Blood pressure measurement method and system based on fingertip photoplethysmography

Abstract

This paper proposes a blood pressure measurement method and system that predicts a subject's arterial blood pressure curve based on fingertip photoplethysmography signals. This system uses a cascaded UNet deep learning model to predict the arterial blood pressure curve and derive three blood pressure physiological indicators from the curve: systolic, diastolic, and mean blood pressure. This method achieves blood pressure measurement accuracy that meets the British Hypertension Society Class A and the Association for the Advancement of Medical Instrumentation (AAMI) standards. It enables noninvasive and continuous blood pressure measurement and facilitates wearable product design.

Description

Technical Field

[0001] The present invention relates to the technical field of blood pressure measurement, and in particular to a blood pressure measurement method and system based on fingertip photoplethysmography. Background Art

[0002] Currently, the common blood pressure measuring instruments on the market are mercury column blood pressure measuring instruments, electronic air pump blood pressure measuring instruments and wristbands with blood pressure detection functions.

[0003] Both of the first two methods use Korotkoff sound blood pressure measurement, a noninvasive, indirect method for measuring blood pressure. The general process involves inflating a cuff. When the pressure inside the cuff exceeds the systolic pressure of the arteries, the arteries close, preventing blood flow. The cuff is then slowly deflated at a rate of 2 to 3 mmHg / s. When the systolic pressure exceeds the pressure inside the cuff, the arteries partially open, and blood ejects, creating vortices or turbulence. This vibrates the blood vessels and is transmitted to the body surface as Korotkoff sounds. The pressure of the remaining air in the cuff at the time of the Korotkoff sounds of different phases is used as the blood pressure measurement.

[0004] When using the Korotkoff sound blood pressure measurement method, traditional mercury column blood pressure meters require medical staff to insert a stethoscope into the cuff, manually judge the different phases of the Korotkoff sounds, and record the reading of the mercury column in the meter when the Korotkoff sounds are produced as the measurement value of the human blood pressure.

[0005] The electronic air pump blood pressure meter is an improvement on the mercury column blood pressure meter. Korotkoff sounds are a type of sound that, when produced in blood vessels, transmit the sound waves to the body surface and the cuff in the form of vibrations. The electronic air pump blood pressure meter detects the sound waves of the propagating Korotkoff sounds and uses the oscillometric method to determine blood pressure by establishing a relationship between systolic pressure, diastolic pressure, mean pressure, and the cuff pressure oscillation wave. Compared to traditional mercury column blood pressure meters, electronic air pump blood pressure meters no longer require mercury to indicate the air pressure in the cuff. Instead, they use a pressure sensor to measure the air pressure in the cuff; the manually inflated airbag device is replaced by an electronic air pump. Blood pressure can be measured by oneself without the assistance of medical personnel.

[0006] In recent years, a new class of blood pressure monitoring electronic devices has emerged: wristbands that predict blood pressure based on wrist photoplethysmography signals. These devices also use a compact photoplethysmography sensor to collect photoplethysmographic signals from the wrist, extracting features that may be related to the user's blood pressure, and using these features to predict the user's blood pressure.

[0007] There are currently four methods for predicting blood pressure from photoplethysmography signals using deep learning methods.

[0008] First, there are methods for predicting blood pressure based on ECG (electrocardiogram) and PPG (plethysmography) signals. Continuous noninvasive blood pressure monitoring technology based on PPG and ECG information possesses significant advantages and is therefore crucial for noninvasive blood pressure monitoring research. In 2015, Kachuee et al. used ECG and PPG signals in their experiments, extracted relevant signal features, and established a blood pressure estimation model through machine learning. While their results were generally promising, the model's estimation accuracy for systolic blood pressure (SBP) was significantly lower than for diastolic blood pressure (DBP), failing to meet full certification standards. In recent years, researchers have used multi-cycle ECG and PPG signals to extract detailed features. Instead of manually constructing characteristic equations, they have adopted a channel-based attention mechanism with adaptive allocation. This has further improved the model's prediction accuracy, achieving SBP and DBP prediction accuracies of 4.70±3.45 mmHg and 2.40±2.99 mmHg, respectively.

[0009] Second, there are methods for predicting blood pressure based on pulse wave characteristics. Many researchers, when analyzing the relationship between PPG signals and blood pressure, have studied and established sets of pulse wave feature values ​​as an aid. By extracting characteristic points from the pulse wave that correlate with blood pressure, and based on measurement principles and relevant blood pressure theory, a characteristic equation is established to estimate blood pressure, enabling noninvasive and continuous blood pressure monitoring.

[0010] Combining the original PPG signal with its first- and second-order derivatives, we extracted the blood pressure-related eigenvalues ​​contained therein and constructed a blood pressure prediction model. The experimental results showed good accuracy. We selected time, amplitude, area, and heart rate as characteristic parameters for regression analysis, establishing eigenvalue relationships for both high and low pressure. This effectively achieved the experimental objectives and the results met AAMI standards.

[0011] Neural networks generally outperform traditional regression models in expressing complex nonlinear relationships. The relationship between blood pressure and PPG signals is not simply linear; it involves a variety of other factors. Therefore, models that utilize neural networks for simultaneous blood pressure prediction have begun to develop in recent years. For example, a method for continuous blood pressure estimation has been developed using deep neural networks. This method uses PTT, age, gender, and other human characteristics of the individual being measured, along with ECG signal features, as training data. This model combines a feedforward neural network with a deep neural network as an experimental model. Experimental results demonstrate that this model has relatively good prediction accuracy. Alternatively, by extracting features from PPG signals, six DBP feature parameters and eight SBP feature parameters were selected, and neural network models for DBP and SBP were developed, respectively. This enabled noninvasive continuous blood pressure measurement based on pulse wave feature parameters. However, the accuracy of the results did not fully meet the AAMI international standard, and further model optimization is needed.

[0012] 3. Then there is the method of predicting blood pressure based on pulse wave transit time and pulse wave velocity. A multivariate linear model was established by combining the time domain characteristics of the accelerated pulse wave, pulse transit time (PTT), and height parameters. The experimental results showed that DBP measurement was much better than the prediction results of SBP. Based on the traditional pulse wave transit time PTT algorithm, the advantages of soft threshold and hard threshold were combined, the noise preprocessing method was improved, the preprocessing steps were refined, and the final measurement results were more accurate. However, since the scope of application of this model is not comprehensive enough, it is not very suitable for patients with hypertension. FENG et al. changed the model of estimating blood pressure based solely on pulse transit time and combined variables such as pulse transit time and ECG signals to establish a model, which greatly improved the accuracy.

[0013] Finally, there are methods for predicting blood pressure based on PPG signals. For example, the PPG2ABP algorithm uses only PPG signals in experiments to adaptively calculate high-level abstract features through deep learning to build a blood pressure prediction model. The experimental results are highly accurate, but overall, the accuracy of SBP prediction is lower, requiring further research to improve SBP accuracy.

[0014] Both mercury-type and electronic-pump blood pressure meters require cuff pressure for measurement. This requires the cuff to apply steady pressure across its entire width, meaning it must not expand or shift during inflation, making it difficult to reduce its volume. Furthermore, whether manually inflated or using an electronic pump, the inflation device is relatively large. These two factors make it difficult for devices using the Korotkoff sound method to achieve optimal portability.

[0015] The blood pressure monitoring device based on photoplethysmography (PPG) signals integrated into the wristband is worn on the back of the user's wrist. Because the wrist muscles are well-developed and there is a lot of surface hair on the back of the wrist, the PPG signals collected are affected by the degree of wrist muscle contraction and the sparseness of the surface hair. Furthermore, the intensity of the PPG signals collected on the back of the wrist is also affected by the skin color of the wrist of different users. For these reasons, the error in measuring blood pressure using PPG signals collected at the wrist is high, and the data stability of multiple measurements is poor.

[0016] The four existing deep learning-based methods for predicting blood pressure all have varying degrees of shortcomings. Experimental results using ECG and PPG signals to predict blood pressure have shown high accuracy, but this requires additional equipment to measure the ECG signal, making measurement inconvenient. Experiments have shown that predicting blood pressure using pulse wave features is feasible. However, manually finding, extracting, and establishing a feature set for the waveform is challenging and requires high data standards, which may affect the accuracy of experimental results. Based on existing experimental conclusions, it is feasible to establish a blood pressure measurement model using only pulse wave time or pulse wave velocity. However, due to significant individual differences between different populations, the same method will exhibit deviations when measuring different populations. Therefore, it is necessary to incorporate individual features to establish a more universal blood pressure measurement model and achieve higher standards of accuracy. Existing deep learning models for predicting blood pressure based on PPG signals still need to be refined and their accuracy improved. Summary of the Invention

[0017] The purpose of the present invention is to address the defect that measuring instruments based on the Korotkoff sound blood pressure measurement method are difficult to achieve better portability. The present invention proposes a method and system for measuring human blood pressure values ​​using fingertip photoplethysmography signals. The system significantly reduces the size of the blood pressure measuring instrument and improves the portability of the blood pressure measuring instrument, thereby making it more convenient for subjects to measure their blood pressure.

[0018] The present invention proposes a blood pressure measurement method based on fingertip photoplethysmography, which includes:

[0019] Step 1: Collecting a photoelectric plethysmography signal at the fingertip position as an original signal through a photoelectric sensor, and performing mean trend line correction processing on the original signal to obtain a stable signal;

[0020] Step 2: After mapping and spline interpolation sampling, the stable signal is mapped into a compliant signal with the same sampling threshold and sampling frequency as the deep learning model training data. The compliant signal is input into the deep learning model to generate an arterial blood pressure curve corresponding to the original signal;

[0021] Step 3: Take the maximum and minimum blood pressure values ​​from the arterial blood pressure curve as the measured values ​​of the user's systolic and diastolic blood pressures, and take the average of the entire blood pressure curve as the measured value of the user's average blood pressure.

[0022] The blood pressure measurement method based on fingertip photoplethysmography, wherein the deep learning model is a cascaded UNet deep learning model, uses U-Net and MultiResUNet deep neural networks for cascade splicing, both use mean absolute error and mean square error as loss functions, respectively, use medical data sets as training data, use Adam optimizer to train the two neural networks, and train the model parameters until convergence.

[0023] In the blood pressure measurement method based on fingertip photoplethysmography, the mean trend line correction process in step 1 includes:

[0024] The window length set for mean smoothing is w, and the mean trend line correction is performed using the following formula:

[0025]

[0026] Where x' i The data in the array index i after the original signal is smoothed, and the x on the right side of the summation symbol i+j is the original signal, and its subscript i+j is the array subscript of the original signal of the array;

[0027] And the stable signal y is obtained by the following formula: i :

[0028]

[0029] y i =x' i -bi.

[0030] The blood pressure measurement method based on fingertip photoplethysmography, wherein step 3 comprises:

[0031] The maximum blood pressure value x is taken from the arterial blood pressure curve max As the measured value of the systolic pressure, the minimum value in the graph between two adjacent maximum values ​​is taken as the measured value of the diastolic pressure.

[0032] The present invention also proposes a blood pressure measurement system based on fingertip photoplethysmography, which includes:

[0033] An initial module is used to collect the photoplethysmography signal at the fingertip position through a photoelectric sensor as the original signal, and perform mean trend line correction processing on the original signal to obtain a stable signal;

[0034] A mapping module is configured to map the stable signal into a compliant signal having the same sampling threshold and sampling frequency as the deep learning model training data after the stable signal undergoes mapping processing and spline interpolation sampling. The compliant signal is input into the deep learning model to generate an arterial blood pressure curve corresponding to the original signal;

[0035] The output module is used to take the maximum and minimum blood pressure values ​​from the arterial blood pressure curve as the measured values ​​of the user's systolic and diastolic blood pressure, and take the average of the entire blood pressure curve as the measured value of the user's average blood pressure.

[0036] The blood pressure measurement system based on fingertip photoplethysmography, wherein the deep learning model is a cascaded UNet deep learning model, uses U-Net and MultiResUNet deep neural networks for cascade splicing, both of which use mean absolute error and mean square error as loss functions, respectively, use medical data sets as training data, and use Adam optimizer to train the two neural networks, and train the model parameters until convergence.

[0037] In the blood pressure measurement system based on fingertip photoplethysmography, the mean trend line correction process in the initial module includes:

[0038] The window length set for mean smoothing is w, and the mean trend line correction is performed using the following formula:

[0039]

[0040] Where x' i The data in the array index i after the original signal is smoothed, and the x on the right side of the summation symbol i+j is the original signal, and its subscript i+j is the array subscript of the original signal of the array;

[0041] And the stable signal y is obtained by the following formula: i :

[0042]

[0043] y i =x' i -bi.

[0044] In the blood pressure measurement method based on fingertip photoplethysmography, the output module is used to:

[0045] The maximum blood pressure value x is taken from the arterial blood pressure curve max As the measured value of the systolic pressure, the minimum value in the graph between two adjacent maximum values ​​is taken as the measured value of the diastolic pressure.

[0046] The present invention also proposes a storage medium for a program of any one of the blood pressure measurement methods based on fingertip photoplethysmography.

[0047] The present invention also proposes a client for use in any of the aforementioned blood pressure measurement systems based on fingertip photoplethysmography.

[0048] According to the British Hypertension Society standards, the cascaded UNet deep learning model used in the blood pressure monitor of the present invention has reached the Grade A medical device standard level, as shown in the table below:

[0049] <=5mm Hg <=10mm Hg <=15mm Hg Grade A requirements 60％ 85％ 95％ DBP diastolic blood pressure 81.4％ 92.4％ 96.2％ SBP systolic blood pressure 69.6％ 86.1％ 95.3％ MAP mean pressure 95.5％ 97.6％ 98.4％

[0050] According to the American Association for the Advancement of Medical Instrumentation standard, the cascaded UNet deep learning model used in the blood pressure monitor of the present invention has reached the medical device standard level, as shown in the following table:

[0051] MAE (mmHg) STD (mmHg) AAMI Medical Supplies Standard <＝5 <＝8 DBP diastolic blood pressure 3.14 6.83 SBP systolic blood pressure 4.75 7.24 MAP mean pressure 2.17 4.48

[0052] To address the shortcomings of existing technologies, this paper proposes a blood pressure measurement method and system that predicts a subject's arterial blood pressure curve based on fingertip photoplethysmography signals. This system uses a cascaded UNet deep learning model to predict the arterial blood pressure curve and derive three blood pressure physiological indicators: systolic, diastolic, and mean blood pressure.

[0053] From the above scheme, it can be seen that the advantages of the present invention are:

[0054] Non-invasive, continuous blood pressure measurement; can predict the arterial blood pressure curve and dynamically output beat-by-beat blood pressure values: systolic, diastolic, and mean pressure; high measurement accuracy, meeting British Hypertension Society Class A and American Association for the Advancement of Medical Instrumentation standards; and can realize wearable product design. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 The overall process of measuring blood pressure using a blood pressure measurement device based on fingertip photoplethysmography is shown in the figure;

[0056] Figure 2 This is a schematic diagram of signal correction;

[0057] Figure 3A and Figure 3B This is a neural network structure diagram of the cascaded UNet deep learning model;

[0058] Figure 4 A graph showing the measurement results. DETAILED DESCRIPTION

[0059] To address the limitations of Korotkoff sound-based blood pressure measurement devices, which lack portability, and the low accuracy of wristbands based on photoplethysmography (PPG) signals, this paper proposes a method and system for measuring blood pressure using fingertip PPG signals. Based on extensive medical data, this system employs deep learning methods to construct and restore a human arterial blood pressure curve using PPG signals, thereby measuring blood pressure.

[0060] Specifically, first, the photoplethysmography sensor is very small and can be easily inserted into devices such as fingertip measuring instruments. In addition, the fingertips have less muscle mass than the wrists, no body hair, and more stable skin color, so more stable and accurate photoplethysmographic signals can be collected here. Secondly, the photoplethysmographic signal contains characteristics that can reflect the changes in the arterial blood pressure curve, that is, the two have sufficient correlation, making it possible to predict the arterial blood pressure curve. Finally, by introducing the cascade UNet deep learning model, the model can automatically learn the correlation between the photoplethysmographic signal and the arterial blood pressure curve, and accurate blood pressure measurement can be achieved based on a large amount of medical measurement data.

[0061] Specifically, the present invention includes the following key technical points:

[0062] Key Point 1: Technology for converting raw fingertip photoplethysmography signals into standard input signals usable by deep learning models. Technical Effect: The original signal with interference noise can be converted into a smooth and stable standard signal.

[0063] Key Point 2: Using a cascaded UNet deep learning model to restore the human arterial blood pressure curve from fingertip photoplethysmography signals; technical performance: meets the British Hypertension Society's Class A standard and the American Association for the Advancement of Medical Instrumentation standard;

[0064] Key point 3: Extracting the measured values ​​of systolic and diastolic blood pressure from the arterial blood pressure curve; Technical effect: Based on the arterial blood pressure curve, the measured values ​​of systolic and diastolic blood pressure can be accurately extracted from the curve.

[0065] In order to make the above features and effects of the present invention more clearly understood, embodiments are given below and described in detail with reference to the accompanying drawings.

[0066] The overall process of measuring blood pressure using a blood pressure measurement device based on fingertip photoplethysmography is as follows: Figure 1 shown.

[0067] First, aim the sensor at the fingertip. The photoplethysmographic signal is collected from the fingertip photoelectric sensor and sent back to the phone via Bluetooth. The sensor's raw photoplethysmographic signal (initial data) needs to be free of sensor noise. The sensor needs to be placed on the fingertip for a period of time. After collecting sufficient initial data, the latest data is corrected using the initial data as a baseline, applying a mean trendline correction to stabilize the signal.

[0068] After normalization (e.g., minimum and maximum normalization), mapping, and spline interpolation sampling, the signal is further mapped into a photoplethysmographic signal with the same sampling threshold and frequency as the deep learning model training data. This converted data is then reconstructed through a cascaded UNet deep learning model to reconstruct the user's current arterial blood pressure curve.

[0069] Then, the maximum and minimum blood pressure values ​​are taken from the arterial blood pressure curve as the measured values ​​of the user's systolic and diastolic blood pressure, and the average of the entire blood pressure curve is taken as the measured value of the user's average blood pressure.

[0070] Figure 2 The original signal is corrected into a stable and straight repair signal using mean smoothing and mean trend line correction. Mean smoothing can filter out high-frequency noise while retaining valid data information, thereby eliminating high-frequency noise signals and obtaining a smooth and effective fingertip photoplethysmography signal. Assuming the window length set during mean smoothing is w, the mean smoothing process follows the following formula:

[0071]

[0072] Where x' i The data in the array index i after smoothing, the x on the right side of the summation symbol i+j is the original signal, and its subscript i+j is the array subscript of the original signal of the array.

[0073] Mean trendline correction removes sensor tilt interference from the graph, resulting in a stable and flat signal. The mean value window must be larger than the user's entire heartbeat cycle. Based on this, the present invention provides recommended window values, assuming a normal human heartbeat frequency of between 0.3 and 1 second. When the sensor sampling frequency is 125Hz, a maximum of 100 data points is recommended for the front and back mean sampling window. The formula for obtaining the mean trendline is the same as that for mean smoothing.

[0074] Based on the obtained mean trend line, the least squares regression method is used to obtain the offset straight line equation that needs to be corrected for the signal. This is done by subtracting the deviation value that needs to be corrected at the current moment from the original signal, ultimately obtaining a stable and flat fingertip photoplethysmography signal. The least squares regression method formula and signal correction formula are as follows:

[0075]

[0076] y i =x' i -bi in the formula y' i is the corrected signal, y i is the original signal data, x i is the array subscript, and m is the length of the signal array to be corrected.

[0077] Figure 3 shows the neural network structure of the cascade UNet deep learning model. The cascade UNet deep model of the present invention uses two deep neural networks, U-Net and MultiResUNet, for cascade splicing, and the mean absolute error (MAE) and mean square error (MSE) are used as loss functions respectively. These two models can be trained step by step and cascaded during prediction. The two models are used because when the two models are stacked, the parameters of the model increase, the model capacity increases, more features can be learned on large-scale data sets, and the learning effect is better. The first model uses UNet with a relatively simple structure, which can roughly fit the curve. The second complex model in the stack is to restore the arterial blood pressure curve as accurately as possible based on the prediction results of the first model. With the restoration of the first model as a foundation, the second model can restore the curve more accurately.

[0078] Using the Cuff-Less Blood Pressure Estimation Data Set, a medical dataset publicly available from the University of California, Irvine, as training data, we trained two neural networks using the Adam optimizer until the model parameters converged. The converged cascade UNet deep model can reconstruct the arterial blood pressure curve based on the fingertip photoplethysmography signal. This neural network is denoted by the transformation function The expression of the arterial blood pressure curve (ABP) obtained from the fingertip photoplethysmography signal (PPG) is as follows:

[0079]

[0080] Figure 4 This article demonstrates how to extract systolic and diastolic blood pressure measurements from an arterial blood pressure curve. The first maximum value to the right of the maximum rising slope point in each cycle is used as the systolic blood pressure measurement. The maximum value is taken using the first-order derivative extreme value measurement method, as shown in the following formula:

[0081]

[0082] Among them, abp max is the maximum value of the entire arterial blood pressure curve, abp min is the minimum value of the entire arterial blood pressure curve, abp i is the arterial blood pressure value at index i in the array. The systolic blood pressure value is the first data point in the arterial blood pressure curve that satisfies the above formula to the right of the point with the maximum upward slope within the cycle. Multiple systolic blood pressure measurements meeting these requirements can be obtained from a single curve. These systolic blood pressure measurements are averaged to represent the subject's observed systolic blood pressure for the entire measurement process.

[0083] After obtaining multiple systolic blood pressure measurements, the minimum value between adjacent systolic blood pressure measurements is taken as the diastolic blood pressure measurement. These diastolic blood pressure measurements are averaged as the subject's diastolic blood pressure observation value during the entire measurement process.

[0084] Finally, take the mean abp of the entire arterial blood pressure curve mean As the mean pressure measurement value and the observed value of the subject's mean pressure during the entire measurement process.

[0085] The following is a system embodiment corresponding to the above method embodiment. This embodiment can be implemented in conjunction with the above embodiment. The relevant technical details mentioned in the above embodiment are still valid in this embodiment and will not be repeated here to reduce repetition. Accordingly, the relevant technical details mentioned in this embodiment can also be applied to the above embodiment.

[0086] The present invention also proposes a blood pressure measurement system based on fingertip photoplethysmography, which includes:

[0087] An initial module is used to collect the photoplethysmography signal at the fingertip position through a photoelectric sensor as the original signal, and perform mean trend line correction processing on the original signal to obtain a stable signal;

[0088] A mapping module is configured to map the stable signal into a compliant signal having the same sampling threshold and sampling frequency as the deep learning model training data after the stable signal undergoes mapping processing and spline interpolation sampling. The compliant signal is input into the deep learning model to generate an arterial blood pressure curve corresponding to the original signal;

[0089] The output module is used to take the maximum and minimum blood pressure values ​​from the arterial blood pressure curve as the measured values ​​of the user's systolic and diastolic blood pressure, and take the average of the entire blood pressure curve as the measured value of the user's average blood pressure.

[0090] The blood pressure measurement system based on fingertip photoplethysmography, wherein the deep learning model is a cascaded UNet deep learning model, uses U-Net and MultiResUNet deep neural networks for cascade splicing, both of which use mean absolute error and mean square error as loss functions, respectively, use medical data sets as training data, and use Adam optimizer to train the two neural networks, and train the model parameters until convergence.

[0091] In the blood pressure measurement system based on fingertip photoplethysmography, the mean trend line correction process in the initial module includes:

[0092] The window length set for mean smoothing is w, and the mean trend line correction is performed using the following formula:

[0093]

[0094] Where x' i The data in the array index i after the original signal is smoothed, and the x on the right side of the summation symbol i+j is the original signal, and its subscript i+j is the array subscript of the original signal of the array;

[0095] And the stable signal y is obtained by the following formula: i :

[0096]

[0097] y i =x' i -bi said blood pressure measurement method based on fingertip photoplethysmography, wherein the output module is used for:

[0098] The maximum blood pressure value x is taken from the arterial blood pressure curve max As the measured value of the systolic pressure, the minimum value in the graph between two adjacent maximum values ​​is taken as the measured value of the diastolic pressure.

[0099] The present invention also proposes a storage medium for a program of any one of the blood pressure measurement methods based on fingertip photoplethysmography.

[0100] The present invention also proposes a client for use in any of the aforementioned blood pressure measurement systems based on fingertip photoplethysmography.

Claims

1. A blood pressure measurement method based on fingertip photoplethysmography, characterized in that: include: Step 1: Collecting a photoelectric plethysmography signal at the fingertip position as an original signal through a photoelectric sensor, and performing mean trend line correction processing on the original signal to obtain a stable signal; Step 2: After mapping and spline interpolation sampling, the stable signal is mapped into a compliant signal with the same sampling threshold and sampling frequency as the deep learning model training data. The compliant signal is input into the deep learning model to generate an arterial blood pressure curve corresponding to the original signal; Step 3: taking the maximum and minimum blood pressure values ​​from the arterial blood pressure curve as the user's systolic and diastolic blood pressure measurements, and taking the average of the entire blood pressure curve as the user's average blood pressure measurement; The mean trend line correction process in step 1 includes: The window length set for mean smoothing is w, and the mean trend line correction is performed using the following formula: Where x' i The data in the array index i after the original signal is smoothed, and the x on the right side of the summation symbol i+j is the original signal, and its subscript i+j is the array subscript of the original signal of the array; And the stable signal y is obtained by the following formula: i : y i =x' i -bi; Where m is the length of the signal array to be corrected, and bi is the deviation value that needs to be corrected at the current moment.

2. The blood pressure measurement method based on fingertip photoplethysmography according to claim 1, wherein: This deep learning model is a cascaded UNet deep learning model, which uses U-Net and MultiResUNet deep neural networks for cascade splicing. The mean absolute error and mean squared error are used as loss functions, respectively, and the medical dataset is used as training data. The Adam optimizer is used to train the two neural networks and train the model parameters until convergence.

3. The blood pressure measurement method based on fingertip photoplethysmography according to claim 1, wherein: This step 3 includes: The first maximum value on the right side of the maximum rising slope point in each cycle of the arterial blood pressure curve is taken as the systolic pressure measurement value, and the minimum value in the graph between adjacent systolic pressure measurement values ​​is taken as the diastolic pressure measurement value.

4. A blood pressure measurement system based on fingertip photoplethysmography, characterized in that: include: An initial module is used to collect the photoplethysmography signal at the fingertip position through a photoelectric sensor as the original signal, and perform mean trend line correction processing on the original signal to obtain a stable signal; A mapping module is configured to map the stable signal into a compliant signal having the same sampling threshold and sampling frequency as the deep learning model training data after the stable signal undergoes mapping processing and spline interpolation sampling. The compliant signal is input into the deep learning model to generate an arterial blood pressure curve corresponding to the original signal; an output module, configured to take the maximum and minimum blood pressure values ​​from the arterial blood pressure curve as the measured values ​​of the user's systolic and diastolic blood pressures, respectively, and take the average of the entire blood pressure curve as the measured value of the user's average blood pressure; The mean trend line correction process in this initial module includes: The window length set for mean smoothing is w, and the mean trend line correction is performed using the following formula: Where x' i The data in the array index i after the original signal is smoothed, and the x on the right side of the summation symbol i+j is the original signal, and its subscript i+j is the array subscript of the original signal of the array; And the stable signal y is obtained by the following formula: i : y i =x' i -bi; Where m is the length of the signal array to be corrected, and bi is the deviation value that needs to be corrected at the current moment.

5. The blood pressure measurement system based on fingertip photoplethysmography according to claim 4, wherein: This deep learning model is a cascaded UNet deep learning model, which uses U-Net and MultiResUNet deep neural networks for cascade splicing. The mean absolute error and mean squared error are used as loss functions, respectively, and the medical dataset is used as training data. The Adam optimizer is used to train the two neural networks and train the model parameters until convergence.

6. The blood pressure measurement method based on fingertip photoplethysmography according to claim 4, wherein: The output module is used to: The maximum blood pressure value x is taken from the arterial blood pressure curve max As the measured value of the systolic pressure, the minimum value in the graph between two adjacent maximum values ​​is taken as the measured value of the diastolic pressure.

7. A storage medium for storing a program for executing any one of the blood pressure measurement methods based on fingertip photoplethysmography according to claims 1 to 3.

8. A client, used in any one of the fingertip photoplethysmography-based blood pressure measurement systems according to claims 4 to 6.

Images