See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/358906734

Intelligent Bio-Impedance System for Personalized Continuous Blood

Pressure Measurement

Article in Biosensors · February 2022

DOI: 10.3390/bios12030150

CITATIONS READS

12 368

9 authors, including:

Ting-Wei Wang Jhen-Yang Syu

National Tsing Hua University National Yang Ming Chiao Tung University
29 PUBLICATIONS 435 CITATIONS 8 PUBLICATIONS 61 CITATIONS

SEE PROFILE SEE PROFILE

Hsiao-Wei Chu Yen Ling Sung

National Yang Ming Chiao Tung University National Yang Ming Chiao Tung University
8 PUBLICATIONS 106 CITATIONS 30 PUBLICATIONS 178 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Ting-Wei Wang on 09 March 2022.

The user has requested enhancement of the downloaded file.

 biosensors
Article
Intelligent Bio-Impedance System for Personalized Continuous
Blood Pressure Measurement
Ting-Wei Wang 1,2,3 , Jhen-Yang Syu 3 , Hsiao-Wei Chu 3 , Yen-Ling Sung 3,4,5,6 , Lin Chou 3 ,
Endian Escott 7 , Olivia Escott 7 , Ting-Tse Lin 5,6,8,9, * and Shien-Fong Lin 3, *

1 Department of Medical Engineering, California Institute of Technology, Pasadena, CA 91125, USA;

[email protected]
2 Department of Electrical Engineering, California Institute of Technology, Pasadena, CA 91125, USA
3 Institute of Biomedical Engineering, National Yang Ming Chiao Tung University, Hsinchu 30010, Taiwan;
[email protected] (J.-Y.S.); [email protected] (H.-W.C.); [email protected] (Y.-L.S.);
[email protected] (L.C.)
4 Department of Cardiology, Smidt Heart Institute, Cedars-Sinai Medical Center, Los Angeles, CA 90048, USA
5 Division of Cardiology, Department of Internal Medicine, National Taiwan University Hospital Hsinchu
Branch, Hsinchu 300195, Taiwan
6 Cardiovascular Center, National Taiwan University Hospital Hsinchu Branch, Hsinchu 300195, Taiwan
7 Department of Electrical Engineering and Computer Science, Florida Atlantic University,
Boca Raton, FL 33431, USA; [email protected] (E.E.); [email protected] (O.E.)
8 College of Medicine, National Taiwan University, Taipei 10617, Taiwan
9 Division of Cardiology, Department of Internal Medicine, National Taiwan University Hospital,
Taipei 10025, Taiwan
* Correspondence: [email protected] (T.-T.L.); [email protected] (S.-F.L.)

Abstract: Continuous blood pressure (BP) measurement is crucial for long-term cardiovascular
monitoring, especially for prompt hypertension detection. However, most of the continuous BP
measurements rely on the pulse transit time (PTT) from multiple-channel physiological acquisition





 systems that impede wearable applications. Recently, wearable and smart health electronics have
Citation: Wang, T.-W.; Syu, J.-Y.; Chu, become significant for next-generation personalized healthcare progress. This study proposes an
H.-W.; Sung, Y.-L.; Chou, L.; Escott, intelligent single-channel bio-impedance system for personalized BP monitoring. Compared to the
E.; Escott, O.; Lin, T.-T.; Lin, S.-F. PTT-based methods, the proposed sensing configuration greatly reduces the hardware complexity,
Intelligent Bio-Impedance System for which is beneficial for wearable applications. Most of all, the proposed system can extract the
Personalized Continuous Blood significant BP features hidden from the measured bio-impedance signals by an ultra-lightweight AI
Pressure Measurement. Biosensors algorithm, implemented to further establish a tailored BP model for personalized healthcare. In the
2022, 12, 150. https://doi.org/ human trial, the proposed system demonstrates the BP accuracy in terms of the mean error (ME) and
10.3390/bios12030150
the mean absolute error (MAE) within 1.7 ± 3.4 mmHg and 2.7 ± 2.6 mmHg, respectively, which
Received: 25 January 2022 agrees with the criteria of the Association for the Advancement of Medical Instrumentation (AAMI).
Accepted: 26 February 2022 In conclusion, this work presents a proof-of-concept for an AI-based single-channel bio-impedance
Published: 28 February 2022 BP system. The new wearable smart system is expected to accelerate the artificial intelligence of
Publisher’s Note: MDPI stays neutral
things (AIoT) technology for personalized BP healthcare in the future.
with regard to jurisdictional claims in
published maps and institutional affil- Keywords: artificial intelligence; bio-impedance measurement; continuous blood pressure measurement;
iations. impedance plethysmography; intelligent system

Copyright: © 2022 by the authors. 1. Introduction

Licensee MDPI, Basel, Switzerland.
Blood pressure (BP) monitoring is an important physiological index for cardiovascular
This article is an open access article
health identification [1–3]. The cuff-based digital electronic sphygmomanometer is a non-
distributed under the terms and
invasive gold standard to detect BP values. However, the device only provides one-shot
conditions of the Creative Commons
Attribution (CC BY) license (https://
systolic BP (SBP) and diastolic BP (DBP) measurements that could inconvenience the users
creativecommons.org/licenses/by/
when they are monitoring their health conditions in real-time, especially in hypertension
4.0/).
patients. Most of all, post-treatment hypertension patients usually need to modify the

Biosensors 2022, 12, 150. https://doi.org/10.3390/bios12030150 https://www.mdpi.com/journal/biosensors

Biosensors 2022, 12, 150 2 of 13

medicine dosage based on their recovery conditions. For the risk management of these
patients, the long-term continuous BP condition recording is important for extracting the
significant symptoms and providing an accurate treatment under a narrow therapeutic
window for medical doctors [4]. Among the existing clinical techniques for continuous BP
measurement, an arterial cannula is a common approach in clinical applications. However,
the measurement procedure is an invasive method that could induce potential risks and
complications to the patient [5]. To solve this clinical unmet need, the cuffless continuous
BP measurement technique is an efficient approach for long-term cardiovascular health-
care. Most of the technique relies on pulse transit time (PTT)-based methods for cuffless
continuous BP measurement, according to the Bramwell–Hill equation. The PTT refers to
the propagation time of the pressure pulse wave between the two measurement locations
by two physiological acquisition systems, such as double photoplethysmography (PPG)
devices [6], and a combination of PPG and electrocardiography (ECG) sensors [7]. However,
the multiple physiological devices could impede wearable applications.
Recently, artificial intelligence (AI) is rapidly evolving for clinical classification and pre-
diction tasks that are helpful for big data analyses for long-term physiological monitoring,
thus providing accurate therapy strategies that can be referenced by physicians [8–10]. For
the AI in the cuffless BP studies, some groups establish the deep learning model between
the physiological signals and arterial BP (ABP) waveforms from the existing databases. For
example, Khalid et al. [11] provided the single-channel PPG-based cuffless BP estimation
model that involved the two databases from the Queensland [12] and the multiparame-
ter intelligent monitoring in intensive care II (MIMIC-II) datasets [13] and satisfied the
Association for the Advancement of Medical Instrumentation (AAMI) standard criteria.
El-Hajj et al. [14] proposed recurrent neural networks (RNN) to establish the correlation
between the PPG and BP signals from the MIMIC-II datasets. Li et al. [15] provided a
long-short-term memory (LSTM)-based deep learning model using the ECG and PPG
signals from the MIMIC-II datasets for a real-time cuffless BP estimation.
Although the aforementioned dataset-based approaches presented the qualified BP
accuracy within the AAMI criteria, some limitations exist in practical applications. First, the
measurement devices with different specifications exist due to the differences in measured
signal morphology that may not be suitable for the developed AI-based BP model in previ-
ous studies. Second, the AI-based regression models were only adaptive for the patients in
the existing datasets, such as the MIMIC-II, restricting the users that are not included in the
datasets from conducting this proposed BP model. Third, personalized health behaviors
affect BP, including lifestyle habits, personal information, and the environment [16]. Nowa-
days, the individual AI-based BP model becomes important in the future implications of
BP management, with an eye towards personalized medicine [17,18]. Personalized BP
healthcare greatly impacts the development of precise therapy, thus providing better BP
control and treatment compliance [19,20]. In consumer health electronics, wearable and
smart functions in devices play an important role in personalized healthcare [21–23].
To this end, this study aims to develop a wearable intelligent bio-impedance system for
personalized continuous BP monitoring. The single-channel impedance plethysmography
(IPG) signal acquisition device was implemented to measure real-time bio-impedance
signals from the pulsation of the carotid artery. Moreover, the proposed system performs
personalized BP model computation by an AI-based algorithm to extract the BP features
hidden from the IPG waveforms, and to further achieve personalized BP healthcare. The
rest of this paper is organized as follows. Section 2 introduces the physiological correlations
between IPG signals and BP, as well as the proposed system design and the ethics statement.
In Section 3, the experiment results for the IPG signal measurement and the BP accuracy
of the proposed system are presented. The discussion and conclusion are provided in
Sections 4 and 5.
 a strong correlation with the cross‐sectional area of the artery. The dP, ρ, D, A, and dA
denote the BP change, blood density, the distance between two physiological measure‐
ment locations of the artery, the arterial cross‐area, and the change in the arterial area,
respectively:
2
Biosensors 2022, 12, 150  D  dA 3 of 13
dP     (1)
 PTT  A
Moreover, the small change in the arterial area can be viewed as the arterial imped‐
2. Materials
ance variation and Methods
[27,28], according to Ohm’s law, as in Equation (2), where Z, L, and σ are
2.1. arterial
the Physiological Correlation
impedance, thebetween
lengthIPG andmeasured
of the BP arterial segment, and the arterial
conductivity, respectively: technique that is commonly applied in noninvasive physiologi-
IPG is a bio-impedance
cal measurements [24]. The IPG technique is based L on the electric impedance measurement
that applies the alternating current into a Zlocal𝜎𝐴
(2)
area of the body, and then measures the
voltage signal.
Thus, In hemodynamic
the arterial pressure canstudies, an IPG-based
be estimated from themeasurement can extract arterial
impedance measurement from
impedance
IPG signals,induced
as shown by in
a small
Figurevariation
1b. Wang in the blood
et al. volume
[29,30] [25,26], as
and Huynh shown
et al. [31] in Figurethe
utilized 1a.
Based
IPG on the Bramwell–Hill
technique to establish theequation, as in model
BP estimation Equation (1), thearterial
between BP hasimpedance
a strong correlation
and pres‐
with
sure bytheacross-sectional area of the
time–domain analysis. artery. The
However, dP, ρ,
the IPG D, A, in
signals andthedA denote the BPanalysis
time–frequency change,
blood density, the distance between two physiological measurement locations
have not been extensively investigated. In this study, the time–frequency analysis‐based of the artery,
the arterial cross-area, and the change in the arterial area, respectively:
continuous wavelet transform (CWT) was performed to extract the BP features in the IPG
signals. Then, the AI‐based regression model was
 used to establish the personalized
D 2 dA

model between the IPG based‐CWT dP features
=ρ and the reference BP from the cuff‐based
(1)
sensor. PTT A

Figure 1. (a)Schematic
1. (a) Schematicofofthe
the IPG
IPG technique
technique forfor hemodynamic
hemodynamic measurement.
measurement. (b) Physiological
(b) Physiological cor‐
relation between arterial impedance and blood pressure.
correlation between arterial impedance and blood pressure.

Moreover,
2.2. Wearable the smallBP
Intelligent change
SysteminDesign
the arterial area can be viewed as the arterial impedance
variation [27,28], according to Ohm’s law, as in Equation (2), where Z, L, and σ are
The wearable intelligent bio‐impedance system was implemented to validate the fea‐
the arterial impedance, the length of the measured arterial segment, and the arterial
sibility of the AI‐based IPG–BP methodology for personalized medicine. The proposed
conductivity, respectively:
system can be divided into two main parts, including an IPG sensing device and the AI‐
L
based BP estimation, as shown in FigureZ2.=The IPG sensing device was installed on the (2)
σA
subject’s neck for the physiological acquisition from the carotid arterial pulsation, owing
Thus, the arterial pressure can be estimated from the impedance measurement from
to the palpable arterial pulsation and the low BP waveform distortion [30]. The measured
IPG signals, as shown in Figure 1b. Wang et al. [29,30] and Huynh et al. [31] utilized the IPG
technique to establish the BP estimation model between arterial impedance and pressure by
a time–domain analysis. However, the IPG signals in the time–frequency analysis have not
been extensively investigated. In this study, the time–frequency analysis-based continuous
wavelet transform (CWT) was performed to extract the BP features in the IPG signals. Then,
the AI-based regression model was used to establish the personalized model between the
IPG based-CWT features and the reference BP from the cuff-based sensor.

2.2. Wearable Intelligent BP System Design

The wearable intelligent bio-impedance system was implemented to validate the
feasibility of the AI-based IPG–BP methodology for personalized medicine. The proposed
system can be divided into two main parts, including an IPG sensing device and the AI-
based BP estimation, as shown in Figure 2. The IPG sensing device was installed on the
subject’s neck for the physiological acquisition from the carotid arterial pulsation, owing to
the palpable arterial pulsation and the low BP waveform distortion [30]. The measured IPG
Biosensors 2022, 12, x FOR PEER REVIEW 4 of 13
Biosensors 2022, 12, 150 4 of 13

IPG signals were transmitted into the proposed AI‐based SSR‐Net model to further com‐
signals
pute thewere transmitted into the proposed AI-based SSR-Net model to further compute
BP information.
the BP information.

Figure 2. Schematic of
Figure of the
theproposed
proposedproof-of-concept
proof‐of‐conceptwearable
wearable intelligent bio-impedance
intelligent system
bio‐impedance for
system
for continuous
continuous BP monitoring,
BP monitoring, including
including IPG sensing,
IPG sensing, IPG excitation
IPG excitation functions,functions, andBP
and AI-based AI‐based BP
estimation.
estimation.
2.2.1. IPG Sensing Device
2.2.1.The
IPGIPGSensing Device
sensing device consisted of four electrodes, an alternating current source,
and aThefront-end analog
IPG sensing device circuit. Two pairs
consisted of of flexible
four electrodes
electrodes, made withcurrent
an alternating a silver-plated
source,
polyester textile with low surface resistivity (<0.05 Ω/inch 2 ) were utilized as the electrical
and a front‐end analog circuit. Two pairs of flexible electrodes made with a silver‐plated
function for
polyester the current
textile with low excitation
surfaceand the physiological
resistivity sensing.
(<0.05 Ω/inch 2) were Each rectangular
utilized electrode
as the electrical
with an area of 2 cm ×
function for the current excitation and the physiological sensing. Each rectangularartery
0.9 cm and a thickness of 0.03 cm was placed on the carotid elec‐
abovewith
trode the neck as the
an area of 2isometric
cm × 0.9distribution with a spacing
cm and a thickness of 0.03 of
cm0.5 cm.placed on the carotid
was
arteryAn alternative
above the neckcurrent
as thesource wasdistribution
isometric implemented by aa combination
with spacing of 0.5ofcm.the Wien–Bridge
oscillator and an improved Howland current pump. The sinusoid waveforms,
An alternative current source was implemented by a combination of the Wien–Bridge with a
frequency of 50 kHz, were produced by the Wen–Bridge oscillator. The
oscillator and an improved Howland current pump. The sinusoid waveforms, with a fre‐ voltage-controlled
current of
quency source (VCCS),
50 kHz, wereimproved
producedby bythe
theHowland
Wen–Bridge current pump,The
oscillator. transforms the signals
voltage‐controlled
from a sinusoid voltage to current waveforms with the amplitude and frequency of 0.14 mA
current source (VCCS), improved by the Howland current pump, transforms the signals
and 50 kHz, which follows the human safety guideline [32,33].
from a sinusoid voltage to current waveforms with the amplitude and frequency of 0.14
To extract the carotid pulse signals in response to the small arterial pulsation, the front-
mA and 50 kHz, which follows the human safety guideline [32,33].
end analog circuit was required to enlarge the small variation in the arterial impedance.
To extract the carotid pulse signals in response to the small arterial pulsation, the
The instrumentation amplifier (AD8421, Analog Devices Inc., Norwood, MA, USA) was
front‐end analog circuit was required to enlarge the small variation in the arterial imped‐
used to provide an amplification gain of 1000 v/v and a high common-mode rejection ratio
ance. The instrumentation amplifier (AD8421, Analog Devices Inc., Norwood, MA, USA)
of 110 dB at the input frequency of 50 kHz. The demodulator (AD8310, Analog Devices
was used to provide an amplification gain of 1000 v/v and a high common‐mode rejection
Inc., Norwood, MA, USA) employs the 50 kHz carrier signal removal from the stage of the
ratio of 110 dB at the input frequency of 50 kHz. The demodulator (AD8310, Analog De‐
instrumentation amplifier. The fourth-order Butterworth bandpass filter, with a narrow
vices Inc., Norwood, MA, USA) employs the 50 kHz carrier signal removal from the stage
bandwidth (0.3–5 Hz), was utilized to cover the typical heart rate ranges from 0.67 Hz to
of the
3.33 Hzinstrumentation
[34]. The analogamplifier. The
IPG signals fourth‐order
were Butterworth
sent into the bandpass
analog-to-digital filter, with
converter a
device
narrow bandwidth (0.3–5 Hz), was utilized to cover the typical
(myDAQ, National Instruments, Austin, TX, USA) for further signal processing. heart rate ranges from 0.67
Hz to 3.33 Hz [34]. The analog IPG signals were sent into the analog‐to‐digital converter
device (myDAQ,
2.2.2. The AI-Based National Instruments, Austin, TX, USA) for further signal processing.
BP Estimation
In this study, we utilized a deep learning architecture-based SSR-Net model for cuffless
2.2.2. The AI‐Based
continuous BP Estimation
BP monitoring. The SSR-Net model is based on the convolutional neural
 with different weighting were implemented, according to Equations (3) and (4). The test
results show that the optimal design for the penalty term that was weighted three times
Biosensors 2022, 12, 150 in the interval below Q1 and above Q3 obtained a lower epoch to reach the model conver‐
5 of 13
gence, as shown in Figure 4d.
Loss  (BPIPG - BPCuff )  |, BPCuff  Q1 or BPCuff  Q3 (3)
network (CNN) architecture [35]. The SSR-Net model has the merit of a lightweight and
complementary two-stream Loss  BPIPG[36]
structure
Cuff |, Q1  BPCuff  Q3
- BPthat is suitable for real-time monitoring and (4)
discrete numerical predictions in BP applications. We also compared the SSR-Net and other
lightweight deep learning
Environment Details:models, as shown in Table 1. Compared to the MobileNet-V2 and
the LSTM approaches
The Python‐based forsoftware
our database
was training,
used to designthe SSR-Net provides an deep
and implement ultra-lightweight
neural net‐
model size of 213 KB and lower model parameters of 0.04 M, resulting
works. The experiments were performed using Python 3.6.8 inside the Windows in a CPU interface
10 En‐
time of 0.17
terprise s. Thewith
computer experimental flowchart
an Intel® Core™ of this
i7‐8700 4.6study is as follows:
GHz processor. IPG signals
Moreover, 64 GBand
of
cuff-based BP measurements, IPG signal pre-processing for BP feature extraction, dataset
RAM and a GeForce GTX 1660 Ti 6 GB GPU were equipped on the computer.
arrangement for model training and testing, and a loss function design for the personalized
BP monitoring,
2.2.3. as shown in Figure 3a.
Ethics Statement
TableThe human experiment
1. Comparison of differentwas permitted
lightweight bylearning
deep the Institutional
models. Review Board of National
Yang Ming Chiao Tung University (NCTU‐REC‐109‐012E). A total of six healthy subjects
(three males and three females) participated
Model SSR-Net in the experiment,
MobileNet-V2 with an age ofLSTM
24 ± 1 years,
a heightModel
of 165size
± 8 cm, and a weight213of
KB66 ± 13 kg. The participants
13,932 KB consented8744
to participate
KB
and Model
provided their written informed
parameters 0.04 M approval. During 3.50 the
M experiment, 215.99
they were
M in‐
Inference
structed time on CPU
to remain 0.17 s for the physiological
in a sitting position 0.29 s measurement. 0.25 s

Figure
Figure 3.
3. (a)
(a)Experiment
Experimentflow chart
flow forfor
chart physiological acquisition,
physiological datadata
acquisition, pre‐processing, and BP
pre-processing, andesti‐
BP
mation. (b) Synchronous measurement between the proposed sensor and cuff‐based
estimation. (b) Synchronous measurement between the proposed sensor and cuff-based device.device. (c) IPG
signals
(c) IPG were
signalspre‐processed for further
were pre-processed for BP feature
further BPextraction, includingincluding
feature extraction, signal segmentation and con‐
signal segmentation
tinuous wavelet transform.
and continuous wavelet transform.

• IPG Signals and Reference BP Acquisition:

To establish the personalized database for the wearable intelligent bio-impedance
BP system, the synchronous measurement for the IPG sensing device and the digital
electronic sphygmomanometer (HEM-1000, OMRON, Osaka, Japan) was conducted for
30 measurement trials in the manner shown in Figure 3b. Each trial consisted of SBP and
DBP from the cuff-based sensor and five consecutive IPG waveforms from the IPG sensing
device, each taking 1 min, due to the 30 s measurement period of the cuff-based BP and
Biosensors 2022, 12, 150 6 of 13

the resting of the arteries for 30 s. Thus, the overall acquisition procedure took 30 min for
30 measurement trials in our experiment.
• IPG Signal Pre-processing and BP Feature Extraction:
For each trial, the five consecutive IPG waveforms were segmented into individual
waveforms for further feature extraction. The individual IPG signal was converted into a
time–frequency analysis by CWT [37]. In this work, the Daubechies 8 (db8) wavelet was
used to transform the IPG signals. As shown in Figure 3c, the five pairs of references and
features in SBP and DBP were obtained in each measurement trial; thereby, the overall
physiological data, with 150 pairs of physiological data for each subject, were acquired.
• Dataset Arrangements for Model Training and Testing:
Overall, 150 pairs of the CWT-IPG based images, and the corresponding reference
SBP and DBP values, were categorized into the training and testing datasets. One hundred
and twenty pairs of all the datasets were selected for personalized BP model training, and
30 pairs were used to test the BP model accuracy.
• Loss Function Design for Personalized BP Monitoring:
In the process of AI training, the loss function is designed to allow the model to learn
the prediction error between the predicted BP by the IPG signal feature (BPIPG ) and the
actual BP by the cuff-based device (BPcuff ) to further obtain model convergence. In this
study, the BP estimation is categorized as a linear regression problem, and it utilizes the
mean absolute error (MAE) as the loss function. To obtain the converged personalized BP
model during the training stage, the penalty term, in terms of loss function, was modified
based on the reference BP distribution of each subject. The penalty term was designed
based on being below quartile 1 (Q1) and above quartile 3 (Q3) of the measured reference
BP, to accelerate the converge time of the personalized model. Figure 4a–c demonstrates all
subjects’ SBP and DBP values, and their quartiles, from the cuff-sensor that was measured
30 times. To evaluate the converge time for the proposed model, the penalty terms with
different weighting were implemented, according to Equations (3) and (4). The test results
show that the optimal design for the penalty term that was weighted three times in the
interval below Q1 and above Q3 obtained a lower epoch to reach the model convergence,
as shown in Figure 4d.

Loss = | (BP IPG − BPCuff ) × α|, BPCuff < Q1 or BPCuff > Q3 (3)

Loss = | BPIPG − BPCuff |, Q1 < BPCuff < Q3 (4)

• Environment Details:
The Python-based software was used to design and implement deep neural networks.
The experiments were performed using Python 3.6.8 inside the Windows 10 Enterprise
computer with an Intel® Core™ i7-8700 4.6 GHz processor. Moreover, 64 GB of RAM and a
GeForce GTX 1660 Ti 6 GB GPU were equipped on the computer.

2.2.3. Ethics Statement

The human experiment was permitted by the Institutional Review Board of National
Yang Ming Chiao Tung University (NCTU-REC-109-012E). A total of six healthy subjects
(three males and three females) participated in the experiment, with an age of 24 ± 1 years,
a height of 165 ± 8 cm, and a weight of 66 ± 13 kg. The participants consented to participate
and provided their written informed approval. During the experiment, they were instructed
to remain in a sitting position for the physiological measurement.
Biosensors 2022, 12, x FOR PEER REVIEW 7 of 13
Biosensors 2022, 12, 150 7 of 13

Figure 4. (a)
Figure 4. (a) Reference
Reference SBP
SBP and
and (b)(b) DBP
DBP distribution
distribution for
for six
six participants.
participants. (c)
(c) Statistical
Statistical BP
BP results
results in
in
quartiles
quartiles 1,
1, 2,
2, and
and 33 within
within six
six subjects.
subjects. (d)
(d) Performance
Performance evaluation
evaluation of model
model convergence
convergence for thethe
penalty
penalty terms
terms with
with different
different weighting
weighting inin the
theinterval
intervalbelow
belowquartile
quartile11and
andabove
abovequartile
quartile3.3.

3.
3. Results
Results
3.1. IPG Signal Measurement and Feature Extraction
3.1. IPG Signal Measurement and Feature Extraction
The IPG sensing device and digital electronic sphygmomanometer were synchronously
The IPG sensing device and digital electronic sphygmomanometer were synchro‐
conducted for 30 min, as shown in Figure 5a. The consecutive IPG waveforms from the
nously conducted
carotid artery abovefor 30subject’s
the min, as shown in Figure
neck were 5a. The
measured by consecutive
the proposedIPG waveforms
system. To alignfrom
the
the carotid artery above the subject’s neck were measured by the proposed
measurement procedure of the cuff-based BP device, the five IPG waveforms were system. To
selected
align the
before andmeasurement procedure
after the operating time ofofthe
thecuff
cuff‐based BPshown
device, as device, the five5b.
in Figure IPG waveforms
In the stage of
were selected before and after the operating time of the cuff device, as
the BP feature extraction, the IPG signals were divided into a single waveform to shown infurther
Figure
perform the time–frequency analysis by the CWT method, as shown in Figure 5c. single
5b. In the stage of the BP feature extraction, the IPG signals were divided into a
waveform to further perform the time–frequency analysis by the CWT method, as shown
3.2.Figure
in 5c.
BP Accuracy Evaluation
The box plot analysis shows the measured SBP and DBP distributions from the cuff
3.2. BP Accuracy
device and EvaluationIPG-based system, as shown in Figure 6a,b. The mean SBP
the proposed
fromThethe box
cuffplot analysis
device shows the
(HEM-1000, measured
OMRON) SBP
and theand DBP distributions
IPG-based fromobtained
system were the cuff
device
with 119.64 ± 4.32
and the proposed
mmHgIPG‐based system, as
(range: 110–130 shownand
mmHg) in Figure ± 3.61
121.346a,b. The mean
mmHg SBP from
(range:
the cuff mmHg);
112–129 device (HEM‐1000,
for the DBP,OMRON)
the mean and
was the IPG‐based
obtained 70.13 ± were
with system obtained
4.44 mmHg with
(range:
119.64mmHg)
56–79 ± 4.32 mmHg
and 71.69 ± 4.17
(range: 110–130
mmHgmmHg) and 121.34
(range: 59–90 mmHg).± 3.61 mmHg (range: 112–129
mmHg); for the DBP, the mean was obtained with 70.13 ± 4.44 mmHg (range: 56–79
mmHg) and 71.69 ± 4.17 mmHg (range: 59–90 mmHg).
To further assess the accuracy performance of the proposed IPG‐based system, the
statistical results in terms of the mean error (ME) and the MAE, using Bland–Altman plots,
were performed as in the evaluation index, according to Equations (5) and (6):
1 n
ME   y i  xi
n i 1
(5)
 posed IPG‐based system, and n is the number of the testing dataset.
The SBP accuracy in the testing results indicated that the ME was 1.69 ± 3.28 mmHg
(Figure 6c) and the MAE was 2.63 ± 2.58 mmHg (Figure 6e). The ME and MAE were 1.56
± 3.32 mmHg and 2.66 ± 2.52 mmHg, respectively, for DBP, as shown in Figure 6d,f. Thus,
Biosensors 2022, 12, 150 the BP performance of the proposed system satisfied with the standard criteria of8 the
of 13
AAMI by less than 5 ± 8 mmHg.

Figure 5. (a) IPG signals from a carotid artery for 30 minutes’ measurement. (b) Five consecutive IPG
waveforms before and after the operating time of the cuff device. (c) IPG signal segmentation and
further continuous wavelet transform.
Biosensors 2022, 12, x FOR PEER REVIEW 9 of 13

Figure 5. (a) IPG signals from a carotid artery for 30 minutes’ measurement. (b) Five consecutive
Biosensors 2022, 12, 150 IPG waveforms before and after the operating time of the cuff device. (c) IPG signal segmentation
9 of 13
and further continuous wavelet transform.

Figure 6. Box
Figure 6. Box plot
plot analysis
analysis for
for subject’s
subject’s (a)
(a) SBP
SBP and
and (b)
(b) DBP
DBP distribution
distribution from
from the
the cuff
cuff device
device and
and
proposed IPG-based system. Bland–Altman plot analysis in terms of ME for (c)
proposed IPG‐based system. Bland–Altman plot analysis in terms of ME for (c) SBP and (d) SBP and (d) DBP,
DBP,
respectively.
respectively. Estimation
Estimation Error
Errorin interms
termsofofMAE
MAEforfor(e)
(e)SBP
SBPand
and(f)
(f)DBP.
DBP.

To further assess the accuracy performance of the proposed IPG-based system, the
4. Discussion
statistical results in terms of the mean error (ME) and the MAE, using Bland–Altman plots,
4.1. Innovation of Proposed Intelligent Bio‐Impedance System
were performed as in the evaluation index, according to Equations (5) and (6):
Wearable and intelligent healthcare greatly impacts the development of therapy and
n
management in BP healthcare, especially in1 personalized medicine [22,38]. The main con‐
ME = ∑ yi − xi (5)
tribution of this study is to provide a single‐channel
n i =1 bio‐impedance‐based intelligent sys‐
tem with sensing and prediction functions for personalized BP applications.
1 n
MAE = ∑ |yi − xi | (6)
n i =1
where xi is the digital electronic sphygmomanometer, yi is the predicted value of the
proposed IPG-based system, and n is the number of the testing dataset.
Biosensors 2022, 12, 150 10 of 13

The SBP accuracy in the testing results indicated that the ME was 1.69 ± 3.28 mmHg
(Figure 6c) and the MAE was 2.63 ± 2.58 mmHg (Figure 6e). The ME and MAE were
1.56 ± 3.32 mmHg and 2.66 ± 2.52 mmHg, respectively, for DBP, as shown in Figure 6d,f.
Thus, the BP performance of the proposed system satisfied with the standard criteria of the
AAMI by less than 5 ± 8 mmHg.

4. Discussion
4.1. Innovation of Proposed Intelligent Bio-Impedance System
Wearable and intelligent healthcare greatly impacts the development of therapy and
management in BP healthcare, especially in personalized medicine [22,38]. The main
contribution of this study is to provide a single-channel bio-impedance-based intelligent
system with sensing and prediction functions for personalized BP applications.
For the novelty of the sensor configuration design, the proposed single-channel IPG-BP
sensor is beneficial for the hardware complexity reduction and for wearable applications,
compared to the PTT-based approaches using a multi-channel physiological measurement
for BP estimation [31,39,40]. For the novelty of personalized BP healthcare, the proposed
intelligent BP system, using an ultra-lightweight AI algorithm, can establish the tailored BP
model from the measured signals from each subject. Compared to the AI-based cuffless BP
algorithm in existing datasets, such as the MIMIC [11,14,15], the proposed system provides
an adaptive BP regression model for each person based on individually measured signals.
Such an intelligent BP system design is suitable for personalized healthcare development.

4.2. BP Measurement Performance

To evaluate the measurement performance of the proposed wearable intelligent system
in BP monitoring applications, the digital electronic sphygmomanometer (as a reference de-
vice) was installed for synchronous measurement. The Bland–Altman plot was performed
to evaluate the difference in the BP measured by the proposed system and the reference BP
by the cuff-based device. Six healthy subjects (three males and three females) with a mean
age of 24 ± 1 participated in the human trial. The statistical results of the ME and MAE
were utilized as the evaluation metrics to assess the BP accuracy. Based on the statistical
results, the accuracy of SBP, in terms of the ME and MAE, was 1.69 ± 3.28 mmHg and
2.63 ± 2.58 mmHg, respectively. The DBP estimation error demonstrated that the ME and
MAE were 1.56 ± 3.32 mmHg and 2.66 ± 2.52 mmHg, respectively. The BP performance of
the proposed system was satisfied with specifications (less than 5 ± 8 mmHg) based on the
standard criteria of the AAMI.

4.3. Comparisons with Previous Cuffless BP Works

We compared the proposed IPG-based BP intelligent system with recent studies, as
shown in Table 2. Several works relied on multiple physiological parameters to establish
the BP model, such as the ECG and the pressure pulse waveform (PPW) [41], two PPG
sensors [6], and a combination of the PPG and phonocardiogram (PCG) [42]. Despite
the multiple physiological acquisition channels, the approach to providing satisfied BP
performance using multi-sensor implementation could increase the complexity of the
measurement procedure in practical applications. Recently, several groups attempted to
develop deep learning-based BP models using a one-channel physiological signal. For
example, Miao et al. [43] proposed deep learning architecture combined with a residual
network with LSTM to establish highly accurate BP modeling using the spatial-temporal
information of the one-channel ECG signal in the database collected from Fuwai Hospital,
Chinese Academy of Medical Sciences. Dal Pont et al. [14] presented an attention-based
RNN for cuffless BP measurements using single-channel PPG signals from the MIMIC-II
database. Compared to existing dataset-based approaches, we provided a deep learning-
based personalized BP model based on the measured signals from an actual single-channel
IPG device. The significance of our work is to develop a single-channel wearable sen-
Biosensors 2022, 12, 150 11 of 13

sor combined with an AI-based personalized BP model that is suitable for personalized
medicine development.

Table 2. Comparison of cuffless continuous BP measurement technologies.

Deep Statistical BP Estimation Error

Author Physiological Learning
Signal Results
Model SBP DBP
Miao et al. [41] ECG, 2-PPW - ME ± SD 1.62 ± 7.76 1.49 ± 5.52
Tabei et al. [6] 2-PPG - MAE ± SD 2.07 ± 2.06 2.12 ± 1.85
Marzorati et al. [42] PPG, PCG - ME ± SD 1.47 ± 3.76 0.01 ± 7.55
Miao et al. [43] ECG Res-LSTM ME ± SD −0.22 ± 5.82 −0.75 ± 5.62
El-Hajj et al. [14] Attention ME ± SD −0.52 ± 4.22 −0.66 ± 2.07
PPG MAE ± SD 2.58 ± 3.35 1.26 ± 1.63
based-RNN
ME ± SD 1.69 ± 3.28 1.56 ± 3.32
Our work IPG SSR-Net MAE ± SD 2.63 ± 2.58 2.66 ± 2.52

4.4. Limitations and Future Works

Although this work presented an intelligent bio-impedance system for personalized BP
monitoring, as well as validating its functional efficacy and BP accuracy, some limitations
require further improvement. First, the participants are young healthy people in our
experiment. More old-aged subjects and patients with cardiovascular disease will be
recruited to make the BP accuracy more reliable in the clinical aspect. Second, the IPG
sensing device and the AI-based model will be integrated into a hardware implementation
for practical applications. Third, the evaluation of the frequency of calibration will be
performed to qualify BP monitoring with changes in measurement conditions, such as an
atmospheric humidity-induced skin-electrode impedance change. Fourth, the ambulatory
BP measurement technique by the analog front-end and post-processing improvements
will be developed for practical applications and daily activities.

5. Conclusions
This study develops a proof-of-concept wearable intelligent system for personalized
BP healthcare. The proposed system integrates a one-channel IPG sensing device and an
AI-based regression model for cuffless continuous BP measurement. Compared to the
PTT-based BP device and the MIMIC series dataset-based BP estimation model, our system
provides a combined solution with the merits of wearable and intelligent properties in
continuous BP measurement. In the accuracy evaluation, the experimental results validated
the feasibility of the proposed system, resulting in qualified BP performance. Overall, our
work develops a novel BP system to present an insightful view towards next-generation
personalized BP healthcare.

Author Contributions: T.-W.W. conceived of the presented idea, supervised the experiment’s
progress, and wrote the manuscript. J.-Y.S. developed artificial intelligence software and data analysis.
H.-W.C. and L.C. implemented the system and performed the human trials and validation. Y.-L.S.
performed the data analysis and graphing. E.E. and O.E. performed investigation and visualiza-
tion. T.-T.L. provided the resources and funding. S.-F.L. supervised the project and reviewed the
manuscript. All authors have read and agreed to the published version of the manuscript.
Funding: This research was funded by Ministry of Science and Technology, Taiwan under fund-
ing number MOST 110-2917-I-564-026, MOST 110-2628-B-002-055, and MOST 109-2628-B-002-033;
University-Industry Collaboration (National Yang Ming Chiao Tung University and Leadtek Research
Inc.) under Grant 109A159.
Institutional Review Board Statement: The study was conducted in accordance with the Declaration
of Helsinki, and approved by the Institutional Review Board of National Yang Ming Chiao Tung
University (approval numbers: NCTU-REC-109-012E).
Informed Consent Statement: Informed consent was obtained from all subjects involved in the study.
Conflicts of Interest: The authors declare no conflict of interest.
Biosensors 2022, 12, 150 12 of 13

References
1. Georgi, N.; Kuchenbuch, M.; Corvol, A.; Jeannés, R.L.B. An Overview of Blood Pressure Measurement in Telemonitoring Context.
IEEE Consum. Electron. Mag. 2020, 9, 42–50. [CrossRef]
2. Yano, Y. Blood pressure management in an ecosystem context. Hypertens. Res. 2020, 43, 989–994. [CrossRef] [PubMed]
3. Fortin, J.; Rogge, D.E.; Fellner, C.; Flotzinger, D.; Grond, J.; Lerche, K.; Saugel, B. A novel art of continuous noninvasive blood
pressure measurement. Nat. Commun. 2021, 12, 1387. [CrossRef] [PubMed]
4. Kario, K. Management of Hypertension in the Digital Era: Small Wearable Monitoring Devices for Remote Blood Pressure
Monitoring. Hypertension 2020, 76, 640–650. [CrossRef]
5. Wang, T.-W.; Lin, S.-F. Wearable Piezoelectric-Based System for Continuous Beat-to-Beat Blood Pressure Measurement. Sensors
2020, 20, 851. [CrossRef] [PubMed]
6. Tabei, F.; Gresham, J.M.; Askarian, B.; Jung, K.; Chong, J.W. Cuff-Less Blood Pressure Monitoring System Using Smartphones.
IEEE Access 2020, 8, 11534–11545. [CrossRef]
7. Ding, X.R.; Zhang, Y.T.; Liu, J.; Dai, W.X.; Tsang, H.K. Continuous Cuffless Blood Pressure Estimation Using Pulse Transit Time
and Photoplethysmogram Intensity Ratio. IEEE Trans. Bio-Med. Eng. 2016, 63, 964–972. [CrossRef]
8. Mateen, B.A.; Liley, J.; Denniston, A.K.; Holmes, C.C.; Vollmer, S.J. Improving the quality of machine learning in health
applications and clinical research. Nat. Mach. Intell. 2020, 2, 554–556. [CrossRef]
9. Yu, K.-H.; Beam, A.L.; Kohane, I.S. Artificial intelligence in healthcare. Nat. Biomed. Eng. 2018, 2, 719–731. [CrossRef]
10. Li, R.C.; Asch, S.M.; Shah, N.H. Developing a delivery science for artificial intelligence in healthcare. NPJ Digit. Med. 2020, 3, 107.
[CrossRef]
11. Khalid, S.G.; Liu, H.; Zia, T.; Zhang, J.; Chen, F.; Zheng, D. Cuffless Blood Pressure Estimation Using Single Channel Photo-
plethysmography: A Two-Step Method. IEEE Access 2020, 8, 58146–58154. [CrossRef]
12. Liu, D.; Görges, M.; Jenkins, S.A. University of Queensland vital signs dataset: Development of an accessible repository of
anesthesia patient monitoring data for research. Anesth. Analg. 2012, 114, 584–589. [CrossRef]
13. Saeed, M.; Villarroel, M.; Reisner, A.T.; Clifford, G.; Lehman, L.W.; Moody, G.; Heldt, T.; Kyaw, T.H.; Moody, B.; Mark, R.G.
Multiparameter Intelligent Monitoring in Intensive Care II: A public-access intensive care unit database. Crit. Care Med. 2011, 39,
952–960. [CrossRef] [PubMed]
14. El-Hajj, C.; Kyriacou, P.A. Deep learning models for cuffless blood pressure monitoring from PPG signals using attention
mechanism. Biomed. Signal Process. Control 2021, 65, 102301. [CrossRef]
15. Li, Y.-H.; Harfiya, L.N.; Purwandari, K.; Lin, Y.-D. Real-Time Cuffless Continuous Blood Pressure Estimation Using Deep Learning
Model. Sensors 2020, 20, 5606. [CrossRef] [PubMed]
16. Chiang, P.; Dey, S. Personalized Effect of Health Behavior on Blood Pressure: Machine Learning Based Prediction and Recom-
mendation. In Proceedings of the 2018 IEEE 20th International Conference on e-Health Networking, Applications and Services
(Healthcom), Ostrava, Czech Republic, 17–20 September 2018; pp. 1–6.
17. Krittanawong, C.; Bomback, A.S.; Baber, U.; Bangalore, S.; Messerli, F.H.; Wilson Tang, W.H. Future Direction for Using Artificial
Intelligence to Predict and Manage Hypertension. Curr. Hypertens. Rep. 2018, 20, 75. [CrossRef] [PubMed]
18. Bhudia, R.P. Treatment of the hypertensive patient in 2030. J. Hum. Hypertens. 2020, 35, 818–820. [CrossRef] [PubMed]
19. Mueller, F.B. AI (Artificial Intelligence) and Hypertension Research. Curr. Hypertens. Rep. 2020, 22, 70. [CrossRef]
20. Krittanawong, C.; Rogers, A.J.; Johnson, K.W.; Wang, Z.; Turakhia, M.P.; Halperin, J.L.; Narayan, S.M. Integration of novel
monitoring devices with machine learning technology for scalable cardiovascular management. Nat. Rev. Cardiol. 2021, 18, 75–91.
[CrossRef]
21. Morra, L.; Mohanty, S.P.; Lamberti, F. Artificial Intelligence in Consumer Electronics. IEEE Consum. Electron. Mag. 2020, 9, 46–47.
[CrossRef]
22. Pirbhulal, S.; Wu, W.; Li, G.; Sangaiah, A.K. Medical Information Security for Wearable Body Sensor Networks in Smart Healthcare.
IEEE Consum. Electron. Mag. 2019, 8, 37–41. [CrossRef]
23. Kirk, S. The Wearables Revolution: Is Standardization a Help or a Hindrance?: Mainstream technology or just a passing phase?
IEEE Consum. Electron. Mag. 2014, 3, 45–50. [CrossRef]
24. Wang, T.-W.; Sung, Y.-L.; Chu, H.-W.; Lin, S.-F. IPG-based field potential measurement of cultured cardiomyocytes for optogenetic
applications. Biosens. Bioelectron. 2021, 179, 113060. [CrossRef] [PubMed]
25. Wang, T.-W.; Chu, H.-W.; Chou, L.; Sung, Y.-L.; Shih, Y.-T.; Hsu, P.-C.; Cheng, H.-M.; Lin, S.-F. Bio-Impedance Measurement
Optimization for High-Resolution Carotid Pulse Sensing. Sensors 2021, 21, 1600. [CrossRef]
26. Pesti, K.; Metshein, M.; Annus, P.; Kõiv, H.; Min, M. Electrode Placement Strategies for the Measurement of Radial Artery
Bioimpedance: Simulations and Experiments. IEEE Trans. Instrum. Meas. 2021, 70, 1–10. [CrossRef]
27. Bera, T.K. Bioelectrical Impedance Methods for Noninvasive Health Monitoring: A Review. J. Med. Eng. 2014, 2014, 381251.
[CrossRef]
28. Swanson, D.; Webster, J. Origin of the Electrical Impedance Pulse in the Limbs. In Proceedings of the 29th Annual Conference on
Engineering in Medicine & Biology, Lyon, France, 22–26 August 2007; p. 324.
29. Wang, T.W.; Chen, W.X.; Chu, H.W.; Lin, S.F. Single-Channel Bioimpedance Measurement for Wearable Continuous Blood
Pressure Monitoring. IEEE Trans. Instrum. Meas. 2021, 70, 1–9. [CrossRef]
Biosensors 2022, 12, 150 13 of 13

30. Wang, T.W.; Chu, H.W.; Chen, W.X.; Shih, Y.T.; Hsu, P.C.; Cheng, H.M.; Lin, S.F. Single-Channel Impedance Plethysmography
Neck Patch Device for Unobtrusive Wearable Cardiovascular Monitoring. IEEE Access 2020, 8, 184909–184919. [CrossRef]
31. Huynh, T.H.; Jafari, R.; Chung, W.Y. Noninvasive Cuffless Blood Pressure Estimation Using Pulse Transit Time and Impedance
Plethysmography. IEEE Trans. Bio-Med. Eng. 2019, 66, 967–976. [CrossRef]
32. Anand, G.; Lowe, A.; Al-Jumaily, A. Simulation of impedance measurements at human forearm within 1 kHz to 2 MHz. J. Electr.
Bioimpedance 2016, 7, 20–27. [CrossRef]
33. Naranjo-Hernández, D.; Reina-Tosina, J.; Min, M. Fundamentals, Recent Advances, and Future Challenges in Bioimpedance
Devices for Healthcare Applications. J. Sens. 2019, 2019, 9210258. [CrossRef]
34. Wang, J.-J.; Wei-Chih, H.; Kao, T.; Liu, C.-P.; Lin, S.-K. Development of forearm impedance plethysmography for the minimally
invasive monitoring of cardiac pumping function. J. Biomed. Sci. Eng. 2011, 4, 122–129. [CrossRef]
35. Yang, T.-Y.; Huang, Y.-H.; Lin, Y.-Y.; Hsiu, P.-C.; Chuang, Y.-Y. SSR-Net: A Compact Soft Stagewise Regression Network for Age
Estimation. In Proceedings of the IJCAI, Stockholm, Sweden, 13–19 July 2018; p. 7.
36. Yang, T.-Y.; Hsu, J.-H.; Lin, Y.-Y.; Chuang, Y.-Y. Deepcd: Learning deep complementary descriptors for patch representations. In
Proceedings of the IEEE International Conference on Computer Vision, Venice, Italy, 22–29 October 2017; pp. 3314–3322.
37. Mallat, S. A Wavelet Tour of Signal Processing; Elsevier: Amsterdam, The Netherlands, 1999.
38. Thapliyal, H. Internet of Things-Based Consumer Electronics: Reviewing Existing Consumer Electronic Devices, Systems, and
Platforms and Exploring New Research Paradigms. IEEE Consum. Electron. Mag. 2018, 7, 66–67. [CrossRef]
39. Ibrahim, B.; Jafari, R. Cuffless Blood Pressure Monitoring from an Array of Wrist Bio-Impedance Sensors Using Subject-Specific
Regression Models: Proof of Concept. IEEE Trans. Biomed. Circuits Syst. 2019, 13, 1723–1735. [CrossRef]
40. Huynh, T.H.; Jafari, R.; Chung, W.-Y. A Robust Bioimpedance Structure for Smartwatch-Based Blood Pressure Monitoring. Sensors
2018, 18, 2095. [CrossRef] [PubMed]
41. Miao, F.; Liu, Z.D.; Liu, J.K.; Wen, B.; He, Q.Y.; Li, Y. Multi-Sensor Fusion Approach for Cuff-Less Blood Pressure Measurement.
IEEE J. Biomed. Health Inform. 2020, 24, 79–91. [CrossRef]
42. Marzorati, D.; Bovio, D.; Salito, C.; Mainardi, L.; Cerveri, P. Chest Wearable Apparatus for Cuffless Continuous Blood Pressure
Measurements Based on PPG and PCG Signals. IEEE Access 2020, 8, 55424–55437. [CrossRef]
43. Miao, F.; Wen, B.; Hu, Z.; Fortino, G.; Wang, X.-P.; Liu, Z.-D.; Tang, M.; Li, Y. Continuous blood pressure measurement from
one-channel electrocardiogram signal using deep-learning techniques. Artif. Intell. Med. 2020, 108, 101919. [CrossRef]

View publication stats