Domain generalization blood pressure estimation training method based on meta learning

Abstract

The invention provides a meta-learning-based domain generalization blood pressure estimation training method, which is characterized in that a learning process is divided into a meta-training stage and a meta-testing stage through a two-stage meta-learning strategy, the meta-training stage is further divided into a training sub-stage and an optimization sub-stage, and in the learning process, the meta-testing stage is further divided into a training sub-stage and an optimization sub-stage; source domains are selected in a transformation mode to form a meta-optimization domain, the other source domains serve as meta-training domains, in each round of meta-training stage, the model is trained on the multiple source domains in the meta-training domains to simulate data distribution of different individuals, and the generalization performance of the model is improved by learning data of the different individuals; in the meta-test stage, the model is finely adjusted by using source domain data so as to improve the prediction precision on a real target domain.

Description

Technical Field

[0001] The present invention relates to the field of data processing, specifically to the field of physiological signal processing, and more specifically to a domain generalization blood pressure estimation training method based on meta-learning. Background Art

[0002] Hypertension is a chronic disease and one of the leading causes of cardiovascular disease and related mortality worldwide. With the aging of the population, the prevalence of hypertension has been increasing year by year. Hypertensive patients need to monitor their blood pressure frequently so that they can adjust their treatment plans in a timely manner to reduce the risk of cardiovascular complications. However, common blood pressure measurement devices (such as mercury sphygmomanometers and cuff sphygmomanometers) currently have many inconveniences in operation, especially when frequent measurements or continuous monitoring are required, the applicability of these devices is limited. Cuff-type blood pressure measurement devices are less applicable in active states, and their measurement method interferes with patients' daily activities, making it difficult to achieve all-weather, non-interference blood pressure monitoring.

[0003] In order to solve the above problems, cuffless blood pressure estimation technology based on photoplethysmography (PPG) and electrocardiogram (ECG) signals has gradually attracted attention in recent years. PPG signals can reflect changes in subcutaneous blood volume and blood flow, while ECG signals reflect the electrical activity of the heart. By analyzing the correlation between PPG and ECG signals, blood pressure can be indirectly inferred. This method is non-invasive, portable and real-time, and is very suitable for real-time blood pressure estimation on portable devices. However, due to the significant differences in individual physiological characteristics, the characteristic distribution of PPG and ECG signals is also different, resulting in unstable performance of blood pressure estimation methods based on machine learning or deep learning on different individuals. To improve accuracy, traditional methods usually require a large amount of personalized data to fine-tune the model, which is difficult to achieve in practical applications. Although transfer learning technology can pre-train models on large-scale public data sets and fine-tune them with a small amount of personalized data, this method still relies heavily on personalized data, resulting in limited generalization ability of the model in practical applications.

[0004] In order to solve the problem of limited generalization ability of the model in practical applications, domain generalization (DG) technology has been introduced into the field of blood pressure estimation. Domain generalization technology trains the model to learn domain-invariant features on multiple source domains, thereby enhancing the model's generalization ability for new domain data.

[0005] However, traditional domain generalization methods have limited performance in the field of physiological signals, especially when dealing with time-series physiological data with high individual differences, it is difficult to obtain ideal estimation results. The temporal characteristics of physiological signals and individual differences lead to inconsistent data distribution in estimation tasks, and existing domain generalization methods are insufficient in dealing with such highly variable data.

[0006] It should be noted that this background technology is only used to introduce the relevant information of the present invention to help understand the technical solution of the present invention, but it does not mean that the relevant information is necessarily the prior art. The relevant information is submitted and disclosed together with the present invention solution. If there is no evidence that the relevant information has been disclosed before the application date of the present invention, the relevant information shall not be regarded as the prior art. Summary of the invention

[0007] Therefore, the object of the present invention is to overcome the above-mentioned defects of the prior art and provide a blood pressure estimation training method.

[0008] The objective of the present invention is achieved through the following technical solutions:

[0009] According to a first aspect of the present invention, a blood pressure estimation training method is provided, comprising: obtaining a blood pressure estimation model, which is used to extract signal features from input sensor data and estimate blood pressure prediction values ​​based on the signal features; obtaining a source domain set including multiple source domains, each source domain including multiple samples of a single object, each sample including sensor data collected from the object and a label indicating the blood pressure value of the object when the sensor data was collected, the sensor data including a PPG signal and an ECG signal; using the source domain set and a preset total loss function, guiding the blood pressure estimation model to perform staged meta-learning to minimize the loss value, wherein the meta-learning The method comprises: performing the learning of the meta-training stage multiple times, each time selecting at least one source domain from the source domain set to form a meta-optimization domain, and the remaining source domains to form a meta-training domain, and each time performing the learning of the meta-training stage comprises: a training sub-stage, comprising: training a blood pressure estimation model using the current meta-training domain and the total loss function; an optimization sub-stage, comprising: training the blood pressure estimation model learned in the sub-stages using the current meta-training domain, the meta-optimization domain and the total loss function; performing the learning of the meta-testing stage, comprising: training the blood pressure estimation model learned in the meta-training stage based on all the source domains in the source domain set and the total loss function. This scheme is divided into a "meta-training stage" and a "meta-testing stage" through a two-stage meta-learning strategy, in which the meta-training stage is further divided into a "training sub-stage" and an "optimization sub-stage". During the learning process, the source domains are selected in a transformation to form the meta-optimization domain, and the remaining source domains are used as meta-training domains. In each round of meta-training, the model is trained on multiple source domains within the meta-training domain to simulate the data distribution of different individuals. By learning from different individual data, the generalization performance of the model is improved; in the meta-testing stage, the model is fine-tuned using the source domain data to improve the prediction accuracy on the real target domain (new individual data). Through the two-stage meta-learning strategy, the model can achieve cross-domain generalization capabilities by simulating the data distribution of different individuals, so that it can achieve high prediction accuracy on new individuals without the need for a large amount of personalized data; compared with traditional methods, this scheme significantly reduces the dependence on personalized data, making the cuffless blood pressure estimation method more practical and efficient in practical applications.

[0010] Optionally, in the meta-training stage, a preset number of unselected source domains are selected from the source domain set each time as meta-optimization domains, wherein the training sub-stage selects batch samples from the current meta-training domain in batches, and iteratively trains the blood pressure estimation model using the batch samples selected from the meta-training domain until the preset requirements are met and then switches to the optimization sub-stage. This solution can at least achieve the following beneficial technical effects: through multiple rounds of iterations of the meta-training domain in the training sub-stage, the model gradually extracts invariant features from different source domains to ensure adaptability to different individual data, and the continuous optimization of the training sub-stage enables the model to have domain invariance, laying the foundation for subsequent optimization in the meta-optimization domain.

[0011] Optionally, the optimization sub-stage is to select batch samples from the current meta-training domain and meta-optimization domain in batches, and iteratively train the blood pressure estimation model using the batch samples selected from the meta-training domain and meta-optimization domain until the preset requirements are met and then transfer to the optimization sub-stage.

[0012] Optionally, the meta-testing phase is performed after the meta-training phase. The meta-testing phase is to select batch samples from all source domains in the source domain set and fine-tune the blood pressure estimation model using the batch samples selected from all source domains.

[0013] Optionally, when training the blood pressure estimation model, a preset training component is used, and the training component includes: a blood pressure estimation model, including: a feature extraction module for extracting signal features from input sensor data, and a prediction module for determining a blood pressure prediction value based on the signal features; a domain-invariant feature extraction module, for extracting domain-invariant features from signal features; wherein, during training, the regression loss of blood pressure, and the alignment loss and distribution difference between domain-invariant features of a specified domain are used to guide the parameter update of the blood pressure estimation model.

[0014] Optionally, the feature extraction module includes: multiple processing branches, each processing branch is used to extract modal features from a modal signal in the sensor data, wherein each processing branch includes multiple stacked time residual blocks and gated recurrent units, and a fusion module is used to fuse the modal features extracted by the multiple processing branches to obtain signal features.

[0015] Optionally, the total loss function is:

[0016]

[0017] in, represents the sub-function used to calculate the blood pressure prediction value and the regression loss between labels, represents a sub-function for calculating the domain-invariant feature alignment loss for a specified domain, represents a subfunction for computing the distribution difference between domain-invariant features of a specified domain, express The weight of express The weight of express The weight of .

[0018] According to a second aspect of the present invention, there is provided a blood pressure estimation method, comprising: obtaining a trained blood pressure estimation model obtained according to the method described in the first aspect; obtaining sensor data of an object to be evaluated, and estimating a blood pressure prediction value based on the sensor data of the object to be evaluated using the trained blood pressure estimation model.

[0019] According to a third aspect of the present invention, there is provided an electronic device comprising: one or more processors; and a memory, wherein the memory is used to store executable instructions; the one or more processors are configured to implement the steps of the method described in the first aspect and / or the second aspect by executing the executable instructions. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The embodiments of the present invention are further described below with reference to the accompanying drawings, in which:

[0021] Figure 1 Schematic diagram of the meta-learning process in the blood pressure estimation training method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0022] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below through specific embodiments in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0023] As mentioned in the background technology section, the temporal characteristics of physiological signals and inter-individual differences lead to inconsistent data distribution in estimation tasks, and existing domain generalization methods are insufficient in dealing with such highly variable data. In this regard, the method of the present invention introduces a two-stage meta-learning strategy in blood pressure estimation, that is, by dynamically transforming the composition of the meta-training domain and the meta-optimization domain during the training process, the distribution of new individual data is simulated, so that the model can be effectively generalized through different individual data during the training phase, thereby improving the model's predictive ability on new individuals. In each iteration of the meta-training phase, the source domain set is divided into a "meta-training domain" and a "meta-optimization domain." In the meta-training domain, the model will learn how to find a common feature representation between multiple source domains, align the feature distributions between source domains, and make the features learned by the model consistent across different source domains. The purpose of this step is to keep the model consistent between source domains and reduce the distribution differences between source domains. After training on the meta-training domain, the meta-optimization domain (i.e., the simulated target domain) is used to optimize the generalization ability of the model; in the meta-optimization phase, the distribution difference between the meta-optimization domain and the meta-training domain can be reduced, with the goal of improving the performance of the model in unseen domains, so that the model can also have better feature alignment in unseen domains. In the meta-testing phase, the model is trained using all samples in the source domain set to improve the prediction accuracy in the real target domain (new individual data).

[0024] In order to better understand the solution of the present invention, the following is an introduction from the aspects of sample definition, model structure, loss function design, training process, and application scenarios.

[0025] 1. Sample Definition

[0026] Samples are data used to train, validate, and / or test models. It is usually necessary to collect sensor data from multiple subjects (people) to build a dataset containing multiple samples. Each sample includes sensor data collected from a subject and a label indicating the blood pressure value of the subject when the sensor data was collected. The sensor data includes PPG signals and ECG signals.

[0027] In addition, the source domain and target domain need to be set so that the model can be trained through the source domain, and the target domain is used to test the generalization ability of the model, especially to measure the performance of the model on unseen data.

[0028] Assume that there is a source domain set consisting of N source domains , each source domain contains multiple samples, each sample includes a pair of sensor data (i.e., input data or input features) and a label (or blood pressure label), expressed as ,in, For the The source domain The sensor data in samples, For the The source domain The source domain set contains data of multiple objects, that is, the data of each object is regarded as an independent source domain. The feature distribution of the source domain Since the individual differences of the subjects vary significantly, the feature distribution of each domain is statistically different. Contains samples of one or more objects, each sample includes a pair of sensor data and blood pressure label, represented as ,in, Target domain set Middle The target domain The sensor data in samples, Target domain set Middle The target domain The labels in the samples. Like the source domain, different target domains also correspond to different objects. The data feature distribution of the target domain is different from that of the source domain, so statistically, the feature distribution of the target domain is Source Domain There is a significant difference. The target domain data is not accessed during the training process, but is only tested after the model training is completed to verify the model's predictive ability and cross-domain generalization ability.

[0029] Implementers can collect data sets by themselves or use existing data sets. Existing data sets include MIMIC III, MIMIC IV, or UQVS. During implementation, samples of some objects in the same data set can be divided into the source domain set, and samples of other objects can be divided into the target domain set. Alternatively, one data set can be regarded as the source domain set, and the other data set can be regarded as the target domain set.

[0030] It should be noted that other implementers may also adopt other methods of dividing the source domain and the target domain, for example, dividing different domains according to the similarity of samples, or using a clustering algorithm to cluster all samples to divide them into different domains, and then further dividing the source domain and the target domain from the divided domains.

[0031] 2. Model Structure

[0032] A blood pressure estimation model needs to be deployed to extract signal features from the input sensor data and estimate the blood pressure prediction value based on the signal features.

[0033] According to one embodiment of the present invention, the blood pressure estimation model includes: a feature extraction module for extracting signal features from input sensor data, and a prediction module for determining a blood pressure prediction value based on the signal features. During implementation, the blood pressure estimation model used in the prior art can be adopted. For example, the model in the existing literature involved in the subsequent comparative experiment.

[0034] In order to estimate blood pressure more accurately, the model can be improved. According to one embodiment of the present invention, in an improved blood pressure estimation model, the feature extraction module includes: multiple processing branches, each processing branch is used to extract modal features from a modal signal in the sensor data, wherein each processing branch includes multiple stacked time residual blocks and gated cycle units, and a fusion module is used to fuse the modal features extracted by multiple processing branches to obtain signal features. For example: when the sensor data has two modal signals, PPG signals and ECG signals, the first processing branch can be set to extract the signal features of the PPG signal, and the second processing branch can be set to extract the signal features of the ECG signal. The fusion module is used to fuse the signal features of the PPG signal and the signal features of the ECG signal to obtain the signal features. It should be understood that if there are more modal signals in the sensor data, more processing branches can be set to extract the signal features of the corresponding modalities.

[0035] Preferably, each processing branch includes: a plurality of stacked temporal residual blocks (Temporal ResidualBlocks) and a gated recurrent unit (GRU), wherein the input of the first temporal residual block is a certain modal signal, the input of the subsequent temporal residual block is its previous temporal residual block, and the input of the last temporal residual block is used as the input of the gated recurrent unit. This scheme designs a novel feature extraction module, which uses a deep feature extraction network to extract effective blood pressure estimation features from PPG and ECG signals. The network combines convolutional layers, temporal residual blocks (Temporal ResidualBlocks) and gated recurrent units (GRU), which can capture the temporal dependencies and local features of physiological signals.

[0036] According to an example of the present invention, the structures of the temporal residual block and the gated cycle unit are schematically introduced below:

[0037] (1) Temporal residual block

[0038] The temporal residual block captures local dependencies in the signal through depthwise separable convolutions. Each residual block consists of a depthwise separable convolution layer, a batch normalization layer, and a ReLU activation function, and is channel-wise adjusted through a 1x1 convolution to ensure the integrity of information flow.

[0039] The introduction of residual blocks helps improve the efficiency of feature extraction, avoid gradient vanishing or exploding problems, and enhance the convergence of the model. The convolution kernel size in the residual block is 5, the padding size is 2, and the number of input and output channels is (1, 32), (32, 64), (64, 128), and (128, 128), respectively.

[0040] (2) Gated Recurrent Unit

[0041] The GRU is a modified recurrent neural network that is used to capture long-term dependencies in time series data. In this example, the GRU uses 128 hidden units and is able to effectively model the temporal relationship between PPG and ECG signals.

[0042] In addition, in order to further improve the performance of the model, other modules that assist in model training may also be set. According to one embodiment of the present invention, when training a blood pressure estimation model, a preset training component is used, and the training component includes: a blood pressure estimation model and a domain-invariant feature extraction module. Among them, the domain-invariant feature extraction module is used to extract domain-invariant features from signal features. The domain-invariant feature extraction module is used to assist in the training of the blood pressure estimation model. After the training is completed, blood pressure estimation can be completed using only the trained blood pressure estimation model. This hybrid deep feature extraction architecture adopted by the training component can ensure the extraction of consistent high-dimensional features between domains while maintaining the time invariance of the features extracted by the model.

[0043] According to an example of the present invention, the domain invariant feature extraction module can be implemented with a fully connected network. Schematically, the domain invariant feature extraction module is set as a three-layer fully connected network, wherein the first layer inputs signal features, the input dimension is set to 256, outputs 128-dimensional features, and uses the ReLU activation function; the input of the second layer is the output of the first layer, the input dimension is 128, and the output is 128-dimensional features, and the ReLU activation function is used; the input of the third layer is the output of the second layer, the input dimension of the third layer is 128, and the output is 512-dimensional domain invariant features. In addition, Dropout regularization (using 0.3) can be added to each layer of the domain invariant feature extraction module to improve the generalization ability of the model and prevent overfitting. Moreover, the implementer can also adjust the feature dimension and / or the activation function used each time to obtain other implementation methods. These fully connected layers can extract features that are independent of the task but have domain consistency from the feature representation, which are used for cross-domain invariant feature learning and enhance the robustness of the model.

[0044] According to one embodiment of the present invention, the prediction module can use a single or multiple fully connected layers. As an example, the prediction module can be configured as a multi-layer fully connected layer, for example, a 2-layer fully connected layer, wherein the first layer: the input dimension is 256, the output is 512-dimensional features, and the layer is used to further process the features to enhance task relevance; the second layer: the 512-dimensional features are further mapped to a 2-dimensional output, corresponding to the predicted values ​​of systolic blood pressure (SBP) and diastolic blood pressure (DBP). This two-layer structure design ensures that the model can make full use of the extracted feature information for blood pressure prediction, while maintaining accuracy and consistency in data from different fields.

[0045] 3. Loss Function Design

[0046] According to one embodiment of the present invention, when an existing model is used and no domain-invariant feature extraction module is added to assist in training, the total loss function may only use the existing regression loss function to directly train the blood pressure prediction capability of the model.

[0047] In order to better train the model, when using the preset training components of the above-mentioned embodiment, an improved loss function can be designed to further improve the performance of the model. According to one embodiment of the present invention, the total loss function is:

[0048]

[0049] in, represents the sub-function used to calculate the blood pressure prediction value and the regression loss between labels, represents a sub-function for calculating the domain-invariant feature alignment loss for a specified domain, represents a subfunction for computing the distribution difference between domain-invariant features of a specified domain, express The weight of express The weight of express The total loss function of the present invention adopts the weighted sum of the above three parts to improve the cross-domain feature alignment effect.

[0050] According to an example of the present invention, the above three sub-loss functions are respectively:

[0051] (1) Regression loss of blood pressure estimation : Regression loss is used to calculate the error of predicting blood pressure, and the more robust Huber loss is used to replace the traditional mean square error (MSE). Huber loss is more robust to outliers and is defined as:

[0052]

[0053] in, represents the label (true blood pressure value) and represents the predicted blood pressure value, Indicates the preset threshold. Can be set according to the needs of the implementer, the default is 1.

[0054] (2) Correlation Alignment Loss : It is used to align the covariance matrix between specified domains and reduce the difference in feature distribution between different domains. The formula is as follows:

[0055]

[0056] in, is the covariance matrix of a specified domain, is the covariance matrix of another specified domain, is the dimension of the feature vector, is the Frobenius norm.

[0057] The covariance matrix measures the relationship between different features in the same domain. It does not directly depend on the predicted value of the model, but is calculated based on the input signal characteristics. Assume that there is a domain-invariant feature of a specified domain and domain-invariant features of another specified domain , the covariance matrix can be calculated by the following steps:

[0058] Calculate the mean of a feature:

[0059] Domain-invariant features for a specified domain Calculate the mean vector :

[0060]

[0061] Domain-invariant features for another specified domain Calculate the mean vector :

[0062]

[0063] Compute the covariance matrix:

[0064] ·right Calculate the covariance matrix :

[0065]

[0066] in, represents the number of domain-invariant features of a specified domain, Indicates domain-invariant features, represents transpose,

[0067] ·right Calculate the covariance matrix :

[0068]

[0069] in, represents the number of domain-invariant features of another specified domain, Indicates the first domain-invariant features, Indicates transpose.

[0070] Related alignment loss The purpose of is to align the feature distributions between different domains by minimizing the difference between the covariance matrices of the two specified domains.

[0071] (3) Maximum mean difference loss : The maximum mean difference loss is used to measure the difference in feature distribution between specified domains. The formula is:

[0072]

[0073] in, is the feature mapping function, Represents a specified domain The The domain-invariant features of samples, Indicates another specified domain The Domain-invariant features of samples. and Domain and domain The number of samples.

[0074] It should be understood that the weighted sum of the above three sub-loss functions as the total loss function is only the optimal implementation of the present invention, and the implementer can adjust the total loss function as needed to achieve similar or suboptimal implementations. For example, the total loss function can be set to:

[0075]

[0076] or

[0077]

[0078] The meaning of each letter refers to the aforementioned total loss function and will not be elaborated here.

[0079] In addition, implementers can also add other sub-loss functions based on the above total loss functions to obtain other implementation methods according to their own needs.

[0080] 4. Training Process

[0081] According to one embodiment of the present invention, a blood pressure estimation training method is provided. Figure 1 , including a two-stage meta-learning strategy, wherein in the meta-training stage, at least one source domain is selected from the source domain set to form a meta-optimization domain each time, and the remaining source domains form a meta-training domain, and the blood pressure estimation model is trained with samples from the meta-training domain and the meta-optimization domain; after the meta-training stage, a meta-testing stage is performed to fine-tune the blood pressure estimation model with samples from the source domain set. Both the training and fine-tuning models use sensor data as network input, and the model outputs blood pressure prediction values ​​and / or extracts relevant domain-invariant features, and then performs loss calculation and parameter update.

[0082] According to an embodiment of the present invention, a blood pressure estimation training method is provided, comprising steps S1, S2 and S3.

[0083] Step S1: Obtain a blood pressure estimation model, which is used to extract signal features from input sensor data and estimate blood pressure prediction values ​​based on the signal features.

[0084] According to an embodiment of the present invention, step S1 is to obtain an initial blood pressure estimation model, the parameters of which can be obtained by random initialization. In addition, if the preset training component of the above embodiment is used, the parameters of the domain invariant feature extraction module need to be randomly initialized.

[0085] Step S2: Acquire a source domain set including multiple source domains, each source domain includes multiple samples of a single object, each sample includes sensor data collected from the object and a label indicating the blood pressure value of the object when the sensor data is collected, and the sensor data includes a PPG signal and an ECG signal.

[0086] S3. Using the source domain set and the preset total loss function, the blood pressure estimation model is guided to perform phased meta-learning to minimize the loss value, wherein the meta-learning includes: a meta-training phase and a meta-testing phase.

[0087] 1) Meta-training phase

[0088] The purpose of the meta-training phase is to extract domain-invariant features from sensor (physiological) data from different source domains, so that the model has better generalization capabilities. In this phase, the training process is divided into two stages: the training sub-stage and the optimization sub-stage, aiming to enhance the generalization ability of the model in the target domain by simulating new domain scenarios.

[0089] During the training process, the meta-training phase is performed multiple times. Each time, at least one source domain is selected from the source domain set to form the meta-optimization domain, and the remaining source domains form the meta-training domain. Each execution of the meta-training phase includes a training sub-phase and an optimization sub-phase. The two phases are introduced below.

[0090] 1.1) Training sub-phase

[0091] The training sub-stage includes: training the blood pressure estimation model using the current meta-training domain and the total loss function;

[0092] In the training sub-stage, the source domain set is divided into "meta-training domain" and "meta-optimization domain". Schematically, in the meta-training stage, a preset number of unselected source domains are selected from the source domain set as meta-optimization domains each time. For example, in each round of iteration, a source domain is randomly selected as the meta-optimization domain (D_optimization), and the remaining source domains are used as meta-training domains (D_train). The purpose of this division method is to simulate the distribution characteristics of the new domain data, and gradually extract domain-invariant features that are robust to different individual data by training on the remaining source domains (i.e., meta-training domains). Among them, the training sub-stage is to select batch samples from the current meta-training domain in batches, and iteratively train the blood pressure estimation model using the batch samples selected from the meta-training domain until the preset requirements are met and then transfer to the optimization sub-stage.

[0093] In each batch-specific training sub-phase, the model reads data from the meta-training domain in mini-batches and performs the following operations:

[0094] Forward propagation: Input small batches of data into the model and obtain the prediction results through forward propagation;

[0095] Calculate training loss: Calculate the loss using the preset total loss function on the meta-training domain data. At this time, the specified domain can be set to the source domain in the meta-training domain. The last two terms in the total loss function are the mean of the alignment loss and the mean of the distribution difference between the domain-invariant features of each two source domains in the meta-training domain.

[0096] Back propagation and parameter update: Calculate the gradient based on the loss and update the model parameters through back propagation to further optimize the model on the current batch of data.

[0097] Through multiple rounds of iterations of the meta-training domain in the training sub-stage, the model gradually extracts invariant features from different source domains to ensure adaptability to different individual data. The continuous optimization of the training sub-stage makes the model domain invariant, laying the foundation for subsequent optimization in the meta-optimization domain.

[0098] 1.2) Optimization sub-stage

[0099] After the initial parameter update in the training sub-phase, the model is further fine-tuned on the meta-optimization domain to simulate the data distribution characteristics of the target domain, so that the model can effectively cope with the sensory data of new individuals.

[0100] The optimization sub-stage includes: using the current meta-training domain, meta-optimization domain and total loss function to train the blood pressure estimation model learned in the sub-stage, wherein the optimization sub-stage selects batch samples from the current meta-optimization domain in batches, and iteratively trains the blood pressure estimation model using the batch samples selected from the meta-optimization domain until the preset requirements are met and then transfers to the optimization sub-stage.

[0101] The specific process of the optimization sub-stage is as follows:

[0102] Read data batch by batch from the meta-training domain and meta-optimization domain: Input small batches of data from the meta-training domain and meta-optimization domain into the model for inference.

[0103] Calculate meta-optimization loss: Calculate the total loss function on the meta-training domain and the meta-optimization domain. The form of this loss function is consistent with that in the training sub-stage, but the specified domains of this stage are set to the meta-training domain and the meta-optimization domain. That is, the last two terms of the total loss function are the alignment loss and distribution difference between the domain-invariant features of the meta-training domain and the meta-optimization domain.

[0104] Parameter update: The gradient is calculated based on the total loss function, and the model parameters are further updated through back propagation, so that the model can achieve better adaptation effect in the meta-optimization domain.

[0105] After the parameter update in the optimization sub-stage, the performance of the model in the meta-optimization domain is further improved, thereby enhancing the domain invariance and cross-domain adaptability of the model. By alternating training and optimization between multiple source domains, each source domain participates in the meta-training and meta-optimization process, ensuring that the model can learn invariant features that are effective for target domain data from different source domain data.

[0106] The meta-training stage uses alternating training of the "training sub-stage" and the "optimization sub-stage" to enable the model to learn domain-invariant features on multiple source domain data, gradually improving the cross-domain adaptability and generalization ability of the model. In practical applications, the meta-training stage not only improves the model's adaptability to different individual data, but also significantly reduces its dependence on personalized data, making the method more flexible in the application of cuffless blood pressure estimation.

[0107] 2) Meta-testing phase

[0108] The meta-testing phase is conducted after the meta-training phase. Its purpose is to conduct a final evaluation and fine-tuning of the model on the entire source domain data to ensure that the model has the ability to generalize efficiently to new domain data. The meta-testing phase enables the model to adapt to more new individual sensor data in practical applications through verification and fine-tuning.

[0109] According to one embodiment of the present invention, performing learning in the meta-testing phase includes: training the blood pressure estimation model learned in the meta-training phase based on all source domains in the source domain set and the total loss function. The meta-testing phase is to select batch samples from all source domains in the source domain set, and fine-tune the training of the blood pressure estimation model using the batch samples selected from all source domains.

[0110] The schematic process of the meta-testing phase is as follows:

[0111] Verify the cross-domain performance of the model: After the meta-training phase, the model has been trained and optimized on data from each source domain. At this point, the model needs to be comprehensively validated on all source domains to evaluate the performance of the model on different source domains. This validation step ensures that the model can achieve high prediction accuracy under multiple data distributions.

[0112] Batch-by-batch validation and fine-tuning: In the meta-testing phase, the model reads data from the source domain collection in small batches for validation. Each small batch of data is input into the model, and the model output is calculated and evaluated on the current source domain data.

[0113] Calculate test loss: Use the preset total loss function to calculate the test loss on all source domain data. The specified domain in this stage is set to the source domain of the source domain set, that is, the last two terms of the total loss function are the mean of the alignment loss and the mean of the distribution difference between the domain invariant features of each two source domains in the source domain set. By calculating the loss on all source domain data, the overall performance of the model can be effectively evaluated.

[0114] Parameter fine-tuning: Based on the test loss, the model parameters are further optimized through back propagation. Since the test loss can comprehensively reflect the performance of the model in multiple source domains, the parameter update at this stage helps to further improve the adaptability of the model in the target domain.

[0115] Comprehensive evaluation of validation results: After the meta-test phase, the test results of all source domain data will be comprehensively evaluated to ensure that the model has good predictive ability for new individuals. The performance of the model on multiple source domain data determines its adaptability and generalization to new domain data in practical applications.

[0116] Through the final verification and fine-tuning in the meta-testing phase, the domain invariance of the model is further strengthened, and it has the ability to maintain stable prediction results on different individual data. The design of the meta-testing phase ensures that the features learned by the model from multiple source domains can be applied to the target domain, improving the prediction accuracy and robustness of the model on new individuals.

[0117] 5. Application Scenarios

[0118] According to one embodiment of the present invention, a blood pressure estimation method is provided, comprising: obtaining a trained blood pressure estimation model obtained according to the method described in the above embodiment; obtaining sensor data of an object to be evaluated, and estimating a blood pressure prediction value according to the sensor data of the object to be evaluated using the trained blood pressure estimation model. The trained blood pressure estimation model is a model finally obtained after training by the blood pressure estimation training method described in the above embodiment.

[0119] In order to verify the effect of the method of the present invention, the inventors also conducted a comparative experiment.

[0120] The existing methods compared in the experiment are as follows:

[0121] [1] Vapnik V N, Vapnik V. Statistical learning theory[J]. 1998.

[0122] [2] Ganin Y, Ustinova E,Ajakan H, et al. Domain-adversarial trainingof neural networks[J]. Journal of machine learning research, 2016, 17(59): 1-35.

[0123] [3] Arjovsky M, Bottou L, Gulrajani I, et al. Invariant riskminimization[J]. arXiv preprint arXiv:1907.02893, 2019.

[0124] [4] Zhang M, Marklund H, Gupta A, et al. Adaptive risk minimization:A meta-learning approach for tackling group shift[J]. arXiv preprint arXiv:2007.02931, 2020, 8(9).

[0125] [5] Peters J, Bühlmann P, Meinshausen N. Causal inference by usinginvariant prediction: identification and confidence intervals[J]. Journal ofthe Royal Statistical Society Series B: Statistical Methodology, 2016, 78(5):947-1012.

[0126] [6] Sagawa S, Koh PW, Hashimoto TB, et al. Distributionally robust neural networks for shifts group: On the importance of regularization for worst-case generalization[J]. arXiv preprint arXiv:1911.08731, 2019.

[0127] [7] Li D, Yang Y, Song YZ, et al. Learning to generalize: Meta-learning for domain generalization[C] / / Proceedings of the AAAI conference on artificial intelligence. 2018, 32(1).

[0128] [8] Leitner J, Chiang PH, Dey S. Personalized blood pressure estimation using photoplethysmography: A transfer learning approach[J]. IEEEJournal of Biomedical and Health Informatics, 2021, 26(1): 218-228.

[0129] [9] Huang B, Chen W, Lin CL, et al. MLP-BP: A novel framework for cuffless blood pressure measurement with PPG and ECG signals based on MLP-Mixer neural networks[J]. Biomedical Signal Processing and Control, 2022, 73:103404.

[0130] The comparison indicators of the experiment are as follows:

[0131] To evaluate the performance, the mean absolute error (MAE) and standard deviation (STD) were calculated. The mean absolute error is calculated as follows:

[0132]

[0133] in, represents the predicted blood pressure value, Indicates the true blood pressure value. Indicates the number of samples.

[0134] The standard deviation is defined as follows:

[0135]

[0136] The meaning of each letter is the same as above.

[0137] Comparison results:

[0138] Table 1 Comparison of models under different data sets

[0139]

[0140] Table 1 compares the blood pressure estimation results of different models on three datasets, including the mean absolute error (MAE) and standard deviation (STD) of systolic blood pressure (SBP) and diastolic blood pressure (DBP) under three experimental settings: MIMIC III→MIMICIII, MIMIC III→MIMIC IV, and MIMIC III→UQVS. The figure shows the cross-domain generalization ability of each model in the target domain dataset (MIMIC III, MIMIC IV, and UQVS) after training on the source domain dataset (MIMIC III). By comparing the MAE and STD, it can be seen that MetaDIBP performs best in all tasks, especially in the cross-domain generalization tasks (MIMICIII→MIMIC IV and MIMIC III→UQVS), where its MAE and STD are better than other methods. Specifically, in the source domain task of MIMICIII→MIMIC III, MetaDIBP has the best performance with MAEs of 12.989 mmHg and 7.391 mmHg for SBP and DBP, respectively. In the cross-domain task MIMIC III→MIMIC IV, MetaDIBP also leads other methods with the lowest MAE of 14.297 mmHg (SBP) and 8.458 mmHg (DBP). In the cross-domain task MIMIC III→UQVS, MetaDIBP further demonstrated its stable cross-domain generalization ability, with MAE of 13.669 mmHg for SBP and 11.392 mmHg for DBP. Compared with traditional methods (such as ERM and IRM), MetaDIBP performs significantly better in dealing with cross-domain data distribution differences, thanks to its two-stage meta-learning strategy and the ability to align feature distributions through CORAL loss and MMD loss. Although other models (such as DANN and GroupDRO) have a slight advantage in some indicators (such as STD), overall, MetaDIBP has the best generalization ability and stability in all tasks, especially in the prediction accuracy of unseen target domains. This shows that MetaDIBP can better handle complex physiological signal data and perform well in cross-individual blood pressure estimation tasks.

[0141] Table 2. Mean absolute error of personalized models pre-trained using different domain generalization methods

[0142]

[0143] Table 2 shows the personalization results, where the models are fine-tuned with different amounts of individual-specific data (5, 10, and 25 samples) after being pre-trained using different domain generalization methods. Since each sample point contains 5 seconds of data and the sampling rate is 125 Hz, each sample contains 625 data points. 5, 10, and 25 sample points were selected to test the adaptability of the model under different data size conditions. In the MIMIC III → MIMIC III setting, the proposed MetaDIBP method achieves the lowest mean absolute error (MAE) for systolic blood pressure (SBP) and diastolic blood pressure (DBP) on all sample sizes. Similarly, in the MIMIC III → MIMIC IV and MIMIC III → UQVS scenarios, MetaDIBP outperforms other methods, especially when fine-tuning with fewer samples (5 and 10 sample points). This consistent performance demonstrates the effectiveness of MetaDIBP in adapting to new individuals when using limited personalized data.

[0144] The superior performance of MetaDIBP in the personalized fine-tuning scenario highlights its strong generalization ability. Traditional domain generalization methods usually require a large amount of individual-specific data to achieve similar accuracy levels. However, our method shows strong adaptability even with very little personalized data by learning universal and invariant features in two stages. This advantage is particularly important in practical applications because collecting large amounts of individual-specific data is often not feasible. The consistent improvement of MAE across different datasets and sample sizes further highlights the potential of MetaDIBP for widespread application in non-invasive blood pressure estimation.

[0145] In general, the two-stage meta-learning strategy proposed in the present invention includes a meta-training stage and a meta-testing stage, which respectively enhance the cross-domain adaptability and generalization of the model by simulating the distribution characteristics of new domain data and verifying and fine-tuning on all source domains. Among them, the meta-training stage extracts domain-invariant features through alternating training sub-stages and optimization sub-stages, so that the model can adapt to the signal characteristics of different individuals and significantly reduce the dependence on personalized data; the meta-testing stage verifies and fine-tunes on all source domain data to ensure that the model can maintain efficient generalization capabilities in practical applications. The method of the present invention effectively solves the problems of poor cross-domain adaptability and high demand for personalized data in cuffless blood pressure estimation, so that the estimation accuracy and stability of the model on new individuals are significantly improved. This strategy makes cuffless blood pressure estimation technology more suitable for practical application scenarios and provides a solid technical foundation for the promotion of health monitoring equipment.

[0146] It should be noted that although the above describes the various steps in a specific order, it does not mean that the various 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 functions can be achieved.

[0147] The present invention may be a system, a method and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for causing a processor to implement various aspects of the present invention.

[0148] A computer-readable storage medium may be a tangible device that holds and stores instructions used by an instruction execution device. Computer-readable storage media may include, for example, but are not limited to, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (non-exhaustive list) of computer-readable storage media include: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a static random access memory (SRAM), a portable compact disk read-only memory (CD-ROM), a digital versatile disk (DVD), a memory stick, a floppy disk, a mechanical encoding device, such as a punch card or a protruding structure in a groove on which instructions are stored, and any suitable combination thereof.

[0149] The embodiments of the present invention have been described above, and the above description is exemplary, not exhaustive, and is not limited to the disclosed embodiments. Many modifications and variations will be apparent to those of ordinary skill in the art without departing from the scope and spirit of the described embodiments. The selection of terms used herein is intended to best explain the principles of the embodiments, practical applications, or technical improvements in the market, or to enable other persons of ordinary skill in the art to understand the embodiments disclosed herein.

Claims

1. A blood pressure estimation training method, characterized in that: include: Obtaining a blood pressure estimation model, which is used to extract signal features from input sensor data and estimate a blood pressure prediction value based on the signal features; Acquire a source domain set including a plurality of source domains, each source domain including a plurality of samples of a single object, each sample including sensor data collected from the object and a label indicating a blood pressure value of the object when the sensor data is collected, the sensor data including a PPG signal and an ECG signal; Using the source domain set and the preset total loss function, the blood pressure estimation model is guided to perform staged meta-learning to minimize the loss value, where the meta-learning includes: The learning of the meta-training stage is performed multiple times, and each time at least one source domain is selected from the source domain set to form a meta-optimization domain, and the remaining source domains form a meta-training domain. The learning of each meta-training stage includes: The training sub-stage includes: training a blood pressure estimation model using the current meta-training domain and the total loss function; The optimization sub-stage includes: using the current meta-training domain, meta-optimization domain and total loss function to train the blood pressure estimation model learned in the sub-stage; Performing learning in the meta-testing phase includes: training the blood pressure estimation model learned in the meta-training phase based on all source domains in the source domain set and the total loss function.

2. The method according to claim 1, characterized in that: In the meta-training phase, a preset number of unselected source domains are selected from the source domain set as meta-optimization domains each time, where: The training sub-stage is to select batch samples from the current meta-training domain in batches, and iteratively train the blood pressure estimation model using the batch samples selected from the meta-training domain until the preset requirements are met and then transfer to the optimization sub-stage.

3. The method according to claim 2, characterized in that The optimization sub-stage is to select batch samples from the current meta-training domain and meta-optimization domain in batches, and iteratively train the blood pressure estimation model using the batch samples selected from the meta-training domain and meta-optimization domain until the preset requirements are met and then transfer to the optimization sub-stage.

4. The method according to claim 3, characterized in that The meta-testing phase is performed after the meta-training phase. In the meta-testing phase, batch samples are selected from all source domains in the source domain set, and the blood pressure estimation model is fine-tuned using the batch samples selected from all source domains.

5. The method according to any one of claims 1 to 4, characterized in that: When training the blood pressure estimation model, a preset training component is used, and the training component includes: A blood pressure estimation model, comprising: a feature extraction module for extracting signal features from input sensor data, and a prediction module for determining a blood pressure prediction value based on the signal features; A domain invariant feature extraction module, used to extract domain invariant features from signal features; During training, the regression loss of blood pressure and the alignment loss and distribution difference between the domain-invariant features of the specified domain are used to guide the parameter update of the blood pressure estimation model; in the training sub-stage, the alignment loss and distribution difference between the domain-invariant features of each source domain in the meta-training domain are calculated; in the optimization sub-stage, the alignment loss and distribution difference between the domain-invariant features of the meta-training domain and the meta-optimization domain are calculated; in the meta-testing stage, the alignment loss and distribution difference between the domain-invariant features of each source domain in the source domain set are calculated.

6. The method according to claim 5, characterized in that The feature extraction module comprises: A plurality of processing branches, each processing branch is used to extract a modal feature from a modal signal in the sensor data, wherein each processing branch includes a plurality of stacked time residual blocks and gated recurrent units, The fusion module is used to fuse the modal features extracted by multiple processing branches to obtain signal features.

7. The method according to claim 5, characterized in that The total loss function is: in, represents the sub-function used to calculate the blood pressure prediction value and the regression loss between labels, represents a sub-function for calculating the domain-invariant feature alignment loss for a specified domain, represents a subfunction for computing the distribution difference between domain-invariant features of a specified domain, express The weight of express The weight of express The weight of .

8. A method for estimating blood pressure, comprising: Obtaining a trained blood pressure estimation model obtained according to the method of any one of claims 1-7; The sensor data of the subject to be evaluated is obtained, and the blood pressure prediction value is estimated according to the sensor data of the subject to be evaluated using the trained blood pressure estimation model.

9. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and the computer program 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; as well as A memory, wherein the memory is used to store executable instructions; The one or more processors are configured to implement the steps of the method of any one of claims 1-8 by executing the executable instructions.

Images