Signal Quality Classification of Impedance Plethysmogram and
Ballistocardiogram for Pulse Transit Time Measurement
SHING-HONG LIU1, TAI-SHEN HUANG2,*, XIN ZHU3, TAN-HSU TAN4, JIA-JUNG WANG5
1Department of Computer Science and Information Engineering, Chaoyang University of Technology,
Taichung City 41349,
TAIWAN
2Department of Industrial Design, Chaoyang University of Technology, Taichung City 41349,
TAIWAN
3Division of Information Systems, School of Computer Science and Engineering, University of Aizu,
Aizu-Wakamatsu 965-8580,
JAPAN
4Department of Electrical Engineering, National Taipei University of Technology,
Taipei 10608,
TAIWAN
5Department of Biomedical Engineering, I-Shou University,
Kaohsiung 84001,
TAIWAN
*Corresponding Author
Abstract: - Mobile health (mHealth) was developed ten years ago, which used wireless wearable devices to
collect the many physiological messages in daily life, regardless of time and place, for some health services
including monitoring chronic diseases and reducing the cost of empowering patients and families for handling
their daily healthcare. However, the challenge for these measurements is the lower signal quality because users
would measure their conditions not on a resting status. Now, the pulse transit time (PTT) is highly related to
blood pressure has been proposed, which is acquired from the impedance plethysmography (IPG) and
ballistocardiogram (BCG) measured by the weight-fat scale. However, the lower signal quality of IPG and
BCG, lowers the accuracy of blood pressure. This study aims to use deep learning techniques to classify the
signal quality of BCG and IPG signals. The reference PTTs were measured by the electrocardiogram (ECG)
and photoplethysmogram (PPG). The signal quality of each segment was labeled with the error between
proposed and reference PTTs. We used three signals, BCG, IPG, and differential IPG, as the input. The
proposed one-dimensional stacking convolutional neural network and gait recursive unit (1-D CNN+GRU)
model to approach the classification. The good performances achieved high accuracy (98.85%), recall (99.4%),
precision (94.29%), and F1-score (96.78%). These results show the potential benefit of the signal quality
classification for the PTT measurement.
Key-Words: - mHealth, signal quality, ballistocardiogram, impedance plethysmogram, pulse transit time,
convolutional neural network.
Received: Septermber 13, 2023. Revised: July 11, 2024. Accepted: August 12, 2024. Published: September 11, 2024.
1 Introduction
Wireless and wearable devices are the major issue in
mobile health (mHealth) because devices deal with
some health services, including monitoring the
conditions of chronic diseases, and reducing the cost
of users for handling their daily healthcare without
the limitations of time and place, [1]. Thus, many
studies have developed innovative wearable devices
in the past 10 years for taking care of patients, or
health management of users. For example, the
Apple watch has the functions of detecting the
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.26
Shing-Hong Liu, Tai-Shen Huang,
Xin Zhu, Tan-Hsu Tan, Jia-Jung Wang
E-ISSN: 2224-2902
242
Volume 21, 2024
arrhythmia by the electrocardiogram, and
monitoring the blood oxygen saturation, [2]. The
electrocardiography (ECG) patch, [3] and
electromyography (EMG) patch, [4] have been
proposed to monitor the condition of the heart and
muscle in real-time. However, innovative mobile
devices for different health care are very important
research.
Blood pressure (BP) is the most important
physiological parameter for healthcare in the home
because it has direct and indirect relations with
many chronic diseases, like hypertension,
hyperlipidemia, heart failure, stroke, and kidney
disease, etc. [5]. According to the World Health
Organization’s report, people must measure their BP
daily and keep their systolic BP lower than 130
mmHg, [6]. The commercial and automatic
sphygmomanometer uses either auscultatory or
oscillometric methods, [7]. These methods all use an
occlusive cuff wrapping around a user’s upper arm
to measure the BP. The disadvantage of these
methods is uncomforting when the BP is measured.
In the recent years, the cuffless BP measurements
have been widely studied, [8]. According to the
Moens-Korteweg equation, the pulse transit time
(PTT) has a high relation with BP, [9]. Reference
[10], showed the PTT by the ECG and
photoplethysmogram (PPG). They found the
relation between the PTT and BP change to be
larger than 0.8. Reference, [11] used
phonocardiography replacing the ECG, and PPG to
measure the PTT and estimate the BP. Reference
[12], used the tonometer measured at the wrist
which was replaced with the PPG, and ECG to
evaluate the blood pressure. Reference [13], used an
impedance plethysmogram (IPG) measured at the
forearm which replaced the PPG, and ECG to
measure the PTT and estimate the BP. Reference
[14], proposed innovative cuffless BP measuring
methods with the ballistocardiogram (BCG) and
IPG. The BCG and IPG signals can be measured
from a commercial weight-fat scale. However, in
these studies, we found a basic problem. The lower
the quality of the signal, the lower the accuracy of
blood pressure.
The quality of physiological signals generally is
labeled by the manual marks of experts, [15].
However, the signal quality would depend on the
experiences of experts. The rule-based method is to
find some waveform characteristics and classify
whether they fit the normal ranges or not, [16]. The
disadvantage is how to define the accurate ranges
which would be affected by the number of samples.
Reference [17], transferred the pulses of PPG and
differential PPG (DPPG) to an image and used a
convolutional neural network (CNN) for the
classification of signal quality. Its advantage was to
transfer a one-dimension signal to a two-dimension
image. Reference [18] showed some methods for
signal quality classifications of ECG. We found that
the deep learning methods classified the signal
quality, which input would be a two-dimensional
image. Thus, the complexity of bringing to practice
will arise.
This study aims to propose a deep learning
model for signal quality classification which uses
the raw BCG and IPG signals as the input. The
signal quality was labeled-labeled by the error of
PTT measured from BCG and IPG. The PTT
measured by the ECG and PPG was the reference.
The BCG and IPG were measured by the self-made
circuits when users were standing on the
commercial weight-fat scale. The deep learning
model was a stacking CNN plus a gate recursive
unit (GRU). The output was one node, one
representing good quality, and zero representing
poor quality.
2 Methods
2.1 Experiment Protocol
This study employed 11 males and 6 females who
were young and healthy subjects. Their ages were
from 22 to 19 years (mean ± standard deviation,
20.2 ± 1.1 years), heights were from 186 to 152 cm
(mean ± standard deviation, 166.1 ± 8.0 cm), and
weights were from 115 to 43 kg (mean ± standard
deviation, 62.8 ± 16.1 kg). The digital
sphygmomanometer (HM-7320, Omron, Osaka,
Japan) was used to measure the BPs as the
reference. The self-made circuit was used to
measure Lead I ECG and finger PPG of the left
hand. PTTECG-PPG was measured from the ECG and
PPG which would be the standard PTT. The self-
made circuits were used to measure the BCG and
IPG signals, and which sensors were at a
commercial body weight-fat scale (HBF-371,
Omron, Osaka, Japan), [18]. The experiment
procedure is mentioned below.
I. ECG, PPG, IPG, and BCG signals were measured
for five minutes, and BP was measured once when
subjects were standing on the weight-fat scale.
II. Subjects were running on a treadmill to raise the
BP until the systolic BP was higher than the 20
mmHg of resting BP.
III. Subjects were standing on the weight-fat scale
again, and ECG, PPG, IPG, and BCG signals were
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.26
Shing-Hong Liu, Tai-Shen Huang,
Xin Zhu, Tan-Hsu Tan, Jia-Jung Wang
E-ISSN: 2224-2902
243
measured for six minutes. Their BPs were measured
once a minute.
IV. Subjects were measured four times. Each
experiment would have a rest for at least a week.
2.2 Signal Processing and Segment
The sampling rate was 500 Hz. The 4th-order
Butterworth bandpass filter with 0.5 Hz to 10 Hz
bandwidth was used to remove the wandering
baseline and high-frequency noise. The group
delays of all signal groups were reduced by an 8th-
order all-pass filter. Figure 1 shows these signals,
ECG (blue), PPG (red), DPPG (pink), BCG (black),
IPG (green), and differential IPG (DIPG, purple).
The PTT1BCG-IPG is the interval between the J wave
of BCG and the foot point of IPG, and the PTT2BCG-
IPG is the interval between the J wave of BCG and
the peak point of DIPG. The PTT1ECG-PPG is the
interval between the R wave of ECG and the foot
point of PPG, and the PTT2ECG-PPG is the interval
between the R wave of ECG and the peak point of
DPPG.
We used the error ratio (E) between PTT2ECG-
PPG and PTT2BCG-IPG of each beat to define the signal
quality.


, (1)
where Bias is the time delay between ECG and
BCG, [14]. By the trial and error method, we
defined 30% of E as the threshold. When E is below
the threshold, the pulse belongs to good quality, this
cycle labeling as 1. Otherwise, the cycle is labeled
as 0. Figure 2 shows the labels of pulses with the red
line. We find that the second pulse belongs to poor
quality because the foot of its IPG is at the wrong
place. Because the PTT1 and PTT2 were extracted
from the BCG, IPG, and DIPG, we used the three
signals directly to classify the signal quality.
In the data segment, the window was 1024
points, the overlap was 512 points. In order to
reduce the personal affection for the classification of
signal quality, the BCG (blue), IPG (red), and DIPG
(orange), were normalized, as shown in Figure 3.
Because one segment has at least two PTTs, the
segment was labeled as good or poor quality
depending on all pulses in it belonging to all good or
poor. The segment would be deleted when the
pulses in it had different qualities. The numbers of
good and poor samples were 3938 and 18682, a total
of 22,620 samples.
Fig. 1: PTT1 and PTT2 are defined by ECG (blue),
PPG (red), DPPG (pink), BCG (black), IPG (green),
and differential IPG (DIPG, purple)
Fig. 2: The label of signal quality is the red line,
When the pulse belongs to the good quality, its
cycle is labeled as 1 (higher horizontal line).
Otherwise, the label is 0 (lower horizontal line)
Fig. 3: The normalized BCG (blue), IPG (red),
and DIPG (orange)
2.3 Signal Quality Classifier
A stacking CNN+GRU model was proposed to
classify the signal quality as shown in Figure 4. The
three channels, BCG, IPG, and DIPG signals, are
the input. A time-distributed layer is separated into
two parts that connect to two one-dimensional
CNNs. The CNN has three layers, a maximal pool
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.26
Shing-Hong Liu, Tai-Shen Huang,
Xin Zhu, Tan-Hsu Tan, Jia-Jung Wang
E-ISSN: 2224-2902
244
layer, and a flattened layer. Then, a GRU is
connected after the flattening layer. A full
connection layer connects with the output layer of
the GRU. In CNN layers, the number of filters is 32,
the kernel sizes are 3, 5, and 13, respectively, and
the stride is 2. In the maximal pooling layer, the
kernel size is 2, and the stride is 2. The activation
function is ReLU. The unit number of GRU is set to
1024. The batch size is set to 512, with the control
reset gate and update gate using a sigmoid function
and the hidden state using a tanh function. One node
is in the output layer, and which activation function
is the sigmoid function. One represents the good
quality, and zero represents the poor quality. The
threshold of output is 0.5. The dropout in the hidden
layer is 0.5. The loss function is the binary Cross-
Entropy function, and the Adam optimizer is used,
with a learning rate of 0.0001.
Fig. 4: The structure of the proposed stacking
CNN+GRU model for the signal quality
classification.
3 Results
An Intel Core i7-8700 CPU and a 186 GeForce
GTX3070 GPU were used to evaluate the
performance of the proposed method. The number
of total samples was 22,620, and the numbers of
training and testing samples were 15,834 and 6,786,
respectively. The training samples were separated as
the training and validation samples with 8 to 2.
The statistic of data is expressed as the mean ±
standard deviation. The sensitivity, specificity, and
accuracy are used to evaluate the performance of the
model. In the fusion matrix, TP is true positive, FN
is false negative, FP is false positive, and TN is true
negative.
󰇛󰇜
󰇛󰇜 (2)
󰇛󰇜 
󰇛󰇜 (3)
󰇛󰇜 
󰇛󰇜. (4)
The accuracy curves of the stacking CNN+GRU
model in the training (blue line) and validation
(orange line) phases are shown in Figure 5(a). The
loss curves are shown in Figure 5(b). We find that
the accuracy approaches 0.97, and the loss value
approaches 0.12 when the epoch is 14. The fusion
matrix in the testing phase is shown in Figure 6, the
numbers of TP, TN, FN, and FP are 1174, 5534, 7,
and 71. The results of accuracy, sensitivity, and
specificity are 98.9%, 99.4%, and 98.7%,
respectively.
(a) (b)
Fig. 5: The results of the stacking CNN+GRU
model in the training (blue) and validation (orange)
phases, (a) the accuracy curves, (b) the loss curves
Fig. 6: The fusion matrix in the testing phase. The
numbers of TP, TN, FN, and FP are 1174, 5534, 7,
and 71
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.26
Shing-Hong Liu, Tai-Shen Huang,
Xin Zhu, Tan-Hsu Tan, Jia-Jung Wang
E-ISSN: 2224-2902
245
4 Discussions
The size of the segment would be an explored issue
in this study. Because the sampling rate was 500 Hz,
the minimum size of the segment was 512 points.
Figure 7(a) shows the accuracy curves in the
training (blue line) and validation (orange line)
phases when the size of the segment is 512, and
Figure 7(b) shows the loss curves. The accuracy
approaches 0.92, and the loss value approaches 0.18
when the epoch is 4. Its performance is lower than
the 1024 points of segment size. The reason would
be that the rate of non-complete cycle is close to the
full cycle. Although the signal quality within the
non-complete cycle is poor, and the signal quality
within the full cycle is good, this segment is also
labeled as good quality. Thus, the model would
recognize the signals with the poor quality as the
good quality.
Figure 8(a) shows the accuracy curves in the
training (blue line) and validation (orange line)
phases when the size of the segment is 2048, and
Figure 8(b) shows the loss curves. The accuracy
approaches 0.98, and the loss value approaches 0.16
when the epoch is 14. This performance is very
close to the 1024 points of segment size. But, the
disadvantage was that the number of samples would
decrease a lot. Because when the segment size is
2048 points, the number of full cycles would be 5 at
least. According to the labeling rules, the rate of
failed segments would increase. Therefore, we
chose the 1024 points of segment size as the sample.
Fig. 7: The results of the stacking CNN+GRU
model in the training (blue) and validation (orange)
phases with a segment size of 512 points, (a) the
accuracy curves, (b) the loss curves
Fig. 8: The results of the stacking CNN+GRU
model in the training (blue) and validation (orange)
phases with a segment size of 2048 points, (a) the
accuracy curves, (b) the loss curves
5 Conclusion
In the development of mHealth, wireless and
wearable devices for monitoring physiological
conditions every day have gotten attention. Now,
PTT can be used to estimate the BP, which usually
is measured by the ECG and PPG signals when
subjects sitting on a chair or lying on the bed. When
users are standing on the weight-fat scale to measure
the PTT, the signals must have larger artificial
motions. In this study, we proposed the stacking
CNN+GRU model to classify the signal quality of
BCG and IPG signals with the one-dimensional
data. The accuracy approached to 98.9%. Thus, it
has the potential benefit for BP measurement with
the weight-fat scale when users standing on it.
Therefore, this method could be applied in the
mHealth in the future.
However, the major limitation of this study is
that the subjects all were young and healthy people.
They stood on the weight-fat scale more stable than
the elderly. When users have Parkinson's disease or
use assistive devices for standing, they cannot suit
this method.
Acknowledgement:
This research was supported by the National
Science and Technology Council, Taiwan, and JST
Japan Collaborative Research Program, Grant
Number: NSTC 113-2923-E-324 -001 -MY3, Grant
Number JPMJKB2311, Japan.
References:
[1] B.M. Silva, J.J. Rodrigues, de la Torre Díez I,
M. López-Coronado, K. Saleem, "Mobile-
health: A review of current state in 2015," J.
Biomed. Inform., vol.56, pp. 265-72,
2015,https://doi.org/10.1016/j.jbi.2015.06.003
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.26
Shing-Hong Liu, Tai-Shen Huang,
Xin Zhu, Tan-Hsu Tan, Jia-Jung Wang
E-ISSN: 2224-2902
246
[2] How to use the Blood Oxygen app on Apple
Watch, [Online].
https://support.apple.com/en-us/HT211027
(Accessed Date: December 2, 2012).
[3] S. S. Lobodzinski, “ECG patch monitors for
assessment of cardiac rhythm abnormalities, ”
Progress in Cardiovascular Diseases, vol. 56,
pp 224–229, 2013,
https://doi.org/10.1016/j.pcad.2013.08.006.
[4] C. Kantarcigil, M. K. Kim, T. Chang, B. A.
Craig, A. Smith, C. H. Lee, G. A. Malandraki,
“Validation of a novel wearable
electromyography patch for monitoring
submental muscle activity during swallowing:
a randomized crossover trial,” Journal of
Speech, Language, and Hearing Research,
vol.63, pp 3293–3310, 2020,
https://doi.org/10.1044/2020_JSLHR-20-
00171.
[5] M.V. Williams, D.W. Baker, R.M. Parker, J.
R. Nurss, “Relationship of functional health
literacy to patients' knowledge of their chronic
disease: a study of patients with hypertension
and diabetes,” Arch. Intern. Med., vol.158,
pp.166-172, 1998,
https://doi.org/10.1001/archinte.158.2.166.
[6] A. M. Master, L.I. Dublin, H.H. Marks. “The
normal blood pressure range and its clinical
implications,” JAMA. vol.143, pp.1464-1470,
1950,
https://doi:10.1001/jama.1950.029105200060
04.
[7] J.J. Carr, J.M. Brown, Introduction to
Biomedical Equipment Technology, 4th ed.;
Prentice Hall: City, Country, 2001; ISBN:0-
13-010492-2.
[8] C. Ahlstrom, A. Johansson, F. Uhlin, T.
Länne, P. Ask, “Noninvasive investigation of
blood pressure changes using the pulse wave
transit time: a novel approach in the
monitoring of hemodialysis patients,” J. Artif.
Organs, vol.8, pp.192–197, 2015,
https://doi.org/10.1007/s10047-005-0301-4.
[9] D.L. Newman, S.E. Greenwald, “Validity of
the Moens-Korteweg equation,” The Arterial
System, pp. 109–115, 1978.
[10] G. Sharwood-Smith, J. Bruce, G. Drummond,
“Assessment of pulse transit time to indicate
cardiovascular changes during obstetric spinal
anesthesia,” Br. J. Anaesth., vol. 96, pp.100–
105, 2006, https://doi.org/10.1093/bja/aei266.
[11] J.Y.A. Foo, C.S. Lim, P. Wang, “Evaluation
of blood pressure changes using vascular
transit time,” Physiol. Meas., vol.27, pp.685–
694, 2006, https://doi.org/10.1088/0967-
3334/27/8/003.
[12] M. Park, H. Kang, Y. Huh, K.C. Kim,
“Cuffless and noninvasive measurement of
systolic blood pressure, diastolic blood
pressure, mean arterial pressure and pulse
pressure using radial artery tonometry
pressure sensor with concept of Korean
traditional medicine,” Proceedings of the 29th
Annual International Conference of the IEEE
Engineering in Medicine and Biology Society,
Lyon, France, 22–26 August 2007, pp. 3597–
3600, 2007, https://doi.org/
10.1109/IEMBS.2007.4353109.
[13] T. H. Huynh, R. Jafari, W.-Y. Chung,
“Noninvasive cuffless blood pressure
estimation using pulse transit time and
impedance plethysmography,” IEEE
Transactions on Biomedical Engineering, vol.
66, pp 967–976, 2018.
[14] S.-H. Liu, Y.-R. Wu, W. Chen, C.-H. Su, C.-
L. Chin, “Using ballistocardiogram and
impedance plethysmogram for minimal
contact measurement of blood pressure based
on a body weight-fat scale,” Sensors, vol.23,
2318, 2023,
https://doi.org/10.3390/s23042318.
[15] C. Fischer, B. Domer, T. Wibmer, T. Penzel,
“An algorithm for real-time pulse waveform
segmentation and artifact detection in
photoplethysmograms,” IEEE J. Biomed.
Health Inform., vol.21, pp. 372–381, 2017.
https://doi.org/10.1109/JBHI.2016.2518202.
[16] W. Karlen, K. Kobayashi, J.M. Ansermino,
G.A. Dumont, “Photoplethysmogram signal
quality estimation using repeated Gaussian
filters and cross-correlation,” Physiol. Meas.,
vol. 33, pp.1617-1629, 2012,
https://doi.org/10.1088/0967-
3334/33/10/1617.
[17] H. Shin, “Deep convolutional neural network-
based signal quality assessment for
photoplethysmogram,” Computers in Biology
and Medicine, vol. 145, 105430, 2022,
https://doi.org/10.1016/j.compbiomed.2022.1
05430.
[18] U. Satija, B. Ramkumar, M.S. Manikandan,
“A review of signal processing techniques for
electrocardiogram signal quality assessment,”
IEEE Reviews in Biomedical Engineering,
vol.11, pp.36-52, 2018,
https://doi.org/10.1109/RBME.2018.2810957.
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.26
Shing-Hong Liu, Tai-Shen Huang,
Xin Zhu, Tan-Hsu Tan, Jia-Jung Wang
E-ISSN: 2224-2902
247
Contribution of Individual Authors to the
Creation of a Scientific Article (Ghostwriting
Policy)
- Sing-Hong Liu carried out the concept and
method.
- Tai-Shen Huang supported the subjects and
prototype.
- Xin Zhu revised the English.
- Tan-Hsu Tan implemented the Algorithm of deep
learning model.
- Jia-Jung Wang write the draft and executed the
experiment.
Sources of Funding for Research Presented in a
Scientific Article or Scientific Article Itself
This research was supported by the National
Science and Technology Council, Taiwan, and JST
Japan Collaborative Research Program, Grant
Number: NSTC 113-2923-E-324 -001 -MY3, Grant
Number JPMJKB2311, Japan.
Conflict of Interest
This experiment was approved by the Research
Ethics Committee of Chung Shan University
Hospital (No. CS2-21194), in Taichung, Taiwan.
Creative Commons Attribution License 4.0
(Attribution 4.0 International, CC BY 4.0)
This article is published under the terms of the
Creative Commons Attribution License 4.0
https://creativecommons.org/licenses/by/4.0/deed.en
_US
WSEAS TRANSACTIONS on BIOLOGY and BIOMEDICINE
DOI: 10.37394/23208.2024.21.26
Shing-Hong Liu, Tai-Shen Huang,
Xin Zhu, Tan-Hsu Tan, Jia-Jung Wang
E-ISSN: 2224-2902
248