WO2021184801A1

WO2021184801A1 - Synchronization signal-based blood pressure prediction method and apparatus

Info

Publication number: WO2021184801A1
Application number: PCT/CN2020/129642
Authority: WO
Inventors: 张碧莹; 曹君
Original assignee: 乐普（北京）医疗器械股份有限公司
Priority date: 2020-03-17
Filing date: 2020-11-18
Publication date: 2021-09-23
Also published as: CN111358452A; CN111358452B

Abstract

A synchronization signal-based blood pressure prediction method and apparatus. The method comprises: obtaining PPG signal data and ECG signal data synchronized with the PPG signal data; respectively performing signal sampling processing on the PPG signal data and the ECG signal data according to a data sampling frequency threshold to generate PPG and ECG one-dimensional data sequences; performing blood pressure CNN input data fusion processing on the PPG and ECG one-dimensional data sequences to generate an input data four-dimensional tensor; according to a convolutional layer number threshold, performing multilayer convolution pooling calculation by using a blood pressure CNN model to generate a characteristic data four-dimensional tensor; constructing a blood pressure ANN input data two-dimensional matrix; performing characteristic data regression calculation by using the blood pressure ANN model to generate a blood pressure regression data two-dimensional matrix; obtaining a prediction mode identifier; when the prediction mode identifier is mean prediction, generating a mean blood pressure prediction data pair; and when the prediction mode identifier is dynamic prediction, generating a dynamic blood pressure prediction one-dimensional data sequence.

Description

Method and device for blood pressure prediction based on synchronization signal

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office on March 17, 2020, with the application number 202010189179.5, and the title of the invention "A method and device for blood pressure prediction based on synchronization signals".

Technical field

The present invention relates to the technical field of electrophysiological signal processing, in particular to a method and device for blood pressure prediction based on a synchronization signal.

Background technique

The heart is the center of human blood circulation. The heart produces blood pressure through regular pulsation, and then supplies blood to the whole body to complete the body's metabolism. Blood pressure is one of the very important physiological signals of the human body. Human blood pressure contains two important values: systolic blood pressure and diastolic blood pressure. Medically, these two quantities are used to judge whether human blood pressure is normal or not. Long-term continuous observation of these two parameters of blood pressure can help people have a clearer understanding of their own heart health. However, most of the current traditional blood pressure measurement methods use external force upward pressure detection methods such as pressure gauges, which are not only cumbersome to operate, but also easily cause discomfort to the subject, so they cannot be used multiple times to achieve the purpose of continuous monitoring. .

Summary of the invention

The purpose of the present invention is to provide a method and device for blood pressure prediction based on synchronization signals in view of the shortcomings of the prior art, by using non-invasive portable signal acquisition equipment to perform non-invasive electrocardiogram (ECG) signals and light volume on the tester Synchronous acquisition of Photoplethysmography (PPG) signals, and data fusion processing of the acquired ECG signals and PPG signals, and then through the blood pressure convolutional neural network (Convolutional Neural Network, CNN) model and blood pressure artificial neural network ( The blood pressure prediction model composed of Artificial Neural Network (ANN) model performs feature calculation and blood pressure data regression calculation on the fusion data to calculate the blood pressure data of the tester, and finally outputs the specific mean blood pressure prediction data pair (systolic blood pressure prediction) through the selection of the prediction mode identifier Data, diastolic blood pressure prediction data) is still a one-dimensional data sequence of ambulatory blood pressure prediction; through the embodiment of the present invention, the cumbersomeness and discomfort of conventional blood pressure collection methods are avoided, and an automatic and intelligent data analysis method is produced, thereby enabling the application It is convenient to carry out multiple continuous monitoring of the measured object.

In order to achieve the foregoing objective, the first aspect of the embodiments of the present invention provides a method for blood pressure prediction based on a synchronization signal, and the method includes:

Acquire the PPG signal data of the photoplethysmography method and the ECG signal data synchronized with it; and according to the data sampling frequency threshold, perform signal sampling processing on the PPG signal data and the ECG signal data to generate a PPG one-dimensional data sequence And generate ECG one-dimensional data sequence;

According to the input data length threshold N of the blood pressure convolutional neural network CNN model, perform blood pressure CNN input data fusion processing on the PPG one-dimensional data sequence and the ECG one-dimensional data sequence to generate a four-dimensional tensor of input data;

According to the threshold of the number of convolutional layers, using the blood pressure CNN model to perform multi-layer convolution pooling calculation on the four-dimensional tensor of the input data to generate a four-dimensional tensor of feature data;

Perform a two-dimensional matrix construction operation of input data of a blood pressure artificial neural network ANN model according to the four-dimensional tensor of the feature data to generate a two-dimensional matrix of input data; and use the blood pressure ANN model to perform feature data regression calculation on the two-dimensional matrix of input data Generate a two-dimensional matrix of blood pressure regression data;

Obtaining a prediction mode identifier; the prediction mode identifier includes two identifiers of mean prediction and dynamic prediction;

When the prediction mode identifier is the average prediction, the average blood pressure calculation operation is performed on the two-dimensional matrix of blood pressure regression data to generate an average blood pressure prediction data pair; the average blood pressure prediction data pair includes diastolic blood pressure prediction data and systolic blood pressure prediction data. Pressure forecast data;

When the prediction mode identifier is the dynamic prediction, perform an ambulatory blood pressure data extraction operation on the blood pressure regression data two-dimensional matrix to generate an ambulatory blood pressure prediction one-dimensional data sequence.

Preferably,

The PPG signal data is generated by using a non-invasive PPG signal acquisition device to perform a preset light source signal acquisition operation on the tester's local skin surface within a signal acquisition time threshold;

The ECG signal data is synchronized with the collection of the PPG signal data, and is generated by using a non-invasive ECG signal collection device to perform an ECG physiological signal collection operation on the tester within the signal collection time threshold;

The PPG one-dimensional data sequence is specifically a PPG one-dimensional data sequence [A]; the PPG one-dimensional data sequence [A] includes the A pieces of PPG data; and the A is the data sampling frequency threshold multiplied by the The product of the signal acquisition time threshold;

The ECG one-dimensional data sequence is specifically an ECG one-dimensional data sequence [A]; the ECG one-dimensional data sequence [A] includes the A pieces of ECG data.

Preferably, according to the input data length threshold N of the blood pressure convolutional neural network CNN model, perform blood pressure CNN input data fusion processing on the PPG one-dimensional data sequence and the ECG one-dimensional data sequence to generate a four-dimensional tensor of input data , Specifically including:

Calculate the total number of fragments M according to the input data length threshold N and the A; when the A is divisible by the input data length threshold N, set the total number of fragments M as the A divided by the input data length The quotient of the threshold N; when the A is not divisible by the input data length threshold N, the total number of fragments M is set as the result of the rounding calculation of the quotient of the A divided by the input data length threshold N;

According to the total number of fragments M and the input data length threshold N, the PPG one-dimensional data sequence [A] is subjected to sequential data fragment division processing to generate a PPG fragment data two-dimensional matrix [M, N], and the ECG one Dimensional data sequence [A] performs sequential data segmentation processing to generate a two-dimensional matrix of ECG segment data [M,N];

Perform input data fusion processing on the two-dimensional matrix of PPG fragment data [M, N] and the two-dimensional matrix of ECG fragment data [M, N] to generate a two-dimensional matrix of fused fragment data [M, 2*N];

Perform a four-dimensional tensor conversion process of blood pressure CNN input data on the two-dimensional matrix [M,2*N] of the fusion segment data to generate the four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ ] of the input data; The B ₁ is the fourth dimension parameter of the input data four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ ], and the B ₁ is the total number of fragments M; the H ₁ is the The input data four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ ] is the third-dimensional parameter, and the _{value of H 1} is 2; the W ₁ is the input data four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ ] second dimension parameter, and the W ₁ is the input data length threshold N; the C ₁ is the input data four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ ] is the first dimension parameter, and the _{value of C 1} is 1.

Further, according to the total number of fragments M and the input data length threshold N, the PPG one-dimensional data sequence [A] is divided into sequential data fragments to generate a PPG fragment data two-dimensional matrix [M, N], Perform sequential data segment division processing on the ECG one-dimensional data sequence [A] to generate a two-dimensional ECG segment data matrix [M, N], which specifically includes:

Generate a temporary data length L according to the product of the total number of fragments M and the input data length threshold N;

From the PPG one-dimensional data sequence [A], taking the first data as the starting data extraction position and the temporary data length L as the data extraction length, extract a piece of continuous data to generate the PPG one-dimensional temporary data sequence [L ]; and according to the input data length threshold N, the PPG one-dimensional temporary data sequence [L] is subjected to continuous data segment division processing; the PPG one-dimensional temporary data sequence [L] includes the total number of segments M PPG one One-dimensional segment data sequence [N]; the PPG one-dimensional segment data sequence [N] includes the input data length threshold N pieces of the PPG data;

From the ECG one-dimensional data sequence [A], taking the first data as the starting data extraction position and the temporary data length L as the data extraction length, extract a piece of continuous data to generate an ECG one-dimensional temporary data sequence [L ]; and according to the input data length threshold N, the ECG one-dimensional temporary data sequence [L] is subjected to continuous data segment division processing; the ECG one-dimensional temporary data sequence [L] includes the total number of segments M ECG one One-dimensional fragment data sequence [N]; the ECG one-dimensional fragment data sequence [N] includes the input data length threshold value N of the ECG data;

Construct the PPG fragment data two-dimensional matrix [M, N], and initialize all matrix elements of the PPG fragment data two-dimensional matrix [M, N] to be empty; from the PPG one-dimensional temporary data sequence [L] , Sequentially extracting the PPG one-dimensional segment data sequence [N] included in the PPG one-dimensional temporary data sequence [L], and assigning values to the matrix elements of the PPG segment data two-dimensional matrix [M, N];

Construct the ECG fragment data two-dimensional matrix [M, N], and initialize all matrix elements of the ECG fragment data two-dimensional matrix [M, N] to be empty; from the ECG one-dimensional temporary data sequence [L] , Sequentially extracting the ECG one-dimensional segment data sequence [N] to perform value assignment processing on the matrix elements of the ECG segment data two-dimensional matrix [M, N].

Further, the input data fusion processing is performed on the two-dimensional matrix of PPG segment data [M, N] and the two-dimensional matrix of ECG segment data [M, N] to generate a two-dimensional matrix of fused segment data [M, 2*N], including:

Step 51: Obtain a preset fusion sorting flag; construct the two-dimensional matrix of fusion fragment data [M, 2*N], and initialize the two-dimensional matrix of fusion fragment data [M, 2*N] to be empty; construct the first A sequence [2*N]; the fusion sequence identifier is one of PPG+ECG sequence and ECG+PPG sequence; the first sequence [2*N] includes 2*N sequence data;

Step 52: Initialize the value of the first index to 1, and initialize the value of the first total to the total number of fragments M;

Step 53: Taking the product of the difference of the first index minus 1 and the input data length threshold N plus 1 as the starting data extraction position, and the input data length threshold N as the extraction data length. Extract a piece of continuous data from the two-dimensional matrix of PPG fragment data [M, N] to generate a second sequence [N], and extract a piece of continuous data from the two-dimensional matrix of ECG fragment data [M, N] to generate a third sequence [ N];

Step 54: Use the second sequence [N] and the third sequence [N] to perform value assignment processing on the first sequence [2*N] according to the fusion order identifier; when the fusion order identifier is When the PPG+ECG is sorted, use the second sequence [N] to perform the assignment processing on the first N sequence data included in the first sequence [2*N], and use the third sequence [N ] Perform value assignment processing on the last N sequence data included in the first sequence [2*N]; when the fusion sequence identifier is the ECG+PPG sequence, use the third sequence [N] pair The first N pieces of the sequence data included in the first sequence [2*N] are assigned values, and the second sequence [N] is used to perform the assignment processing on the last N pieces of data included in the first sequence [2*N]. Assign value to the sequence data;

Step 55: Use the first sequence [2*N] to perform a sequence data addition operation on the two-dimensional matrix [M, 2*N] of fusion fragment data;

Step 56: Add 1 to the first index;

Step 57: Determine whether the first index is greater than the first total, if the first index is greater than the first total, go to step 58, if the first index is less than or equal to the first total, go to Step 53;

Step 58: Return the two-dimensional matrix [M, 2*N] of the fused segment data to the upper application.

Further, the four-dimensional tensor conversion process of blood pressure CNN input data is performed on the two-dimensional matrix [M, 2*N] of the fused segment data to generate the four-dimensional tensor of the input data [B ₁ , H ₁ , W ₁ ,C ₁ ], specifically including:

Construct the input data four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ ], and initialize the input data four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ ] to be empty; then extract sequentially The matrix elements included in the two-dimensional matrix [M, 2*N] of the fusion fragment data _{perform a data item addition operation on the four-dimensional tensor [B 1} , H ₁ , W ₁ , C ₁ ] of the input data; the input data The four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ ] is specifically the input data four-dimensional tensor [M,2,N,1].

Preferably, according to the threshold of the number of convolutional layers, the blood pressure CNN model is used to perform multi-layer convolution pooling calculation on the four-dimensional tensor of the input data to generate a four-dimensional tensor of characteristic data, which specifically includes:

Step 71: Initialize the value of the second index to 1; initialize the second total to the threshold of the number of convolutional layers; initialize the temporary four-dimensional tensor of the second index to the four-dimensional tensor of the input data [B ₁ , H ₁ , W ₁ ,C ₁ ];

Step 72: Use the second index layer convolution layer of the blood pressure CNN model to perform convolution calculation processing on the second index temporary four-dimensional tensor to generate a second index convolution output data four-dimensional tensor; use the blood pressure The second index pooling layer of the CNN model performs pooling calculation processing on the four-dimensional tensor of the second index convolution output data to generate the four-dimensional tensor of the second index pooling output data; the blood pressure CNN model includes multiple layers. Said convolutional layer and multiple layers of said pooling layer;

Step 73: Set the temporary four-dimensional tensor of the second index as the four-dimensional tensor of the second index pooling output data;

Step 74: Add 1 to the second index;

Step 75: Determine whether the second index is greater than the second total, if the second index is greater than the second total, go to step 76, if the second index is less than or equal to the second total, go to Step 72;

Step 76: Set the feature data four-dimensional tensor as the second index temporary four-dimensional tensor; the feature data four-dimensional tensor is specifically the feature data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C ₂ ]; The B2 is the fourth dimension parameter of the characteristic data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C ₂ ] and the B ₂ is the B ₁ ; the H ₂ is the characteristic data four-dimensional The third dimension parameter of the tensor [B ₂ , H ₂ , W ₂ , C ₂ _{]; the W 2} is the second dimension of the feature data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C ₂ ] Parameters; The C ₂ is the first dimension parameter of the characteristic data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C _{2 ].}

Preferably, the two-dimensional matrix construction operation of the input data of the blood pressure artificial neural network ANN model is performed according to the four-dimensional tensor of the characteristic data to generate a two-dimensional matrix of input data; and the two-dimensional matrix of input data is calculated by the blood pressure ANN model Perform feature data regression calculation to generate a two-dimensional matrix of blood pressure regression data, which specifically includes:

According to the feature data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C ₂ ], perform tensor data dimensionality reduction processing on the feature data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C _2] Construct the input data two-dimensional matrix; the input data two-dimensional matrix is specifically the input data two-dimensional matrix [W ₃ , C ₃ ]; the W ₃ is the input data two-dimensional matrix [W ₃ , C ₃ ] The second dimension parameter of the W ₃ is the B ₂ ; the C ₃ is the first dimension parameter of the input data two-dimensional matrix [W ₃ , C ₃ ] and the C ₃ is the H ₂ Multiply by the product of W ₂ and then by the product of _{C 2;}

Using the blood pressure ANN model, perform characteristic data regression calculation on the input data two-dimensional matrix [W ₃ , C ₃ ] to generate a blood pressure regression data two-dimensional matrix; the blood pressure regression data two-dimensional matrix is specifically blood pressure regression data two-dimensional Matrix [X, 2]; the X is the second dimension parameter of the blood pressure regression data two-dimensional matrix [X, 2] and the X is the W ₃ ; the blood pressure regression data two-dimensional matrix [X, 2] includes the X regression data one-dimensional data sequence [2]; the regression data one-dimensional data sequence [2] includes the fragment systolic blood pressure data and the fragment diastolic blood pressure data.

Preferably, when the prediction mode identifier is the average prediction, performing an average blood pressure calculation operation on the two-dimensional matrix of blood pressure regression data to generate an average blood pressure prediction data pair specifically includes:

When the prediction mode identifier is the average prediction, the average blood pressure prediction data pair is set; and the diastolic blood pressure prediction data of the average blood pressure prediction data pair is initialized to be empty, and the average blood pressure prediction data pair is initialized The systolic blood pressure data of is empty;

The sum of the diastolic blood pressure data of all the pieces of the regression data one-dimensional data sequence [2] included in the blood pressure regression data two-dimensional matrix [X, 2] is calculated to generate the sum of diastolic blood pressure, which is divided by the sum of diastolic blood pressure Use the quotient of X to generate the first mean value; generate the sum of the systolic blood pressure data of all the regression data one-dimensional data sequence [2] included in the blood pressure regression data two-dimensional matrix [X, 2] The sum of systolic blood pressures, the second mean value is generated according to the quotient of the sum of systolic blood pressures divided by the X;

Set the diastolic blood pressure prediction data of the average blood pressure prediction data pair as the first average value; set the systolic blood pressure prediction data of the average blood pressure prediction data pair as the second average value.

Preferably, when the prediction mode identifier is the dynamic prediction, performing an ambulatory blood pressure data extraction operation on the two-dimensional matrix of blood pressure regression data to generate a one-dimensional ambulatory blood pressure prediction data sequence specifically includes:

When the prediction mode identifier is the dynamic prediction, initialize the ambulatory blood pressure prediction one-dimensional data sequence to be empty; set the blood pressure data group; initialize the diastolic blood pressure data of the blood pressure data group to be empty; initialize the blood pressure data The systolic blood pressure data of the group is empty;

The one-dimensional data sequence [2] of the regression data included in the two-dimensional matrix [X, 2] of the blood pressure regression data is sequentially extracted to generate the current data sequence [2]; the systolic blood pressure data of the blood pressure data group is set to all The segment systolic blood pressure data of the current data sequence [2], the diastolic blood pressure data of the blood pressure data group is set to the segment diastolic blood pressure data of the current data sequence [2]; and the blood pressure data The group performs a data group addition operation to the ambulatory blood pressure prediction one-dimensional data sequence.

The first aspect of the embodiments of the present invention provides a method for blood pressure prediction based on synchronization signals, which performs feature data fusion on the synchronized ECG signal and PPG signal, and then uses a blood pressure prediction model composed of a blood pressure CNN model and a blood pressure ANN model to fuse the data Perform feature calculation and blood pressure data regression calculation to calculate the tester's blood pressure data, and finally output the mean blood pressure prediction data pair (systolic blood pressure prediction data, diastolic blood pressure prediction data) or ambulatory blood pressure prediction one-dimensional data through the selection of the prediction mode identifier sequence.

A second aspect of the embodiments of the present invention provides a device, the device including a memory and a processor, the memory is used to store a program, and the processor is used to execute the first aspect and the methods in each implementation manner of the first aspect.

A third aspect of the embodiments of the present invention provides a computer program product containing instructions, which when the computer program product runs on a computer, causes the computer to execute the first aspect and the methods in the implementation manners of the first aspect.

A fourth aspect of the embodiments of the present invention provides a computer-readable storage medium on which a computer program is stored. The computer program implements the first aspect and the methods in the first aspect when executed by a processor. .

Description of the drawings

FIG. 1 is a schematic diagram of a method for blood pressure prediction based on a synchronization signal according to Embodiment 1 of the present invention;

2 is a schematic diagram of a method for blood pressure prediction based on a synchronization signal according to Embodiment 2 of the present invention;

FIG. 3 is a schematic diagram of the device structure of an apparatus for blood pressure prediction based on a synchronization signal according to Embodiment 3 of the present invention.

Detailed ways

In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. . Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of the present invention.

Before further elaborating the present invention through the embodiments, some technical means mentioned in the text will be briefly described.

The PPG signal is a set of signals that uses the light sensor to identify and record the change in light intensity of a specific light source. When the heart beats, the blood flow per unit area in the blood vessel changes periodically, and the corresponding blood volume also changes accordingly, resulting in a periodic change trend in the PPG signal reflecting the amount of light absorbed by the blood. A cardiac cycle consists of two time periods: systolic and diastolic; during systole, the heart does work on the blood throughout the body, causing continuous and periodic changes in intravascular pressure and blood flow volume. The absorption of light is the most; when the heart is in diastole, the pressure on the blood vessels is relatively small. At this time, the blood pushed out to the whole body from the last systole hits the heart valve through the circulation, which produces a certain reflection and refraction effect on the light, resulting in the diastolic cycle At this time, the absorption of light energy by blood in the blood vessel decreases. Therefore, the PPG signal waveform reflecting the light energy absorbed by the blood in the blood vessel is composed of two signals: the systolic period signal and the diastolic period signal; the common PPG signal has two peaks, the first one belongs to the systolic period, and the diastolic period signal. The latter belongs to the diastolic period. The blood pressure can be predicted directly by using the PPG signal.

The ECG signal is a group of electrophysiological signals of the heart's cardiac cycle collected from body surface records using an electrocardiographic signal acquisition device. The ECG signal used in the embodiment of the present invention is a group of single-lead ECG signals. Because the ECG signal reflects the activity of the heart, and the PPG reflects the blood turbulence, this blood turbulence is closely related to each contraction of the heart: there is a pulse propagation time between the PPG signal cycle and the ECG signal cycle (Pulse Translation Time, PTT) time difference. After years of research, it has been found that under certain conditions, there is a linear relationship between PTT and arterial blood pressure. In addition, about one-third of hypertensive patients have different degrees of high heart disease, and some heart diseases can also cause blood pressure changes when they occur. Therefore, we believe that the introduction of an ECG signal synchronized with the PPG signal will provide additional information for the blood pressure prediction model, which will help improve the accuracy of blood pressure prediction.

By performing further feature recognition and regression classification processing on the synchronized ECG signal and PPG signal using the blood pressure prediction model, we can obtain the predicted value of the systolic and diastolic blood pressure. A brief summary of the operation steps is: First, by using the feature calculation network, perform feature calculation on the PPG signal according to the relationship between the PPG signal and the blood pressure, output a set of blood pressure feature values calculated by the PPG, and feature the ECG signal according to the relationship between the PTT and the blood pressure Calculate and output a set of blood pressure characteristic values calculated by PTT; then the obtained two sets of blood pressure characteristic data can be predicted by using blood pressure regression calculation: diastolic blood pressure data and systolic blood pressure data, where the systolic blood pressure data is greater than the diastolic blood pressure data. The blood pressure prediction models mentioned in the embodiment of the present invention include a blood pressure CNN model and a blood pressure ANN model. The former is responsible for feature calculation and the latter is responsible for regression calculation. Among them, in the input fragment data of the blood pressure CNN model, there are two sorting methods for the sorting of the PPG signal data and the ECG signal data (PPG+ECG or ECG+PPG), and the ratio between the two is 1: 1 (Complete the fusion processing of input data through blood pressure CNN).

Regarding feature calculation, we know that CNN has long been one of the core algorithms in the field of feature recognition. Used in image recognition, it can be used in fine classification and recognition to extract the discriminative features of the image for learning by other classifiers. In the field of application of blood pressure feature recognition, the blood pressure feature extraction calculation is performed on the input data: after the input data is convolved and pooled, the feature information related to blood pressure is retained for other networks to learn. The blood pressure CNN model mentioned in the article is a CNN model that has been trained through blood pressure feature extraction. It is specifically composed of a convolutional layer and a pooling layer. The convolutional layer is responsible for performing blood pressure feature extraction calculations on the input data of the CNN model. , The pooling layer is to down-sample the extraction results of the convolutional layer; the blood pressure CNN model in this article is divided into multiple CNN network layers, and each CNN network layer includes a convolutional layer and a pooling layer. The input data and output data format of the blood pressure CNN model are both in the form of a 4-dimensional tensor: [B, H, W, C]. After each layer of convolutional layer or pooling layer, the value of some dimensional parameters of the output data will change, that is, the total data length of the tensor will be shortened. The characteristic of the change is: B as the fourth dimension parameter (PPG The total number of fragments of one-dimensional data sequence or ECG one-dimensional data sequence) will not change; H and W are the third and two-dimensional parameters in four dimensions. The setting of the step size is also related to the pooling window size and sliding step size of the pooling layer; C is the first dimension parameter in the four-dimensional, and its change is related to the selected output space dimension in the convolutional layer (volume The number of product cores) is related.

ANN refers to a complex network structure formed by interconnecting a large number of processing units (neurons). It is a certain abstraction, simplification and simulation of the human brain tissue structure and operating mechanism. ANN simulates neuron activity with mathematical models. It is an information processing system based on imitating the structure and function of the brain's neural network. A common application of ANN is to perform classification and regression calculations on data. The blood pressure ANN model mentioned in the article is an ANN model that has been trained through blood pressure classification regression; specifically, the blood pressure ANN model consists of a fully connected layer, where each node of the fully connected layer is connected to the previous layer. All nodes are connected and used to integrate the features extracted in the front to perform a regression calculation and use the calculation result as the input of the regression calculation of the next layer until the stop condition is met and the final calculation result is output to the outside of the network. Here, the input of the blood pressure ANN model is a two-dimensional matrix, so the output of the CNN needs to be converted from a four-dimensional tensor [B, H, W, C] to a two-dimensional matrix; the output of the blood pressure ANN model is also a two-dimensional Matrix [X, 2], in which the second dimension parameter X and B are equal to indicate the total number of segments, and the first dimension parameter of 2 means that the length of the X one-dimensional data sequences included in the matrix are all 2. Each one-dimensional data sequence [2] includes two values, the higher value is the systolic blood pressure predicted by the corresponding PPG+ECG (or ECG+PPG) segment, and the lower value is the corresponding PPG+ECG (or ECG) +PPG) Diastolic blood pressure predicted by the fragment. Finally, the output of the blood pressure ANN model is a pair of blood pressure prediction values (systolic blood pressure and diastolic blood pressure) corresponding to each PPG+ECG (or ECG+PPG) segment. For these blood pressure prediction values, different processing methods can be adopted, such as averaging Value, so as to obtain the average blood pressure data within the signal acquisition time threshold; or directly output the blood pressure value sequence to obtain a segment of ambulatory blood pressure signal.

FIG. 1 is a schematic diagram of a method for blood pressure prediction based on a synchronization signal according to Embodiment 1 of the present invention. The method mainly includes the following steps:

Step 1. Obtain the PPG signal data of the photoplethysmography method and the ECG signal data synchronized with it; and according to the data sampling frequency threshold, perform signal sampling processing on the PPG signal data and the ECG signal data to generate PPG one-dimensional data Sequence and ECG one-dimensional data sequence;

Among them, the PPG signal data is generated by using a non-invasive PPG signal acquisition device to perform a preset light source signal acquisition operation on the tester’s local skin surface within the signal acquisition time threshold; here, when the PPG signal is acquired, the preset The set light source signal includes at least one of a red light source signal, an infrared light source signal, and a green light source signal;

The ECG signal data is synchronized with the acquisition of the PPG signal data, and is generated by using a non-invasive ECG signal acquisition device to perform the ECG physiological signal acquisition operation on the tester within the signal acquisition time threshold; here, when the ECG signal is acquired, It is a single-lead ECG signal; and the ECG signal must be collected synchronously with the PPG signal, so that the PTT time difference between the two is true and reliable;

The PPG one-dimensional data sequence is specifically the PPG one-dimensional data sequence [A]; the PPG one-dimensional data sequence [A] includes A pieces of PPG data; where A is the product of the data sampling frequency threshold multiplied by the signal acquisition time threshold; here, Suppose the signal acquisition time threshold is 10 seconds, and the data sampling frequency threshold is 125Hz, then A=125*10=1250, which means that there are 1250 collected data, and the PPG one-dimensional data sequence [A] is one that includes 1250 PPG acquisitions. One-dimensional data sequence of data;

The one-dimensional ECG data sequence is specifically the one-dimensional ECG data sequence [A]; the one-dimensional ECG data sequence [A] includes A pieces of ECG data; here, in the same way, the signal acquisition time threshold is set to 10 seconds, and the data sampling frequency threshold is 125Hz, then A=125*10=1250, which means that there are 1250 collected data. The ECG one-dimensional data sequence [A] is a one-dimensional data sequence including 1250 ECG collected data.

Step 2: According to the input data length threshold N of the blood pressure convolutional neural network CNN model, perform blood pressure CNN input data fusion processing on the PPG one-dimensional data sequence and the ECG one-dimensional data sequence to generate a four-dimensional tensor of input data;

Specifically: Step 21: Calculate the total number of fragments M based on the input data length threshold N and A; when A is divisible by the input data length threshold N, set the total number of fragments M as the quotient of A divided by the input data length threshold N; when A cannot When divisible by the input data length threshold N, the total number of fragments M is set as the result of rounding the quotient of A divided by the input data length threshold N;

Here, because the blood pressure CNN model will be used to perform feature calculations on the data in the PPG (or ECG) one-dimensional data sequence, in view of the input requirements of the blood pressure CNN (input data length threshold N), according to the input data length of the blood pressure CNN The threshold N divides the PPG (or ECG) one-dimensional data sequence into segments. The method for setting the total number of fragments here: If the total data length of the PPG (or ECG) one-dimensional data sequence can be evenly divided by the input data length threshold N, then the total number of fragments is the quotient of the division of the two; if the PPG (or ECG) is one The total data length of the one-dimensional data sequence cannot be divisible by the input data length threshold N, then the total number of segments is the rounded result of the quotient of the division of the two. The last segment of the PPG (or ECG) one-dimensional data sequence is regarded as the segment with insufficient length. Discard the fragments with incomplete data. For example, assuming that the input data length threshold N is 250, the total number of fragments here is 1250/250=5; assuming the input data length threshold N is 200, the total number of fragments here is int(1250/200)=6, int() Is the rounding function;

Step 22: According to the total number of fragments M and the input data length threshold N, the PPG one-dimensional data sequence [A] is divided into sequential data fragments to generate the PPG fragment data two-dimensional matrix [M, N], and the ECG one-dimensional data sequence [A] ] Perform sequential data segmentation processing to generate a two-dimensional matrix of ECG segment data [M, N];

Specifically, it includes: step 221, generating a temporary data length L according to the product of the total number of fragments M and the input data length threshold N;

Here, from the calculation method of the total number of fragments above, we can obtain the effective data length L=M*N of the two-dimensional matrix [M,N] of the fragment data actually included in the one-dimensional data sequence of PPG and ECG; The length is 1250, and the input data length threshold N is 250, then M=1250/250=5, L=250*5=1250, which means that all data will be included in the fragment data two-dimensional matrix [M, N]; assuming the total length of the data Is 1250, the input data length threshold N is 200, M=int(1250/200)=6, L=200*6=1200, which means that only 1200 data are included in the two-dimensional matrix of fragment data [M, N], and the rest 50 data will be discarded;

Step 222: From the PPG one-dimensional data sequence [A], take the first data as the starting data extraction position and the temporary data length L as the data extraction length, extract a piece of continuous data to generate the PPG one-dimensional temporary data sequence [L] ; And according to the input data length threshold N, the PPG one-dimensional temporary data sequence [L] is divided into continuous data segments;

Among them, the PPG one-dimensional temporary data sequence [L] includes the total number of segments M PPG one-dimensional segment data sequences [N]; the PPG one-dimensional segment data sequence [N] includes the input data length threshold N PPG data;

Here, it is to extract the PPG one-dimensional temporary data sequence [L] (effective data sequence) from the original PPG one-dimensional data sequence [A] according to the temporary data length L (also the effective data length) obtained by the upper calculation; The method is to extract backwards from the earliest data;

In addition, while extracting data, use the extraction unit (input data length threshold N) as the segment length to divide the PPG one-dimensional temporary data sequence [L]; suppose A=1250, and the input data length threshold N is 250, M =5, PPG one-dimensional data sequence [1250]={D ₁ , D ₂ , D ₃ ,...D _i ,...D ₁₂₅₀ } (the value of i is from 1 to 1250), then PPG one-dimensional temporary data sequence [L ] Is the PPG one-dimensional temporary data sequence [1250], including 5 PPG one-dimensional fragment data sequences [250]: the first PPG one-dimensional fragment data sequence [250]={D ₁ ,...D ₂₅₀ }, the second PPG one One-dimensional segment data sequence [250] = {D ₂₅₁ ,...D ₅₀₀ }, the third PPG one-dimensional segment data sequence [250] = {D ₅₀₁ ,...D ₇₅₀ }, the fourth PPG one-dimensional segment data sequence [250] = {D ₇₅₁ ,...D ₁₀₀₀ }, the fifth PPG one-dimensional segment data sequence [250]={D ₁₀₀₁ ,...D ₁₂₅₀ };

Step 223: From the ECG one-dimensional data sequence [A], take the first data as the starting data extraction position and the temporary data length L as the data extraction length, extract a piece of continuous data to generate the ECG one-dimensional temporary data sequence [L] ; And according to the input data length threshold N, the ECG one-dimensional temporary data sequence [L] is divided into continuous data segments;

Among them, the ECG one-dimensional temporary data sequence [L] includes the total number of segments M ECG one-dimensional segment data sequences [N]; the ECG one-dimensional segment data sequence [N] includes the input data length threshold N ECG data;

Here, the one-dimensional ECG temporary data sequence [L] (effective data sequence) is extracted from the original ECG one-dimensional data sequence [A] according to the temporary data length L (also the effective data length) obtained by the upper calculation; The method is to extract backwards from the earliest data;

In addition, while extracting data, take the extraction unit (input data length threshold N) as the segment length to complete the segmentation of the ECG one-dimensional temporary data sequence [L]; suppose A=1250, the input data length threshold N is 250, M =5, ECG one-dimensional data sequence [1250]={E ₁ , E ₂ , E ₃ ,...E _i ,...E ₁₂₅₀ } (the value of i is from 1 to 1250), then the ECG one-dimensional temporary data sequence [L ] Is an ECG one-dimensional temporary data sequence [1250], including 5 ECG one-dimensional fragment data sequences [250]: the first ECG one-dimensional fragment data sequence [250] = {E ₁ ,...E ₂₅₀ }, the second ECG one One-dimensional segment data sequence [250]={E ₂₅₁ ,...E ₅₀₀ }, the third ECG one-dimensional segment data sequence [250]={E ₅₀₁ ,...E ₇₅₀ }, the fourth ECG one-dimensional segment data sequence [250]= {E ₇₅₁ ,...E ₁₀₀₀ }, the fifth ECG one-dimensional fragment data sequence [250]={E ₁₀₀₁ ,...E ₁₂₅₀ };

Step 224: Construct a two-dimensional PPG fragment data matrix [M, N], and initialize all matrix elements of the PPG fragment data two-dimensional matrix [M, N] to be empty; extract sequentially from the PPG one-dimensional temporary data sequence [L] The PPG one-dimensional segment data sequence [N] included in the PPG one-dimensional temporary data sequence [L] assigns values to the matrix elements of the PPG segment data two-dimensional matrix [M, N];

Here, the PPG fragment data two-dimensional matrix [M, N], although it is a two-dimensional matrix, the actual number of matrix elements M*N it contains is equal to the total number of data included in the PPG one-dimensional temporary data sequence [L]. Here is to take the shape of the PPG one-dimensional temporary data sequence [L] from the one-dimensional sequence to a tensor upgrade process; for example, there is a PPG one-dimensional temporary data sequence [4]={0,1,2,3}, change it The two-dimensional matrix [2,2] generated after the two-dimensional tensor is upgraded will not adjust the original data sorting, so the two-dimensional matrix [2,2]={{0,1},{2,3} };

Step 225: Construct a two-dimensional ECG fragment data matrix [M, N], and initialize all matrix elements of the ECG fragment data two-dimensional matrix [M, N] to be empty; extract sequentially from the ECG one-dimensional temporary data sequence [L] The ECG one-dimensional segment data sequence [N] assigns values to the matrix elements of the ECG segment data two-dimensional matrix [M, N];

Similar to step 224, the two-dimensional matrix of ECG fragment data [M,N], although it is a two-dimensional matrix, the actual number of matrix elements M*N contained in it is the total number of data included in the ECG one-dimensional temporary data sequence [L] Equal, here is the tensor upscaling process of the shape of the ECG one-dimensional temporary data sequence [L] from the one-dimensional sequence;

Step 23: Perform input data fusion processing on the two-dimensional matrix of PPG fragment data [M, N] and the two-dimensional matrix of ECG fragment data [M, N] to generate a two-dimensional matrix of fused fragment data [M, 2*N];

It specifically includes: step 231, obtaining a preset fusion sorting flag; constructing a two-dimensional matrix of fusion fragment data [M, 2*N], and initializing the two-dimensional matrix of fusion fragment data [M, 2*N] to be empty; constructing the first Sequence [2*N];

Wherein, the fusion sorting identifier is one of PPG+ECG sorting and ECG+PPG sorting, and the first sequence [2*N] includes 2*N sequence data;

Here, an identifier that specifies the sorting order during data fusion is provided: the fusion sorting identifier; the identifier includes two possible values: PPG+ECG sorting or ECG+PPG sorting; when the fusion sorting identifier is PPG+ECG sorting, fusion In each one-dimensional vector [2*N] (first sequence [2*N]) included in the two-dimensional matrix of fragment data, the first N are PPG data and the last N are ECG data; when the fusion sorting flag is ECG +When PPG sorting, in each one-dimensional vector [2*N] (first sequence [2*N]) included in the two-dimensional matrix of fusion fragment data, the first N are ECG data, and the last N are PPG data;

Step 232: Initialize the value of the first index to 1, and initialize the value of the first total to the total number of fragments M;

Step 233, taking the product of the first index minus 1 and the input data length threshold N plus 1 as the starting data extraction position, and the input data length threshold N as the extraction data length, from the two-dimensional matrix of PPG fragment data [ Extract a piece of continuous data from M, N] to generate a second sequence [N], and extract a piece of continuous data from a two-dimensional matrix of ECG fragment data [M, N] to generate a third sequence [N];

Here, we have obtained the two-dimensional matrix of PPG fragment data [M, N] and the two-dimensional matrix of ECG fragment data [M, N]; before using the blood pressure CNN model, the two sets of data need to be fused and assembled into a complete set The fusion principle of the embodiment of the present invention is to keep all the data without modification while keeping the total number of fragments unchanged; then, we know that the total number of the two sets of data is 2*M*N, and the next step is to generate one The two-dimensional matrix of fusion fragment data [M,2*N] with the shape of M*(2*N) includes all PPG and ECG data; for the two-dimensional matrix of fusion fragment data [M,2*N] we can regard it as M data sequences with a length of 2N, each data sequence includes N PPG data and N ECG data; here, the embodiment of the present invention provides a fusion order identifier to identify the front and back of N PPG data and N ECG data Order: When the fusion sorting flag is PPG+ECG sorting, set the first N as PPG data and the last N as ECG data. When the fusion sorting flag is PPG+ECG, set the first N as ECG data and the last N Is PPG data;

Suppose A=1250, N=250, M=5, PPG one-dimensional data sequence [1250], then PPG segment data two-dimensional matrix [M, N] is PPG segment data two-dimensional matrix [5,250], ECG segment data two-dimensional Matrix [M, N] is the two-dimensional matrix of ECG fragment data [5,250]; the final fused two-dimensional matrix of fused fragment data [M, 2*N] is the two-dimensional matrix of fused fragment data [5,500];

Step 234: Use the second sequence [N] and the third sequence [N] to assign values to the first sequence [2*N] according to the fusion order identifier; when the fusion order identifier is PPG+ECG order, use the second sequence [N] Perform assignment processing on the first N sequence data included in the first sequence [2*N], and use the third sequence [N] to perform assignment processing on the last N sequence data included in the first sequence [2*N] ; When the fusion sorting flag is ECG+PPG sorting, use the third sequence [N] to assign values to the first N sequence data included in the first sequence [2*N], and use the second sequence [N] to The last N sequence data included in sequence [2*N] are assigned values;

Here, it is the operation of assigning a value to each data sequence (first sequence [2*N]) column with a length of 2N in the two-dimensional matrix [M,2*N] of the fusion fragment data: when the fusion sequence flag is PPG+ In ECG sorting, the first N are PPG data (second sequence [N]) and the next N are ECG data (third sequence [N]); when the fusion sorting flag is ECG+PPG sorting, the first N are ECG data (The third sequence [N]) The last N are PPG data (the second sequence [N]);

Step 235, using the first sequence [2*N] to perform a sequence data addition operation on the two-dimensional matrix [M, 2*N] of the fusion fragment data;

Here, the 2N-length data sequence (the first sequence [2*N]) at the end of each assignment is added to the two-dimensional matrix [M, 2*N] of the fusion fragment data, and finally after M times are added Complete the whole construction process of the two-dimensional matrix [M,2*N] of the fusion fragment data;

Suppose A=4, N=2, M=2, PPG one-dimensional data sequence [4]={0,1,2,3}, ECG one-dimensional data sequence [4]={10,11,12,13} , Then the two-dimensional matrix of PPG fragment data [M, N] is the two-dimensional matrix of PPG fragment data [2,2]={{0,1},{2,3}}, the two-dimensional matrix of ECG fragment data [M, N ] Is the two-dimensional matrix of ECG fragment data [2,2]={{10,11},{12,13}}, and the final fused fragment data two-dimensional matrix [M,2*N] is the fused fragment data Two-dimensional matrix [2,4]={{0,1,10,11},{2,3,12,13}};

Step 236: Add 1 to the first index;

Here, the first index is the segment index extracted from the two-dimensional matrix [M, N] of PPG (ECG) segment data;

Step 237: Determine whether the first index is greater than the first total, if the first index is greater than the first total, go to step 24, and if the first index is less than or equal to the first total, go to step 233.

Step 24: Perform a four-dimensional tensor conversion process of blood pressure CNN input data on the two-dimensional matrix [M,2*N] of the fused segment data to generate a four-dimensional tensor of input data [B ₁ , H ₁ , W ₁ , C ₁ ];

It specifically includes: constructing the input data four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ ], and initializing the input data four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ ] to be empty; then extract the fusion in turn The matrix elements included in the two-dimensional matrix of fragment data [M,2*N] add data items to the _{four-dimensional tensor [B 1} , H ₁ , W ₁ , C _{1] of the input data;}

Among them, the input data four-dimensional tensor [B ₁ ,H ₁ ,W ₁ ,C ₁ ] is specifically the input data four-dimensional tensor [M,2,N,1]; B ₁ is the input data four-dimensional tensor [B ₁ ,H ₁ ,W ₁ ,C ₁ ], the fourth dimension parameter, and B ₁ is the total number of fragments M; H ₁ is the third dimension parameter of the input data four-dimensional tensor [B ₁ ,H ₁ ,W ₁ ,C _{1 ], and} The value of H ₁ is 2; W ₁ is the second dimension parameter of the input data four-dimensional tensor [B ₁ , H ₁ , W ₁ , C ₁ _{], and W 1} is the input data length threshold N; C ₁ is the input data four-dimensional The first dimension parameter of the tensor [B ₁ ,H ₁ ,W ₁ ,C ₁ ], and the value of _{C 1 is 1.}

As mentioned in the foregoing technical introduction, the input and output parameters of the blood pressure CNN model used in the embodiment of the present invention are all in the form of a four-dimensional tensor, so here is a four-dimensional tensor increase for the two-dimensional matrix [M,2*N] of the fusion fragment data. Two-dimensional operation, assuming A=1250, N=250, M=5, PPG one-dimensional data sequence [1250] and ECG one-dimensional data sequence [1250], then the two-dimensional matrix of fused fragment data after fusion [M,2* N] is the two-dimensional matrix of fused segment data [5,500]; the two-dimensional matrix of fused segment data [5,500] is upscaled into a four-dimensional tensor of input data [B ₁ ,H ₁ ,W ₁ ,C ₁ ], which is Is the input data four-dimensional tensor [M,2,N,1] (input data four-dimensional tensor [5,2 ₁ ,250,1]);

Step 3. According to the threshold of the number of convolutional layers, use the blood pressure CNN model to perform multi-layer convolution pooling calculation on the four-dimensional tensor of the input data to generate the four-dimensional tensor of feature data;

Specifically: Step 31: Initialize the value of the second index to 1; initialize the second total to the threshold of the number of convolutional layers; initialize the temporary four-dimensional tensor of the second index to the four-dimensional tensor of the input data [B ₁ ,H ₁ ,W ₁ , C ₁ ];

Step 32: Use the second index layer convolution layer of the blood pressure CNN model to perform convolution calculation processing on the second index temporary four-dimensional tensor to generate the second index convolution output data four-dimensional tensor; use the second index of the blood pressure CNN model Pooling layer, which performs pooling calculation processing on the four-dimensional tensor of the second index convolution output data to generate a four-dimensional tensor of second index pooling output data; the blood pressure CNN model includes a multi-layer convolutional layer and a multi-layer pooling layer;

Here, the preprocessed data is input into the trained blood pressure CNN model to extract features. The blood pressure CNN model consists of a multi-layer convolutional layer and a pooling layer. The general structure is one layer of convolution and one layer of pooling. Connected to the next convolutional layer, the final number of layers of the blood pressure CNN model is determined by the number of convolutional layer thresholds. For example, a network with 4 convolutional layers and 4 pooling layers is called a 4-layer convolutional network, where convolution The layer performs convolution operations to convert the input into outputs of different dimensions. These outputs can be regarded as another way of expressing the input, and the pooling layer is used to control the number of outputs, simplify the operation and prompt the network to extract more effective information. ；

Step 33: Set the temporary four-dimensional tensor of the second index as the four-dimensional tensor of the second index pooling output data;

Step 34: Add 1 to the second index;

Step 35: Determine whether the second index is greater than the second total, if the second index is greater than the second total, go to step 36, if the second index is less than or equal to the second total, go to step 32;

Step 36: Set the feature data four-dimensional tensor as the second index temporary four-dimensional tensor;

Among them, the four-dimensional tensor of feature data is specifically the four-dimensional tensor of feature data [B ₂ , H ₂ , W ₂ , C ₂ ]; B2 is the fourth tensor of feature data [B ₂ , H ₂ , W ₂ , C ₂ ] Four-dimensional parameters and B ₂ is B ₁ ; H ₂ is the third-dimensional parameter of the feature data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C ₂ _{]; W 2} is the feature data four-dimensional tensor [B ₂ ,H ₂ , W ₂ , C ₂ ] the second dimension parameter; C ₂ is the first dimension parameter of the feature data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C ₂ ].

Here, the convolution principle of each layer of the blood pressure CNN model is the same as the two-dimensional convolution principle. The difference from image convolution is that the height H of the PPG signal and the ECG signal is 1, so the convolution kernel in the convolution layer is the first All dimensions are 1, such as [1x3], [1x5], [1x7] and so on. After a layer of convolutional layer or pooling layer, the shape of the input data will change, but it still maintains the 4-dimensional tensor form. Among them, the fourth dimension parameter (the total number of fragments) will not change, and the third and two-dimensional parameters (H and W) change are related to the size of the convolution kernel of each convolutional layer and the setting of the sliding step length, and it is also related to the pool The pooling window size of the transformation layer is related to the sliding step size. The first dimension parameter (the number of channels) is related to the selected output space dimension (the number of convolution kernels) in the convolution layer, and the number of layers in the network is set , The setting of various parameters of each layer must be determined based on experience and experimental results, not a fixed value, here it is assumed that after several layers of the network, the output of the network becomes 4 of the shape [5,2,20,64] Dimension tensor

Step 4. According to the four-dimensional tensor of the characteristic data, perform the operation of constructing a two-dimensional matrix of blood pressure artificial neural network ANN input data to generate a two-dimensional matrix of input data; and use the blood pressure ANN model to perform characteristic data regression calculation on the two-dimensional matrix of input data to generate blood pressure regression data. Dimensional matrix

Specifically: Step 41, according to the feature data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C ₂ ], perform tensor data reduction on the feature data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C _2] Dimensional processing to construct a two-dimensional matrix of input data;

Among them, the input data two-dimensional matrix is specifically the input data two-dimensional matrix [W ₃ , C ₃ ]; W ₃ is the second dimension parameter of the input data two-dimensional matrix [W ₃ , C ₃ _{] and W 3} is B ₂ ; C ₃ is the first dimension parameter of the input data two-dimensional matrix [W ₃ , C ₃ _{] and C 3} is the product of H ₂ multiplied by W ₂ and then multiplied by C _2;

Here, because the input and output data structure of the blood pressure ANN model is defined as a two-dimensional matrix tensor structure, the characteristic data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C ₂ ] is input into the blood pressure ANN for regression The shape of the tensor needs to be reduced before calculation. For example, the feature data four-dimensional tensor [B ₂ ,H ₂ ,W ₂ ,C ₂ ] is the feature data four-dimensional tensor [5,2,20,64]. After its shape is reduced in dimensionality, it becomes a two-dimensional matrix of input data [5,2*20*64], which is a two-dimensional matrix of input data [5,2560];

Step 42, using the blood pressure ANN model to perform feature data regression calculation on the two-dimensional matrix of input data [W ₃ , C ₃ ] to generate a two-dimensional matrix of blood pressure regression data;

Among them, the blood pressure regression data two-dimensional matrix is specifically the blood pressure regression data two-dimensional matrix [X, 2]; X is the second dimension parameter of the blood pressure regression data two-dimensional matrix [X, 2] and X is W ₃ ; the blood pressure regression data two The dimensional matrix [X, 2] includes X regression data one-dimensional data sequence [2]; the regression data one-dimensional data sequence [2] includes fragment systolic blood pressure data and fragment diastolic blood pressure data.

Here, the blood pressure ANN model is composed of fully connected layers. Each node of the fully connected layer is connected to all the nodes in the previous layer. It is used to integrate the features extracted from the front. Each fully connected layer can be set The number of nodes in this layer and the activation function (there are more ReLUs, you can also change to other), for example, if the number of nodes in the current connection layer is set to 512, the output of the current connection layer becomes [5,512], after several layers The connection layer converts the final output into a matrix of shape [X, 2]. The second dimension parameter X represents the total number of fragments, and the first dimension parameter 2 means that two regression calculation values are output for each fragment: respectively represent The systolic and diastolic blood pressure; the blood pressure ANN model is a mature model after training. Generally, the blood pressure CNN model and the blood pressure ANN model are connected together and trained with the same batch of training data;

Step 5. Obtain the prediction mode identifier;

Among them, the prediction mode identifier includes two types of identifiers: mean prediction and dynamic prediction.

Here, the prediction mode identifier is a system variable. The predicted value of blood pressure in multiple segments of time after the regression calculation of the blood pressure ANN model is completed. This variable can be used to clarify the further prediction output content: when the prediction mode identifier is the mean value When predicting, it means that the tester's average blood pressure data is required to be output within the acquisition time threshold; when the prediction mode identifier is dynamic prediction, it means that the tester's blood pressure change data sequence within the acquisition time threshold is required to be output.

Step 6. When the prediction mode identifier is average prediction, perform an average blood pressure calculation operation on the two-dimensional matrix of blood pressure regression data to generate an average blood pressure prediction data pair;

Specifically: Step 61, when the prediction mode identifier is average prediction, set the average blood pressure prediction data pair; initialize the diastolic blood pressure prediction data of the average blood pressure prediction data pair to be empty, and initialize the systolic blood pressure data of the average blood pressure prediction data pair to be empty ；

Step 62: Calculate the sum of the diastolic blood pressure data of all regression data one-dimensional data sequence [2] included in the two-dimensional matrix of blood pressure regression data [X,2] to generate the sum of diastolic blood pressure, which is generated according to the quotient of the sum of diastolic blood pressure divided by X The first mean; the systolic blood pressure data of all regression data one-dimensional data sequence [2] included in the two-dimensional matrix of blood pressure regression data [X,2] is calculated to generate the total systolic blood pressure, and the quotient of the total systolic blood pressure divided by X Generate the second mean;

Step 63: Set the diastolic blood pressure prediction data of the average blood pressure prediction data pair as the first average value; set the systolic blood pressure prediction data of the average blood pressure prediction data pair as the second average value.

Here, suppose that the two-dimensional matrix of blood pressure regression data [X,2] is the two-dimensional matrix of blood pressure regression data [5,2]={[D ₁₁ ,D ₁₂ ],[D ₂₁ ,D ₂₂ ],[D ₃₁ ,D ₃₂ ],[D ₄₁ ,D ₄₂ ],[D ₅₁ ,D ₅₂ ]}, then the one-dimensional data sequence of the regression data included in the two-dimensional matrix of blood pressure regression data [5,2] is: the first regression data one-dimensional Data sequence [2]={D ₁₁ ,D ₁₂ }, one-dimensional data sequence of the second regression data [2]={D ₂₁ ,D ₂₂ }, one-dimensional data sequence of the third regression data [2]={D ₃₁ , D ₃₂ }, one-dimensional data sequence of the fourth regression data [2]={D ₄₁ ,D ₄₂ }, one-dimensional data sequence of the fifth regression data [2]={D ₅₁ ,D ₅₂ }; where, each regression data The two values in the one-dimensional data sequence are the segment diastolic blood pressure data (smaller value) and segment systolic blood pressure data (larger value) corresponding to the current segment; when the prediction mode identifier is mean prediction, it means that only external The average blood pressure value within the output signal acquisition time threshold. Assuming that the acquisition time threshold is 10 seconds, the calculated mean blood pressure data in these 10 seconds is calculated at this time; assuming that D _X1 is the segment diastolic blood pressure data, the mean diastolic blood pressure data is predicted =(D ₁₁ +D ₂₁ +D ₃₁ +D ₄₁ +D ₅₁ )/5; assuming that DX2 is all segmental systolic blood pressure data, the predicted mean systolic blood pressure data = (D ₁₂ +D ₂₂ +D ₃₂ +D ₄₂ +D ₅₂ )/5.

FIG. 2 is a schematic diagram of a method for blood pressure prediction based on a synchronization signal according to Embodiment 2 of the present invention. The method mainly includes the following steps:

Step 101: Obtain the PPG signal data of the photoplethysmography method and the ECG signal data synchronized with it; and according to the data sampling frequency threshold, perform signal sampling processing on the PPG signal data and the ECG signal data to generate PPG one-dimensional data Sequence and ECG one-dimensional data sequence;

Step 102: Perform blood pressure CNN input data fusion processing on the PPG one-dimensional data sequence and the ECG one-dimensional data sequence according to the input data length threshold N of the blood pressure convolutional neural network CNN model to generate a four-dimensional tensor of input data;

Specifically: Step 1021: Calculate the total number of fragments M based on the input data length threshold N and A; when A is divisible by the input data length threshold N, set the total number of fragments M as the quotient of A divided by the input data length threshold N; when A cannot When divisible by the input data length threshold N, the total number of fragments M is set as the result of rounding the quotient of A divided by the input data length threshold N;

Step 1022: According to the total number of fragments M and the input data length threshold N, the PPG one-dimensional data sequence [A] is divided into sequential data fragments to generate a PPG fragment data two-dimensional matrix [M, N], and the ECG one-dimensional data sequence [A] ] Perform sequential data segmentation processing to generate a two-dimensional matrix of ECG segment data [M, N];

Specifically, it includes: Step 10221, generating a temporary data length L according to the product of the total number of fragments M and the input data length threshold N;

Step 10222: From the PPG one-dimensional data sequence [A], take the first data as the starting data extraction position and the temporary data length L as the data extraction length, extract a piece of continuous data to generate the PPG one-dimensional temporary data sequence [L] ; And according to the input data length threshold N, the PPG one-dimensional temporary data sequence [L] is divided into continuous data segments;

Step 10223: From the ECG one-dimensional data sequence [A], take the first data as the data start extraction position and the temporary data length L as the data extraction length, extract a piece of continuous data to generate the ECG one-dimensional temporary data sequence [L] ; And according to the input data length threshold N, the ECG one-dimensional temporary data sequence [L] is divided into continuous data segments;

Step 10224, construct a two-dimensional PPG fragment data matrix [M, N], and initialize all matrix elements of the PPG fragment data two-dimensional matrix [M, N] to be empty; extract sequentially from the PPG one-dimensional temporary data sequence [L] The PPG one-dimensional segment data sequence [N] included in the PPG one-dimensional temporary data sequence [L] assigns values to the matrix elements of the PPG segment data two-dimensional matrix [M, N];

Step 10225, construct a two-dimensional matrix of ECG fragment data [M, N], and initialize all matrix elements of the two-dimensional matrix of ECG fragment data [M, N] to be empty; extract sequentially from the ECG one-dimensional temporary data sequence [L] The ECG one-dimensional segment data sequence [N] assigns values to the matrix elements of the ECG segment data two-dimensional matrix [M, N];

Similar to step 10224, the two-dimensional matrix of ECG fragment data [M,N], although it is a two-dimensional matrix, the actual number of matrix elements M*N contained in it is the total number of data included in the ECG one-dimensional temporary data sequence [L] Equal, here is the tensor upscaling process of the shape of the ECG one-dimensional temporary data sequence [L] from the one-dimensional sequence;

Step 1023: Perform input data fusion processing on the two-dimensional matrix of PPG fragment data [M, N] and the two-dimensional matrix of ECG fragment data [M, N] to generate a two-dimensional matrix of fused fragment data [M, 2*N];

It specifically includes: Step 10231: Obtain a preset fusion sorting flag; construct a two-dimensional matrix of fusion fragment data [M, 2*N], and initialize the two-dimensional matrix of fusion fragment data [M, 2*N] to be empty; construct the first Sequence [2*N];

Step 10232: Initialize the value of the first index to 1, and initialize the value of the first total to the total number of fragments M;

Step 10233, using (first index-1)*input data length threshold N+1 as the starting data extraction position, and input data length threshold N as the extraction data length, from the two-dimensional matrix of PPG fragment data [M, N] Extract a piece of continuous data to generate the second sequence [N], and extract a piece of continuous data from the two-dimensional matrix of ECG fragment data [M, N] to generate the third sequence [N];

Step 10234: Use the second sequence [N] and the third sequence [N] to assign values to the first sequence [2*N] according to the fusion order identifier; when the fusion order identifier is PPG+ECG order, use the second sequence [N] Perform assignment processing on the first N sequence data included in the first sequence [2*N], and use the third sequence [N] to perform assignment processing on the last N sequence data included in the first sequence [2*N] ; When the fusion sorting flag is ECG+PPG sorting, use the third sequence [N] to assign values to the first N sequence data included in the first sequence [2*N], and use the second sequence [N] to The last N sequence data included in sequence [2*N] are assigned values;

Step 10235, using the first sequence [2*N] to perform a sequence data addition operation on the two-dimensional matrix [M, 2*N] of the fusion fragment data;

Step 10236, add 1 to the first index;

Here, the first index is the slice index extracted from the two-dimensional matrix [M, N] of PPG (ECG) slice data;

Step 10237: Determine whether the first index is greater than the first total. If the first index is greater than the first total, go to step 1024; if the first index is less than or equal to the first total, go to step 10233.

Step 1024: Perform four-dimensional tensor conversion processing of blood pressure CNN input data on the two-dimensional matrix [M,2*N] of the fused segment data to generate a four-dimensional tensor of input data [B ₁ , H ₁ , W ₁ , C ₁ ];

Step 103: According to the threshold of the number of convolution layers, use the blood pressure CNN model to perform multi-layer convolution pooling calculation on the four-dimensional tensor of the input data to generate a four-dimensional tensor of feature data;

Specifically: Step 1031, initialize the value of the second index to 1; initialize the second total to the threshold of the number of convolutional layers; initialize the temporary four-dimensional tensor of the second index to the four-dimensional tensor of the input data [B ₁ ,H ₁ ,W ₁ , C ₁ ];

Step 1032: Use the second index layer convolution layer of the blood pressure CNN model to perform convolution calculation processing on the second index temporary four-dimensional tensor to generate the second index convolution output data four-dimensional tensor; use the second index of the blood pressure CNN model Pooling layer, which performs pooling calculation processing on the four-dimensional tensor of the second index convolution output data to generate a four-dimensional tensor of second index pooling output data; the blood pressure CNN model includes a multi-layer convolutional layer and a multi-layer pooling layer;

Step 1033: Set the temporary four-dimensional tensor of the second index as the four-dimensional tensor of the second index pooling output data;

Step 1034, add 1 to the second index;

Step 1035: Determine whether the second index is greater than the second total, if the second index is greater than the second total, go to step 1036, if the second index is less than or equal to the second total, go to step 1032;

Step 1036: Set the four-dimensional tensor of feature data as a temporary four-dimensional tensor of the second index;

Step 104: Perform a blood pressure artificial neural network ANN input data two-dimensional matrix construction operation based on the feature data four-dimensional tensor to generate a two-dimensional matrix of input data; and use the blood pressure ANN model to perform feature data regression calculation on the two-dimensional matrix of input data to generate blood pressure regression data. Dimensional matrix

Specifically: Step 1041, according to the feature data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C ₂ ], perform tensor data reduction on the feature data four-dimensional tensor [B ₂ , H ₂ , W ₂ , C _2] Dimensional processing to construct a two-dimensional matrix of input data;

Step 1042, using the blood pressure ANN model to perform characteristic data regression calculation on the two-dimensional matrix of input data [W ₃ , C ₃ ] to generate a two-dimensional matrix of blood pressure regression data;

Here, the blood pressure ANN model is composed of fully connected layers. Each node of the fully connected layer is connected to all the nodes in the previous layer. It is used to integrate the features extracted from the front. Each fully connected layer can be set The number of nodes in this layer and the activation function (there are more ReLUs, you can also change to other), for example, if the number of nodes in the current connection layer is set to 512, the output of the current connection layer becomes [5,512], after several layers The connection layer converts the final output into a matrix of shape [X, 2]. The second dimension parameter X represents the total number of fragments, and the first dimension parameter 2 means that two regression calculation values are output for each fragment: respectively represent The systolic and diastolic blood pressure; the blood pressure ANN model is a mature model after training. Generally, the blood pressure CNN model and the blood pressure ANN model are connected together and trained with the same batch of training data.

Step 105: Obtain a prediction mode identifier;

Step 106: When the prediction mode identifier is dynamic prediction, perform an ambulatory blood pressure data extraction operation on a two-dimensional matrix of blood pressure regression data to generate a one-dimensional ambulatory blood pressure prediction data sequence;

Specifically include: step 1061, when the prediction mode identifier is dynamic prediction, initialize the ambulatory blood pressure prediction one-dimensional data sequence to be empty; set the blood pressure data group; initialize the diastolic blood pressure data of the blood pressure data group to be empty; initialize the systolic blood pressure of the blood pressure data group Data is empty;

Step 1062, sequentially extract the regression data one-dimensional data sequence [2] included in the blood pressure regression data two-dimensional matrix [X, 2] to generate the current data sequence [2]; set the systolic blood pressure data of the blood pressure data group as the current data sequence [2] Set the diastolic blood pressure data of the blood pressure data group as the fragment diastolic blood pressure data of the current data sequence [2]; add the blood pressure data group to the ambulatory blood pressure prediction one-dimensional data sequence for data group addition operation.

Here, suppose that the two-dimensional matrix of blood pressure regression data [X,2] is the two-dimensional matrix of blood pressure regression data [5,2]={[D ₁₁ ,D ₁₂ ],[D ₂₁ ,D ₂₂ ],[D ₃₁ ,D ₃₂ ],[D ₄₁ ,D ₄₂ ],[D ₅₁ ,D ₅₂ ]}, then the one-dimensional data sequence of the regression data included in the two-dimensional matrix of blood pressure regression data [5,2] is: the first regression data one-dimensional Data sequence [2]={D ₁₁ ,D ₁₂ }, one-dimensional data sequence of the second regression data [2]={D ₂₁ ,D ₂₂ }, one-dimensional data sequence of the third regression data [2]={D ₃₁ , D ₃₂ }, one-dimensional data sequence of the fourth regression data [2]={D ₄₁ ,D ₄₂ }, one-dimensional data sequence of the fifth regression data [2]={D ₅₁ ,D ₅₂ }; where, each regression data The two values in the one-dimensional data sequence are the segment diastolic blood pressure data (smaller value) and segment systolic blood pressure data (larger value) corresponding to the current segment;

After extraction, the one-dimensional data sequence of ambulatory blood pressure prediction should be the one-dimensional data sequence of ambulatory blood pressure prediction [5]={1st blood pressure data group, 2nd blood pressure data group, 3rd blood pressure data group, 4th blood pressure data group, No. 5 blood pressure data group}; wherein the first data diastolic blood pressure data group is [D _11, ₁₂ D] fragment diastolic pressure data, the first data systolic blood pressure data group is [D _11, D _12] of the The segmental systolic blood pressure data;...The diastolic blood pressure data of the fifth blood pressure data group is the segment systolic blood pressure data in [D ₅₁ ,D ₅₂ ], and the systolic blood pressure data of the fifth blood pressure data group is [D ₅₁ ,D ₅₂ ] Fragment systolic blood pressure data.

FIG. 3 is a schematic diagram of a device structure of an apparatus for performing blood pressure prediction based on a synchronization signal according to Embodiment 3 of the present invention. The device includes a processor and a memory. The memory can be connected to the processor through a bus. The memory may be a non-volatile memory, such as a hard disk drive and a flash memory, and a software program and a device driver program are stored in the memory. The software program can execute various functions of the above method provided by the embodiments of the present invention; the device driver may be a network and interface driver. The processor is used to execute a software program, and when the software program is executed, the method provided in the embodiment of the present invention can be implemented.

It should be noted that the embodiment of the present invention also provides a computer-readable storage medium. A computer program is stored on the computer-readable storage medium, and when the computer program is executed by a processor, the method provided in the embodiment of the present invention can be implemented.

The embodiment of the present invention also provides a computer program product containing instructions. When the computer program product runs on the computer, the processor is caused to execute the above method.

The embodiment of the present invention provides a method and device for blood pressure prediction based on synchronization signals, which perform feature data fusion on synchronized ECG signals and PPG signals, and then perform fusion data on the fusion data through a blood pressure prediction model composed of a blood pressure CNN model and a blood pressure ANN model Feature calculation and blood pressure data regression calculation to calculate the blood pressure data of the tester, and finally output the mean blood pressure prediction data pair (systolic blood pressure prediction data, diastolic blood pressure prediction data) or ambulatory blood pressure prediction one-dimensional data sequence through the selection of the prediction mode identifier . Through the embodiments of the present invention, the cumbersomeness and discomfort of conventional blood pressure collection methods are avoided, and an automatic and intelligent data analysis method is produced, so that the application party can conveniently perform multiple continuous monitoring of the measured object.

Professionals should be further aware that the units and algorithm steps of the examples described in the embodiments disclosed in this article can be implemented by electronic hardware, computer software, or a combination of the two, in order to clearly illustrate the hardware and software Interchangeability, in the above description, the composition and steps of each example have been generally described in accordance with the function. Whether these functions are executed by hardware or software depends on the specific application and design constraint conditions of the technical solution. Professionals and technicians can use different methods for each specific application to implement the described functions, but such implementation should not be considered as going beyond the scope of the present invention.

The steps of the method or algorithm described in combination with the embodiments disclosed herein can be implemented by hardware, a software module executed by a processor, or a combination of the two. The software module can be placed in random access memory (RAM), internal memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disks, removable disks, CD-ROMs, or all areas in the technical field. Any other known storage media.

The specific embodiments described above further describe the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention, and are not intended to limit the scope of the present invention. The scope of protection, any modification, equivalent replacement, improvement, etc., made within the spirit and principle of the present invention shall be included in the scope of protection of the present invention.

Claims

A method for predicting blood pressure based on a synchronization signal, characterized in that the method includes:

Acquire the PPG signal data of the photoplethysmography method and the ECG signal data synchronized with it; and according to the data sampling frequency threshold, perform signal sampling processing on the PPG signal data and the ECG signal data to generate a PPG one-dimensional data sequence And generate ECG one-dimensional data sequence;

According to the input data length threshold N of the blood pressure convolutional neural network CNN model, perform blood pressure CNN input data fusion processing on the PPG one-dimensional data sequence and the ECG one-dimensional data sequence to generate a four-dimensional tensor of input data;

According to the threshold of the number of convolutional layers, using the blood pressure CNN model to perform multi-layer convolution pooling calculation on the four-dimensional tensor of the input data to generate a four-dimensional tensor of feature data;

Perform a two-dimensional matrix construction operation of input data of a blood pressure artificial neural network ANN model according to the four-dimensional tensor of the feature data to generate a two-dimensional matrix of input data; and use the blood pressure ANN model to perform feature data regression calculation on the two-dimensional matrix of input data Generate a two-dimensional matrix of blood pressure regression data;

Obtaining a prediction mode identifier; the prediction mode identifier includes two identifiers of mean prediction and dynamic prediction;

When the prediction mode identifier is the average prediction, the average blood pressure calculation operation is performed on the two-dimensional matrix of blood pressure regression data to generate an average blood pressure prediction data pair; the average blood pressure prediction data pair includes diastolic blood pressure prediction data and systolic blood pressure prediction data. Pressure forecast data;

When the prediction mode identifier is the dynamic prediction, performing an ambulatory blood pressure data extraction operation on the two-dimensional matrix of blood pressure regression data to generate an ambulatory blood pressure prediction one-dimensional data sequence.
The method for blood pressure prediction based on a synchronization signal according to claim 1, wherein:

The PPG signal data is generated by using a non-invasive PPG signal acquisition device to perform a preset light source signal acquisition operation on the tester's local skin surface within a signal acquisition time threshold;

The ECG signal data is synchronized with the collection of the PPG signal data, and is generated by using a non-invasive ECG signal collection device to perform an electrophysiological signal collection operation on the tester within the signal collection time threshold;

The PPG one-dimensional data sequence is specifically a PPG one-dimensional data sequence [A]; the PPG one-dimensional data sequence [A] includes the A pieces of PPG data; and the A is the data sampling frequency threshold multiplied by the The product of the signal acquisition time threshold;

The ECG one-dimensional data sequence is specifically an ECG one-dimensional data sequence [A]; the ECG one-dimensional data sequence [A] includes the A pieces of ECG data.
The method for blood pressure prediction based on synchronization signals according to claim 2, characterized in that, according to the input data length threshold N of the blood pressure convolutional neural network CNN model, the PPG one-dimensional data sequence and the ECG one Dimensional data sequence, perform blood pressure CNN input data fusion processing to generate a four-dimensional tensor of input data, including:

Calculate the total number of fragments M according to the input data length threshold N and the A; when the A is divisible by the input data length threshold N, set the total number of fragments M as the A divided by the input data length The quotient of the threshold N; when the A is not divisible by the input data length threshold N, the total number of fragments M is set as the result of the rounding calculation of the quotient of the A divided by the input data length threshold N;

According to the total number of fragments M and the input data length threshold N, the PPG one-dimensional data sequence [A] is subjected to sequential data fragment division processing to generate a PPG fragment data two-dimensional matrix [M, N], and the ECG one Dimensional data sequence [A] performs sequential data segmentation processing to generate a two-dimensional matrix of ECG segment data [M,N];

Perform input data fusion processing on the two-dimensional matrix of PPG fragment data [M, N] and the two-dimensional matrix of ECG fragment data [M, N] to generate a two-dimensional matrix of fused fragment data [M, 2*N];

Perform a four-dimensional tensor conversion process of blood pressure CNN input data on the two-dimensional matrix [M,2*N] of the fusion segment data to generate the four-dimensional tensor [B 1 , H 1 , W 1 , C 1 ] of the input data; The B 1 is the fourth dimension parameter of the input data four-dimensional tensor [B 1 , H 1 , W 1 , C 1 ], and the B 1 is the total number of fragments M; the H 1 is the The input data four-dimensional tensor [B 1 , H 1 , W 1 , C 1 ] is the third-dimensional parameter, and the value of H 1 is 2; the W 1 is the input data four-dimensional tensor [B 1 , H 1 , W 1 , C 1 ] second dimension parameter, and the W 1 is the input data length threshold N; the C 1 is the input data four-dimensional tensor [B 1 , H 1 , W 1 , C 1 ] is the first dimension parameter, and the value of C 1 is 1.
The method for blood pressure prediction based on a synchronization signal according to claim 3, wherein the sequence of the PPG one-dimensional data sequence [A] is performed according to the total number of fragments M and the input data length threshold N. The data segment division processing generates a two-dimensional PPG segment data matrix [M, N], and the sequential data segment division processing is performed on the ECG one-dimensional data sequence [A] to generate a two-dimensional ECG segment data matrix [M, N], which specifically includes:

Generate a temporary data length L according to the product of the total number of fragments M and the input data length threshold N;

From the PPG one-dimensional data sequence [A], taking the first data as the starting data extraction position and the temporary data length L as the data extraction length, extract a piece of continuous data to generate the PPG one-dimensional temporary data sequence [L ]; and according to the input data length threshold N, the PPG one-dimensional temporary data sequence [L] is subjected to continuous data segment division processing; the PPG one-dimensional temporary data sequence [L] includes the total number of segments M PPG one One-dimensional segment data sequence [N]; the PPG one-dimensional segment data sequence [N] includes the input data length threshold N pieces of the PPG data;

From the ECG one-dimensional data sequence [A], taking the first data as the starting data extraction position and the temporary data length L as the data extraction length, extract a piece of continuous data to generate an ECG one-dimensional temporary data sequence [L ]; and according to the input data length threshold N, the ECG one-dimensional temporary data sequence [L] is subjected to continuous data segment division processing; the ECG one-dimensional temporary data sequence [L] includes the total number of segments M ECG one One-dimensional fragment data sequence [N]; the ECG one-dimensional fragment data sequence [N] includes the input data length threshold value N of the ECG data;

Construct the PPG fragment data two-dimensional matrix [M, N], and initialize all matrix elements of the PPG fragment data two-dimensional matrix [M, N] to be empty; from the PPG one-dimensional temporary data sequence [L] , Sequentially extracting the PPG one-dimensional segment data sequence [N] included in the PPG one-dimensional temporary data sequence [L], and assigning values to the matrix elements of the PPG segment data two-dimensional matrix [M, N];

Construct the ECG fragment data two-dimensional matrix [M, N], and initialize all matrix elements of the ECG fragment data two-dimensional matrix [M, N] to be empty; from the ECG one-dimensional temporary data sequence [L] , Sequentially extracting the ECG one-dimensional segment data sequence [N] to perform value assignment processing on the matrix elements of the ECG segment data two-dimensional matrix [M, N].
The method for blood pressure prediction based on a synchronization signal according to claim 3, wherein said pair of said PPG segment data two-dimensional matrix [M, N] and said ECG segment data two-dimensional matrix [M, N] , Perform input data fusion processing to generate a two-dimensional matrix of fused fragment data [M,2*N], which specifically includes:

Step 51: Obtain a preset fusion sorting flag; construct the two-dimensional matrix of fusion fragment data [M, 2*N], and initialize the two-dimensional matrix of fusion fragment data [M, 2*N] to be empty; construct the first A sequence [2*N]; the fusion sequence identifier is one of PPG+ECG sequence and ECG+PPG sequence; the first sequence [2*N] includes 2*N sequence data;

Step 52: Initialize the value of the first index to 1, and initialize the value of the first total to the total number of fragments M;

Step 53: Taking the product of the difference of the first index minus 1 and the input data length threshold N plus 1 as the starting data extraction position, and the input data length threshold N as the extraction data length. Extract a piece of continuous data from the two-dimensional matrix of PPG fragment data [M, N] to generate a second sequence [N], and extract a piece of continuous data from the two-dimensional matrix of ECG fragment data [M, N] to generate a third sequence [ N];

Step 54: Use the second sequence [N] and the third sequence [N] to perform value assignment processing on the first sequence [2*N] according to the fusion order identifier; when the fusion order identifier is When the PPG+ECG is sorted, use the second sequence [N] to perform the assignment processing on the first N sequence data included in the first sequence [2*N], and use the third sequence [N ] Perform value assignment processing on the last N sequence data included in the first sequence [2*N]; when the fusion sequence identifier is the ECG+PPG sequence, use the third sequence [N] pair The first N sequence data included in the first sequence [2*N] are assigned values, and the second sequence [N] is used to perform the assignment processing on the last N data included in the first sequence [2*N]. Assign value to the sequence data;

Step 55: Use the first sequence [2*N] to perform a sequence data addition operation on the two-dimensional matrix [M, 2*N] of fusion fragment data;

Step 56: Add 1 to the first index;

Step 57: Determine whether the first index is greater than the first total, if the first index is greater than the first total, go to step 58, if the first index is less than or equal to the first total, go to Step 53;

Step 58: Return the two-dimensional matrix [M, 2*N] of the fused segment data to the upper application.
The method for blood pressure prediction based on a synchronization signal according to claim 3, characterized in that the four-dimensional tensor conversion processing of blood pressure CNN input data is performed on the two-dimensional matrix [M, 2*N] of the fusion segment data, Generating the input data four-dimensional tensor [B 1 , H 1 , W 1 , C 1 ] specifically includes:

Construct the input data four-dimensional tensor [B 1 , H 1 , W 1 , C 1 ], and initialize the input data four-dimensional tensor [B 1 , H 1 , W 1 , C 1 ] to be empty; then extract sequentially The matrix elements included in the two-dimensional matrix [M, 2*N] of the fusion fragment data perform a data item addition operation on the four-dimensional tensor [B 1 , H 1 , W 1 , C 1 ] of the input data; the input data The four-dimensional tensor [B 1 , H 1 , W 1 , C 1 ] is specifically the input data four-dimensional tensor [M,2,N,1].
The method for blood pressure prediction based on synchronization signals according to claim 3, characterized in that, according to the threshold of the number of convolutional layers, the blood pressure CNN model is used to perform multi-layer convolution pooling on the four-dimensional tensor of the input data Calculate and generate a four-dimensional tensor of feature data, including:

Step 71: Initialize the value of the second index to 1; initialize the second total to the threshold of the number of convolutional layers; initialize the temporary four-dimensional tensor of the second index to the four-dimensional tensor of the input data [B 1 , H 1 , W 1 ,C 1 ];

Step 72: Use the second index layer convolution layer of the blood pressure CNN model to perform convolution calculation processing on the second index temporary four-dimensional tensor to generate a second index convolution output data four-dimensional tensor; use the blood pressure The second index pooling layer of the CNN model performs pooling calculation processing on the four-dimensional tensor of the second index convolution output data to generate the four-dimensional tensor of the second index pooling output data; the blood pressure CNN model includes multiple layers. Said convolutional layer and multiple layers of said pooling layer;

Step 73: Set the temporary four-dimensional tensor of the second index as the four-dimensional tensor of the second index pooling output data;

Step 74: Add 1 to the second index;

Step 75: Determine whether the second index is greater than the second total, if the second index is greater than the second total, go to step 76, if the second index is less than or equal to the second total, go to Step 72;

Step 76: Set the feature data four-dimensional tensor as the second index temporary four-dimensional tensor; the feature data four-dimensional tensor is specifically the feature data four-dimensional tensor [B 2 , H 2 , W 2 , C 2 ]; The B2 is the fourth dimension parameter of the characteristic data four-dimensional tensor [B 2 , H 2 , W 2 , C 2 ] and the B 2 is the B 1 ; the H 2 is the characteristic data four-dimensional The third dimension parameter of the tensor [B 2 , H 2 , W 2 , C 2 ]; the W 2 is the second dimension of the feature data four-dimensional tensor [B 2 , H 2 , W 2 , C 2 ] Parameters; The C 2 is the first dimension parameter of the characteristic data four-dimensional tensor [B 2 , H 2 , W 2 , C 2 ].
The method for blood pressure prediction based on a synchronization signal according to claim 7, wherein the two-dimensional matrix construction operation of the input data of the blood pressure artificial neural network ANN model is performed according to the four-dimensional tensor of the characteristic data to generate the two-dimensional input data Matrix; and using the blood pressure ANN model to perform feature data regression calculation on the input data two-dimensional matrix to generate a blood pressure regression data two-dimensional matrix, which specifically includes:

According to the feature data four-dimensional tensor [B 2 , H 2 , W 2 , C 2 ], perform tensor data dimensionality reduction processing on the feature data four-dimensional tensor [B 2 , H 2 , W 2 , C 2] Construct the input data two-dimensional matrix; the input data two-dimensional matrix is specifically the input data two-dimensional matrix [W 3 , C 3 ]; the W 3 is the input data two-dimensional matrix [W 3 , C 3 ] The second dimension parameter of the W 3 is the B 2 ; the C 3 is the first dimension parameter of the input data two-dimensional matrix [W 3 , C 3 ] and the C 3 is the H 2 Multiply by the product of W 2 and then by the product of C 2;

Using the blood pressure ANN model, perform characteristic data regression calculation on the input data two-dimensional matrix [W 3 , C 3 ] to generate a blood pressure regression data two-dimensional matrix; the blood pressure regression data two-dimensional matrix is specifically blood pressure regression data two-dimensional Matrix [X, 2]; the X is the second dimension parameter of the blood pressure regression data two-dimensional matrix [X, 2] and the X is the W 3 ; the blood pressure regression data two-dimensional matrix [X, 2] includes the X regression data one-dimensional data sequence [2]; the regression data one-dimensional data sequence [2] includes the fragment systolic blood pressure data and the fragment diastolic blood pressure data.
The method for blood pressure prediction based on a synchronization signal according to claim 8, wherein when the prediction mode identifier is the average prediction, the average blood pressure is calculated on the two-dimensional matrix of blood pressure regression data Operate to generate mean blood pressure prediction data pair, including:

When the prediction mode identifier is the average prediction, the average blood pressure prediction data pair is set; and the diastolic blood pressure prediction data of the average blood pressure prediction data pair is initialized to be empty, and the average blood pressure prediction data pair is initialized The systolic blood pressure data of is empty;

The sum of the diastolic blood pressure data of all the pieces of the regression data one-dimensional data sequence [2] included in the blood pressure regression data two-dimensional matrix [X, 2] is calculated to generate the sum of diastolic blood pressure, which is divided by the sum of diastolic blood pressure Use the quotient of X to generate the first mean value; generate the sum of the systolic blood pressure data of all the regression data one-dimensional data sequence [2] included in the blood pressure regression data two-dimensional matrix [X, 2] The sum of systolic blood pressures, the second mean value is generated according to the quotient of the sum of systolic blood pressures divided by the X;

Set the diastolic blood pressure prediction data of the average blood pressure prediction data pair as the first average value; set the systolic blood pressure prediction data of the average blood pressure prediction data pair as the second average value.
The method for blood pressure prediction based on a synchronization signal according to claim 8, wherein when the prediction mode identifier is the dynamic prediction, perform dynamic blood pressure data on the two-dimensional matrix of blood pressure regression data The extraction operation generates a one-dimensional data sequence of ambulatory blood pressure prediction, which specifically includes:

When the prediction mode identifier is the dynamic prediction, initialize the ambulatory blood pressure prediction one-dimensional data sequence to be empty; set the blood pressure data group; initialize the diastolic blood pressure data of the blood pressure data group to be empty; initialize the blood pressure data The systolic blood pressure data of the group is empty;

The one-dimensional data sequence [2] of the regression data included in the two-dimensional matrix [X, 2] of the blood pressure regression data is sequentially extracted to generate the current data sequence [2]; the systolic blood pressure data of the blood pressure data group is set to all The segment systolic blood pressure data of the current data sequence [2], the diastolic blood pressure data of the blood pressure data group is set to the segment diastolic blood pressure data of the current data sequence [2]; and the blood pressure data The group performs a data group addition operation to the ambulatory blood pressure prediction one-dimensional data sequence.
A device comprising a memory and a processor, wherein the memory is used to store a program, and the processor is used to execute the method according to any one of claims 1 to 10.
A computer program product containing instructions that, when run on a computer, causes the computer to execute the method according to any one of claims 1 to 10.
A computer-readable storage medium, comprising instructions, which when run on a computer, cause the computer to execute the method according to any one of claims 1 to 10.