A method for training a blood pressure prediction model based on meta learning

Abstract

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

Description

Technical Field

[0001] This invention relates to the field of deep learning, specifically to the field of blood pressure prediction within deep learning, and more specifically, to a method for training a blood pressure prediction model based on meta-learning for blood pressure prediction. Background Technology

[0002] Blood pressure refers to the force that propels blood into the arteries when the heart contracts and the pressure of blood on the arteries when the heart relaxes. Normal blood pressure should fall within a certain range. Both excessively high and low blood pressure can negatively impact health. When blood pressure consistently rises to a certain level, it can damage the cardiovascular system, leading to diseases such as coronary heart disease, cardiomegaly, and myocardial infarction. It also increases the risk of stroke, retinopathy, and kidney damage, and may even cause short-term or long-term damage to organs such as the heart and brain, or even death. Conversely, when blood pressure consistently drops to a certain level, it can cause discomfort such as dizziness, weakness, and fainting. This is especially true for the elderly or those with underlying medical conditions. Therefore, regular blood pressure monitoring is crucial for maintaining good health. Two important indicators of blood pressure are systolic blood pressure and diastolic blood pressure. Systolic blood pressure refers to the pressure of blood in the arteries when the left ventricle contracts, and blood flows through the arteries to the whole body. This is usually called the highest blood pressure, and the normal range of systolic blood pressure is between 90-140 mmHg. Diastolic blood pressure refers to the pressure in the arteries when the heart relaxes, and blood flows back into the systemic circulation. This is usually called the lowest blood pressure, and the normal range of diastolic blood pressure is between 60-90 mmHg.

[0003] Existing methods for measuring blood pressure are mainly divided into two types: invasive and non-invasive. Invasive blood pressure measurement involves inserting a device directly into the body through a catheter or needle to measure the actual blood pressure value. This method can measure very accurate blood pressure values, but it is difficult to operate, carries a high risk of infection, and is not suitable for routine use. Non-invasive blood pressure measurement, on the other hand, involves using a pressure cuff or flow cuff, relying on the device itself to change the air pressure to measure blood pressure indirectly. While it does not have a high risk of infection, the process is often more complex, limiting the frequency of blood pressure measurements. Furthermore, this method cannot continuously monitor patient data, failing to provide a reliable dataset for blood pressure prediction models. Subsequently, wearable devices with sensors have been developed to continuously collect patient data, providing reliable blood pressure data for these models.

[0004] Existing blood pressure prediction models are typically built upon large-scale population blood pressure data and trained using conventional methods. However, due to individual differences, these models often fail to accurately reflect each individual's health status and risks when predicting blood pressure. Therefore, personalizing blood pressure models can better address individual differences in prediction and intervention, improving the accuracy and efficiency of blood pressure monitoring. Furthermore, analyzing and modeling individual data can identify more accurate predictive factors and individual characteristics, leading to more precise predictions and interventions.

[0005] However, the current method for personalizing and fine-tuning existing models mainly involves collecting a large number of data samples from target individuals and then fine-tuning some layers of the model to make the existing model adapt to the target individual data in order to obtain good prediction results. However, this method is extremely prone to data overfitting and has the problem of inaccurate measurement. In addition, collecting a large number of data samples is a very complex process. Therefore, existing blood pressure prediction models have certain limitations in predicting blood pressure for individuals. Summary of the Invention

[0006] Therefore, the purpose of this invention is to overcome the shortcomings of the prior art and provide a method for training a blood pressure prediction model based on meta-learning.

[0007] According to one aspect of the present invention, a method for training a blood pressure prediction model based on meta-learning is provided. The method includes: 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 ratio; performing a first training stage, including: pre-training the blood pressure prediction model using the first training set to obtain a pre-trained blood pressure prediction model; performing a second training stage, including: acquiring the pre-trained blood pressure prediction model and initializing an initial meta-learner using the parameters of the pre-trained blood pressure prediction model; acquiring a second training set, the second training set including multiple training tasks, and training the initial meta-learner using the multiple 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 the training data of one patient; and performing a third training stage, including: initializing the target meta-learner for each patient in the third training set to obtain an initial personalized blood pressure prediction model corresponding to each patient, and training the initial personalized blood pressure prediction model corresponding to each patient using the training data of each patient in the third training set to obtain a personalized blood pressure prediction model corresponding to that patient.

[0008] In some embodiments of the present invention, the second training stage includes: acquiring a pre-trained blood pressure prediction model and initializing an initial meta-learner using the parameters of the pre-trained blood pressure prediction model; acquiring a second training set, the second training set including multiple training tasks, wherein the data in one training task corresponds to the training data of one patient, and the training data of each patient includes multiple pulse wave signal segments and their corresponding systolic and diastolic blood pressures, and dividing the training data of each patient into a first support set and a first query set; training the initial meta-learner based on the meta-learning algorithm using multiple training tasks in the second training set, and obtaining a target meta-learner after the completion of the last training task, wherein each training includes: using the meta-learner obtained after the completion of the previous training task as the initial meta-learner for the current training task. The parameters of the initial meta-learner are assigned to the first personalized model of the current training task. The first personalized model is trained iteratively multiple times using the first support set of the current training task to obtain the second personalized model under the current training task. The first query set of the current training task is input into the second personalized model to obtain systolic and diastolic blood pressure. The loss value is determined using the obtained systolic and diastolic blood pressure and the systolic and diastolic blood pressure in the current training task. The gradient is calculated based on the loss value to back-update the parameters of the initial meta-learner of the current training task, thus obtaining the meta-learner under the current training task. The meta-learner under the current training task is used as the initial meta-learner for the next training task. The initial meta-learner initialized with the parameters of the pre-trained blood pressure prediction model is used as the initial meta-learner for the first training task.

[0009] In some embodiments of the present invention, the third training stage includes: acquiring a third training set, the third training set including training data of multiple patients, each patient's training data including multiple pulse wave signal segments and their corresponding systolic and diastolic blood pressures, and dividing each patient's training data into a second support set and a second query set; acquiring a target meta-learner, initializing it for each patient in the third training set using the target meta-learner to obtain an initial personalized blood pressure prediction model for each patient; training the corresponding initial personalized blood pressure prediction model using the second support set of each patient in the third training set to obtain a corresponding personalized blood pressure prediction model; and evaluating the performance of the personalized blood pressure prediction model using the second query set of each patient in the third training set.

[0010] In some embodiments of the present invention, the blood pressure prediction model includes a feature extraction module, a gated loop module, and a regression module, wherein: the feature extraction module includes three extraction layers; the first extraction layer is configured to extract multiple low-level blood pressure features from pulse wave signal segments; the second extraction layer is configured to extract multiple mid-level blood pressure features from the multiple low-level blood pressure features; and the third extraction layer is configured to extract multiple deep-level blood pressure features from the multiple mid-level blood pressure features; and multiple cascaded blood pressure features obtained by superimposing the multiple low-level blood pressure features extracted by the first layer and the multiple deep-level blood pressure features extracted by the third layer according to channels are used as inputs to the gated loop module; the gated loop module is used to model the multiple cascaded blood pressure features according to the time sequence of pulse wave signal segment acquisition, and output multiple blood pressure features with time sequence; the regression module is used to predict blood pressure from the multiple blood pressure features with time sequence, and output the predicted systolic and diastolic blood pressure.

[0011] 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 to convolve the pulse wave signal segment to extract multiple blood pressure features from the pulse wave signal segment; the dimension adjustment unit is used to normalize the dimensions of the extracted multiple blood pressure features so that the dimensions of the multiple blood pressure features tend to be the same; and the mapping unit is used to perform nonlinear mapping on the normalized multiple blood pressure features so that the output multiple blood pressure features are nonlinear.

[0012] In some embodiments of the present invention, the training set includes blood pressure data from multiple patients, which includes pulse wave signals and blood pressure wave signals. The training set is obtained by processing the following methods: deleting signals in the pulse wave signals and blood pressure wave signals whose continuous acquisition duration is less than a preset duration; applying a filter to the pulse wave signals and blood pressure wave signals after signal deletion to obtain filtered pulse wave signals and blood pressure wave signals; segmenting the filtered pulse wave signals and blood pressure wave signals into segments of equal length; using an autocorrelation filter to eliminate damaged pulse wave signal segments in the pulse wave signal segments, and processing the eliminated pulse wave signal segments using mean-variance normalization to obtain pulse wave signal segments in the training set; using a multi-scale peak detection algorithm to remove segments in the blood pressure wave signal segments with a preset number of systolic peaks, and calculating the systolic and diastolic blood pressure of the blood pressure wave signal segments to obtain the systolic and diastolic blood pressure corresponding to the pulse wave signal segments in the training set.

[0013] In some embodiments of the present invention, the step of using filters to filter the pulse wave signal and blood pressure wave signal after signal removal to obtain filtered pulse wave signal and blood pressure wave signal includes: using a fourth-order Butterworth bandpass filter to filter the pulse wave signal and blood pressure wave signal after signal removal to remove noise from the pulse wave signal and blood pressure wave signal; and using a Hamper filter to filter the noise-removed pulse wave signal and blood pressure wave signal to remove outliers from the pulse wave signal and blood pressure wave signal.

[0014] According to a second aspect of the present invention, a blood pressure prediction method is provided, 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 using the method described in the first aspect of the present invention for prediction, and outputting the systolic blood pressure and diastolic blood pressure of the target patient.

[0015] Compared with the prior art, the advantages of the present invention are as follows:

[0016] 1) By training the blood pressure prediction model through the first and second training phases, the blood pressure prediction model trained in these two phases has a good learning ability. Furthermore, in the subsequent third training phase, only a small amount of data from target patients is needed to enable the blood pressure prediction model to achieve good personalized prediction capabilities, truly reflecting each person's health status and risk. At the same time, it improves the prediction effect of the blood pressure prediction model on target individual data. This not only solves the problem in existing technologies that require the collection of a large number of target individual data samples to fine-tune some layers of the model so that the existing model can adapt to the target individual data in order to obtain good prediction results, but also solves the problem of needing to collect a large number of data samples. It also overcomes the limitations of existing blood pressure prediction models in predicting blood pressure for individuals.

[0017] 2) By training the blood pressure prediction model based on the meta-learning algorithm in the second training stage, the trained blood pressure prediction model can be fine-tuned with a small number of target patients to obtain a personalized blood pressure prediction model, without the parameters of the trained model becoming overfitted to the target patients due to a large number of data samples. This solves the problem of data overfitting that is easy to occur in the existing methods of training models. Attached Figure Description

[0018] The embodiments of the present invention will be further described below with reference to the accompanying drawings, wherein:

[0019] Figure 1 This is a schematic diagram of a method for training a blood pressure prediction model based on meta-learning according to an embodiment of the present invention.

[0020] Figure 2This is a schematic diagram illustrating the process of processing the training set according to an embodiment of the present invention;

[0021] Figure 3 This is a schematic diagram of the blood pressure prediction model structure according to an embodiment of the present invention;

[0022] Figure 4 This is a schematic diagram of the second training stage according to an embodiment of the present invention;

[0023] Figure 5 This 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.

[0024] Figure 6 This is a schematic diagram illustrating the technical implementation process of training a blood pressure prediction model based on meta-learning according to an embodiment of the present invention.

[0025] Figure 7 This is a schematic diagram illustrating the blood pressure prediction effect of subject number one according to an embodiment of the present invention;

[0026] Figure 8 This is a schematic diagram illustrating the blood pressure prediction effect of subject number two according to an embodiment of the present invention;

[0027] Figure 9 This is a schematic diagram of the systolic blood pressure error value as a function of the number of samples according to an embodiment of the present invention;

[0028] Figure 10 This is a schematic diagram of the diastolic blood pressure error value as a function of the number of samples, according to an embodiment of the present invention. Detailed Implementation

[0029] To make the objectives, technical solutions, and advantages of this invention clearer, the invention is further described in detail below through specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0030] As described in the background section, existing blood pressure prediction models are extremely prone to overfitting and inaccurate measurement.

[0031] To address the aforementioned problems, this invention proposes a method for training a blood pressure prediction model based on meta-learning, such as... Figure 1As shown, the method includes the following steps: acquiring a training set, dividing the training set into a first training set, a second training set, and a third training set according to a preset ratio; performing a first training stage, including: pre-training a blood pressure prediction model using the first training set to obtain a pre-trained blood pressure prediction model; performing a second training stage, including: acquiring the pre-trained blood pressure prediction model and initializing an initial meta-learner using the parameters of the pre-trained blood pressure prediction model; acquiring a second training set, the second training set including multiple training tasks, and training the initial meta-learner using the multiple 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 the training data of one patient; performing a third training stage, including: initializing the target meta-learner for each patient in the third training set to obtain an initial personalized blood pressure prediction model corresponding to each patient, and training the initial personalized blood pressure prediction model corresponding to each patient using the training data of each patient in the third training set to obtain a personalized blood pressure prediction model corresponding to that patient. This invention proposes a method for training a blood pressure prediction model based on meta-learning. This method involves three training stages: pre-training, meta-learning training, and fine-tuning with a small amount of target individual data. This improves the prediction performance of the trained blood pressure prediction model on target individual data, solving the problem that existing blood pressure prediction models trained using large-scale population blood pressure data through conventional training methods do not achieve ideal prediction results for target individuals. It also addresses the need to collect a large number of target individual data samples to fine-tune certain layers of the model to adapt it to target individual data and achieve good prediction results. Furthermore, it overcomes the problem of data overfitting that often occurs in existing model training methods, thus overcoming the limitations of existing blood pressure prediction models in predicting individual blood pressure.

[0032] Furthermore, before specifically describing the embodiments of the present invention, some of the terms used herein are explained as follows:

[0033] Meta-learning refers to the learning of learning mechanisms by an individual, that is, a type of machine learning algorithm that learns how to learn. In traditional machine learning, features and models are usually designed manually for each task, and the prediction results are optimized and adjusted using the selected models and their parameters. This process requires experience and expertise, and the model and corresponding parameters need to be redesigned for new tasks. Meta-learning, on the other hand, learns information and patterns from a set of tasks to obtain some general learning strategies, making it more adaptable to new tasks and more generalizable.

[0034] The Medical Information Set for Critical Care Medicine III (MIMIC III Waveform Dataset) is a large-scale medical dataset containing physiological monitoring signals collected from 28,000 adult patients in multiple hospitals in Boston, USA, during their time in the Intensive Care Unit (ICU) from 2001 to 2012. This waveform dataset includes various types of physiological monitoring signals, such as photoplethysmography (PPG), electrocardiogram (ECG), pulse oxygen saturation, respiratory rate, and blood pressure. Each patient's waveform data records historical values ​​of their heart rate, respiration, and various other physiological variables.

[0035] Photoplethysmography (PPG) is a non-invasive physiological signal monitoring technology. It uses photoelectric sensors to detect photoelectric signals on the skin surface, obtaining the weak signals generated by the minute expansion and contraction of the skin during a heartbeat. This technology is primarily based on optical principles. A beam of infrared light is shone onto the skin surface, and the photoelectric sensor measures the light transmitted through the skin after the light is shone, obtaining the reflected light signal from the skin. This allows for the detection of the pulse wave morphology. The pulse wave is the waveform formed by the blood flow within blood vessels caused by the heartbeat, including peaks and troughs. By detecting the pulse wave, physiological parameters such as heart rate and blood pressure can be calculated, providing an assessment indicator of the cardiovascular system's functional status.

[0036] Mean-variance normalization, also known as Z-score standardization, is a data preprocessing method used to transform the original dataset into a normally distributed dataset with a mean of 0 and a standard deviation of 1. Its basic idea is to perform a linear transformation on the original dataset so that the transformed dataset has a mean of 0 and a variance of 1.

[0037] To better understand the method for training a blood pressure prediction model based on meta-learning proposed in this invention, a detailed description is provided below with reference to specific embodiments, accompanying drawings, and examples.

[0038] I. Training Set

[0039] First, a training set is acquired and divided into a first training set, a second training set, and a third training set according to a preset ratio. According to some embodiments of the present invention, the training set can be an open-source dataset, such as the MIMIC I dataset, MIMIC II dataset, MIMIC III waveform dataset, UCI-BP dataset, etc.; it can also be a dataset autonomously collected by some wearable sensors. In this invention, the training set uses the MIMIC III dataset, which includes photoplethysmography (PPG) pulse wave signals and arterial blood pressure waveform signals (in the following description, PPG pulse wave signals are used to represent PPG pulse wave signals, and blood pressure waveform signals are used to represent arterial blood pressure waveform signals). The data in the training set is preprocessed, and the processed training set is then divided into a first training set, a second training set, and a third training set according to a preset ratio. Figure 2As shown, the method for preprocessing the data in the training set includes the following steps:

[0040] Step S1: Delete the signals in the pulse wave signal and blood pressure wave signal whose continuous acquisition duration is less than the preset duration.

[0041] According to one embodiment of the present invention, since the acquisition frequency of the MIMIC III waveform dataset is 125Hz, that is, 125 data points are acquired per second, it can be determined whether the continuous acquisition duration of each signal is less than a preset duration. If it is less than the preset duration, the signal is deleted. Indicatively, the preset duration can be set to 10 minutes. Then, the length of a signal that meets the requirements is greater than or equal to 125*10*60. If the signal length is less than 125*10*60, it can be determined that the continuous acquisition duration of this signal is less than the preset duration, and it is deleted. It should be understood that the above is only an illustrative example, and those skilled in the art can adjust the preset duration, for example, by adjusting it to 12 minutes, 14 minutes, etc., to obtain other embodiments.

[0042] Step S2: Use filters to process the pulse wave signal and blood pressure wave signal after signal deletion to obtain filtered pulse wave signal and blood pressure wave signal.

[0043] According to one embodiment of the present invention, since anything below 0.5Hz can be attributed to baseline drift, and anything above 8Hz is high-frequency noise, a fourth-order Butterworth bandpass filter with cutoff frequencies of 0.5Hz and 8Hz can be used first to filter the pulse wave and blood pressure signals after signal removal, removing noise from the pulse wave and blood pressure signals. Then, a Hampel filter is used to filter the noise-removed pulse wave and blood pressure signals to remove outliers. The outlier removal method is as follows: based on the size of the sliding window, the median of the data points in each window is calculated, and its standard deviation is calculated based on the calculated median. If the first value in each window exceeds a threshold, the median of that window is used instead of this value. To illustrate, let's set the window size to 3 and the threshold to 3 times the standard deviation. Assuming the dataset is [5, 8, 12, 10, 6, 15, 7, 9, 13, 20, 25, 30, 17], for the value 10, the window data is [5, 8, 12, 10, 6, 15, 7], with a median of 8 and a standard deviation of 3.72. Since the value 10 does not exceed 3 times its standard deviation, it is not an outlier and does not require replacement. For the value 15, the window data is [12, 10, 6, 15]. The dataset is set as follows: [[7,9,13], with a median of 10 and a standard deviation of 3.27. The value 15 exceeds three times its standard deviation, indicating it is an outlier. Therefore, it is replaced with the median of 10. The median and standard deviation of each value within the corresponding window are calculated sequentially using the above method and then replaced. Since the first three and last three data points in the dataset cannot be calculated based on the data in the corresponding window, the final corrected dataset is [*,*,*,10,6,10,7,9,13,20,*,*,*]. It should be understood that the above is merely an illustrative example. Those skilled in the art can adjust the sliding window size and threshold, for example, by adjusting the window size to 5 or 7, and the threshold to 2 or 4, to obtain other embodiments.

[0044] Step S3: Divide the filtered pulse wave signal and blood pressure wave signal into segments of the same length.

[0045] According to one embodiment of the present invention, a preset sliding window is used to divide the pulse wave signal and blood pressure wave signal into signal segments of equal length. Indicatively, the duration of each signal segment can be set to 5 seconds. Since the frequency of the acquired signal is 125Hz, the length of each signal segment is 625. Therefore, a sliding window with a size of 625 and a step size of 250 can be used for signal segmentation. It should be understood that the above is only an illustrative example, and those skilled in the art can adjust the duration of each signal segment and the sliding window, for example, setting the duration to 4 seconds, 6 seconds, etc., and adjusting the sliding window size to 600, the step size to 240, etc., to obtain other embodiments.

[0046] Step S4: Use an autocorrelation filter to eliminate damaged pulse wave signal segments in the pulse wave signal segments, and use mean-variance normalization to process the eliminated pulse wave signal segments to obtain the pulse wave signal segments in the training set.

[0047] According to one embodiment of the present invention, in order to ensure the high periodicity of a normal photoplethysmography (PPG) pulse wave signal segment, the autocorrelation signal of the pulse wave signal segment can be calculated using an autocorrelation filter, and a maximum autocorrelation threshold can be set. If the calculated autocorrelation signal is greater than the maximum autocorrelation threshold, it indicates that the pulse wave signal segment is damaged, and this signal segment is eliminated. After eliminating the damaged pulse wave signal segment, mean-variance normalization is used to process the eliminated pulse wave signal segment. Schematic, the mean and standard deviation of all pulse wave signal segments in the dataset are first calculated using the mean calculation formula and the standard deviation calculation formula, and then the following transformation is performed on each pulse wave signal segment using the following formula: Where x represents the original pulse wave signal segment, μ represents the mean, σ represents the standard deviation, and x′ represents the pulse wave signal segment after mean-variance normalization. It should be understood that the above is merely an illustrative example, and those skilled in the art can employ other normalization methods for data processing, such as maximum-minimum value normalization, decimal scaling, linear proportional normalization, etc., to obtain other embodiments.

[0048] Step S5: Use a multi-scale peak detection algorithm to remove segments with a preset number of systolic peaks from the blood pressure wave signal segments, and calculate the systolic and diastolic blood pressure of the blood pressure wave signal segments to obtain the systolic and diastolic blood pressure corresponding to the pulse wave signal segments in the training set.

[0049] According to one embodiment of the present invention, since a signal segment contains multiple peaks and troughs, a multi-scale peak detection algorithm can be used to find the peaks and troughs of the blood pressure wave signal segment. The average value of the peaks is calculated as the systolic blood pressure of the blood pressure wave signal segment, and the average value of the troughs is calculated as the diastolic blood pressure of the blood pressure wave signal segment. This algorithm analyzes the periodicity information of the signal by calculating the distance between the local minimum and the global minimum, thereby determining the peak position of the signal. A normal heart rate is generally above 60 beats per minute, but since the algorithm may not have a complete cardiac cycle at the beginning or end of the segment, it cannot detect the first or last systolic peak of the segment. Therefore, signal segments with fewer than a preset number of systolic peaks are removed. For example, for a blood pressure wave signal segment with a continuous acquisition time of 5 seconds, there are more than 5 detectable systolic peaks. To ensure the data meets the requirements, the preset number can be set to 4, meaning that blood pressure wave signal segments with fewer than 4 systolic peaks are removed. It should be understood that the above is only an illustrative example. Those skilled in the art can set the preset number to other values ​​according to the acquisition time of the signal segment. For example, if the acquisition time is 6 seconds, the preset number can be set to 5, etc., to obtain other embodiments.

[0050] Based on the above data processing method, a preprocessed training set is obtained. Then, taking individual patients as the unit, the preprocessed training set is divided into a first training set, a second training set, and a third training set in a certain proportion (e.g., 4:4:2). The three training sets are used for the first training stage, the second training stage, and the third training stage to train the blood pressure prediction model, respectively.

[0051] II. Model Structure

[0052] Secondly, a blood pressure prediction model is obtained based on the training set acquired in the first part. In some embodiments of the present invention, the blood pressure prediction model can use a fully connected neural network model, a convolutional neural network model, an adversarial network model, a decision tree, and a support vector machine, etc. Existing conventional blood pressure prediction models generally include a feature extraction model for extracting blood pressure features and a prediction module for predicting blood pressure based on blood pressure features. However, to better illustrate the technical solution of the present invention, the present invention constructs a blood pressure prediction model, the model structure of which is as follows: Figure 3As shown, the blood pressure prediction model includes a feature extraction module, a gated loop module, and a regression module. The feature extraction module comprises three extraction layers: a first extraction layer is configured to extract multiple low-level blood pressure features from pulse wave signal segments; a second extraction layer is configured to extract multiple mid-level blood pressure features from the multiple low-level blood pressure features; and a third extraction layer is configured to extract multiple deep-level blood pressure features from the multiple mid-level blood pressure features. Multiple cascaded blood pressure features obtained by superimposing the multiple low-level blood pressure features extracted by the first layer and the multiple deep-level blood pressure features extracted by the third layer are used as input to the gated loop module. The gated loop module is used to model the multiple cascaded blood pressure features according to the time sequence of pulse wave signal segment acquisition, outputting multiple blood pressure features with a time sequence. The regression module is used to predict blood pressure based on the multiple blood pressure features with a time sequence, outputting the predicted systolic and diastolic blood pressure.

[0053] According to an embodiment of the present invention, each extraction layer in the feature extraction module includes: a convolution unit, a dimension adjustment unit, and a mapping unit; wherein: the convolution unit is used to convolve the pulse wave signal segment to extract multiple blood pressure features from the pulse wave signal segment; the dimension adjustment unit is used to normalize the dimensions of the extracted multiple blood pressure features so that the dimensions of the multiple blood pressure features tend to be the same; the mapping unit is used to perform nonlinear mapping on the normalized multiple blood pressure features so that the output multiple blood pressure features are nonlinear. The mapping unit uses activation functions to perform nonlinear mapping on the multiple blood pressure features, and the activation functions include, but are not limited to: ReLU activation function, Sigmoid activation function, Softmax activation function, etc.

[0054] According to one embodiment of the present invention, the regression module includes two fully connected layers. The first fully connected layer transforms multiple blood pressure features with time sequence and calculates systolic and diastolic blood pressure. The second fully connected layer reduces the dimension of the calculated systolic and diastolic blood pressure and outputs the final systolic and diastolic blood pressure.

[0055] It is worth noting that the blood pressure prediction model constructed above can be used not only to extract features of photoplethysmography (PPG) signals, but also to extract features of electrocardiogram (ECG) signals. Furthermore, depending on the application scenario, the number of units in each layer of the model can be adjusted to obtain a model structure applicable to other scenarios. Of course, a suitable prediction model can also be built based on the different sample sizes in the application scenario to improve the accuracy of blood pressure prediction.

[0056] III. Training Process

[0057] The first, second, and third training sets from the first part are used to train the blood pressure prediction model constructed in the second part.

[0058] 1. First training phase

[0059] The blood pressure prediction model was pre-trained using the first training set to obtain the pre-trained blood pressure prediction model.

[0060] According to one embodiment of the present invention, a first training set is obtained, comprising multiple training samples and corresponding labels. The training samples include pulse wave signal segments, and the labels are the systolic and diastolic blood pressure corresponding to the training samples. Multiple samples from the first training set are input into the feature extraction module of the blood pressure prediction model to extract multiple low-level and multiple high-level blood pressure features from the pulse wave signal segments. These low-level and high-level blood pressure features are then superimposed by channel and input into the gated loop module of the blood pressure prediction model for modeling, outputting multiple time-ordered blood pressure features. The regression module of the blood pressure prediction model is used to predict blood pressure using these time-ordered features, outputting the predicted systolic and diastolic blood pressures. A loss value is determined based on the predicted systolic and diastolic blood pressures and the labels, and a gradient is calculated based on the loss value. The parameters of the blood pressure prediction model are updated based on the obtained gradient and a preset first learning rate to obtain a pre-trained blood pressure prediction model. Indicatively, the first learning rate can be set to 0.001 or 0.002.

[0061] According to one embodiment of the present invention, in the first training phase, a loss value is calculated using a preset loss function, which is as follows:

[0062]

[0063] Where n represents the number of training samples, This represents the systolic and diastolic blood pressure corresponding to the i-th pulse wave signal segment output. This represents the label corresponding to the i-th pulse wave signal segment.

[0064] 2. Second training phase

[0065] A pre-trained blood pressure prediction model is obtained, and the parameters of the pre-trained blood pressure prediction model are used to initialize the initial meta-learner; a second training set is obtained, which includes multiple training tasks. The initial meta-learner is trained using the multiple training tasks in the second training set based on the meta-learning algorithm to obtain the target meta-learner; wherein, the data in one training task corresponds to the training data of one patient.

[0066] According to one embodiment of the present invention, firstly, a pre-trained blood pressure prediction model completed in the first training phase is obtained, and the parameters of the pre-trained blood pressure prediction model are used to initialize an initial meta-learner. Secondly, a second training set is obtained, which includes multiple training tasks. The data in one training task corresponds to the training data of one patient. Each patient's training data includes multiple pulse wave signal segments and their corresponding systolic and diastolic blood pressures. The training data of each patient is divided into a first support set and a first query set. Based on the initial meta-learner, a meta-learning algorithm is used to iteratively train the model multiple times using the multiple training tasks in the second training set. The target meta-learner is obtained after the last training task is completed. Each training step includes: using the meta-learner obtained from the previous training task as the initial meta-learner for the current training task, and assigning the parameters of the initial meta-learner to the first personalized set of the current training task. The model iterates through the first support set of the current training task to train the first personalized model multiple times, resulting in a second personalized model for the current training task. The first query set of the current training task is input into the second personalized model to obtain systolic and diastolic blood pressure. The loss value is determined using the obtained systolic and diastolic blood pressure, along with the systolic and diastolic blood pressure from the current training task. The gradient is then used to back-update the parameters of the initial meta-learner for the current training task, resulting in a meta-learner for the current training task. This meta-learner is then used as the initial meta-learner for the next training task. Specifically, the initial meta-learner initialized with the parameters of the pre-trained blood pressure prediction model is used as the initial meta-learner for the first training task. For example, if a patient has 1000 training data points, each including a pulse wave signal segment and its corresponding systolic and diastolic blood pressure, these 1000 training data points are divided into a first support set and a query set. Assuming x training data points form the first support set, the remaining 1000-x training data points form the first query set.

[0067] According to an embodiment of the present invention, the specific process of training a first personalized model using a first support set in a training task includes: inputting a pulse wave signal fragment from the first support set in the current training task into the feature extraction module of the first personalized model to extract multiple low-level blood pressure features and multiple high-level blood pressure features from the pulse wave signal fragment; superimposing the multiple low-level blood pressure features and multiple high-level blood pressure features by channel into the gated loop module of the first personalized model for modeling, and outputting multiple blood pressure features with a time order; using the regression module of the first personalized model to predict blood pressure from the multiple blood pressure features with a time order, and outputting the predicted systolic and diastolic blood pressure; determining the loss value based on the predicted systolic and diastolic blood pressure and the label, and calculating the gradient based on the loss value; and updating the parameters of the first personalized model in reverse based on the obtained gradient and a preset second learning rate to obtain a second personalized model. Indicatively, the second learning rate can be set to 0.0001 or 0.0002. During training, the loss value is still calculated using the loss function as in the first training stage.

[0068] According to an embodiment of the present invention, the specific process of training a personalized model using a first query set in a training task includes: inputting a pulse wave signal fragment from the first query set in the current training task into the feature extraction module of the second personalized model to extract multiple low-level blood pressure features and multiple high-level blood pressure features from the pulse wave signal fragment; superimposing the multiple low-level blood pressure features and multiple high-level blood pressure features by channel into the gated loop module of the second personalized model for modeling, and outputting multiple blood pressure features with a time order; using the regression module of the second personalized model to predict blood pressure from the multiple blood pressure features with a time order, and outputting the predicted systolic and diastolic blood pressure; determining the loss value based on the predicted systolic and diastolic blood pressure and the label, and calculating the gradient based on the loss value; updating the parameters of the 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, and the third learning rate is greater than the second learning rate. For example, the third learning rate can be set to 0.001 or 0.002. During training, the loss value is still calculated using the loss function as in the first training stage. The technical solution of this embodiment can achieve at least the following beneficial technical effects: setting a smaller second learning rate enables the model to learn how to learn, and setting the third learning rate to be greater than the second learning rate enables the model that has already learned to learn through the second learning rate to learn knowledge more quickly, thereby improving the convergence speed of the model.

[0069] According to one example of the invention, such as Figure 4As shown, the second training set includes n training tasks, with each training task corresponding to the training data of a patient. First, the initial meta-learner is initialized using the parameters of the pre-trained blood pressure prediction model to obtain the initial meta-learner for the first training task. The parameters of the initial meta-learner are assigned to the first personalized model in the first training task. The first personalized model is iteratively trained using the first support set of the first training task to obtain the second personalized model in the first training task. The first query set of the first training task is input into the second personalized model to obtain systolic and diastolic blood pressure. The loss value is determined using the obtained systolic and diastolic blood pressure and the systolic and diastolic blood pressure in the first training task. The parameters of the initial meta-learner are then updated in reverse based on the gradient of the loss value to obtain the meta-learner under the first training task. The meta-learner obtained after the completion of the first training task is used as the initial meta-learner in the second training task. The above operations are performed to obtain the meta-learner under the second training task until the nth training task is completed, resulting in the final target meta-learner.

[0070] According to some embodiments of the present invention, based on the different distributions of data samples in the training set, a suitable meta-learning algorithm can be selected to perform meta-learning training on the blood pressure prediction model, so that the model trained by meta-learning has better prediction performance. The meta-learning algorithms include, but are not limited to: model-agnostic meta-learning (MAML), meta-gradient descent (Reptile), and meta-stochastic gradient descent (Meta-SGD). The technical solution of this embodiment can achieve at least the following beneficial technical effects: re-training the pre-trained model using meta-learning not only improves the model's prediction accuracy but also allows the model to be trained using only a small amount of data from target patients during the subsequent third training stage, achieving excellent personalized prediction capabilities.

[0071] 3. Third training phase

[0072] For each patient in the third training set, the target meta-learner is initialized to obtain an initial personalized blood pressure prediction model for each patient. The initial personalized blood pressure prediction model for each patient in the third training set is then trained using the training data of each patient in the third training set to obtain a personalized blood pressure prediction model for that patient.

[0073] According to one embodiment of the present invention, firstly, a third training set is obtained, which includes training data from multiple patients. Each patient's training data includes multiple pulse wave signal segments and their corresponding systolic and diastolic blood pressures. The training data for each patient is then divided into a second support set and a second query set. Next, a target meta-learner is obtained. Based on the training data of each patient in the third training set, the target meta-learner obtained in the second training phase is initialized for each patient to obtain an initial personalized blood pressure prediction model. The initial personalized blood pressure prediction model for each patient in the third training set is trained using the second support set of each patient in the third training set to obtain the corresponding personalized blood pressure prediction model. Finally, the performance of the personalized blood pressure prediction model is evaluated using the second query set of each patient in the third training set.

[0074] According to an embodiment of the present invention, the process of training the corresponding initial personalized blood pressure prediction model using the second support set of each patient includes: inputting the pulse wave signal fragments from the second support set of each patient into the feature extraction module of the corresponding initial personalized blood pressure prediction model to extract multiple low-level blood pressure features and multiple high-level blood pressure features from the pulse wave signal fragments; superimposing the multiple low-level blood pressure features and multiple high-level blood pressure features by channel into the gated loop module of the corresponding initial personalized blood pressure prediction model for modeling, and outputting multiple blood pressure features with time order; using the regression module of the corresponding initial personalized blood pressure prediction model to predict blood pressure from the multiple blood pressure features with time order, and outputting the predicted systolic blood pressure and diastolic blood pressure; determining the loss value based on the predicted systolic blood pressure and diastolic blood pressure and the label, and back-updating the corresponding initial personalized blood pressure prediction model based on the gradient of the loss value to obtain the personalized blood pressure prediction model corresponding to the patient.

[0075] According to one embodiment of the present invention, the process of evaluating the performance of the personalized blood pressure prediction model using a second query set for each patient includes: inputting pulse wave signal segments from the second query set of each patient into the feature extraction module of the corresponding personalized blood pressure prediction model to extract multiple low-level blood pressure features and multiple high-level blood pressure features from the pulse wave signal segments; superimposing the multiple low-level blood pressure features and multiple high-level blood pressure features by channel into the gated loop module of the corresponding personalized blood pressure prediction model for modeling, and outputting multiple blood pressure features with time order; using the regression module of the corresponding personalized blood pressure prediction model to predict blood pressure from the multiple blood pressure features with time order, and outputting the predicted systolic and diastolic blood pressure; and evaluating the performance of the model based on the predicted systolic and diastolic blood pressure and the systolic and diastolic blood pressure from the second query set.

[0076] It is worth noting that, compared to the number of iterations in which the first support set trains the first personalized model in the second training phase, the number of iterations in which the second support set trains the corresponding initial personalized blood pressure prediction model in this phase can be appropriately increased or decreased based on the model's performance evaluation results.

[0077] IV. Application Scenarios

[0078] This invention uses a personalized blood pressure model obtained from the training set, model structure, and training process described above to predict blood pressure from the pulse wave signal of a target patient.

[0079] The pulse wave signal of the target patient is acquired; the pulse wave signal of the target patient is input into a personalized blood pressure prediction model for prediction, and the systolic and diastolic blood pressure of the target patient are output. More specifically, the feature extraction module of the personalized blood pressure prediction model extracts multiple low-level blood pressure features and multiple high-level blood pressure features from the pulse wave signal. The multiple low-level blood pressure features and multiple high-level blood pressure features are superimposed by channels and input into the gated loop module of the personalized blood pressure prediction model for modeling, and multiple blood pressure features with time order are output. The regression module of the personalized blood pressure prediction model is used to predict the multiple blood pressure features with time order, and the systolic and diastolic blood pressure of the target patient are output.

[0080] To better demonstrate the overall technical solution of this invention, the following description is in conjunction with the appendix. Figure 5 and attached Figure 6 Please provide an explanation.

[0081] Figure 5 This paper illustrates a schematic diagram of the technical implementation framework for training a blood pressure prediction model based on meta-learning, as proposed in this invention. First, an open-source dataset is acquired and preprocessed. The preprocessed dataset is divided into a first training set, a second training set, and a third training set. These three sets are then used for the pre-training, meta-learning, and testing phases, respectively. In the pre-training phase, the initial neural network model (blood pressure prediction model) is trained using the first training set to obtain a pre-trained model. In the meta-learning phase, the pre-trained model is acquired and used as the initial meta-learner. The initial meta-learner is trained using the second training set to obtain a target meta-learner. In the testing phase, the third training set is divided into a support set and a query set. The target meta-learner is fine-tuned using the support set to obtain a personalized model. This personalized model is then used to predict blood pressure based on data in the query set, outputting systolic and diastolic blood pressure.

[0082] Figure 6 This invention presents a schematic diagram illustrating the technical implementation process of training a blood pressure prediction model based on meta-learning, which includes the following steps:

[0083] T1. Obtain an open-source dataset containing photoplethysmography pulse wave information and arterial blood pressure waveform signals;

[0084] T2. Preprocess the photoplethysmography pulse wave signal and arterial blood pressure waveform signal, and divide the preprocessed dataset into the first training set, the second training set and the third training set.

[0085] T3. Construct a blood pressure prediction model based on photoplethysmography pulse wave signal to predict systolic and diastolic blood pressure, and determine the initialization parameters of this model;

[0086] T4. Pre-training phase: The blood pressure prediction model is trained and updated using the data samples from the first training set.

[0087] In the T5 meta-learning stage, the pre-trained model is used as the initial meta-learner, and the initial meta-learner is trained and updated using the second training set to obtain the target meta-learner.

[0088] T6. During the testing phase, the target meta-learner is fine-tuned using the third training set data samples, and its training effect is evaluated.

[0089] To better demonstrate the predictive performance of the personalized blood pressure prediction model trained using the above method, experimental tests were conducted using the personalized blood pressure prediction model. The experimental results are as follows: Figure 7-10 As shown, where:

[0090] Figure 7 The image shows the effect of using the personalized blood pressure prediction model trained by this invention to predict the blood pressure of subject number one. Figure 8 The image shows the effect of using the personalized blood pressure prediction model trained according to this invention to predict the blood pressure of subject number two. Figure 7 and Figure 8 The red line represents the actual systolic and diastolic blood pressure values ​​of the subjects, while the blue line represents the predicted systolic and diastolic blood pressure values. In this experiment, only 10 samples from the subjects were used for personalized training of the meta-learner, with each sample collected over a 5-second period. Therefore, only 50 seconds of data from the subjects was needed to obtain a good personalized prediction model. Figure 7 and Figure 8 As can be seen, the blood pressure prediction model obtained by this invention shows a small difference between the predicted systolic and diastolic blood pressure and the actual values, and accurately reflects the trend of blood pressure changes. This also demonstrates that the personalized blood pressure prediction model obtained through the training method proposed in this invention performs excellently in personalized prediction, exhibiting high accuracy in predicting blood pressure for different individuals.

[0091] Figure 9The graphs show the prediction error values ​​of systolic blood pressure using a conventional pre-training and fine-tuning method and the training method proposed in this invention, under different target subject sample sizes. Figure 10 The diagram shows the prediction error values ​​of diastolic blood pressure using a conventional pre-training and fine-tuning method and the training method proposed in this invention, with different numbers of target subjects. In the experiments, different numbers of target subject training samples (50, 25, 10, and 5) were used to test the performance of the models trained by the two methods. Figure 9 and Figure 10 The + curve represents the prediction error value of the model obtained by the training method mentioned in this invention, and the × curve represents the prediction error value of the model obtained by the conventional pre-training and fine-tuning method; the horizontal axis represents the number of training samples from the target subjects, and the vertical axis represents the error value. Figure 9 and Figure 10 It is evident that the error value of the blood pressure prediction model obtained by the training method of this invention is smaller than that of the conventional method, indicating that the model obtained by the method proposed in this invention has better prediction performance and better predictive ability for blood pressure of target subjects with a small sample size.

[0092] In summary, the method for training a blood pressure prediction model based on meta-learning proposed in this invention has the following advantages:

[0093] 1) By training the blood pressure prediction model through the first and second training phases, the blood pressure prediction model trained in these two phases has a good learning ability. Furthermore, in the subsequent third training phase, only a small amount of data from target patients is needed to enable the blood pressure prediction model to achieve good personalized prediction capabilities, truly reflecting each person's health status and risk. At the same time, it improves the prediction effect of the blood pressure prediction model on target individual data. This not only solves the problem in existing technologies that require the collection of a large number of target individual data samples to fine-tune some layers of the model so that the existing model can adapt to the target individual data in order to obtain good prediction results, but also solves the problem of needing to collect a large number of data samples. It also overcomes the limitations of existing blood pressure prediction models in predicting blood pressure for individuals.

[0094] 2) By training the blood pressure prediction model based on the meta-learning algorithm in the second training stage, the trained blood pressure prediction model can be fine-tuned with a small number of target patients to obtain a personalized blood pressure prediction model, without the parameters of the trained model becoming overfitted to the target patients due to a large number of data samples. This solves the problem of data overfitting that is easy to occur in the existing methods of training models.

[0095] It should be noted that although the steps are described in a specific order above, it does not mean that the steps must be executed in the above specific order. In fact, some of these steps can be executed concurrently, or even in a different order, as long as the required function can be achieved.

[0096] This invention can be a system, method, and / or computer program product. A computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for causing a processor to implement various aspects of the invention.

[0097] Computer-readable storage media can be tangible devices that hold and store instructions for use by an instruction execution device. Computer-readable storage media can be, for example, including but not limited to, electrical storage devices, magnetic storage devices, optical storage devices, electromagnetic storage devices, semiconductor storage devices, or any suitable combination thereof. More specific examples (a non-exhaustive list) of computer-readable storage media include: 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 disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, such as punch cards or recessed protrusions storing instructions thereon, and any suitable combination thereof.

[0098] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for training a blood pressure prediction model based on meta-learning, characterized in that, The method includes: Obtain the training set and divide it into the first training set, the second training set, and the third training set according to a preset ratio; The first training phase includes: pre-training the blood pressure prediction model using the first training set to obtain the pre-trained blood pressure prediction model; The second training phase includes: acquiring a pre-trained blood pressure prediction model and initializing an initial meta-learner using the parameters of the pre-trained blood pressure prediction model; acquiring a second training set, which includes multiple training tasks; and training the initial meta-learner using the multiple 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 the training data of one patient. The third training phase includes: initializing each patient in the third training set using the target meta-learner to obtain an initial personalized blood pressure prediction model for each patient; and training the initial personalized blood pressure prediction model for each patient using the training data from the third training set to obtain a personalized blood pressure prediction model for that patient.

2. The method according to claim 1, characterized in that, The second training phase includes: Obtain a pre-trained blood pressure prediction model, and initialize the initial meta-learner with the parameters of the pre-trained blood pressure prediction model; Obtain a second training set, which includes multiple training tasks. The data in one training task corresponds to the training data of one patient. The training data of each patient includes multiple pulse wave signal segments and their corresponding systolic and diastolic blood pressures. The training data of each patient is divided into a first support set and a first query set. Based on the initial meta-learner, it is trained using multiple training tasks in the second training set based on the meta-learning algorithm. The target meta-learner is obtained after the last training task is completed. Each training step includes: The meta-learner obtained from the previous training task is used as the initial meta-learner for the current training task, and the parameters of the initial meta-learner are assigned to the first personalized model of the current training task. The first personalized model is trained iteratively multiple times using the first support set of the current training task to obtain the second personalized model under the current training task. The first query set of the current training task is input into the second personalized model to obtain systolic and diastolic blood pressure. The loss value is determined using the obtained systolic and diastolic blood pressure and the systolic and diastolic blood pressure in the current training task. The gradient is then calculated based on the loss value to back-update the parameters of the initial meta-learner for the current training task, thus obtaining the meta-learner for the current training task. The meta-learner for the current training task is then used as the initial meta-learner for the next training task. The initial meta-learner initialized with the parameters of the pre-trained blood pressure prediction model is used as the initial meta-learner for the first training task.

3. The method according to claim 2, characterized in that, The third training phase includes: Obtain a third training set, which includes training data from multiple patients. Each patient's training data includes multiple pulse wave signal segments and their corresponding systolic and diastolic blood pressures. The training data of each patient is then divided into a second support set and a second query set. Obtain the target meta-learner, and initialize it for each patient in the third training set to obtain the initial personalized blood pressure prediction model for each patient. The second support set of each patient in the third training set is used to train the corresponding initial personalized blood pressure prediction model to obtain the corresponding personalized blood pressure prediction model. The performance of the personalized blood pressure prediction model is evaluated using the second query set for each patient in the third training set.

4. The method according to claim 1, characterized in that, The blood pressure prediction model includes a feature extraction module, a gated recurrent module, and a regression module, wherein: The feature extraction module includes three extraction layers. The first extraction layer is configured to extract multiple low-level blood pressure features from the pulse wave signal segment; the second extraction layer is configured to extract multiple mid-level blood pressure features from the multiple low-level blood pressure features; and the third extraction layer is configured to extract multiple deep-level blood pressure features from the multiple mid-level blood pressure features. The multiple cascaded blood pressure features obtained by superimposing the multiple low-level blood pressure features extracted by the first layer and the multiple deep-level blood pressure features extracted by the third layer are used as the input of the gated loop module. The gated loop module is used to model the multiple cascaded blood pressure features according to the time sequence of pulse wave signal segment acquisition, and output multiple blood pressure features with time sequence. The regression module is used to predict blood pressure based on the multiple time-sequential blood pressure features and output the predicted systolic and diastolic blood pressure.

5. The method according to claim 4, characterized in that, Each extraction layer includes convolutional units, dimension adjustment units, and mapping units, where: Convolutional units are used to convolve pulse wave signal segments to extract multiple blood pressure features from the pulse wave signal segments; The dimension adjustment unit is used to normalize the dimensions of multiple extracted blood pressure features so that the dimensions of multiple blood pressure features tend to be the same. The mapping unit is used to perform nonlinear mapping on multiple normalized blood pressure features so that the output blood pressure features are nonlinear.

6. The method according to claim 1, characterized in that, The training set includes blood pressure data from multiple patients, comprising pulse wave signals and blood pressure wave signals. The training set is obtained by processing the data in the following manner: Delete the pulse wave signal and blood pressure wave signal whose continuous acquisition duration is less than the preset duration; The pulse wave signal and blood pressure wave signal after signal deletion are processed by using a filter to obtain the filtered pulse wave signal and blood pressure wave signal. The filtered pulse wave signal and blood pressure wave signal are divided into segments of the same length; An autocorrelation filter is used to eliminate damaged pulse wave signal segments in the pulse wave signal segment, and mean-variance normalization is used to process the eliminated pulse wave signal segments to obtain the pulse wave signal segments in the training set. A multi-scale peak detection algorithm is used to remove segments with a preset number of systolic peaks from the blood pressure wave signal segments, and the systolic and diastolic blood pressures of the blood pressure wave signal segments are calculated to obtain the systolic and diastolic blood pressures corresponding to the pulse wave signal segments in the training set.

7. The method according to claim 6, characterized in that, The step of using a filter to filter the pulse wave signal and blood pressure wave signal after signal deletion to obtain the filtered pulse wave signal and blood pressure wave signal includes: A fourth-order Butterworth bandpass filter is used to filter the pulse wave signal and blood pressure wave signal after signal deletion to remove noise from the pulse wave signal and blood pressure wave signal. A Hamper filter was used to filter the noise-removed pulse wave signal and blood pressure wave signal to remove outliers from the pulse wave signal and blood pressure wave signal.

8. A method for predicting blood pressure, characterized in that, The method includes: Acquire the pulse wave signal of the target patient; The pulse wave signal of the target patient is input into the blood pressure prediction model trained by any of the methods described in claims 1-7, and the systolic and diastolic blood pressure of the target patient are output.

9. A computer-readable storage medium, characterized in that, It stores a computer program that can be executed by a processor to implement the steps of the method according to any one of claims 1 to 8.

10. An electronic device, characterized in that, include: One or more processors; A storage device for storing one or more programs that, when executed by one or more processors, cause the electronic device to perform the steps of the method as described in any one of claims 1 to 8.

Images