Blood Pressure Monitoring Based on Flexible Encapsulated Sensors

Abstract

Blood pressure monitoring is a significant concern in the field of healthcare, and the utilization of flexible encapsulated sensors presents a promising solution for achieving noninvasive and comfortable monitoring. This paper presents a study on the flexible encapsulation of MEMS pressure sensors and the development of an enhanced arterial tonometry method for blood pressure measurement, ultimately leading to the realization of a blood pressure monitoring system based on flexible encapsulated sensors. To improve wearer comfort and acquire reliable pulse signals, a flexible encapsulation sensor combining parylene and PDMS materials was fabricated. Additionally, to address the issue of low accuracy in blood pressure measurement, various machine learning algorithms were compared and analyzed, leading to the identification of the random forest model as the optimal regressor. Consequently, a blood pressure monitoring system based on the improved arterial tension method was designed and implemented. The experimental results demonstrate that the proposed system achieved a significant enhancement of 31.4% and 21% in the accuracy of systolic and diastolic blood pressure measurements, respectively, compared with the arterial tension method.

1. Introduction

Hypertension is a common chronic disease that gradually causes irreversible damage to the heart, brain, kidneys, and other organs of affected individuals. Currently, there are 245 million hypertensive patients in China, but the awareness, treatment, and control rates among patients are only 51.6%, 45.8%, and 16.8%, respectively [1]. Hypertension is difficult to detect, develops rapidly, and has a high mortality rate, requiring long-term and continuous blood pressure monitoring for disease tracking and treatment [2,3,4].

Continuous blood pressure measurement requires high accuracy and sensitivity from the sensor. Pressure sensors are known for their high accuracy and good sensitivity, and they are not affected by changes in light intensity. However, traditional pressure sensor chips are typically made of rigid materials such as a metal or semiconductor. Prolonged wearing of these sensors can lead to corrosion from sweat, which affects measurement accuracy and may even result in damage to the sensor chip. Based on the above situation, Peng et al. discovered that Mxene/rGO/P(VDF-TrFE) hybrid material exhibits excellent electrical performance and mechanical strength. Therefore, they fabricated a flexible sensor based on this material [5]. Furthermore, scholars have utilized polydimethylsiloxane (PDMS) as a flexible material for sensor fabrication. Among them, Sylgard 184 PDMS has been widely employed in the field of chip technology [6,7,8,9,10,11]. This material exhibits excellent biocompatibility; and its stiffness, elasticity, and other properties are close to those of human skin [12,13]. Kumar A et al. fabricated high-performance flexible sensors using deionized water (DIW) and PDMS [14]. Choong C et al. developed pyramid microstructured sensors based on PDMS [15]. Xin Li et al. prepared a flexible dual interdigital electrode sensor using PDMS, which exhibited high sensitivity, stretchability, a large measurement range, and outstanding stability [16]. Some researchers have found that poly-para-xylylene (parylene) possesses excellent corrosion resistance, moisture barrier, biocompatibility, and thermal stability [17,18,19,20]. This material is widely utilized in fields such as biology and healthcare. For instance, in references [21,22], parylene was employed to encapsulate bioelectrodes. References [23,24,25] applied parylene in retinal transplantation surgeries. Wu P et al. used parylene for the water sealing of micro-electromechanical system (MEMS) sensors and achieved devices with excellent waterproofing performance and high sensitivity [26]. Wang Y et al. summarized the applications of parylene in MEMS [27].

Continuous blood pressure measurement methods can be classified as invasive or noninvasive, with invasive measurement being the gold standard but requiring high operational standards and potentially causing infections, bleeding, blood clots, and other complications [28]. The noninvasive continuous blood pressure monitoring methods mainly include the arterial tonometry method and the pulse wave parameter method, among others [29,30]. Of these methods, pulse wave parameter measurement extracts feature parameters from pulse waveforms, uses regression analysis to establish the relationship between blood pressure and pulse wave, and thus realizes continuous noninvasive blood pressure measurement [31,32]. Many products on the market now use photoplethysmography sensors to collect pulse signals and obtain pulse wave feature parameters to calculate blood pressure values [33], such as the Mi bracelet and HUAWEI smartwatch. Photoplethysmography sensors are small, highly integrated, easy to use and flexible, and can be integrated into portable devices to enable wearable blood pressure measurement. However, photoplethysmography sensors are sensitive to changes in light intensity, skin color, and sweat, which can lead to poor device accuracy based on this monitoring principle, making it difficult to meet medical standards (for example, AAMI standards) [34,35,36]. Arterial tonometry is a method of gradually applying external pressure to achieve equal internal and external pressure in the blood vessels and then using a pressure sensor placed above the artery to measure external pressure values to calculate arterial blood pressure [37]. Currently, blood pressure monitoring devices based on this method, such as the US TL-300 and Japan CBM-7000 noninvasive blood pressure measurement systems, have good blood pressure measurement accuracy [38,39].

Existing studies have achieved good results. However, the current fabrication processes for related flexible sensors are relatively complex and still in the developmental stage. Therefore, in this study, a flexible packaging sensor combining parylene and PDMS was fabricated to enhance the comfort of the wearer and prevent sweat erosion, while achieving the stability and reliability of traditional rigid pressure sensors. The arterial tonometry method, which relies solely on the peak and trough values of the pulse waveform for blood pressure calculation, fails to comprehensively consider other pulse wave parameters. This limitation leaves room for improvement in the accuracy of blood pressure measurement using this method. Therefore, we compared and analyzed various machine learning algorithms and determined that the random forest model was found to be the optimal regressor. A blood pressure monitoring system based on an improved arterial tonometry method was designed and implemented. The proposed blood pressure monitoring system, based on flexible packaging sensors, provides a theoretical foundation and has application value for achieving wearable blood pressure measurement and precise diagnosis and treatment of conditions such as hypertension.

2. Analysis and Research of Blood Pressure Measurement Methods

2.1. The Arterial Tonometry Method

For the arterial tonometry method for measuring blood pressure, superficial arteries, such as the radial artery, femoral artery, and carotid artery, are usually selected as measurement sites. In this study, the radial artery was chosen as the measurement location because it has a larger diameter and is supported by rigid bone tissue at the bottom [40].

As shown in Figure 1, the pressure sensor is placed above the radial artery and an external pressure F_c is applied to the artery to obtain the blood pressure F_p. The red region represents the blood vessel, T represents the circumferential tension in arterial wall, and θ is the angle between tension T and the horizontal direction. The blood pressure F_p and the external force F_c satisfy the equation:

$$F_p=F_c+2T\sin \theta$$

(1)

During the measurement process, the external pressure F_c gradually increases until it equals the blood pressure F_p. When the external pressure F_c is too small, the blood vessel does not reach a flattened state, resulting in a smaller pulse peak value and unclear waveform feature points. As the external pressure F_c gradually increases to the blood pressure level F_p, the blood vessel becomes flattened, resulting in clear waveform feature points and a maximum pulse peak value. At this point, the pulse peak value corresponds to the systolic blood pressure (SBP), and the pulse trough value corresponds to the diastolic blood pressure (DBP).

The arterial tonometry method requires the long-term application of external dynamic pressure to keep the blood vessels in a flattened state. Therefore, existing monitoring devices rely on air pumps or control motors to determine the pressure range, resulting in a bulky, complex, and expensive measurement system. In addition, the arterial tonometry method only uses the peak and trough values of the pulse waveform for calibration and does not fully utilize the features that are highly correlated with blood pressure values for calibration.

2.2. The Pulse Wave Parameter Method

The pulse wave parameter method is a continuous and noninvasive blood pressure measurement technique that involves extracting characteristic parameters from the pulse waveform and establishing a regression analysis between blood pressure and the pulse wave. Some characteristic parameters of the pulse signal are shown in Figure 2, including the pulse onset; the percussion peak; the preceding tidal wave trough; the tidal wave; the dicrotic notch; the dicrotic wave peak; systolic time (St); diastolic time (Dt); time required to reach 10%, 25%, and 50% of the pulse peak height during systole (St10, St25, and St50, respectively); and the time required to reach 10%, 25%, and 50% of the pulse amplitude during diastole (Dt10, Dt25, and Dt50, respectively), among others [41]. Currently, most blood pressure monitoring devices based on the pulse wave parameter method use photoelectric sensors to collect pulse signals. However, photoelectric sensors are susceptible to changes in light intensity, skin color, sweat, and other factors that can lead to low measurement accuracy.

2.3. An Improved Arterial Tonometry Blood Pressure Measurement Method

To address the issues of the arterial tonometry method, this paper proposes an improved arterial tonometry blood pressure measurement method. This method combines the preliminary blood pressure values calculated using arterial tonometry with other extracted pulse wave parameters to form a new feature set. Through machine learning algorithms, more accurate blood pressure measurement values are obtained.

The steps of the proposed blood pressure measurement method are as follows: (1) by adjusting the tightness of the cuff to control the magnitude of the external pressure and observing the pulse wave amplitude to adjust the external pressure, the vessel is considered flat when the pulse wave amplitude is maximized; (2) after the pulse wave amplitude reaches its maximum, the pulse signal is collected, and the pulse wave parameters are extracted. Based on the extracted pulse wave parameters such as peak and trough values, the preliminary blood pressure value is calculated using the arterial tonometry method. The obtained preliminary blood pressure value is then fused with other pulse wave parameters such as peak and trough values, tidal wave anterior valley value, pre-rebound wave value, descent and inflection valley value, rebound wave peak value, systolic period St, diastolic period Dt, etc., to form a new feature set. The machine learning algorithm is then used to achieve more accurate blood pressure measurements.

3. Flexible Packaging and Testing of Sensors

Sensors are a critical factor affecting the quality of pulse signal acquisition. Therefore, after designing a reliable measurement method, we developed a packaging method based on parylene and polydimethylsiloxane (PDMS) to address the limitations of using micro-pressure sensors for blood pressure monitoring. The packaged sensor exhibited high accuracy and sensitivity, and pulse signal testing was successfully conducted using the packaged sensor.

3.1. MEMS Silicon Piezoresistive Pressure Sensors

The MEMS silicon piezoresistive pressure sensor utilizes the piezoresistive effect of silicon to convert pressure signals into electrical signals through a Wheatstone measurement bridge. The physical photo of the MEMS silicon piezoresistive pressure sensor used in this study is shown in Figure 3, which is highly sensitive, small (2 mm × 2.5 mm × 0.9 mm), low cost, and highly precise, and its power consumption is low. Based on these features, this type of pressure sensor was applied to pulse testing in this study.

3.2. Flexible Packaging Method for Sensors

If pressure sensors are directly placed on the radial artery for long-term measurements, the sensors may be affected by sweat and/or damage and cause discomfort to the wearer. Therefore, this paper proposes a flexible packaging method for MEMS pressure sensors.

Polydimethylsiloxane (PDMS), poly-para-xylylene (parylene), polyurethane (PU), and polyimide (PI) are commonly used flexible materials. Among them, PDMS more easily bends and deforms, has stable chemical properties, is easy to process, has a low manufacturing cost, and has good biocompatibility. Its stiffness, elasticity, and other properties are close to those of human skin, so can greatly reduce discomfort during wear [12,13]. However, PDMS also has some drawbacks. Its thermal expansion coefficient significantly differs from that of silicon, which can cause cracks at the junction of the PDMS and the silicon chip during the sensor packaging process due to temperature changes, affecting the quality of signal acquisition. Meanwhile, PDMS is porous and easily absorbs moisture during use, affecting measurement accuracy. Parylene-C, as a commonly used organic thin film coating material, also possesses desired characteristics such as biocompatibility, chemical inertness, and transparency. It has fond extensive applications in fields such as biology and medicine. Compared with PDMS, parylene exhibits better adhesion to silicon wafers and is impermeable to moisture, sweat, and other substances. It possesses excellent compactness and corrosion resistance. However, the fabrication process of parylene films is relatively complex, requiring specialized deposition equipment. For thick substrates, a prolonged deposition time is necessary. The resulting films are highly rigid and do not meet the requirements for flexible packaging.

Based on these considerations, we first applied a layer of parylene on the sensor surface through a vapor deposition process to increase the sensor’s corrosion resistance. Then, the sensor was flexibly packaged with PDMS to ensure the subject’s comfort. The schematic diagram of the sensor packaging scheme is shown in Figure 4.

3.3. Flexible Packaging Process for Sensors

The flexible packaging process of the sensor was as follows:

(1) A 0.5 μm thick parylene-C layer was deposited on the surface of the sensor using vapor phase deposition. The specific deposition method was as follows: The parylene precursor material was heated to 150 °C, causing it to sublimate from a solid state to a gas phase. The resulting sublimated dimer gas was placed in a pyrolysis chamber, where the temperature was further increased to 650 °C. At this temperature, the bonds of the p-xylene dimer were broken, resulting in the formation of free radical monomers. Finally, the free radical monomers were deposited onto the sensor surface at room temperature, forming a thin parylene film.

(2) The curing agent and PDMS base were added to plastic cups in mass ratios of 1:30, 1:35, 1:40, and 1:45, separately, to achieve the desired thickness at the bottom of the cups.

(3) The PDMS mixture was stirred for about five minutes to ensure a uniform mixture.

(4) The sensor was placed on the surface of the PDMS mixture in the plastic cup, allowed to stand for 15 min, and then placed it in a vacuum drying oven to vacuumize the mixture and remove any air bubbles.

(5) The prepared PDMS mixture was poured onto the surface of the sensor, and a spin coater was used to spin-coat the mixture to a thickness of 0.3 mm, evenly coating the surface of the sensor.

(6) The temperature of the drying oven was set to 80 °C and maintained for 2 h to cure the PDMS mixture.

(7) The pressure sensor with the PDMS film covering the sensitive area was removed, appropriate adjustments were made, and the production was completed.

The flexibly packaged sensor can effectively prevent sweat corrosion of the chip and adhere well to the skin, so the subjects do not feel uncomfortable during long-term measurement.

The sensing mechanism of the sensor is based on the silicon piezoresistive effect. Specifically, when external pressure is applied to the surface of the sensor, the pressure is transmitted from the flexible film composed of PDMS and parylene to the sensor chip. Upon sensing the external pressure, the internal resistance of the sensor changes, resulting in a variation in the sensor output.

3.4. Performance Testing of Sensors

In order to compare the performance of the sensor before and after packaging (the ratio of PDMS curing agent to base was 1:40), this study conducted a cyclic experiment of applying and releasing pressure to the sensor three times while controlling the external pressure. During the application of pressure, the external pressure was increased from 104 kPa to 114 kPa in 2 kPa steps. During the release of pressure, the external pressure was decreased from 114 kPa to 104 kPa in 2 kPa steps. The relationship curve between the external applied pressure and the output value of the packaged sensor is shown in Figure 5.

From the above figure, it can be observed that the encapsulated sensors exhibited hysteresis and good repeatability. Furthermore, the output values of the sensors were directly proportional to the applied external pressure. According to our calculations, the sensitivity of the encapsulated sensor was 0.9972, linearity was 0.26%, hysteresis was 0.0217%, repeatability was 0.127%, and accuracy was 0.1962. To compare the performance of the sensors before and after encapsulation, the accuracy and sensitivity of sensors before and after encapsulation were calculated and are presented in

Table 1. Comparison table of accuracy and sensitivity of the sensor before and after encapsulation.

	Before Encapsulation	After Encapsulation	Change
Accuracy/mmHg	0.0017	0.1962	0.1945
Sensitivity	0.9996	0.9972	0.0024

.

According to Table 1, it can be seen that the changes in the precision and sensitivity of the sensor after encapsulation were small, and the encapsulated sensor still retained high precision and sensitivity.

3.5. Testing of Pulse Signals

By installing sensors in a wristwatch and dynamically adjusting the external pressure to observe the peak-to-peak values of the pulse, when the peak-to-peak value was at its maximum, maintaining the pressure constant was considered to result in a flattened state of the blood vessels. At this point, the relationship between blood pressure and the output of the pressure sensor was linear. The pulse waveform collected by the unpackaged MEMS pressure sensor is shown in Figure 6a, and the pulse waveform plots collected by the sensors encapsulated with PDMS in different ratios (with a uniform thickness of 0.3 mm) are shown in Figure 6b–d.

From Figure 6a, it can be seen that the pulse signal directly collected by the MEMS pressure sensor was of good quality; the relevant characteristics of the pulse wave signal, such as the pulse onset, the percussion peak, the preceding tidal wave trough, and tidal wave In order to prevent device damage caused by the direct contact between the skin and the sensitive area of the sensor chip, the hardness characteristics of PDMS with different mass ratios were studied. The experimental results showed that when the packaged PDMS was relatively hard (the ratio of curing agent to this agent was 1:30 or 1:35), the sensor chip was easily damaged by hard PDMS, and it was difficult for the harder PDMS to fully contact the curved skin surface to collect a good pulse waveform, as shown in Figure 6b. When the encapsulated PDMS had a softer texture (with a curing agent to base ratio of 1:45 or 1:50), the acquired pulse signal by the sensor, as shown in Figure 6d, was affected due to the close contact between PDMS and the skin surface. The signal was susceptible to interference from microarteries near the radial artery and other tissues, leading to a poor-quality collected pulse signal When the hardness of the packaged PDMS was moderate (the ratio of curing agent to this agent was 1:40), PDMS not only had good adhesion but could also fully adhere to the skin surface, so as to ensure that the sensor chip could be protected by the soft PDMS and, at the same time, collect good-quality pulse wave signals.

4. Blood Pressure Measurement

After flexible packaging, the sensors captured good pulse signals, but noise interference still remained. Therefore, to obtain accurate blood pressure values, further processing of the pulse signals is necessary, including filtering and extraction of peak and trough values. Subsequently, the preliminary blood pressure values are obtained based on arterial tonometry using the peak and trough values. Then, data preprocessing is performed by extracting other pulse wave feature parameters and normalizing the feature set. Finally, multiple machine learning algorithms are applied for regression analysis to achieve precise blood pressure measurement. The overall flowchart of the blood pressure calculation is shown in Figure 7.

4.1. Pulse Signal Processing

Power frequency interference, respiratory movements, and bodily shaking introduced noise to pulse waves, necessitating the use of a third-order Butterworth bandpass filter and a 50 Hz notch filter to eliminate it. Figure 8 displays the pulse wave signals before and after filtering, while their amplitude-frequency characteristics are presented in Figure 9.

As demonstrated in Figure 8 and Figure 9, the filter design effectively removed baseline noise and power frequency interference while retaining the essential feature points of the pulse signal.

Extracting these feature points is vital for accurate calculation of blood pressure values. To locate the peak points, the findpeaks function in Matrix Laboratory (MATLAB, version 2021b) was employed, and a new waveform was obtained by subtracting the existing pulse data from the maximum peak value. The trough value of the original waveform was determined by locating the peak points in the new waveform. Figure 10 shows the localization of the peak and trough points, and this method was effective in identifying them accurately.

4.2. Preliminary Blood Pressure Calculation Based on Arterial Tonometry

Prior to measurement, calibration using a standard sphygmomanometer was necessary to convert the pressure measurement values obtained from the MEMS pressure sensor into the corresponding blood pressure values. The specific method was as follows: First, a commercial blood pressure monitor was used to measure the systolic blood pressure (SYS) and diastolic blood pressure (DIA) on the left wrist of the subject for a certain period of time (e.g., 30 s), while the packaged sensor was used to perform arterial tonometry blood pressure testing on the right wrist of the subject. Based on the pulse waveform, the average values of the peak and trough values during this period, denoted by P_p and P_t, respectively, were calculated; a fitted relationship between the systolic and diastolic blood pressure and P_p and P_t was obtained, as follows:

$$SYS=k_1{P}_p+b_1$$

(2)

$$DIA=k_2{P}_t+b_2$$

(3)

At any given time, the output of the sensor is y(t), and the preliminary blood pressure value BP(t) after calibration is given by:

$$BP(t)=\frac{k_1+k_2}{2}y(t)+\frac{b_1+b_2}{2}$$

(4)

where k₁, k₂, b₁, and b₂ are constants. After blood pressure calibration, continuous long-term monitoring of blood pressure could be achieved.

4.3. Data Preprocessing

Data Preprocessing mainly includes two major steps: feature extraction and data normalization.

4.3.1. Feature Extraction

To reduce computational cost, this study extracted feature parameters (shown in

Table 2. Extracted feature parameters.

Feature Parameters	Definition
St	Systolic time
Dt	Diastolic time
HR	Heart rate
S-PTT	Transit time between two peaks
St10 + Dt10	Addition of systolic time and diastolic time @ 10% of the pulse amplitude
St10/Dt10	Division of diastolic time and systolic time @ 10% of the pulse amplitude
St25 + Dt25	Addition of systolic time and diastolic time @ 25% of the pulse amplitude
St25/Dt25	Division of diastolic time and systolic time @ 25% of the pulse amplitude
St50 + Dt50	Addition of systolic time and diastolic time @ 50% of the pulse amplitude
St50/Dt50	Division of diastolic time and systolic time @ 50% of the pulse amplitude
Pre-sbp	SBP obtained by arterial tonometry method
Pre-sdp	DBP obtained by arterial tonometry method

) for blood pressure prediction.

4.3.2. Data Normalization

The preliminary blood pressure values obtained through arterial tonometry were combined with the extracted pulse wave feature parameters to form a new feature set. Considering the significant differences in feature values, the data were normalized using the following formula:

$$\overline{X_{ij}}=\frac{X_{ij}-\mu \left(X_i\right)}{\sigma (X_i)}$$

(5)

In the formula, $\overline{X_{ij}}$ represents the normalized value; $X_{ij}$ represents the jth value of the ith feature; $\mu \left(X_i\right)$ and $\sigma (X_i)$ denote the mean and standard deviation of the ith feature, respectively.

4.4. Regression Prediction Based on Machine Learning Algorithms

The preprocessed feature set was divided into training and testing sets (in this study, the ratio of training set to testing set was 8:2). Machine learning regression models were used to train the feature set and perform regression prediction. Common machine learning regression algorithms include linear regression, ridge regression, support vector regression, random forest regression, and eXtreme Gradient Boosting (XGBoost) regression.

4.5. Experimental Method

A blood pressure measurement device was developed and used along with a commercial blood pressure monitor to measure the blood pressure of 20 volunteers in different states (10 measurements at rest and 10 measurements taken after running for 5 min and resting for 10 min). The specific method involved the use of a commercial blood pressure monitor to measure the systolic pressure and diastolic pressure (SYS and DIA, respectively) of the subject’s left wrist while using an encapsulated sensor to collect the pulse waveforms from the subject’s right wrist to extract the peak and valley values and other pulse wave parameter features. Based on Formulas (2) and (3), the blood pressure systolic and diastolic values were calculated using the arterial tonometry method. Then, all the extracted features were input into a machine learning regression model. The values measured by the commercial blood pressure monitor were considered as the true blood pressure values, while those calculated by the self-made blood pressure device were considered as the predicted values. To test the effectiveness of the blood pressure measurement method combining arterial tonometry and pulse wave parameter, this study calculated the blood pressure prediction error of the arterial tonometry method, the pulse wave parameters method, and the method combining arterial tonometry and pulse wave parameters.

The blood pressure prediction error refers to the difference between the true blood pressure value and the predicted value. To evaluate the prediction accuracy, commonly used regression prediction evaluation metrics were used, including the mean absolute error (MAE) and standard deviation (SD).

$$MAE=\frac{1}{n}{\displaystyle \sum _{i=1}^n\left|y_i-\widehat{y_i}\right|}$$

(6)

$$z_i=y_i-\widehat{y_i}$$

(7)

$$z_{mean}=\frac{1}{n}{\displaystyle \sum _{i=1}^n{z}_i}$$

(8)

$$SD=\sqrt{{\displaystyle \sum _{i=1}^n{(z_i-z_{mean})}^2}}$$

(9)

In the equations, n represents the number of samples, $y_i$ represents the true blood pressure value, $\widehat{y_i}$ represents the predicted blood pressure value, z_i is the difference between the predicted and true values of blood pressure for the ith sample, and z_mean is the average of z_i.

4.6. Results

The MAE and SD between the preliminary values of systolic and diastolic blood pressure obtained by the arterial tonometry method and those measured by the commercial electronic blood pressure monitor are shown in

Table 3. Results of arterial tonometry measurement.

Method	SBP		DBP
	MAE	SD	MAE	SD
AT	4.72	4.77	5.38	5.66

. The MAE and SD of the blood pressure values obtained by the different machine learning regression algorithms are shown in

Table 4. Different regression methods.

	Method	Machine Learning Algorithms
		Linear Regression		Ridge Regression		Support Vector Regression		XGBoost		Random Forest Regression
		MAE	SD	MAE	SD	MAE	SD	MAE	SD	MAE	SD
SBP	PWP	6.95	8.26	6.25	7.66	6.03	7.44	7.05	8.58	6.58	7.89
	AT-PWP	3.64	4.75	3.96	4.92	3.29	4.22	3.24	4.26	3.24	4.04
DBP	PWP	7.87	9.94	7.51	9.64	7.73	9.78	8.55	10.98	7.57	9.88
	AT-PWP	4.38	5.28	5.11	6.46	4.96	6.35	4.33	5.38	4.25	5.15

, where AT represents the arterial tonometry method; PWP represents the pulse wave parameter method, which means that the pulse wave parameter method was the feature set obtained by extracting features (excluding the preliminary values of systolic and diastolic blood pressure obtained by the arterial tonometry method) and using different regression algorithms to obtain the MAE and SD between the true and predicted blood pressure values. AT-PWP represents the blood pressure measurement method proposed in this paper, which combines the arterial tonometry method with the pulse wave parameter method. This method used the preliminary blood pressure values obtained by the arterial tonometry method and other extracted pulse feature parameters as input features and then uses different machine learning regression algorithms to calculate the MAE and SD between the predicted and true blood pressure values.

According to Table 3 and Table 4, the arterial-tonometry-based measurement method exhibited good measurement accuracy. Combining the arterial tonometry method with the pulse wave parameters method effectively improved blood pressure measurement accuracy. The accuracy of systolic and diastolic blood pressure obtained with this method improved by 31.4% and 21%, respectively, compared with those of the arterial tonometry method. Among them, the use of the random forest algorithm as the regressor achieved the highest blood pressure prediction accuracy. The mean absolute deviation for systolic blood pressure was 3.24 mmHg with a standard deviation of 4.04 mmHg, while the mean absolute deviation for diastolic blood pressure was 4.25 mmHg with a standard deviation of 5.15 mmHg. Therefore, for the blood pressure monitoring system developed in this study, we selected the random forest algorithm as the regressor.

To verify the consistency between the blood pressure values measured by the self-made blood pressure monitor and the commercial blood pressure monitor, a Bland–Altman plot was drawn, as shown in Figure 11. The horizontal axis of the plot represents the average of the two blood pressure measurement values, and the vertical axis represents the difference between the two blood pressure measurement values. The solid black line represents the mean difference (mean), and the red dashed lines represent the 95% confidence interval (mean ± 1.96 SD).

According to Figure 11, there were a few data points for systolic and diastolic blood pressure measured by this system and by the commercial blood pressure monitor that fell outside the limit lines (i.e., the 95% confidence interval). However, the majority of the data points were within the limit lines. Among the 80 sets of test data, the average deviation of systolic blood pressure measured by this system was 0.86 mmHg with a standard deviation of 4.04 mmHg, while the average deviation of diastolic blood pressure was 0.63 mmHg with a standard deviation of 5.15 mmHg, which satisfies the AAMI standard for clinical use (mean deviation < 5 mmHg, standard deviation < 8 mmHg).

5. Conclusions

This study investigated a sensor packaging method based on the combination of parylene and PDMS to address the issues of traditional pressure sensors in blood pressure monitoring. The encapsulated sensor exhibited high sensitivity and accuracy and could capture good pulse signals when installed in a wristband. Additionally, a blood pressure measurement method based on the combination of arterial tonometry and pulse wave analysis was developed. Specifically, a new feature set was formed by combining the extracted pulse wave features with the initial blood pressure values obtained from arterial tonometry, and multiple machine learning algorithms were used for regression analysis to achieve accurate blood pressure measurement. Among them, the random forest regression algorithm produced the smallest mean absolute error and standard deviation values for systolic and diastolic blood pressure and met the AAMI standard commonly used in clinical practice. The proposed method blood pressure monitoring based on flexible packaged sensors has great potential for precise the diagnosis and treatment of diseases such as hypertension.

This study addressed the limitations of traditional pressure sensors in the field of blood pressure monitoring by designing and implementing a novel blood pressure monitoring system using flexible encapsulated sensors. The following achievements were obtained:

(1) A sensor encapsulation method was developed by combining parylene and PDMS, resulting in highly sensitive and accurate encapsulated sensors. These sensors were integrated into wristbands, allowing for the reliable acquisition of pulse signals.

(2) An improved arterial tonometry blood pressure measurement method was proposed. It involved combining extracted pulse wave feature parameters with preliminary blood pressure values obtained through arterial tonometry. Regression analysis using various machine learning algorithms was performed to achieve accurate blood pressure measurement.

(3) The results from multiple experiments demonstrated that the proposed system improved the accuracy of systolic and diastolic blood pressure measurements by 31.4% and 21%, respectively, compared with those of the arterial tonometry method. Among the regression models used, the random forest algorithm exhibited the highest precision in predicting blood pressure, and the measured systolic and diastolic pressures met the widely accepted AAMI standards.

In summary, the proposed blood pressure monitoring system based on flexible encapsulated sensors shows potential for wearable blood pressure measurement and holds significant application prospects in the precise diagnosis and treatment of conditions such as hypertension.