A Novel Cuff-Less and Calibration-Free Blood Pressure Estimation

Framework Using Single Photoplethysmogram

Yusuf Ziya Hayirlioglu and Beren Semiz

Department of Electrical and Electronics Engineering, Koc¸ University, Istanbul, Turkey

Keywords:

Blood Pressure, Photoplethysmogram, Deep Learning, Continuous Monitoring.

Abstract:

Blood pressure (BP) is one of the four main vital signs and is a key indicator of cardiovascular health. However,

monitoring of BP is not regularly done in most of the population until health problems arise. Continuous

and convenient monitoring of blood pressure is thus needed to address this issue. We propose a novel BP

estimation algorithm without calibration to estimate BP from a cuff-less photoplethysmogram (PPG) system.

Data from a total of 219 subjects, which underwent only simple preprocessing steps, was used to train and

evaluate a hybrid Convolutional Long Short-Term Memory Neural Network (CNN-LSTM) model. The model

was trained using the preprocessed PPG signal as the only input. The model had two neurons in the last

layer to output systolic blood pressure (SBP) and diastolic blood pressure (DBP) values. The model was

optimized by conducting a random search on its hyperparameters for better performance. The model resulted

in a comparable performance to those in the literature, with mean absolute errors (MAEs) of 14.13 mmHg

and 8.80 mmHg for SBP and DBP, respectively. To assess generalizability, we also tested the trained model

on a second dataset collected from 20 subjects using a custom wearable system, which was again resulted

in MAEs of 10.71 mmHg and 10.09 mmHg, respectively. Overall, our results show that such a pipeline

could potentially be leveraged in the design of wearable systems to achieve cuff-less and calibration-free BP

monitoring in ambulatory settings.

1 INTRODUCTION

Blood pressure (BP) is one of the four main vital

signs along with pulse rate, respiration rate, and tem-

perature, serving as a critical parameter in assess-

ing cardiovascular health (Sapra et al., 2020). Accu-

rate evaluation of BP levels contributes substantially

to the early detection and management of hyperten-

sion. Hypertension is a major cause of premature

death around the world and is a major risk factor for

conditions and diseases such as stroke, heart failure,

and kidney disease. Only 42% of adults with hyper-

tension are properly diagnosed and treated (Lackland

and Weber, 2015). Continuous blood pressure mon-

itoring systems hold the potential to facilitate timely

disease detection and intervention, facilitate the for-

mulation of individualized therapeutic regimens, and

afford proactive and preemptive healthcare strategies

for individuals susceptible to ailments resulting from

abnormal blood pressure (Sana et al., 2020).

BP monitoring methods can be classiﬁed into two

groups: non-invasive and invasive. The gold standard

https://orcid.org/0000-0002-7544-5974

for non-invasive blood pressure monitoring involves

using the auscultatory technique (Korotkoff sounds)

through oscillometric sphygmomanometers and cuffs

(MHRA, 2019). This method provides accurate but

intermittent readings of BP (Mukherjee et al., 2018).

Invasive methods require the cannulation of an artery

with a stiff catheter to insert a transducer for BP

measurement. This method provides continuous and

signiﬁcantly accurate BP readings; however, it is

not preferred unless necessary due to potential harm

and high levels of discomfort to the patient. Other

BP measurement methods include ultrasound sens-

ing, volume clamping, tactile sensing, pulse tran-

sit time (PTT)-based measurement, and photophle-

tysmography (PPG)-based measurement (Mukherjee

et al., 2018).

Among all the methods, PPG-based BP measure-

ment is the most promising method for daily, contin-

uous BP monitoring due to its low cost, accessibility,

unobtrusive nature, and easy integration into wearable

devices such as smartwatches. The PPG signal is a re-

sult of the variance in the amount of light absorbed in

the arteries due to the changes in arterial blood vol-

Hayirlioglu, Y. Z. and Semiz, B.

A Novel Cuff-Less and Calibration-Free Blood Pressure Estimation Framework Using Single Photoplethysmogram.

DOI: 10.5220/0013055600003911

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 18th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2025) - Volume 1, pages 697-703

ISBN: 978-989-758-731-3; ISSN: 2184-4305

697

ume during the cardiac cycle. Its working principle is

based on optical detection, consisting of a light source

and a photosensor. Depending on the working mode

of the photosensor, a PPG signal is generated by mea-

suring either the reﬂected or transmitted light from the

relevant region. In the literature, PPG signals have

been analyzed to provide crucial information regard-

ing vascular resistance, blood oxygen level, and blood

pressure (cheol Jeong et al., 2018).

Studies involving PPG developed different tech-

niques to analyze the signal to estimate blood pres-

sure. Pulse arrival time (PAT)- and PTT-based blood

pressure calculations are among the most common

methods to estimate the blood pressure from PPG

signals since there is an established relationship be-

tween the two (Kim et al., 2015). Other commonly

used techniques involve extracting several frequency

and time domain features and using machine learn-

ing algorithms to estimate the blood pressure (Maq-

sood et al., 2021). Most work in the domain also re-

quires personal calibration over short time intervals,

which would necessitate access to information about

the subject beforehand, which might not always be

available (Elgendi et al., 2019). For PAT- and PTT-

based estimation methods, the requirement to use two

synchronized sensors to accurately measure these fea-

tures is their biggest disadvantage. For methods that

involve manual feature extraction, a careful and time-

consuming analysis of the signals is required. Ad-

ditionally, it is difﬁcult to standardize the steps in

feature extraction procedures. On the other hand,

deep learning-based methods can have feature ex-

traction capabilities embedded in their architecture.

They are supposed to generalize over the dataset they

are trained on, but training them requires a substan-

tial amount of data. The availability of numerous

databases with PPG and BP measurements trivializes

this problem.

Considering the points above, we propose a deep

learning-based regression model for BP estimation

using a single PPG measurement. Although methods

to estimate or measure BP from certain physiological

markers exist, a vast majority of them require mul-

tiple sensor modalities, bulky and inconvenient mea-

surement devices, or exhaustive signal analysis. This

work aims to develop a novel, robust, and convenient

estimation of BP from a single PPG signal without the

use of a cuff or any kind of personal calibration proce-

dure. Our main contributions are that we demonstrate

a novel approach to BP estimation, utilizing a Convo-

lutional Long Short-Term Memory Neural Network

(CNN-LSTM) hybrid architecture to estimate BP di-

rectly from ﬁltered PPG signals with no personal cal-

ibration procedure. The proposed model is also sufﬁ-

ciently lightweight and can easily be trained on com-

mercial PCs with a GPU.

This paper is organized as follows: The dataset,

the general structure of the preprocessing steps, and

selection and evaluation methods of our proposed al-

gorithm are described in Section 2. Section 3 presents

a comparative evaluation of our algorithm’s perfor-

mance. Concluding remarks have been presented in

Section 4.

2 METHODS

2.1 Dataset

This study primarily uses the PPG-BP dataset pub-

lished by Liang et al. (Liang et al., 2022) for training

and validation. The dataset consists of 657 recordings

from 219 subjects. Each subject ﬁrst had their arte-

rial blood pressure measured using the Omron HEM-

7201 (Omron Company, Kyoto, Japan) followed by

three, 2.1-second-long PPG recordings in the span of

three minutes. The PPG recording quality was eval-

uated by the authors of the dataset using a skewness

signal quality index.

2.2 Preprocessing

The raw PPG values from the dataset had high fre-

quency noise contaminating the signal. Low fre-

quency baseline wander was also present in the

recordings. Therefore, we conducted a ﬁltering op-

eration before using the dataset to train our BP esti-

mation algorithm. A 4th order Butterworth ﬁlter was

used with cut-offs at 0.4 Hz and 20 Hz. The resulting

clean signal is shown in Figure 1. In the literature, it

has been shown that the PPG signal can be adequately

analyzed within these frequency ranges (Reali et al.,

2022).

Also, deep learning algorithms greatly beneﬁt

from normalization. We normalized the data to limit

the scale and reduce the variance according to the

equation below, where µ is the mean and σ is the stan-

dard deviation (Equation 1).

normalized

x − µ

(1)

2.3 Machine Learning Algorithms

1. CNNs and their application over 1D data is not

an unexplored topic. However, due to the pop-

ularity and success of 2D CNNs, conventionally,

1D data is transformed into 2D graphs (such as

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

698

Figure 1: PPG signal after preprocessing.

.…

LSTM

.…

LSTM

1D CNN

Max

Pooling

1D CNN

Max

Pooling

Flatten

Input Convolutional Layers

LSTM Layers

Dense

Layer

Output

Figure 2: General Structure of the Model.

spectrograms) for analysis. Recently, the use of

1D CNNs for 1D data has become more com-

mon in the literature. This is in part due to 1D

CNNs requiring less computational complexity

and memory, making them a suitable candidate to

use for lightweight, real-time applications with-

out requiring specialized hardware. They are also

easier to train than 2D CNNs, often requiring a

smaller dataset. They have been used in time-

series data analysis such as speech recognition,

ECG monitoring, and stock value forecasting (Ki-

ranyaz et al., 2021).

2. LSTMs are proposed as a solution to the explod-

ing/vanishing gradient problem of Recurrent Neu-

ral Networks (RNNs). Its ability to remember

longer dependencies is a result of its recurrently

connected memory blocks that regulate the ﬂow

of information via non-linear gating units. For this

reason, it’s a popular choice for handling time se-

ries data (Van Houdt et al., 2020).

3. CNN-LSTM hybrid models take advantage of

both architectures. CNNs provide LSTMs with

extracted features with a reduced dimensional-

ity that adequately represents the input, while

LSTMs capture temporal dependencies over long

sequences. The hybrid architecture consistently

outperformed conventional unmixed CNN and

LSTM architectures (Van Houdt et al., 2020). In

the literature, the hybrid model has been used for

classiﬁcation of plant growth status (Xing et al.,

2023), forecasting of photovoltaic power produc-

tion, emotion identiﬁcation, etc. (Van Houdt et al.,

2020; Agga et al., 2022). Bao et al. have success-

fully demonstrated that a hybrid model can suc-

cessfully estimate wrist angles from electromyo-

gram (EMG) signals (Bao et al., 2020). They

compared their hybrid model with support vec-

tor regression (SVR), Random Forest, CNNs, and

LSTMs. Their hybrid model outperformed all

other models in all trials and protocols.

We propose a CNN-LSTM hybrid network to

solve the BP estimation problem. LSTMs and CNNs

are highly popular deep learning techniques and have

a variety of use cases ranging from natural language

processing to image processing. LSTMs’ ability

to capture dependencies in long temporal sequences

makes them a good match for time series estimation

applications. CNNs can extract learned features from

the raw data. In theory, a hybrid model consisting of

both should be able to extract the local features of a

signal and capture the long-term dependencies.

2.4 Hyperparameter Selection

The general structure of the model is presented in Fig-

ure 2. To tune the hyperparameters for our model, we

conducted a random search on the hyperparameters in

Table 1, training 100 different models over 30 epochs

A Novel Cuff-Less and Calibration-Free Blood Pressure Estimation Framework Using Single Photoplethysmogram

699

Table 1: List of selectable hyperparameters.

Hyperparameters Hyperparameter Values

1D CNN Layers 1, 3

1D CNN Filters 16, 32, 64

1D CNN Kernel Sizes 5, 15, 63

Lstm Layers 1, 3, 5

Lstm Units 32, 64, 128

Dense Layers 0, 1, 3

Dense Units 16, 64, 128

Table 2: Hyperparameter selection results.

Hyperparameters Hyperparameter Values

1D CNN Layers 3

1D CNN Filters

Layer 1: 16

Layer 2: 32

Layer 3: 64

1D CNN Kernel Sizes

Layer 1: 63

Layer 2: 15

Layer 3: 15

Lstm Layers 5

Lstm Units

Layer 1: 128

Layer 2: 128

Layer 3: 32

Layer 4: 64

Layer 5: 128

Dense Layers 0

Dense Units –

on the dataset.

The resulting architecture with the smallest val-

idation root mean square error was selected as the

candidate to be further trained on the dataset. The

random search yielded the parameters in Table 2 for

the model.

We have two neurons as the last layer of our ar-

chitecture to output systolic blood pressure (SBP) and

diastolic blood pressure (DBP) values.

2.5 Model Evaluation

For a more reliable estimation of our model’s perfor-

mance, we used 10-fold cross-validation. The met-

rics used to evaluate the algorithm’s performance are

presented below, where y is the ground truth, x is the

prediction, and n is the number of samples.

1. Mean Absolute Error (MAE): The mean of abso-

lute errors (Equation 2).

MAE =

∑

i=1

− x

| (2)

2. Root Mean Square Error (RMSE): The standard

deviation of prediction errors. It measures how

spread the residuals are (Equation 3).

RMSE =

∑

i=1

− x

)

(3)

3. Mean Absolute Error Percentage (MAPE): The

percentage mean of absolute errors (Equation 4).

MAPE =

∑

i=1

− x

| (4)

We used MSE (square of RMSE) to monitor the

prediction success of the algorithm during training.

2.6 Experiments to Assess

Generalizability

To assess generalizability, we also tested the trained

model on the dataset we collected from 20 subjects

using a custom wearable system. Data collection was

conducted upon approval by the Koc University Insti-

tutional Review Board, and all participants provided

written consent. Subject demographics were as fol-

lows: 8 females and 12 males, Age: 23.8 ± 4.2,

Height: 172.6 ± 9.5 cm, Weight: 70.9 ± 16.1 kg.

Before the experimental protocol, the subjects’

SBP and DBP were measured using Omron M2 de-

vice. Following that, the subjects were asked to stand

still for two minutes while their PPG signals were be-

ing collected from their wrist area using a MAX30102

sensor at 50 Hz. The IR signal was pre-processed in

a similar manner (detailed in Section 2.2) and used as

the test data for the pretrained CNN-LSTM model.

3 RESULTS AND DISCUSSION

In this work, we focused on BP estimation from raw

PPG signals without using handcrafted features. Only

the ﬁltered PPG recording was fed to the CNN-LSTM

hybrid model. Most works in the literature propose a

calibration method to improve their results. Slapniv-

car et al. used 20% of the test subjects to train their

network for personalization (Slapni

car et al., 2019).

Their MAE improved as much as 5.98 mmHg and 5.5

mmHg, respectively, for their best-performing algo-

rithm. In another work, Xing et al. compared their

calibration-free algorithm with the same algorithm

with a calibration factor (Xing et al., 2019). Their

calibration factor was the median of a person’s previ-

ous ﬁtting errors. Their calibration factor improved

their estimation results by 2.0±4.1 mmHg for SBP,

2.2±0.5 mmHg for DPB in the young (≤50 yo), and

5.5±4.3 mmHg for SBP, 2.4±2.1 mmHg for DPB in

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

700

Figure 3: SBP Trend.

Figure 4: DBP Trend.

the old (>50 yo) population. They argue that the cali-

bration process reduced biases. However, we ﬁnd this

a bit counterintuitive since the population size for the

older population was higher than the younger popu-

lation by 81%. In fact, the estimations should have

been more biased towards the older population when

the calibration-free algorithm was used; however, this

was not the case. Depending on how the calibration

is made, calibrations might introduce even more bi-

ases to compensate for the errors. Considering the

previous point and how personal calibration requires

subject-speciﬁc information that might not always be

available, we decided to go with a calibration-free

method in designing our algorithm.

3.1 Model Evaluation Results

We selected two papers using the same dataset as

ours; therefore, the evaluation criteria that would nor-

mally be dependent on data distribution to draw con-

clusions, such as MAE, makes sense. Although most

BP datasets have a similar range, depending on the

subject demographics, the difference in the distribu-

tion of BP can be signiﬁcant enough to skew con-

clusions drawn from different datasets. Our moti-

vation was to prevent such an occasion from occur-

ring. Table 3 lists some of the best algorithms be-

longing to the two papers that used the same dataset

as ours. Among them, only ours and (Gonz

alez et al.,

2023)’s ResNet approach don’t use any feature extrac-

Figure 5: SBP Error Histogram.

Figure 6: DBP Error Histogram.

tion prior to training. Our approach and (Gonz

alez

et al., 2023)’s ResNet approach show very similar per-

formances, with ours having better mean errors but

slightly worse standard deviations. The rest of the

machine learning approaches that utilize some kind of

feature extraction method prior to training have better

results in general. This might be due to deep learning-

based approaches requiring a larger dataset to better

understand the relationship between the input and out-

put.

Figures 3 - 4 show trends in four folds of the test

data. The ﬁgures show that our algorithm can prop-

erly encapsulate the upward and downward move-

ments in the blood pressure. The data was smoothed

with a moving average ﬁlter (n = 30) to make the trend

clearer in the visualization.

Figures 5 - 6 visualize the histograms of resid-

ual errors. The error histograms indicate that the er-

rors are normally distributed around 0. The error his-

togram of SBP has a higher range. This is because

SBP has a higher variance as a result of its larger mag-

nitude in comparison to DBP.

3.2 Experiments to Assess

Generalizability

As detailed in Section 2.6, we also tested the trained

model on the IR PPG signals we collected from 20

subjects. For the SBP measurements, the MAE and

MAPE were calculated to be 10.71 mmHg and 9.45%,

A Novel Cuff-Less and Calibration-Free Blood Pressure Estimation Framework Using Single Photoplethysmogram

701

Table 3: Comparison of model performance.

Model Performance Criteria Systolic BP Diastolic BP

RelieF GPR (w/o opt.)

MAE (mmHg)

RMSE (mmHg)

ME+STD (mmHg)

MAPE (%)

10.08

14.80

7.87

9.83

CFS GPR (w/o opt.)

MAE

RMSE

ME+STD

MAPE

11.91

16.05

7.64

9.16

RelieF GPR (w opt.)

MAE

RMSE

ME+STD

MAPE

3.02

6.74

1.74

3.59

LightBGM

MAE

RMSE

ME+STD

MAPE

13.06

0.00±16.95

8.16

-0.04±10.30

AdaBoost

MAE

RMSE

ME+STD

MAPE

13.22

-0.56±16.95

8.04

-0.16±10.25

ResNet

MAE

RMSE

ME+STD

MAPE

13.62

-1.85±17.45

8.61

-2.17±10.81

Ours

MAE

RMSE

ME+STD

MAPE

14.13

18.13

0.71±18.23

11.24%

8.80

10.96

-0.79±11.08

12.54%

respectively. For the DBP measurements, the MAE

and MAPE were 10.09 mmHg and 12.02%, respec-

tively. Obtaining high performance from a different

test dataset justiﬁed that our pipeline was indeed gen-

eralizable regardless of the instrumentation used.

4 CONCLUSION

In this study, we proposed and implemented a

calibration-free BP estimation algorithm and demon-

strated that it is possible to estimate SBP and DBP

from a single PPG without manual feature extrac-

tion. The preprocessing steps to prepare the data for

training were discussed. The general structure of the

proposed CNN-LSTM model and its hyperparame-

ters were described in detail. The performance of

the model was evaluated with the metrics indicated

in Section 2.5 and compared to similar works in the

literature.

Different from the other works, our algorithm

does not rely on multiple sensor modalities, nor does

it rely on personal calibration to estimate the BP. Our

CNN-LSTM hybrid model is also a novel approach

to BP estimation. An advantage of our model is that

it is sufﬁciently lightweight to run on commercial

old GPUs such as GTX1070. It completes a 10-fold

cross-validation under 10 minutes on the full dataset,

and only takes seconds to predict BP given the data.

Single PPG, cuff-less BP monitoring is cheap, ac-

cessible, and user-friendly. In the near future, it might

be possible to integrate algorithms of this kind into

smartwatches or other wearables. Similar to how a

ﬁngertip pulse oximeter is able to measure blood oxy-

gen levels, devices of the same kind might be able

to provide BP readings. This way, BP-related risk

factors could be detected and serve to prevent or di-

agnose certain conditions and diseases in the general

populace, increasing quality of life.

BIOSIGNALS 2025 - 18th International Conference on Bio-inspired Systems and Signal Processing

702

REFERENCES

Agga, A., Abbou, A., Labbadi, M., El Houm, Y., and Ali,

I. H. O. (2022). Cnn-lstm: An efﬁcient hybrid deep

learning architecture for predicting short-term photo-

voltaic power production. Electric Power Systems Re-

search, 208:107908.

Bao, T., Zaidi, S. A. R., Xie, S., Yang, P., and Zhang, Z.-Q.

(2020). A cnn-lstm hybrid model for wrist kinemat-

ics estimation using surface electromyography. IEEE

Transactions on Instrumentation and Measurement,

70:1–9.

cheol Jeong, I., Bychkov, D., and Searson, P. C. (2018).

Wearable devices for precision medicine and health

state monitoring. IEEE Transactions on Biomedical

Engineering, 66(5):1242–1258.

Elgendi, M., Fletcher, R., Liang, Y., Howard, N., Lovell,

N. H., Abbott, D., Lim, K., and Ward, R. (2019). The

use of photoplethysmography for assessing hyperten-

sion. NPJ digital medicine, 2(1):60.

Gonz

alez, S., Hsieh, W.-T., and Chen, T. P.-C. (2023). A

benchmark for machine-learning based non-invasive

blood pressure estimation using photoplethysmogram.

Scientiﬁc Data, 10(1):149.

Kim, C.-S., Carek, A. M., Mukkamala, R., Inan, O. T.,

and Hahn, J.-O. (2015). Ballistocardiogram as prox-

imal timing reference for pulse transit time measure-

ment: Potential for cufﬂess blood pressure monitor-

ing. IEEE Transactions on Biomedical Engineering,

62(11):2657–2664.

Kiranyaz, S., Avci, O., Abdeljaber, O., Ince, T., Gabbouj,

M., and Inman, D. J. (2021). 1d convolutional neu-

ral networks and applications: A survey. Mechanical

systems and signal processing, 151:107398.

Lackland, D. T. and Weber, M. A. (2015). Global burden

of cardiovascular disease and stroke: hypertension at

the core. Canadian Journal of Cardiology, 31(5):569–

571.

Liang, Y., Liu, G., Chen, Z., and Elgendi, M. (2022). PPG-

BP Database.

Maqsood, S., Xu, S., Springer, M., and Mohawesh, R.

(2021). A benchmark study of machine learning for

analysis of signal feature extraction techniques for

blood pressure estimation using photoplethysmogra-

phy (ppg). Ieee Access, 9:138817–138833.

MHRA (2019). Blood pressure measurement devices.

Mukherjee, R., Ghosh, S., Gupta, B., and Chakravarty, T.

(2018). A literature review on current and proposed

technologies of noninvasive blood pressure measure-

ment. Telemedicine and e-Health, 24(3):185–193.

Reali, P., Lolatto, R., Coelli, S., Tartaglia, G., and Bianchi,

A. M. (2022). Information retrieval from photo-

plethysmographic sensors: A comprehensive compar-

ison of practical interpolation and breath-extraction

techniques at different sampling rates. Sensors,

22(4):1428.

Sana, F., Isselbacher, E. M., Singh, J. P., Heist, E. K.,

Pathik, B., and Armoundas, A. A. (2020). Wearable

devices for ambulatory cardiac monitoring: Jacc state-

of-the-art review. Journal of the American College of

Cardiology, 75(13):1582–1592.

Sapra, A., Malik, A., and Bhandari, P. (2020). Vital sign

assessment.

Slapni

car, G., Mlakar, N., and Lu

strek, M. (2019). Blood

pressure estimation from photoplethysmogram using

a spectro-temporal deep neural network. Sensors,

19(15):3420.

Van Houdt, G., Mosquera, C., and N

apoles, G. (2020). A

review on the long short-term memory model. Artiﬁ-

cial Intelligence Review, 53:5929–5955.

Xing, D., Wang, Y., Sun, P., Huang, H., and Lin, E. (2023).

A cnn-lstm-att hybrid model for classiﬁcation and

evaluation of growth status under drought and heat

stress in chinese ﬁr (cunninghamia lanceolata). Plant

Methods, 19(1):66.

Xing, X., Ma, Z., Zhang, M., Zhou, Y., Dong, W., and Song,

M. (2019). An unobtrusive and calibration-free blood

pressure estimation method using photoplethysmog-

raphy and biometrics. Scientiﬁc reports, 9(1):8611.

A Novel Cuff-Less and Calibration-Free Blood Pressure Estimation Framework Using Single Photoplethysmogram

703