A cuffless blood pressure prediction system based on deep neural network model

Abstract

The present invention discloses a cuffless blood pressure prediction system based on a deep neural network model, the system is used to perform the following steps: receiving a data set including a pulse wave signal and an electrocardiogram signal; preprocessing the received pulse wave signal and electrocardiogram signal to obtain a usable signal; the preprocessing includes signal filtering, fixed-length cutting, peak cutting / trough cutting, resampling and obtaining label values; inputting the usable signal into a deep neural network model to obtain predicted systolic and diastolic pressure values; the deep neural network model includes a convolutional adaptation layer, a residual network with a compression excitation module and a fully connected layer connected in sequence; the input of the convolutional adaptation layer is the usable signal; the output of the fully connected layer is the systolic and diastolic pressure values. The present invention realizes cuffless prediction of continuous dynamic blood pressure values ​​by analyzing the pulse wave signal and the electrocardiogram signal.

Description

Technical Field

[0001] The present invention belongs to the field of biological signal processing, and in particular relates to a cuffless blood pressure prediction system based on a deep neural network model. Background Art

[0002] Hypertension is the main factor leading to most cardiovascular diseases. Blood pressure assessment and dynamic detection are of great significance for timely understanding of the incidence of hypertension. In the study of blood pressure estimation, the existing blood pressure measurement with a cuff cannot obtain continuous blood pressure values. Therefore, the continuous blood pressure estimation technology based on cuffless measurement has become a current technical research hotspot. A large number of studies have shown that the combined input of photoelectric plethysmography and electrocardiogram signals can effectively predict ambulatory blood pressure. The pulse arrival time is defined as the time delay from the peak of the electrocardiogram signal to the next valley of the pulse wave within a specific time. Pulse wave velocity, pulse transit time, and pulse arrival time are also used to establish complex linear or nonlinear relationships with blood pressure to calculate ambulatory blood pressure values.

[0003] Existing techniques include direct linear regression of blood pressure and pulse wave transmission time or pulse wave transmission velocity. This direct method often results in large prediction errors due to inaccurate positioning of specific points on the waveform. Although using filtered signals can alleviate this situation to a certain extent, the performance of this method is often poor because simple linear and nonlinear models are not sufficient to evaluate complex blood pressure changes.

[0004] In the application of machine learning in this field, early experimental conclusions found that using artificially calibrated feature points with medical significance as model input can often achieve better prediction accuracy. However, since the feature point extraction positions need to be manually calibrated and certain knowledge of biomedical signals is required, this method is too expensive.

[0005] Currently, in the field of cuffless blood pressure estimation, the most widely used combination structure is a model based on feature extraction and time correlation, such as a combination of fully connected / convolutional neural network-recurrent neural network / long short-term memory network (single / bidirectional). Due to the feature input based on the previous and next order and the time series correlation, this model combination is easily affected by the instability of the waveform, resulting in inaccurate predictions, and the sample capacity of the test is often relatively low. Summary of the invention

[0006] In order to solve the problem that cuffless blood pressure estimation has a low sample capacity due to reliance on a timing network and is susceptible to overall waveform baseline drift and high-frequency noise, the present invention provides a cuffless blood pressure prediction system based on a deep neural network model, which can stably and reliably estimate dynamic blood pressure values ​​through pulse wave signals and electrocardiogram signals on a data set containing a large number of people.

[0007] The present invention adopts the following technical solution:

[0008] A cuffless blood pressure prediction system based on a deep neural network model, the system is used to perform the following steps:

[0009] Step 1: signal preprocessing.

[0010] A. Construction of training data set:

[0011] The data of the subjects are extracted from the database, which contains a large amount of waveform information. Since the hemodynamic information of each sample is unique, the observation and analysis of the cycle morphology also confirms that the same blood pressure value may correspond to different waveform information. Therefore, each subject will provide part of the data for training. This classification method ensures that the model can extract enough waveform feature information from each subject.

[0012] B, signal preprocessing:

[0013] Since pulse waves and ECG signal waveforms contain more high- and low-frequency noises, and the signal cycles intercepted by each person within a specific time are different, biological signals need to be preprocessed.

[0014] Step B1, signal filtering, includes filtering out high-frequency noise and low-frequency noise (baseline drift).

[0015] Step B2: cutting the signal into fixed length.

[0016] Step B3, performing peak cutting and resampling on the fixed-length cutting signal in step B2.

[0017] Step B4, taking the peak cutting signal range in step B3 to calculate the corresponding blood pressure value as the model label.

[0018] In the above step B1, the filtering method can be a combination of low-pass and high-pass filtering methods, or a band-pass filtering method can be directly used. In the process of further filtering out noise, the remaining noise can be further filtered by wavelet transform or smoothing filtering to make the signal smoother.

[0019] In the above step B2, it is necessary to select a suitable fixed-length window size to cut the waveform signal. The specific window size is determined by the number of resampling. Usually, a window with non-overlapping fixed points has a better effect.

[0020] In the above step B3, the fixed-length signal obtained from B2 is subjected to peak cutting once for each fixed number of peaks, so that two groups of signal segments with a fixed number of peaks (respectively pulse wave and electrocardiogram signal) are obtained each time.

[0021] In the above step B4, the peak range calibrated in step B3 is taken. Since the peaks of the pulse wave and the ECG signal may not coincide, the range should be set to the union of the peak ranges in step B3. In this union, the peaks and troughs of the dynamic blood pressure values, i.e., systolic pressure and diastolic pressure, are taken as the label values ​​of the model.

[0022] Step 2: Construction of deep neural network model:

[0023] The model is mainly divided into three parts, including the convolutional neural network-adaptation layer module (generating input), the residual network module with a compression excitation module (feature extraction), and the fully connected layer (generating output). Since the convolution receptive field and parameters can be adjusted through the convolution adaptation layer, this composite model structure can effectively utilize the residual network enhancement model with a compression excitation module to extract features in real-time signals, and can adjust the appropriate number of layers to adapt to the complexity of the signal.

[0024] C. Convolutional neural network-adaptation layer module (generating input):

[0025] The residual network structure is generally used to process input signals with the same length and width, so the input needs to adapt to a fixed format. The convolution adaptation layer used in this design mainly includes the convolution layer, batch normalization layer, and maximum pooling layer. The convolution kernel size of the adaptation layer needs to be adjusted according to the length of the feature points of the pulse wave and ECG signal. The specific size is determined by the resampling point, that is, the receptive field should contain the feature points of these two signals as much as possible.

[0026] D. Residual network structure with compression excitation module (feature extraction):

[0027] The residual network was originally designed to handle image classification problems, so the adaptation layer in step C is required to produce a suitable output format. The compressed excitation module is not a complete network structure, but an attention mechanism that can improve the interdependence between channels and is an effective way to improve the feature extraction ability of the convolutional neural network module. Using the residual network structure with a compressed excitation module, the channel weights based on global information are redistributed, which selectively enhances useful information and suppresses useless information.

[0028] E. Fully connected layer (producing output):

[0029] The output generated by the feature extraction layer cannot be used directly for error calculation. The output needs to be further integrated, and the feature data generated by step D is integrated using full connection.

[0030] In the above step E, the output dimension of the full connection is 2, corresponding to the systolic and diastolic blood pressure values ​​respectively.

[0031] Step 3: Analysis of model training results

[0032] The training results of the model are evaluated by comparing the model predictions with the actual blood pressure measurements to achieve the required prediction accuracy.

[0033] The beneficial effects of the present invention are as follows:

[0034] (1) The present invention proposes a new signal preprocessing method. The specific implementation method is to use fixed-length cutting first and then peak cutting / trough cutting. The advantage of fixed-length cutting first is that it can effectively prevent the influence of abnormal points on the entire signal. Then, the peak cutting / trough cutting method is used to further process the fixed-length signal. Each available signal segment is a periodic signal containing a fixed number of peaks. This combination of available signals is used as the final input of the model.

[0035] (2) The present invention builds a model framework based entirely on real-time waveform information, and independently processes and analyzes the cut input signal segments, that is, completely ignores the up-down temporal correlation of different signal segments, so that the output result is not easily affected by the noise in the biological signal;

[0036] (3) The present invention proposes to use a deep neural network model to predict cuffless dynamic blood pressure. The deep neural network model adopts a convolutional adaptation layer, a residual network with a compression excitation module and a fully connected layer network structure to further improve the accuracy of cuffless blood pressure measurement; in the convolutional adaptation layer, the receptive field is adjusted to enable the model to fully extract the feature points of the pulse wave and electrocardiogram signal; the extracted feature information is further sent to the residual network structure with a compression excitation module. This design can not only effectively prevent overfitting of the input signal by flexibly adjusting the number of model layers, but also enhance the model's ability to extract features, so that it does not rely on timing information, but only relies on the current input waveform information to perform stably in a data set containing a large number of subjects.

[0037] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments, but the cuffless blood pressure prediction system based on a deep neural network model of the present invention is not limited to the embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 It is a processing flow chart of the execution steps of the cuffless blood pressure prediction system based on the deep neural network model according to an embodiment of the present invention;

[0039] Figure 2It is a framework flow chart of a cuffless blood pressure prediction system based on a deep neural network model according to an embodiment of the present invention;

[0040] Figure 3 The graphs are comparisons of the correlation results between the predicted values ​​and the true values ​​of systolic and diastolic pressures of the neural network model of the embodiment of the present invention; wherein (a) represents the comparison of the correlation results between the predicted value and the true value of systolic pressure; (b) represents the comparison of the correlation results between the predicted value and the true value of diastolic pressure;

[0041] Figure 4 The box plots of the predicted values ​​of systolic and diastolic blood pressure and the true values ​​of the neural network model of the embodiment of the present invention are compared; wherein (a) represents the box plot comparison of the predicted value of systolic blood pressure and the true value; (b) represents the box plot comparison of the predicted value of diastolic blood pressure and the true value; the horizontal line in the box is the median in the statistical data, and the cross mark is the outlier point in the data. DETAILED DESCRIPTION

[0042] The present invention is further described below through specific implementation modes. It should be noted that the specific embodiments described herein are only used to facilitate the description and explanation of the specific implementation modes of the present invention, and are not used to limit the present invention.

[0043] In order to make the purpose and technical solution of the present invention clearer, the present invention is further described below in conjunction with the accompanying drawings and examples. It should be understood that the examples described herein are only used to explain the present invention and are not used to limit the present invention.

[0044] The data set of the embodiment of the present invention is collected from the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC-II) database, which includes data sets of pulse wave signals and electrocardiogram signals collected by wearable devices and / or medical monitoring devices.

[0045] See also Figure 1 and Figure 2 As shown, the present invention provides a cuffless blood pressure prediction system based on a deep neural network model, and the system is used to perform the following steps:

[0046] Step 1: signal preprocessing.

[0047] A. Construction of the dataset:

[0048] The dataset is divided into a training set and a test set. A dataset containing 1,200 people is extracted from the database. The dataset of each subject is long enough to meet the requirements of the training set and the test set. Both the training set and the test set contain the data of each person, but these data do not overlap. The dataset of each subject is divided into 70% and 30% segments and put into the training set and the test set respectively.

[0049] The present invention trains the deep neural network model through the training set, and after the training is completed, the test set is verified based on the trained deep neural network model.

[0050] B, Pulse wave / ECG signal preprocessing.

[0051] Before deep learning, the signal is preprocessed by filtering, cutting to fixed length, cutting to peak value, etc. The specific steps are as follows:

[0052] Step B1, low-pass and high-pass combined filtering is performed on the original signal waveform, and then wavelet transform is used to make the signal smoother to prevent noise from affecting the learning of the model.

[0053] Step B2, extract training signals from the sample set, and perform signal cutting every 1000 fixed-length sampling points. The original sampling rate is 125Hz, and this time the scheme of cutting every 1000 points is adopted, so each cutting represents an 8-second segment length of the pulse wave and ECG signal. If cutting every 2000 points corresponds to 16 seconds, and so on.

[0054] Step B3, in step B2, there are continuous peaks corresponding to the r peak of the ECG signal and the systolic peak of the pulse wave. A section of the signal is extracted for the next preprocessing stage for every three peaks. The number of peaks taken in the signal is variable, such as 3, 4, 5, etc. The signal is then resampled to 512 points, and the resampling points can also be adjusted arbitrarily according to the input of the model, such as 512, 1024, etc., and then normalized.

[0055] Step B4, using the maximum range of the pulse wave and the ECG signal within a cycle to extract the boundary of dynamic blood pressure. Where x1…xn represents the sampling point range of the pulse wave, and z1…zn represents the sampling point range of the ECG signal. The extraction boundary of dynamic blood pressure is [min(x1, z1), max(xn, zn)]

[0056] In the above step B2, the reason for first performing fixed-length segmentation is to prevent global peak extraction errors caused by individual abnormal points. Using this cutting method can minimize the impact of abnormal points.

[0057] In the above step B3, the number of selected peaks during peak cutting is determined manually, and a trough cutting method may also be selected to pre-process the signal.

[0058] In the above step B4, this dynamic blood pressure extraction method can ensure the one-to-one correspondence between the label value and the input signal as much as possible.

[0059] Step 2: Construction of deep neural network model

[0060] C. Convolutional neural network-adaptation layer module (generating input):

[0061] The specific parameters of the convolution layer are determined by the resampling points. Since 512 resampling points are used, which contain three peak information, in order to ensure that each peak is included in the convolution receptive field as much as possible. The number of available signals input to the model each time is 32. The input signal is regarded as a two-dimensional sequence on a channel. The convolution window is 64*2, the step size is 2, the input channel is 1, and the output channel is 57. The maximum pooling convolution kernel is 3*3, and the step size is 2. The final input size fed to the next layer of the network is the batch size (batchsize)*1*114*114. The activation function is ReLu.

[0062] In the above step C, the output channel of the convolution adaptation layer or the size of the maximum pooling layer may change due to setting different resampling points. The activation function can be replaced by sigmoid, tanh, etc.

[0063] D. Residual network with compressed excitation module (feature extraction):

[0064] The convolutional blocks of the residual network are implemented with 1*1, 3*3, and 1*1 structures. These modules are all equipped with a compression excitation module structure to enhance their channel attention capabilities. The number of channel growth is designed to be 1 to 128 to prevent overfitting caused by complex models. The activation function can be selected as ReLu, sigmoid, tanh, etc., and the ratio r is set to 16.

[0065] In the above step D, the ratio parameter setting, activation function selection, etc. can be adjusted as needed. A specific model structure example is Figure 2 The specific number of convolutional layers is set to 25 this time, and the number of layers can be more or less. The convolutional block is implemented using 1*1, 3*3, 1*1 structures this time, and can also use common residual block structures such as 3*3, 3*3, etc.

[0066] The specific parameter values ​​of the neural network model of this specific implementation scheme are shown in Table 1.

[0067] Table 1

[0068]

[0069] E. Fully connected layer (producing output):

[0070] The output of the feature extraction dimension of the previous layer is flattened and sent to the fully connected layer. The output dimension is 2, corresponding to systolic pressure and diastolic pressure respectively. The activation function uses sigmoid to normalize the output and make it easier to calculate the error with the label.

[0071] Step 3: Accuracy analysis of model training results

[0072] According to the requirements of the Association for the Advancement of Medical Instrumentation and the British Hypertension Society, mean error, mean absolute error, root mean square error and Pearson correlation coefficient were used to evaluate the training performance of the model.

[0073] The following is a formula for quantitatively evaluating the predicted level of blood pressure:

[0074]

[0075]

[0076]

[0077]

[0078] Where N is equal to 2; i takes different values, P i They represent the systolic and diastolic blood pressure values ​​output by the deep neural network model, respectively. i denote the label value of systolic pressure and the label value of diastolic pressure respectively; ME denotes mean error; MAE denotes mean absolute error; RMSE denotes root mean square error; R denotes Pearson correlation coefficient; conv(P i ,Y i ) represents covariance; Indicates P i The standard deviation of Represents Y i The standard deviation of .

[0079] The error performance of the model in a data set containing 1200 subjects is shown in Table 2. The correlation between the predicted values ​​of systolic and diastolic blood pressure of the neural network model and the true values ​​is shown in Figure 3 , the box plots of the model systolic and diastolic pressure predictions and true values ​​are shown in Figure 4 .

[0080] Table 2

[0081]

[0082] It should be understood that those skilled in the art can make improvements and changes based on the above description, and all these improvements and changes should fall within the scope of protection of the appended claims of the present invention.

Claims

1. A cuffless blood pressure prediction system based on a deep neural network model, characterized in that: The system is used to perform the following steps: receiving a data set including a pulse wave signal and an electrocardiogram signal; Preprocessing the received pulse wave signal and ECG signal to obtain a usable signal; the preprocessing includes signal filtering, fixed-length cutting, peak cutting / trough cutting, resampling and obtaining label values; Inputting the available signal into a deep neural network model to obtain predicted systolic and diastolic blood pressure values; the deep neural network model comprises a convolutional adaptation layer, a residual network with a compression excitation module, and a fully connected layer connected in sequence; the input of the convolutional adaptation layer is the available signal; The output of the fully connected layer is the systolic blood pressure value and the diastolic blood pressure value; The preprocessing specifically includes: Step B1, signal filtering; Step B2, cutting the filtered signal into fixed lengths to obtain fixed length cutting signals; Step B3, performing peak cutting and resampling on the fixed-length cutting signal; Step B4, obtaining the peak cutting signal range after processing in step B3, and calculating the corresponding blood pressure value as a model label; In step B1, the signal filtering method includes a combination of low-pass filtering and high-pass filtering, or a band-pass filtering method; and also includes a wavelet transform or a smoothing filtering method; In step B2, a suitable fixed-length window size is selected to cut the waveform signal; In step B3, the fixed-length cut signal obtained in step B2 is subjected to peak cutting once for each fixed number of peaks; each time two groups of signal segments with a fixed number of peaks are obtained; the two groups of signal segments with a fixed number of peaks are the pulse wave signal and the electrocardiogram signal respectively; In step B4, the peak range calibrated in step B3 is obtained. Since the peak values ​​of the pulse wave and the electrocardiogram signal do not necessarily coincide, the range is set as the union of the peak ranges in step B3. In this union, the values ​​at the peak and trough of the dynamic blood pressure value, namely the systolic pressure and diastolic pressure, are taken as the label values ​​of the model.

2. The cuffless blood pressure prediction system based on the deep neural network model according to claim 1 is characterized in that: The data set includes data sets of several different subjects, and the length of the data set of each subject can meet the segmentation requirements.

3. The cuffless blood pressure prediction system based on a deep neural network model according to claim 1, characterized in that: In step B2, a window with non-overlapping fixed points is used.

4. The cuffless blood pressure prediction system based on a deep neural network model according to claim 1, characterized in that: In step B4, the maximum range of the pulse wave signal and the electrocardiogram signal within one cycle is used to extract the boundary of the dynamic blood pressure; let x1…xn represent the sampling point range of the pulse wave, and z1…zn represent the sampling point range of the electrocardiogram signal, then the extraction boundary of the dynamic blood pressure is [min(x1, z1), max(xn, zn)], and the diastolic pressure and systolic pressure label values ​​correspond to the trough and peak within the dynamic blood pressure range, respectively.

5. The cuffless blood pressure prediction system based on a deep neural network model according to claim 1, characterized in that: The convolution adaptation layer includes a convolution layer, a batch normalization layer and a maximum pooling layer connected in sequence; the convolution kernel size of the convolution adaptation layer needs to be adjusted according to the length of the feature points of the pulse wave signal and the electrocardiogram signal; the specific size is determined by the resampling point, that is, the receptive field includes the feature points of the pulse wave signal and the electrocardiogram signal.

6. The cuffless blood pressure prediction system based on a deep neural network model according to claim 1, characterized in that: The convolution blocks of the residual network with compressed excitation module are implemented with 1*1, 3*3 and 1*1 structures. These modules all have compressed excitation module structures to enhance their attention capabilities on the channel; the number of channel growth is designed to be 1 to 128 to prevent overfitting caused by complex models; the activation functions of the residual network with compressed excitation module include ReLu, sigmoid or tanh; the ratio r can be set.

7. The cuffless blood pressure prediction system based on a deep neural network model according to claim 1, characterized in that: A fully connected layer is used to integrate the feature data output by the residual network with a compressed excitation module. The output dimension of the fully connected layer is 2, corresponding to systolic and diastolic pressures respectively. The activation function of the fully connected layer uses sigmoid to normalize the output to calculate the error with the label value.

8. The cuffless blood pressure prediction system based on a deep neural network model according to claim 1, characterized in that: Also includes: The training performance of the model is evaluated using mean error, mean absolute error, root mean square error, and Pearson correlation coefficient, as follows: Where N is equal to 2; i takes different values, P i Respectively represent the systolic and diastolic blood pressure values ​​output by the deep neural network model, Y i denote the label value of systolic pressure and the label value of diastolic pressure respectively; ME denotes mean error; MAE denotes mean absolute error; RMSE denotes root mean square error; R denotes Pearson correlation coefficient; conv(P i ,Y i ) represents the covariance; σ Pi Indicates P i The standard deviation of Yi Represents Y i The standard deviation of .

Images