Small-sample non-invasive continuous blood pressure monitoring method based on ballistocardiogram signals

Abstract

The present invention discloses a small sample non-invasive continuous blood pressure monitoring method based on ballistocardiogram signals. The method selects the ballistocardiogram signals as input data to construct a blood pressure prediction model based on the strong physiological correlation characteristics between the ballistocardiogram signals and blood pressure; constructs multiple small sample tasks, each task corresponds to a support set and a query set; combines the meta-learning framework with the blood pressure prediction model, and uses the support set and query set corresponding to the small task for training; uses labeled samples of new subjects to fine-tune the trained blood pressure prediction model, and then inputs unknown ballistocardiogram signal data to achieve continuous monitoring of individual blood pressure. The present invention is the first to create a small sample blood pressure regression method based on ballistocardiogram signals, breaking through the traditional model's neglect of individual vascular system differences, overcoming the bottleneck of the lack of ballistocardiogram signal samples with blood pressure monitoring, and providing a new paradigm for personalized blood pressure monitoring; it can achieve higher blood pressure prediction performance with a small amount of data.

Description

Technical Field

[0001] The present invention relates to the technical field of biomedical signal processing, and in particular to a small-sample non-invasive continuous blood pressure monitoring method based on ballistocardiogram signals. Background Art

[0002] Hypertension is one of the most prevalent chronic diseases worldwide. Reports indicate that over 40% of adults are likely to be affected by hypertension, meaning its prevalence is extremely high. Without proper treatment, hypertension can lead to serious complications such as stroke, heart disease, and kidney failure. Therefore, early blood pressure monitoring to alleviate problems and avoid further complications is crucial.

[0003] Blood pressure monitoring methods can be divided into three categories: direct blood pressure monitoring method, intermittent blood pressure monitoring method, and continuous blood pressure monitoring method. Among them, the measurement result of the direct blood pressure monitoring method is considered the true value of blood pressure. However, due to its invasive monitoring drawbacks and high technical requirements, it is only applicable to patients undergoing major surgeries or intensive care. The main drawback of the intermittent blood pressure monitoring method is its insufficient ability to track and reflect instantaneous and continuous blood pressure changes. In contrast, the continuous blood pressure monitoring method can monitor continuous blood pressure changes in daily environments, and the measurement results are closer to the real human situation, with higher physiological validity and monitoring accuracy.

[0004] Ballistocardiogram (BCG) signals contain rich physiological information, including the rhythm, intensity of cardiac activity, and cardiac pumping function, etc., which have potential value for the identification of hypertension. BCG measurement technology is a non-invasive technology that analyzes the minute vibrations of the human body caused by heartbeats and creates corresponding waveforms, with advantages such as simple operation, stable performance, and strong adaptability. Using BCG for continuous blood pressure monitoring overcomes the drawbacks of the mainstream methods that require deploying multiple sensors or are inconvenient to deploy, thereby further improving the continuous blood pressure monitoring effect.

[0005] In classical regression methods for biomedical signal analysis, a large amount of instance-labeled data is usually required to train the model, which often faces many challenges in non-invasive blood pressure monitoring based on BCG signals. First, the long-tailed distribution characteristics of the real world lead to the scarcity of data for some special groups, such as rare pathological abnormal blood pressure data. Second, the collection of synchronized measurement data of blood pressure and BCG signals requires professional equipment and manpower, and it is also difficult to ensure the accuracy of annotation. In addition, model training requires a large amount of computing resources and professional knowledge, for example, it is necessary to overcome the domain shift phenomenon that will affect the model's effect in new scenarios. Therefore, how to achieve accurate blood pressure prediction with a small number of target samples has become a research problem of great practical significance. Summary of the Invention

[0006] The object of the present invention is to provide a small-sample non-invasive continuous blood pressure monitoring method based on ballistocardiogram signals in view of the deficiencies of the prior art. Considering the differences in the vascular system and anatomy among individuals and the interference of slow and subtle cardiac movements, the initial parameters are trained by meta-learning, enabling the model to optimize to a better performance after a small number of updates when facing new data, thereby modeling the vascular system and achieving continuous monitoring of individual blood pressure.

[0007] The object of the present invention is achieved through the following technical solutions: A small-sample non-invasive continuous blood pressure monitoring method based on ballistocardiogram signals, comprising the following steps:

[0008] S1. Collect ballistocardiogram signals and corresponding blood pressure signals, and slice them according to a preset time window to obtain a plurality of ballistocardiogram slice signals and a plurality of blood pressure slice signals; construct a ballistocardiogram signal data set according to the ballistocardiogram slice signals and their corresponding blood pressure slice signals;

[0009] S2. Preprocess the ballistocardiogram slice signals and their corresponding blood pressure slice signals in the ballistocardiogram signal data set to construct a training set;

[0010] S3. Based on the requirements of small-sample learning, select sample data from the training set to construct a plurality of small-sample tasks, each task consisting of a support set and a query set, to complete the construction of task sample pairs;

[0011] S4. Construct a blood pressure prediction model based on a deep learning network using the gradient descent algorithm, with the preprocessed ballistocardiogram slice signals as the input and the predicted blood pressure values as the output;

[0012] S5. Combine the meta-learning framework with the blood pressure prediction model, and train the blood pressure prediction model through the meta-learning framework using the support sets and query sets corresponding to a plurality of small-sample tasks to obtain a trained blood pressure prediction model;

[0013] S6. After fine-tuning the trained blood pressure prediction model using the labeled samples of a new subject, input the unknown ballistocardiogram signal data of this subject to achieve continuous monitoring of individual blood pressure.

[0014] Further, in the step S2, the preprocessing of the ballistocardiogram slice signals specifically includes:

[0015] Taking different subjects as processing units, perform downsampling processing on the multiple ballistocardiogram slice signals of different subjects by using sliding window downsampling;

[0016] Taking different subjects as processing units, for each post-downsampling processed ballistocardiogram slice signal, perform Z-score normalization processing to obtain the normalized ballistocardiogram slice signal.

[0017] Further, in step S3, the construction of the task sample pair is specifically implemented by the following method:

[0018] First, select sample data from the training set to construct multiple small sample tasks. Each task consists of a support set and a query set corresponding to that task. Among them, 70%-90% of the training samples in the support set corresponding to each task and all the test samples in the query set come from the sample data of the same subject in the training set, and the remaining training samples in the support set come from the sample data of other different subjects in the training set. The sample data includes the preprocessed ballistocardiogram slice signal and blood pressure slice signal, and the systolic or diastolic blood pressure of the blood pressure slice signal corresponding to the preprocessed ballistocardiogram slice signal is used as the blood pressure label.

[0019] Then, for the sample data from different subjects in the support set, artificially add perturbations to obtain the final task sample pair.

[0020] Further, the blood pressure prediction model uses ResNet-18. The blood pressure prediction model includes an input module, multiple residual modules, and an output module. Among them, the input module includes a convolutional layer and a max pooling layer. Each residual module includes two residual blocks, and each residual block includes two convolutional layers. The output module includes a global average pooling layer, a fully connected layer, and a linear layer.

[0021] The preprocessed ballistocardiogram slice signal is input into the blood pressure prediction model. First, it passes through the convolutional layer and max pooling layer of the input module to obtain an initial feature representation. Then, the initial feature representation successively passes through multiple residual modules to extract different levels of features of the ballistocardiogram slice signal. In multiple sequentially connected residual modules, the output of the previous residual block is added to the input of the previous residual block as the input of the current residual block through a skip connection mechanism. The input of the first residual block of the first residual module is the initial feature representation output by the input module. The output of the last residual block of the last residual module is added to the input of this residual block to obtain the final feature vector. The final feature vector is used as the input of the output module. After passing through the global average pooling layer and the fully connected layer, the final feature vector is reduced to a fixed-length embedding vector, and the predicted blood pressure value is obtained through the linear layer.

[0022] Further, step S5 specifically includes the following sub-steps:

[0023] S51. Initialize model parameters: Randomly initialize the parameters of the blood pressure prediction model and use them as the initial state of the meta-model. Consider the blood pressure prediction model as the meta-model.

[0024] S52. Define the loss function: Calculate the mean squared error between the predicted blood pressure value and the true blood pressure value as the loss function.

[0025] S53. Task training and update: For each small sample task, clone the meta-model for the inner-loop training of the meta-learning framework. Perform several steps of gradient descent update on the support set of each task to obtain the model parameters corresponding to that task. Each step of gradient descent update is carried out with the optimization goal of minimizing the calculated loss function.

[0026] S54. Meta-update: During the outer-loop training of the meta-learning framework, calculate the loss function on the query set corresponding to each task based on the meta-model obtained in the inner loop for that task. Accumulate the loss functions of all tasks to obtain the total loss. Using the Adam optimizer, optimize the model parameters of the meta-model with the optimization goal of minimizing the total loss through backpropagation.

[0027] S55. Train the meta-model in the outer loop: Repeat steps S53 - S54, update the parameters of the meta-model using the loss function passed out from the inner loop to obtain the trained meta-model, which is the trained blood pressure prediction model.

[0028] Furthermore, step S6 specifically includes the following sub-steps:

[0029] S61. Freeze the parameters of the first N residual modules of the blood pressure prediction model, where the value of N is less than or equal to half of the total number of residual modules.

[0030] S62. Fine-tune the trained blood pressure prediction model using the pre-processed labeled partial sample data of the new subject to obtain a fine-tuned blood pressure prediction model suitable for this subject.

[0031] S63. Input the unknown ballistocardiogram signal data of the subject into the fine-tuned blood pressure prediction model suitable for this subject to obtain the predicted blood pressure value, realizing continuous monitoring of individual blood pressure.

[0032] The beneficial effects of the present invention are as follows: The present invention simultaneously considers the differences in the vascular system and anatomy between individuals, as well as the slow and subtle interference of cardiac motion, and also takes into account the signal characteristics of the BCG signal itself, enabling the finally learned blood pressure prediction model to have the ability to analyze the BCG signal, with high robustness, capable of performing personalized modeling after inputting new samples, and improving the accuracy of blood pressure prediction. Description of the Drawings

[0033] Figure 1Flow chart of the small-sample non-invasive continuous blood pressure monitoring method based on BCG signals according to the present invention;

[0034] Figure 2 Schematic diagram of the network structure of the blood pressure prediction model according to the present invention;

[0035] Figure 3 Example diagram of the fitting effect of the model output for one task according to the present invention. Detailed implementation manners

[0036] Here, exemplary embodiments will be described in detail, and the examples are shown in the drawings. When the following description refers to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. On the contrary, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0037] The terms used in the present invention are for the purpose of describing specific embodiments only and are not intended to limit the present invention. The singular forms "a", "the", and "said" used in the present invention and the appended claims are also intended to include the plural forms unless the context clearly dictates otherwise. It should also be understood that the term "and / or" used herein refers to and encompasses any and all possible combinations of one or more of the associated listed items.

[0038] It should be understood that although the terms first, second, third, etc. may be used in the present invention to describe various information, such information should not be limited to these terms. These terms are only used to distinguish the same type of information from each other. For example, without departing from the scope of the present invention, the first information may also be referred to as the second information, and similarly, the second information may also be referred to as the first information. Depending on the context, the word "if" as used herein may be interpreted as "when" or "while" or "in response to determining".

[0039] The present invention will be described in detail below with reference to the drawings. Without conflict, the features in the following embodiments and implementation manners can be combined with each other.

[0040] See Figure 1 , the small-sample non-invasive continuous blood pressure monitoring method based on ballistocardiography (BCG) signals of the present invention specifically includes the following steps:

[0041] S1. Collect the BCG signal and the corresponding blood pressure signal, and slice them according to a preset time window to obtain multiple BCG slice signals and multiple blood pressure slice signals; construct a BCG signal dataset based on the BCG slice signals and their corresponding blood pressure slice signals.

[0042] Specifically, collect the BCG signal and the corresponding blood pressure signal, and slice the BCG signal according to a preset time window. For example, slice the BCG signal and its corresponding blood pressure signal according to a time window of 10 s or 30 s as the data unit for subsequent processing. After slicing, multiple BCG slice signals and multiple blood pressure slice signals are obtained; finally, construct a BCG signal dataset required for training based on the multiple BCG slice signals and their corresponding blood pressure slice signals. In this BCG signal dataset, data of multiple subjects are included, and each subject contains its corresponding multiple BCG slice signals and their corresponding multiple blood pressure slice signals.

[0043] S2. Preprocess the BCG slice signals and their corresponding blood pressure slice signals in the BCG signal dataset to construct a training set, that is, construct a training set based on the preprocessed BCG slice signals and blood pressure slice signals.

[0044] Further, in step S2, preprocess the BCG slice signals, which specifically includes the following steps:

[0045] S21. Take different subjects as processing units, and perform downsampling processing on the multiple BCG slice signals of different subjects by using sliding window downsampling to achieve downsampling processing of each BCG slice signal of each subject.

[0046] Specifically, take different subjects as processing units, and perform downsampling processing on the BCG slice signals of different subjects, which is specifically achieved by sliding window downsampling. Specifically, set the size and stride of the sliding window to 10, and take the average value of the sampling points in the same time window as the result after downsampling processing of this time window. Finally, obtain the BCG slice signals after downsampling processing. Through sliding window downsampling, the original data volume can be compressed to one-tenth of the original, realizing data downsampling and increasing the calculation efficiency of the model. For example, for each BCG slice signal, there are originally 100 sampling points. Since the size and stride of the sliding window are set to 10, for the first time window, that is, the data of the first 10 sampling points are averaged as the result after downsampling processing corresponding to the first time window. Similarly, according to the set stride, and so on. Finally, 10 results corresponding to 10 time windows can be obtained, that is, after downsampling, a result of 10 sampling points is obtained, and the original data volume is compressed to one-tenth of the original, realizing downsampling of the entire segment of data.

[0047] It should be noted that the size and stride of the sliding window are determined according to the required amount of data. For example, if there is a large amount of original data and it is required to compress it to one-tenth of the original amount of data, then the size and stride of the sliding window are set to 10; if it is required to compress the amount of data to one-twentieth of the original amount of data, then the size and stride of the sliding window are set to 20.

[0048] S22. Taking different subjects as processing units, for each downsampled BCG slice signal, perform Z-score normalization processing to obtain the normalized BCG slice signal, which is the preprocessed BCG slice signal.

[0049] It should be noted that considering the differences in the vascular system and anatomy among individuals and the interference of slow and subtle cardiac movements, taking different subjects as processing units, perform Z-score normalization on the data of the BCG slice signal of each subject to eliminate the signal amplitude differences caused by different subjects.

[0050] Furthermore, the Z-score normalization is implemented according to the following formula:

[0051]

[0052] In the formula, represents the data value of the normalized BCG slice signal, which is used to represent the deviation degree of the data point from the mean value, and its unit is the standard deviation; represents the original data to be normalized, that is, the data of the downsampled BCG slice signal; represents the mean value (mean) of the BCG signal dataset, that is, the average value of all data points of all downsampled BCG slice signals in the BCG signal dataset; represents the standard deviation (standard deviation) of the BCG signal dataset, that is, the standard deviation of all data points of all downsampled BCG slice signals in the BCG signal dataset, which is used to represent the degree of dispersion or fluctuation range of the data.

[0053] It should be understood that the Z-score normalization method is a commonly used method for data processing. Through it, data of different magnitudes can be transformed into data with a unified measurement, making the data standard unified and improving the comparability of the data.

[0054] Furthermore, in step S2, preprocess the blood pressure slice signal, which specifically includes: taking different subjects as processing units, for each blood pressure slice signal, select the maximum value of the corresponding time window of the blood pressure slice signal as the systolic blood pressure of the blood pressure slice signal, and the minimum value as the diastolic blood pressure of the blood pressure slice signal, which is the preprocessed blood pressure slice signal.

[0055] S3. Based on the specific requirements of few-shot learning, select sample data from the training set to construct multiple few-shot tasks, each task consisting of a support set and a query set, to complete the construction of task sample pairs. Here, a task sample pair refers to the support set and the query set corresponding to the task.

[0056] Furthermore, the construction of task sample pairs is specifically achieved through the following method: First, select sample data from the training set to construct multiple few-shot tasks, each task consisting of a support set and a query set corresponding to that task; among them, 70%-90% of the training samples in the support set corresponding to each task and all the test samples in the query set come from the sample data of the same subject in the training set, and the remaining training samples in the support set come from the sample data of other different subjects in the training set; the sample data includes preprocessed BCG slice signals and blood pressure slice signals, and the systolic or diastolic blood pressure of the blood pressure slice signal corresponding to the preprocessed BCG slice signal is used as the blood pressure label for subsequent model training. To ensure the training effect of the model, only one of the systolic or diastolic blood pressure is selected as the blood pressure label during the training process. Then, for the sample data in the support set from different subjects, add an appropriate amount of artificial perturbation to obtain the final task sample pair. The purpose is to introduce the information of different subjects during the subsequent model training process, hoping that the model can find commonalities; by adding an appropriate amount of artificial perturbation, the model can be made more robust and stable to the possible minor fluctuations when learning the blood pressure pattern of a specific patient. This data perturbation method helps the model better handle signal deviations caused by measurement conditions or patient state changes in practical applications, thereby improving prediction accuracy and practicality.

[0057] Exemplarily, for the support set and the query set corresponding to the same task, the support set and the query set respectively contain 10 training samples and 10 test samples. Among them, the first 8 training samples in the support set corresponding to the same task and the 10 test samples in the query set are both selected from the sample data of the same subject in the training set to evaluate the adaptability of the model to this specific patient; the last 2 training samples in the support set are randomly selected from the sample data of different subjects in the training set to introduce data diversity. It should be noted that the number of samples in the support set and the query set can be equal or unequal. Generally speaking, the number of training samples in the support set is less, and the number of test samples in the query set is more.

[0058] S4. Construct a blood pressure prediction model based on a deep learning network using the gradient descent algorithm, with the input being the preprocessed BCG slice signal and the output being the predicted blood pressure value.

[0059] In the present invention, a blood pressure prediction model can be constructed based on any deep learning network that adopts the gradient descent algorithm. In this embodiment, ResNet-18 is selected to construct the blood pressure prediction model, and its structure is as Figure 2 shown. The blood pressure prediction model includes an input module, multiple residual modules, and an output module. Among them, the input module includes a convolutional layer and a max pooling layer; each residual module includes two residual blocks, and each residual block includes two convolutional layers; the output module includes a global average pooling layer, a fully connected layer, and a linear layer. Specifically, the preprocessed BCG slice signal is input into the blood pressure prediction model. First, it passes through the convolutional layer and the max pooling layer of the input module to obtain an initial feature representation; then the initial feature representation successively passes through multiple residual modules to extract different levels of features of the BCG slice signal layer by layer. In multiple successively connected residual modules, the output of the previous residual block is added to the input of the previous residual block through a skip connection mechanism as the input of the current residual block. The input of the first residual block of the first residual module is the initial feature representation output by the input module. The output of the last residual block of the last residual module is added to the input of this residual block to obtain the final feature vector; the final feature vector is used as the input of the output module, and the final feature vector is reduced to a fixed-length embedding vector through the global average pooling layer and the fully connected layer. The fixed-length embedding vector is used to obtain the predicted blood pressure value through the linear layer.

[0060] Specifically, the blood pressure prediction model has strong temporal and frequency domain feature capture capabilities and is used to process the preprocessed BCG slice signal, that is, the preprocessed BCG slice signal is input into the blood pressure prediction model for blood pressure regression to obtain the predicted blood pressure value. In Figure 2In the shown blood pressure prediction model, it includes an input module, four residual modules, and an output module. The input module includes a convolutional layer and a max pooling layer. Each residual module includes two residual blocks, and each residual block includes two convolutional layers. The output module includes a global average pooling layer, a fully connected layer, and a linear layer. The preprocessed BCG slice signal is input into the blood pressure prediction model. First, it enters the input module. The input signal passes through a 7×7 convolutional layer and a 3×3 max pooling layer to obtain an initial feature representation. The number of convolutional kernels in the convolutional layer is 64, the stride is 2, and the padding is 3. The stride of the max pooling layer is 2. Then, the initial feature representation successively passes through four residual modules to gradually extract high-level and low-level features of the BCG slice signal layer by layer. In the first residual module, the initial feature representation enters a residual block containing two 3×3 convolutional layers. The number of convolutional kernels in each convolutional layer is 64. Through the skip connection mechanism, the output of the first residual block is added to the input of the first residual block and used as the input of the second residual block. Through the skip connection mechanism, the input of the second residual block and the output of the second residual block are added and used as the input of the next residual block, which is the output of the first residual module, enabling the blood pressure prediction model to gradually learn and extract low-level features in the BCG slice signal. In the second residual module, the output features of the first residual module are input into a residual block containing two 3×3 convolutional layers, and the number of convolutional kernels in each convolutional layer is extended to 128. In this residual module, deeper features are gradually extracted through feature weighting and fusion. In the third and fourth residual modules, the number of convolutional kernels in the convolutional layers is extended to 256 and 512 respectively, gradually enhancing the expression ability of the blood pressure prediction model for the global features of the BCG slice signal. It should be noted that the skip connection mechanism is used in each residual module to gradually extract features at different levels. Finally, the output of the last residual block of the last residual module is added to the input of this residual block to obtain the final feature vector. The final feature vector is reduced to a fixed-length embedding vector with a length of 512 through the global average pooling layer and the fully connected layer, which can retain rich features. The fixed-length embedding vector is mapped to a single dimension through the linear layer to obtain the predicted blood pressure value for blood pressure regression prediction.

[0061] S5. Combine the meta-learning (MAML) framework with the blood pressure prediction model. Train the blood pressure prediction model using the support sets and query sets corresponding to multiple few-shot tasks through the meta-learning framework to improve the fast adaptation ability of the blood pressure prediction model under few-shot scenarios and obtain a trained blood pressure prediction model. It specifically includes the following sub-steps:

[0062] S51. Initialize model parameters: Randomly initialize the parameters of the blood pressure prediction model and use them as the initial state of the meta-model. Consider the blood pressure prediction model as the meta-model.

[0063] S52. Define the loss function: Calculate the mean squared error (MSE) between the predicted blood pressure value and the true blood pressure value as the loss function. Its calculation formula is expressed as:

[0064]

[0065] In the formula, represents the loss function, represents the predicted blood pressure value corresponding to the i-th sample, represents the true blood pressure value (i.e., blood pressure label) of the i-th sample, and N represents the total number of samples.

[0066] S53. Task training and update: For each small-sample task, clone the meta-model for the inner-loop training of the meta-learning framework. Perform several steps of gradient descent update on the support set of each task to obtain the model parameters corresponding to that task. These model parameters are task-specific model parameters, and each step of gradient descent update is carried out with the goal of minimizing the calculated loss function. In this way, through rapid fine-tuning of a small number of samples for each task, it can adapt to different data distributions.

[0067] It should be noted that the meta-learning framework is a training framework with a double-loop structure, including an inner loop and an outer loop. First, divide the task samples into pairs, which has been completed through step S3, making the application scenario of the blood pressure prediction model a small-sample scenario. Since different tasks simulate actual application scenarios, each time it is applied, only a limited support set is used for fine-tuning. Secondly, it is a double-loop structure. The inner loop calculates the loss after fine-tuning each task separately and passes out the loss. The outer loop adds up the losses of each task in the inner loop to fine-tune the model.

[0068] Specifically, during the inner-loop training process of the meta-learning framework, traverse all small-sample tasks. For each small-sample task, it is necessary to clone the meta-model once and perform several steps of gradient descent update on the support set of that task, that is, divide the training samples in the support set of each task into several batches for training and update. Calculate the loss function during the training process, and use the goal of minimizing the calculated loss function value for each step of gradient descent update. The number of loops in the inner loop is very small, generally 3 to 5 rounds. The purpose is to simulate the fine-tuning scenario during application, so rapid fine-tuning is carried out on a small number of samples for each task. After the update is completed, calculate the loss for the query set and record the loss. It should be noted that the gradient descent update in the inner loop does not update the outer meta-model, which is also the reason for cloning. In order to achieve isolation, the only impact of the inner loop is to record the loss.

[0069] Therefore, the inner loop means traversing all tasks. For each task, the cloned meta-model in the outer loop is used to perform several steps of loop update on the support set, and then the loss of the query set of each task is passed out. Finally, meta-update is performed in the outer loop based on the losses of each task (as described in step S54 below). The updated meta-model obtained finally can be applied to different subjects, that is, it can be applied to different task scenarios. The meta-model cloned in the inner loop does not affect the meta-model in the outer loop. The fine-tuning of the model parameters of the meta-model during the inner loop training process is to obtain the loss of the meta-model on the corresponding query set of each task, so as to update the model parameters of the meta-model in the outer loop.

[0070] S54. Meta-update: During the outer loop training process of the meta-learning framework, calculate the loss function on the query set corresponding to each task according to the meta-model obtained in the inner loop for each task. Accumulate the loss functions of all tasks to obtain the total loss. Taking minimizing the total loss as the optimization goal, use the Adam optimizer to optimize the model parameters of the meta-model through backpropagation, which can help improve the generalization ability of the meta-model.

[0071] S55. Train the meta-model in the outer loop: Repeat steps S53 - S54, and use the loss function passed out from the inner loop to update the parameters of the meta-model to obtain a trained meta-model, which is the trained blood pressure prediction model. This trained blood pressure prediction model can fully adapt to the task after a small number of samples are quickly fine-tuned when facing different tasks.

[0072] It should be understood that during the inner loop process, the meta-model is copied, which does not affect the meta-model in the outer loop. Therefore, during the outer loop process, it is necessary to update the model parameters of the meta-model according to the losses of each task in the inner loop. Specifically, refer to the meta-update described in step S54. After repeating steps S53 - S54 and updating the meta-model through the operation of such a double-loop structure, the value of the inner loop loss function obtained in the next loop becomes smaller, indicating that the meta-model has improved its fast adaptation ability for small-sample tasks.

[0073] S6. After fine-tuning the trained blood pressure prediction model with the labeled samples of the new subject, input the unknown BCG signal data of this subject to realize continuous monitoring of individual blood pressure.

[0074] S61. Freeze the parameters of the first N residual modules of the blood pressure prediction model. For example, when N = 2, that is, freeze the parameters of the first two residual modules of the blood pressure prediction model, and do not fine-tune the parameters of the first N residual modules in the following operations. In this way, the overall feature extraction ability of the blood pressure prediction model can be not changed. The value of N is less than or equal to half of the total number of residual modules.

[0075] S62. Fine-tune the trained blood pressure prediction model using the pre-processed and labeled partial sample data of the new subject, and model the blood pressure monitoring system of this subject to obtain a fine-tuned blood pressure prediction model applicable to this subject.

[0076] Specifically, use the pre-processed and labeled partial sample data of the new subject, such as 10 sample data (a total of 100 s). Take the BCG signal in each sample data as the input of the trained blood pressure prediction model to obtain the predicted blood pressure value; calculate the loss function defined in step S52 based on the predicted blood pressure value and the blood pressure label (i.e., the true blood pressure value) in this sample data. With minimizing this loss function as the optimization goal, fine-tune the model parameters of the trained blood pressure prediction model to obtain a fine-tuned blood pressure prediction model applicable to this subject.

[0077] S63. Input the unknown BCG signal data of the subject into the fine-tuned blood pressure prediction model applicable to this subject to obtain the predicted blood pressure value, and realize continuous monitoring of individual blood pressure.

[0078] In this embodiment, the technical effect of the method of the present invention is tested based on the open-source dataset KSU. This open-source dataset KSU was collected by Carlson et al., and it includes short-term blood pressure measurement data of 40 subjects and the corresponding ballistocardiogram signal data. The blood pressure prediction model in the present invention is constructed in the Pytorch framework; the model input is the BCG signal of three channels: film 0, film 1, film 2; the model output is the blood pressure value corresponding to each time slice of the input signal. An example of the output fitting effect diagram of a task is as Figure 3 shown, where the query set subscript represents different time slices.

[0079] The batch size used for training is 18, which corresponds to the number of samples included in a task. To prevent overfitting, an early stopping mechanism is set, that is, training is terminated if there is no improvement in 30 batches. The optimizer used during training is Adam. To further control overfitting, the weight decay hyperparameter is set to 0.0001.

[0080] In this embodiment, the leave-one-out cross-validation method is also used to verify the technical effects of the method of the present invention, and the mean absolute error (MAE) and root mean square error (RMSE) are used to evaluate the performance of the method of the present invention. The results are shown in Table 1. In addition, in order to evaluate the industrial application value of the method of the present invention, the prediction results of the blood pressure prediction model in the present invention are also verified against the AAMI (Advancement of Medical Instrumentation) standard and the BHS (British Hypertension Society) standard. The results are shown in Tables 2 and 3 respectively.

[0081] Table 1: Performance Results of the Model

[0082]

[0083] Table 2: AAMI Standard Comparison Table

[0084]

[0085] Table 3: BHS Standard Comparison Table

[0086]

[0087] The AAMI standard is used for non-invasive blood pressure monitoring, and the standard deviation (SD) and mean difference (MD) are used as evaluation criteria. According to this standard, the mean difference must be less than 5 mmHg, and the standard deviation must be less than 8 mmHg. As shown in Table 2, the mean differences of the systolic blood pressure (SBP) and diastolic blood pressure (DBP) of this model are -1.91 mmHg and -0.46 mmHg respectively, and the standard deviations are 5.63 mmHg and 1.58 mmHg respectively, all meeting the requirements of the AAMI standard. The BHS standard evaluates the model according to the percentage of samples within three different error thresholds. It divides the performance into three grades: A, B, and C, and the thresholds are 5, 10, and 15 mmHg respectively. As shown in Table 3, the effect of this model fully passes the B grade for systolic blood pressure and the A grade for diastolic blood pressure, achieving the B / A grade standard. It can be seen from this that the present invention can effectively realize the prediction of blood pressure and thus realize the continuous monitoring of blood pressure.

[0088] In summary, the present invention trains a highly robust blood pressure prediction model based on the meta-learning framework of few-shot learning, enabling the trained blood pressure prediction model to be optimized to a better performance after a small amount of update when facing new data, thereby modeling the vascular system of a new individual and realizing continuous monitoring of individual blood pressure. The present invention simultaneously considers the vascular system and anatomical differences between individuals as well as the slow and subtle cardiac motion interference, and also considers the signal characteristics of the BCG signal itself, enabling the finally learned blood pressure prediction model to have the ability to analyze the BCG signal and have high robustness, capable of performing personalized modeling after inputting new samples, improving the accuracy of blood pressure prediction; and the existing few-shot meta-learning framework has less application in the regression field, while the present invention realizes an effective application in the few-shot regression field.

[0089] The above embodiments are only used to illustrate the design concept and characteristics of the present invention, and the purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The protection scope of the present invention is not limited to the above embodiments. Therefore, all equivalent changes or modifications made according to the principles and design ideas disclosed by the present invention are within the protection scope of the present invention.

[0090] The above-described embodiments only represent several implementation manners of the present application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of the present application, several deformations and improvements can still be made, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

Claims

1. A small-sample non-invasive continuous blood pressure monitoring method based on ballistocardiogram signals, characterized in that, It includes the following steps: S1. Collect the ballistocardiogram (BCG) signals and the corresponding blood pressure signals, and slice them according to a preset time window to obtain a plurality of BCG slice signals and a plurality of blood pressure slice signals; construct a BCG signal dataset based on the BCG slice signals and their corresponding blood pressure slice signals; S2. Preprocess the BCG slice signals and their corresponding blood pressure slice signals in the BCG signal dataset to construct a training set; S3. Based on the need for few-shot learning, select sample data from the training set to construct a plurality of few-shot tasks, each task consisting of a support set and a query set, to complete the construction of task sample pairs; S4. Construct a blood pressure prediction model based on a deep learning network using the gradient descent algorithm, with the preprocessed BCG slice signals as the input and the predicted blood pressure values as the output; S5. Combine the meta-learning framework with the blood pressure prediction model, and use the support sets and query sets corresponding to the plurality of few-shot tasks in the meta-learning framework to train the blood pressure prediction model to obtain a trained blood pressure prediction model; S6. After fine-tuning the trained blood pressure prediction model using the labeled samples of a new subject, input the unknown BCG signal data of this subject to achieve continuous monitoring of the individual's blood pressure.

2. The non-invasive continuous blood pressure monitoring method for small samples based on ballistocardiogram signals according to claim 1, characterized in that, In the step S2, the preprocessing of the BCG slice signals specifically includes: Taking different subjects as processing units, perform downsampling on the plurality of BCG slice signals of different subjects using sliding window downsampling; Taking different subjects as processing units, for each downsampled BCG slice signal, perform Z-score normalization to obtain the normalized BCG slice signal.

3. The small-sample non-invasive continuous blood pressure monitoring method based on ballistocardiogram signals according to claim 1, wherein In the step S3, the construction of the task sample pairs is specifically achieved by the following method: First, select sample data from the training set to construct a plurality of few-shot tasks, each task consisting of the support set and the query set corresponding to this task; among them, 70%-90% of the training samples in the support set corresponding to each task and all the test samples in the query set come from the sample data of the same subject in the training set, and the remaining training samples in the support set come from the sample data of other different subjects in the training set; the sample data includes the preprocessed BCG slice signals and blood pressure slice signals, and take the systolic or diastolic blood pressure of the blood pressure slice signal corresponding to the preprocessed BCG slice signal as the blood pressure label; Then, for the sample data from different subjects in the support set, add perturbations artificially to obtain the final task sample pairs.

4. The non-invasive continuous blood pressure monitoring method for small samples based on ballistocardiogram signals according to claim 1, characterized in that, The blood pressure prediction model uses ResNet-18, and this blood pressure prediction model includes an input module, a plurality of residual modules, and an output module. Among them, the input module includes a convolutional layer and a max pooling layer; each residual module includes two residual blocks, and each residual block includes two convolutional layers; the output module includes a global average pooling layer, a fully connected layer, and a linear layer; The preprocessed ballistocardiogram slice signal is input into the blood pressure prediction model. First, it passes through the convolutional layer and the max pooling layer of the input module to obtain an initial feature representation. Then, the initial feature representation successively passes through multiple residual modules to extract different levels of features of the ballistocardiogram slice signal. In multiple successively connected residual modules, through the skip connection mechanism, the output of the previous residual block is added to the input of the previous residual block as the input of the current residual block. The input of the first residual block of the first residual module is the initial feature representation output by the input module. The output of the last residual block of the last residual module is added to the input of this residual block to obtain the final feature vector. The final feature vector is used as the input of the output module. Through the global average pooling layer and the fully connected layer, the final feature vector is reduced to a fixed-length embedding vector, and the predicted blood pressure value is obtained through the linear layer.

5. The non-invasive continuous blood pressure monitoring method for small samples based on ballistocardiogram signals according to claim 1, characterized in that The specific steps of step S5 are as follows: S51. Initialize model parameters: Randomly initialize the parameters of the blood pressure prediction model and use them as the initial state of the meta-model. Regard the blood pressure prediction model as the meta-model. S52. Define the loss function: Calculate the mean square error between the predicted blood pressure value and the true blood pressure value as the loss function. S53. Task training and update: For each few-shot task, clone the meta-model for the inner-loop training of the meta-learning framework. Perform several steps of gradient descent update on the support set of each task to obtain the model parameters corresponding to this task. Each step of gradient descent update is performed with the optimization goal of minimizing the calculated loss function. S54. Meta-update: During the outer-loop training of the meta-learning framework, calculate the loss function based on the meta-model obtained in the inner-loop for each task on the query set corresponding to this task. Accumulate the loss functions of all tasks to obtain the total loss. With the optimization goal of minimizing the total loss, use the Adam optimizer to optimize the model parameters of the meta-model through backpropagation. S55. Outer-loop train the meta-model: Repeat steps S53 - S54, and use the loss function output from the inner-loop to update the parameters of the meta-model to obtain the trained meta-model, which is the trained blood pressure prediction model.

6. The non-invasive continuous blood pressure monitoring method for small samples based on ballistocardiogram signals according to claim 1, characterized in that, The specific steps of step S6 are as follows: S61. Freeze the parameters of the first N residual modules of the blood pressure prediction model, where the value of N is less than or equal to half of the total number of residual modules. S62. Use the preprocessed labeled partial sample data of the new subject to fine-tune the trained blood pressure prediction model to obtain a fine-tuned blood pressure prediction model suitable for this subject. S63. Input the unknown ballistocardiogram signal data of the subject into the fine-tuned blood pressure prediction model suitable for this subject to obtain the predicted blood pressure value and realize continuous monitoring of the individual's blood pressure.

Images