waveforms can reveal important information about the underlying (dys-)function of
Accepted 27 November 2020
the heart. Determining these features requires the detection of fiducial points that mark the on- and
MSC: offset as well as the peak of each ECG waveform (P wave, QRS complex, T wave). Manually setting
92C55 these points is time-consuming and requires a physician’s expert knowledge. Therefore, the highly
94A12 modular ECGdeli toolbox for MATLAB was developed, which is capable of filtering clinically recorded
12-lead ECG signals and detecting the fiducial points, also called delineation. It is one of the few
Keywords:
open toolboxes offering ECG delineation for P waves, T Waves and QRS complexes. The algorithms
ECG waveform boundary detection
provided were evaluated with the QT database, an ECG database comprising 105 signals with fiducial
ECG delineation
12 lead ECG processing points annotated by clinicians. The median difference between the fiducial points set by the boundary
detection algorithm and the clinical annotations serving as a ground truth is less than 4 samples
(16 ms) for the P wave and the QRS complex markers. The T wave onset, peak and offset were detected
with a median difference of 5, 2 and 7 samples, respectively. Results were compared to two free
algorithms available on PhysioNet. Our results show that ECGdeli can reliably detect P waves, QRS
complexes and T waves. Thus, it can contribute to diagnose specific cardiac diseases by analyzing the
ECG signal. As ECGdeli is published under GNU GPLv3 and thanks to its modularity, it can be used to
extend existing algorithms or as a benchmark for new algorithms.
© 2020 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license
(http://creativecommons.org/licenses/by/4.0/).

Code metadata
This software repository is listed in Zenodo with the DOI 10.5281/zenodo.3977971 [1].

Current code version v1.0.2

Permanent link to code/repository used for this code version https://github.com/ElsevierSoftwareX/SOFTX-D-20-00010
Code Ocean compute capsule doi 10.24433/CO.9115981.v1
Legal Code License GNU GPLv3
Code versioning system used git
Software code languages, tools, and services used MATLAB
Compilation requirements, operating environments & dependencies MATLAB and the following toolboxes: Signal Processing Toolbox, Image
Processing Toolbox, Statistics and Machine Learning Toolbox, Wavelet Toolbox
If available Link to developer documentation/manual https://github.com/KIT-IBT/ECGdeli/blob/master/ README.md
Support email for questions [email protected]

1. Motivation and significance data, too. Especially the electrocardiogram (ECG) as a cheap and
available standard heart activity monitoring device [2] is used
As the amount of recorded data increases in all fields, auto- in medicine around the world. Automatic ECG processing can be
matic processing is inevitably needed. This is the case for health applied to batch process many ECGs in a short amount of time. In
this way, manual annotation work of physicians can be avoided.
∗ Corresponding author. Furthermore, retrospective analysis especially plays an important
E-mail address: [email protected] (N. Pilia). role in clinical research to reveal effects of new drugs on the

https://doi.org/10.1016/j.softx.2020.100639
2352-7110/© 2020 The Author(s). Published by Elsevier B.V. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
N. Pilia, C. Nagel, G. Lenis et al. SoftwareX 13 (2021) 100639

Fig. 1. Flowchart of QRS detection algorithm. WT: wavelet transform, SWDN : detail coefficients of the SWT at level N. FPT: fiducial point table, twindow : search window
length, candQRSx : QRSx candidates, n: sample index, Qj : threshold, Qj,k : adapted threshold.

heart, optimize diagnosis and treatment strategies of cardiac dis- research use, freely available on github. We call this open source
eases [3–8]. The automatic analysis of the ECG usually starts with toolbox ECGdeli. The aim of this work is not to present new
determining wave types (P waves, T waves and QRS complexes), approaches for wave delineation as the algorithms or slightly
followed by finding the peaks and boundaries, i.e., wave onsets modified versions were already used in several studies [11–15].
and offsets. Starting from that, further temporal parameters, like Rather, this work should be seen as accompanying documen-
QT intervals, RR intervals, etc. can be derived. As well, these an- tation of the open software resource. We will describe shortly
notations are needed to calculate further features like amplitudes, the main ideas behind the algorithms and compare them to two
slopes, and parameters like heart rate variability [9]. existing open implementations on PhysioNet: ecg-kit [16] and
In this work, we present a toolbox for MATLAB comprising ecgpuwave [17]. The algorithms for comparison, especially ecg-
algorithms for ECG pre-processing and ECG wave delineation for kit, offer more options for evaluation and data import/export,
2
N. Pilia, C. Nagel, G. Lenis et al. SoftwareX 13 (2021) 100639

Fig. 2. T wave (A) and P wave (B) detection for signal sel100 from [10]. Extrema of WT in line 3 correspond to the T and P wave boundaries and peaks shown in
line 4 in A and B, respectively.

however, at the cost of a more complicated non-modular frame- In the following, we present a short description of the sin-
work or perform worse. The advantages of ECGdeli in comparison gle algorithms provided with the toolbox and discuss the main
with the existing implementations are a simple and modular advantages of ECGdeli.
design concentrating on the essential algorithms needed for ECG
evaluation yielding an easy to use software; ECGdeli focuses 2. Software description
on one concrete purpose: ECG wave detection and delineation.
File input/output, post processing like feature calculation and a 2.1. ECG preprocessing
graphical user interface are intentionally not included. Thus, users
can easily integrate the single algorithms or the whole toolbox The ECG preprocessing functionalities are shown in Fig. 5,
in their evaluation frameworks, replace single algorithms, all by lines one to four and include a baseline removal technique [15]
using few clear and standardized interfaces. ECG_Baseline_Removal.m, a bandpass filter
3
N. Pilia, C. Nagel, G. Lenis et al. SoftwareX 13 (2021) 100639

Fig. 3. A: Flowchart of the T wave detection algorithm. B: Flow chart of the P wave detection algorithm. WT: wavelet transform, AUC: area under the curve, min:
minimum, max: maximum, fid: fiducial.

ECG_High_Low_Filter.m and a notch filter Notch_Filter.m, To detect intervals containing the QRS complexes, a threshold
as well as a method for correcting the electrical isoline Iso- based approach is chosen. This threshold is adapted dependent
line_Correction.m. All preprocessing steps are performed on a changing time windows which is applied to SWDN and a
lead by lead.
signal dependent statistical measure. Among all sets of solutions
calculated with different thresholds, a voting algorithm yields the
2.2. QRS detection
most consistent solution.
In the QRS detection algorithm QRS_detection.m (an The R peak is annotated at the position with the highest
overview is shown in Fig. 1), the input signal is first bandpass amplitude within the bounds of the QRS interval. The Q- and S
filtered (5 Hz to 100 Hz) in order to reduce high frequency noise, peak are subsequently marked at the minimum amplitude in the
baseline wander, to attenuate P and T waves and to obtain QRS regions from QRS onset to R peak and from R peak to QRS offset,
complexes with pronounced Q and S waves. Then, a phase-free respectively. The onset and the offset of the QRS complex are
stationary wavelet transformation (SWT) using the Haar wavelet
afterwards marked 20 ms prior the Q-peak and 20 ms after the S-
is applied [12]. The detail coefficients SWDN of the SWT are
calculated at the level including frequencies around 45 Hz, which peak, respectively. In the end, all detected peaks and boundaries
is the frequency range where spectral components of the QRS are stored in a fiducial point table (FPT) in columns five to eight.
complex being relevant for delineation can be found [18]. The structure is shown in Table 1.
4
N. Pilia, C. Nagel, G. Lenis et al. SoftwareX 13 (2021) 100639

2.3. T wave detection

The implemented T wave detection method T_Detection.m

relies on QRS detection and is performed for each lead separately.
Fig. 3A shows an overview of the algorithm. To detect the T
waves, the following pre-processing steps are needed: QRS re-
placement, filtering and SWT. First, a fixed part of each RR interval
is replaced by a sigmoid function, such that QRS complexes and
P waves are blanked (see Fig. 2A, second line). Subsequently,
phase-free Butterworth band-pass filtering (passband from 0.3 Hz
to 20 Hz) of order 4 is performed. This QRS-free signal is fed
into a phase-free SWT [12] with the rbio3.3 wavelet yielding a
wavelet signal hardly showing activity in the QRS segment (see
Fig. 2A, third line). The level N of the detail coefficients to be
analyzed is determined based on the central frequency of the T
wave (7 Hz) [18].
After preparing the signal, the polarity of the T wave is de-
termined considering all positive and negative extrema in the
SWT signal. In a search window dependent on the RR interval
and QRS offset, we determine the peak of the T wave (Tpeak )
in each RR interval by choosing the time instant of maximum
amplitude (for positive polarity) or of the minimum amplitude
(for negative polarity). According to the polarity, a correction of
this peak is made in the ECG signal (time domain). Afterwards,
T wave boundaries are determined. Around every T peak in each
RR interval, we calculate the area under the curve of the SWT at
level N − 1 within a search window dependent on the respective
RR interval. With two preset thresholds for the area distribution
and the positions of the peaks in the interval, we estimate the
position of T wave onset (Ton ) and T wave offset (Toff ). Finally, all
detected peaks and boundaries are stored in the FPT (columns ten
to twelve as shown in Table 1).

2.4. P wave detection

P wave detection P_Detection.m also relies on the phase-

free SWT as well as on QRS detection and T wave detection.
Fig. 3B provides an overview of the algorithm which was pre-
sented in [12]. The single lead input signal is band-pass filtered
with a Gaussian filter (passband 1 Hz to 15 Hz). QRS complexes
and T waves are replaced by sigmoid functions (see Fig. 2B,
second line). The SWT is performed with the quadratic spline
wavelet at the detail coefficients of level N corresponding to the
P wave center frequency of 7 Hz [18] (see Fig. 2B, third line).
The peaks of the SWT in each RR interval are determined by
annotating the absolute maximum of the SWT as P wave peak in
a search window dependent on each RR interval and the Q peak
location. The user can also choose to set a nearby extremum in
the filtered ECG signal as peak.
After peak detection, P wave boundaries are determined. Since Fig. 4. A possible pipeline for processing an ECG signal with ECGdeli. After
preprocessing, ECG wave peak and boundary detection are performed. The
P waves are low in amplitude, a template of the SWT in a window
arrows pointing into the box determine input parameters, arrows pointing away
of 50 waves around the current P wave peak is built to estimate output parameters, respectively.
the boundaries. Hotelling’s T-squared method is applied to ob-
tain a clean template. P wave onset and offset are estimated by
initial guesses from the template considering the nearest wavelet
extremum as the respective wave boundary. 2.6. Software architecture and functionality
The found peaks and wave boundaries are stored in the FPT,
columns one to three as shown in Table 1.
A high-level view on the pipeline for the ECG preprocessing
2.5. Multilead processing and annotation is shown in Fig. 4. First, baseline removal, filtering
and isoline correction are performed (see Fig. 5, lines one to
As already mentioned, P wave, QRS complex, and T wave four). Following that, the signal can be annotated by launching
detection are generally performed lead by lead. To achieve a mul- the script Annotate_ECG_Multi.m. Here, QRS detection, T and
tilead delineation, a voting algorithm (Sync_Beats.m) discards
P wave detection are run (see Fig. 5, line five). Afterwards, in the
possible beats in the FPT if they are not visible in the majority
of all leads and averages over the found positions to deliver a case of a multilead signal, the voting algorithm (Sync_Beats.m)
overall delineation result. Innately, a reasonable result can only would discard possible beats in the FPT if they are not visible in
be achieved with more than two leads. the majority of all leads.
5
N. Pilia, C. Nagel, G. Lenis et al. SoftwareX 13 (2021) 100639

Fig. 5. Application of the filtering and wave detection to an ECG signal for signal sel100 from [10]. Line 1: Estimated baseline wander and ECG. Line 2: ECG without
baseline wander. Line 3: Bandpass filtered signal and estimated isoline. Line 4: Bandpass filtered signal with isoline correction. Line 5: The annotated on- and offsets
as well as the peaks for the P and T wave and the QRS complex.

Table 1
Structure of the FPT. Lines in the FPT represent the number of the detected beat. Column 9 is reserved (res.) for
the J point, column 13 for a beat classification.
1 2 3 4 5 6 7 8 9 10 11 12 13
Beat number Pon Ppeak Poff QRSon Q R S QRSoff res. Ton Tpeak Toff res.

3. Performance evaluation available on github (version 1.0, Commit c8e3de4) and the ecg-
puwave [17] algorithm as provided in PhysioNet. With ecg-kit,
The algorithms provided in the toolbox were evaluated with we used the wavedet delineation algorithm which is the default
the QT database, an ECG database comprising 105 signals that selection for ECG delineation. Detection errors for all types of
were either recorded during normal sinus rhythm or represent annotations for all algorithms were calculated, including median
one of six selected cardiac diseases with fiducial points annotated errors and the interquartile ranges, mean errors and standard
by clinicians [10]. We extracted annotations from the *.q1c and deviations of the signed and unsigned (absolute) errors.
*.q2c files of the database. To obtain the detection errors, we
analyzed both of the two leads separately. For each annotation, 3.1. Results
we subsequently took the best result among the two leads mean-
ing choosing the annotation being closest to the annotation. To Fig. 6 shows the differences between the calculated fiducial
get the clostest annotation, every annotation available in one points with the ECGdeli and the manually annotated points by
beat was considered finding the respective annotation beat. We clinicians for all annotated beats in the QT database. Especially
carried out the exact same evaluation for ECGdeli, ecg-kit [16] with the detection of QRS and T wave annotations, outliers were
6
N. Pilia, C. Nagel, G. Lenis et al. SoftwareX 13 (2021) 100639

Table 2
Detection errors of ECGdeli, ecg-kit and ecgpuwave compared to manual expert annotations in samples as well as number of detected points/available annotations.
med: median, iqr: interquartile range, m: mean, std: standard deviation, signed: signed errors, abs: absolute errors.

QRS Offset
QRS Onset

QRS Peak
P Offset

T Offset
P Onset

T Onset
P Peak

T Peak
med±iqr ECGdeli 2.00 ± 3.00 1.00 ± 2.00 3.00 ± 4.00 2.00 ± 4.00 1.00 ± 1.00 3.00 ± 4.00 5.00 ± 9.00 2.00 ± 3.00 3.00 ± 7.00
m±std abs ECGdeli 3.87 ± 6.14 2.46 ± 5.70 4.17 ± 6.38 4.01 ± 8.50 2.46 ± 8.46 4.39 ± 8.53 8.86 ± 9.98 6.24 ± 12.29 7.18 ± 11.24
m±std signed ECGdeli −1.51 ± 7.10 0.65 ± 6.18 2.26 ± 7.27 −1.31 ± 9.30 −0.60 ± 8.79 2.19 ± 9.34 −3.65 ± 12.84 −0.72 ± 13.76 −1.92 ± 13.20
med±iqr ecg-kit 2.00 ± 4.00 1.00 ± 2.00 2.00 ± 3.00 1.00 ± 4.00 2.00 ± 2.00 1.00 ± 3.00 5.00 ± 14.00 1.00 ± 14.00 3.00 ± 11.00
m±std abs ecg-kit 3.33 ± 6.66 2.71 ± 6.57 3.30 ± 5.64 1.66 ± 2.14 3.36 ± 4.37 1.91 ± 2.40 9.82 ± 13.43 4.62 ± 12.24 6.67 ± 14.38
m±std signed ecg-kit −1.10 ± 7.36 −1.17 ± 7.01 −2.18 ± 6.15 0.54 ± 2.65 −3.13 ± 4.54 −0.46 ± 3.04 −1.11 ± 16.60 −2.18 ± 12.90 −2.81 ± 15.60
med±iqr ecgpuwave 3.00 ± 5.00 2.00 ± 2.00 2.00 ± 4.00 2.00 ± 4.00 3.00 ± 2.00 2.00 ± 3.00 8.00 ± 14.00 2.00 ± 14.00 6.00 ± 11.00
m±std abs ecgpuwave 4.52 ± 6.05 2.96 ± 4.62 3.89 ± 4.91 3.46 ± 3.55 3.54 ± 4.32 3.51 ± 5.48 13.10 ± 13.98 10.50 ± 17.66 11.57 ± 16.56
m±std signed ecgpuwave 3.33 ± 6.78 −0.24 ± 5.48 −1.23 ± 6.14 −1.61 ± 4.69 −1.71 ± 5.31 −0.38 ± 6.49 9.96 ± 16.37 7.51 ± 19.12 3.86 ± 19.84
Detected points ECGdeli 3194 3194 3194 4019 4019 4019 1414 3936 3936
Detected points ecg-kit 3096 3096 3096 4019 4013 4018 1332 3801 3812
Detected points ecgpuwave 2127 2127 2127 4018 4019 4017 1315 3865 3864
Number of clinical annotations 3194 3194 3194 4019 4019 4019 1414 3936 3936

3.2. Discussion

In this section, three main points connected with the results

from Section 3.1 will be discussed.
First, we decided to evaluate the two leads separately and
then take the lowest error per annotation as the annotator had
also both leads at hands during annotation and we did not know
which lead was chosen. We therefore did not evaluate a possible
improvement by using a multilead approach which would have
been hardly possible for ECGdeli since a voting between two
different annotations does not deliver an advantage.
Second, we want to reference to further publications compar-
ing more closed source implementations [19,20]. We intention-
ally did not repeat the results of the closed source implemen-
tations here for two main reasons: first, we wanted to compare
our results to the two most visible algorithms for ECG delineation
offering P wave, T wave and QRS detection. Second, as there is no
standard procedure to generate the final evaluation results, it is
hardly possible to guarantee the comparability.
Third, we want to highlight the possibility of a postprocessing
step for ECGdeli and a possible improvement of the average
detection errors. With the current implementation, and as already
stated, the algorithm is forced to detect a P and T wave in
each RR interval. On the one hand, this implies that every wave
is detected. Nevertheless, an adequate detection of the outliers
(visible e.g. for the T wave in Fig. 6) and a subsequent correction
or dropping of those, could have lowered the average errors of
ECGdeli.

4. Impact

ECG delineation algorithms are important in clinical and re-

search practice. As stated in Section 1, many parameters depend
on the result of the wave detection. With ECGdeli we offer one
Fig. 6. Boxplot of the absolute errors obtained when applying ECGdeli on the of the few open toolboxes to solve this relevant problem. By
QT database. N is the number of detected waves. comparing the toolbox with two alternative implementations for
wave delineation, we showed that ECGdeli is already at its current
state delivering results en par with existing open approaches.
visible. Detection errors for all three algorithms are given in Ta- However, several partly unique features of ECGdeli should be
ble 2. ECGdeli and ecg-kit performed comparably, dependent on highlighted:
The simple input/output and modular structure make the tool-
the performance measure and the annotation type, one was out-
box functions easy and intuitive to use. In this way, the anno-
performing the other. Regarding only median and interquartile
tation functions can also be executed separately if for example
ranges, ecgpuwave was always outperformed by ecg-kit.
only a P wave detection is necessary. This goes hand in hand with
A particular difference was visible in the number of detected the fact that by making ECGdeli freely available under the GPLv3,
waveforms. This number was highest in the case of ECGdeli (see single algorithms can be extracted and incorporated into existing
last lines in Table 2). ecg-kit and ecgpuwave discarded waves projects to extend them.
even though they should ideally be detected due to the fact that Furthermore, others can easily apply ECGdeli as a benchmark
clinicians clearly found and annotated the respective wave in the for new algorithms as we did with ecg-kit and ecgpuwave. Usu-
signal. ECGdeli did not so. ally, the evaluation of the algorithms with standard databases,
7
N. Pilia, C. Nagel, G. Lenis et al. SoftwareX 13 (2021) 100639

like the QT database, allows to compare new work to already Appendix A. Supplementary data
existing. However, manual annotations in these databases can be
prone to error and different ways of calculating performance pa- Supplementary material related to this article can be found
rameters impair comparability (as discussed in Section 3). More- online at https://doi.org/10.1016/j.softx.2020.100639.
over, there might be pathologies not represented in a freely
available database that can however be relevant for the intended References
application of a new algorithm. The best way to compare the
own work to existing algorithms is to have benchmark algorithms [1] Pilia N, Nagel C. KIT-ibt/ecgdeli: Ecgdeli v1.0.2. Zenodo; 2020, http://dx.
doi.org/10.5281/zenodo.3977971.
available.
[2] Fye W. A History of the origin, evolution, and impact of electrocardiogra-
At last, ECGdeli is capable of detecting all P and T waves in phy. Amer J Cardiol 1994;73(13):937–49. http://dx.doi.org/10.1016/0002-
a provided input ECG signal if the respective QRS complex was 9149(94)90135-X.
detected. Thus, for example a P wave gets detected in any case [3] Lyon A, Mincholé A, Martínez JP, Laguna P, Rodriguez B. Computational
between two subsequent R peaks. In this way, no false negative techniques for ECG analysis and interpretation in light of their contribution
to medical advances. J R Soc Interface 2018;15(138). http://dx.doi.org/10.
detected waves are possible and false positive waves can still 1098/rsif.2017.0821.
be eliminated in a postprocessing step that is designed for the [4] Majumder S, Chen L, Marinov O, Chen C-H, Mondal T, Deen MJ. Noncontact
specific problem. wearable wireless ECG systems for long-term monitoring. IEEE Rev Biomed
Eng 2018;11:306–21. http://dx.doi.org/10.1109/RBME.2018.2840336.
[5] Nielsen JB, Kuhl JT, Pietersen A, Graff C, Lind B, Struijk JJ, et al. P-wave
5. Conclusions
duration and the risk of atrial fibrillation: Results from the Copenhagen
ECG study. Heart Rhythm : Off J Heart Rhythm Soc 2015;12(9):1887–95.
In this work, we present the ECGdeli MATLAB toolbox for ECG http://dx.doi.org/10.1016/j.hrthm.2015.04.026.
wave delineation. The modularity is the major advantage as single [6] Loewe A, Krueger MW, Holmqvist F, Dossel O, Seemann G, Platonov PG.
algorithms can be used and adapted independently from the Influence of the earliest right atrial activation site and its proximity to
interatrial connections on P-wave morphology. Europace 2016;18(Suppl.
others (e.g., use another QRS detector as long as the standardized 4):iv35–43. http://dx.doi.org/10.1093/europace/euw349.
FPT is used). By publishing ECGdeli under GPLv3, the toolbox [7] Andlauer R, Seemann G, Baron L, Dössel O, Kohl P, Platonov P, et al.
can be freely used for wave detection, serve as a benchmark Influence of left atrial size on P-wave morphology: differential effects of
method for future works on non-standard datasets or as a basis dilation and hypertrophy. Europace 2018;20(S3):iii36–44. http://dx.doi.org/
10.1093/europace/euy231.
for extended feature extraction algorithms.
[8] Jahmunah V, Oh SL, Wei JKE, Ciaccio EJ, Chua K, San TR, et al. Computer-
aided diagnosis of congestive heart failure using ECG signals - A review.
CRediT authorship contribution statement Phys Med Off J Ital Assoc Biomed Phys 2019;62:95–104. http://dx.doi.org/
10.1016/j.ejmp.2019.05.004.
N. Pilia: Methodology, Software, Conceptualization, Investiga- [9] Clifford GD, Azuaje F, McSharry P. Advanced methods and tools for ECG
data analysis. Boston, Mass.[u.a.]: Artech House; 2006.
tion, Validation, Visualization, Writing - original draft, Writing [10] Laguna P, Mark RG, Goldberg A, Moody GB. A database for evaluation of
- review & editing. C. Nagel: Methodology, Conceptualization, algorithms for measurement of QT and other waveform intervals in the
Validation, Visualization, Writing - original draft, Writing - review ECG. In: Computers in cardiology 1997; 1997. p. 673–76.
& editing. G. Lenis: Methodology, Software, Conceptualization, [11] Lenis G, Dössel O. T wave morphology during heart rate turbulence in
patients with chronic heart failure. Biomed Tech Biomed Eng 2013;58(s1).
Investigation, Validation. S. Becker: Methodology, Software, In-
http://dx.doi.org/10.1515/bmt-2013-4171.
vestigation, Validation, Visualization, Writing - original draft. O. [12] Lenis G, Pilia N, Oesterlein T, Luik A, Schmitt C, Dössel O. P wave detection
Dössel: Funding acquisition, Project administration, Supervision, and delineation in the ECG based on the phase free stationary wavelet
Conceptualization. A. Loewe: Funding acquisition, Project admin- transform and using intracardiac atrial electrograms as reference. Biomed
istration, Data curation, Supervision, Conceptualization, Writing - Tech Biomed Eng 2016;61(1):37–56. http://dx.doi.org/10.1515/bmt-2014-
0161.
review & editing..
[13] Lenis G, Reichensperger P, Sommer D, Heinze C, Golz M, Dössel O.
Detection of microsleep events in a car driving simulation study using
Declaration of competing interest electrocardiographic features. Curr Direct Biomed Eng 2016;2(1):283–7.
[14] Lenis G, Kircher M, Lazaro J, Bailon R, Gil E, Dössel O. Separating the
effect of respiration on the heart rate variability using Granger’s causality
The authors declare that they have no known competing finan-
and linear filtering. Biomed Signal Process Control 2017;31:272–8. http:
cial interests or personal relationships that could have appeared //dx.doi.org/10.1016/j.bspc.2016.07.014.
to influence the work reported in this paper. [15] Lenis G, Pilia N, Loewe A, Schulze WHW, Dössel O. Comparison of
baseline wander removal techniques considering the preservation of ST
Acknowledgments changes in the ischemic ECG: A simulation study. Comput Math Methods
Med 2017;2017:13. http://dx.doi.org/10.1155/2017/9295029, (Article ID
9295029).
We thank several colleagues, collaborators and students at our [16] Demski A, Llamedo Soria M. Ecg-kit a Matlab toolbox for cardiovascular
institute for using the software, reporting bugs and proposing signal processing. J Open Res Softw 2016;4. http://dx.doi.org/10.5334/jors.
improvements. Furthermore, we acknowledge support by the 86.
KIT-Publication Fund of the Karlsruhe Institute of Technology, [17] Laguna P, Jané R, Caminal P. Automatic detection of wave boundaries in
multilead ECG signals: Validation with the CSE database. Comput Biomed
Germany. Res 1994;27(1):45–60.
We gratefully acknowledge financial support by the Euro- [18] Sörnmo L, Laguna P. Bioelectrical signal processing in cardiac and
pean Metrology Programme for Innovation and Research, Ger- neurological applications. Amsterdam: Elsevier; 2005.
many [18HLT07 MedalCare]. We acknowledge support by the [19] Di Marco LY, Chiari L. A wavelet-based ECG delineation algorithm for
32-bit integer online processing. Biomed Eng Online 2011;10:23. http:
KIT-Publication Fund of the Karlsruhe Institute of Technology,
//dx.doi.org/10.1186/1475-925X-10-23.
Germany. [20] Spicher N, Kukuk M. Delineation of electrocardiograms using multiscale
We also gratefully acknowledge support by Deutsche parameter estimation. IEEE J Biomed Health Inf 2020;24(8):2216–29. http:
Forschungsgemeinschaft (DFG), Germany [project ID 391128822, //dx.doi.org/10.1109/JBHI.2019.2963786.
LO2093/1-1].