A blood pressure measurement method, system, device and medium

Abstract

The present invention relates to a blood pressure measurement method, system, device and medium, wherein the method comprises: collecting SCG signals and blood pressure data of a human body; performing three-level noise reduction processing on the collected SCG signals, including: performing preliminary noise reduction on the SCG signals by an ICEEMDAN method to obtain a first signal; performing further noise reduction on the first signal by an RLS adaptive filter to obtain a second signal; performing three noise reduction on the second signal by a soft threshold method to obtain a noise-reduced SCG signal; training a pre-set deep learning model using the blood pressure data and the noise-reduced SCG signal to obtain a trained deep learning model; and measuring the blood pressure of a subject to be measured by using the trained deep learning model. The present invention can accurately measure human blood pressure.

Description

Technical Field

[0001] The present invention relates to the technical field of electrocardiogram measurement, and in particular to a blood pressure measurement method, system, equipment and medium. Background Art

[0002] Measuring blood pressure is an important means of monitoring human vital signs. Through effective blood pressure measurement methods, we can understand whether the blood pressure level is normal, whether there is hypertension or hypotension, so as to provide timely treatment.

[0003] Currently, some blood pressure measurement solutions combine traditional electrocardiogram (ECG) measurement devices, which are used in professional medical facilities. Although these devices can accurately measure ECG signals to infer blood pressure, these devices are usually costly, bulky, and require professional operation, making them unsuitable for blood pressure monitoring everywhere.

[0004] Based on the need to measure blood pressure anywhere, the most convenient and mainstream blood pressure measurement methods currently use systems based on photoplethysmography (PPG) signals, which use the built-in camera and LED of smartphones to capture pulse waveforms, and then process and analyze the pulse waveforms through algorithms to estimate blood pressure. However, these systems are susceptible to a variety of interference factors, such as finger movement, ambient lighting fluctuations, etc., which cause signal distortion and thus affect the accuracy of blood pressure estimation.

[0005] In summary, under the condition of being suitable for blood pressure monitoring anywhere, the existing technology mainly measures blood pressure through a smartphone blood pressure measurement system based on PPG signals. However, they have problems of signal distortion and insufficient blood pressure measurement accuracy. Summary of the invention

[0006] To this end, the technical problem to be solved by the present invention is to overcome the problem in the prior art that the smartphone blood pressure measurement system based on PPG signals is easily interfered with, thereby reducing the accuracy of blood pressure measurement.

[0007] In order to solve the above technical problems, the present invention provides a blood pressure measurement method, comprising:

[0008] Step S1: Collecting the SCG signal and blood pressure data of the human body, and performing three-level noise reduction processing on the collected SCG signal, wherein the three-level noise reduction processing includes: performing preliminary noise reduction on the SCG signal by the ICEEMDAN method to obtain a first signal; performing further noise reduction on the first signal by the RLS adaptive filter to obtain a second signal; performing three-times noise reduction on the second signal by the soft threshold method to obtain a noise-reduced SCG signal;

[0009] Step S2: training a preset deep learning model using the blood pressure data and the denoised SCG signal to obtain a trained deep learning model;

[0010] Step S3: Measure the blood pressure of the subject using the trained deep learning model.

[0011] In one embodiment of the present invention, in step S1, the SCG signal is preliminarily denoised by the ICEEMDAN method to obtain a first signal, and the method includes:

[0012] The SCG signal is decomposed into several different intrinsic mode functions IMFs by the ICEEMDAN method, and the intrinsic mode function IMF with the most significant periodicity of the heartbeat signal is selected, and the intrinsic mode function IMF with the most significant periodicity of the heartbeat signal is used as the first signal.

[0013] In one embodiment of the present invention, in step S1, the first signal is subjected to further denoising by an RLS adaptive filter to obtain a second signal, and the method includes:

[0014] Define the most significant intrinsic mode function IMF of the heartbeat signal as x, and divide x into L segments, namely x1, ..., x L , calculate the i-th segment x in the L segment i and segment j x j The similarity is:

[0015]

[0016] Among them, μ i and μ j ,δ i and δ j Respectively represent the i-th segment x i and segment j x j The mean and standard deviation of N is the signal length, k represents the signal sampling time when calculating the similarity, m represents the element offset when calculating the similarity, and mod represents the remainder;

[0017] Count the number of similarities between each segment and other segments that exceed the preset threshold, and use the segment with the largest number of similarities exceeding the similarity threshold as the time reference sample;

[0018] A mean square error is constructed according to x and the time reference sample, and the RLS adaptive filter performs denoising on the first signal again by minimizing the mean square error to obtain a second signal. The mean square error formula is:

[0019]

[0020] Where n is the signal sampling time when calculating the mean square error, y is the time reference sample, w(n) is the weight vector of the RLS adaptive filter, T is the matrix transpose, || is the absolute value, and λ(n) is the forgetting factor and satisfies:

[0021]

[0022] Among them, δ(n) and δ r (n) represent the standard deviation of the input signal and the time reference sample, respectively, ξ and ρ are the first constant and the second constant;

[0023] Define R(n) as the inverse of the autocorrelation matrix of the input signal, expressed as: R(n) = [x·x T ] -1 ;

[0024] At each signal sampling time n, the process of updating w(n) is expressed as:

[0025] Calculate the prior error, expressed as: η(n) = yw T (n-1)x(n);

[0026] Determine the gain vector:

[0027] Update the inverse matrix:

[0028] Update weight: w(n)=w(n-1)+η(n)e(n).

[0029] In one embodiment of the present invention, in step S1, the second signal is subjected to three noise reduction operations using a soft threshold method to obtain a noise-reduced SCG signal, and the method includes:

[0030] Design empirical global threshold, expressed as:

[0031]

[0032] Where N is the signal length; δ n represents the noise level, expressed as:

[0033]

[0034] Among them, S n represents the value of the sequence in the second signal obtained by denoising through the RLS adaptive filter; median() represents taking the median value;

[0035] For each value in the sequence of the second signal, if δ n -ψ>0, the value remains unchanged; if δ n-ψ<0, the value is 0; the second signal is subjected to three noise reduction operations to obtain a noise-reduced SCG signal.

[0036] In one embodiment of the present invention, in step S3, the blood pressure of the subject to be measured is measured by using the trained deep learning model, and the method includes:

[0037] The SCG signal of the object to be tested is collected and subjected to three-level noise reduction processing, and the noise-reduced SCG signal is used as the input signal of the trained deep learning model. Assuming that the input signal is a 1-dimensional feature vector 1*640 with a length of 640, the 1-dimensional feature vector 1*640 with a length of 640 is converted from 1 dimension to 64 dimensions through 64 1-dimensional convolution kernels to obtain a 64-dimensional feature vector 64*640 with a length of 640;

[0038] The 64-dimensional feature vector 64*640 with a length of 640 is passed through the first residual module, the second residual module, and the third residual module respectively to obtain a 256-dimensional feature vector 256*160 with a length of 160, and the 256-dimensional feature vector 256*160 with a length of 160 is converted into a first 1-dimensional feature vector 1*160 with a length of 160 through 256 one-dimensional convolution kernels. The first 1-dimensional feature vector 1*160 with a length of 160 is passed through the Transformer layer to obtain a second 1-dimensional feature vector 1*160 with a length of 160. The second 1-dimensional feature vector 1*160 with a length of 160 is passed through the first fully connected layer and randomly inactivated to obtain a 1-dimensional feature vector 1*110 with a length of 110. The 1-dimensional feature vector 1*110 with a length of 110 is passed through the second fully connected layer to obtain blood pressure data.

[0039] In one embodiment of the present invention, the first residual module includes a first depth-wise separable convolutional layer and a second depth-wise separable convolutional layer connected in sequence, and the input of the first depth-wise separable convolutional layer and the output of the second depth-wise separable convolutional layer are shortcut-connected as the output of the first residual module; wherein a 64-dimensional feature vector 64*640 with a length of 640 is passed through the first depth-wise separable convolutional layer to obtain a first 128-dimensional feature vector 128*320 with a length of 320, and the first 128-dimensional feature vector 128*320 with a length of 320 is passed through the second depth-wise separable convolutional layer to obtain a second 128-dimensional feature vector 128*320 with a length of 320;

[0040] The second residual module includes a third depth-separable convolution layer and a fourth depth-separable convolution layer connected in sequence, and the input of the third depth-separable convolution layer and the output of the fourth depth-separable convolution layer are quickly connected as the output of the second residual module; wherein the second 128-dimensional feature vector 128*320 with a length of 320 is passed through the third depth-separable convolution layer to obtain a third 128-dimensional feature vector 128*320 with a length of 320, and the third 128-dimensional feature vector 128*320 with a length of 320 is passed through the fourth depth-separable convolution layer to obtain a fourth 128-dimensional feature vector 128*320 with a length of 320;

[0041] The third residual module includes a fifth depth-wise separable convolutional layer and a sixth depth-wise separable convolutional layer connected in sequence, and the input of the fifth depth-wise separable convolutional layer and the output of the sixth depth-wise separable convolutional layer are shortcut-connected as the output of the third residual module; wherein a fourth 128-dimensional feature vector 128*320 with a length of 320 is passed through the fifth depth-wise separable convolutional layer to obtain a first 256-dimensional feature vector 256*160 with a length of 160, and the first 256-dimensional feature vector 256*160 with a length of 160 is passed through the sixth depth-wise separable convolutional layer to obtain a second 256-dimensional feature vector 256*160 with a length of 160.

[0042] In one embodiment of the present invention, when the deep learning model is trained using the blood pressure data and the denoised SCG signal in step S2, the loss function formula used is:

[0043]

[0044] Among them, y represents the training label, represents the model output, cosh( ) represents the hyperbolic function, λ * represents the regularization weight factor, W represents the model parameter, and || ||1 represents the one-norm.

[0045] In order to solve the above technical problems, the present invention provides a blood pressure measurement system, comprising:

[0046] The acquisition and noise reduction module is used to acquire the SCG signal and blood pressure data of the human body, and perform three-level noise reduction processing on the acquired SCG signal, wherein the three-level noise reduction processing includes: performing preliminary noise reduction on the SCG signal by the ICEEMDAN method to obtain a first signal; performing further noise reduction on the first signal by the RLS adaptive filter to obtain a second signal; performing three-time noise reduction on the second signal by the soft threshold method to obtain a noise-reduced SCG signal;

[0047] Training module: used to train a preset deep learning model using the blood pressure data and the denoised SCG signal to obtain a trained deep learning model;

[0048] Detection module: used to measure the blood pressure of the subject through the trained deep learning model.

[0049] To solve the above technical problems, the present invention provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above blood pressure measurement method when executing the computer program.

[0050] In order to solve the above technical problem, the present invention provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the blood pressure measurement method as described above are implemented.

[0051] The above technical solution of the present invention has the following advantages compared with the prior art:

[0052] The present invention collects the human body's SCG signals and performs three-level noise reduction processing on the signals, which can eliminate noise to the greatest extent and facilitate the subsequent deep learning model to effectively identify blood pressure;

[0053] The present invention uses an RLS adaptive filter to reduce noise on the signal and improves the forgetting factor therein, so that the noise reduction effect is better; the present invention also designs a soft threshold method to reduce noise on the signal, which can further reduce the interference of noise;

[0054] The present invention constructs a deep learning model that can accurately measure blood pressure;

[0055] The present invention also designs a loss function for the deep learning model, through which the model can be effectively trained. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] In order to make the contents of the present invention more clearly understood, the present invention is further described in detail below based on specific embodiments of the present invention in conjunction with the accompanying drawings.

[0057] Figure 1 is a flow chart of the method of the present invention;

[0058] Figure 2 is a schematic diagram of three-axis signals collected by an accelerometer in a mobile phone in an embodiment of the present invention;

[0059] Figure 3 is a schematic diagram of Y-axis data in a three-axis accelerometer signal waveform in an embodiment of the present invention;

[0060] Figure 4It is a schematic diagram of decomposing the target SCG signal into multiple different IMFs using the ICEEMDAN method in an embodiment of the present invention;

[0061] Figure 5 is a schematic diagram of a signal after noise reduction by an RLS adaptive filter in an embodiment of the present invention;

[0062] Figure 6 is a schematic diagram of a signal after noise reduction by a soft threshold method in an embodiment of the present invention;

[0063] Figure 7 It is a schematic diagram of the deep learning model structure in an embodiment of the present invention. DETAILED DESCRIPTION

[0064] The present invention is further described below in conjunction with the accompanying drawings and specific embodiments so that those skilled in the art can better understand the present invention and implement it, but the embodiments are not intended to limit the present invention.

[0065] Embodiment 1

[0066] Reference Figure 1 As shown, the present invention relates to a blood pressure measurement method, comprising:

[0067] Step S1: Collecting the SCG signal and blood pressure data of the human body, and performing three-level noise reduction processing on the collected SCG signal, wherein the three-level noise reduction processing includes: performing preliminary noise reduction on the SCG signal by the ICEEMDAN method to obtain a first signal; performing further noise reduction on the first signal by the RLS adaptive filter to obtain a second signal; performing three-times noise reduction on the second signal by the soft threshold method to obtain a noise-reduced SCG signal;

[0068] Step S2: training a preset deep learning model using the blood pressure data (as a label) and the denoised SCG signal to obtain a trained deep learning model;

[0069] Step S3: Measure the blood pressure of the subject using the trained deep learning model.

[0070] The following is a detailed introduction to this embodiment:

[0071] Step 1: Collect SCG signals and blood pressure data: Specifically, use the built-in accelerometer of the mobile phone to collect chest vibration (SCG) signals on the chest and collect blood pressure data.

[0072] Step 2: Three-stage noise reduction: Use the ICEEMDAN (Improved Complete Ensemble EmpiricalMode Decomposition with Adaptive Noise) method for preliminary noise reduction. Then apply the RLS (Recursive Least Squares) adaptive filter to further filter the noise. Finally, use the soft threshold method to perform three-stage noise reduction to reconstruct a high-quality heartbeat waveform from the original SCG signal.

[0073] Step 3: Put the denoised heart vibration signal and blood pressure data (as labels) into the deep learning model for training, and finally establish a deep learning model that converts SCG signals to blood pressure (i.e., the trained deep learning model).

[0074] Step 4: Deploy the trained deep learning model to the mobile phone, collect the SCG signal from the chest of the subject to be tested, and measure the blood pressure data of the subject to be tested through the trained deep learning model.

[0075] In step 1, the accelerometer data of the smartphone placed on the chest is collected, and the blood pressure data is collected using a blood pressure monitor on the arm. The accelerometer of the smartphone returns the values ​​of three axes (including the X-axis, Y-axis, and Z-axis). Figure 2 The three-axis accelerometer signal waveform is shown in FIG. 1. In this embodiment, only the Y-axis data is used as the data required by this embodiment. Figure 3 It is the Y-axis data in the three-axis accelerometer signal waveform collected in this embodiment, which is used as the target SCG signal (with large fluctuation noise from 2.5 to 3.2 seconds).

[0076] In step 2, the SCG signal is initially denoised using the ICEEMDAN method. Specifically, empirical mode decomposition (EMD) is an adaptive technique well suited for the analysis of non-stationary and nonlinear time-varying series. Instead of using the specified basis functions in wavelet-based methods, EMD decomposes the time series into a series of independent time-frequency components, called intrinsic mode functions (IMFs), which help to extract the heartbeat patterns. This is mainly because the IMF components are instantaneous frequency components that represent the instantaneous components of the local characteristics of the origin, which makes EMD very suitable for analyzing non-stationary and nonlinear data. However, this introduces a potential challenge in EMD, known as “mode mixing”. Here, different scale oscillations may occur in the same mode, or similar scale oscillations may manifest in different modes, complicating the extraction of heartbeat patterns. Based on empirical mode decomposition, the fully integrated empirical mode decomposition with adaptive noise (CEEMDAN) technique is proposed as a solution to the modulation and demodulation problem achieved by integrating adaptive white noise. However, this approach addresses two important issues: residual noise in the mode and spurious modes that appear during the decomposition process. The EMD-based technology was further improved to obtain ICEEMDAN, which has been proven to be effective in solving the above challenges. Therefore, this embodiment finally uses the ICEEMDAN method to extract heartbeat patterns. Figure 4 As shown in Figure 2, the target SCG signal is decomposed into five different IMFs (IMF1-IMF5) using the ICEEMDAN method. Figure 4 As can be seen from the figure, the heartbeat signal in IMF1 is obviously prominent and shows obvious periodic characteristics. Therefore, this embodiment selects IMF1 to reconstruct the heartbeat signal. The results show that the ICEEMDAN method can effectively eliminate most of the noise, especially Figure 3 The noise generated by the sudden fluctuations between 2.5 and 3.2 seconds in the heartbeat is removed. However, due to the pseudo-periodicity of the heartbeat and the instability of the noise, the ICEEMDAN method cannot completely eliminate all the noise and further noise filtering is required.

[0077] Although the ICEEMDAN method significantly reduces the impact of hand shaking, it does not completely successfully eliminate hand shaking, resulting in the presence of residual noise. As a result, this results in segments of the signal that lack clearly distinguishable patterns. The following will focus on how to further filter out the impact of residual motion artifacts.

[0078] In step 2, an RLS adaptive filter is applied to further filter the noise. Before using the RLS filter, a reference signal (time reference sample) that reflects the ideal output or expected state of the system must be selected. This embodiment requires accurate identification and extraction of segments with similar patterns: similar segments show high correlation, while different segments produce low correlation. Based on this concept, the method of this embodiment is summarized as follows: First, the output of the ICEEMDAN method, i.e., x, is divided into L segments

[0079] (i.e. L fragments), we get x1, ..., x L , where L is determined by the signal length N. To ensure that each segment fully captures the overall properties of the SCG morphology, the segment length is specified to be between one and two cycles. Conventionally, the heartbeat cycle of an adult at rest is approximately 0.6-1.0 seconds. Therefore, guided by empirical data, this embodiment determines the segment length to be 1 second. Then, for any two segments x i and x j , and the similarity is obtained by calculating the Pearson correlation coefficient:

[0080]

[0081] Among them, μ i and μ j ,δ i and δ j Respectively represent the i-th segment x i and segment j x j The average value and standard deviation of k represents the signal sampling moment when calculating the similarity, that is, the kth point in the sequence corresponding to the segment; m represents the element offset when calculating the similarity, which varies from 0 to N-1 and is used to find the maximum similarity between two segments; mod represents the remainder. In this embodiment, segments whose similarity exceeds a preset threshold (0.7) are considered candidate segments. Finally, the segment that appears most frequently in the candidate set is identified as the time reference sample. Figure 4 As shown, after calculation, the segment from 8 seconds to 9 seconds in IMF1 is used as the time reference sample.

[0082] The mean square error is constructed based on the output of ICEEMDAN, that is, x, and the time reference sample y. The RLS adaptive filter further reduces the noise of the first signal by minimizing the mean square error. The mean square error formula is:

[0083]

[0084] Where n is the signal sampling time when calculating the mean square error, y is the time reference sample, w(n) is the weight vector of the RLS adaptive filter, T is the matrix transpose, and || is the absolute value; λ(n) is the forgetting factor, the purpose of which is to assign a larger weight to the mean square error close to time n, and a smaller weight to other times. The larger the forgetting factor λ(n), the better the stability of the system, and the smaller the forgetting factor λ(n), the stronger the adaptability of the system. When a sudden interference occurs, by reducing the forgetting factor λ(n), the weight of the historical data can be reduced, so that the filter can adapt to new system conditions more quickly, making it more sensitive in tracking system changes. Based on this idea, the forgetting factor λ(n) is designed in this embodiment as:

[0085]

[0086] Among them, δ(n) and δ r (n) represents the standard deviation of the input signal and the time reference sample, respectively. This embodiment uses the ratio of the two to quantify the motion noise level; ξ and ρ are the first constant and the second constant, and ρ is limited to the range of 0 to 1. As the motion noise level increases, the forgetting factor decreases accordingly, which enables faster tracking of high-frequency and non-stationary motion and reduces distortion in the input signal. Next, define R(n) as the inverse of the autocorrelation matrix of the input signal, expressed as: R(n) = [x·x T ] -1 , e(n) is the gain vector. To calculate the optimal weight vector w(n), the weight vector w(0) is defined as 0, and the autocorrelation matrix R(0) is defined as μ·i*, where μ represents a custom positive number (for the purpose of accelerating convergence), and u* is the unit matrix. At each signal sampling time n, the process of updating w(n) is expressed as:

[0087] Calculate the prior error, expressed as: η(n) = yw T (n-1)x(n);

[0088] Determine the gain vector:

[0089] Update the inverse matrix:

[0090] Update weight: w(n)=w(n-1)+η(n)e(n).

[0091] like Figure 5As shown in the figure, compared with the results obtained by using only the ICEEMDAN method, the method of this embodiment significantly reduces motion noise and obtains a more discernible heartbeat pattern. Although the previous technology has been proven to be very effective in eliminating most motion noise, there is still a small amount of residual tiny spike noise. Therefore, this embodiment designs a lightweight denoising algorithm, called the soft threshold method, to further extract effective heart vibration signals. The method is as follows:

[0092] Design empirical global threshold, expressed as:

[0093]

[0094] Where N is the signal length; δ n represents the noise level, expressed as:

[0095]

[0096] Among them, S n represents the value of the sequence in the second signal obtained by noise reduction through the RLS adaptive filter (for example, obtained by sampling the signal at a sampling frequency of 100 Hz);

[0097] For each value in the sequence of the second signal, if δ n -ψ>0, the value remains unchanged; if δ n -ψ<0, the value is 0.

[0098] The above process effectively filters out low-amplitude noise components and retains the main features of the signal. That is, the soft threshold method further refines the SCG signal and produces a more obvious heartbeat pattern. Figure 6 shown.

[0099] In step 3 and step 4, in order to solve the problems of resource overhead and low precision, this embodiment constructs a deep learning model and trains the deep learning model, and finally implements a lightweight model deployment with high precision and robustness on a smart phone. Specifically, the SCG signal of the object to be tested after noise reduction is input into the trained deep learning model, and the blood pressure of the object to be tested is detected by the trained deep learning model, and the method includes:

[0100] Collect the SCG signal of the object to be tested and perform three-level noise reduction processing, and use the SCG signal after noise reduction as the input signal of the trained deep learning model. Figure 7, assuming that the input signal length N is a 1-dimensional feature vector 1*640 of 640, the 1-dimensional feature vector 1*640 of length 640 is converted from 1 dimension to 64 dimensions through 64 1-dimensional convolution kernels to obtain a 64-dimensional feature vector 64*640 of length 640; the 64-dimensional feature vector 64*640 of length 640 is respectively passed through the first residual module, the second residual module, and the third residual module to obtain a 256-dimensional feature vector 256*160 of length 160, and the 256-dimensional feature vector 256*160 of length 160 is converted into the first length through 256 one-dimensional convolution kernels. The first 1-dimensional feature vector 1*160 with a length of 160 is passed through the Transformer layer to obtain a second 1-dimensional feature vector 1*160 with a length of 160, the second 1-dimensional feature vector 1*160 with a length of 160 is passed through the first fully connected layer (the output of the first fully connected layer is a 1-dimensional feature vector 1*110 with a length of 110), and random inactivation is performed to obtain a 1-dimensional feature vector 1*110 with a length of 110, and the 1-dimensional feature vector 1*110 with a length of 110 is passed through the second fully connected layer to obtain the blood pressure data.

[0101] The Transformer layer of this embodiment can obtain results with relatively strong correlation of temporal features by capturing the temporal relationship between data.

[0102] In this embodiment, the random dropout process is to set some values ​​in the feature vector to 0 to reduce the overfitting of the model, but the length of the feature vector remains unchanged.

[0103] Furthermore, the first residual module includes a first depth-wise separable convolution layer and a second depth-wise separable convolution layer connected in sequence, and the input of the first depth-wise separable convolution layer and the output of the second depth-wise separable convolution layer are quickly connected as the output of the first residual module; wherein the step size of the first depth-wise separable convolution layer is 2, the convolution kernel size is 1*3 (including 1*3 depth-wise convolution DwConv, 1*1 point-wise convolution PwConv), and the padding of the signal is 1 (1 indicates padding of the left and right sides of the signal A 0 is added to both the end and the tail); the step size of the second depth-separable convolution layer is 2, the convolution kernel size is 1*3, and the padding of the signal is 1; the 64-dimensional feature vector 64*640 with a length of 640 passes through the first depth-separable convolution layer to obtain the first 128-dimensional feature vector 128*320 with a length of 320, and the first 128-dimensional feature vector 128*320 with a length of 320 passes through the second depth-separable convolution layer to obtain the second 128-dimensional feature vector 128*320 with a length of 320;

[0104] In this embodiment, the quick connection is used to build a deeper network, which improves the network feature extraction capability and reduces overfitting.

[0105] Further, the second residual module includes a third depth-separable convolution layer and a fourth depth-separable convolution layer connected in sequence, and the input of the third depth-separable convolution layer and the output of the fourth depth-separable convolution layer are quickly connected as the output of the second residual module; wherein the step size of the third depth-separable convolution layer is 1, the convolution kernel size is 1*3, and the padding of the signal is 1; the step size of the fourth depth-separable convolution layer is 1, the convolution kernel size is 1*3, and the padding of the signal is 1; the second 128-dimensional feature vector 128*320 with a length of 320 is passed through the third depth-separable convolution layer to obtain a third 128-dimensional feature vector 128*320 with a length of 320, and the third 128-dimensional feature vector 128*320 with a length of 320 is passed through the fourth depth-separable convolution layer to obtain a fourth 128-dimensional feature vector 128*320 with a length of 320;

[0106] Furthermore, the third residual module includes a fifth depth-wise separable convolution layer and a sixth depth-wise separable convolution layer connected in sequence, and the input of the fifth depth-wise separable convolution layer and the output of the sixth depth-wise separable convolution layer are shortcut-connected as the output of the third residual module; wherein the step size of the fifth depth-wise separable convolution layer is 2, the convolution kernel size is 1*3, and the padding of the signal is 1; the step size of the sixth depth-wise separable convolution layer is 1, the convolution kernel size is 1*3, and the padding of the signal is 1; the fourth 128-dimensional feature vector 128*320 with a length of 320 is passed through the fifth depth-wise separable convolution layer to obtain a first 256-dimensional feature vector 256*160 with a length of 160, and the first 256-dimensional feature vector 256*160 with a length of 160 is passed through the sixth depth-wise separable convolution layer to obtain a second 256-dimensional feature vector 256*160 with a length of 160. The depthwise separable convolution in this embodiment can achieve the same feature extraction capability as conventional convolution, but the number of parameters is significantly reduced, which speeds up the calculation.

[0107] Furthermore, in the corresponding residual module, BN layers and RELU layers are connected in sequence between the depthwise separable convolutional layers, and BN layers and RELU layers are connected in sequence between the depthwise convolution DwConv and the pointwise convolution PwConv in each depthwise separable convolutional layer. The BN layer can normalize the neural network parameters to avoid overfitting, and the RELU layer can enhance the nonlinear ability of the neural network.

[0108] Furthermore, when the deep learning model is trained by using the blood pressure data and the heartbeat waveform after data enhancement, a loss function is designed in this embodiment. The loss function formula is as follows:

[0109]

[0110] Among them, y represents the training label, represents the model output, cosh( ) represents the hyperbolic function, λ * represents the regularization weight factor, W represents the model parameter, and || ||1 represents the one-norm.

[0111] Embodiment 3

[0112] This embodiment provides a blood pressure measurement system, including:

[0113] The acquisition and noise reduction module is used to acquire the SCG signal and blood pressure data of the human body, and perform three-level noise reduction processing on the acquired SCG signal, wherein the three-level noise reduction processing includes: performing preliminary noise reduction on the SCG signal by the ICEEMDAN method to obtain a first signal; performing further noise reduction on the first signal by the RLS adaptive filter to obtain a second signal; performing three-time noise reduction on the second signal by the soft threshold method to obtain a noise-reduced SCG signal;

[0114] Training module: used to train a preset deep learning model using the blood pressure data and the denoised SCG signal to obtain a trained deep learning model;

[0115] Detection module: used to measure the blood pressure of the subject through the trained deep learning model.

[0116] Embodiment 3

[0117] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the steps of the blood pressure measurement method described in the first embodiment are implemented.

[0118] The electronic device in this embodiment may be a mobile device, such as a mobile phone.

[0119] Embodiment 4

[0120] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the blood pressure measurement method described in the first embodiment are implemented.

[0121] Those skilled in the art will appreciate that the embodiments of the present application can be provided as methods, systems, or computer program products. Therefore, the present application can adopt the form of complete hardware embodiments, complete software embodiments, or embodiments in combination with software and hardware. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that contain computer-usable program code. The scheme in the embodiments of the present application can be implemented in various computer languages, for example, object-oriented programming language Java and literal scripting language JavaScript, etc.

[0122] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to generate a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0123] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0124] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0125] Although the preferred embodiments of the present application have been described, those skilled in the art may make other changes and modifications to these embodiments once they have learned the basic creative concept. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications falling within the scope of the present application.

[0126] Obviously, the above embodiments are merely examples for clear explanation and are not intended to limit the implementation methods. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the implementation methods here. The obvious changes or modifications derived from these are still within the protection scope of the invention.

Claims

1. A blood pressure measurement method, characterized in that: include: Step S1: Collect the SCG signal and blood pressure data of the human body, and perform three-level noise reduction processing on the collected SCG signal, wherein the three-level noise reduction processing includes: performing preliminary noise reduction on the SCG signal by the ICEEMDAN method to obtain a first signal; performing further noise reduction on the first signal by the RLS adaptive filter to obtain a second signal; performing three times of noise reduction on the second signal by the soft threshold method to obtain a noise-reduced SCG signal, wherein, Collecting accelerometer data of a smartphone placed on the chest, wherein the accelerometer data is a three-axis accelerometer signal including an X-axis, a Y-axis, and a Z-axis, and collecting Y-axis data in the three-axis accelerometer signal waveform as a target SCG signal; The ICEEMDAN method is an improved complete ensemble empirical mode decomposition method with adaptive noise; The RLS adaptive filter is a recursive least squares adaptive filter; In step S1, the SCG signal is subjected to preliminary noise reduction by the ICEEMDAN method to obtain a first signal, and the method includes: Decomposing the SCG signal into several different intrinsic mode functions IMFs by using the ICEEMDAN method, selecting the intrinsic mode function IMF with the most significant periodicity of the heartbeat signal, and taking the intrinsic mode function IMF with the most significant periodicity of the heartbeat signal as the first signal; In step S1, the first signal is subjected to further noise reduction by an RLS adaptive filter to obtain a second signal, and the method includes: Define the most significant intrinsic mode function IMF of the heartbeat signal as x, and divide x into L segments, namely x1, ..., x L , calculate the i-th segment x in the L segment i and segment j x j The similarity is: Among them, μ i and μ j , δ i and δ j Respectively represent the i-th segment x i and segment j x j The mean and standard deviation of N is the signal length, k represents the signal sampling time when calculating the similarity, m represents the element offset when calculating the similarity, and mod represents the remainder; Count the number of similarities between each segment and other segments that exceed the preset threshold, and use the segment with the largest number of similarities exceeding the similarity threshold as the time reference sample; A mean square error is constructed according to x and the time reference sample, and the RLS adaptive filter performs denoising on the first signal again by minimizing the mean square error to obtain a second signal. The mean square error formula is: Where n is the signal sampling time when calculating the mean square error, y is the time reference sample, w(n) is the weight vector of the RLS adaptive filter, T is the matrix transpose, || is the absolute value, and λ(n) is the forgetting factor and satisfies: Among them, δ(n) and δ r (n) represent the standard deviation of the input signal and the time reference sample, respectively, ξ and ρ are the first constant and the second constant; Define R(n) as the inverse of the autocorrelation matrix of the input signal, expressed as: R(n) = [x·x T ] -1 ; At each signal sampling time n, the process of updating w(n) is expressed as: Calculate the prior error, expressed as: η(n) = yw T (n-1)x(n); Determine the gain vector: Update the inverse matrix: Update weight: w(n) = w(n-1) + η(n)e(n); Step S2: training a preset deep learning model using the blood pressure data and the denoised SCG signal to obtain a trained deep learning model; Step S3: measuring the blood pressure of the subject using the trained deep learning model; In step S3, the blood pressure of the subject to be measured is measured by using the trained deep learning model, and the method includes: The SCG signal of the object to be tested is collected and subjected to three-level noise reduction processing, and the noise-reduced SCG signal is used as the input signal of the trained deep learning model. Assuming that the input signal is a 1-dimensional feature vector 1*640 with a length of 640, the 1-dimensional feature vector 1*640 with a length of 640 is converted from 1 dimension to 64 dimensions through 64 1-dimensional convolution kernels to obtain a 64-dimensional feature vector 64*640 with a length of 640; The 64-dimensional feature vector 64*640 with a length of 640 is passed through the first residual module, the second residual module, and the third residual module respectively to obtain a 256-dimensional feature vector 256*160 with a length of 160, and the 256-dimensional feature vector 256*160 with a length of 160 is converted into a first 1-dimensional feature vector 1*160 with a length of 160 through 256 one-dimensional convolution kernels. The first 1-dimensional feature vector 1*160 with a length of 160 is passed through the Transformer layer to obtain a second 1-dimensional feature vector 1*160 with a length of 160. The second 1-dimensional feature vector 1*160 with a length of 160 is passed through the first fully connected layer and randomly inactivated to obtain a 1-dimensional feature vector 1*110 with a length of 110. The 1-dimensional feature vector 1*110 with a length of 110 is passed through the second fully connected layer to obtain blood pressure data; The first residual module includes a first depth-separable convolution layer and a second depth-separable convolution layer connected in sequence, and the input of the first depth-separable convolution layer and the output of the second depth-separable convolution layer are quickly connected as the output of the first residual module; wherein, a 64-dimensional feature vector 64*640 with a length of 640 is passed through the first depth-separable convolution layer to obtain a first 128-dimensional feature vector 128*320 with a length of 320, and the first 128-dimensional feature vector 128*320 with a length of 320 is passed through the second depth-separable convolution layer to obtain a second 128-dimensional feature vector 128*320 with a length of 320; The second residual module includes a third depth-separable convolution layer and a fourth depth-separable convolution layer connected in sequence, and the input of the third depth-separable convolution layer and the output of the fourth depth-separable convolution layer are quickly connected as the output of the second residual module; wherein the second 128-dimensional feature vector 128*320 with a length of 320 is passed through the third depth-separable convolution layer to obtain a third 128-dimensional feature vector 128*320 with a length of 320, and the third 128-dimensional feature vector 128*320 with a length of 320 is passed through the fourth depth-separable convolution layer to obtain a fourth 128-dimensional feature vector 128*320 with a length of 320; The third residual module includes a fifth depth-wise separable convolutional layer and a sixth depth-wise separable convolutional layer connected in sequence, and the input of the fifth depth-wise separable convolutional layer and the output of the sixth depth-wise separable convolutional layer are shortcut-connected as the output of the third residual module; wherein a fourth 128-dimensional feature vector 128*320 with a length of 320 is passed through the fifth depth-wise separable convolutional layer to obtain a first 256-dimensional feature vector 256*160 with a length of 160, and the first 256-dimensional feature vector 256*160 with a length of 160 is passed through the sixth depth-wise separable convolutional layer to obtain a second 256-dimensional feature vector 256*160 with a length of 160.

2. The blood pressure measurement method according to claim 1, characterized in that: In the step S1, the second signal is subjected to three noise reduction operations using a soft threshold method to obtain a noise-reduced SCG signal, and the method includes: Design empirical global threshold, expressed as: Where N is the signal length; δ n represents the noise level, expressed as: Among them, S n represents the value of the sequence in the second signal obtained by denoising through the RLS adaptive filter; median() represents taking the median value; For each value in the sequence of the second signal, if δ n -ψ>0, the value remains unchanged; if δ n -ψ<0, the value is 0; the second signal is subjected to three noise reduction operations to obtain a noise-reduced SCG signal.

3. The blood pressure measurement method according to claim 1, characterized in that: When the blood pressure data and the denoised SCG signal are used to train the deep learning model in step S2, the loss function formula used is: Among them, y represents the training label, represents the model output, cosh( ) represents the hyperbolic function, λ * represents the regularization weight factor, W represents the model parameter, and || ||1 represents the one-norm.

4. A blood pressure measurement system, characterized in that: include: The acquisition and noise reduction module is used to acquire the SCG signal and blood pressure data of the human body, and perform three-level noise reduction processing on the acquired SCG signal, wherein the three-level noise reduction processing includes: performing preliminary noise reduction on the SCG signal by the ICEEMDAN method to obtain a first signal; performing further noise reduction on the first signal by the RLS adaptive filter to obtain a second signal; performing three-time noise reduction on the second signal by the soft threshold method to obtain a noise-reduced SCG signal; Collecting accelerometer data of a smartphone placed on the chest, wherein the accelerometer data is a three-axis accelerometer signal including an X-axis, a Y-axis, and a Z-axis, and collecting Y-axis data in the three-axis accelerometer signal waveform as a target SCG signal; The ICEEMDAN method is an improved complete ensemble empirical mode decomposition method with adaptive noise; The RLS adaptive filter is a recursive least squares adaptive filter; The acquisition and denoising module performs preliminary denoising on the SCG signal by using the ICEEMDAN method to obtain a first signal, including: Decomposing the SCG signal into several different intrinsic mode functions IMFs by using the ICEEMDAN method, selecting the intrinsic mode function IMF with the most significant periodicity of the heartbeat signal, and taking the intrinsic mode function IMF with the most significant periodicity of the heartbeat signal as the first signal; The acquisition and noise reduction module further reduces the noise of the first signal by using an RLS adaptive filter to obtain a second signal, including: Define the most significant intrinsic mode function IMF of the heartbeat signal as x, and divide x into L segments, namely x1, ..., x L , calculate the i-th segment x in the L segment i and segment j x j The similarity is: Among them, μ i and μ j , δ i and δ j Respectively represent the i-th segment x i and segment j x j The mean and standard deviation of N is the signal length, k represents the signal sampling time when calculating the similarity, m represents the element offset when calculating the similarity, and mod represents the remainder; Count the number of similarities between each segment and other segments that exceed the preset threshold, and use the segment with the largest number of similarities exceeding the similarity threshold as the time reference sample; A mean square error is constructed according to x and the time reference sample, and the RLS adaptive filter performs denoising on the first signal again by minimizing the mean square error to obtain a second signal. The mean square error formula is: Where n is the signal sampling time when calculating the mean square error, y is the time reference sample, w(n) is the weight vector of the RLS adaptive filter, T is the matrix transpose, || is the absolute value, and λ(n) is the forgetting factor and satisfies: Among them, δ(n) and δ r (n) represent the standard deviation of the input signal and the time reference sample, respectively, ξ and ρ are the first constant and the second constant; Define R(n) as the inverse of the autocorrelation matrix of the input signal, expressed as: R(n) = [x·x T ] -1 ; At each signal sampling time n, the process of updating w(n) is expressed as: Calculate the prior error, expressed as: η(n) = yw T (n-1)x(n); Determine the gain vector: Update the inverse matrix: Update weight: w(n) = w(n-1) + η(n)e(n); Training module: used to train a preset deep learning model using the blood pressure data and the denoised SCG signal to obtain a trained deep learning model; Detection module: used to measure the blood pressure of the subject through the trained deep learning model; The detection module measures the blood pressure of the object to be measured by using the trained deep learning model, and the method includes: The SCG signal of the object to be tested is collected and subjected to three-level noise reduction processing, and the noise-reduced SCG signal is used as the input signal of the trained deep learning model. Assuming that the input signal is a 1-dimensional feature vector 1*640 with a length of 640, the 1-dimensional feature vector 1*640 with a length of 640 is converted from 1 dimension to 64 dimensions through 64 1-dimensional convolution kernels to obtain a 64-dimensional feature vector 64*640 with a length of 640; The 64-dimensional feature vector 64*640 with a length of 640 is passed through the first residual module, the second residual module, and the third residual module respectively to obtain a 256-dimensional feature vector 256*160 with a length of 160, and the 256-dimensional feature vector 256*160 with a length of 160 is converted into a first 1-dimensional feature vector 1*160 with a length of 160 through 256 one-dimensional convolution kernels. The first 1-dimensional feature vector 1*160 with a length of 160 is passed through the Transformer layer to obtain a second 1-dimensional feature vector 1*160 with a length of 160. The second 1-dimensional feature vector 1*160 with a length of 160 is passed through the first fully connected layer and randomly inactivated to obtain a 1-dimensional feature vector 1*110 with a length of 110. The 1-dimensional feature vector 1*110 with a length of 110 is passed through the second fully connected layer to obtain blood pressure data; The first residual module includes a first depth-separable convolution layer and a second depth-separable convolution layer connected in sequence, and the input of the first depth-separable convolution layer and the output of the second depth-separable convolution layer are quickly connected as the output of the first residual module; wherein, a 64-dimensional feature vector 64*640 with a length of 640 is passed through the first depth-separable convolution layer to obtain a first 128-dimensional feature vector 128*320 with a length of 320, and the first 128-dimensional feature vector 128*320 with a length of 320 is passed through the second depth-separable convolution layer to obtain a second 128-dimensional feature vector 128*320 with a length of 320; The second residual module includes a third depth-separable convolution layer and a fourth depth-separable convolution layer connected in sequence, and the input of the third depth-separable convolution layer and the output of the fourth depth-separable convolution layer are quickly connected as the output of the second residual module; wherein the second 128-dimensional feature vector 128*320 with a length of 320 is passed through the third depth-separable convolution layer to obtain a third 128-dimensional feature vector 128*320 with a length of 320, and the third 128-dimensional feature vector 128*320 with a length of 320 is passed through the fourth depth-separable convolution layer to obtain a fourth 128-dimensional feature vector 128*320 with a length of 320; The third residual module includes a fifth depth-wise separable convolutional layer and a sixth depth-wise separable convolutional layer connected in sequence, and the input of the fifth depth-wise separable convolutional layer and the output of the sixth depth-wise separable convolutional layer are shortcut-connected as the output of the third residual module; wherein a fourth 128-dimensional feature vector 128*320 with a length of 320 is passed through the fifth depth-wise separable convolutional layer to obtain a first 256-dimensional feature vector 256*160 with a length of 160, and the first 256-dimensional feature vector 256*160 with a length of 160 is passed through the sixth depth-wise separable convolutional layer to obtain a second 256-dimensional feature vector 256*160 with a length of 160.

5. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the blood pressure measurement method according to any one of claims 1 to 3 are implemented.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the blood pressure measurement method according to any one of claims 1 to 3 are implemented.

Images