A wearable continuous blood pressure measurement method and system based on removal of age confounding factors

Abstract

The application discloses a wearable continuous blood pressure measurement method and system based on removal of an age confounding factor, wherein the system specifically comprises a data acquisition module, a data preprocessing module, a neural network blood pressure estimation model, and a blood pressure display module. The electrocardiogram (ECG) signal and the photoplethysmogram (PPG) signal can be collected by a wearable device to estimate blood pressure. The signals collected by the wearable device are first denoised, time-synchronized, normalized and subjected to a series of operations, and then the signals are input into the wearable continuous blood pressure estimation model based on removal of an age confounding factor for training, and finally the trained model is used for blood pressure estimation. It is proved by example data that the model accuracy and generalization are improved, and the robustness is better.

Description

Technical Field

[0001] The present invention belongs to the field of blood pressure measurement of medical devices, and mainly relates to a wearable continuous blood pressure measurement system based on removing age confounding factors. Technical Background

[0002] Blood pressure is the lateral pressure on the inner wall of blood vessels when blood flows continuously. It is the driving force that promotes blood flow and maintains blood perfusion and circulation in tissues and organs. The blood pressure value is an important basis for judging various cardiovascular diseases. Therefore, the measurement of blood pressure is of great significance in medicine.

[0003] Currently, there are two methods for measuring blood pressure: direct and indirect. Direct measurement involves inserting a device directly into a human artery to measure arterial blood pressure. Its advantages include continuous and highly accurate blood pressure measurements, making it the "gold standard" for blood pressure measurement. However, its disadvantages are invasive and cumbersome to perform, primarily used in critical care and during open surgery. Indirect measurement methods primarily include Korotkoff sound measurement, oscillometric measurement, pulse wave velocity measurement, radial artery applanation tonometry, and arterial volume compensation. While indirect measurement methods are widely used in clinical practice due to their non-invasive and relatively simple implementation, they still have some drawbacks. For example, Korotkoff sound measurement and oscillometric measurement cannot obtain continuous blood pressure measurements and require the inflation and deflation of a cuff, which can be uncomfortable for the patient. Radial artery applanation tonometry and arterial volume compensation methods can obtain continuous blood pressure measurements, but they require invasive measurement, complex measurement equipment, and specialized personnel, resulting in numerous inconveniences. While pulse wave velocity measurement can measure blood pressure continuously and non-invasively, it suffers from low accuracy and poor generalizability.

[0004] In recent years, the development of blood pressure measurement has focused on non-invasive, non-intrusive, and continuous measurement, which is of great significance for cardiovascular disease monitoring. The development of wearable sensors and artificial intelligence algorithms has provided a new approach to cuffless blood pressure measurement. This method primarily uses electrocardiogram (ECG) or pulse wave signals as input into a neural network, which then predicts blood pressure. However, due to the numerous factors that influence blood pressure, such as the user's age, weight, height, and gender, this type of blood pressure measurement suffers from low accuracy and poor generalization.

[0005] Age is a major factor that plagues blood pressure measurement. With age, the waveforms of both the ECG signal and the pulse wave signal change, and blood pressure also increases with age. Based on the premise that ECG and pulse wave signals can reflect changes in blood pressure, artificial intelligence-related algorithms can measure blood pressure through ECG (Electrocardiogram, ECG) signals or photoplethysmogram (Photoplethysmogram, PPG) signals. Because age affects both the cause—ECG and pulse wave signals—and the result—blood pressure, the estimation suffers from attribution confusion, which in turn reduces the accuracy of blood pressure measurement and reduces generalization. To solve the above problems, it is necessary to study and improve cuffless blood pressure measurement from a physiological mechanism perspective.

[0006] Causal representation learning has been a rapidly developing area of ​​research in recent years and has been the subject of extensive research in the deep learning field. Its advantage lies in reducing spurious correlations between learned data when handling complex application scenarios, making the model more robust and generalizable. Its key feature is that it considers potential attribution confounding caused by variables, generally referred to as confounding factors, when modeling the problem. Summary of the Invention

[0007] The technical problem to be solved by the present invention is to provide a wearable cuffless continuous blood pressure measurement system based on removing the age confounding factor, which can realize continuous blood pressure measurement with higher accuracy.

[0008] In order to solve the above problems, the present invention provides a wearable cuffless continuous blood pressure measurement system based on removing age confounding factors, the system includes: a data acquisition module, a data preprocessing module, a neural network blood pressure estimation model, and a blood pressure display module;

[0009] The data acquisition module collects reflected light signals or pressure signals from a certain part of the human body, and simultaneously collects electrocardiogram signals and blood pressure; and transmits all collected signals to the data preprocessing module;

[0010] The data preprocessing module filters noise and amplifies all signals collected by the data collector, and then converts the obtained reflected light signal or pressure signal into a pulse wave; normalizes the preprocessed electrocardiogram signal and pulse wave signal, and resamples them.

[0011] A wearable cuffless continuous blood pressure measurement method based on removing age confounding factors mainly includes the following steps:

[0012] Step 1: Use a wearable cuffless measurement device to obtain pulse wave information and electrocardiogram information, and use wearable electrical, optical, and pressure sensors to measure heart and arterial pulsation information in a certain part of the human body to obtain electrocardiogram, photoelectric volume pulse wave, and surface arterial pressure signals, and measure blood pressure changes in the signals;

[0013] Step 2: Preprocess the ECG signal and pulse wave signal, such as Figure 2 As shown;

[0014] In step 2, “preprocessing of the ECG signal and pulse wave signal” specifically includes the following steps:

[0015] Step 2.1: Use a Butterworth filter to remove high-frequency noise from the ECG signal and pulse wave signal;

[0016] Step 2.2: Align the ECG signal and the pulse wave signal cycles, and divide the ECG signal and the pulse wave signal corresponding to five consecutive cardiac cycles into a group of input signals;

[0017] Step 2.3: Normalize the ECG signal and pulse wave signal;

[0018] Step 2.4: Interpolate and resample the ECG signal and pulse wave signal, sampling 100 points per cardiac cycle;

[0019] Step 2.5: Calculate the signal quality index of the ECG signal and the pulse wave signal respectively with the reference ECG signal and the pulse wave signal, and remove the signals whose signal quality index is less than the set threshold;

[0020] Step 3: Build a blood pressure estimation network to remove the age confounding factor, such as Figure 1 As shown, the network outputs the estimated systolic or diastolic blood pressure through the fully connected layer;

[0021] In step 3, "building a blood pressure estimation network to remove age confounding factors" specifically includes the following steps:

[0022] Step 3.1: The network's input size is 2*500, specifically the ECG signal and pulse wave signal of 5 cardiac cycles, with 100 sampling points in each cycle;

[0023] The neural network model includes two layers of ANN, an age confounding factor removal layer, a GRU layer, and a fully connected layer;

[0024] Step 3.2: The first layer is an ANN layer with an input size of 2*500 and an output size of 2*1000, and the activation function is ReLu. The second layer is an ANN layer with an input size of 2*1000 and an output size of 2*250, and the activation function is ReLu.

[0025] Step 3.3: The third layer removes the age confounding factor from the features output by the second layer, and the expression is as follows:

[0026]

[0027] wherein is the age-confounded feature, q t is the feature output by the second layer, K is the average feature of each age group, Z is the average feature, p(z) is the probability of the input feature q t length;

[0028] The average feature of each age group is obtained by dividing the samples of 20-89 years old into three groups, i.e., 20-39 years old, 40-59 years old, and 60-89 years old, and then pre-training a model for each group. p(z) is obtained by connecting a 5-layer perception machine with a softmax layer to predict the probability, and the input size is 2*500, specifically, the electrocardiogram signal and the pulse wave signal of 5 cardiac cycles; the output is the probability of the age being 20-39 years old, 40-59 years old, and 60-89 years old, respectively.

[0029] Step 3.4: The obtained age group probability is input into an ANN layer, and the output size is 2*100. The final age-confounded factor removal layer has an input size of 2*250 and an output size of 2*100.

[0030] Step 3.5: The fourth layer concatenates the output of the third layer and the output of the second layer, and the output size is 2*350.

[0031] Step 3.6: The fifth layer is a GRU layer, and the input size is 2*350 and the output size is 2*128.

[0032] Step 3.7: The sixth layer is a fully connected layer with an input size of 256 and an output size of 50, and the activation function is ReLu. The seventh layer is a fully connected layer with an input size of 50 and an output size of 1 to predict systolic pressure or diastolic pressure, and the activation function is ReLu.

[0033] Step 4: The sample data is input into the age-confounded factor removal blood pressure estimation network for training, and the parameters of the blood pressure estimation network are updated, and finally the age-confounded factor removal blood pressure estimation network is obtained.

[0034] Step 5: The data obtained by the smart wearable device is input into the trained age-confounded factor removal blood pressure estimation network, and the blood pressure estimation value is displayed to the blood pressure display module.

[0035] In step 1, “wearing a cuffless continuous blood pressure measurement device” specifically includes the following steps:

[0036] Step 1.1: Select the subject's surface artery as the object to be examined;

[0037] Step 1.2: Continuously measure the blood pressure at the subject's selected surface artery using a cuff-type continuous blood pressure measurement device;

[0038] Step 1.3: While measuring the blood pressure with a cuff, place the wearable measurement device on the skin surface of the other artery of the same subject.

[0039] Step 1.4: Simultaneously record the arterial pulse and the heart-generated signal;

[0040] In step 4, "removing the age confounding factor and training the blood pressure estimation network" specifically includes:

[0041] Use the root mean square error function to calculate the loss value, and use the Adam algorithm back propagation to update the network weight parameters to make the model gradually converge;

[0042] The learning rate used is 0.0001, the weight decay is 0.0001, and the batch_size is 128.

[0043] The present invention adopts the above solution and has the following beneficial effects:

[0044] 1. Compared with traditional inflatable cuff blood pressure measurement, the present invention can achieve continuous, non-invasive blood pressure prediction and is more suitable for continuous blood pressure monitoring scenarios.

[0045] 2. Compared with traditional artificial intelligence blood pressure estimation algorithms, the present invention has better generalization.

[0046] 3. The present invention is based on deep learning algorithm and does not require manual feature extraction.

[0047] 4. Compared with the algorithm that estimates blood pressure using only the pulse wave signal, the present invention uses both the electrocardiogram signal and the pulse wave signal, and has better robustness.

[0048] 5. The signal acquisition and algorithm processing of the present invention can be completed by a microcontroller chip, which can realize miniaturized wearable device monitoring and can be applied to daily life scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 Diagram of the network structure model for wearable continuous blood pressure estimation after removing age confounding factors;

[0050] Figure 2 It is the signal preprocessing flow chart;

[0051] Figure 3 a set of pulse wave signals inputted to the network.

[0052] Figure 4 a set of pulse wave signals inputted to the network. DETAILED DESCRIPTION

[0053] In order to achieve the above-mentioned purpose, the wearable device measures the physiological information of the human body's artery and heart, obtains the electrocardiogram signal and pulse wave signal, and then the arterial blood pressure, and then the signal is preprocessed. The electrocardiogram signal and pulse wave signal of 5 cardiac cycles are a group, which is inputted to the age confusion factor removal model for training to estimate the average blood pressure of 5 cycles.

[0054] The wearable continuous blood pressure measurement system for removing age confusion factor specifically comprises the following steps:

[0055] Step one: using a continuous blood pressure measurement device to measure the radial artery blood pressure of the subject;

[0056] Step two: wearing the wearable device on the surface of the subject's skin to record the subject's electrocardiogram signal and pulse wave signal, and synchronously measuring the arterial blood pressure;

[0057] Step three: preprocessing the signals collected in steps one and two, mainly including noise filtering, normalization, cycle alignment, resampling, etc.

[0058] Step four: according to the idea of removing confusion factors, an age confusion factor removal blood pressure estimation model is established, as shown in Figure 1 for real-time estimation of arterial blood pressure.

[0059] The age confusion factor removal blood pressure estimation model in step four is specifically as follows:

[0060] 1) The input size of the network is 2*500, specifically the electrocardiogram signal and pulse wave signal of 5 cardiac cycles, wherein each cycle has 100 sampling points.

[0061] 2) The first layer is an ANN layer with an input size of 2*500 and an output size of 2*1000, and the activation function is ReLu, and the second layer is an ANN layer with an input size of 2*1000 and an output size of 2*250, and the activation function is ReLu;

[0062] 3) The third layer processes the features outputted by the second layer to remove the age confusion factor, and then puts them into the ANN layer, and the output size is 2*100.

[0063] 4) The fourth layer splices the output of the third layer and the output of the second layer together, and the output size is 2*350;

[0064] 5) The fifth layer is the GRU layer, with an input size of 2*350 and an output size of 2*128;

[0065] 6) The sixth layer is a fully connected layer with an input size of 256, an output size of 50, and an activation function of ReLu;

[0066] 7) The seventh layer is a fully connected layer with an input size of 50 and an output size of 1 to predict systolic or diastolic blood pressure. The activation function is ReLu.

[0067] Step 5: The sample data is input into the blood pressure estimation model after removing the age confounding factor for training, specifically including: using the RMSE root mean square error as the loss function to calculate the loss value:

[0068]

[0069] Where Y is the reference blood pressure value, To predict blood pressure values, N is the number of samples input to the model, i is each prediction sample, and σ is the input feature q t length.

[0070] The Adam algorithm is used to back-propagate and update the network weights, so that the network model gradually converges. The learning rate is 0.0001, the weight decay is 0.0001, and the batch_size is 128.

[0071] Step 6: Input the real-time pulse wave signal into the trained blood pressure estimation model that removes the age confounding factor to obtain the real-time predicted blood pressure value

[0072] To verify the method proposed in the present invention, the electrocardiogram (ECG) signals and pulse wave signals of 96 patients aged 20 to 89 were collected experimentally.

[0073] The specific experimental steps are as follows:

[0074] Step 1: Use a wearable device to collect the patient's ECG signal and pulse wave signal, and simultaneously collect systolic and diastolic blood pressure values;

[0075] Step 2: Input ECG signal Figure 3 As shown, the input pulse wave signal is Figure 4 As shown, preprocessing is performed, such as Figure 2 As shown in the figure, a Butterworth filter is used to filter out high-frequency noise from the ECG signal and the pulse wave signal. The ECG signal and the pulse wave signal are resampled and time-synchronized to eliminate the time delay between the two signals. Finally, the signal quality index is calculated with the reference signal, and signals with a signal quality index less than 0.75 are removed.

[0076] Step 3: Put the preprocessed signal and reference blood pressure values ​​(systolic pressure, diastolic pressure) into the blood pressure estimation model after removing the age confounding factor for training, and use the root mean square error (RMSE) as the loss function to calculate the loss value:

[0077]

[0078] Where Y is the reference blood pressure value, To predict blood pressure values, N is the number of samples input to the model, and i is each predicted sample. The Adam algorithm is then used to backpropagate and update the network weights, gradually converging the network model. The learning rate is 0.0001, the weight decay is 0.0001, and the batch size is 128.

[0079] Step 4: The trained model is used to estimate blood pressure from 20 individuals with a total of 2000 input segments of ECG and pulse wave signals. The accuracy of the model in estimating systolic blood pressure is shown in Table 1.

[0080] Table 1 Comparison of the measurement performance of the traditional neural network model and the model with the age confounding factor removed (BHS standard)

[0081]

Claims

1. A wearable cuffless continuous blood pressure measurement method based on removing age confounding factors mainly includes the following steps: Step 1: Use a wearable cuffless measurement device to obtain pulse wave information and electrocardiogram information, and use wearable electrical, optical, and pressure sensors to measure heart and arterial pulsation information in a certain part of the human body to obtain electrocardiogram, photoelectric volume pulse wave, and surface arterial pressure signals, and measure blood pressure changes in the signals; Step 2: Preprocess the ECG signal and pulse wave signal; In step 2, "preprocessing of the electrocardiogram signal and pulse wave signal" specifically includes the following steps: Step 2.1: Use a Butterworth filter to remove high-frequency noise from the ECG signal and pulse wave signal; Step 2.2: Align the ECG signal and the pulse wave signal cycles, and divide the ECG signal and the pulse wave signal corresponding to five consecutive cardiac cycles into a group of input signals; Step 2.3: Normalize the ECG signal and pulse wave signal; Step 2.4: Interpolate and resample the ECG signal and pulse wave signal, sampling 100 points per cardiac cycle; Step 2.5: Calculate the signal quality index of the ECG signal and the pulse wave signal respectively with the reference ECG signal and the pulse wave signal, and remove the signals whose signal quality index is less than the set threshold; Step 3: Build a blood pressure estimation neural network model that removes the age confounding factor. The neural network model outputs the estimated systolic or diastolic blood pressure through a fully connected layer. In step 3, "building a blood pressure estimation neural network model by removing age confounding factors" specifically includes the following steps: Step 3.1: The input_size of the neural network model is 2*500, specifically the ECG signal and pulse wave signal of 5 cardiac cycles, with 100 sampling points in each cycle; The neural network model includes two layers of ANN, an age confounding factor removal layer, a GRU layer, and a fully connected layer; Step 3.2: The first layer is an ANN layer with an input size of 2*500 and an output size of 2*1000, and the activation function is ReLu. The second layer is an ANN layer with an input size of 2*1000 and an output size of 2*250, and the activation function is ReLu. Step 3.3: The third layer removes the age confusion factor from the features output by the second layer. The expression is as follows: in To remove the age-confounding feature, q t is the second layer output feature, K is the average feature of each age group, Z is the average feature, p(z) input feature is the probability of each age group, σ is the input feature q t length; For the average features of each age group, the samples aged 20-89 were first divided into three groups: 20-39, 40-59, and 60-89. An average feature extraction model was pre-trained for each group to extract the average features of each age group. p(z) was predicted probabilistically using a 5-layer perceptron connected to a softmax layer. The input_size was 2*500, specifically the ECG and pulse wave signals of 5 cardiac cycles. The output was the probability of being 20-39, 40-59, and 60-89 years old, respectively. Step 3.4: Put the obtained age group probabilities into an ANN layer with an output size of 2*100. The final layer constructed to remove the age confounding factor has an input size of 2*250 and an output size of 2*100. Step 3.5: The fourth layer concatenates the output of the third layer and the output of the second layer, and the output size is 2*350; Step 3.6: The fifth layer is the GRU layer, with an input size of 2*350 and an output size of 2*128; Step 3.7: The sixth layer is a fully connected layer with an input size of 256, an output size of 50, and an activation function of ReLu. The seventh layer is a fully connected layer with an input size of 50, an output size of the predicted systolic or diastolic blood pressure, and an activation function of ReLu. Step 4: Put the sample data into the blood pressure estimation neural network model after removing the age confounding factor for training, update the parameters of the blood pressure estimation neural network model, and finally obtain the blood pressure estimation neural network model after removing the age confounding factor; Step 5: Input the data obtained by the smart wearable device into the trained blood pressure estimation neural network model after removing the age confounding factor, and obtain the blood pressure estimation value to display on the blood pressure display module.

2. A wearable cuffless continuous blood pressure measurement method based on removing age confounding factors as claimed in claim 1, characterized in that: The method in step 1 specifically comprises the following steps: Step 1.1: Select the subject's surface artery as the object to be examined; Step 1.2: Continuously measure the blood pressure at the subject's selected surface artery using a cuff-type continuous blood pressure measurement device; Step 1.3: While measuring the blood pressure with a cuff, place the wearable measurement device on the skin surface of the other artery of the same subject. Step 1.4: Simultaneously record the arterial pulse and the signals generated by the heart.

3. A wearable cuffless continuous blood pressure measurement method based on removing age confounding factors as claimed in claim 1, characterized in that: The step 4 specifically includes: Use the root mean square error function to calculate the loss value, and use the Adam algorithm back propagation to update the neural network model weight parameters to make the model gradually converge; The learning rate used is 0.0001, the weight decay is 0.0001, and the batch_size is 128.

4. A system using the method of claim 1, comprising: Data acquisition module, data preprocessing module, blood pressure estimation neural network model, blood pressure display module; The data acquisition module collects reflected light signals or pressure signals from a certain part of the human body, and simultaneously collects electrocardiogram signals and blood pressure; and transmits all collected signals to the data preprocessing module; The data preprocessing module filters out noise and amplifies all signals collected by the data collector, and then converts the obtained reflected light signal or pressure signal into a pulse wave; normalizes and resamples the preprocessed ECG signal and pulse wave signal; The blood pressure estimation neural network model estimates blood pressure based on the preprocessed data; Finally, there is a blood pressure display module for display.

Images