Noninvasive blood pressure prediction method based on signal dimension expansion and lightweight network

Abstract

The invention discloses a non-invasive blood pressure prediction method based on signal dimension extension and a lightweight network, which comprises the following steps: firstly collecting PPG blood pressure data and ABP data, then denoising the collected PPG blood pressure data and ABP data, and cleaning the denoised data. The processed data is used for data signal dimension expansion, a PPG blood pressure measurement model is constructed, and the PPG blood pressure measurement model comprises a basic convolution layer and a residual grouping convolution module Group ResidualBock; and then deploying the constructed PPG blood pressure measurement model into the wearable device to obtain the wearable blood pressure monitoring device. And finally, measuring the processed PPG signal by using wearable blood pressure monitoring equipment to obtain a blood pressure value. According to the method, the PPG signal dimension is expanded, the data processing link is optimized, and the model blood pressure prediction accuracy is improved; and meanwhile, a lightweight algorithm architecture is adopted, so that resource occupation is reduced, the system can be deployed in wearable embedded equipment, and the security of personal privacy data is guaranteed.

Description

Technical Field

[0001] The present invention relates to the technical field of digital signal processing, and in particular to a non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network. Background Art

[0002] Hypertension is a common cardiovascular disease and one of the top risk factors for death. Long-term hypertension can cause damage to the heart, brain, kidneys and other organs, and lead to coronary heart disease, myocardial infarction and other diseases. Therefore, it is important to always understand your blood pressure and take early measures to prevent and control hypertension, which is of great significance to the prevention of cardiovascular diseases.

[0003] At present, there are two methods for measuring blood pressure: invasive blood pressure (ABP, aortic pressure) measurement and non-invasive blood pressure measurement. Invasive measurement methods measure accurate data, but they can cause damage to the human body and easily lead to a series of risks such as bleeding and infection. Non-invasive blood pressure measurement is safer. Currently, there are mainly arterial tension method, volume compensation method, pulse wave velocity measurement method, photoplethysmography method, etc. Among them, photoplethysmograph (PPG) sensors are small and low-cost, and are often used in wearable devices to measure blood pressure by analyzing human PPG signals.

[0004] In order to realize continuous blood pressure monitoring based on PPG signals, some scholars use neural network models to extract important pulse signal parameters related to blood pressure monitoring, such as PWV (Pulse Wave Velocity), PTT (Pulse Transit Time) and PAT (Pulse Arrival Time), and establish a linear regression model by linking them with the continuous changes in blood pressure waveform characteristics to realize the prediction of blood pressure values. However, when the feature vector and the target have a strong nonlinear relationship, the linear regression model is not applicable.

[0005] Since the linear relationship of the traditional linear regression model is not obvious, deep neural networks have unique advantages in portable non-invasive blood pressure prediction without causing damage to the human body. However, traditional deep neural networks consume large resources such as memory, consume high cost resources, and are difficult to deploy. In addition, due to the lack of samples, the selection of feature parameters is subjective and prone to overfitting. Secondly, the processing of PPG blood pressure data needs to be optimized. At the same time, by reducing data transmission and implementing strict access control, the risk of data leakage is reduced, and decentralization reduces dependence on a single data center, further reducing the risk of attack. Summary of the invention

[0006] The purpose of the present invention is to overcome the above-mentioned shortcomings and deficiencies of the prior art and provide a non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network. The present invention is applicable to wearable devices.

[0007] The present invention expands the one-dimensional PPG signal into multi-dimensional data by combining time and optimizes the data processing link, thereby improving the accuracy of model blood pressure prediction and reducing overfitting. At the same time, a lightweight algorithm architecture is adopted to reduce resource usage so that it can be deployed in wearable embedded devices.

[0008] The present invention is achieved through the following technical solutions:

[0009] A non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network includes the following steps:

[0010] S1: Collect PPG blood pressure signals and ABP data;

[0011] S2: De-noising the collected PPG blood pressure data and ABP data, and performing data cleaning on the de-noised data;

[0012] S3: Expand the dimension of PPG blood pressure signals and build a deep learning model for PPG blood pressure monitoring using processed PPG blood pressure data and ABP data;

[0013] S4: Deploy the constructed PPG blood pressure monitoring deep learning model into the wearable device to obtain a wearable blood pressure monitoring device;

[0014] S5: Use a wearable blood pressure monitoring device to detect the processed PPG data to obtain a blood pressure value.

[0015] In step S2, the collected blood pressure PPG data and ABP data are denoised, including using a Butterworth bandpass filter with a frequency range of 0.5-10 Hz to remove the high-frequency noise and baseline drift of the PPG data and using a low-pass filter with a cutoff frequency of 5 Hz to process the high-frequency noise of the aortic pressure data.

[0016] In step S2, the denoised data is cleaned, including: dividing PPG and ABP into 8s windows with a step size of 2s, resulting in a 6s overlap between two consecutive windows, obtaining the time series of PPG and ABP, and the maximum and minimum values ​​in each window of the aortic pressure signal (ABP) are the SBP (Systolic Blood Pressure) and DBP (Diastolic Blood Pressure) of the interval, respectively. PPG data and ABP data that do not meet the preset requirements are filtered or deleted.

[0017] Step S3 expands the dimension of the PPG blood pressure signal, including further calculating the first-order gradient information, second-order gradient information, and third-order gradient information from the original one-dimensional PPG data in each window (as shown in the figure) using the intermediate difference method, and using Z-score normalization to limit the data size while retaining the trend information and zero point information, and merging the gradient information of each node into a multi-dimensional data matrix.

[0018] Furthermore, the gradient is calculated using the central difference method:

[0019]

[0020]

[0021]

[0022] Where T represents the signal value at a specific time point;

[0023] t represents the measurement time interval;

[0024] O represents truncation error;

[0025] The ABP data of the corresponding time period is used for deep network training. The PPG blood pressure monitoring deep learning model described in step S3 includes an input layer, a convolution layer, a pooling layer, and a fully connected layer. The input layer is used to expand the dimension of the PPG signal and receive the input PPG blood pressure data. The convolution layer uses the focus residual group convolution module GroupResidualBock to extract features from the input data. The pooling layer is used to extract preliminary features from the input PPG blood pressure data. The fully connected layer is used to perform numerical prediction on the processed feature map and output the blood pressure value.

[0026] Before step S5, the method further includes: determining whether the quality of the cleaned data meets the requirements.

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

[0028] (1) The present invention uses the central difference method to expand the dimension of the PPG signal, uses Z-score normalization, limits the data size while retaining trend information and zero point information, and merges it into a multidimensional data matrix. Compared with before processing, the prediction accuracy of the PPG blood pressure monitoring deep learning model using this method is higher.

[0029] (2) The present invention uses grouped convolution and depthwise separable convolution to reduce the parameters of the entire model. Grouped convolution divides the input channels into multiple groups, and then performs convolution operations on each group separately. Since the convolution kernel is only calculated within the corresponding group instead of operating on all channels simultaneously, the number of convolution kernels is reduced; depthwise separable convolution decomposes the convolution process into two steps: depthwise convolution and pointwise convolution. Depthwise convolution performs convolution operations on each input channel separately, and its convolution kernel size is the same as the number of input channels, which greatly reduces the number of parameters. Pointwise convolution is used to combine the output results of depthwise convolution, and its number of parameters is relatively small. The number of parameters of the entire model is only 800,000, which is convenient for deployment on wearable embedded devices, greatly reducing the difficulty of deployment.

[0030] (3) In the method provided by the present invention, the PPG signal data is processed locally, reducing the need to transmit data to a remote server, thereby reducing the risk of data leakage, so as to implement a stricter access control policy to ensure that only authorized users can access or process the data. At the same time, its decentralized nature reduces the dependence on a single data center, thereby reducing the risk of data leakage caused by attacks on the data center.

[0031] (4) The present invention adopts the method of channel cleaning and residual combination to ensure the accuracy of the model, while alleviating the problem of gradient disappearance during training and speeding up the network training. In order to better train the network, the Hswish activation function is used, which has better gradient propagation and faster convergence speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 This is a flow chart of blood pressure measurement in a specific embodiment of the present invention.

[0033] Figure 2 It is an overall network structure diagram in a specific embodiment of the present invention.

[0034] Figure 3 It is a schematic diagram of the GroupResidualBock module in a specific embodiment of the present invention.

[0035] Figure 4 It is a specific application flow chart of the blood pressure measurement model constructed in a specific embodiment of the present invention.

[0036] Figure 5 Schematic diagram of a standard convolution kernel in a specific embodiment of the present invention.

[0037] Figure 6 It is a loss value image during the training process in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0038] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only provided for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0039] The present invention discloses a non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network, which can be implemented by the following steps:

[0040] S1, collect PPG blood pressure signals and ABP (aortic pressure) data;

[0041] S2: De-noising the collected PPG blood pressure data and ABP data, and performing data cleaning on the de-noised data;

[0042] S3: Expand the dimension of PPG blood pressure signals and build a deep learning model for PPG blood pressure monitoring using processed PPG blood pressure data and ABP data;

[0043] S4: Deploy the constructed PPG blood pressure monitoring deep learning model into the wearable device to obtain a wearable blood pressure monitoring device;

[0044] S5: Use a wearable blood pressure monitoring device to detect the processed PPG data to obtain a blood pressure value.

[0045] The overall process of blood pressure measurement is as follows Figure 1 As shown, after signal denoising, the corresponding PPG waveform is drawn on the client side for the processed PPG signal, and the Pearson correlation coefficient can be used to determine whether the data quality is good. For suitable PPG data, the signal dimension is expanded to improve the accuracy of the blood pressure model in predicting blood pressure. Finally, after the expanded data is tested using a wearable embedded blood pressure monitoring system, the blood pressure value and classification results are displayed on the client side.

[0046] The step S1 uses a PPG sensor probe and a blood pressure sensor to collect PPG blood pressure signals and ABP signals.

[0047] In step S2, the collected PPG blood pressure data and ABP data are denoised, including: using a Butterworth bandpass filter with a frequency range of 0.5-10 Hz to remove high-frequency noise and baseline drift of the PPG data, and using a low-pass filter with a cutoff frequency of 5 Hz to process the high-frequency noise of the aortic pressure data.

[0048] Specifically, the voltage signal output by the I / V conversion circuit is accompanied by a large number of noise signals, including: baseline drift, power frequency interference, etc. The effective frequency of the human PPG signal is usually concentrated between 0Hz and 10Hz. The baseline drift is generally caused by breathing, and its frequency is generally lower than 0.5Hz. Power frequency interference mostly comes from the working current, and the noise signal frequency is generally around 50Hz. For the aortic pressure signal, it can be considered as high-frequency noise above 5Hz. Therefore, this embodiment uses a Butterworth bandpass filter with a frequency range of 0.5-10Hz to simultaneously remove high-frequency noise and low-frequency baseline drift in the PPG signal, and uses a low-pass filter with a cutoff frequency of 5Hz to remove high-frequency noise from the ABP signal.

[0049] Step S2 performs data cleaning on the denoised data, including: dividing PPG and ABP into 8s windows with a step size of 2s, resulting in a 6s overlap between two consecutive windows, obtaining a time series of the PPG signal, and the maximum and minimum values ​​in each window of the aortic pressure signal (ABP) are the SBP and DBP of the interval, respectively. PPG data and ABP data that do not meet the preset requirements are filtered or deleted.

[0050] Data that does not meet the preset requirements include data with a duration less than the preset duration, data with missing or identical preset ratio lengths, and data with values ​​outside the set range. In this embodiment, according to the ABP data situation, abnormal blood pressure values ​​(SBP≥200, DBP≥120, SBP≤80 and DBP≤40) and the corresponding PPG blood pressure signals are eliminated. According to the PPG blood pressure data situation, the preset duration is 10 minutes and the preset ratio length is 5%. Therefore, PPG blood pressure data with a length of less than 10 minutes is filtered out. If there is data with a length of 5% that is missing or identical, delete the data. If there is a problem with some cycles, delete the PPG blood pressure data and the corresponding blood pressure signal for that cycle.

[0051] Step S3 described in expanding the dimension of the PPG blood pressure signal includes further converting the original one-dimensional data of the PPG in each window (such as Figure 2 ) The first-order gradient information, second-order gradient information, and third-order gradient information are calculated using the middle difference method, and Z-score normalization is used to limit the data size while retaining the trend information and zero-point information, and the gradient information of each node is merged into a multi-dimensional data matrix.

[0052] First, it is necessary to further optimize the original one-dimensional PPG data in each window, and the intermediate difference method is used to calculate the first-order, second-order and third-order derivative information. Since the calculated derivative information has significant numerical differences, the Z-score normalization method is used for normalization. This process not only retains the trend and zero-point characteristics of the data, but also effectively controls the magnitude of the data. Finally, these normalized gradient information are spliced ​​at each node to construct a multidimensional data matrix, which is then input into the PPG blood pressure monitoring deep learning model, and the corresponding processed ABP data is used for training, and the training batch is selected as 5. After the training is completed, it can be found by testing with the test set that the final error is very small, and the MAE (mean error) is only less than 0.02;

[0053] Step S3 expands the dimension of the PPG blood pressure signal, which includes the following mathematical knowledge: suppose the input is x, the output is y, and the entire model is a function Φ, then y=Φ(x). As we know, blood pressure information is included in the trend of the PPG signal, so Φ must contain a gradient extraction operation. Get gradient information Then perform feature operations on the gradient information. This will waste the training process and parameter utilization. Therefore, we pre-extract the gradient information and then perform feature extraction on the whole. That is Therefore, we compare the results under the two modes and compare the advantages and disadvantages of expanding the performance dimension.

[0054] The PPG blood pressure monitoring deep learning model described in step S3 is the DESNet model, which includes a unique dimension expansion convolution module (whose parameters remain fixed), together with a standard convolution layer and a residual grouping convolution module. The dimension expansion convolution module is to expand the signal dimension, and the standard convolution is to extract the standard feature map. The residual grouping convolution module is based on the depth separable convolution, and the number of channels of the feature map is divided into 2 parts, one of which is first convolved and then convolved by depth separable convolution, and the other is directly merged with the other group when the step size is 1; when the step size is 2, a depth separable convolution is performed and merged with the other group. The merging is used to achieve residual addition Residual to prevent the problem of gradient disappearance during the training process. Finally, the merged feature matrix is ​​channel cleaned and mixed, and the channels are rearranged to enhance the information flow between different groups, which helps to reduce the correlation within the group and improve the network expression ability.

[0055] After the data processing stage, the processed blood pressure PPG data matrix is ​​input into the PPG blood pressure monitoring deep learning model. The standard convolution layer first extracts the multi-channel feature matrix from the PPG data matrix, in which the convolution kernel (3, 7) is used. The reason is to ensure the integrity of the information. We can know that before the aortic valve closes, due to the appearance of the pressure gradient opposite to the blood flow direction, the blood volume of the capillaries temporarily increases again, and there will be a brief rise and fall in the diastolic PPG waveform. At this time, the concave point is defined as a bidirectional concave trace, and the point where the first-order derivative of the waveform is closest to zero after the systolic peak is defined as the diastolic peak. This means that the trend change of the PPG signal contains information about blood pressure, so the gradient information dimension is expanded to better extract information and predict blood pressure.

[0056] The gradient is calculated using the central difference method:

[0057]

[0058]

[0059]

[0060] It can be seen that it is a second-order error about time. When the time is very small, the error can be ignored. At the same time, from the formula of central difference, it can be seen that at any point, the length of the maximum strongly correlated term is 7, so the convolution kernel of (3, 7) is used.

[0061] The role of the convolution layer is to learn and extract useful features from the input data. By performing convolution operations with the convolution kernel and the input data, the patterns and features of the local area can be captured. In the entire DESNet model, a total of one dimension expansion convolution, two standard feature map extraction standard convolutions, and 16 group convolutions in the GroupResidualBock module are performed. In the group convolution, the group convolution rules are designed according to the step size. The step size is 1, and two depth-separable convolutions and two standard convolutions are performed, and the residuals are finally merged; the step size is 2, and three depth-separable convolutions and two standard convolutions are performed, and the residuals are also merged at the end.

[0062] The pooling layer uses average pooling to ensure more relevant information. The activation function uses the Hswish activation function, which helps to improve the efficiency and speed of the model in applications on lightweight networks and mobile devices. The above process is mainly used to extract preliminary features of the input PPG blood pressure data.

[0063]

[0064] The initially extracted feature matrix is ​​input into the residual group convolution module GroupResidualBock, and convolution and channel cleaning are repeated in the module to enhance the relevance of information. Finally, it is fully connected and output as a unique value as the predicted blood pressure values ​​SBP and DBP.

[0065] Before step S5, the method further includes: determining whether the quality of the cleaned data meets the requirements.

[0066] The training process of the entire model uses the mean square error (MSE). Predicting blood pressure values ​​is a regression-like problem. The MSE loss function can amplify the penalty, and the positive error and negative error contribute equally to the loss.

[0067] The training batch size is set small during training, because MSE is a loss function. When the training batch size is large, the adjustment of parameters is small, and it is difficult to converge. It is set to 5. Training and prediction are performed on an independent personal data set. The test set is used for testing and it is found that the standard error MAE reaches 0.016329091, the mean square error MSE reaches 0.00054526, and the standard error MAE without dimensional expansion reaches 0.059687917, and the mean square error MSE reaches 0.004190504.

[0068] If used as a general model and trained and predicted on the public dataset MIMIC-II (M.Kachuee,MMKiani,H.Mohammadzade,M.Shabany,Cuff-Less High-Accuracy Calibration-Free BloodPressure Estimation Using Pulse Transit Time,IEEE International Symposium onCircuits and Systems(ISCAS'15),2015.), the standard error MAE of diastolic blood pressure reaches 7.23445834, the root mean square error RMSE reaches 10.29902335, the standard error MAE of systolic blood pressure reaches 13.18022772, and the root mean square error RMSE reaches 16.78072161. For the general convolutional neural network model (CNN), the standard error MAE for diastolic blood pressure is 12.38, and the standard error MAE for systolic blood pressure is 15.41 (reference https: / / doi.org / 10.1016 / j.bspc.2022.103891).

[0069] The blood pressure monitoring model constructed by the present invention and the specific application flow chart are as follows: Figure 4 As shown, the present invention provides a lightweight neural network based on DESNet.

[0070] Based on this neural network, combined with the PPG signal expansion method, a more accurate blood pressure prediction is provided. First, the PPG signal data transmitted through the PPG sensor and the ABP data transmitted by the blood pressure sensor are cleaned and filtered to reduce the impact of data noise on the model accuracy, and a Butterworth bandpass filter with a frequency range of 0.5-10Hz is used to remove high-frequency noise and low-frequency baseline drift in the PPG signal at the same time, and a low-pass filter with a cutoff frequency of 5Hz is used to remove high-frequency noise in the ABP signal; then the Pearson correlation coefficient is used to judge the quality of the collected PPG signal. If the coefficient is less than 0.75, it is considered that the signal quality is low and the signal needs to be re-collected. If the coefficient is greater than 0.75, it is considered that the quality of the collected PPG signal meets the requirements; then the first-order gradient information, second-order gradient information, and third-order gradient information are calculated for the PPG signal that meets the requirements using the intermediate difference method, and the Z score is normalized to merge the gradient information of each node into a multi-dimensional data matrix. Finally, the blood pressure of the collected PPG signal is predicted by the DESNet blood pressure monitoring model deployed on the wearable embedded device.

[0071] according to Figure 6 As shown, loss indicates that dimension expansion is used, and loss2 indicates that dimension expansion is not used. In the training process of the specific embodiment, the loss value of the training result using signal dimension expansion processing is lower than the loss value of the training result not using signal dimension processing. It can be seen that signal dimension expansion improves the accuracy of blood pressure prediction.

[0072] The present invention improves the prediction accuracy of the blood pressure of the neural network blood pressure monitoring model by expanding the dimension of the PPG signal; at the same time, a lightweight neural network model is constructed to make it suitable for deployment in wearable embedded devices, which can provide all-weather blood pressure monitoring.

[0073] The above embodiments are preferred implementation modes of the present invention, but the implementation modes of the present invention are not limited to the embodiments. 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 protection scope of the present invention.

Claims

1. A non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network, characterized in that The steps include: S1: Collect PPG blood pressure signals and ABP data; S2: De-noising the collected PPG blood pressure data and ABP data, and performing data cleaning on the de-noised data; S3: Expand the dimension of PPG blood pressure signals and build a deep learning model for PPG blood pressure monitoring using processed PPG blood pressure data and ABP data; S4: Deploy the constructed PPG blood pressure monitoring deep learning model into wearable devices; S5: Use a wearable blood pressure monitoring device to detect the processed PPG data to obtain a blood pressure value.

2. The non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network according to claim 1, characterized in that: In step S2, the collected blood pressure PPG data and ABP data are denoised, including: using a Butterworth bandpass filter with a frequency range of 0.5-10 Hz to remove the high-frequency noise and baseline drift of the PPG data, and using a low-pass filter with a cutoff frequency of 5 Hz to process the high-frequency noise of the aortic pressure data.

3. The non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network according to claim 1 is characterized by: In step S2, the denoised data is cleaned, including: dividing PPG and ABP into 8-s windows with a step size of 2s, resulting in a 6-s overlap between two consecutive windows, and obtaining the time series of PPG and ABP. The maximum and minimum values ​​in each window of the aortic pressure signal are the SBP and DBP in that interval, respectively.

4. The non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network according to claim 1, characterized in that: In step S3, the dimension of the PPG blood pressure signal is expanded, including using the intermediate difference method to calculate the first-order gradient information, second-order gradient information, and third-order gradient information of the original one-dimensional PPG data in each window, and using Z-score normalization to limit the data size while retaining the trend information and zero point information, and merging the gradient information of each node into a multi-dimensional data matrix.

5. The non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network according to claim 4 is characterized by: The calculation formula for calculating the gradient uses the central difference method: Where T represents the signal value at a specific time point; t represents the measurement time interval; O represents the truncation error.

6. The non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network according to claim 1, characterized in that: In step S3, the PPG blood pressure monitoring deep learning model includes an input layer, a convolutional layer, a pooling layer, and a fully connected layer; The input layer is used to expand the dimension of the PPG signal and receive the input PPG blood pressure data; The convolution layer uses the focused residual group convolution module GroupResidualBock to extract features from the input data; The pooling layer is used to extract preliminary features of the input PPG blood pressure data; The fully connected layer is used to make numerical predictions on the processed feature graph and output the blood pressure value.

7. The non-invasive blood pressure prediction method based on signal dimension expansion and lightweight network according to claim 1, characterized in that: In step S5, before using the wearable blood pressure monitoring device to detect the processed PPG data, it is necessary to determine whether the quality of the cleaned data meets the requirements.

Images