1

Cuff-less Arterial Blood Pressure Waveform

Synthesis from Single-site PPG using Transformer
& Frequency-domain Learning
Muhammad Ahmad Tahir∗ , Ahsan Mehmood∗ , Muhammad Mahboob Ur Rahman∗ , Muhammad Wasim Nawaz† ,
Kashif Riaz∗ , Qammer H. Abbasi‡
∗ Electrical engineering department, Information Technology University, Lahore 54000, Pakistan
† Electrical engineering department, University of Lahore, Lahore, Pakistan
‡ James Watt School of Engineering, University of Glasgow, Glasgow, G12 8QQ, UK
∗ {msds21039, mahboob.rahman}@itu.edu.pk, ‡ [email protected]
arXiv:2401.05452v1 [eess.SP] 9 Jan 2024

Abstract—We propose two novel purpose-built deep learning less frequent than hypertension, is also widespread as approx-
(DL) models for synthesis of the arterial blood pressure (ABP) imately (10-20)% of people of age over 65 years experience
waveform in a cuff-less manner, using a single-site photoplethys- hypotension. Hypotension occurs when blood pressure falls
mography (PPG) signal. We utilize the public UCI dataset on
cuff-less blood pressure (CLBP) estimation to train and evaluate below 90/60, which could lead to dizziness and nausea.
our DL models. Firstly, we implement a transformer model Hypotension becomes perilous when BP falls below 60/45,
that incorporates positional encoding, multi-head attention, layer causing blurriness, confusion, coma, or even death [4]. Thus,
normalization, and dropout techniques, and synthesizes the ABP regular and frequent monitoring of blood pressure by the el-
waveform with a mean absolute error (MAE) of 14. Secondly, we derly is of utmost importance for early diagnosis and efficient
implement a frequency-domain (FD) learning approach where we
first obtain the discrete cosine transform (DCT) coefficients of management of hypertension and hypotension [5].
the PPG and ABP signals corresponding to two cardiac cycles, The existing clinical-grade methods for blood pressure
and then learn a linear/non-linear (L/NL) regression between estimation (including ambulatory devices) could be classified
them. We learn that the FD L/NL regression model outperforms into two categories: i) invasive methods, ii) non-invasive but
the transformer model by achieving an MAE of 11.87 and 8.01, cuff-based methods. Invasive methods are the gold standard
for diastolic blood pressure (DBP) and systolic blood pressure
(SBP), respectively. Our FD L/NL regression model also fulfills methods which measure the centralized aortic blood pressure
the AAMI criterion of utilizing data from more than 85 subjects, by inserting catheters in the aorta (main artery near the heart)
and achieves grade B by the BHS criterion. [6]. Need not to say that this method is painful, prone to infec-
Index Terms—PPG, arterial blood pressure, systolic, diastolic, tions, and required trained professionals in a clinical setting.
transformer, ridge regression, discrete cosine transform. Next, non-invasive but cuff-based methods work by inflating
the cuff around the arm and rely upon manual auscultation and
Sphygmomanometers for peripheral blood pressure estimation.
I. I NTRODUCTION These methods, though highly accurate, suffer from a num-
ber of limitations, e.g., inconvenience of regular monitoring,
Hypertension (high blood pressure) and hypotension (low requirement of trained personnel, Sphygmomanometers use
blood pressure) are two common blood pressure (BP) irreg- toxic mercury, etc [7]–[9]. Further, the need for frequent cuff
ularities that are of great clinical significance. According to inflation and deflation makes continuous BP measurement
the World Health Organization (WHO), approximately 1.28 impractical, especially for individuals with weaker arms, or
billion people aged 30-79 years suffer from hypertension, with the elderly [10]. Oscillometric devices, though do not require
46% unaware of it and only 42% receiving treatment [1]. manual auscultation, have relatively lower accuracy as they
Further, hypertension is more prevalent in low and middle rely on mean arterial blood pressure (MABP) for estimating
income countries [2]. Hypertension (also known as the silent both SBP and DBP.
killer) is a major risk factor for a range of cardiovascular The limitations of the existing invasive methods and cuff-
diseases, stroke, chronic kidney disease and pre-mature death based non-invasive methods have thus prompted the re-
[3]. Hypertension, if not treated, could elevate both the systolic searchers to design novel cuff-less blood pressure (CLBP)
BP (SBP) and diastolic BP (DBP) from normal range (below estimation methods which are meant to be user-friendly,
120/below 80) to stage 1 hypertension (130-139/80-89), to reliable, and support continuous and at-home BP measure-
stage 2 hypertension (above 140/above 90), to hypertensive ment [7], [11]–[13]. Additionally, advancements in wearable
crisis (above 180/above 120), leading to health emergency technology (e.g., smart watches, smart bands, fitness trackers,
necessitating immediate hospitalization. Hypotension, though internet of medical things, flexible electronic tattoos, etc.)
and assistive gadgets (e.g., smartphones, airpods, etc.) have
This work is supported in part by UK Engineering and Physical Sci-
ences Research Council (EPSRC) under the grants: EP/T021063/1 and further accelerated the development of the CLBP estimation
EP/T021020/1. methods [14], [15]. CLBP estimation methods either rely upon
bio-signals such as photoplethysmography (PPG), electrocar- wave velocity (PWV). Before we delve into the discussion of
diography (ECG), phonocardiography (PCG) acquired non- the related work, it is imperative to formally define the three
invasively from one or more body sites [12], or innovative indices (BP surrogates) as follows:
sensor technology, e.g., tonometry [16]. • PTT is the time taken by a pulse pressure wave to travel
Contributions. Inline with the recent works on CLBP from one point of an artery to another. PTT is inversely
estimation, we propose two novel purpose-built deep learning proportional to the blood pressure. A typical measurement
(DL) models for synthesis of the arterial blood pressure (ABP) system for PTT includes two PPG sensors placed at two
waveform in a cuff-less manner, using a single-site PPG signal. different locations close to the artery [19], [20].
We utilize the public UCI dataset on CLBP estimation to train • PAT is the time delay between the R-peak in the ECG
and evaluate our DL models. The main contribution of this signal and the time of arrival of the corresponding (PPG)
work is two-fold: pulse at the peripheral site, during the same heart beat.
• We implement a transformer model that incorporates In other words, PAT is the time taken by the pulse
positional encoding, multi-head attention, layer normal- pressure wave to travel from the heart to the peripheral
ization, and dropout techniques, and synthesizes the ABP measurement site. PAT is related to PTT as follows: PAT
waveform with a mean absolute error (MAE) of 14. = PTT + pre-ejection period [12].
• We implement a frequency-domain (FD) learning ap- • PWV is the velocity of a pulse pressure wave that travels
proach where we first obtain the discrete cosine trans- a distance L from a proximal point to a distal point. PWV
form (DCT) coefficients of the PPG and ABP signals is related to PTT as follows: P W V = P L T T [12].
corresponding to two cardiac cycles, and then learn a Despite the good correlation of PTT, PAT and PWV with
linear/non-linear (L/NL) regression between them. the blood pressure, these indices have a number of limitations
The transformer model (the FD L/NL regression model) as well, e.g., need for simultaneous and synchronous measure-
achieves an MAE of 19 and 9 (11.87 and 8.01), for diastolic ment of two waveforms from two different sites, calibration is-
blood pressure (DBP) and systolic blood pressure (SBP), sues, lack of consensus on translation of PTT/PAT/PWV to BP
respectively. Our FD L/NL model achieves grade B by the (i.e., there exist a number of equations to map PTT/PAT/PWV
British Hypertension Society (BHS) criterion. to BP, each with many unknowns that are subject-specific),
Feasibility. PPG and ABP signals, though of different etc. [21], [12], [19]. This highlights the need for the design of
origins, have a very similar morphology (because both capture novel single-site methods for cuff-less blood pressure (CLBP)
the pressure exerted by the blood volume on arterial walls). estimation.
In other words, the morphology of the two signals is tightly Regulatory criteria for CLBP estimation methods: Over
binded to each other, at the sub-cardiac cycle resolution. time, two regulatory bodies have developed standards to grade
Furthermore, the recent advances in the generative artificial the performance of the emerging CLBP estimation methods:
intelligence (AI) has made it possible to translate one biosignal 1) association for the advancement of medical instrumentation
to another, e.g., [17] translates a PPG signal to an ECG signal, (AAMI), and 2) British hypertension society (BHS)1 .
[18] translates a PPG signal to an ABP signal. This motivates • The AAMI requires a dataset consisting of at least 85
us to propose purpose-built DL models for synthesis of the individuals having low, high, and normal blood pressure.
ABP waveform from a single-site PPG waveform. Further, according to AAMI, the mean error (and standard
Outline. Section II provides some necessary background deviation of the error) of a CLBP estimation method
and summarizes selected related work. Section III provides should be within 5 mmHg (and 8 mmHg) [23].
compact details of the public UCI dataset that we have used, • The BHS assigns one of the three letter grades (A, B, C)
and the key data pre-processing steps that we have performed. to each CLBP estimation method. If the model error is
Section IV describes the two deep learning models we have less than or equal to 5 mmHg 60% of time, 10 mmHg
implemented for cuff-less arterial blood pressure waveform 85% of time, and 15 mmHg 95% of time, then the model
synthesis. Section V presents some selected results. Section is given grade A. Similarly, if the model error is less than
VI concludes the paper. 5 mmHg 50% of time, 10 mmHg 75% of time, and 15
mmHg 90% of time, then the model is given grade B.
II. BACKGROUND & R ELATED W ORK Finally, if the model error is less than 5 mmHg 40% of
A. Background time, 10 mmHg 65% of time, and 15 mmHg 85% of time,
Pulse wave analysis methods: During the last decade, then the model is given grade C [24].
researchers have made extensive investigation of the physi-
ological waveforms acquired in a non-invasive manner (e.g., B. Related work
ECG, PPG, PCG, etc.), under the umbrella term of pulse wave
Since the literature on CLBP estimation is quite vast, we
analysis (PWA) [12]. The aim of PWA methods is to develop a
could only summarize selected related works due to space
number of indices to indirectly measure the blood pressure in
constraints (see [12], [19] and the references therein, which
a cuff-less manner using wearables (e.g., smart watches) and
provide a more detailed treatment of the subject). The CLBP
assistive gadgets (e.g., smartphones) [14]. Some prominent
indices (markers for blood pressure) include the following: 1 Additionally, IEEE has also developed 1708-2014 (and 1708a-2019)
pulse transit time (PTT), pulse arrival time (PAT), and pulse standard for wearable cuffless blood pressure measuring devices [22].

2
estimation methods are mostly PWA methods which are ei- pre-processed version of) the MIMIC-II waveform dataset
ther two-sites/two-waveforms methods or single-site methods. from Physionet databank, which is a large database of elec-
More recently, researchers have extensively used various deep tronic health records of the patients from the intensive care
learning (DL) models for ABP waveform synthesis and CLBP unit of the Beth Israel Deaconess Medical Center in Boston,
estimation. Massachusetts. This UCI dataset contains 12,000 records of
Single-site-based methods: [25] utilizes single-site PPG (942 patients) that consists of the physiological waveforms
data collected from 150 subjects, computes PWV, and feeds carrying information about the vital signs, with length of each
it to a learning-based non-parametric regression method in record varying between 8 sec and 10 min. Each record is
order to estimate MABP, SBP and DBP. [26] collects multi- sampled at 125 Hz and contains fingertip PPG, ECG (channel
wavelength, single-site PPG data from 10 subjects, computes II), and invasive ABP signal (in mmHg). Further, we note
the PTT using the IR-PPG and blue-PPG, in order to estimate that the dataset labels (i.e., DBP and SBP values) lie in the
the DBP and SBP. [27] recruits 22 young healthy subjects to following range: (60≤DBP≤130) mmHg and (80≤SBP≤180)
measure their ballistocardiogram (BCG) using the foot-based mmHg. In this work, we mainly focus on the PPG and ABP
pressure sensor/force plate, in order to estimate the DBP (SBP) waveforms for CLBP estimation.
using I-J interval (J-K amplitude) of a BCG. [18] utilizes the
UCI dataset and implements the well-known U-Net model to B. Data Preprocessing
synthesize an ABP waveform from a single PPG waveform.
Two-sites-based methods: [28] proposes a two-sites-based In order to prepare the data for our DL models, we im-
method for CLBP estimation that utilizes a PPG sensor at plemented the following data preprocessing steps on the PPG
the fingertip and an impedance plethysmography (IPG) sensor and ABP signals. 1) We began by removing out-of-band noise
at the wrist in order to compute the PTT which is then from the data using a Butterworth low-pass filter with a cut-
translated to SBP and DBP values. [29] recruits 12 young off frequency of 10 Hz. 2) We then proceeded to remove
healthy subjects, records their ECG and fingertip PPG data the residual artifacts manually (by discarding the anomalous
at multiple regular intervals after treadmill exercise, computes samples). 3) Next, we removed the baseline from the data
the PTT and translates it to the SBP and DBP values. [8] using the Wavelet filtering. 4) We then normalized the data
utilizes the ECG and PPG signals from MIMIC-II dataset, using z-score normalization. 5) Further, the process of ABP
pre-processes them, extracts various physiological features, waveform synthesis from PPG waveform required beat-level
does dimensionality reduction, and implements a number of synchronization between the PPG and ABP signals. To this
machine learning regression models to estimate SBP, MABP end, we utilized time shifting approach to align the systolic
and DBP. [30] adopts a rather different approach for CLBP es- peaks of the PPG signal with the systolic peaks of the ABP
timation whereby authors utilize data from the PPG, pressure, signal. 6) We then did segmentation (with a stride of 0) to
and height sensors (accelerometers), exploiting the fact that construct a number of segments from each PPG and ABP
movement of the arm changes the natural hydrostatic pressure. recording, in order to increase the size of the dataset. Note
Deep learning-based methods: [31] utilizes the PPG and that each segment consists of two cardiac cycles.
ABP signals corresponding to 20 subjects from the MIMIC
II dataset, and implements a convolutional neural network IV. M ETHODOLOGY
(CNN) for feature extraction, followed by an long short- Keeping in mind that the problem at hand—ABP waveform
term memory (LSTM) network for CLBP prediction. [32] synthesis from PPG waveform—is a sequence-to-sequence
considers 3 subjects from MIMIC II dataset, and utilizes a translation (i.e., regression) problem, we have implemented
CNN-LSTM model for ABP waveform synthesis. Paviglianiti two novel deep learning methods using Tensorflow and Keras
et al. [33] utilize both PPG and ECG signals form the MIMIC frameworks in Python: 1) a transformer model, 2) a linear/non-
dataset, and implement a number of DL models, i.e., ResNet, linear frequency-domain regression model. Once the ABP
LSTM, WavNet, and more, for CLBP prediction. Brophy et waveform is synthesized, we sample its maximum and min-
al. [34] utilize PPG signals from the UCI dataset, implement imum values which yield the systolic and diastolic blood
a number of generative adversarial networks (GAN) models pressure, respectively. Below, we discuss the two methods in
in order to realize a federated learning approach for ABP detail, one by one.
waveform synthesis. They evaluate the performance of their
GAN-based federated learning model on a new unseen dataset,
the University of Queensland vital signs dataset. Finally, [35] A. Transformer model for ABP waveform synthesis
utilizes a CNN, while [36] implements a feedforward neural The motivation for implementing a transformer model for
network for CLBP estimation. ABP waveform synthesis from PPG waveform comes from
the following: i) both the PPG and the ABP waveforms
III. T HE UCI DATASET & DATA P RE - PROCESSING represent the time-series data, ii) each of the two time-series
consists of an ordered set of data points that capture the
A. The UCI public dataset on CLBP estimation temporal dynamics of the heart, i.e., the heart physiology,
We utilize the cuff-less blood pressure estimation dataset at different time-scales and on its own, iii) there is one-to-
from the University of California Irvine (UCI) machine learn- one correspondence (relationship) between the (ordered) data
ing repository [37]. This dataset is basically a subset of (and points (at the same instant) in the PPG and ABP time-series,

3
 are implemented as follows:
• Frequency-domain representation of pre-processed data:
Firstly, we obtain the frequency-domain representation
of the pre-processed data, for both signals of interest,
i.e., PPG and ABP signals. The motivation for this
comes from the fact that both PPG and ABP signals
are quasi-periodic due to the rhythmic activity of the
heart. Specifically, we consider one PPG-ABP segment-
pair corresponding to a single cardiac cycle of duration
Q′ samples, at a time. We then compute the discrete
Fig. 1. Architecture of Transformer model for ABP waveform synthesis.
cosine transform (DCT) (type-II) of each of the PPG
and ABP segments in the single PPG-ABP segment-
due to the fact that they both capture the same phenomenon pair. Further, it is well-known that the PPG and ABP
(of heart physiology), though independently. Keeping in mind signals are sparse in frequency domain (with only a few
all these strong analogies with the natural language processing low-frequency components with significant energy). This
(NLP) problems (e.g., language translation, paraphrasing, etc.), allows us to have a compact representation of each of the
we solve the ABP waveform synthesis problem using the two segments in the given PPG-ABP segment-pair by re-
transformer model—a powerful tool for solving a range of taining first QX (QY ) significant DCT coefficients of the
NLP problems. PPG (ABP) segment only. This in turn helps in reducing
Our transformer model implementation consists of the the computationally complexity of our linear/non-linear
following discrete components: positional encoding, multi- (L/NL) regression model by reducing the dimension of
head self-attention, feedforward network, layer normalization, the feature vector it takes as input. Note that we zero-pad
dropout, and a dense output layer. Fig. 1 shows the architecture both DCT vectors of size QX and QY to resize them to
of the transformer implemented in this work in detail. Q in order to do the L/NL regression in the next step
At this point, it is imperative to revisit the individual (where Q represents the duration of a cardiac cycle, in
building blocks of the transformer model to see how each samples).
component helps solve the problem at hand. 1) First, the • Linear/non-linear regression: Secondly, we learn the

positional encoding (often implemented by means of sine and linear/non-linear regression between the size-Q DCT
cosine functions) keeps the transformer informed about the coefficients vector x (that represents the most impor-
order and position of data points (i.e., the notion of time) tant features of the PPG signal) and the size-Q DCT
in each of the two (i.e., PPG and ABP) sequences. 2) Next, coefficients vector y (that represents the most important
the multi-head self-attention mechanism, as the name implies, features of the ABP signal). For linear regression, we
focuses on capturing relationship between every pair of data have implemented ridge regression with α = 0.52 . For
points (by means of attention scores) in both the PPG and non-linear regression, we have implemented a shallow
ABP sequences, using multiple heads that attend to different feedforward neural network (FFNN) with two hidden
parts of the PPG and ABP sequences in order to capture the layers and a recurrent neural network (RNN) with two
short term and the long term dependencies within the PPG hidden layers.
and ABP signals. 3) Next, the layer Normalization reduces Next, the testing phase. We compute DCT (type-II) of a
the internal covariate shift, prevents the issue of vanishing previously unseen size-Q′ PPG segment, resize it to QX and
or exploding gradients, helps prevent overfitting, all by nor- then Q. We then feed the size-Q DCT vector (of PPG) to
malizing the activations. 4) The dropout regularization layer the L/NL frequency-domain regression model which outputs
randomly sets a fraction of the values to zero during training, a size-Q DCT vector (of ABP). Then, ABP waveform is
which helps to prevent overfitting. 5) Element-wise addition synthesized by taking the inverse DCT (I-DCT) of the output
between the original input and the output of the dropout layer of the L/NL regression model. Fig. 2 provides a pictorial
is performed in the next block. This residual connection allows overview of the proposed FD L/NL regression model.
information to bypass the self-attention and normalization
layers, helping the model retain valuable information from V. R ESULTS AND D ISCUSSION
the original input. 6) Next, the feedforward network applies a Performance Metrics: We have utilized the following two
set of linear transformations and non-linear activations to the metrics to evaluate the performance of our DL models for
input data, allowing the model to learn complex patterns and CLBP estimation: mean absolute error (MAE),
1
PN and root mean
relationships. 7) Finally, the dense layer produces the output squared error (RMSE), where MAE = N i=1 |ŷi − yi |, and
q
of the model. PN
i i (y −ŷ )2
RMSE(y, ŷ) = i=1
N . Here, N represents the batch
size, y represents the reference value/ground truth, and ŷ
B. Linear/non-linear frequency-domain regression model for represents the model prediction.
ABP waveform synthesis
2 Ridge regression is basically linear regression but with an additional term
In this method, we follow a two-step approach for ABP for L2 regularization that helps prevent overfitting and stabilize the regression
waveform synthesis. During the training phase, the two steps coefficients.

4
Fig. 2. Architecture of FD L/NL regression model for ABP waveform
synthesis. Fig. 3. Two examples of ABP waveform synthesis by our transformer model.

A. Performance evaluation of Transformer model and the morphology of the reference ABP waveforms (on the
left) are highly similar. Also, the synthetic ABP waveform is
Setting of hyperparameters: We used Adam optimizer, and
again shifted in amplitude, compared to the reference ABP
used a learning rate of 1e-3. Further, we set segment size to
waveform. But luckily, this time this shift in amplitude of the
2 seconds, i.e., 250 samples. For training and evaluation of
synthetic ABP is less pronounced. This helps us get a better
transformer model, we utilized an 80-20 train-test split. In
MAE performance when we measure the SBP and DBP.
multi-head attention layer, we set number of heads to 8, with
For our FD L/NL regression model, after converting the
the size of each attention head for query and key being 64.
PPG signal to frequency-domain by means of the DCT,
Next, the feedforward network consisted of 64 neurons and a
we have implemented ridge regression for linear regression
relu activation function. The number of transformer layers was
purpose, and a feedforward neural network (FFNN) and a
varied between 7-15. For layer normalization, we set epsilon
recurrent neural network (RNN) for non-linear regression
to 1e-6. Finally, we set the dropout ratio to 0.1.
purposes. The resulting MAE performance (for ABP waveform
Fig. 3 is a qualitative result that showcases two examples
synthesis) is summarized in Table I. We learn from Table I
where the transformer model synthesizes the ABP signal
that the ridge regression outperforms the FFNN and RNN by
from a PPG signal. We make the following observations: 1)
a large margin, which indicates that the relationship between
It is clear that the morphology of the two synthetic ABP
the PPG signal and the ABP signal is perhaps linear.
waveforms constructed by the transformer model (on the right)
Finally, our FD L/NL regression model (with ridge regres-
is quite similar to the morphology of the reference ABP
sion) achieves an MAE of 8.01 mmHg for DBP and an MAE
waveforms (on the left). 2) We note that the synthetic ABP
of 11.87 mmHg for SBP.
waveform is shifted in amplitude, compared to the reference
ABP waveform. This is the root cause of the error when we
TABLE I
measure the SBP and DBP later by sampling the maximum P ERFORMANCE OF DIFFERENT ARCHITECTURES FOR OUR FD L/NL
and minimum point of the synthetic ABP waveform. REGRESSION MODEL
Now, the quantitative results. The lowest MAE (for ABP
FD L/NL regression model MAE
waveform synthesis) achieved by the transformer model is
DCT + Ridge Regression 8.1
9.10, with 15 layers. That is, the transformer model’s per- DCT + RNN 17.23
formance increases with increase in number of layers, and DCT + FFNN 20.24
reaches a maximum for 15 layers (as further increasing the
depth does not result in significant improvements). Further,
the transformer model achieved an MAE of 19 and 9, for
C. Performance comparison with related work
DBP and SBP, respectively.
Table II provides a detailed performance comparison of our
work with the selected related works by providing a quick
B. Performance evaluation of FD L/NL Regression model summary of different deep learning models used for CLBP
Setting of hyperparameters: For ridge regression, we set estimation, their corresponding datasets, and the performance
α = 0.5. The rest of the hyperparameters remain the same. achieved by each model. Table II efficiently highlights the
Fig. 4 is again a qualitative result that demonstrates by variability in the performance of the related works, which is
means of two examples the quality of the ABP signal syn- in turn due to the change of data distribution across datasets,
thesized by the FD L/NL regression model. It is again evident variable sizes of the datasets, model architectures, etc. Our
that the morphology of the two synthetic ABP waveforms first approach, i.e., the transformer based model achieved an
produced by the FD L/NL regression model (on the right) MAE of 9.10 for ABP waveform synthesis which is at par with

5
 TABLE III
P ERFORMANCE COMPARISON OF OUR WORK WITH RELATED WORK
ACCORDING TO THE BHS CRITERION

Work Grade Cumulative Error

≤ 5% ≤ 10% ≤ 15%
BHS A 60% 85% 95%
B 50 75 90
C 40 65 85
N. Ibtehaz et. al [18] A 87.4 95.2 97.7%
Tazarv et. al [31] A 75 92 96%
Li Y-H et. al B 59.46 79.97 88.45%
Kauchee et. al [8] C 44.7 71.6 86.7%
Yang et. al C 42.1 94.7 100%
Our work (FD L/NL regression) B 51.69 76.79 100%
Our work (transformer) C 36.47 71.98 100%

grades obtained by the selected related works, for comparison

Fig. 4. Two examples of ABP waveform synthesis by our FD L/NL regression purposes. We also want to highlight that nearly all the works
model.
suffer from the issue of lack of generalization, i.e., as the
number of subjects increases, the MAE of the SBP and DBP
the related works. However, the transformer model achieved increases, while the BHS grade tends to fall. On a side note,
a relatively large MAE of 19, for DBP estimation, which it is also worth mentioning that both of our models models
suggests that our model is well-trained for ABP synthesis and fulfill the AAMI criterion of utilizing more than 85 subjects
SBP estimation only. Our 2nd approach, i.e., the FD L/NL for training purposes.
regression model achieved a MAE of 8.01 for DBP and 11.87
for SBP, which is again at par with the performance reported VI. C ONCLUSION
by the related works. We evaluated two novel purpose-built DL models on public
UCI dataset for CLBP estimation using a single-site PPG
TABLE II signal. The transformer model (the FD L/NL regression model)
P ERFORMANCE COMPARISON OF OUR WORK WITH THE RELATED WORK : achieved an MAE of 19 and 9 (11.87 and 8.01), for DBP
SBP MAE AND DBP MAE ARE IN MM H G . N REPRESENTS NUMBER OF and SBP, respectively. Our FD L/NL regression model also
SUBJECTS . * REPRESENTS RMSE IN MM H G .
fulfilled the AAMI criterion of utilizing data from more than
Work Dataset N Algorithm SBP DBP 85 subjects, and achieved grade B by the BHS criterion.
Tazarv et al. [31] MIMIC 20 CNN-LSTM 3.70 2.02 Now, some parting thoughts. The UCI dataset consists of
II
data from patients in intensive care units, who are often under
Paviglianiti et al. MIMIC 40 ResNet, 4.118 2.228
[33] LSTM, the effects of drugs and other medical interventions. Thus,
WavNet, it is difficult to generalize the results of studies–that use the
ResNet+LS- UCI dataset–to other populations, such as healthy adults. Thus,
TM
Treebup-achatsa- UCI 812 CAN 7.14* 6.08* there is a need to evaluate our DL models on larger and diverse
kul et al. datasets in order to increase their generalization capability.
Slapnicar et al. MIMIC 510 Spectrotem- 9.43 6.88
III poral ResNet
Wang et al. UCI 348 Pretrained 6.17 3.16 R EFERENCES
AlexNet, [1] World Health Organization, “First who report details devastating impact
Inception- of hypertension and ways to stop it.” https://www.who.int/news/item/
V3, VGG-19 19-09-2023-first-who-report-details-devastating-impact-of-hypertension-and-ways-to-s
Aguirre et al. MIMIC 1131 RNN 14.39 14.39 2023. Accessed on: Jan. 2, 2024.
Encoder- [2] P. Nugroho, H. Andrew, K. Kohar, C. A. Noor, and A. L. Sutranto,
Decoder + “Comparison between the world health organization (who) and interna-
attention tional society of hypertension (ish) guidelines for hypertension,” Annals
Our work (trans- UCI 209 Transformer 19.0 9.0 of Medicine, vol. 54, no. 1, pp. 837–845, 2022.
former) [3] B. Zhou, P. Perel, G. A. Mensah, and M. Ezzati, “Global epidemiology,
Our work (FD UCI 209 DCT+L/NL 11.87 8.01 health burden and effective interventions for elevated blood pressure and
L/NL regression) Regression hypertension,” Nature Reviews Cardiology, vol. 18, no. 11, pp. 785–802,
2021.
[4] N. I. Saedon, M. Pin Tan, and J. Frith, “The prevalence of orthostatic
Finally, Table III outlines the performance grades of our hypotension: a systematic review and meta-analysis,” The Journals of
transformer model and FD L/NL regression model, according Gerontology: Series A, vol. 75, no. 1, pp. 117–122, 2020.
to the BHS grading criterion. Specifically, Table III shows [5] K. L. Tucker, J. P. Sheppard, R. Stevens, H. B. Bosworth, A. Bove,
E. P. Bray, K. Earle, J. George, M. Godwin, B. B. Green, et al., “Self-
the cumulative percentages (CP) of the error achieved by monitoring of blood pressure in hypertension: a systematic review and
the two models on the UCI test data. We can see that individual patient data meta-analysis,” PLoS medicine, vol. 14, no. 9,
our transformer model achieves the grade C, while our FD p. e1002389, 2017.
[6] B. Saugel, K. Kouz, A. S. Meidert, L. Schulte-Uentrop, and S. Ro-
L/NL regression model achieves the grade B. For the sake of magnoli, “How to measure blood pressure using an arterial catheter: a
completeness, Table III also summarizes the BHS performance systematic 5-step approach,” Critical Care, vol. 24, pp. 1–10, 2020.

6
 [7] J. A. Pandit, E. Lores, and D. Batlle, “Cuffless blood pressure monitor- ing: Proof of concept and potential challenges,” IEEE Transactions on
ing: promises and challenges,” Clinical Journal of the American Society Biomedical Engineering, vol. 65, no. 11, pp. 2384–2391, 2018.
of Nephrology: CJASN, vol. 15, no. 10, p. 1531, 2020. [28] T. H. Huynh, R. Jafari, and W.-Y. Chung, “Noninvasive cuffless blood
[8] M. Kachuee, M. M. Kiani, H. Mohammadzade, and M. Shabany, “Cuf- pressure estimation using pulse transit time and impedance plethysmog-
fless blood pressure estimation algorithms for continuous health-care raphy,” IEEE Transactions on Biomedical Engineering, vol. 66, no. 4,
monitoring,” IEEE Transactions on Biomedical Engineering, vol. 64, pp. 967–976, 2019.
no. 4, pp. 859–869, 2016. [29] X. F. Teng and Y. T. Zhang, “An evaluation of a ptt-based method
[9] P. M. Nabeel, J. Jayaraj, K. Srinivasa, S. Mohanasankar, and M. Chen- for noninvasive and cuffless estimation of arterial blood pressure,” in
niappan, “Bi-modal arterial compliance probe for calibration-free cuf- 2006 International Conference of the IEEE Engineering in Medicine
fless blood pressure estimation,” IEEE Transactions on Biomedical and Biology Society, pp. 6049–6052, 2006.
Engineering, vol. 65, no. 11, pp. 2392–2404, 2018. [30] P. A. Shaltis, A. T. Reisner, and H. H. Asada, “Cuffless blood pressure
[10] P. Shaltis, A. Reisner, and H. Asada, “A hydrostatic pressure approach monitoring using hydrostatic pressure changes,” IEEE Transactions on
to cuffless blood pressure monitoring,” in The 26th Annual International Biomedical Engineering, vol. 55, no. 6, pp. 1775–1777, 2008.
Conference of the IEEE Engineering in Medicine and Biology Society, [31] A. Tazarv and M. Levorato, “A deep learning approach to predict
vol. 1, pp. 2173–2176, 2004. blood pressure from ppg signals,” in 2021 43rd Annual International
[11] J. A. Pandit, E. Lores, and D. Batlle, “Cuffless blood pressure monitor- Conference of the IEEE Engineering in Medicine & Biology Society
ing,” Clinical Journal of the American Society of Nephrology, vol. 15, (EMBC), pp. 5658–5662, 2021.
no. 10, p. 1531–1538, 2020. [32] H. Mou and J. Yu, “Cnn-lstm prediction method for blood pressure based
[12] J. Solà and R. Delgado-Gonzalo, “The handbook of cuffless blood on pulse wave,” Electronics, vol. 10, no. 14, p. 1664, 2021.
pressure monitoring,” Cham: Springer, 2019. [33] A. Paviglianiti, V. Randazzo, E. Pasero, and A. Vallan, “Noninvasive
[13] L.-Y. Shyu, Y.-L. Kao, W.-Y. Tsai, and W. Hu, “Development of a cuf- arterial blood pressure estimation using abpnet and vital-ecg,” in 2020
fless blood pressure measurement system,” in 2012 Annual International IEEE International Instrumentation and Measurement Technology Con-
Conference of the IEEE Engineering in Medicine and Biology Society, ference (I2MTC), pp. 1–5, IEEE, 2020.
pp. 2040–2043, 2012. [34] E. Brophy, M. De Vos, G. Boylan, and T. Ward, “Estimation of
[14] H. Y. Lee, D.-J. Lee, J. Seo, S.-H. Ihm, K.-i. Kim, E. J. Cho, H. C. Kim, continuous blood pressure from ppg via a federated learning approach,”
J. Shin, S. Park, I.-S. Sohn, and et al., “Smartphone / smartwatch-based Sensors, vol. 21, no. 18, p. 6311, 2021.
cuffless blood pressure measurement : A position paper from the korean [35] S. Shimazaki, H. Kawanaka, H. Ishikawa, K. Inoue, and K. Oguri, “Cuf-
society of hypertension,” Clinical Hypertension, vol. 27, no. 1, 2021. fless blood pressure estimation from only the waveform of photoplethys-
[15] J. Baek, J. Kim, N. Kim, D. Lee, and S.-M. Park, “Validation of cuffless mography using cnn,” in 2019 41st Annual International Conference
blood pressure monitoring using wearable device,” in TENCON 2018 - of the IEEE Engineering in Medicine and Biology Society (EMBC),
2018 IEEE Region 10 Conference, pp. 0416–0419, 2018. pp. 5042–5045, 2019.
[36] U. Senturk, I. Yucedag, and K. Polat, “Cuff-less continuous blood
[16] M. Park, H. Kang, Y. Huh, and K.-C. Kim, “Cuffless and noninvasive
pressure estimation from electrocardiogram(ecg) and photoplethysmog-
measurement of systolic blood pressure, diastolic blood pressure, mean
raphy (ppg) signals with artificial neural network,” in 2018 26th Signal
arterial pressure and pulse pressure using radial artery tonometry pres-
Processing and Communications Applications Conference (SIU), pp. 1–
sure sensor with concept of korean traditional medicine,” in 2007 29th
4, 2018.
Annual International Conference of the IEEE Engineering in Medicine
[37] M. Kachuee, M. Kiani, H. Mohammadzade, and M. Shabany, “Cuff-
and Biology Society, pp. 3597–3600, 2007.
Less Blood Pressure Estimation.” UCI Machine Learning Repository,
[17] A. Mehmood, A. Sarouji, M. M. U. Rahman, and T. Y. Al-Naffouri,
2015. DOI: https://doi.org/10.24432/C5B602.
“Your smartphone could act as a pulse-oximeter and as a single-lead
ecg,” Scientific Reports, vol. 13, no. 1, p. 19277, 2023.
[18] N. Ibtehaz, S. Mahmud, M. E. Chowdhury, A. Khandakar,
M. Salman Khan, M. A. Ayari, A. M. Tahir, and M. S. Rahman,
“Ppg2abp: Translating photoplethysmogram (ppg) signals to arterial
blood pressure (abp) waveforms,” Bioengineering, vol. 9, no. 11, p. 692,
2022.
[19] R. Mukkamala, J.-O. Hahn, O. T. Inan, L. K. Mestha, C.-S. Kim,
H. Töreyin, and S. Kyal, “Toward ubiquitous blood pressure monitoring
via pulse transit time: Theory and practice,” IEEE Transactions on
Biomedical Engineering, vol. 62, no. 8, pp. 1879–1901, 2015.
[20] R. Shriram, A. Wakankar, N. Daimiwal, and D. Ramdasi, “Continuous
cuffless blood pressure monitoring based on ptt,” in 2010 International
Conference on Bioinformatics and Biomedical Technology, pp. 51–55,
IEEE, 2010.
[21] K. Shimura, M. Hori, T. Dohi, and H. Takao, “The calibration method
for blood pressure pulse wave measurement based on arterial tonometry
method,” in 2018 40th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society (EMBC), pp. 1–4, 2018.
[22] “Ieee standard for wearable cuffless blood pressure measuring devices,”
IEEE P1708/D04, June 2014, pp. 1–32, 2014.
[23] Electronic or Automated Sphygmomanometers. ANSI/AAMI Standard
SP10-192, ed. Arlington, VA, USA, 1993.
[24] E. O’Brien, J. Petrie, W. Littler, M. De Swiet, P. L. Padfield, D. Altman,
M. Bland, A. Coats, N. Atkins, et al., “The british hypertension society
protocol for the evaluation of blood pressure measuring devices,” J
hypertens, vol. 11, no. Suppl 2, pp. S43–S62, 1993.
[25] A. Chakraborty, D. Goswami, J. Mukhopadhyay, and S. Chakrabarti,
“Measurement of arterial blood pressure through single-site acquisition
of photoplethysmograph signal,” IEEE Transactions on Instrumentation
and Measurement, vol. 70, pp. 1–10, 2021.
[26] J. Liu, Y.-T. Zhang, X.-R. Ding, W.-X. Dai, and N. Zhao, “A preliminary
study on multi-wavelength ppg based pulse transit time detection for
cuffless blood pressure measurement,” in 2016 38th Annual International
Conference of the IEEE Engineering in Medicine and Biology Society
(EMBC), pp. 615–618, 2016.
[27] C.-S. Kim, A. M. Carek, O. T. Inan, R. Mukkamala, and J.-O. Hahn,
“Ballistocardiogram-based approach to cuffless blood pressure monitor-