CN117442173A

CN117442173A - Method for training blood pressure prediction model based on meta learning

Info

Publication number: CN117442173A
Application number: CN202311329592.7A
Authority: CN
Inventors: 范非易; 张继灵; 陈益强
Original assignee: Institute of Computing Technology of CAS
Current assignee: Institute of Computing Technology of CAS
Priority date: 2023-10-13
Filing date: 2023-10-13
Publication date: 2024-01-26

Abstract

The invention provides a method for training a blood pressure prediction model based on meta learning, which comprises the following steps: acquiring a training set, and dividing the training set into a first training set, a second training set and a third training set; pre-training the blood pressure prediction model by using a first training set to obtain a pre-trained blood pressure prediction model; initializing an initial element learner by using parameters of the pre-training blood pressure prediction model; training the initial meta learner by utilizing a plurality of training tasks in a second training set based on a meta learning algorithm to obtain a target meta learner; and initializing each patient in the third training set by using the target element learner to obtain an initial personalized blood pressure prediction model corresponding to each patient, and training the corresponding initial personalized blood pressure prediction model by using training data of each patient in the third training set to obtain the personalized blood pressure prediction model corresponding to the patient.

Description

Method for training blood pressure prediction model based on meta learning

Technical Field

The invention relates to the field of deep learning, in particular to the field of blood pressure prediction in the field of deep learning, and more particularly relates to a meta-learning technology aiming at blood pressure prediction and based on a deep neural network, namely a method for training a blood pressure prediction model based on meta-learning.

Background

Blood pressure refers to the force required to push blood into an arterial vessel during systole and the pressure of blood against an arterial vessel during diastole. The blood pressure of normal persons should be within a certain range, and if the blood pressure is too high or too low, the physical health may be adversely affected. When the blood pressure is continuously increased to a certain degree, the damage of the cardiovascular system can be caused, and not only can diseases such as coronary heart disease, cardiac hypertrophy, myocardial infarction and the like be caused; and the risk of cerebral apoplexy, retinopathy, kidney function damage and other diseases is increased, and short-term or long-term damage and death of heart, brain and other organs are more likely to occur. When the blood pressure is continuously reduced to a certain degree, the symptoms such as dizziness, hypodynamia, fainting and the like can also be caused, and especially the situation that the old or the patients suffering from other basic diseases are more likely to suffer from hypotension is caused. Therefore, the regular measurement of blood pressure is very important for maintaining physical health. Two important indicators of blood pressure are systolic pressure and diastolic pressure, systolic pressure refers to the pressure of blood in an artery when the left ventricle contracts, blood flows to the whole body through the artery, and is commonly called as the highest blood pressure, and normal systolic pressure ranges from 90mmHg to 140 mmHg; diastolic pressure refers to the pressure in the arteries at diastole, at which time the heart relaxes and blood flows back to the systemic circulation of the heart, commonly referred to as the lowest blood pressure, with normal diastolic pressures ranging between 60-90 mmHg.

The existing blood pressure measuring method is mainly divided into two types, one is an invasive measuring method and the other is a non-invasive measuring method. The invasive blood pressure measuring method is to directly insert the blood pressure measuring equipment into the human body by inserting a catheter or a needle head and the like to measure the real blood pressure value; the method can measure very accurate blood pressure values, but has high operation difficulty and high infection risk, and is not suitable for routine use. The noninvasive blood pressure measurement method refers to indirect measurement of blood pressure by means of pressure or flow type cuffs and by means of automatic air pressure change of equipment, and although the method has no higher infection risk, the process is complex, and the frequency of blood pressure measurement is limited. However, the method cannot continuously monitor the data of the patient, cannot provide a reliable data set for the blood pressure prediction model to predict the blood pressure, and continuously acquire the data of the patient by using the sensor of the wearable device, so that reliable blood pressure data is provided for the blood pressure prediction model.

The existing blood pressure prediction model is usually obtained by modeling based on blood pressure data of a large number of people and training and converging in a common training mode, but due to individual variability, the health condition and risk of each person cannot be truly reflected when the existing blood pressure prediction model predicts the blood pressure of a target individual. Therefore, the blood pressure model is personalized, prediction and intervention can be better carried out according to individual variability, the accuracy and efficiency of blood pressure monitoring are improved, and more accurate prediction factors and individual characteristics can be found through analysis and modeling of individual data, so that more accurate prediction and intervention are realized.

However, the current method for performing personalized fine tuning on the existing model mainly performs fine tuning on a part of layers of the model by collecting a large number of data samples of a target individual so that the existing model can adapt to the data of the target individual to obtain a good prediction effect, but the method is extremely easy to have the condition of data overfitting, meanwhile, has the problem of inaccurate measurement, and is a very complex process of collecting a large number of data samples, so that the existing blood pressure prediction model has a certain limitation on individual blood pressure prediction.

Disclosure of Invention

It is therefore an object of the present invention to overcome the above-mentioned drawbacks of the prior art and to provide a method for training a blood pressure prediction model based on meta-learning.

According to one aspect of the present invention, there is provided a method of training a blood pressure prediction model based on meta learning, the method comprising: acquiring a training set, and dividing the training set into a first training set, a second training set and a third training set according to a preset proportion; performing a first training phase comprising: pre-training the blood pressure prediction model by using a first training set to obtain a pre-trained blood pressure prediction model; performing a second training phase comprising: acquiring a pre-training blood pressure prediction model, and initializing an initial element learner by using parameters of the pre-training blood pressure prediction model; acquiring a second training set, wherein the second training set comprises a plurality of training tasks, and training the initial meta learner by utilizing the plurality of training tasks in the second training set based on a meta learning algorithm to obtain a target meta learner, wherein data in one training task corresponds to training data of one patient; performing a third training phase comprising: and initializing each patient in the third training set by using the target element learner to obtain an initial personalized blood pressure prediction model corresponding to each patient, and training the corresponding initial personalized blood pressure prediction model by using training data of each patient in the third training set to obtain the personalized blood pressure prediction model corresponding to the patient.

In some embodiments of the invention, the second training phase comprises: acquiring a pre-training blood pressure prediction model, and initializing an initial element learner by using parameters of the pre-training blood pressure prediction model; acquiring a second training set, wherein the second training set comprises a plurality of training tasks, data in one training task corresponds to training data of one patient, the training data of each patient comprises a plurality of pulse wave signal fragments and systolic pressure and diastolic pressure corresponding to the pulse wave signal fragments, and the training data of each patient is divided into a first supporting set and a first inquiring set; based on the initial meta learner, training the initial meta learner by using a plurality of training tasks in the second training set based on a meta learning algorithm, and completing the last training task to obtain a target meta learner, wherein each training comprises the following steps: taking a meta learner obtained by completing the last training task as an initial meta learner of the current training task, assigning parameters of the initial meta learner to a first personalized model of the current training task, and performing repeated iterative training on the first personalized model by utilizing a first support set of the current training task to obtain a second personalized model under the current training task; inputting the first query set of the current training task into the second personalized model to obtain systolic pressure and diastolic pressure, determining a loss value by using the obtained systolic pressure and diastolic pressure and the systolic pressure and diastolic pressure in the current training task, and solving the gradient according to the loss value to reversely update the parameters of an initial meta-learner of the current training task to obtain the meta-learner under the current training task; taking the element learner under the current training task as an initial element learner of the next training task; the initial element learner which is initialized by the parameters of the pre-training blood pressure prediction model is used as the initial element learner of the first training task.

In some embodiments of the invention, the third training phase comprises: acquiring a third training set, wherein the third training set comprises training data of a plurality of patients, the training data of each patient comprises a plurality of pulse wave signal fragments and systolic pressure and diastolic pressure corresponding to the pulse wave signal fragments, and the training data of each patient is divided into a second support set and a second query set; the target element learner is acquired, and each patient in the third training set is initialized by the target element learner respectively to obtain an initial personalized blood pressure prediction model corresponding to each patient; training the corresponding initial personalized blood pressure prediction model by using a second support set of each patient in the third training set to obtain a corresponding personalized blood pressure prediction model; evaluating the performance of the personalized blood pressure predictive model using a second set of queries for each patient in the third training set.

In some embodiments of the invention, the blood pressure prediction model includes a feature extraction module, a gating loop module, and a regression module, wherein: the feature extraction module comprises three extraction layers, wherein a first extraction layer is configured to extract a plurality of low-layer blood pressure features in a pulse wave signal segment; the second extraction layer is configured to extract a plurality of middle layer blood pressure features of the plurality of low layer blood pressure features; the third extraction layer is configured to extract a plurality of deep blood pressure features of the plurality of middle blood pressure features; and a plurality of serial blood pressure characteristics obtained by superposing a plurality of low-layer blood pressure characteristics extracted from the first layer and a plurality of deep-layer blood pressure characteristics extracted from the third layer according to channels are used as input of the gating circulation module; the gating circulation module is used for modeling the plurality of serial blood pressure characteristics according to the time sequence acquired by the pulse wave signal segments and outputting a plurality of blood pressure characteristics with time sequence; the regression module is used for predicting the blood pressure of the plurality of time-ordered blood pressure characteristics and outputting predicted systolic pressure and diastolic pressure.

In some embodiments of the present invention, each extraction layer includes a convolution unit, a dimension adjustment unit, and a mapping unit, wherein: the convolution unit is used for convolving the pulse wave signal segments to extract a plurality of blood pressure characteristics in the pulse wave signal segments; the dimension adjustment unit is used for normalizing the dimensions of the extracted blood pressure features so that the dimensions of the blood pressure features tend to be the same; and the mapping unit is used for carrying out nonlinear mapping on the plurality of blood pressure characteristics subjected to normalization processing so as to enable the plurality of output blood pressure characteristics to have nonlinearity.

In some embodiments of the present invention, the training set includes blood pressure data of a plurality of patients, the blood pressure data including pulse wave signals and blood pressure wave signals, and the training set is processed as follows: deleting signals with the duration of continuously collecting signals being less than a preset duration from the pulse wave signals and the blood pressure wave signals; the pulse wave signals and the blood pressure wave signals which are deleted by the signals are filtered by a filter, so that filtered pulse wave signals and filtered blood pressure wave signals are obtained; dividing the filtered pulse wave signal and blood pressure wave signal into segments with the same length; eliminating damaged pulse wave signal segments in the pulse wave signal segments by adopting an autocorrelation filter, and processing the eliminated pulse wave signal segments by adopting mean variance normalization to obtain pulse wave signal segments in the training set; and removing the segments with the preset number of systolic peak values in the blood pressure wave signal segments by adopting a multiscale peak detection algorithm, and calculating the systolic pressure and the diastolic pressure of the blood pressure wave signal segments to obtain the systolic pressure and the diastolic pressure corresponding to the pulse wave signal segments in the training set.

In some embodiments of the present invention, the filtering the pulse wave signal and the blood pressure wave signal with the filter to obtain a filtered pulse wave signal and a filtered blood pressure wave signal includes: filtering the pulse wave signals and the blood pressure wave signals subjected to signal deletion by adopting a fourth-order Butterworth band-pass filter, and removing noise in the pulse wave signals and the blood pressure wave signals; the pulse wave signal and the blood pressure wave signal with noise removed are filtered by adopting a Hanpel filter, and abnormal values in the pulse wave signal and the blood pressure wave signal are removed.

According to a second aspect of the present invention, there is provided a blood pressure prediction method, the method comprising: acquiring a pulse wave signal of a target patient; inputting the pulse wave signal of the target patient into a blood pressure prediction model trained by the method according to the first aspect of the invention for prediction, and outputting the systolic pressure and the diastolic pressure of the target patient.

Compared with the prior art, the invention has the advantages that:

1) The blood pressure prediction model is trained through the first training stage and the second training stage, so that the blood pressure prediction model obtained through training in the two training stages has good learning ability, the blood pressure prediction model can achieve good individuation prediction ability only by using a small amount of data of a target patient during the subsequent third training stage, the health condition and risk of each person are truly reflected, the prediction effect of the blood pressure prediction model on target individual data is improved, the problem that in the prior art, the good prediction effect can be obtained only by acquiring a large amount of data samples of the target individual to finely adjust part layers of the model so that the existing model can be suitable for the target individual data is solved, the problem that a large amount of data samples need to be acquired is solved, and meanwhile, the limitation of the existing blood pressure prediction model on individual blood pressure prediction is overcome.

2) The blood pressure prediction model is trained based on a meta-learning algorithm in the second training stage, so that the trained blood pressure prediction model can be subjected to fine adjustment to obtain a personalized blood pressure prediction model through a small number of samples of a target patient, parameters of the trained model cannot be over-fitted to the target patient due to a large number of data samples, and the problem that the method for training the model in the prior art is easy to over-fit data is solved.

Drawings

Embodiments of the invention are further described below with reference to the accompanying drawings, in which:

FIG. 1 is a flowchart of a method for training a blood pressure prediction model based on meta learning according to an embodiment of the present invention;

FIG. 2 is a schematic flow chart of processing a training set according to an embodiment of the present invention;

FIG. 3 is a schematic view of a blood pressure prediction model according to an embodiment of the present invention;

FIG. 4 is a flow chart of a second training phase according to an embodiment of the present invention;

FIG. 5 is a schematic diagram of a technical implementation framework for training a blood pressure prediction model based on meta-learning according to an embodiment of the present invention;

FIG. 6 is a schematic flow diagram of a technique implementation for training a blood pressure prediction model based on meta-learning according to an embodiment of the present invention;

FIG. 7 is a diagram showing the effect of predicting the blood pressure of the first person under test according to the embodiment of the present invention;

FIG. 8 is a schematic diagram showing the effect of predicting blood pressure of a second subject according to an embodiment of the present invention;

FIG. 9 is a graph showing the variation of the systolic pressure error value with the number of samples according to an embodiment of the present invention;

FIG. 10 is a graph showing the variation of the diastolic blood pressure error value with the number of samples according to the embodiment of the present invention.

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described in detail by means of specific examples. It should be understood that the specific embodiments described herein are for purposes of illustration only and are not intended to limit the scope of the invention.

As described in the background art, the existing training method of the blood pressure prediction model is extremely easy to have the condition of data over-fitting, and meanwhile, has the problem of inaccurate measurement.

In order to solve the above problems, the present invention proposes a method for training a blood pressure prediction model based on meta learning, as shown in fig. 1, the method includes the following steps: acquiring a training set, and dividing the training set into a first training set, a second training set and a third training set according to a preset proportion; performing a first training phase comprising: pre-training the blood pressure prediction model by using a first training set to obtain a pre-trained blood pressure prediction model; performing a second training phase comprising: acquiring a pre-training blood pressure prediction model, and initializing an initial element learner by using parameters of the pre-training blood pressure prediction model; acquiring a second training set, wherein the second training set comprises a plurality of training tasks, and training the initial meta learner by utilizing the plurality of training tasks in the second training set based on a meta learning algorithm to obtain a target meta learner, wherein data in one training task corresponds to training data of one patient; performing a third training phase comprising: and initializing each patient in the third training set by using the target element learner to obtain an initial personalized blood pressure prediction model corresponding to each patient, and training the corresponding initial personalized blood pressure prediction model by using training data of each patient in the third training set to obtain the personalized blood pressure prediction model corresponding to the patient. According to the method for training the blood pressure prediction model based on meta learning, the blood pressure prediction model can be pre-trained, meta-learning trained and a small amount of target individual data are subjected to fine adjustment through three training stages, so that the blood pressure prediction model obtained through training can improve the prediction effect on the target individual data, the problem that the blood pressure prediction model obtained through training the blood pressure data of a large number of people in the prior art is not ideal in the blood pressure prediction effect on the target individual through a common training mode is solved, the problem that a part of layers of the model are subjected to fine adjustment by acquiring a large number of data samples of the target individual is solved, the problem that the existing model can adapt to the target individual data to obtain good prediction effect is solved, the problem that the method for training the model in the prior art is prone to data overfitting is solved, and the limitation of the existing blood pressure prediction model in blood pressure prediction of individuals is overcome.

Furthermore, before describing embodiments of the present invention in detail, some of the terms used therein are explained as follows:

meta learning refers to learning in which an individual obtains a learning mechanism, i.e., a class of machine learning algorithms that learn how to learn. In traditional machine learning, it is generally necessary to manually design features and select models for each task, and perform optimization adjustment on prediction results through the selected models and parameters thereof, such a process requires experience and expertise, and new tasks need to redesign models and corresponding various parameters, whereas meta learning is to obtain some general learning strategies by learning information and rules on a group of tasks, so that adaptability to the new tasks is higher and generalization is achieved.

Intensive care medical information set III (MIMIC III waveform data set) is a large-scale medical data set containing physiological monitoring signals collected during an Intensive Care Unit (ICU) from 2001 to 2012 in 28,000 adult patients in boston, multi-family hospitals in us, which waveform data set includes various types of physiological monitoring signals such as photoplethysmogram, electrocardiogram, pulse oxygen saturation, respiratory rate, blood pressure, and the like. The waveform data for each patient records historical values of their heart rate, respiration, and various other physiological variables.

The photoelectric volume pulse wave is a non-invasive physiological signal monitoring technology, and photoelectric signal detection is carried out on the surface of skin through a photoelectric sensor to obtain weak signals generated by tiny expansion and contraction of the skin during heartbeat. The technology is mainly based on an optical principle, irradiates a beam of infrared light on the surface of skin, measures the light transmitted through the skin after irradiation by a photoelectric sensor, and is used for obtaining a reflected light signal of the skin, so that the detection of pulse wave forms is realized, the pulse wave is a waveform formed in blood vessels by blood flow liquid generated by the beating of a heart, the waveform comprises wave crests and wave troughs, physiological parameters such as heart rate, blood pressure and the like can be calculated by detecting the pulse wave, and an evaluation index for the functional condition of a cardiovascular system is provided.

The mean variance normalization, also called Z-score normalization, is a data preprocessing method, used for converting an original data set into a normal distribution data set with a mean value of 0 and a standard deviation of 1, and the basic idea is that the original data set is subjected to linear transformation, so that the mean value of the converted data set is 0 and the variance is 1.

For better understanding of the method for training the blood pressure prediction model based on meta learning provided by the invention, the following detailed description is made with reference to specific embodiments, drawings and examples.

1. Training set

Firstly, a training set is obtained, and the training set is divided into a first training set, a second training set and a third training set according to a preset proportion. According to some embodiments of the invention, the training set may be an open source data set, such as: MIMIC I data sets, MIMIC II data sets, MIMIC III waveform data sets, UCI-BP data sets, etc.; or a data set autonomously acquired by some wearable sensors. In the present invention, the training set uses a MIMIC III data set including a photoplethysmogram signal and an arterial blood pressure waveform signal (in the following description, the photoplethysmogram signal is represented by a pulse wave signal and the arterial blood pressure waveform signal is represented by a blood pressure wave signal), data in the training set is preprocessed, and the data-processed training set is proportionally divided into a first training set, a second training set, and a third training set. As shown in fig. 2, the method for preprocessing the data in the training set includes the following steps:

and S1, deleting signals with the duration of continuously collecting signals being less than a preset duration from the pulse wave signals and the blood pressure wave signals.

According to one embodiment of the invention, since the acquisition frequency of the MIMIC III waveform dataset is 125Hz, that is, 125 data points are acquired in one second, it can be determined whether the duration of acquisition of each signal is less than a preset duration, and if so, the signal is deleted. Schematically, the preset duration may be set to 10 minutes, and if the length of the signal meeting the requirement is greater than or equal to 125×10×60, and if the signal length is less than 125×10×60, the continuous acquisition duration of the signal may be determined to be less than the preset duration, and the signal may be deleted. It should be appreciated that the above is only one illustrative example, and that one skilled in the art may adjust the preset time period, for example, to 12 minutes, 14 minutes, etc., to obtain other embodiments.

And S2, adopting a filter to delete the pulse wave signals and the blood pressure wave signals to obtain filtered pulse wave signals and blood pressure wave signals.

According to one embodiment of the present invention, since any signal lower than 0.5Hz can be reduced to baseline drift and all higher than 8Hz is high-frequency noise, the pulse wave signal and the blood pressure wave signal after signal deletion can be filtered by using a fourth-order butterworth band-pass filter with cutoff frequencies of 0.5Hz and 8Hz, so that noise in the pulse wave signal and the blood pressure wave signal is removed. And then, a Hanpel filter is used for filtering the pulse wave signal and the blood pressure wave signal with noise removed, and abnormal values in the pulse wave signal and the blood pressure wave signal are removed. The way to remove the outliers is as follows: according to the size of the sliding window, the median of the data points in each window is calculated, the standard deviation is calculated according to the calculated median, and if the first value in each window exceeds the threshold value, the median under the window is used for replacing the value. Illustratively, the window size is set to 3, the threshold is set to 3 standard deviations; assuming the data set is [5,8,12,10,6,15,7,9,13,20,25,30,17], for the value 10, the window data formed by the data set is [5,8,12,10,6,15,7], wherein the number of bits is 8, the standard deviation is 3.72, and the value 10 does not exceed 3 times of the standard deviation, namely the value is not an abnormal value and does not need replacement; for a value 15, the window data of which is [12,10,6,15,7,9,13], wherein the number of bits is 10, the standard deviation is 3.27, and the value 15 exceeds 3 times of the standard deviation, namely the value is an abnormal value, and the value is replaced by the median 10; the median and standard deviation of the data in the corresponding window for each value are sequentially calculated and replaced in the above manner, and since the first 3 data and the last three data in the dataset cannot be calculated according to the data in the corresponding window, the corrected dataset is finally obtained as [ #, 10,6,10,7,9,13,20, # ]. It should be appreciated that the above is only one illustrative example, and that one skilled in the art may adjust the sliding window size, the threshold, e.g., adjust the window size to 5,7, etc., the threshold to 2, 4, etc., to arrive at other embodiments.

Step S3, the filtered pulse wave signal and the blood pressure wave signal are divided into segments with the same length.

According to one embodiment of the present invention, the pulse wave signal and the blood pressure wave signal are divided into signal segments of the same length by using a preset sliding window. Illustratively, the duration of each signal segment may be set to 5 seconds, and since the frequency of the acquired signal is 125Hz, i.e., the length of each signal segment is 625, a sliding window of 625 size and step size 250 may be used for signal segmentation. It should be appreciated that the above is only one illustrative example, and that one skilled in the art may adjust the duration of each signal segment, the sliding window, e.g., set the duration to 4 seconds, 6 seconds, etc., adjust the sliding window size to 600, adjust the step size to 240, etc., to arrive at other embodiments.

And S4, eliminating damaged pulse wave signal segments in the pulse wave signal segments by adopting an autocorrelation filter, and processing the eliminated pulse wave signal segments by adopting mean variance normalization to obtain the pulse wave signal segments in the training set.

According to an embodiment of the present invention, in order to ensure a normal high periodicity of the photoplethysmogram pulse wave signal segment, an autocorrelation signal of the pulse wave signal segment may be calculated by using an autocorrelation filter, and a maximum autocorrelation threshold is set, and if the calculated autocorrelation signal is greater than the maximum autocorrelation threshold, it is indicated that the pulse wave signal segment is damaged, and the signal segment is eliminated. After the damaged pulse wave signal segments are eliminated, the pulse wave signal segments after elimination are processed by adopting mean variance normalization. Illustratively, the mean and standard deviation of all pulse wave signal segments in the dataset are calculated using a mean calculation formula and a standard deviation calculation formula, and then the following formula is used The following conversion is performed for each pulse wave signal segment:wherein x represents an original pulse wave signal segment, mu represents a mean value, sigma represents a standard deviation, and x' represents a pulse wave signal segment normalized by mean variance. It should be appreciated that the above is only one illustrative example, and that other normalization processing methods may be used by those skilled in the art to perform data processing, such as maximum-minimum normalization, fractional scaling, linear scale normalization, etc., to obtain other embodiments.

And S5, removing the segments with the preset number of systolic peak values in the blood pressure wave signal segments by adopting a multi-scale peak detection algorithm, and calculating the systolic pressure and the diastolic pressure of the blood pressure wave signal segments to obtain the systolic pressure and the diastolic pressure corresponding to the pulse wave signal segments in the training set.

According to one embodiment of the present invention, since one signal segment includes a plurality of peaks and valleys, a multiscale peak detection algorithm may be used to find peaks and valleys of a blood pressure wave signal segment, calculate an average value of the peaks as a systolic pressure of the blood pressure wave signal segment, and calculate an average value of the valleys as a diastolic pressure of the blood pressure wave signal segment. The algorithm analyzes the periodic information of the signal by calculating the distance between the local minimum to the global minimum of the signal to determine the peak position of the signal. While normal human heart rate is typically above 60 times per minute, signal segments less than a predetermined number of systolic peaks are selected for removal because the algorithm may not have a complete cardiac cycle at the beginning or end of the segment, and the first or last systolic peak of the segment cannot be detected. Illustratively, for a blood pressure wave signal segment with a duration of 5 seconds, there are more than 5 detectable systolic peaks corresponding thereto, and in order to make the data fit, the preset number may be set to 4 at this time, i.e. the blood pressure wave signal segment with less than 4 systolic peaks is removed. It should be understood that the foregoing is only an illustrative example, and those skilled in the art may set the preset number to other values according to the acquisition time of the signal segment, for example, the acquisition time is 6 seconds, and the preset number may be set to 5, etc., to obtain other embodiments.

Based on the data processing mode, a preprocessed training set is obtained, the data processed training set is divided into a first training set, a second training set and a third training set according to a proportion (for example, 4:4:2) by taking an individual patient as a unit, and the three training sets are respectively used for a first training stage, a second training stage and a third training stage for training a blood pressure prediction model.

2. Model structure

Second, a blood pressure prediction model is obtained from the training set obtained in the first part, and in some embodiments of the present invention, the blood pressure prediction model may use a fully connected neural network model, a convolutional neural network model, an antagonistic network model, a decision tree, a support vector machine, and the like. The existing conventional blood pressure prediction model generally comprises a feature extraction model for extracting blood pressure features and a prediction module for performing blood pressure prediction according to the blood pressure features. However, in order to better explain the technical scheme of the invention, the invention constructs a blood pressure prediction model, the model structure of which is shown in fig. 3, wherein the blood pressure prediction model comprises a feature extraction module, a gating circulation module and a regression module, and the method comprises the following steps: the feature extraction module comprises three extraction layers, wherein a first extraction layer is configured to extract a plurality of low-layer blood pressure features in a pulse wave signal segment; the second extraction layer is configured to extract a plurality of middle layer blood pressure features of the plurality of low layer blood pressure features; the third extraction layer is configured to extract a plurality of deep blood pressure features of the plurality of middle blood pressure features; and a plurality of serial blood pressure characteristics obtained by superposing a plurality of low-layer blood pressure characteristics extracted from the first layer and a plurality of deep-layer blood pressure characteristics extracted from the third layer according to channels are used as input of the gating circulation module; the gating circulation module is used for modeling the plurality of serial blood pressure characteristics according to the time sequence acquired by the pulse wave signal segments and outputting a plurality of blood pressure characteristics with time sequence; the regression module is used for predicting the blood pressure of the plurality of time-ordered blood pressure characteristics and outputting predicted systolic pressure and diastolic pressure.

According to an embodiment of the present invention, each extraction layer in the feature extraction module includes: the device comprises a convolution unit, a dimension adjustment unit and a mapping unit; wherein: the convolution unit is used for convolving the pulse wave signal segments to extract a plurality of blood pressure characteristics in the pulse wave signal segments; the dimension adjustment unit is used for normalizing the dimensions of the extracted blood pressure features so that the dimensions of the blood pressure features tend to be the same; and the mapping unit is used for carrying out nonlinear mapping on the plurality of blood pressure characteristics subjected to normalization processing so as to enable the plurality of output blood pressure characteristics to have nonlinearity. Wherein the plurality of blood pressure characteristics are non-linearly mapped at the mapping unit using an activation function, the activation function including but not limited to: reLu activation function, sigmoid activation function, softmax activation function, etc.

According to one embodiment of the invention, the regression module comprises two fully connected layers, wherein the first fully connected layer converts a plurality of time-ordered blood pressure characteristics and calculates systolic pressure and diastolic pressure; the second fully-connected layer reduces the dimension of the calculated systolic pressure and diastolic pressure and outputs the final systolic pressure and diastolic pressure.

It should be noted that, the above-mentioned constructed blood pressure prediction model may be used to extract characteristics of the photoplethysmogram signal, and may also be used to extract characteristics of the electrocardiograph signal, etc., and may be used to obtain a model structure applicable to other scenes by adjusting the number of units in each layer of the model according to different application scenes, and of course, may also be used to construct a suitable prediction model according to different numbers of samples in the application scenes, so as to improve blood pressure prediction accuracy.

3. Training process

And acquiring a first training set, a second training set and a third training set in the first part to train the blood pressure prediction model constructed by the second part.

1. First training stage

And pre-training the blood pressure prediction model by using the first training set to obtain a pre-trained blood pressure prediction model.

According to one embodiment of the invention, a first training set is obtained, wherein the first training set comprises a plurality of training samples and corresponding labels, the training samples comprise pulse wave signal fragments, and the labels are systolic pressure and diastolic pressure corresponding to the training samples; inputting a plurality of samples in the first training set into a feature extraction module of the blood pressure prediction model to extract a plurality of low-layer blood pressure features and a plurality of deep-layer blood pressure features of a pulse wave signal segment, superposing the plurality of low-layer blood pressure features and the plurality of high-layer blood pressure features according to channels, inputting the superposition of the multiple low-layer blood pressure features and the superposition of the multiple high-layer blood pressure features into a gating circulation module of the blood pressure prediction model to perform modeling, and outputting a plurality of time-ordered blood pressure features; carrying out blood pressure prediction on the plurality of time-ordered blood pressure characteristics by using a regression module of the blood pressure prediction model, and outputting predicted systolic pressure and diastolic pressure; and determining a loss value according to the predicted systolic pressure, the diastolic pressure and the label, solving a gradient according to the loss value, and updating parameters of the blood pressure prediction model based on the obtained gradient and a preset first learning rate to obtain a pre-training blood pressure prediction model. Illustratively, the first learning rate may be set to 0.001, 0.002.

According to one embodiment of the invention, in the first training phase, the loss value is calculated using a preset loss function, which is shown as follows:

where n represents the number of training samples,indicates the systolic pressure and the diastolic pressure corresponding to the ith pulse wave signal segment of the output, +.>And the label corresponding to the ith pulse wave signal segment is represented.

2. Second training stage

Acquiring a pre-training blood pressure prediction model, and initializing an initial element learner by using parameters of the pre-training blood pressure prediction model; acquiring a second training set, wherein the second training set comprises a plurality of training tasks, and training the initial meta-learner by utilizing the plurality of training tasks in the second training set based on a meta-learning algorithm to obtain a target meta-learner; wherein the data in one training task corresponds to training data of one patient.

According to one embodiment of the invention, a pre-training blood pressure prediction model trained in a first training stage is firstly obtained, and an initial meta-learner is initialized by parameters of the pre-training blood pressure prediction model. Secondly, a second training set is obtained, wherein the second training set comprises a plurality of training tasks, data in one training task corresponds to training data of one patient, the training data of each patient comprises a plurality of pulse wave signal fragments and systolic pressure and diastolic pressure corresponding to the pulse wave signal fragments, and the training data of each patient is divided into a first supporting set and a first inquiring set; based on the initial meta learner, performing iterative training on the initial meta learner by using a plurality of training tasks in the second training set based on a meta learning algorithm, and completing the last training task to obtain a target meta learner, wherein each training comprises: taking a meta learner obtained by completing the last training task as an initial meta learner of the current training task, assigning parameters of the initial meta learner to a first personalized model of the current training task, and performing repeated iterative training on the first personalized model by utilizing a first support set of the current training task to obtain a second personalized model under the current training task; inputting the first query set of the current training task into the second personalized model to obtain systolic pressure and diastolic pressure, determining a loss value by using the obtained systolic pressure and diastolic pressure and the systolic pressure and diastolic pressure in the current training task, and solving the gradient according to the loss value to reversely update the parameters of an initial meta-learner of the current training task to obtain the meta-learner under the current training task; taking the element learner under the current training task as an initial element learner of the next training task; the initial element learner which is initialized by the parameters of the pre-training blood pressure prediction model is used as the initial element learner of the first training task. Illustratively, if a patient has 1000 training data, each comprising a pulse wave signal segment and its corresponding systolic and diastolic pressures, the 1000 training data are divided into a first support set and a query set, assuming x training data as the first support set, then the remaining 1000-x training data are the first query set.

According to one embodiment of the invention, training a first personalized model using a first set of support in a training task comprises: inputting a pulse wave signal segment of a first support set in a current training task into a feature extraction module of the first personalized model to extract a plurality of low-layer blood pressure features and a plurality of deep blood pressure features of the pulse wave signal segment, superposing and inputting the plurality of low-layer blood pressure features and the plurality of high-layer blood pressure features into a gating circulation module of the first personalized model according to channels to perform modeling, and outputting a plurality of time-ordered blood pressure features; carrying out blood pressure prediction on the plurality of time-ordered blood pressure characteristics by using a regression module of the first personalized model, and outputting predicted systolic pressure and diastolic pressure; determining a loss value according to the predicted systolic pressure and diastolic pressure and the label, solving a gradient according to the loss value, and reversely updating parameters of the first personalized model based on the obtained gradient and a preset second learning rate to obtain a second personalized model. Illustratively, the second learning rate may be set to 0.0001, 0.0002. In training, the loss value is still calculated using the loss function as in the first training phase.

According to one embodiment of the invention, training a first personalized model using a first set of queries in a training task comprises: inputting pulse wave signal fragments of a first query set in a current training task into a characteristic extraction module of the second personalized model to extract a plurality of low-layer blood pressure characteristics and a plurality of deep blood pressure characteristics of the pulse wave signal fragments, superposing the plurality of low-layer blood pressure characteristics and the plurality of high-layer blood pressure characteristics according to channels, inputting the superposed low-layer blood pressure characteristics and the plurality of high-layer blood pressure characteristics into a gating circulation module of the second personalized model for modeling, and outputting a plurality of time-sequence blood pressure characteristics; carrying out blood pressure prediction on the plurality of time-ordered blood pressure characteristics by using a regression module of the second personalized model, and outputting predicted systolic pressure and diastolic pressure; and determining a loss value according to the predicted systolic pressure, the diastolic pressure and the label, solving a gradient according to the loss value, and updating parameters of an initial meta-learner of the current training task based on the obtained gradient and a preset third learning rate to obtain the meta-learner of the current training task. Illustratively, the third learning rate is different from the second learning rate, the third learning rate being greater than the second learning rate. For example, the third learning rate may be set to 0.001 or 0.002. In training, the loss value is still calculated using the loss function as in the first training phase. The technical scheme of the embodiment at least can realize the following beneficial technical effects: the second learning rate which is smaller is set to enable the model to learn how to learn, the third learning rate is set to be larger than the second learning rate, so that the model which is learned by the second learning rate can learn knowledge more quickly, and the convergence rate of the model is improved.

According to an example of the present invention, as shown in fig. 4, the second training set includes n training tasks, the data in each training task corresponds to training data of a patient, first, the initial meta learner is initialized by using the parameters of the pre-training blood pressure prediction model to obtain an initial meta learner for a first training task, the parameters of the initial meta learner are assigned to a first personalized model in the first training task, and the first personalized model is iteratively trained for multiple times by using a first support set of the first training task to obtain a second personalized model in the first training task; inputting a first query set of a first training task into the second personalized model to obtain systolic pressure and diastolic pressure, determining a loss value by using the obtained systolic pressure and diastolic pressure and the systolic pressure and diastolic pressure in the first training task, and solving the gradient according to the loss value to reversely update the parameters of an initial meta-learner to obtain the meta-learner under the first training task; and taking the meta learner obtained by completing the first training task as an initial meta learner in the second training task, and performing the operation to obtain the meta learner in the second training task until the nth training task is completed, so as to obtain a final target meta learner.

According to some embodiments of the present invention, according to the difference of data sample distribution in the training set, a proper meta-learning algorithm may be selected to perform meta-learning training on the blood pressure prediction model, so that the model after meta-learning training has a better prediction effect, where the meta-learning algorithm includes, but is not limited to: model-independent Meta-Learning (MAML), meta-gradient descent (Reptile), random gradient descent (Meta Stochastic Gradient Descent, meta-SGD), and the like. The technical scheme of the embodiment at least can realize the following beneficial technical effects: and the model after the pre-training is subjected to meta-learning training again, so that the prediction precision of the model can be improved, and the model can be trained by using a small amount of data of a target patient in the subsequent third training stage, so that the good personalized prediction capability is achieved.

3. Third training stage

And initializing each patient in the third training set by using the target element learner to obtain an initial personalized blood pressure prediction model corresponding to each patient, and training the corresponding initial personalized blood pressure prediction model by using training data of each patient in the third training set to obtain the personalized blood pressure prediction model corresponding to the patient.

According to one embodiment of the invention, a third training set is firstly obtained, wherein the third training set comprises training data of a plurality of patients, the training data of each patient comprises a plurality of pulse wave signal fragments and corresponding systolic pressure and diastolic pressure, and the training data of each patient is divided into a second support set and a second query set; secondly, acquiring a target element learner, and respectively initializing the target element learner obtained by utilizing the second training stage for each patient based on the training data of each patient in the third training set to obtain an initial personalized blood pressure prediction model corresponding to each patient; training the corresponding initial personalized blood pressure prediction model by using a second support set of each patient in the third training set to obtain a corresponding personalized blood pressure prediction model; evaluating the performance of the personalized blood pressure predictive model using a second set of queries for each patient in the third training set.

According to one embodiment of the invention, training its corresponding initial personalized blood pressure prediction model with each patient's second support set includes: inputting the pulse wave signal segments in the second support set of each patient into a feature extraction module of the corresponding initial personalized blood pressure prediction model to extract a plurality of low-layer blood pressure features and a plurality of deep blood pressure features of the pulse wave signal segments, and superposing the low-layer blood pressure features and the high-layer blood pressure features according to channels to input a gating circulation module of the corresponding initial personalized blood pressure prediction model for modeling, and outputting a plurality of time-ordered blood pressure features; carrying out blood pressure prediction on the plurality of time-ordered blood pressure characteristics by using a regression module of the corresponding initial personalized blood pressure prediction model, and outputting predicted systolic pressure and diastolic pressure; and determining a loss value according to the predicted systolic pressure, the diastolic pressure and the label, and solving the gradient according to the loss value to reversely update the corresponding initial personalized blood pressure prediction model to obtain the personalized blood pressure prediction model corresponding to the patient.

According to one embodiment of the invention, the process of evaluating the performance of the personalized blood pressure prediction model using the second query set for each patient comprises: inputting pulse wave signal segments in a second query set of each patient into a feature extraction module of a corresponding personalized blood pressure prediction model to extract a plurality of low-layer blood pressure features and a plurality of deep blood pressure features of the pulse wave signal segments, and superposing the plurality of low-layer blood pressure features and the plurality of high-layer blood pressure features according to channels to input a gating circulation module of the corresponding personalized blood pressure prediction model for modeling, and outputting a plurality of time-ordered blood pressure features; carrying out blood pressure prediction on the plurality of time-ordered blood pressure characteristics by using a regression module of a corresponding personalized blood pressure prediction model, and outputting predicted systolic pressure and diastolic pressure; and performing performance evaluation of the model according to the predicted systolic pressure and the diastolic pressure and the systolic pressure and the diastolic pressure in the second query set.

It should be noted that, compared to the number of iterations of training the first personalized model by the first support set in the second training stage, in this stage, the number of iterations of training the corresponding initial personalized blood pressure prediction model by the second support set may be increased or decreased appropriately according to the performance evaluation result of the model.

4. Application scenario

The invention predicts the blood pressure of the pulse wave signal of the target patient based on the training set, the model structure and the personalized blood pressure model obtained in the training process:

acquiring a pulse wave signal of a target patient; inputting the pulse wave signals of the target patient into a personalized blood pressure prediction model for prediction, and outputting the systolic pressure and the diastolic pressure of the target patient. More specifically, the feature extraction module of the personalized blood pressure prediction model extracts a plurality of low-layer blood pressure features and a plurality of deep blood pressure features of the pulse wave signal, superimposes and inputs the plurality of low-layer blood pressure features and the plurality of high-layer blood pressure features into the gating circulation module of the personalized blood pressure prediction model according to channels to model, outputs a plurality of time-ordered blood pressure features, predicts the plurality of time-ordered blood pressure features by using the regression module of the personalized blood pressure prediction model, and outputs the systolic pressure and the diastolic pressure of a target patient.

In order to better demonstrate the overall technical solution of the present invention, the following description is given with reference to fig. 5 and 6.

Fig. 5 shows a schematic diagram of a technical implementation framework for training a blood pressure prediction model based on meta learning, which is provided by the invention, firstly, an open source data set is obtained, the open source data set is preprocessed, the preprocessed open source data set is divided into a first training set, a second training set and a third training set, and the first training set, the second training set and a testing set are respectively used in a pre-training stage, a meta learning stage and a testing stage. In the pre-training stage, training an initial neural network model (blood pressure prediction model) by using a first training set to obtain a pre-training model; in the meta learning stage, a pre-training model obtained in the pre-training stage is obtained, the pre-training model is used as an initial meta learner, and the second training set is utilized to train the initial meta learner, so that a target meta learner is obtained; in the test stage, the third training set is divided into a support set and a query set, the support set is utilized to finely tune the target element learner to obtain a personalized model, the personalized model is utilized to conduct blood pressure prediction on data in the query set, and the systolic pressure and the diastolic pressure are output.

Fig. 6 shows a schematic diagram of a technical implementation process for training a blood pressure prediction model based on meta learning, which comprises the following steps:

t1, acquiring an open source data set containing photoplethysmogram information and arterial blood pressure waveform signals;

t2, preprocessing the photo-capacitive pulse wave signal and the arterial blood pressure waveform signal, and dividing the preprocessed data set into a first training set, a second training set and a third training set;

t3, constructing a blood pressure prediction model for predicting systolic pressure and diastolic pressure based on the photoelectric volume pulse wave signals, and determining initialization parameters of the model;

a pre-training stage, namely training and updating the blood pressure prediction model by utilizing a first training set data sample;

in the stage of element learning, taking the pre-training model as an initial element learner, and training and updating the initial element learner by utilizing a second training set to obtain a target element learner;

and T6, in the test stage, the target element learner is finely adjusted by utilizing the data sample of the third training set, and the training effect of the target element learner is evaluated.

In order to better show the prediction effect of the personalized blood pressure prediction model obtained based on the training in the above manner, experimental tests are performed by using the personalized blood pressure prediction model, and experimental results are shown in fig. 7-10, wherein:

Fig. 7 shows an effect diagram of blood pressure prediction of a first subject using the personalized blood pressure prediction model obtained by training of the present invention, fig. 8 shows an effect diagram of blood pressure prediction of a second subject using the personalized blood pressure prediction model obtained by training of the present invention, wherein red broken lines in fig. 7 and 8 represent actual values of systolic and diastolic blood pressure of the subject, and blue broken lines represent predicted values of systolic and diastolic blood pressure of the subject. In the experiment, only 10 samples of the tested person are used for personalized training of the meta learner, and the acquisition period of each sample is 5 seconds, so that a good personalized prediction model can be obtained only by 50 seconds of data of the tested person. As can be seen from fig. 7 and 8, the difference between the systolic pressure and the diastolic pressure predicted by the blood pressure prediction model obtained by the invention and the actual value is small, and the change trend of the blood pressure can be accurately reflected. The personalized blood pressure prediction model obtained by the training method provided by the invention is excellent in personalized prediction, and has higher accuracy for blood pressure prediction of different individuals.

Fig. 9 shows the effect of the predicted systolic pressure error value using the conventional pre-training and fine-tuning method and the training method according to the present invention for different target subject sample numbers, and fig. 10 shows the effect of the predicted diastolic pressure error value using the conventional pre-training and fine-tuning method and the training method according to the present invention for different target subject sample numbers. In the experiments, different numbers of training samples (50, 25, 10, 5) of the target testees were used to test the performance of the models obtained by training the two methods, wherein the + curves in fig. 9 and 10 represent the prediction error values of the models obtained by the training method mentioned in the present invention, and the x curves represent the prediction error values of the models obtained by the conventional pre-training plus fine tuning method; the horizontal axis represents the number of training samples of the target subject, and the vertical axis represents the error value. As is apparent from fig. 9 and fig. 10, the error value of the result of the prediction performed by the blood pressure prediction model obtained by the training method of the present invention is smaller than that of the conventional method, which indicates that the model obtained by the method of the present invention has better prediction performance and better blood pressure predictability for the target subject with fewer samples.

In summary, the method for training the blood pressure prediction model based on meta learning provided by the invention has the following advantages:

It should be noted that, although the steps are described above in a specific order, it is not meant to necessarily be performed in the specific order, and in fact, some of the steps may be performed concurrently or even in a changed order, as long as the required functions are achieved.

The present invention may be a system, method, and/or computer program product. The computer program product may include a computer readable storage medium having computer readable program instructions embodied thereon for causing a processor to implement aspects of the present invention.

The computer readable storage medium may be a tangible device that retains and stores instructions for use by an instruction execution device. The computer readable storage medium may include, for example, but is not limited to, an electronic storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination of the foregoing. More specific examples (a non-exhaustive list) of the computer-readable storage medium would include the following: portable computer disks, hard disks, random Access Memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static Random Access Memory (SRAM), portable compact disk read-only memory (CD-ROM), digital Versatile Disks (DVD), memory sticks, floppy disks, mechanical coding devices, punch cards or in-groove structures such as punch cards or grooves having instructions stored thereon, and any suitable combination of the foregoing.

The foregoing description of embodiments of the invention has been presented for purposes of illustration and description, and is not intended to be exhaustive or limited to the embodiments disclosed. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the various embodiments described. The terminology used herein was chosen in order to best explain the principles of the embodiments, the practical application, or the technical improvements in the marketplace, or to enable others of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A method for training a blood pressure prediction model based on meta learning, the method comprising:

acquiring a training set, and dividing the training set into a first training set, a second training set and a third training set according to a preset proportion;

performing a first training phase comprising: pre-training the blood pressure prediction model by using a first training set to obtain a pre-trained blood pressure prediction model;

performing a second training phase comprising: acquiring a pre-training blood pressure prediction model, and initializing an initial element learner by using parameters of the pre-training blood pressure prediction model; acquiring a second training set, wherein the second training set comprises a plurality of training tasks, and training the initial meta learner by utilizing the plurality of training tasks in the second training set based on a meta learning algorithm to obtain a target meta learner, wherein data in one training task corresponds to training data of one patient;

Performing a third training phase comprising: and initializing each patient in the third training set by using the target element learner to obtain an initial personalized blood pressure prediction model corresponding to each patient, and training the corresponding initial personalized blood pressure prediction model by using training data of each patient in the third training set to obtain the personalized blood pressure prediction model corresponding to the patient.

2. The method of claim 1, wherein the second training phase comprises:

acquiring a pre-training blood pressure prediction model, and initializing an initial element learner by using parameters of the pre-training blood pressure prediction model;

acquiring a second training set, wherein the second training set comprises a plurality of training tasks, data in one training task corresponds to training data of one patient, the training data of each patient comprises a plurality of pulse wave signal fragments and systolic pressure and diastolic pressure corresponding to the pulse wave signal fragments, and the training data of each patient is divided into a first supporting set and a first inquiring set;

based on the initial meta learner, training the initial meta learner by using a plurality of training tasks in the second training set based on a meta learning algorithm, and completing the last training task to obtain a target meta learner, wherein each training comprises the following steps:

Taking a meta learner obtained by completing the last training task as an initial meta learner of the current training task, assigning parameters of the initial meta learner to a first personalized model of the current training task, and performing repeated iterative training on the first personalized model by utilizing a first support set of the current training task to obtain a second personalized model under the current training task;

inputting the first query set of the current training task into the second personalized model to obtain systolic pressure and diastolic pressure, determining a loss value by using the obtained systolic pressure and diastolic pressure and the systolic pressure and diastolic pressure in the current training task, and solving the gradient according to the loss value to reversely update the parameters of an initial meta-learner of the current training task to obtain the meta-learner under the current training task; taking the element learner under the current training task as an initial element learner of the next training task;

the initial element learner which is initialized by the parameters of the pre-training blood pressure prediction model is used as the initial element learner of the first training task.

3. The method according to claim 2, wherein the third training phase comprises:

Acquiring a third training set, wherein the third training set comprises training data of a plurality of patients, the training data of each patient comprises a plurality of pulse wave signal fragments and systolic pressure and diastolic pressure corresponding to the pulse wave signal fragments, and the training data of each patient is divided into a second support set and a second query set;

the target element learner is acquired, and each patient in the third training set is initialized by the target element learner respectively to obtain an initial personalized blood pressure prediction model corresponding to each patient;

training the corresponding initial personalized blood pressure prediction model by using a second support set of each patient in the third training set to obtain a corresponding personalized blood pressure prediction model;

evaluating the performance of the personalized blood pressure predictive model using a second set of queries for each patient in the third training set.

4. The method of claim 1, wherein the blood pressure prediction model comprises a feature extraction module, a gating loop module, and a regression module, wherein:

the feature extraction module comprises three extraction layers, wherein a first extraction layer is configured to extract a plurality of low-layer blood pressure features in a pulse wave signal segment; the second extraction layer is configured to extract a plurality of middle layer blood pressure features of the plurality of low layer blood pressure features; the third extraction layer is configured to extract a plurality of deep blood pressure features of the plurality of middle blood pressure features; and a plurality of serial blood pressure characteristics obtained by superposing a plurality of low-layer blood pressure characteristics extracted from the first layer and a plurality of deep-layer blood pressure characteristics extracted from the third layer according to channels are used as input of the gating circulation module;

The gating circulation module is used for modeling the plurality of serial blood pressure characteristics according to the time sequence acquired by the pulse wave signal segments and outputting a plurality of blood pressure characteristics with time sequence;

the regression module is used for predicting the blood pressure of the plurality of time-ordered blood pressure characteristics and outputting predicted systolic pressure and diastolic pressure.

5. The method of claim 4, wherein each extraction layer comprises a convolution unit, a dimension adjustment unit, and a mapping unit, wherein:

the convolution unit is used for convolving the pulse wave signal segments to extract a plurality of blood pressure characteristics in the pulse wave signal segments;

the dimension adjustment unit is used for normalizing the dimensions of the extracted blood pressure features so that the dimensions of the blood pressure features tend to be the same;

and the mapping unit is used for carrying out nonlinear mapping on the plurality of blood pressure characteristics subjected to normalization processing so as to enable the plurality of output blood pressure characteristics to have nonlinearity.

6. The method of claim 1, wherein the training set comprises blood pressure data of a plurality of patients, the blood pressure data comprising pulse wave signals and blood pressure wave signals, the training set being processed as follows:

Deleting signals with the duration of continuously collecting signals being less than a preset duration from the pulse wave signals and the blood pressure wave signals;

the pulse wave signals and the blood pressure wave signals which are deleted by the signals are filtered by a filter, so that filtered pulse wave signals and filtered blood pressure wave signals are obtained;

dividing the filtered pulse wave signal and blood pressure wave signal into segments with the same length;

eliminating damaged pulse wave signal segments in the pulse wave signal segments by adopting an autocorrelation filter, and processing the eliminated pulse wave signal segments by adopting mean variance normalization to obtain pulse wave signal segments in the training set;

and removing the segments with the preset number of systolic peak values in the blood pressure wave signal segments by adopting a multiscale peak detection algorithm, and calculating the systolic pressure and the diastolic pressure of the blood pressure wave signal segments to obtain the systolic pressure and the diastolic pressure corresponding to the pulse wave signal segments in the training set.

7. The method of claim 6, wherein filtering the deleted pulse wave signal and the deleted blood pressure wave signal with a filter to obtain a filtered pulse wave signal and a filtered blood pressure wave signal comprises:

filtering the pulse wave signals and the blood pressure wave signals subjected to signal deletion by adopting a fourth-order Butterworth band-pass filter, and removing noise in the pulse wave signals and the blood pressure wave signals;

The pulse wave signal and the blood pressure wave signal with noise removed are filtered by adopting a Hanpel filter, and abnormal values in the pulse wave signal and the blood pressure wave signal are removed.

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

acquiring a pulse wave signal of a target patient;

inputting the pulse wave signals of the target patient into a blood pressure prediction model trained by the method of any one of claims 1-7 for prediction, and outputting the systolic pressure and the diastolic pressure of the target patient.

9. A computer readable storage medium, having stored thereon a computer program executable by a processor to implement the steps of the method of any one of claims 1 to 8.

10. An electronic device, comprising:

one or more processors;

storage means for storing one or more programs which, when executed by the one or more processors, cause the electronic device to perform the steps of the method of any of claims 1-8.