CN111248882A

CN111248882A - Method and device for predicting blood pressure

Info

Publication number: CN111248882A
Application number: CN202010110287.9A
Authority: CN
Inventors: 张碧莹; 曹君
Original assignee: Lepu Medical Technology Beijing Co Ltd
Current assignee: Lepu Medical Technology Beijing Co Ltd
Priority date: 2020-02-21
Filing date: 2020-02-21
Publication date: 2020-06-09
Anticipated expiration: 2040-02-21
Also published as: CN111248882B; WO2021164346A1

Abstract

The embodiment of the invention relates to a method and a device for predicting blood pressure, wherein the method comprises the following steps: converting the PPG signal to generate a pulse wave one-dimensional vector; dividing the pulse wave one-dimensional vector into a plurality of pulse wave one-dimensional segments and acquiring the total number of the segments; dividing the pulse wave one-dimensional segment into a plurality of pulse wave one-dimensional sub-segments and acquiring the total number of the sub-segments; acquiring a CNN scheme identifier; when the CNN scheme identifier is a first scheme, performing all-direction volume convolution pooling calculation on the pulse wave one-dimensional vector and reducing the dimension of the result to generate a first LSTM input three-dimensional tensor; when the CNN scheme identifier is a second scheme, carrying out segmented vector convolution pooling calculation on the pulse wave one-dimensional vector and reducing the dimension of the result to generate a second LSTM input three-dimensional tensor; performing blood pressure long-term and short-term memory calculation on the first or second LSTM input three-dimensional tensor to generate an LSTM output three-dimensional tensor; performing blood pressure regression calculation on the LSTM output three-dimensional tensor to generate a blood pressure prediction three-dimensional tensor [ X, Y, 2 ]; a set of blood pressure prediction data is generated.

Description

Method and device for predicting blood pressure

Technical Field

The invention relates to the technical field of electrophysiological signal processing, in particular to a method and a device for predicting blood pressure.

Background

The heart is the center of human blood circulation, and the heart produces blood pressure through regular pulsation, and then supplies blood to the whole body to complete the metabolism of the human body, and blood pressure is one of the very important physiological signals of the human body. The blood pressure in the normal range can ensure the normal circulation and flow of blood, and the blood pressure can be kept normal under the combined action of a plurality of factors, so that each organ and tissue of the human body can obtain enough blood volume, and the normal operation of the human body is further kept. Human blood pressure contains two important values: systolic pressure and diastolic pressure, and whether the blood pressure of a human body is normal or not is judged by the two quantities medically. The long-term continuous observation of the two parameters of the blood pressure can help people to have clear understanding on the health state of the heart. However, most of the conventional blood pressure measuring methods adopt an external force pressing detection method such as a pressure gauge, which is not only cumbersome to operate, but also easily causes discomfort to the person to be measured, and thus cannot be used for multiple times to achieve the purpose of continuous monitoring.

Disclosure of Invention

The invention aims to provide a method and a device for predicting blood pressure aiming at the defects of the prior art, firstly, a Photoplethysmography (PPG) method is used for carrying out data acquisition on a tester and converting the data to generate pulse wave signal data, then a feature extraction regression fusion Network consisting of a blood pressure Convolutional Neural Network (CNN) model and a blood pressure Long-Short Term Memory (LSTM) model is used for carrying out feature extraction operation on the pulse wave signal data to generate blood pressure feature data, and the blood pressure feature data is subjected to regression calculation so as to predict the blood pressure data (diastolic pressure and systolic pressure) of the tester; the convolution scheme of the embodiment of the invention has two types, one type is that a full vector convolution mode is used if the vector length is not enough, and a segmented vector convolution mode can be adopted if the vector length is long enough; the embodiment of the invention not only avoids the complexity and the uncomfortable feeling of the conventional testing means, but also generates an automatic intelligent data analysis method, thereby leading an application party to conveniently and continuously monitor the tested object for many times.

In order to achieve the above object, a first aspect of the embodiments of the present invention provides a method for predicting blood pressure, where the method includes:

carrying out pulse wave conversion and sampling processing on PPG signal data by a photoplethysmography to generate a pulse wave one-dimensional vector; dividing the pulse wave one-dimensional vector into a plurality of pulse wave one-dimensional segments and acquiring the total number of the segments; dividing the pulse wave one-dimensional segment into a plurality of pulse wave one-dimensional sub-segments and acquiring the total number of the sub-segments;

acquiring a Convolutional Neural Network (CNN) scheme identifier; the CNN scheme identifier comprises two identifiers of a first scheme and a second scheme;

when the CNN scheme identifier is the first scheme, performing full-vector convolution pooling calculation operation on the pulse wave one-dimensional vector by using a blood pressure CNN model, and performing tensor dimensionality reduction processing on a calculation result according to an input parameter format of a blood pressure long-short term memory network (LSTM) to generate a first LSTM input three-dimensional tensor;

when the CNN scheme identifier is the second scheme, performing segmented vector convolution pooling calculation operation on the pulse wave one-dimensional vector by using the blood pressure CNN model, and performing tensor dimension reduction processing on a calculation result according to an input parameter format of the blood pressure LSTM to generate a second LSTM input three-dimensional tensor;

according to the CNN scheme identifier, performing blood pressure long-term and short-term memory calculation operation on the first LSTM input three-dimensional tensor or the second LSTM input three-dimensional tensor by using an LSTM network layer of the blood pressure LSTM model to generate an LSTM output three-dimensional tensor; the blood pressure LSTM network model comprises an LSTM network layer and a full connection layer;

performing blood pressure regression calculation operation on the LSTM output three-dimensional tensor by using the full connection layer to generate a blood pressure prediction three-dimensional tensor [ X, Y, 2 ]; the X is the total number of the fragments; the Y is the total number of the sub-segments;

and according to the sequence of the pulse wave one-dimensional segments and the sequence of the pulse wave one-dimensional sub-segments, sequentially extracting predicted blood pressure data from the blood pressure prediction three-dimensional tensor [ X, Y, 2] to generate a blood pressure prediction data set.

Preferably, the pulse wave conversion and sampling processing are carried out on the photoplethysmography PPG signal data to generate a pulse wave one-dimensional vector; dividing the pulse wave one-dimensional vector into a plurality of pulse wave one-dimensional segments and acquiring the total number of the segments; dividing the pulse wave one-dimensional segment into a plurality of pulse wave one-dimensional sub-segments and acquiring the total number of the sub-segments, specifically comprising:

calling PPG signal acquisition equipment, carrying out light intensity signal acquisition on a preset light source signal on the local skin surface of the organism, and generating PPG signal data with a length of a signal acquisition time threshold; performing pulse wave data conversion operation on the PPG signal data to generate pulse wave signal data; sampling feature data of the pulse wave signal data according to a feature sampling frequency threshold value to generate a pulse wave one-dimensional vector; the preset light source signals at least comprise one type of red light source signals, infrared light source signals and green light source signals;

performing data segment division on the pulse wave one-dimensional vector according to a segment length threshold value to generate a plurality of pulse wave one-dimensional segments, and taking the total number of the pulse wave one-dimensional segments included in the pulse wave one-dimensional vector as the total number of the segments;

and performing data sub-segment division on the pulse wave one-dimensional segments according to sub-segment length thresholds to generate a plurality of pulse wave one-dimensional sub-segments, and taking the total number of the pulse wave one-dimensional sub-segments included in the pulse wave one-dimensional segments as the total number of the sub-segments.

Preferably, when the CNN scheme identifier is the first scheme, the performing, by using a blood pressure CNN model, a full-vector convolution pooling calculation operation on the pulse wave one-dimensional vector, and performing tensor dimensionality reduction on a calculation result according to an input parameter format of a blood pressure long-short term memory network LSTM to generate a first LSTM input three-dimensional tensor specifically includes:

when the CNN scheme identifier is the first scheme, performing a first blood pressure CNN input parameter setting operation according to the total number of the fragments and the pulse wave one-dimensional vector to generate a first CNN input four-dimensional tensor;

according to a preset convolution layer number threshold value, performing multilayer convolution pooling calculation on the first CNN input four-dimensional tensor by using the blood pressure CNN model to generate a first CNN output four-dimensional tensor;

and carrying out tensor dimensionality reduction on the four-dimensional tensor output by the first CNN according to the input parameter format of the blood pressure long-short term memory network LSTM to generate the first LSTM input three-dimensional tensor.

Further, when the CNN scheme identifier is the first scheme, performing a first blood pressure CNN input parameter setting operation according to the total number of the segments and the pulse wave one-dimensional vector to generate a first CNN input four-dimensional tensor, specifically including:

when the CNN scheme identifier is the first scheme, setting the first CNN input four-dimensional tensor to be specifically the first CNN input four-dimensional tensor [ B ]₁，1，W₁，1](ii) a The first CNN inputs a four-dimensional tensor [ B ]₁，1，W₁，1]Comprising the said B₁The first CNN inputs the three-dimensional tensor [1, W₁，1](ii) a B is₁Inputting a four-dimensional tensor [ B ] for the first CNN₁，1，W₁，1]And said B is a fourth dimension parameter of₁Is the total number of fragments; the W is₁Inputting a four-dimensional tensor [ B ] for the first CNN₁，1，W₁，1]And said W is a second dimension parameter of₁Is the segment length threshold;

sequentially extracting the pulse wave one-dimensional segments included in the pulse wave one-dimensional vector, and inputting a four-dimensional tensor [ B ] to the first CNN₁，1，W₁，1]The corresponding first CNN inputs a three-dimensional tensor [1, W₁，1]And carrying out matrix element assignment processing.

Further, the performing, according to a preset convolution layer number threshold, a multilayer convolution pooling calculation on the first CNN input four-dimensional tensor by using the blood pressure CNN model to generate a first CNN output four-dimensional tensor specifically includes:

step 51, initializing the value of the first index to 1; initializing a first total number as the threshold of the number of convolution layers; initializing a first indexed temporary four-dimensional tensor as the first CNN input four-dimensional tensor;

step 52, performing convolution calculation processing on the first index temporary four-dimensional tensor by using a first index layer convolution layer of the blood pressure CNN model to generate a first index convolution output data four-dimensional tensor; performing pooling calculation processing on the first index convolution output data four-dimensional tensor by using a first index layer pooling layer of the blood pressure CNN model to generate a first index pooling output data four-dimensional tensor; the blood pressure CNN model comprises a plurality of layers of the convolutional layer and a plurality of layers of the pooling layer;

step 53, setting the first index temporary four-dimensional tensor as the first index pooled output data four-dimensional tensor;

step 54, adding 1 to the first index;

step 55, determining whether the first index is greater than the first total number, if the first index is greater than the first total number, going to step 56, and if the first index is less than or equal to the first total number, going to step 52;

step 56, setting the first CNN output four-dimensional tensor as the first index temporary four-dimensional tensor; the first CNN output four-dimensional tensor is specifically a first CNN output four-dimensional tensor [ B ]₂，1，W₂，C₂](ii) a B is₂Outputting a four-dimensional tensor [ B ] for the first CNN₂，1，W₂，C₂]And said B is a fourth dimension parameter of₂Is the total number of fragments; the W is₂Outputting a four-dimensional tensor [ B ] for the first CNN₂，1，W₂，C₂]And said W is a second dimension parameter of₂Is a preset LSTM layer neuron total number threshold value; said C is₂Outputting a four-dimensional tensor [ B ] for the first CNN₂，1，W₂，C₂]And said C is a first dimension parameter of₂Is a preset LSTM layer neuron length threshold.

Further, the first LSTM input three-dimensional tensor is specifically a first LSTM input three-dimensional tensor [ H₃，W₃，C₃](ii) a Said H₃Inputting a three-dimensional tensor [ H ] for the first LSTM₃，W₃，C₃]And the third dimension of (2), and the H₃Is the total number of fragments; the W is₃Inputting a three-dimensional tensor [ H ] for the first LSTM₃，W₃，C₃]And said W is a second dimension parameter of₃Is the said W₂(ii) a Said C is₃Inputting a three-dimensional tensor [ H ] for the first LSTM₃，W₃，C₃]And said C is a first dimension parameter of₃Is the C₂。

Preferably, when the CNN scheme identifier is the second scheme, the blood pressure CNN model is used to perform a segmented vector convolution pooling calculation operation on the pulse wave one-dimensional vector, and tensor dimensionality reduction processing is performed on a calculation result according to an input parameter format of the blood pressure LSTM to generate the second LSTM input three-dimensional tensor, specifically including:

when the CNN scheme identifier is the second scheme, generating a tensor total number according to a product of the total number of the fragments multiplied by the total number of the sub-fragments;

according to the total number of the tensors and the pulse wave one-dimensional vector, performing a second blood pressure CNN input parameter setting operation to generate a second CNN input four-dimensional tensor group; the set of second CNN input four-dimensional tensors comprises a total number of second CNN input four-dimensional tensors of the tensor;

according to a preset convolution layer number threshold value, utilizing the blood pressure CNN model to respectively perform multilayer convolution pooling calculation on all second CNN input four-dimensional tensors in the second CNN input four-dimensional tensor group to generate a second CNN output four-dimensional tensor group; the set of second CNN output four-dimensional tensors includes a total number of second CNN output four-dimensional tensors of the tensor;

performing four-dimensional tensor merging operation on all the second CNN output four-dimensional tensors in the second CNN output four-dimensional tensor group to generate a third CNN output four-dimensional tensor;

and carrying out tensor dimensionality reduction on the fourth-dimensional tensor output by the third CNN according to the input parameter format of the blood pressure LSTM to generate the second LSTM input three-dimensional tensor.

Further, the setting operation of the input parameters of the second blood pressure CNN is performed according to the total number of tensors and the one-dimensional vector of the pulse wave to generate a second CNN input four-dimensional tensor group, which specifically includes:

sequencing all pulse wave one-dimensional sub-segments of the pulse wave one-dimensional vector to generate a sub-segment complete sequence; the full sequence of sub-segments comprises a total number of the pulse wave one-dimensional sub-segments of the tensor;

setting the second CNN input four-dimensional tensor group; setting the second CNN input four-dimensional tensor as the second CNN inputTensor [1, 1, W ] in four dimensions₄，1](ii) a The set of second CNN-input four-dimensional tensors includes a total number of the tensors of the second CNN-input four-dimensional tensor [1, 1, W₄，1](ii) a The W is₄Inputting a four-dimensional tensor [1, 1, W ] for the second CNN₄，1]And said W is a second dimension parameter of₄Is the sub-segment length threshold;

sequentially extracting the pulse wave one-dimensional sub-segments in the sub-segment complete sequence, and inputting a four-dimensional tensor [1, 1, W ] into a corresponding second CNN in a four-dimensional tensor group to the second CNN₄，1]And carrying out matrix element assignment processing.

Further, the second CNN output four-dimensional tensor group specifically includes a total number of the tensors of the second CNN output four-dimensional tensor; the second CNN output four-dimensional tensor is specifically a second CNN output four-dimensional tensor [1, 1, 1, C₅](ii) a Said C is₅Outputting a four-dimensional tensor [1, 1, 1, C for the second CNN₅]Is measured.

Further, the performing a four-dimensional tensor merging operation on all the second CNN output four-dimensional tensors in the second CNN output four-dimensional tensor group to generate a third CNN output four-dimensional tensor specifically includes:

setting the third CNN output four-dimensional tensor as a third CNN output four-dimensional tensor [ B ]₆，1，1，C₆](ii) a B is₆Outputting a four-dimensional tensor [ B ] for the third CNN₆，1，1，C₆]And said B is a fourth dimension parameter of₆Is the total number of tensors; said C is₆Outputting a four-dimensional tensor [ B ] for the third CNN₆，1，1，C₆]And said C is a first dimension parameter of₆Is the C₅；

Sequentially extracting the four-dimensional tensor [1, 1, 1, C ] output by the second CNN from the four-dimensional tensor group output by the second CNN₅]Outputs a four-dimensional tensor [ B ] to the third CNN₆，1，1，C₆]And carrying out matrix element assignment processing.

Further, the second LSTM input three-dimensional tensor is specifically an LSTM input three-dimensional tensor[H₇，W₇，C₇](ii) a Said H₇Inputting a three-dimensional tensor [ H ] for the second LSTM₇，W₇，C₇]And the third dimension of (2), and the H₃The value of (d) is the total number of segments; the W is₇Inputting a three-dimensional tensor [ H ] for the second LSTM₇，W₇，C₇]And said W is a second dimension parameter of₇Is the said B₆Is divided by the H₃Quotient of (d); said C is₇Inputting a three-dimensional tensor [ H ] for the second LSTM₇，W₇，C₇]And said C is a first dimension parameter of₇Is the C₆。

Preferably, the generating an LSTM output three-dimensional tensor by performing a blood pressure long-term and short-term memory calculation operation on the first LSTM input three-dimensional tensor or the second LSTM input three-dimensional tensor by using the LSTM network layer of the blood pressure LSTM model according to the CNN scheme identifier specifically includes:

when the CNN scheme identifier is the first scheme, performing the blood pressure long-term and short-term memory calculation operation on the first LSTM input three-dimensional tensor by using an LSTM network layer of the blood pressure LSTM model to generate an LSTM output three-dimensional tensor;

and when the CNN scheme identifier is the second scheme, performing the blood pressure long-term and short-term memory calculation operation on the second LSTM input three-dimensional tensor by using the LSTM network layer of the blood pressure LSTM model to generate the LSTM output three-dimensional tensor.

Preferably, the sequentially extracting predicted blood pressure data from the three-dimensional tensor [ X, Y, 2] for blood pressure prediction according to the sequence of the pulse wave one-dimensional segments and the sequence of the pulse wave one-dimensional subfragments to generate a set of blood pressure prediction data specifically includes:

initializing the blood pressure prediction data set to be empty; setting a blood pressure data set; initializing the diastolic blood pressure data of the blood pressure data set to null; initializing the systolic blood pressure data of the blood pressure data group to be null;

sequentially extracting a blood pressure prediction one-dimensional vector [2] included in the blood pressure prediction three-dimensional tensor [ X, Y, 2] to generate a current one-dimensional vector [2 ]; setting the systolic pressure data of the blood pressure data group as sub-segment systolic pressure data in the current one-dimensional vector [2], and setting the diastolic pressure data of the blood pressure data group as sub-segment diastolic pressure data in the current one-dimensional vector [2 ]; performing a data set addition operation on the blood pressure data set to the blood pressure prediction data set; said blood pressure prediction three-dimensional tensor [ X, Y, 2] comprises X X Y of said blood pressure prediction one-dimensional vectors [2 ]; the blood pressure prediction one-dimensional vector [2] includes the sub-segment systolic pressure data and the sub-segment diastolic pressure data.

In the method for predicting blood pressure provided by the first aspect of the embodiment of the invention, PPG is used for carrying out data acquisition on a tester and converting the data to generate pulse wave signal data, then a feature extraction regression fusion network consisting of a blood pressure CNN model and a blood pressure LSTM model is used for carrying out feature extraction operation on the pulse wave signal data to generate blood pressure feature data, and regression calculation is carried out on the blood pressure feature data so as to predict the blood pressure data (diastolic pressure and systolic pressure) of the tester; the convolution scheme of the embodiment of the invention has two types, one type is that a full vector convolution mode is used if the vector length is not enough, and a segmented vector convolution mode can be adopted if the vector length is long enough; the embodiment of the invention not only avoids the complexity and the uncomfortable feeling of the conventional testing means, but also generates an automatic intelligent data analysis method, thereby leading an application party to conveniently and continuously monitor the tested object for many times.

A second aspect of an embodiment of the present invention provides an apparatus, which includes a memory and a processor, where the memory is used to store a program, and the processor is used to execute the first aspect and the method in each implementation manner of the first aspect.

A third aspect of embodiments of the present invention provides a computer program product including instructions, which, when run on a computer, cause the computer to perform the first aspect and the method in each implementation manner of the first aspect.

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

Drawings

Fig. 1 is a schematic diagram of a method for predicting blood pressure according to an embodiment of the present invention;

FIG. 2 is a schematic diagram of a method for predicting blood pressure according to a second embodiment of the present invention;

fig. 3 is a schematic structural diagram of an apparatus for predicting blood pressure according to a third embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention clearer, the present invention will be described in further detail with reference to the accompanying drawings, and it is apparent that the described embodiments are only a part of the embodiments of the present invention, not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

Before the present invention is explained in further detail by examples, some technical means mentioned in the text will be briefly explained.

The PPG signal is a set of signals that uses a light-sensitive sensor to record the light intensity changes for light intensity identification of a particular 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, so that the PPG signal, which reflects the amount of light absorbed by the blood, also shows a periodic change tendency. A cardiac cycle comprises two time periods: systolic and diastolic phases; when the heart contracts, the heart does work on the whole body, so that the pressure in the blood vessel and the volume of blood flow generate continuous periodic change, and the blood in the blood vessel absorbs the light most; when the heart is in diastole, the pressure to the blood vessel is relatively small, and the blood pushed out to the whole body by the last systole strikes a heart valve through circulation so as to generate certain reflection and refraction effects on light, so that the absorption of the blood in the blood vessel to the light energy is reduced during the diastole period. Therefore, the waveform of the PPG signal, which reflects the absorption energy of blood in the blood vessel, is formed by the superposition of two signals: a systolic phase signal and a diastolic phase signal; a common PPG signal has two peaks of magnitude, the first belonging to the systolic phase and the second to the diastolic phase.

In the initial collected PPG signal, there are many noise and interference sources, so a certain filtering noise reduction conversion needs to be performed on the PPG signal after collection, and the converted signal is regarded as a pulse wave signal which can normally reflect the pulse fluctuation cycle characteristics of a tester. By further carrying out feature identification and regression classification processing on the pulse wave signals, the predicted values of the systolic pressure and the diastolic pressure of the blood pressure can be further obtained. Specifically, the method comprises the following steps: firstly, effective data extraction, namely, feature extraction or feature calculation is carried out on pulse wave data; and then obtaining regression data as a prediction result by using a blood pressure regression calculation mode on the obtained characteristic data. Because the systolic feature and the diastolic feature are included in the single pulse wave signal, the regression data in the blood pressure regression calculation correspond to two: diastolic pressure data and systolic pressure data, wherein the systolic pressure data is greater than the diastolic pressure data.

With respect to feature computation, we know that CNN has long been one of the core algorithms in the field of feature recognition. The method is applied to image recognition, and can be used for extracting the discriminant features of the image in fine classification recognition for other classifiers to learn. The method is applied to the field of blood pressure feature identification, and comprises the following steps of performing pulse wave feature extraction calculation on input one-dimensional pulse wave data: after the convolution and pooling of the input raw pulse wave data, feature data that conform to the pulse wave characteristics are retained for learning by other networks. The blood pressure CNN model mentioned in the text is a CNN model which is trained by blood pressure feature extraction and specifically comprises a convolutional layer and a pooling layer, wherein the convolutional layer is responsible for performing blood pressure feature extraction calculation on input data of the CNN model, and the pooling layer is used for performing downsampling on an extraction result of the convolutional layer; the blood pressure CNN model is divided into a plurality of CNN network layers, and each CNN network layer comprises a convolutional layer and a pooling layer. The formats of input data and output data of the blood pressure CNN model are both 4-dimensional tensor forms: [ B, H, W, C ]. Every time the data passes through a convolutional layer or a pooling layer, the values of certain dimension parameters of the output data can change, namely the total data length of the tensor can be shortened, and the change is characterized in that: b is taken as a fourth dimension parameter (the total number of segments of the pulse wave one-dimensional data sequence) in the four dimensions and does not change; H. w is a third and a two-dimensional parameter in four dimensions, and the change of the first and the second parameters is related to the setting of the convolution kernel size and the sliding step length of each convolution layer and also related to the pooling window size and the sliding step length of each pooling layer; c is the first dimension parameter in the four dimensions, and its variation is related to the selected output spatial dimension (number of convolution kernels) in the convolutional layer.

Further, the embodiment of the present invention provides two convolution schemes for CNN feature calculation, where the two convolution schemes are respectively: sending pulse wave data as a whole into a blood pressure CNN model for integral convolution and pooling calculation and outputting a feature tensor; and a second scheme is that the pulse wave is divided according to segments, each segment is divided into a plurality of sub-segments, all the sub-segments are sent to a blood pressure CNN model according to the sequence of the segments to be respectively subjected to convolution pooling calculation, the total number of the segments and the total number of the sub-segments are output, and all the output feature tensors are combined to generate the feature tensor of the integral pulse wave. According to practical engineering experience, the two schemes are provided, and the fact that when the data volume is insufficient is found that the feature extraction effect can be improved by adjusting the input and output parameter structures of the blood pressure CNN model, and the loss caused by insufficient data volume is made up.

The LSTM network is suitable for processing and predicting important events with very long intervals and delays in a time sequence, information of the previous step on the time sequence can be selectively memorized or forgotten, and the model effect of regression classification can be achieved by training the LSTM network through time sequence feature data. The embodiment of the invention uses the blood pressure LSTM network model which is trained by batch pulse wave signal characteristic data and corresponding batch actually measured blood pressure data to carry out further regression classification processing on the characteristic data output by the blood pressure CNN model according to the time sequence of the data, and finally outputs a predicted blood pressure result. The blood pressure LSTM network model in the embodiment of the invention comprises a plurality of layers of LSTM networks and a full connection layer. Each layer of LSTM network is composed of a plurality of processor units which are connected in sequence and can transmit information according to a specified direction so that the network becomes a memory function, parameters are set in the processor units to control the dimension size of output, and the time sequence transmitted from front to back is called forward direction and from back to front is called reverse direction. The first layer of LSTM network of the LSTM network model used in this embodiment is a bidirectional LSTM network, and the subsequent layers are all unidirectional LSTM networks. And the data format output by the last processor unit of the last layer of the LSTM network is a three-dimensional tensor form, the three-dimensional tensor is input into the full-connection layer to perform blood pressure regression calculation, and finally the predicted required data is obtained, wherein the predicted required data comprises diastolic pressure data and systolic pressure data.

As shown in fig. 1, which is a schematic diagram of a method for predicting blood pressure according to an embodiment of the present invention, the method mainly includes the following steps:

step 1, performing pulse wave conversion and sampling processing on PPG signal data of a photoplethysmography to generate a pulse wave one-dimensional vector; dividing the pulse wave one-dimensional vector into a plurality of pulse wave one-dimensional segments and acquiring the total number of the segments; dividing the pulse wave one-dimensional segment into a plurality of pulse wave one-dimensional sub-segments and acquiring the total number of the sub-segments;

the method specifically comprises the following steps: step 11, calling PPG signal acquisition equipment, carrying out light intensity signal acquisition on a preset light source signal on the local skin surface of the organism, and generating PPG signal data with a length being a signal acquisition time threshold; performing pulse wave data conversion operation on the PPG signal data to generate pulse wave signal data; sampling the characteristic data of the pulse wave signal data according to a characteristic sampling frequency threshold value to generate a pulse wave one-dimensional vector; the preset light source signals at least comprise one type of red light source signals, infrared light source signals and green light source signals;

here, the pulse wave one-dimensional vector is specifically a pulse wave one-dimensional vector [ a ]; a is a first dimension parameter of a pulse wave one-dimensional vector [ A ], and the value of A is the product of a signal acquisition time threshold value and a characteristic sampling frequency threshold value. For example, if the signal acquisition time threshold is 10 seconds and the characteristic sampling frequency threshold is 125Hz, a is 125 × 10 — 1250, which means that there are 1250 acquired data. The pulse wave one-dimensional vector [ A ] becomes a pulse wave one-dimensional vector [1250] which is a one-dimensional vector comprising 1250 independent pulse wave data;

step 12, performing data segment division on the pulse wave one-dimensional vector according to segment length thresholds to generate a plurality of pulse wave one-dimensional segments, and taking the total number of the pulse wave one-dimensional segments included in the pulse wave one-dimensional vector as the total number of the segments;

and step 13, dividing the pulse wave one-dimensional segment into a plurality of pulse wave one-dimensional sub-segments according to the sub-segment length threshold, and taking the total number of the pulse wave one-dimensional sub-segments included in the pulse wave one-dimensional segment as the total number of the sub-segments.

Here, since the blood pressure CNN is subsequently used to perform feature calculation on the data in the pulse wave one-dimensional vector, in view of the requirement for the input of the blood pressure CNN, the pulse wave one-dimensional vector is segmented according to the maximum data length input by the blood pressure CNN, where the segment length threshold is the maximum data length input by the blood pressure CNN.

Step 2, acquiring the identifier value of the convolutional neural network CNN scheme as a first scheme;

the CNN scheme identifier comprises a first scheme identifier and a second scheme identifier.

Here, the CNN scheme identifier is used to distinguish two processing modes of CNN: when the CNN scheme identifier is the first scheme, the pulse wave one-dimensional vector is input into the CNN as a whole for carrying out the omnidirectional volume convolution pooling calculation, and the output result is used as the input of the next LSTM calculation.

Step 3, when the CNN scheme identifier is a first scheme, performing full-vector convolution pooling calculation operation on the pulse wave one-dimensional vector by using a blood pressure CNN model, and performing tensor dimensionality reduction processing on a calculation result according to an input parameter format of a blood pressure long-short term memory network (LSTM) to generate a first LSTM input three-dimensional tensor;

the method specifically comprises the following steps: step 31, when the CNN scheme identifier is the first scheme, performing a first blood pressure CNN input parameter setting operation according to the total number of the segments and the pulse wave one-dimensional vector to generate a first CNN input four-dimensional tensor;

the method specifically comprises the following steps: step 311, when the CNN scheme identifier is the first scheme, setting the first CNN input four-dimensional tensor as the first CNN input four-dimensional tensor [ B ]₁，1，W₁，1]；

Wherein the first CNN inputs a four-dimensional tensor [ B ]₁，1，W₁，1]Comprising B₁The first CNN inputs the three-dimensional tensor [1, W₁，1]；B₁Inputting a four-dimensional tensor [ B ] for the first CNN₁，1，W₁，1]And B is a fourth dimensional parameter of₁Is the total number of fragments; w₁Inputting a four-dimensional tensor [ B ] for the first CNN₁，1，W₁，1]A second dimension parameter of, and W₁Is a segment length threshold;

here, the obtained pulse wave one-dimensional segment is converted into 4-dimensional data, and four dimensions [ B, H, W, C ] represent the number of segments (batch), the height of the segment data (height), the width of the segment data (width), and the number of channels of the segment data (channel), respectively. In processing color image data, the height, width and number of channels of a segment correspond to the height, width and RGB channels of the image, respectively. Since the pulse wave data is a one-dimensional vector, the number B of segments should be set as the total number of segments, the height H should be set as 1, the number C of channels should be set as 1, and the width W should be a segment length threshold previously set, for example: assuming that the segment length threshold is 250, the four-dimensional tensor of the first CNN input obtained by converting the pulse wave one-dimensional vector [1250] is [5, 1, 250, 1 ];

step 312, sequentially extracting the pulse wave one-dimensional segments included in the pulse wave one-dimensional vector, and inputting the four-dimensional tensor [ B ] to the first CNN₁，1，W₁，1]The corresponding first CNN inputs the three-dimensional tensor [1, W₁，1]Carrying out matrix element assignment processing;

here, assume a pulse wave one-dimensional vector [1250]＝{D₁，……D₁₂₅₀And if the total number of the segments is 5, dividing the pulse wave one-dimensional vector into 5 segments: { D₁，……D₂₅₀}、{D₂₅₁，……D₅₀₀}、{D₅₀₁，……D₇₅₀}、{D₇₅₁，……D₁₀₀₀}、{D₁₀₀₁，……D₁₂₅₀}; assume that the first CNN input four-dimensional tensor is [5, 1, 250, 1]]That is, the first CNN input four-dimensional tensor includes 5 first CNN input three-dimensional tensors [1, 250, 1]]_i(i has a value of 1 to 5); then it is determined that,

the first CNN inputs the three-dimensional tensor [1, 250, 1]]₁＝{D₁，……D₂₅₀}、

The first CNN inputs the three-dimensional tensor [1, 250, 1]]₂＝{D₂₅₁，……D₅₀₀}、

The first CNN inputs the three-dimensional tensor [1, 250, 1]]₃＝{D₅₀₁，……D₇₅₀}、

The first CNN inputs the three-dimensional tensor [1, 250, 1]]₄＝{D₇₅₁，……D₁₀₀₀}、

The first CNN inputs the three-dimensional tensor [1, 250, 1]]₅＝{D₁₀₀₁，……D₁₂₅₀}；

After the data addition is completed, respectively, the first CNN input four-dimensional tensor is [5, 1, 250, 1]]Expressed in data sequence is: [ { D)₁，……D₂₅₀}，{D₂₅₁，……D₅₀₀}，{D₅₀₁，……D₇₅₀}，{D₇₅₁，……D₁₀₀₀}，

{D₁₀₀₁，……D₁₂₅₀}]；

Step 32, performing multilayer convolution pooling calculation on the first CNN input four-dimensional tensor by using the blood pressure CNN model according to a preset convolution layer number threshold to generate a first CNN output four-dimensional tensor;

the method specifically comprises the following steps: step 321, initializing the value of the first index to be 1; initializing a first total number as a convolution layer number threshold; initializing a first index temporary four-dimensional tensor as a first CNN input four-dimensional tensor;

step 322, performing convolution calculation processing on the first index temporary four-dimensional tensor by using a first index layer convolution layer of the blood pressure CNN model to generate a first index convolution output data four-dimensional tensor; performing pooling calculation processing on the first index convolution output data four-dimensional tensor by using a first index layer pooling layer of the blood pressure CNN model to generate a first index pooling output data four-dimensional tensor;

the blood pressure CNN model comprises a plurality of convolution layers and a plurality of pooling layers;

step 323, setting the first index temporary four-dimensional tensor as a first index pooling output data four-dimensional tensor;

step 324, add 1 to the first index;

step 325, determine whether the first index is greater than the first total number, go to step 326 if the first index is greater than the first total number, go to step 322 if the first index is less than or equal to the first total number;

step 326, setting the first CNN output four-dimensional tensor as a first index temporary four-dimensional tensor;

wherein the first CNN output four-dimensional tensor is specifically the first CNN output four-dimensional tensor [ B ]₂，1，W₂，C₂]；B₂Outputting a four-dimensional tensor [ B ] for the first CNN₂，1，W₂，C₂]And B is a fourth dimensional parameter of₂Is the total number of fragments; w₂Outputting a four-dimensional tensor [ B ] for the first CNN₂，1，W₂，C₂]A second dimension parameter of, and W₂Is a preset LSTM layer neuron total number threshold value; c₂Outputting a four-dimensional tensor [ B ] for the first CNN₂，1，W₂，C₂]A first dimension parameter of, and C₂Is a preset LSTM layer neuron length threshold;

here, step 32 is to perform a multilayer convolution pooling calculation on the four-dimensional tensor of the first input data for the blood pressure CNN to generate an explanation of the four-dimensional tensor of the first CNN output; here, the blood pressure CNN model is composed of convolutional layers and pooling layers, and has a general structure that one layer of convolution is collocated with one layer of pooling and then connected with the next convolutional layer, and the final layer number of the network is determined by the number of convolutional layers, i.e. a convolutional layer number threshold; assuming that the threshold of the number of convolution layers is 4, 4 times of convolution and pooling calculation need to be completed here; the output of each layer of computation will be the input of the next layer of computation.

Here, the first CNN output four-dimensional tensor is specifically the first CNN output four-dimensional tensor [ B ]₂，1，W₂，C₂]；B₂Output four for the first CNNDimension tensor [ B₂，1，W₂，C₂]And B is a fourth dimensional parameter of₂The value of (d) is the total number of fragments; w₂Outputting a four-dimensional tensor [ B ] for the first CNN₂，1，W₂，C₂]A second dimension parameter of, and W₂Is a preset LSTM layer neuron total number threshold value; c₂Outputting a four-dimensional tensor [ B ] for the first CNN₂，1，W₂，C₂]A first dimension parameter of, and C₂Is a preset LSTM layer neuron length threshold; assuming a total number of fragments of 5, a threshold of 5 for the total number of LSTM layer neurons, and a threshold of 64 for the length of LSTM layer neurons, the first CNN outputs a four-dimensional tensor of [5, 1, 5, 64]。

Step 33, carrying out tensor dimensionality reduction on the four-dimensional tensor output by the first CNN according to the input parameter format of the blood pressure long-term and short-term memory network LSTM to generate a first LSTM input three-dimensional tensor;

wherein the first LSTM input three-dimensional tensor is specifically the first LSTM input three-dimensional tensor [ H₃，W₃，C₃]；H₃Inputting a three-dimensional tensor [ H ] for a first LSTM₃，W₃，C₃]A third dimension parameter of, and H₃Is the total number of fragments; w₃Inputting a three-dimensional tensor [ H ] for a first LSTM₃，W₃，C₃]A second dimension parameter of, and W₃Is W₂；C₃Inputting a three-dimensional tensor [ H ] for a first LSTM₃，W₃，C₃]A first dimension parameter of, and C₃Is C₂。

Assuming that the first CNN outputs a four-dimensional tensor of [5, 1, 5, 64], then the LSTM input three-dimensional tensor is [5, 5, 64 ]. Here, because the four-dimensional tensor output by the first CNN keeps the four-dimensional parameters unchanged, and the three-dimensional parameter is removed for dimensionality reduction, the three-dimensional parameter after actual dimensionality reduction should be equal to the four-dimensional parameter before dimensionality reduction, the two-dimensional parameter after dimensionality reduction should be equal to the product of the three-dimensional parameter before dimensionality reduction multiplied by the two-dimensional parameter, and because the third-dimensional parameter of the four-dimensional tensor output by the first CNN is actually 1, the two-dimensional parameter after dimensionality reduction is not changed.

Step 4, performing blood pressure long-term and short-term memory calculation operation on the first LSTM input three-dimensional tensor by using an LSTM network layer of the blood pressure LSTM model to generate an LSTM output three-dimensional tensor;

the blood pressure LSTM network model comprises an LSTM network layer and a full connection layer;

the method specifically comprises the following steps: step 41, when the CNN scheme identifier is the first scheme, performing blood pressure long-term and short-term memory calculation operation on the first LSTM input three-dimensional tensor by using the LSTM network layer of the blood pressure LSTM model to generate an LSTM output three-dimensional tensor;

and 42, when the CNN scheme identifier is the second scheme, performing blood pressure long-term and short-term memory calculation operation on the second LSTM input three-dimensional tensor by using the LSTM network layer of the blood pressure LSTM model to generate an LSTM output three-dimensional tensor.

Step 5, performing blood pressure regression calculation operation on the LSTM output three-dimensional tensor by using the full connection layer to generate a blood pressure prediction three-dimensional tensor [ X, Y, 2 ];

wherein X is the total number of fragments; y is the total number of sub-segments.

Here, assuming that the LSTM input three-dimensional tensor is [5, 5, 64], the three-dimensional tensor for blood pressure prediction should be [5, 5, 2], that is, 5 × 5 ═ 25 one-dimensional tensors [2], and each one-dimensional tensor [2] includes two pieces of blood pressure prediction data: a diastolic pressure reference value and a systolic pressure reference value; the three-dimensional tensor of blood pressure prediction should be [5, 5, 2], that is, 25 pairs of blood pressure prediction values are included.

Step 6, sequentially extracting predicted blood pressure data from the three-dimensional tensor [ X, Y, 2] of blood pressure prediction according to the sequence of the pulse wave one-dimensional segments and the sequence of the pulse wave one-dimensional sub-segments, and generating a blood pressure prediction data set;

the method specifically comprises the following steps: step 61, initializing a blood pressure prediction data set to be empty; setting a blood pressure data set; initializing the diastolic blood pressure data of the blood pressure data set to null; initializing the systolic blood pressure data of the blood pressure data group to be null;

step 62, sequentially extracting a blood pressure prediction one-dimensional vector [2] included in the blood pressure prediction three-dimensional tensor [ X, Y, 2] to generate a current one-dimensional vector [2 ]; setting systolic pressure data of a blood pressure data group as sub-segment systolic pressure data in the current one-dimensional vector [2], and setting diastolic pressure data of the blood pressure data group as sub-segment diastolic pressure data in the current one-dimensional vector [2 ]; performing data set addition operation on the blood pressure data set to the blood pressure prediction data set; the three-dimensional tensor [ X, Y, 2] of the blood pressure prediction comprises X X Y one-dimensional vectors [2] of the blood pressure prediction; the blood pressure prediction one-dimensional vector [2] includes sub-segment systolic pressure data and sub-segment diastolic pressure data.

In step 6, the blood pressure prediction three-dimensional tensors of 25 pairs of blood pressure prediction values in [5, 5, 2] are extracted one by one, and the extracted blood pressure prediction values are added to the blood pressure prediction data set in the form of 25 blood pressure data sets. The blood pressure prediction data set is the calculation result of blood pressure prediction of the PPG signal data by using a CNN and LSTM fusion network.

As shown in fig. 2, which is a schematic diagram of a method for predicting blood pressure according to a second embodiment of the present invention, the method mainly includes the following steps:

step 101, performing pulse wave conversion and sampling processing on PPG signal data of a photoplethysmography to generate a pulse wave one-dimensional vector; dividing the pulse wave one-dimensional vector into a plurality of pulse wave one-dimensional segments and acquiring the total number of the segments; dividing the pulse wave one-dimensional segment into a plurality of pulse wave one-dimensional sub-segments and acquiring the total number of the sub-segments;

the method specifically comprises the following steps: step 1011, calling PPG signal acquisition equipment, and acquiring light intensity signals of preset light source signals on the local skin surface of the organism to generate PPG signal data with a length of a signal acquisition time threshold; performing pulse wave data conversion operation on the PPG signal data to generate pulse wave signal data; sampling the characteristic data of the pulse wave signal data according to a characteristic sampling frequency threshold value to generate a pulse wave one-dimensional vector; the preset light source signals at least comprise one type of red light source signals, infrared light source signals and green light source signals;

step 1012, performing data segment division on the pulse wave one-dimensional vector according to segment length thresholds to generate a plurality of pulse wave one-dimensional segments, and taking the total number of the pulse wave one-dimensional segments included in the pulse wave one-dimensional vector as the total number of the segments;

and 1013, dividing the pulse wave one-dimensional segment into a plurality of pulse wave one-dimensional sub-segments according to the sub-segment length threshold, and taking the total number of the pulse wave one-dimensional sub-segments included in the pulse wave one-dimensional segment as the total number of the sub-segments.

Step 102, acquiring a value of a Convolutional Neural Network (CNN) scheme identifier as a second scheme;

Here, the CNN scheme identifier is used to distinguish two processing modes of CNN: when the CNN scheme identifier is a second scheme, the pulse wave one-dimensional vector is divided into a plurality of segments and sub-segments, each sub-segment is subjected to segmented vector convolution pooling calculation in a blood pressure CNN model, and an output result is used as input of next LSTM calculation.

103, when the CNN scheme identifier is a second scheme, performing segmented vector convolution pooling calculation operation on the pulse wave one-dimensional vector by using a blood pressure CNN model, and performing tensor dimension reduction processing on the calculation result according to the input parameter format of the blood pressure LSTM to generate a second LSTM input three-dimensional tensor;

the method specifically comprises the following steps: step 1031, when the CNN scheme identifier is the second scheme, generating a total number of tensors according to a product of the total number of fragments multiplied by the total number of sub-fragments;

step 1032, according to the total number of tensors and the pulse wave one-dimensional vector, performing a second blood pressure CNN input parameter setting operation to generate a second CNN input four-dimensional tensor group;

the second CNN input four-dimensional tensor group comprises a total number of second CNN input four-dimensional tensors;

the method specifically comprises the following steps: step 10321, sequencing all pulse wave one-dimensional sub-segments of the pulse wave one-dimensional vector to generate a sub-segment complete sequence;

wherein the full sequence of sub-segments comprises a tensor total number of pulse wave one-dimensional sub-segments;

step 10322, setting a second CNN input four-dimensional tensor group; setting the second CNN input four-dimensional tensor to be the [1, 1, W ] four-dimensional tensor of the second CNN input₄，1]；

Wherein the second CNN input four-dimensional tensor group comprises a total number of second CNN input four-dimensional tensors [1, 1, W ]₄，1]；W₄Inputting a four-dimensional tensor [1, 1, W ] for the second CNN₄，1]A second dimension parameter of, and W₄A sub-segment length threshold;

here, the four-dimensional tensor group structure input by the second CNN is set; assuming that the pulse wave one-dimensional vector [ a ] is specifically the pulse wave one-dimensional vector [1250] divided into 5 segments, and each segment is divided into 5 sub-segments, then performing four-dimensional tensor conversion on the 5 × 5-25 sub-segment one-dimensional data to generate a second CNN input four-dimensional tensor group; here, the second CNN input four-dimensional tensor is a four-dimensional tensor for each sub-segment, and the second CNN input four-dimensional tensor group includes 25 second CNN input four-dimensional tensors in total;

step 10323, sequentially extracting pulse wave one-dimensional sub-segments in the sub-segment complete sequence, inputting a four-dimensional tensor [1, 1, W ] to the second CNN corresponding to the four-dimensional tensor group₄，1]Carrying out matrix element assignment processing;

here, the specific values of the four-dimensional tensor group input by the second CNN are assigned; suppose a pulse wave one-dimensional vector [ A ]]Specifically, it becomes a pulse wave one-dimensional vector [1250]Divided into 5 segments, each divided into 5 sub-segments, then the second CNN input four-dimensional tensor group includes 25 total four-dimensional tensorsThe second CNN inputs a four-dimensional tensor; assuming a sub-segment length threshold of 50, the second CNN input four-dimensional tensor is specifically [1, 1, 50, 1]](ii) a Suppose a pulse wave one-dimensional vector [1250]＝{D₁，……D₁₂₅₀Then the second CNN inputs the set of four-dimensional tensors as: {

The second CNN inputs the four-dimensional tensor [1, 1, 50, 1]]₁＝{D₁，……D₅₀}，

……

The second CNN inputs the four-dimensional tensor [1, 1, 50, 1]]₂₅＝{D₁₂₂₆，……D₁₂₅₀}}；

1033, according to a preset convolution layer number threshold, respectively performing multilayer convolution pooling calculation on all second CNN input four-dimensional tensors in the second CNN input four-dimensional tensor group by using the blood pressure CNN model to generate a second CNN output four-dimensional tensor group;

the second CNN output four-dimensional tensor group comprises a total number of second CNN output four-dimensional tensors; the second CNN output four-dimensional tensor group specifically comprises a total number of second CNN output four-dimensional tensors; the second CNN output four-dimensional tensor is specifically the second CNN output four-dimensional tensor [1, 1, 1, C₅]；C₅Outputting the four-dimensional tensor [1, 1, 1, C for the second CNN₅]A first dimension parameter of;

here, the convolution calculation is performed on 25 second CNN input four-dimensional tensors in the second CNN input four-dimensional tensor group to generate a second CNN output four-dimensional tensor group, which corresponds to the input and also includes 25 second CNN output four-dimensional tensors [1, 1, 1, C ]₅](ii) a Assuming that the length of the LSTM neural unit is 64, the second CNN outputs a four-dimensional tensor of [1, 1, 1, 64%]；

Step 1034, performing four-dimensional tensor merging operation on all the second CNN output four-dimensional tensors in the second CNN output four-dimensional tensor group to generate a third CNN output four-dimensional tensor;

the method specifically comprises the following steps: step 10341, setting the third CNN output four-dimensional tensor as the third CNN output four-dimensional tensor [ B ]₆，1，1，C₆]；

Wherein, B₆Is the third CNNOutputting a four-dimensional tensor [ B ]₆，1，1，C₆]And B is a fourth dimensional parameter of₆Is the total number of tensors; c₆Outputting a four-dimensional tensor [ B ] for the third CNN₆，1，1，C₆]A first dimension parameter of, and C₆Is C₅；

Step 10342, sequentially extracting the four-dimensional tensor [1, 1, 1, C ] output by the second CNN from the four-dimensional tensor group output by the second CNN₅]Outputs a four-dimensional tensor [ B ] to the third CNN₆，1，1，C₆]Carrying out matrix element assignment processing;

here, because LSTM input is subsequently performed, a plurality of second CNN output four-dimensional tensors of the second CNN output four-dimensional tensor group need to be combined to generate one four-dimensional tensor; assuming that the second CNN output four-dimensional tensor group includes 25 second CNN output four-dimensional tensors [1, 1, 1, 64], the combined third CNN output four-dimensional tensor is [25, 1, 1, 64 ]; the specific data sequence of each tensor in the corresponding second CNN output four-dimensional tensor group also needs to be sequentially added to the third CNN output four-dimensional tensor;

1035, performing tensor dimensionality reduction on the fourth-dimensional tensor output by the third CNN according to the input parameter format of the blood pressure LSTM to generate a second LSTM input three-dimensional tensor;

wherein the second LSTM input three-dimensional tensor is specifically LSTM input three-dimensional tensor [ H₇，W₇，C₇]；H₇Inputting a three-dimensional tensor [ H ] for the second LSTM₇，W₇，C₇]A third dimension parameter of, and H₃The value of (d) is the total number of fragments; w₇Inputting a three-dimensional tensor [ H ] for the second LSTM₇，W₇，C₇]A second dimension parameter of, and W₇Is B₆Is divided by H₃Quotient of (d); c₇Inputting a three-dimensional tensor [ H ] for the second LSTM₇，W₇，C₇]A first dimension parameter of, and C₇Is C₆。

The input parameter structure of the LSTM network layer is a three-dimensional tensor form, so that in this step, tensor dimension reduction processing needs to be performed on the third CNN output four-dimensional tensor generated by merging, the principle of this dimension reduction is to set the third dimension parameter of the three-dimensional tensor after dimension reduction as the total number of segments, the product of the third and second dimension parameters of the tensor after dimension reduction is equal to the fourth dimension parameter before dimension reduction, and the value of the first dimension parameter after dimension reduction is unchanged; assuming that the fourth CNN output four-dimensional tensor is [25, 1, 1, 64], then the reduced-dimension second LSTM input three-dimensional tensor is [5, 5, 64 ]; this shows that the data value sequence of the actual tensor is not increased or decreased, but only the tensor shape is transformed.

104, performing blood pressure long-term and short-term memory calculation operation on a second LSTM input three-dimensional tensor by using an LSTM network layer of the blood pressure LSTM model to generate an LSTM output three-dimensional tensor;

the method specifically comprises the following steps: step 1041, when the CNN scheme identifier is the first scheme, performing a blood pressure long and short term memory calculation operation on the first LSTM input three-dimensional tensor by using the LSTM network layer of the blood pressure LSTM model, and generating an LSTM output three-dimensional tensor;

and 1042, when the CNN scheme identifier is the second scheme, performing blood pressure long-term and short-term memory calculation operation on the second LSTM input three-dimensional tensor by using the LSTM network layer of the blood pressure LSTM model to generate an LSTM output three-dimensional tensor.

105, performing blood pressure regression calculation operation on the LSTM output three-dimensional tensor by using the full connection layer to generate a blood pressure prediction three-dimensional tensor [ X, Y, 2 ];

Here, assuming that the second LSTM input three-dimensional tensor is [5, 5, 64], the blood pressure prediction three-dimensional tensor should be [5, 5, 2], that is, 5 × 5 ═ 25 one-dimensional tensors [2], and each one-dimensional tensor [2] includes two blood pressure prediction data: a diastolic pressure reference value and a systolic pressure reference value; the three-dimensional tensor of blood pressure prediction should be [5, 5, 2], that is, 25 pairs of blood pressure prediction values are included.

Step 106, sequentially extracting predicted blood pressure data from the three-dimensional tensor [ X, Y, 2] of blood pressure prediction according to the sequence of the pulse wave one-dimensional segments and the sequence of the pulse wave one-dimensional sub-segments, and generating a blood pressure prediction data set;

the method specifically comprises the following steps: step 1061, initializing a blood pressure prediction data set to be empty; setting a blood pressure data set; initializing the diastolic blood pressure data of the blood pressure data set to null; initializing the systolic blood pressure data of the blood pressure data group to be null;

step 1062, sequentially extracting a blood pressure prediction one-dimensional vector [2] included in the blood pressure prediction three-dimensional tensor [ X, Y, 2] to generate a current one-dimensional vector [2 ]; setting systolic pressure data of a blood pressure data group as sub-segment systolic pressure data in the current one-dimensional vector [2], and setting diastolic pressure data of the blood pressure data group as sub-segment diastolic pressure data in the current one-dimensional vector [2 ]; performing data set addition operation on the blood pressure data set to the blood pressure prediction data set; the three-dimensional tensor [ X, Y, 2] of the blood pressure prediction comprises X X Y one-dimensional vectors [2] of the blood pressure prediction; the blood pressure prediction one-dimensional vector [2] includes sub-segment systolic pressure data and sub-segment diastolic pressure data.

Here, the blood pressure predicted values of 25 pairs of the three-dimensional tensors [5, 5, 2] for blood pressure prediction are extracted one by one, and are added to the blood pressure prediction data set in the form of 25 blood pressure data sets. The blood pressure prediction data set is the calculation result of blood pressure prediction of the PPG signal data by using a CNN and LSTM fusion network.

As shown in fig. 3, which is a schematic structural diagram of an apparatus for predicting blood pressure according to a third embodiment of the present invention, the apparatus includes: a processor and a memory. The memory may be connected to the processor by a bus. The memory may be a non-volatile memory such as a hard disk drive and a flash memory, in which a software program and a device driver are stored. The software program is capable of performing various functions of the above-described methods provided by embodiments of the present invention; the device drivers may be network and interface drivers. The processor is used for executing a software program, and the software program can realize the method provided by the embodiment of the invention when being executed.

It should be noted that the embodiment of the present invention also provides a computer-readable storage medium. The computer readable storage medium has a computer program stored thereon, and the computer program can realize the method provided by the embodiment of the invention when being executed by a processor.

The embodiment of the invention also provides a computer program product containing the instruction. The computer program product causes a processor to perform the above-mentioned method when run on a computer.

According to the method and the device for predicting the blood pressure, provided by the embodiment of the invention, firstly, collected PPG signal data is converted into pulse wave signal data, then, a fusion network consisting of blood pressure CNN and LSTM is adopted to carry out feature extraction operation on the pulse wave signal data to generate blood pressure feature data, and regression calculation is carried out on the blood pressure feature data so as to predict the blood pressure data (diastolic pressure and systolic pressure) of a tester. The embodiment of the invention not only avoids the complexity and the uncomfortable feeling of the conventional testing means, but also generates an automatic intelligent data analysis method, thereby leading an application party to conveniently and continuously monitor the tested object for many times.

Those of skill would further appreciate that the various illustrative components and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware, computer software, or combinations of both, and that the various illustrative components and steps have been described above generally in terms of their functionality in order to clearly illustrate this interchangeability of hardware and software. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present invention.

The steps of a method or algorithm described in connection with the embodiments disclosed herein may be embodied in hardware, a software module executed by a processor, or a combination of the two. A software module may reside in Random Access Memory (RAM), memory, Read Only Memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, a removable disk, a CD-ROM, or any other form of storage medium known in the art.

The above-mentioned embodiments are intended to illustrate the objects, technical solutions and advantages of the present invention in further detail, and it should be understood that the above-mentioned embodiments are merely exemplary embodiments of the present invention, and are not intended to limit the scope of the present invention, and any modifications, equivalent substitutions, improvements and the like made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims

1. A method of predicting blood pressure, the method comprising:

2. A method of predicting blood pressure as recited in claim 1 in which said photoplethysmography PPG signal data is pulse wave converted and sampled to produce a pulse wave one-dimensional vector; dividing the pulse wave one-dimensional vector into a plurality of pulse wave one-dimensional segments and acquiring the total number of the segments; dividing the pulse wave one-dimensional segment into a plurality of pulse wave one-dimensional sub-segments and acquiring the total number of the sub-segments, specifically comprising:

3. The method of predicting blood pressure according to claim 2, wherein when the CNN scheme identifier is the first scheme, performing a full-vector convolution pooling calculation operation on the pulse wave one-dimensional vector by using a blood pressure CNN model, and performing tensor dimensionality reduction on the calculation result according to an input parameter format of a blood pressure long-short term memory network LSTM to generate a first LSTM input three-dimensional tensor specifically comprises:

4. The method of claim 3, wherein when the CNN scheme identifier is the first scheme, performing a first blood pressure CNN input parameter setting operation according to the total number of segments and the pulse wave one-dimensional vector to generate a first CNN input four-dimensional tensor comprises:

sequentially extracting the pulse wave one-dimensional segments included by the pulse wave one-dimensional vectors, namely the first one-dimensional segmentOne CNN inputs four-dimensional tensor [ B ]₁，1，W₁，1]The corresponding first CNN inputs a three-dimensional tensor [1, W₁，1]And carrying out matrix element assignment processing.

5. The method according to claim 3, wherein the generating a first CNN output four-dimensional tensor by performing a multi-layer convolution pooling calculation on the first CNN input four-dimensional tensor using the blood pressure CNN model according to a preset convolution layer number threshold includes:

step 54, adding 1 to the first index;

6. The method of predicting blood pressure according to claim 5,

the first LSTM input three-dimensional tensor is specifically a first LSTM input three-dimensional tensor [ H ]₃，W₃，C₃](ii) a Said H₃Inputting a three-dimensional tensor [ H ] for the first LSTM₃，W₃，C₃]And the third dimension of (2), and the H₃Is the total number of fragments; the W is₃Inputting a three-dimensional tensor [ H ] for the first LSTM₃，W₃，C₃]And said W is a second dimension parameter of₃Is the said W₂(ii) a Said C is₃Inputting a three-dimensional tensor [ H ] for the first LSTM₃，W₃，C₃]And said C is a first dimension parameter of₃Is the C₂。

7. The method according to claim 2, wherein when the CNN scheme identifier is the second scheme, performing a piecewise vector convolution pooling calculation operation on the pulse wave one-dimensional vector by using the blood pressure CNN model, and performing tensor dimensionality reduction on the calculation result according to the input parameter format of the blood pressure LSTM to generate the second LSTM input three-dimensional tensor specifically comprises:

8. The method of predicting blood pressure according to claim 7, wherein the performing a second blood pressure CNN input parameter setting operation according to the total number of tensors and the pulse wave one-dimensional vector to generate a second CNN input four-dimensional tensor group specifically includes:

setting the second CNN input four-dimensional tensor group; setting the second CNN input four-dimensional tensor to be the second CNN input four-dimensional tensor [1, 1, W₄，1](ii) a The set of second CNN-input four-dimensional tensors includes a total number of the tensors of the second CNN-input four-dimensional tensor [1, 1, W₄，1](ii) a The W is₄Inputting a four-dimensional tensor [1, 1, W ] for the second CNN₄，1]And said W is a second dimension parameter of₄Is the sub-segment length threshold;

9. The method of predicting blood pressure as set forth in claim 7,

the second CNN output four-dimensional tensor group specifically includes a total number of the tensors of the second CNN output four-dimensional tensor; the second CNN output four-dimensional tensor is specifically a second CNN output four-dimensional tensor [1, 1, 1, C₅](ii) a Said C is₅Outputting a four-dimensional tensor [1, 1, 1, C for the second CNN₅]Is measured.

10. The method of predicting blood pressure according to claim 9, wherein the performing a four-dimensional tensor merge operation on all the second CNN output four-dimensional tensors in the second CNN output four-dimensional tensor group to generate a third CNN output four-dimensional tensor specifically includes:

11. The method of predicting blood pressure as set forth in claim 7,

the second LSTM input three-dimensional tensor is specifically LSTM input three-dimensional tensor [ H₇，W₇，C₇](ii) a Said H₇Inputting a three-dimensional tensor [ H ] for the second LSTM₇，W₇，C₇]Third of (2)Dimension parameter, and the H₃The value of (d) is the total number of segments; the W is₇Inputting a three-dimensional tensor [ H ] for the second LSTM₇，W₇，C₇]And said W is a second dimension parameter of₇Is the said B₆Is divided by the H₃Quotient of (d); said C is₇Inputting a three-dimensional tensor [ H ] for the second LSTM₇，W₇，C₇]And said C is a first dimension parameter of₇Is the C₆。

12. The method of predicting blood pressure according to claim 1, wherein the performing, according to the CNN scheme identifier, a blood pressure long-short term memory calculation operation on the first LSTM input three-dimensional tensor or the second LSTM input three-dimensional tensor using an LSTM network layer of the blood pressure LSTM model to generate an LSTM output three-dimensional tensor specifically comprises:

13. The method of predicting blood pressure according to claim 1, wherein the step of sequentially extracting predicted blood pressure data from the blood pressure predicted three-dimensional tensor [ X, Y, 2] in the order of the pulse wave one-dimensional segments and the pulse wave one-dimensional sub-segments to generate a blood pressure predicted data set comprises:

14. An apparatus comprising a memory for storing a program and a processor for performing the method of any one of claims 1 to 13.

15. A computer program product comprising instructions which, when run on a computer, cause the computer to perform the method of any one of claims 1 to 13.

16. A computer-readable storage medium comprising instructions that, when executed on a computer, cause the computer to perform the method of any of claims 1 to 13.