A Novel Contactless Blood Pressure Measurement System and Algorithm Based on Vision Intelligence

Abstract

The measurement of vital signs such as blood pressure plays a key role in human health. Usually, we encounter some problems when we check them in the traditional way; for example, it is impossible to check continuously, and measuring vital signs requires direct contact with the patient, which can be uncomfortable for certain individuals. In this research, we present a vision-based system for estimating blood pressure using pulse transit time (PTT) and the Eulerian video magnification (EVM) technique to amplify tiny color variations caused by blood flow to calculate arterial pulse waves traveling between two arterial sites. Calculating the PTT by processing the video signal for each subject, an oscillometer BP device was used to evaluate the performance between measurements in different conditions, including rest, exercise, and during recovery. Mean systolic BP was 115 mmHg at rest, 137 mmHg during high-intensity exercise, and 114 mmHg during recovery, respectively. The average value of diastolic blood pressure did not change significantly before, during, and after exercise. When we compared the systolic and diastolic blood pressure with ground-truth results, our system showed an accuracy of 91% for systolic blood pressure and 90% for diastolic blood pressure.

1. Introduction

Blood pressure level is one of the critical factors that significantly impact human health. Normal systolic BP is considered <120 mmHg and diastolic BP < 80 mmHg, and high blood pressure, also known as hypertension, is systolic BP ≥ 130 mmHg or diastolic BP ≥ 80 mmHg [1]. Uncontrolled high blood pressure can cause serious health problems, increasing the risk of heart attack, kidney failure, and damage to the brain and eyes. Early detection of high BP plays an important role in preventing the progression of hypertension and its associated complications [2]. Low BP, or hypotension, in contrast, is frequently overlooked due to often presenting no symptoms. However, if left untreated, hypotension can lead to dangerous consequences. Hypotension is a decrease in arterial blood pressure, typically below 90 mmHg for systolic and 60 mmHg for diastolic. Severe hypotension can lead to shock and multiple organ failure, and could increase mortality risk [3]. Cuff devices commonly used to measure blood pressure rely on the Korotkoff method and require direct contact with the patient’s arm. This method blocks the arterial blood flow by inflating the cuff and then releasing the pressure, while the change in blood flow is determined by listening to specific sounds in the arteries. One notable drawback of cuff-based devices is their inability to provide continuous blood pressure monitoring. Furthermore, in clinical settings wherein cuff-based devices are commonly used, there is a potential concern regarding the spread of infections. These devices come into direct contact with the patient’s skin and may be used on multiple individuals throughout the day. Proper cleaning and disinfection procedures are difficult and time-consuming. In addition, contact-based devices can be uncomfortable and challenging, particularly for specific groups of individuals. For instance, people with certain disabilities might find it challenging to wear or adjust these devices. Elderly individuals (who may have fragile skin) and infants and babies might not always react well to having unfamiliar objects attached to them. Vision-based non-contact monitoring may be a solution to all the above-mentioned problems, as it allows the checking of vital signs without physical contact with the individual. By enabling the detection of blood pressure fluctuations in real time, Vision-based non-contact monitoring can alert individuals and healthcare providers to potential health issues before they become critical. This early detection can prevent the development of more severe conditions associated with hypertension, such as stroke or heart attack. In addition, by measuring blood pressure continuously, we can better understand the characteristics and behavior of blood pressure.

One widely and highly correlated approach to estimating blood pressure non-invasively is through pulse transit time (PTT) [4,5]. PTT refers to the time it takes for the arterial pulse wave to travel between two distinct points. In the study conducted by Martin et al. [6], the researchers used ballistocardiography and foot PPG waveforms to calculate PTT and pulse arrival time (PAT). The study demonstrated that PTT is a more advantageous marker of BP than PAT. The correlation coefficient between PTT and diastolic BP was −0.80 ± 0.02, with a root-mean-squared error (RMSE) of 7.6 ± 0.5 mmHg. In contrast, the correlation coefficient for PAT and diastolic BP was −0.60 ± 0.04, with an RMSE of 14.6 ± 1.5 mmHg. In practice, PTT is extracted from the time delay between the R wave of the electrocardiogram (ECG) and the photoplethysmogram (PPG) pulse wave measured from the finger. Wibmer et al. [7] explored the potential of PTT as a surrogate marker for blood pressure changes during cardiopulmonary exercise tests. Analyzing data from 20 patients, the researchers recorded ECG and finger-photoplethysmography to calculate PTT at multiple stages: at rest, peak exercise, and recovery. They reported mean systolic BP and PTT at rest as 128 mmHg and 366 ms, which changed significantly to 197 mmHg and 289 ms at peak exercise, returning to 128 mmHg and 371 ms during the recovery phase. Their findings revealed a strong correlation between PTT and systolic BP, with a weaker correlation observed with diastolic BP. In finding the difference between ECG and PPG signals using the PTT method, researchers have achieved very good results. HT Ma et al. [8] used ECG and PPG recordings to estimate BP. Signals received from the two parts of the body were firstly filtered through a sliding window, and then the time interval between the R-wave of the ECG and the peak of the PPG pulse was calculated. The proposed method was compared with the electronic sphygmomanometer device, and the results showed a mean error of −0.2 ± 2.4 mmHg for systolic BP and 0.5 ± 3.9 mmHg for diastolic BP. Chen et al. [9] demonstrated a linear relationship between systolic BP and pulse arrival time. In their study, the proposed method was tested on 20 patients in cardiovascular surgery. They used an ECG and a finger blood oxygen sensor to determine pulse arrival time. They calculated blood pressure assuming that the vascular properties are constant. When the obtained result was compared with an invasive radial arterial catheter, the estimated systolic BP value showed a high correlation with the reference values, reaching a mean of 0.97 ± 0.02. Additionally, the error range was within ±10%. Although the study showed promising results, it is important to note that this blood pressure measurement method is contact-based. To obtain PTT, some studies used a bioimpedance plethysmogram to replace ECG, or PPG, or used a ballistocardiogram signal, which reflects the mechanical vibration caused by the heart beating [10,11].

Another method is the camera-based photoplethysmography technique. In obtaining non-contact blood pressure signals, researchers use different algorithms based on camera-based photoplethysmography, also known as remote photoplethysmography (rPPG) or image photoplethysmography (iPPG). Researchers selected regions of interest (ROI) from different body sites to obtain PTT. Murakami et al. [12] selected the wrist and ankle as regions of interest (ROIs) to investigate the relationship between (BP) and (PTT). The signal peaks were observed in both regions through the use of finite impulse response (FIR) low-pass filtering and phase delay compensation. In this research, PTT showed a significant correlation of −0.88 with blood pressure levels. In the study conducted by Fan et al. [13], the researchers employ a tracking algorithm to continuously detect face and palm areas through video captured with a camera. They used the chrominance-based method to extract signals and to enhance the quality of the rPPG signals; the researchers implemented an adaptive Gaussian model. The result of the experiment shows a high correlation between PTT and BP, and the proposed method showed that the camera is capable of estimating blood pressure. Liu et al. [14] used two cameras by placing one camera on the forehead and another camera on the finger; they detected maxima and minima in iPPG signals and found BP by calculating PTT. The proposed method showed a PTT correlation coefficient (0.86 ± 0.06). To study the effect of finger temperature on the PPG signal, they measured the temperature of the finger before and after the experiment. The results showed that there was no significant difference in the finger’s temperature during the experiment. Al-Naji et al. [15] selected both ROIs from the forehead. They measured systolic and diastolic blood pressures by extracting the features from selected locations using the intrinsic mode functions (IMFs) method. When they compared the experimental results with the commercial digital sphygmomanometer method, they showed promising results. In another study, researchers Oiwa et al. [16] examined the relationship between skin temperature and mean BP. They estimated blood pressure using an infrared thermography camera. This camera, operating at 1 frame per second (FPS) with a temperature resolution of less than 0.1 °C, processed video to measure facial skin temperature. The right and left cheek, right and left forehead, and nasal and background regions were defined as ROIs, and facial skin temperature was measured from them. According to the results of temperature measured from all ROIs, nasal skin temperature and BP were correlated, the correlation coefficient between nasal skin temperature and measured mean BP was 0.768, and nasal skin temperature showed a minimum error rate compared to other regions of the face. Finkelstein et al. [17] selected the forehead and palm as ROIs for video recordings. They employed a high-speed infrared camera with a frame rate of 420 FPS. After extracting the red channel and averaging the pixels, they detrended the signal using a curve fitting method, and using a low-pass filter they identified a maximum point of the signal. PTT was then calculated based on the maximum derivative points obtained from each ROI. In this study, it was concluded that BP can be estimated by calculating PTT from the video captured by the camera. Takahashi et al. [18] presented the relationship between PPT and blood pressure using an RGB camera. A high-speed camera with a frame rate of 500 fps was used in the experiment. They studied the relationship between PPT and iPPT. Next, they studied the relationship between iPPT and BP from the face, and the result indicated that it is possible to obtain BP from the face using an RGB camera. In the study conducted by Kaito et al. [19] to estimate the blood pressure on the faces of the participants, a video was taken at a speed of 160 frames, and then the video was converted into a spatial pulse wave signal to find special features from the face video. They also conducted an experiment to explore how the frame rate affects signal quality. They demonstrated that the correlation coefficient is reduced at lower frame rates due to the loss of signal characteristics. This is in contrast to the higher frame rates wherein the correlation coefficient remains higher.

Some researchers have used deep learning methods to predict BP. Patil et al. [20] selected the forehead region as the ROI; they extracted the following features from the ROI: systolic amplitude, pulse interval, systolic slope, diastolic slope, peak interval, crest time, and delta time, using independent component analysis as the preprocessing method for rPPG. The authors used features extracted from rPPG to train the model, and they used a feed-forward neural network with one hidden layer for regression. The experiment was conducted in the afternoon and evening, and with the participation of 20 people. The morning video was used to train the neural network and the rest was used for testing purposes. The average error rates for systolic BP were 9.62% for the afternoon and 8.4% for the evening session, while for diastolic BP, the error rates were 11.63% for the afternoon and 11.18% for the evening session.

Zou et al. [21] selected the forehead and palm as an ROI to obtain signals. After extracting the green channel from the selected ROIs and processing the signal, they found the time interval obtained by two signals from the forehead and the palm. To predict blood pressure, they used a 2-layer neural network with 128 neurons in each layer; 90% of the data are used as the training set, and 10% as the test set. As a reference device, they used an Omron wrist electronic sphygmomanometer. The method achieves an average error rate of −9.28~3.16% for systolic BP, and the error for diastolic BP is −9.84~4.35%. However, one of the major drawbacks of deep learning methods is their sensitivity to the size of the data set. This limitation can affect the ability to achieve high levels of accuracy, as these models typically require large amounts of data for effective training. When the available dataset is limited in size or diversity, several issues can arise.

A noticeable limitation observed in a majority of these research studies is that they measure blood pressure under resting conditions. In this study, we examined BP not only under resting conditions, but also in the after-active-exercise state. The reason for checking BP during active exercise is to measure high blood pressure, because typically, blood pressure rises during exercise. It is normal for a healthy person to have a systolic blood pressure of 150 and above during exercise [22]. The basic approach of this paper is to estimate BP based on amplifying the subtle changes in the video, using EVM [23] to find PTT from two different ROIs. EVM has already demonstrated efficient and positive results in pulse detection [24].

2. Materials and Methods

2.1. System Setup

Five young healthy subjects participated in this study. The experiment was performed with a full explanation of the experimental procedure and written informed consent. The average room temperature was 25 °C. Before the test, the participants rested for 5 min. To measure high blood pressure, we measured one participant’s blood pressure one minute after high-intensity exercise. This process of measuring high blood pressure was carried out for 2 days at different times of the day. The participant’s BP was measured under the same conditions before and after physical activity.

In the experiment process, participants were asked to rest for a few minutes while sitting in a chair in order to receive signals from their faces. A camera (Sony RX100 VI, Sony Corporation, Tokyo, Japan, 20.1 MP, 960 fps HFR mode) and artificial light were installed approximately 60 cm from them, as shown in Figure 1. The camera needs artificial light; proper lighting is essential, as we measured blood pressure by detecting the PTT at two points: the forehead and the chin. Overhead lighting was insufficient, as it did not adequately illuminate the chin area. A total of 25 videos were collected, each 1 min in length, with a resolution of 1920 × 1080 pixels and 460 FPS. As a reference device, standard blood pressure was measured with an oscillometer (OMRON HEM-790IT, OMRON HEALTHCARE, Bannockburn, IL, USA) on the left upper arm of the subjects. The mean systolic blood pressure for all participants was 115 mmHg at rest; subjects participating in high-intensity exercise had a systolic BP of 137 mmHg and 114 mmHg during recovery, as shown in Figure 2.

2.2. The Proposed Algorithm

The proposed system includes the measurement of blood pressure by extracting the plethysmography signals from the face and finding the difference of the pulses in it. For this, we first detect the facial region automatically. After face recognition, we can extract the ROIs we need. The coordinates of specific facial areas, forehead, and chin, are isolated for further analysis. Figure 3 shows the forehead and chin signals. We selected the forehead and chin areas of the face for our study because these regions minimally influence facial expressions of emotion. Another reason is the relatively long distance between these two points on the face, as a longer distance allows more accurate detection of the PTT.

In our PTT measurement process, we are specifically analyzing the time interval between pulses in two distinct areas of the face. Ensuring the accurate identification of the peak of each pulse wave is crucial in this method. One way to enhance the precision of our measurements is by using a camera with a higher frames-per-second (FPS) rate. An increased FPS ensures that we capture more frames in the same time span, reducing the time lapse between each captured frame and consequently providing smoother and more detailed information about the pulse waves.

After detecting the face, we represent the frame at multiple scales, which is applied in the process of resizing down and resizing up frames from their original format of a Gaussian pyramid. We utilize a Gaussian pyramid approach to process the frames, reducing high-frequency noise and preserving the essential facial features necessary for accurate pulse signal extraction. The amplification process involves the enhancement of the signal in a detected frame, which helps to reveal the changes in blood flow through the detected area; this allows us to enhance subtle changes in skin color and monitor blood vessel movement in real time, providing valuable information on how the cardiovascular system works. It is crucial to carefully select the appropriate amplification factor, and attention must be taken to balance between the desired level of amplification because it causes increased noise and artifacts. After obtaining the signal datathe fast Fourier transform (FFT) was used to transform the data from the time domain to the frequency domain. Bandpass filters were used to separate the signals to differentiate them. With the help of these filters, the acceptable frequency range was separated from the undesirable frequency components.

ROI1 and ROI2 values were calculated based on the detection of peaks in selected signals. The maximal power value is selected as an indicator of the ROIs, as it represents the peak intensity of the signal. PTT is then calculated by measuring the time delay between the peaks detected in the pulse signals from the ROI1 and ROI2. In summary, EVM amplifies subtle color changes in the facial video, which correspond to blood flow. PTT is then measured as the time difference between pulse waves in two ROIs. These data, combined with mathematical models, allow us to estimate BP. A brief description of the overall algorithm is shown in Figure 4.

2.3. Measurement of Pulse Transit Time

During systole, the central arteries expand, driving blood toward the narrower distal arteries, and they contract during diastole [25]. The Moens–Korteweg equation [26] describes the relationship of the pressure wave velocity (PWV) with the blood vessels. Researchers have used different models to explore the mathematical relationship between BP and PTT. Proença et al. [27] obtained signals from the finger and earlobe, and proposed a logarithmic equation to estimate BP from PTT.

BP = a × ln(PTT) + b

(1)

Secerbegovic et al. [28] calculated PTT from the forehead and palm, and they used a linear regression model to estimate the BP:

$$\mathrm{B}\mathrm{P}=\mathrm{a}\times \mathrm{P}\mathrm{T}\mathrm{T}+\mathrm{b}$$

(2)

where a and b are constants calculated from the first post-exercise recording.

The relationship between blood pressure and HR has been studied by many researchers, and integrating heart rate as an additional parameter when finding BP using PTT improves accuracy [29,30,31].

$$BP=aPTT+bHR+c$$

(3)

a, b, and c are coefficients found by calibration, and are estimated using the appropriate linear equation. Equation (3) is used in this study to estimate BP. From this formula, we calculate systolic BP and diastolic BP separately.

$$\begin{gathered} SBP=a_{SBP}PTT+b_{SBP}HR+c_{SBP} \\ DBP=a_{DBP}PTT+b_{DBP}HR+c_{DBP} \end{gathered}$$

(4)

In the calibration process, we collected initial data using our camera-based system and ground-truth BP. From regression analysis, we find the best-fit linear relationship between variables. After the regression analysis, we obtained the coefficients a, b, and c. These coefficients represent how changes in PTT and HR correlate with changes in blood pressure.

Coefficient a represents the sensitivity of blood pressure changes to changes in PTT.

Coefficient b represents the correlation of blood pressure with heart rate.

Coefficient c is the intercept term, accounting for baseline blood pressure. After determining initial values for a, b, and c, we conducted further testing to validate and adjust these coefficients. It is important to note that while this calibration worked well in our controlled experimental setup, different setups or other external factors may require recalibration and could lead to different coefficient values; this may affect the accuracy of the system.

3. Experimental Results

In order to evaluate the performance of the proposed method, we analyzed 25 video recordings of five participants at rest and during high-intensity exercise, each video being one minute in length. We compared our results with those of a standard blood pressure oscillometer device, and we calculated the accuracy, mean absolute error (MAE), and mean absolute percentage error (MAPE). Figure 5 shows a comparison between the BP ground-truth result and BP estimated by our system. In Figure 6, the box plot provides a visual representation of the data, showing the results obtained from resting BP and after-exercise BP measurements with the reference device. Comparing resting BP and after-exercise BP is crucial. It allows us to assess how physical activity impacts blood pressure levels, which is essential for understanding the cardiovascular response to exercise. Furthermore, by comparing these measurements to the reference device, we can evaluate the accuracy of the proposed method.

The proposed system demonstrated high accuracy in various tests. Under resting conditions, the system’s accuracy for measuring systolic blood pressure was 91%, and for diastolic blood pressure, it was 90%. Even during high-intensity exercise, the system maintained good performance, with accuracy rates of 90% for systolic and 89% for diastolic measurements. In terms of the system’s MAE and MAPE, the results are as follows. For systolic blood pressure, the MAE was 9.4 mmHg, with a MAPE of 8.2%. Meanwhile, for diastolic pressure, we found an MAE of 8 mmHg and a MAPE of 9.8%. Compared to other studies, the research in [15] showed an average error of 9.62% SBP and 11.63% DBP, the study in [10] reported errors of 8.42% SBP and 12.34% DBP, and the findings in [16] presented an error average of 9.28% SBP and 9.84% DBP, as shown in

Table 1. Performance comparison of the system based on mean absolute percentage error (MAPE).

Authors	SBP (mmHg)	DBP (mmHg)
Fan et al. [13]	8.42%	12.34%
Omkar et al. [20]	9.62%	11.63%
Zou et al. [21]	9.28%	9.84%
Ours	8.2%	9.8%

.

4. Discussion

In this study, we focused on measuring blood pressure from the facial region. We used EVM for signal processing and used PTT for blood pressure measurement. The presented camera-based non-invasive blood pressure monitoring system showed promising results in terms of accuracy and reliability, both at rest and during high-intensity exercise. This methodology allowed us to test the reliability of the system under non-static conditions. While most studies are limited to more controlled and relaxed states, we tested how well our system performs when the subject is in a post-exercise state. Non-invasive continuous monitoring systems are convenient for checking BP at any time of the day. In addition, continuous monitoring provides a comprehensive study of blood pressure, which can be particularly useful for detecting hypertension or evaluating the effectiveness of antihypertensive treatment [32].

Future research will focus on expanding to include participants with diverse demographics, various age groups, skin types, and individuals with different health conditions, including hypertension and hypotension. In addition, signal processing will be improved, and a real-time monitoring system developed for artifacts in motion, in different lighting conditions, and in the wider environment. The ultimate goal is to make this technology a reliable tool for home health monitoring and telemedicine applications, thereby increasing access to BP monitoring for populations worldwide. In the future, with the collection of large amounts of data, we will be able to monitor not only BP, but also a wide range of heart diseases. Certain diseases affecting the cardiovascular system can result in increased arterial stiffness. When the arteries become stiffer, the heart has to work with extra pressure. This directly affects the dynamics of blood flow. The heart has to work harder to pump blood through these less flexible vessels, which then increases the pressure within the arterial system. This phenomenon is one of the factors leading to high blood pressure [33]. The presence of cardiovascular disease may influence the relationship between PTT and systemic blood pressure. This may provide additional information about the functioning of the cardiovascular system [34], and analyzing patterns and anomalies will allow for the early prediction of various heart-related illnesses.

Although the current study shows promising results, there are some limitations that affect this system’s accuracy. One noteworthy external factor is the presence of makeup on the forehead or chin. When makeup is applied to these areas, it creates a layer that can reduce the system’s ability to capture signals precisely.

5. Conclusions

This research demonstrated the usefulness of the camera-based algorithm in monitoring blood pressure. The proposed method is non-contact, which helps individuals who are particularly sensitive to devices that require physical contact, and it prevents the spread of infectious diseases. The results showed that our model can estimate blood pressure under different conditions. The system could be beneficial in telemedicine and for at-home use, especially for individuals with chronic conditions that require regular BP monitoring. Additionally, the system could be adapted for use in fitness and wellness centers, helping individuals track their blood pressure response to exercise in real time. In addition, this approach to blood pressure monitoring gives people the opportunity to be informed about their health. They can change their diet, increase physical activity, or manage stress to reduce the risk of the recurrence of hypertension.