SCIENCE CHINA

Information Sciences
. LETTER . April 2020, Vol. 63 149402:1–149402:3
https://doi.org/10.1007/s11432-018-9719-9

A digital signal processor (DSP)-based system for

embedded continuous-time cuffless blood pressure
monitoring using single-channel PPG signal

Qirui ZHANG† , Qingsong XIE† , Kefeng DUAN, Bo LIANG,

Min WANG & Guoxing WANG∗
Department of Micro-Nano Electronics, Shanghai Jiao Tong University, Shanghai 200240, China

Received 10 July 2018/Revised 11 September 2018/Accepted 19 December 2018/Published online 5 March 2020

Citation Zhang Q R, Xie Q S, Duan K F, et al. A digital signal processor (DSP)-based system for embedded
continuous-time cuffless blood pressure monitoring using single-channel PPG signal. Sci China Inf Sci, 2020, 63(4):
149402, https://doi.org/10.1007/s11432-018-9719-9

Dear editor, photoplethysmograph (PPG) signal. Figure 1(a)

According to a 2015 World Health Organization illustrates the overall architecture of the system.
(WHO) report, cardiovascular diseases accounted First, the analog PPG signal from the sensor mod-
for the death of approximately 17 million peo- ule is quantized by an analog-to-digital converter
ple worldwide each year, i.e., 37% of the annual (ADC) in the DSP system. Further, to improve
global deaths [1]. Therefore, the prevention and signal quality, a morphological filter (MF) and a
prediction of cardiovascular diseases are urgent low-pass filter (LPF) are applied to the PPG sig-
global need to ease future diagnosis and treatment. nal. After preprocessing, just two features are
Blood pressure (BP) is an important indicator for extracted and normalized from the PPG signal.
cardiovascular conditions. Continuous ambulatory These features are then fed into the inference func-
BP is a powerful indicator of immanent cardio- tion of a least squares support vector machine
vascular events than clinic or office BP measure- (LSSVM) [4] to calculate the BP estimation re-
ments [2]. Additionally, at-home BP monitoring sults. The LSSVM model parameters are stored
prevents white-coat hypertension and is superior in the read-only memory (ROM) of the DSP sys-
to clinic BP measurement in predicting important tem and are loaded into the DSP system from a
end points like mortality and geriatric functional personal computer (PC) using a universal serial
decline [3]. Although conventional cuff-based BP bus (USB) when the DSP system is initialized.
monitoring provides accurate measurements, the The estimated systolic BP (SBP) and diastolic
method cannot be used for continuous-time BP BP (DBP) values and heart rate (HR) together
monitoring and can cause discomfort and incon- with preprocessed PPG signal are then transmit-
venience to the user. Therefore, new methods are ted to the PC via a universal asynchronous receiver
needed for user-friendly, continuous, and cuffless and transmitter (UART) interface. The PPG sig-
at-home BP monitoring. nal, HR, and BP values are then displayed using
This study proposes a user-friendly digital a graphical user interface (GUI) to allow real-time
signal processing (DSP) system for embedded monitoring.
continuous-time cuffless blood pressure (BP) mon- Activities such as respiration or slight move-
itoring. The device utilizes only a single-channel ments of the system user can cause baseline wan-
* Corresponding author (email: guoxing@sjtu.edu.cn)
† Zhang Q R and Xie Q S have equal contribution to this work.

c Science China Press and Springer-Verlag GmbH Germany, part of Springer Nature 2020
info.scichina.com link.springer.com
 Zhang Q R, et al. Sci China Inf Sci April 2020 Vol. 63 149402:2

Digital signal processor

Personal computer
Feature extractor
Preprocessing LSSVM inference UART GUI
PPG sensor filter P T
pd function
ADC D Trained model
ROM
O0 O1 USB parameters
Tc

(a)
Estimated blood pressure and heart rate
(updated every 10 s) Preprocessed PPG signal
GUI for preprocessed PPG
signal, estimated blood
pressure and heart rate

Commercial
DSP BP meter

PPG sensor

DBP SBP
(c)
(b)

Figure 1 (Color online) (a) Overall system architecture; (b) prototype for the proposed system; (c) GUI of the proposed
system.

dering (BW) interference, which in turn causes respectively.

severe distortion in the original PPG signal. To
mitigate BW interference, the MF from [5] is ap- f − LB(f )
fscaled = , (1)
plied to the PPG signal. The BW is calculated UB(f ) − LB(f )
from the MF and subtracted from the PPG sig-
N
nal. Further, a finite impulse response (FIR) LPF X
f (x) = αk K(x, xk ) + b, (2)
with 5 Hz cut-off frequency is applied to eliminate
k=1
high-frequency noise.
kx − xk k2
 
When implementing a machine learning-based K(x, xk ) = exp − . (3)
estimation model in embedded systems, using ex- 2σ 2
travagant features might lead to an impractical The regression version of the LSSVM algo-
computational burden. To balance the trade-off rithm [4] is applied in the proposed system to esti-
between computational complexity and model ac- mate the BP value from these extracted features.
curacy, only two features are extracted from the The inference function of LSSVM is displayed in
PPG signal in our proposed system shown in Fig- (2), where f (x) is the estimation result for a given
ure 1(a). The first feature f1 = 1/Tc is the instan- feature-input vector x. αk is the Lagrange multi-
taneous HR, which is the inverse of PPG period plier or weight coefficient, and xk is the support
Tc . The second feature f2 is Tpd , which is the time vector, which also serves as the feature-input vec-
delay between the percussion and dicrotic peaks in tor for corresponding training samples. K is the
the PPG signal. HR is related to BP, and Tpd is kernel function and b is a bias. Here, the popu-
correlated to transit time of blood-pressure waves, lar Gaussian kernel shown in (3) is used, where
which is also a strong indicator of BP value [6]. σ is the training parameter. N is the number of
To normalize the extracted features and sup- support vectors, which is typically also the total
press outliers, the feature sets are sorted offline number of training samples used in the LSSVM
to find the first and third quartiles Q1 and Q3 , algorithm.
and the second quartile Q2 is taken as the median. Training and testing samples of the LSSVM
Then the features are normalized in the DSP sys- model are taken from the University of Queensland
tem using (1), with lower and upper bounds (LB Vital Signs data set [7]. The ranges of SBP and
and UB) of the training set for a certain feature DBP labels are 70–150 mmHg and 40–110 mmHg,
f set to Q1 − 2(Q3 − Q1 ) and Q3 + 2(Q3 − Q1 ), respectively. 1962 samples were used for testing
 Zhang Q R, et al. Sci China Inf Sci April 2020 Vol. 63 149402:3

the model, which was implemented in MATLAB. 5.6 mmHg for SBP and DBP, respectively, which
The mean absolute errors (MAE) and standard de- satisfies the medical device standards of the As-
viations (SD) (in the form of MAE ± SD) of SBP sociation for the Advancement of Medical Instru-
and DBP estimation results are 7.41 ± 9.93 mmHg mentation (AAMI) [9].
and 6.10 ± 8.37 mmHg, respectively. Though the estimation accuracy needs to be im-
The proposed BP estimation method was im- proved further, the proposed system requires only
plemented using C programming language on the a single-channel PPG signal and two classifying
TMS320C5535 eZdspTM platform from Spectrum features. This simplicity reduces the hardware
Digital, Inc., which features a low-power fixed- complexity and cost of the system. The user expe-
point DSP chip (TMS320C5535 from Texas In- rience is also improved as users only need to wear
struments) designed for biomedical and wearable one PPG sensor on a finger. Furthermore, LSSVM
applications. To record the PPG signal, we used a is a non-parametric model, so frequent calibration
customized high-sensitivity sensor chip with a wide is not needed, which increases user convenience of
dynamic range [8]. The analog PPG signal from BP monitoring.
the sensor chip is sampled by a 10-bit ADC on the
DSP platform at 100 Hz. The 16-bit fixed-point Acknowledgements This work was supported by Na-
arithmetic routine in DSP is used to calculate the tional Key Research and Development Program of China
LSSVM decision function in (2). (Grant No. 2016YFC0105502).

A photo of the DSP-based cuffless BP monitor- References

ing system prototype is displayed in Figure 1(b). 1 Mozaffarian D, Benjamin E J, Go A S, et al. Execu-
For comparison, the commercial cuff-based BP me- tive summary: heart disease and stroke statistics–2016
ter from OMRON was used to record the ground- update: a report from the american heart association.
truth BP values. Figure 1(c) shows the GUI used Circulation, 2016, 133: 447–454
2 Hoshide S, Cheng H M, Huang Q, et al. Role of ambu-
with the proposed system. The BP and HR esti- latory blood pressure monitoring for the management
mation results are averaged for a 10 s window and of hypertension in Asian populations. J Clin Hyper-
are updated every 10 s. Also, the preprocessed tens, 2017, 19: 1240–1245
PPG signal and BP curves are continuously dis- 3 Sheikh S, Sinha A D, Agarwal R. Home blood pressure
monitoring: how good a predictor of long-term risk?
played. Curr Hypertens Rep, 2011, 13: 192–199
An experiment comprising 20 subjects (20– 4 Suykens J A K, Vandewalle J. Least squares support
60 years of age) was conducted to test the perfor- vector machine classifiers. Neural Process Lett, 1999,
9: 293–300
mance of the proposed system in realistic applica- 5 Sun P, Wu Q H, Weindling A M, et al. An improved
tions. All measurements were conducted at room morphological approach to background normalization
temperature. The subject’s finger was placed di- of ECG signals. IEEE Trans Biomed Eng, 2003, 50:
rectly above the PPG sensor. Subjects were seated 117–121
6 Cattivelli F S, Garudadri H. Noinvasive cuffless esti-
on a chair with back support, feet positioned flat mation of blood pressure from pulse arrival time and
on the floor and forearms supported at the height heart rate with adaptive calibration. In: Proceedings
of the heart. Three BP estimation results and cor- of the International Workshop on Wearable and Im-
responding ground-truth values were recorded for plantable Body Sensor Networks, Berkeley, 2009. 114–
119
every subject. Then the recorded values were av- 7 Liu D, Görges M, Jenkins S A. University of Queens-
eraged to give the final BP estimation and corre- land vital signs dataset: development of an accessible
sponding ground-truth result for each subject. Ab- repository of anesthesia patient monitoring data for
research. Anesthesia Analgesia, 2012, 114: 584–589
solute error of the BP estimation was calculated as
8 Atef M, Wang M, Wang G X. A fully integrated high-
the absolute difference between the averaged BP sensitivity wide dynamic range PPG sensor with an
estimation and the averaged ground-truth value. integrated photodiode and an automatic dimming con-
The MAEs between the 20 estimated BP results trol LED driver. IEEE Senss J, 2018, 18: 652–659
9 White W B, Berson A S, Robbins C, et al. Na-
and BP ground-truth values are 5.5 and 4.8 mmHg
tional standard for measurement of resting and am-
for SBP and DBP, respectively. The mean errors bulatory blood pressures with automated sphygmo-
and their SDs are 1.3 ± 6.1 mmHg and −0.1 ± manometers.. Hypertension, 1993, 21: 504–509