Non-invasive blood pressure monitoring system and method based on pulse waves

Abstract

The invention discloses a non-invasive blood pressure monitoring system and method.The non-invasive blood pressure monitoring system comprises a pulse wave collecting module, an electrocardio collecting module and a control module.The pulse wave collecting module is used for collecting pulse wave signals; the electrocardio acquisition module is used for acquiring electrocardio signals; the control module is used for obtaining blood pressure according to the pulse wave signals and the electrocardiosignals and based on a built-in blood pressure prediction model. The non-invasive blood pressure monitoring system can collect pulse waves and electrocardiosignals to predict blood pressure, the defects that traditional blood pressure measuring equipment is inconvenient to wear and poor in real-time performance can be overcome, and meanwhile the problem that the error is large when the blood pressure is predicted through the pulse waves is solved. The invention further discloses a non-invasive blood pressure monitoring method based on the pulse waves.

Description

Technical Field

[0001] The present invention belongs to the field of medical detection, and in particular relates to a pulse wave-based non-invasive blood pressure monitoring system and method. Background Art

[0002] Blood pressure is a crucial biomedical signal, providing crucial physiological information related to cardiac function, vascular status, organ perfusion, and hemodynamics, and reflecting the health of the cardiovascular system. Hypertension is a major contributing factor to various cardiovascular diseases, while hypotension can result in insufficient blood pressure supply to various parts of the body, posing serious risks to patients, particularly during medical interventions. Therefore, real-time blood pressure monitoring is crucial for early detection of potential threats and timely intervention.

[0003] Currently, the traditional method for regular blood pressure measurement involves inflating and deflating a cuff to stop blood flow, and then judging blood pressure by the sound of blood flow or pressure changes. This method is medically known as cuff blood pressure measurement. However, the traditional cuff blood pressure measurement method has the disadvantages of being inconvenient to wear and unable to predict blood pressure in real time, making it unsuitable for early detection and diagnosis of hypertension. While invasive arterial cannulation can achieve continuous monitoring of blood pressure, its inherent invasive nature imposes significant limitations: on the one hand, the puncture operation may increase the risk of infection and may also cause complications such as vascular damage and thrombosis; on the other hand, this invasive arterial cannulation method relies on professional medical staff for operation, and the equipment and consumables are expensive. Therefore, it is only suitable for special medical scenarios such as operating rooms and intensive care units (ICUs) for real-time monitoring of critically ill patients, and is completely unable to meet the blood pressure measurement needs of daily life.

[0004] In recent years, with the rapid development of wearable sensing technology and the growing demand for continuous blood pressure monitoring, cuffless, non-invasive blood pressure measurement methods have become a research focus in this field. Among them, photoplethysmography (PPG) signals are widely recognized as the most promising non-invasive continuous blood pressure monitoring solution due to their suitability for continuous monitoring of multiple physiological indicators, including blood pressure, oxygen saturation, heart rate, respiration, and blood glucose. PPG technology uses a photoelectric device to detect changes in tissue blood volume without intruding the human body. Its measurement principle is based on the Lambert-Beer law, which states that light attenuates as it passes through blood due to factors such as path length, tissue density, and light absorption characteristics. As blood is pumped out and returned with the heartbeat, blood volume exhibits periodic fluctuations. The PPG waveform captures these fluctuations, providing a crucial basis for assessing a subject's cardiovascular function. It not only reflects the dynamic changes in blood volume and blood flow velocity, but also indirectly reflects the overall state of the cardiovascular system. Early studies focused on manually extracting features from PPG signals and then combining them with machine learning methods to predict blood pressure. However, the accuracy of such methods heavily relied on the researchers' experience and often produced poor results. With the rise of deep learning technology, "end-to-end" blood pressure prediction models have gradually become mainstream: without the need for manual feature extraction, blood pressure can be output directly using the PPG sequence as input, significantly improving the automation and accuracy of monitoring. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the present invention provides a non-invasive blood pressure monitoring system based on pulse waves. The non-invasive blood pressure monitoring system can collect pulse waves and electrocardiogram signals to predict blood pressure, and can overcome the shortcomings of traditional blood pressure measurement equipment that are inconvenient to wear and have poor real-time performance, while solving the problem of large errors in predicting blood pressure through pulse waves.

[0006] The second object of the present invention is to provide a non-invasive blood pressure monitoring method based on pulse wave.

[0007] The technical solution of the present invention to solve the above technical problems is:

[0008] A pulse wave-based non-invasive blood pressure monitoring system includes a pulse wave acquisition module, an electrocardiogram acquisition module, and a control module, wherein the pulse wave acquisition module is used to acquire pulse wave signals; the electrocardiogram acquisition module is used to acquire electrocardiogram signals; and the control module is used to obtain blood pressure based on the pulse wave signal and the electrocardiogram signal and its built-in blood pressure prediction model.

[0009] Preferably, the control module is a single chip microcomputer, and its model is STM32F407.

[0010] Preferably, the pulse wave acquisition module is a pulse wave acquisition chip, and its model is MAX30101.

[0011] Preferably, the ECG acquisition module is an ECG acquisition chip, model ADS1292R.

[0012] Preferably, the control module communicates with the pulse wave acquisition module through the I2C protocol to control the working mode of the pulse wave acquisition module and receive the pulse wave signal returned by the pulse wave acquisition module, and store the received pulse wave data into the SD card.

[0013] Preferably, the control module communicates with the ECG acquisition module via the SPI protocol, thereby controlling the ECG acquisition module to acquire two-lead ECG signals and receive ECG signals returned by the ECG acquisition module, and storing the ECG data in an SD card.

[0014] A non-invasive blood pressure monitoring method based on pulse wave comprises the following steps:

[0015] Step 1: Collect PPG signals through the pulse wave acquisition module, collect ECG signals through the electrocardiogram acquisition module, and collect the corresponding blood pressure to construct a data set;

[0016] Step 2: Preprocess the data in the dataset;

[0017] Step 3: Divide the preprocessed dataset into training set, test set and validation set;

[0018] Step 4: Construct a blood pressure prediction model and train it using the training set. Simultaneously record the training set loss and validation set loss after each round of training. Select the number of training rounds that results in the minimum validation set loss within the training cycle, and use the weight parameters of the blood pressure prediction model at the end of that round as the final weights. After loading the final weights into the blood pressure prediction model, input the test set for prediction. Quantify the model's blood pressure prediction performance using preset evaluation indicators, and select a blood pressure prediction model that meets the accuracy requirements.

[0019] Step 5: The PPG signal collected by the pulse wave acquisition module and the ECG signal collected by the electrocardiogram acquisition module are sent to the trained blood pressure prediction model to obtain the blood pressure.

[0020] Preferably, in step 1, the PPG signal within 3 minutes is collected by the pulse wave acquisition module, and the ECG signal within 3 minutes is collected by the electrocardiogram acquisition module; the systolic blood pressure SBP and the diastolic blood pressure DBP are measured using a cuff sphygmomanometer, and the systolic blood pressure SBP and the diastolic blood pressure DBP are used as reference blood pressure in the subsequent 3 minutes.

[0021] Preferably, in step 2, the preprocessing process is: the collected PPG signal and ECG signal are filtered through a 0.5 Hz-8 Hz Butterworth bandpass filter, normalized, and the normalized signal is cut into 10 s windows to obtain multiple signal segments.

[0022] Preferably, in step 4, the data processing process of the blood pressure prediction model is:

[0023] Inputting the PPG signal into a first MultiResUnet network and the ECG signal into a second MultiResUnet network, reconstructing the ECG signal through the first MultiResUnet network, and reconstructing the PPG signal through the second MultiResUnet network;

[0024] Get the output of the first downsampling layer, the output of the bottom layer, and the input before the last upsampling of the first MultiResUnet network, and the output of the first downsampling layer, the output of the bottom layer, and the input before the last upsampling of the second MultiResUnet network respectively;

[0025] The obtained outputs are spliced ​​together and depth-wise separable convolution is performed on the spliced ​​results;

[0026] The results of the depthwise separable convolution are spliced ​​again, and after one layer of convolution, the predicted systolic blood pressure SBP and diastolic blood pressure DBP are obtained.

[0027] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0028] 1. The pulse wave-based non-invasive blood pressure monitoring system of the present invention can collect the pulse wave of the fingertips and the two-lead ECG signal, and transmit the above-mentioned pulse wave and two-lead ECG signal to the control device. The control device obtains the actual blood pressure based on its built-in blood pressure prediction model.

[0029] 2. The pulse wave-based non-invasive blood pressure monitoring system of the present invention can overcome the shortcomings of traditional cuff blood pressure monitors, which are inconvenient to wear and have poor real-time performance, so as to facilitate real-time monitoring of blood pressure.

[0030] 3. The pulse wave-based non-invasive blood pressure monitoring method of the present invention uses PPG signals and ECG signals as multimodal inputs and constructs a blood pressure prediction model to reduce the error of blood pressure prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Schematic diagram of the flow of the pulse wave-based non-invasive blood pressure monitoring method of the present invention.

[0032] Figure 2 This is the network structure diagram of the blood pressure prediction model.

[0033] Figure 3 This is the structural diagram of the coding layer.

[0034] Figure 4 This is the structural diagram of the residual block.

[0035] Figure 5 This is the loss curve for one-fold cross-validation training. DETAILED DESCRIPTION

[0036] The present invention will be described in further detail below with reference to the embodiments and drawings, but the embodiments of the present invention are not limited thereto.

[0037] The pulse wave-based non-invasive blood pressure monitoring system of the present invention includes a pulse wave acquisition module, an electrocardiogram acquisition module and a control module, wherein:

[0038] The pulse wave acquisition module is used to acquire pulse wave (PPG) signals. The pulse wave acquisition module is a pulse wave acquisition chip, model MAX30101. The MAX30101 is a heart rate and blood oxygen sensor that integrates an optical sensor and a signal processor. It is widely used in medical fields such as heart rate monitoring, pulse oxygen saturation monitoring, and other biological parameter detection. It has the characteristics of high integration, low power consumption, and high precision, and can detect heart rate and blood oxygen saturation in real time. The heart rate and blood oxygen sensor integrates a green LED, an infrared LED, a red light LED, and a photodetector. The LED illuminates the skin, and the photodetector collects the reflected light signal, thereby collecting the pulse wave.

[0039] The ECG acquisition module is used to acquire ECG signals. The ECG acquisition module is an ECG acquisition chip, model ADS1292R. The ADS1292R is a high-performance analog front-end chip designed by Texas Instruments specifically for biopotential measurement. It integrates a 2-channel 24-bit delta-sigma analog-to-digital converter (ADC), a programmable gain amplifier (PGA), an internal reference voltage source, and an onboard oscillator. Its core function is to acquire ECG and respiratory signals with high precision, making it particularly suitable for portable medical devices and wearable health monitoring scenarios.

[0040] The control module is used to obtain blood pressure according to the pulse wave signal and the electrocardiogram signal based on its built-in blood pressure prediction model. The control module is a single-chip microcomputer, and its model is STM32F407.

[0041] In this embodiment, the control module communicates with the pulse wave acquisition module through the I2C protocol to control the working mode of the pulse wave acquisition module and receive the pulse wave signal returned by the pulse wave acquisition module, and stores the received pulse wave data in the SD card, wherein the sampling rate is 125HZ; at the same time, the control module communicates with the ECG acquisition module through the SPI protocol to control the ECG acquisition module to acquire two-lead ECG signals and receive the ECG signals returned by the ECG acquisition module, and store the ECG data in the SD card, wherein the sampling rate is 125HZ.

[0042] See also Figure 1-Figure 5 The non-invasive blood pressure monitoring method based on pulse wave of the present invention comprises the following steps:

[0043] Step 1: Collect PPG signals through the pulse wave acquisition module, collect ECG signals through the electrocardiogram acquisition module, and collect the corresponding blood pressure to construct a data set;

[0044] In this embodiment, the PPG signal within 3 minutes is collected by the pulse wave acquisition module, and the ECG signal within 3 minutes is collected by the electrocardiogram acquisition module; since the blood pressure change within a period of time will not be large, the systolic pressure SBP and diastolic pressure DBP can be measured using a cuff sphygmomanometer, and the systolic pressure SBP and diastolic pressure DBP are used as the reference blood pressure within the subsequent 3 minutes.

[0045] Step 2: Preprocess the data in the dataset, specifically:

[0046] The collected PPG and ECG signals were filtered through a 0.5-8 Hz Butterworth bandpass filter and then normalized. The normalized signals were cut into 10-s windows to obtain multiple signal segments.

[0047] Step 3: Divide the preprocessed dataset into training set, test set and validation set;

[0048] In this embodiment, the obtained data set is divided into training set, validation set and test set according to the ratio of 8:1:1, and within the training set divided by the above basic division, ten-fold cross-validation is further used to optimize model training, that is, the training set is randomly divided into 10 subsets (folds), and 9 folds are selected as "sub-training sets" each time, and the remaining 1 fold is used as "sub-validation set"; the training is repeated 10 times, and the number of folds of the sub-validation set is rotated each time. Finally, the average performance of the 10 trainings is taken as the comprehensive performance of the model in the training stage. This can reduce the impact of sample distribution deviation caused by a single division on the model, which is particularly suitable for scenarios with limited sample size or uneven distribution.

[0049] Step 4: Construct a blood pressure prediction model and train the constructed blood pressure prediction model using the training set. Simultaneously record the training set loss value and the validation set loss value after each round of training. Select the number of training rounds corresponding to the minimum validation set loss value within the training cycle, and use the weight parameters of the blood pressure prediction model at the end of the round as the final weights. After loading the final weights into the blood pressure prediction model, input the test set for prediction. Quantify the blood pressure prediction performance of the model using preset evaluation indicators (such as MAE and RMSE) to screen out a blood pressure prediction model that meets the accuracy requirements.

[0050] In this embodiment, the blood pressure prediction model inputs the PPG signal into the first MultiResUnet network and the ECG signal into the second MultiResUnet network; reconstructs the ECG signal through the first MultiResUnet network and reconstructs the PPG signal through the second MultiResUnet network; then obtains the output of the first downsampling layer, the output of the bottom layer, and the input before the last upsampling of the first MultiResUnet network, and the output of the first downsampling layer, the output of the bottom layer, and the input before the last upsampling of the second MultiResUnet network; then splices the obtained outputs and performs depthwise separable convolution on the spliced ​​results; finally, the results after the depthwise separable convolution are spliced ​​again, and after one layer of convolution, the predicted systolic blood pressure SBP and diastolic blood pressure DBP are obtained.

[0051] The first MultiResUnet network and the second MultiResUnet network have the same network structure, both including a coding layer, a decoding layer and a skip layer, wherein:

[0052] See also Figure 3 , the coding layer ( Figure 2 The red square in the figure adopts the residual convolutional network structure, whose core function is to solve the degradation problem in deep neural network training: In theory, deep neural networks can improve feature expression capabilities by increasing the number of layers, but in actual training, increasing the number of layers may cause the gradient to disappear or explode, which in turn increases the training error; the residual network constructs residual blocks by introducing identity mapping, so that input information can be directly transmitted across layers (residual connection), thereby effectively alleviating the gradient propagation problem, ensuring that the network can still be stably optimized as it deepens, and accelerating the training convergence speed; the encoding layer achieves downsampling by cascading multiple groups of residual blocks, thereby gradually extracting high-level semantic features of the input data, providing basic feature support for the subsequent decoding layer;

[0053] The decoding layer ( Figure 2 The green square in the figure also uses the residual convolutional network structure. The core difference from the encoding layer is:

[0054] (1) The decoding layer adds upsampling operations (such as transposed convolution or interpolation upsampling) based on the residual convolution operation, gradually restoring the spatial resolution of the feature map and realizing the mapping from abstract features to specific details;

[0055] (2) The reuse of the residual structure ensures that feature information is not easily lost during the decoding process, while upsampling provides a basis for scale matching for subsequent feature fusion with the coding layer.

[0056] See also Figure 4 , the jump layer ( Figure 2 The arrow between the red and green squares in the figure is the key bridge connecting the encoding layer and the decoding layer. It adopts a multi-scale convolution design: multi-receptive field features are extracted from the corresponding levels of the encoding layer through convolution kernels of different sizes (such as 1×1, 3×3, and 5×5), capturing the scale changes, shape details, and local texture information of the target; the multi-scale features are transmitted to the decoding layer via skip connections and fused with the features in the decoding process to compensate for the loss of details caused by downsampling and enhance the model's perception of target diversity (such as scale and morphological changes).

[0057] In addition to the first MultiResUnet network and the second MultiResUnet network, the blood pressure prediction model in this embodiment also includes a multimodal connection layer and an output layer, wherein:

[0058] The multimodal connection layer ( Figure 2 The purple square in the figure uses depthwise separable convolution to achieve cross-modal / cross-level feature fusion. Its structure is divided into two stages:

[0059] Depthwise convolution: Use a single convolution kernel to perform convolution on each input channel independently, extract spatial features only within the channel, and do not perform inter-channel fusion, which can significantly reduce the amount of calculation;

[0060] Pointwise convolution: The output of depthwise convolution is fused in the channel dimension through a 1×1 convolution kernel, integrating the feature information of different channels to generate the final fused feature;

[0061] The above design can significantly reduce the scale of network parameters and improve computational efficiency while ensuring the effect of feature fusion.

[0062] The output layer ( Figure 2 The yellow square in the figure is used to concatenate the fused features output by the multimodal connection layer, complete the dimension mapping through the 1×1 convolution layer, and finally output the target prediction value (i.e., blood pressure sequence). This output layer simplifies the convolution operation (without spatial receptive field expansion) and focuses on the final dimensional adaptation of the features to ensure that the output is consistent with the task target (such as the label dimension).

[0063] Step 5: The PPG signal collected by the pulse wave acquisition module and the ECG signal collected by the electrocardiogram acquisition module are sent to the trained blood pressure prediction model to obtain the blood pressure.

[0064] Figure 5 The figure shows the loss curve of one-fold cross-validation training. It can be seen from the figure that the final error of the blood pressure prediction model is: the mean absolute error (MAE) of systolic blood pressure SBP is 5.61, and the standard deviation is 8.03; the mean absolute error of diastolic blood pressure DBP is 3.08, and the standard deviation is 5.35.

[0065] The above is a preferred embodiment of the present invention, but the embodiment of the present invention is not limited to the above content. Any other changes, modifications, substitutions, combinations, and simplifications that do not deviate from the spirit and principles of the present invention should be equivalent replacement methods and are included in the scope of protection of the present invention.

Claims

1. A non-invasive blood pressure monitoring system based on pulse wave, characterized in that: It includes a pulse wave acquisition module, an electrocardiogram acquisition module and a control module, wherein the pulse wave acquisition module is used to acquire pulse wave signals; the electrocardiogram acquisition module is used to acquire electrocardiogram signals; and the control module is used to obtain blood pressure based on the pulse wave signal and the electrocardiogram signal and its built-in blood pressure prediction model.

2. The pulse wave-based non-invasive blood pressure monitoring system according to claim 1, characterized in that: The control module is a single chip microcomputer, and its model is STM32F407.

3. The pulse wave-based non-invasive blood pressure monitoring system according to claim 1, characterized in that: The pulse wave acquisition module is a pulse wave acquisition chip, and its model is MAX30101.

4. The pulse wave-based non-invasive blood pressure monitoring system according to claim 1, characterized in that: The ECG acquisition module is an ECG acquisition chip, and its model is ADS1292R.

5. The pulse wave-based non-invasive blood pressure monitoring system according to claim 1, characterized in that: The control module communicates with the pulse wave acquisition module through the I2C protocol to control the working mode of the pulse wave acquisition module and receive the pulse wave signal returned by the pulse wave acquisition module, and store the received pulse wave data into the SD card.

6. The pulse wave-based non-invasive blood pressure monitoring system according to claim 1, characterized in that: The control module communicates with the ECG acquisition module via the SPI protocol to control the ECG acquisition module to acquire two-lead ECG signals and receive ECG signals returned by the ECG acquisition module, and stores the ECG data in the SD card.

7. A non-invasive blood pressure monitoring method for the pulse wave-based non-invasive blood pressure monitoring system according to any one of claims 1 to 6, characterized in that: The following steps are involved: Step 1: Collect PPG signals through the pulse wave acquisition module, collect ECG signals through the electrocardiogram acquisition module, and collect the corresponding blood pressure to construct a data set; Step 2: Preprocess the data in the dataset; Step 3: Divide the preprocessed dataset into training set, test set and validation set; Step 4: Construct a blood pressure prediction model and train it using the training set. Simultaneously record the training set loss value and validation set loss value after each round of training. Select the number of training rounds that corresponds to the minimum validation set loss value within the training cycle, and use the weight parameter of the blood pressure prediction model at the end of the round as the final weight. After loading the final weights into the blood pressure prediction model, the test set is input for prediction. The blood pressure prediction performance of the model is quantified using preset evaluation indicators, and a blood pressure prediction model that meets the accuracy requirements is screened out. Step 5: The PPG signal collected by the pulse wave acquisition module and the ECG signal collected by the electrocardiogram acquisition module are sent to the trained blood pressure prediction model to obtain the blood pressure prediction value.

8. The non-invasive blood pressure monitoring method according to claim 7, characterized in that: In step 1, the PPG signal within 3 minutes is collected by the pulse wave acquisition module, and the ECG signal within 3 minutes is collected by the electrocardiogram acquisition module; the systolic blood pressure SBP and diastolic blood pressure DBP are measured using a cuff sphygmomanometer, and the systolic blood pressure SBP and diastolic blood pressure DBP are used as reference blood pressure for the subsequent 3 minutes.

9. The non-invasive blood pressure monitoring method according to claim 7, wherein: In step 2, the preprocessing process is as follows: the collected PPG signal and ECG signal are filtered through a 0.5 Hz-8 Hz Butterworth bandpass filter, normalized, and the normalized signal is cut into 10 s windows to obtain multiple signal segments.

10. The non-invasive blood pressure monitoring method according to claim 7, characterized in that: In step 4, the data processing process of the blood pressure prediction model is as follows: Inputting the PPG signal into a first MultiResUnet network and the ECG signal into a second MultiResUnet network, reconstructing the ECG signal through the first MultiResUnet network, and reconstructing the PPG signal through the second MultiResUnet network; Get the output of the first downsampling layer, the output of the bottom layer, and the input before the last upsampling of the first MultiResUnet network, and the output of the first downsampling layer, the output of the bottom layer, and the input before the last upsampling of the second MultiResUnet network respectively; The obtained outputs are spliced ​​together and depth-wise separable convolution is performed on the spliced ​​results; The results of the depthwise separable convolution are spliced ​​again, and after one layer of convolution, the predicted systolic blood pressure SBP and diastolic blood pressure DBP are obtained.

Images