Blood pressure measuring method, wearable device and cloud device

Abstract

The invention provides a blood pressure measuring method, wearable equipment and cloud equipment. The method comprises the following steps: in response to a blood pressure measurement operation of a user, acquiring a photoplethysmography (PPG) signal of the user to obtain PPG signal data; based on the PPG signal data and a trained blood pressure prediction model, obtaining a blood pressure measurement result of the user; the target training sample is obtained by sampling an original training sample set, the difference degree between any two pieces of sample PPG signal data in the target training sample set is greater than a preset threshold value, the target training sample set contains target sample PPG signal data, and the original training sample set does not contain the target sample PPG signal data. Therefore, the training set is sampled, the data redundancy of the training set is reduced, and the gaps among the data in the training set are filled, so that the effect of improving the generalization of the blood pressure prediction model is achieved, and the blood pressure prediction precision of the blood pressure prediction model on unseen individuals is improved.

Description

Technical Field

[0001] The present application relates to the field of smart terminal technology, and in particular to a blood pressure measurement method, a wearable device, and a cloud device. Background Art

[0002] Blood pressure measurement is a common means of health assessment. With the development of technology, blood pressure measurement equipment is becoming smaller and lighter. For example, blood pressure measurement equipment can be a smart wearable device. Users who need to measure blood pressure can wear the smart wearable device for a long time and use the smart wearable device to measure blood pressure at any time.

[0003] Existing smart wearable devices generally obtain the user's blood pressure value by collecting relevant physiological signals from the user and inputting the relevant physiological signals into a trained blood pressure prediction model. The blood pressure prediction model is usually trained on a large-scale training set, and tends to have a strong memory for individuals that have been seen, so the prediction accuracy is high. However, this memory is associated with the individual, not the blood pressure, which results in the blood pressure prediction ability of the blood pressure prediction model being very limited for users that have never been seen. Summary of the invention

[0004] The present application provides a blood pressure measurement method, a wearable device and a cloud device. In the method, by sampling the training set, reducing the data redundancy of the training set and filling the gaps between the data in the training set, the generalization of the blood pressure prediction model is improved, thereby improving the blood pressure prediction accuracy of the blood pressure prediction model for unseen individuals.

[0005] In a first aspect, the present application provides a blood pressure measurement method, which is applied to a wearable device, the method comprising: in response to a user's blood pressure measurement operation, collecting the user's photoplethysmography (PPG) signal to obtain PPG signal data; based on the PPG signal data and a trained blood pressure prediction model, obtaining the user's blood pressure measurement result, and displaying the blood pressure measurement result; the blood pressure prediction model is trained based on a target training sample set, the target training sample is obtained by sampling an original training sample set, the difference between any two sample PPG signal data in the target training sample set is greater than a preset threshold, and the target training sample set contains the target sample PPG signal data, while the original training sample set does not contain the target sample PPG signal data.

[0006] Through the above scheme, after obtaining the photoplethysmography PPG signal data of the user to be tested, the present application inputs the PPG signal data into the trained blood pressure prediction model to obtain the blood pressure measurement result of the user to be tested. Since the blood pressure prediction model of the present application is trained based on the target training sample set, rather than the original training sample set. The target training sample set is obtained after sampling the original training samples. The present application effectively screens and augments the original training sample set, so that the target training sample set is richer and more representative than the original training sample set. Therefore, compared with the related art that directly uses the original training samples to train the blood pressure prediction model, the present application can suppress the overfitting phenomenon of the blood pressure prediction model by optimizing the training samples, enhance the generalization of the blood pressure prediction model, and thus improve the accuracy of the blood pressure prediction model.

[0007] In one possible implementation, the wearable device is deployed with a blood pressure prediction model; based on the PPG signal data and the trained blood pressure prediction model, the user's blood pressure measurement result is obtained, including: inputting the PPG signal data into the blood pressure prediction model to obtain the blood pressure measurement result of the user to be measured.

[0008] Through the above solution, after the wearable device obtains the user's PPG signal data, it can directly input the PPG signal data into the locally deployed blood pressure prediction model, thereby achieving the effect of quickly measuring blood pressure.

[0009] In one possible implementation, a user's blood pressure measurement result is obtained based on PPG signal data and a trained blood pressure prediction model, including: sending a blood pressure measurement request containing PPG signal data to the cloud; a blood pressure prediction model is deployed on the cloud; receiving the blood pressure measurement result sent by the cloud; the blood pressure measurement result is obtained by the cloud inputting the PPG signal data into the blood pressure prediction model.

[0010] Through the above solution, wearable devices use the model deployed in the cloud to measure the user's blood pressure, which reduces the local deployment pressure of wearable devices and saves the computing power of wearable devices.

[0011] In a second aspect, the present application provides a blood pressure measurement method, which is applied to the cloud, and the method includes: receiving a blood pressure measurement request containing PPG signal data sent by a wearable device; inputting the PPG signal data into a trained blood pressure prediction model to obtain a blood pressure measurement result, and sending the blood pressure measurement result to the wearable device; the blood pressure prediction model is trained based on a target training sample set, the target training sample is obtained by sampling the original training sample set, the difference between any two sample PPG signal data in the target training sample set is greater than a preset threshold, and the target training sample set contains the target sample PPG signal data, and the original training sample set does not contain the target sample PPG signal data.

[0012] Through the above scheme, after obtaining the photoplethysmography PPG signal data of the user to be tested, the present application inputs the PPG signal data into the trained blood pressure prediction model to obtain the blood pressure measurement result of the user to be tested. Since the blood pressure prediction model of the present application is trained based on the target training sample set, rather than the original training sample set. The target training sample set is obtained after sampling the original training samples. The present application effectively screens and augments the original training sample set, so that the target training sample set is richer and more representative than the original training sample set. Therefore, compared with the related art that directly uses the original training samples to train the blood pressure prediction model, the present application can suppress the overfitting phenomenon of the blood pressure prediction model by optimizing the training samples, enhance the generalization of the blood pressure prediction model, and thus improve the accuracy of the blood pressure prediction model.

[0013] In one possible implementation, the method also includes: obtaining an original training sample set; sampling the original training sample set to obtain a target training sample set; using the sample PPG signal data in the target training sample set as input, and using the blood pressure corresponding to the sample PPG signal data in the target training sample set as a label, to train a blood pressure prediction model.

[0014] In one possible implementation, sampling the original training sample set to obtain a target training sample set includes: performing a first sampling on the original training sample set based on a first sampling strategy to obtain a first training sample set; the number of sample PPG signal data in the first training sample set is less than the number of sample PPG signal data in the original training sample, and the degree of difference between any two sample PPG signal data in the first training sample set is greater than a preset threshold; performing a second sampling on the first training sample set based on a second sampling strategy to obtain a second training sample set; the second training sample set contains target sample PPG signal data, and the original training sample set and the first training sample set do not contain target sample PPG signal data; and the second training sample set is used as the target training sample set.

[0015] Through the above scheme, the cloud server first performs the first sampling of the original training sample set through the first sampling strategy. Selecting samples with large differences can reduce the redundant information between samples, avoid overfitting due to too much redundant information, and cover the original feature space as much as possible without losing personal information, which can fully improve the training efficiency. At the same time, when processing unbalanced data sets, it can help select minority class samples and improve the model's recognition ability for minority classes. Furthermore, the cloud server performs the second sampling of the first training sample set through the second sampling strategy to fill in some samples that are missing in the original training sample set, which can enhance the generalization of the model in subsequent training.

[0016] In a possible implementation, sampling the original training sample set to obtain a target training sample set includes: performing a first sampling on the original training sample set based on a first sampling strategy to obtain a first training sample set; the number of sample PPG signal data in the first training sample set is less than the number of sample PPG signal data in the original training sample, and the degree of difference between any two sample PPG signal data in the first training sample set is greater than a preset threshold; performing a second sampling on the original training sample set based on a second sampling strategy to obtain a second training sample set; the second training sample set contains target sample PPG signal data, and the original training sample set does not contain target sample PPG signal data; and using the first training sample set and the second training sample set as target training sample sets.

[0017] Through the above scheme, the cloud server can perform the first sampling and the second sampling on the original training samples. The first sampling can select samples with large differences to reduce the redundant information between samples and avoid overfitting due to excessive redundant information. The first sampling can fill in some samples that are missing in the original training sample set. Furthermore, the cloud server uses the training sample set with two advantages as the target training sample set, which can better enhance the generalization ability of the model.

[0018] In a possible implementation, the first sampling strategy includes diversity sampling, and the second sampling strategy includes interpolation sampling.

[0019] Through the above scheme, this application aims at the technical defects of the original training sample set, designs a targeted sampling strategy, and performs targeted sampling processing on the original training sample set to effectively screen and expand the original training sample set, so that the target training sample set is richer and more representative than the original training sample set. Therefore, compared with the related technology of directly training the blood pressure prediction model with the original training samples, this application can suppress the overfitting phenomenon of the blood pressure prediction model by optimizing the training samples, enhance the generalization of the blood pressure prediction model, and thus improve the accuracy of the blood pressure prediction model.

[0020] In a possible implementation, a first sampling is performed on the original training sample set based on a first sampling strategy to obtain a first training sample set, including: inputting the sample PPG signal data in the original training sample set into a trained classification model to obtain a classification result of the sample PPG signal data; the classification model is used to extract the identity features of each sample PPG signal data, and classify each sample PPG signal data according to the identity features; sample PPG signal data of the same category in the classification result corresponds to the same sample user; and first sampling is performed on the sample PPG signal data of each category based on the first sampling strategy to obtain a first training sample set.

[0021] In a possible implementation, first sampling is performed in sample PPG signal data of each category based on a first sampling strategy to obtain a first training sample set, including: for sample PPG signal data of any category, a target number of sample PPG signal data are collected from sample PPG signal data of the current category to obtain a first sampling data set; the cosine similarity between the identity features of any two sample PPG signal data in the first sampling data set is greater than a first preset similarity; and each first sampling data set is used as a first training sample set.

[0022] In one possible implementation, a target number of sample PPG signal data are collected from the sample PPG signal data of the current category to obtain a first sampling data set, including: for any collected sample PPG signal data, respectively calculating the cosine similarity between the uncollected sample PPG signal data and the collected sample PPG signal data; and adding the sample PPG signal data with the largest cosine similarity among the uncollected sample PPG signal data to the first sampling data set.

[0023] In a possible implementation, the method also includes: dividing the sample PPG signal data in the original training sample set according to the sample users to which they belong, to obtain sample PPG signal data of multiple sample users; taking the sample PPG signal data of multiple sample users as input, and setting the cosine residual normalization loss function Angular-margin Softmax to train a classification model; Angular-margin Softmax is used to instruct the classification model to maximize the probability of the category to which it belongs and minimize the probability gap with adjacent categories.

[0024] In a possible implementation, performing a second sampling on the first training sample set based on a second sampling strategy to obtain a second training sample set includes: inputting the first training sample set into a trained sampling model to obtain a second training sample set.

[0025] In one possible implementation, the sampling model includes an encoder, a sampler, and a decoder; the first training sample set is input into the trained sampling model to obtain a second training sample set, including: encoding the sample PPG signal data in the first training sample set according to the sample users to which they belong by the encoder to obtain multiple first codes; one first code corresponds to the sample PPG signal data of a sample user in the first training sample set; interpolating sampling is performed between the multiple first codes by the sampler to obtain multiple second codes; and the multiple second codes are decoded by the decoder to obtain a second training sample set.

[0026] In a possible implementation, a sampler performs interpolation sampling between multiple first codes to obtain multiple second codes, including: constructing a data structure graph corresponding to the multiple first codes; the data structure graph includes multiple nodes, one node corresponds to one first code, each node is connected by an edge, and the length of the edge between any two nodes is used to reflect the coding similarity between the two nodes; determining candidate node pairs in the data structure graph; the candidate node pairs are any two nodes in the data structure graph whose edge lengths are greater than a preset edge length; based on the candidate node pairs, determining the target node pairs, and inserting new nodes between the target node pairs; the target node pairs are candidate node pairs where no other candidate nodes exist between any two candidate nodes.

[0027] In a possible implementation, inserting a new node between a target node pair includes: determining the code corresponding to the node to be inserted according to the code corresponding to the first target node and the code corresponding to the second target node, and obtaining the node to be removed; the first target node and the second target node are any two target node pairs; and inserting the node to be removed between the first target node and the second target node.

[0028] In a possible implementation, the code corresponding to the node to be inserted is determined according to the code corresponding to the first target node and the code corresponding to the second target node to obtain the node to be inserted, including: weighting the code corresponding to the first target node and the code corresponding to the second target node to obtain the node to be inserted.

[0029] In a third aspect, a wearable device is provided, comprising a processor coupled to a memory, the processor being configured to execute a computer program or instruction stored in the memory so that the wearable device implements a blood pressure measurement method as described in any one of the first aspects.

[0030] In a fourth aspect, a cloud device is provided, comprising a processor coupled to a memory, the processor being configured to execute a computer program or instruction stored in the memory, so that the cloud device implements a blood pressure measurement method as described in any one of the first aspects.

[0031] In a fifth aspect, a chip is provided, which is coupled to a memory and is used to read and execute a computer program stored in the memory to implement a blood pressure measurement method as described in any one of the first aspects.

[0032] In a sixth aspect, a computer-readable storage medium is provided, which stores a computer program. When the computer program is executed on a terminal device, the terminal device executes a blood pressure measurement method as described in any one of the first aspects.

[0033] In a seventh aspect, a computer program product is provided. When the computer program product is run on a computer, the computer is caused to execute the blood pressure measurement method as described in any one of the first aspects.

[0034] It can be understood that the beneficial effects of the second to seventh aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A schematic diagram of the PPG acquisition principle provided in the embodiment of the present application; Figure 2 A PPG signal waveform diagram provided in an embodiment of the present application; Figure 3 A schematic diagram of a blood pressure prediction model training provided in an embodiment of the present application; Figure 4 Schematic diagram of sample data division provided in the embodiment of the present application Figure 1 ; Figure 5 Schematic diagram of sample data division provided in the embodiment of the present application Figure 2 ; Figure 6 Schematic diagram of the application scenario of the blood pressure measurement method provided in the embodiment of the present application Figure 1 ; Figure 7 Schematic diagram of the application scenario of the blood pressure measurement method provided in the embodiment of the present application Figure 2 ; Figure 8 Schematic diagram of the application scenario of the blood pressure measurement method provided in the embodiment of the present application Figure 3 ; Fig. 9 A schematic diagram of the hardware structure of a wearable device provided in an embodiment of the present application; Fig.10 A schematic diagram of the software structure of a wearable device provided in an embodiment of the present application; Fig.11 A schematic diagram of the hardware structure of a cloud device provided in an embodiment of the present application; Fig.12 Schematic diagram of the blood pressure measurement method provided in the embodiment of the present application Figure 1 ; Fig.13 Schematic diagram of the blood pressure measurement method provided in the embodiment of the present application Figure 2 ; Fig.14 A schematic diagram of the distribution process of the blood pressure prediction model provided in the embodiment of the present application; Fig.15 A flow chart of the training process of the blood pressure prediction model provided in the embodiment of the present application; Fig.16 A schematic diagram of the structure of the classification model provided in the embodiment of the present application; Fig.17 A schematic diagram of the classification effect of the classification model provided in the embodiment of the present application; Fig.18 A schematic diagram of the structure of the sampling model provided in the embodiment of the present application; Fig.19 A schematic diagram of interpolation conditions provided in an embodiment of the present application; Fig. 20 This is a diagram showing the effect of sampling after interpolation by the sampling model provided in the embodiment of the present application; Fig.21 A schematic diagram of the module structure of a wearable device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present application to clearly and completely describe the technical solutions in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of this application.

[0037] The term "and / or" in this article is merely a description of the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone.

[0038] The terms "first" and "second" in the description and claims of the embodiments of the present application are used to distinguish different objects rather than to describe a specific order of objects. For example, a first target object and a second target object are used to distinguish different target objects rather than to describe a specific order of target objects.

[0039] In the embodiments of the present application, words such as "exemplary" or "for example" are used to indicate examples, illustrations or descriptions. Any embodiment or design described as "exemplary" or "for example" in the embodiments of the present application should not be interpreted as being more preferred or more advantageous than other embodiments or designs. Specifically, the use of words such as "exemplary" or "for example" is intended to present related concepts in a specific way.

[0040] In the description of the embodiments of the present application, unless otherwise specified, the meaning of "multiple" refers to two or more than two. For example, multiple processing units refer to two or more processing units; multiple systems refer to two or more systems.

[0041] For ease of understanding, the relevant technical terms and related technologies involved in the embodiments of the present application are first introduced below.

[0042] 1. Blood pressure (BP).

[0043] Blood pressure refers to the lateral pressure acting on the blood vessel wall per unit area when blood flows in the blood vessels. It is the driving force for blood to flow in the blood vessels. Among them, blood vessels are divided into arteries, capillaries and veins. Correspondingly, blood pressure can be divided into arterial blood pressure, capillary blood pressure and venous blood pressure. The blood pressure mentioned in daily life refers to the blood pressure of the brachial artery in the artery. It should be understood that the blood pressure measured in this article is also the blood pressure of the brachial artery.

[0044] 2. Systolic blood pressure (SBP) and diastolic blood pressure (DBP).

[0045] Blood pressure further includes systolic pressure (which is what people call high pressure in daily life) and diastolic pressure (which is what people call low pressure in daily life). Systolic pressure refers to the blood pressure when the heart contracts and pumps blood into the arteries, and the arteries expand to buffer the pressure. The blood pressure at this time is systolic pressure. Diastolic pressure refers to the blood pressure when the heart relaxes and the arteries contract to allow blood to flow back to the heart. The blood pressure at this time is diastolic pressure.

[0046] 3. Photoplethysmographic (PPG).

[0047] Photoplethysmography is a technology that non-invasively and continuously monitors changes in blood volume. It uses photoelectric means to detect changes in blood volume in living tissues, thereby indirectly reflecting the pulsation of the heart. PPG technology is usually achieved by emitting a certain wavelength of light (usually green or red light) to the skin and measuring the changes in light intensity absorbed and reflected by the tissue. When light hits the skin, part of the light is absorbed by tissues such as the skin, muscles, and bones, while the other part is reflected back. Due to the pulsation of blood in the arteries, the blood volume under the skin changes periodically, which affects the intensity of the reflected light. The PPG sensor receives the intensity changes of this reflected light and converts it into an electrical signal to obtain a pulse waveform.

[0048] 4.PPG sensor.

[0049] A PPG sensor is a sensor that shines light into the skin and measures the light scattered by blood flow. Figure 1 , PPG sensors usually include a light emitter (usually an LED) and a light detector. The light emitter emits a light beam of a certain wavelength onto the skin, and the light passes through the skin (such as Figure 1The light will be attenuated to a certain extent after it is reflected to the light detector after passing through the epidermis, dermis and / or subcutaneous tissue in the artery. It should be understood that when the measurement site does not change significantly, the absorption of light by muscles, bones and veins is basically unchanged. However, during cardiac contraction and diastole, the absorption of light by arteries will be different due to blood flow. It is precisely because the absorption of light by arteries changes while the absorption of light by other tissues (such as the aforementioned muscles, bones and veins) remains basically unchanged, then, after measuring the light signal for a period of time through the light detector, the AC component of the electrical signal corresponding to the light signal (referred to as the PPG signal in this article) is extracted, and a waveform reflecting the characteristics of blood flow (referred to as the PPG waveform in this article) can be obtained. It is thus used to determine physiological parameters such as blood oxygen saturation, blood pressure, and heart rate.

[0050] The PPG sensor uses LED to illuminate human skin and uses a photodiode to measure the change in reflected light intensity caused by blood flow. The resulting periodic waveform contains information about the pulsating changes in blood volume during the heart beat cycle. The blood volume is the largest when the heart contracts, and the diastolic volume is the smallest when the heart relaxes. Therefore, the systolic and diastolic pressures can be obtained through the relationship between volume pulse blood flow information and blood pressure. Figure 2 The waveform of the photoplethysmography (PPG) signal is shown in Figure 1, where the horizontal axis is the timestamp and the vertical axis is the waveform of the PPG signal over time. The peak corresponds to the contraction of the human heart and the trough corresponds to the diastole of the human heart. Therefore, the peak of the PPG waveform can correspond to the systolic pressure and the trough of the PPG waveform can correspond to the diastolic pressure.

[0051] 5. Machine learning model.

[0052] A machine learning model is a mathematical model that makes predictions or inferences based on data. These models learn patterns in data to automatically complete tasks such as classification, regression, clustering, generation, etc. In practical applications, training can be done using labeled data, which contains inputs and corresponding correct outputs (labels). Common models include linear regression, logistic regression, support vector machine (SVM), decision tree, random forest, etc.

[0053] Therefore, the relationship between PPG signals and blood pressure can be used to train a blood pressure prediction model to achieve blood pressure prediction effects.

[0054] like Figure 3 As shown, the related technology collects data of PPG signals and blood pressure values ​​of different sample users to obtain training sets and validation sets, and uses the training sets and validation sets to train a blood pressure prediction model. During the training process, a large number of sample users' PPG signals are input, and the blood pressure of each sample user is used as a label to adjust the model parameters. When the parameters in the model basically no longer change, that is, the training reaches a stable state, the training process ends.

[0055] 6. Training Set, Validation Set and Test Set.

[0056] In machine learning and data science, it is a common practice to divide the dataset into training set, validation set, and test set, which helps to evaluate the performance of the model, prevent overfitting, and optimize model parameters. The following are the specific functions and usage of these datasets: Training Set Function: The training set is used to train the machine learning model, that is, the model fits the data distribution and learns the parameters by learning the data in the training set.

[0057] How to use: Input data into the model for training. The model minimizes the loss function by adjusting parameters to better fit the training data.

[0058] Validation Set Function: The validation set is used to evaluate the performance of the model during the training process, help select the best model parameters and prevent overfitting. The validation set does not participate in the model training process and is only used to evaluate and adjust the model.

[0059] How to use: During the training process, use the validation set regularly to evaluate the performance of the model (such as accuracy, loss, etc.). By observing the performance changes on the validation set, you can perform early stopping, adjust the learning rate, select the best model architecture, etc. During the model selection and parameter adjustment stage, use the validation set to select the best performing model.

[0060] Test Set Function: The test set is used to evaluate the generalization ability of the model after training, that is, the performance of the model on unseen data. The test set provides an unbiased performance evaluation to measure the performance of the model in practical applications.

[0061] Usage: After the model training is completed, use the test set for a one-time evaluation to obtain the final performance indicators of the model. The results of the test set should not be used to adjust model parameters or perform any form of model selection.

[0062] In the blood pressure prediction scenario, the data set partitioning method has a great impact on the blood pressure prediction results. The relevant technology usually includes two partitioning methods, which are introduced below.

[0063] Partitioning method 1 (also known as intra-subject): divide the data of each individual into training set, validation set and test set. Figure 4 As shown, the existing data set includes a first sample group (including sample user 1, sample user 2, sample user 3), a second sample group (including sample user 4, sample user 5, sample user 6) and a third sample group (including sample user 7, sample user 8, sample user 9). The data of each sample user is divided into a training set, a validation set and a test set according to the dotted lines.

[0064] Since the training set and the test set share the data of the same individual in this way, this division method is somewhat deceptive to the model. For example, the first part of the data of sample user 1 is the training set, the second part of the data is the validation set, and the third part of the data is the test set. After the model is trained with the first and second parts of the data of sample user 1, it is easy to make blood pressure predictions for the third part of the data of sample user 1 because they are the data of the same user. However, this does not mean that the model meets the standard. If a sample user 10 (the data of this user does not exist in the above sample data) is tested, the test results show that the model cannot accurately predict the blood pressure of sample user 10.

[0065] Division method 2 (also known as inter-subject): Divide different individual data into training sets, validation sets, and test sets to ensure that there is no overlap in the individual data in each data set. Figure 5 As shown, the existing data set includes a first sample group (including sample user 1, sample user 2, sample user 3), a second sample group (including sample user 4, sample user 5, sample user 6), and a third sample group (including sample user 7, sample user 8, sample user 9). The data of different sample users are divided into a training set, a validation set, and a test set according to the dotted lines.

[0066] Under this division method, the algorithm and training method of the model remain unchanged. Tests show that the blood pressure prediction performance of the model is worse than that of division method 1. This further illustrates that the current blood pressure prediction model has very limited prediction capabilities for unseen data, that is, the model obtained by sampling in related technologies on the entire training set has a serious overfitting phenomenon and poor generalization.

[0067] In view of the above problems, the embodiment of the present application provides a blood pressure measurement method, which considers using PPG signals to perform individual recognition training on a large-scale data set to obtain the identity codes of different individuals in the latent space. On the basis of obtaining the latent space codes of the individuals in the training set, the data set is subjected to diversity sampling and significant gap interpolation sampling based on the generative model. The purpose is to suppress overfitting through such data augmentation strategies, improve the generalization of the blood pressure model, and thus improve the blood pressure prediction accuracy of unseen individuals.

[0068] Before describing the technical solution of the embodiments of the present application, the application scenarios of the embodiments of the present application are first described with reference to the accompanying drawings.

[0069] See also Figure 6 , is a schematic diagram of an application scenario provided in an embodiment of the present application. The application scenario includes a wearable device 11 and a cloud 12. Figure 6 As shown, the wearable device 11 is connected to the cloud 12 for communication. In the specific implementation of the embodiment of the present application, the cloud 12 may include one or more cloud servers for providing services for the wearable device 11. The wearable device 11 may be a smart watch, a smart bracelet, a smart wristband, smart jewelry, etc., which is not limited in the embodiment of the present application. The specific technology and specific device form used by the wearable device are not limited in the embodiment of the present application. It should be noted that in actual applications, the number of clouds can be one or more, Figure 6 The number of cloud and wearable devices in the application scenario shown is only an adaptive example and is not limited in this application.

[0070] In a specific application, after the user wears the wearable device 11, the wearable device 11 can be instructed to measure blood pressure. After receiving the user's blood pressure measurement instruction, the wearable device 11 can collect the user's PPG signal and upload the collected PPG signal to the cloud 12 to request the cloud 12 to predict the user's blood pressure. Correspondingly, the cloud 12 is deployed with a blood pressure prediction model. After receiving the PPG signal uploaded by the wearable device 11, the cloud 12 inputs the PPG signal into the blood pressure prediction model to obtain the blood pressure value, and sends the blood pressure value to the wearable device 11. After receiving the blood pressure value sent by the cloud 12, the wearable device 11 can save and display the blood pressure value.

[0071] Take the wearable device as a smart watch as an example. Figure 7 As shown in a, a blood pressure detection APP is installed on the smart watch, and the user can click the blood pressure detection APP to enter the blood pressure detection interface. Figure 7 As shown in Figure b, the blood pressure detection interface includes a blood pressure display box and a measurement button. The blood pressure display box includes a high pressure (systolic pressure) display box and a low pressure (diastolic pressure) display box. The user can click the measurement button in the blood pressure detection interface to enter the blood pressure measurement interface, and the smart watch starts to obtain the user's PPG signal. Figure 7 As shown in c, the blood pressure measurement interface includes a prompt box and a time box. The prompt box is used to remind the user to "keep still during measurement and pay attention to wearing the watch correctly"; the time box is used to display the countdown of the blood pressure measurement time. After the measurement is completed, Figure 7As shown in d, it will automatically return to the blood pressure detection interface, and the high pressure (systolic pressure) and low pressure (diastolic pressure) currently measured by the user will be displayed in the blood pressure display box.

[0072] See also Figure 8 , is a schematic diagram of another application scenario provided by an embodiment of the present application. The application scenario includes a wearable device 21, a mobile phone 22, and a cloud 23. Figure 8 As shown, the wearable device 21 is connected to the mobile phone 22 for communication, and the mobile phone 22 is connected to the cloud 23 for communication. In the specific implementation of the embodiment of the present application, the wearable device 21 and the mobile phone 22 can be connected via Bluetooth, and the mobile phone 22 and the cloud 23 can be connected via the network.

[0073] In a specific application, the wearable device 21 can establish a Bluetooth connection with the mobile phone 22 in response to the user's Bluetooth pairing operation. Or the mobile phone 22 establishes a Bluetooth connection with the wearable device 21 in response to the user's Bluetooth pairing operation. After the wearable device 21 and the mobile phone 22 establish a Bluetooth connection for the first time, as long as the wearable device 21 and the mobile phone 22 both have the Bluetooth function turned on and the distance is less than a certain threshold, a Bluetooth connection will be automatically established. After the user wears the wearable device 21, the user can instruct the wearable device 21 to measure blood pressure through the mobile phone 22. After receiving the blood pressure measurement instruction, the wearable device 21 can collect the user's PPG signal and send the collected PPG signal to the mobile phone 22. After receiving the PPG signal collected by the wearable device 21, the mobile phone 22 uploads the PPG signal to the cloud 23 to request the cloud 23 to predict the user's blood pressure. Accordingly, the cloud 23 is deployed with a blood pressure prediction model. After receiving the PPG signal uploaded by the mobile phone 22, the cloud 23 inputs the PPG signal into the blood pressure prediction model to obtain the blood pressure value, and sends the blood pressure value to the mobile phone 22. After the mobile phone 22 receives the blood pressure value sent by the cloud 23, the blood pressure value can be saved and displayed.

[0074] Of course, the present application can also execute the blood pressure measurement method in other scenarios, and the embodiments of the present application are not limited to this.

[0075] Combine the following Fig. 9 Introduce the hardware structure of wearable devices.

[0076] See also Fig. 9 , is a hardware structure diagram of a wearable device (such as a smart watch) that can be worn on a wrist provided in an embodiment of the present application. Fig. 9As shown, the wearable device includes: a processor 310, a display screen 320, a PPG sensor 330, a pressure sensor 340, an accelerometer (ACC) sensor 350, a memory 360 and a wireless communication module 370. The processor 310 may include one or more interfaces for connecting to other components of the electronic device. Among them, the one or more interfaces may include: an input / output (I / O) interface (also called an I / O pin), an interrupt pin and a data bus interface, etc. Among them, the data bus interface may include: one or more of the interfaces such as a serial peripheral interface (SPI), an integrated circuit (I2C) interface, etc.

[0077] The processor 310 may include one or more processing units, for example, the processor 310 may include an application processor (AP), a modem processor, a graphics processor (GPU), an image signal processor (ISP), a controller, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural-network processing unit (NPU), etc. Different processing units may be independent devices or integrated into one or more processors.

[0078] The processor 310 can generate an operation control signal according to the instruction operation code and the timing signal to complete the control of fetching and executing instructions.

[0079] The processor 310 may also be provided with a memory for storing instructions and data. In some embodiments, the memory in the processor 310 may be a cache memory. The memory may store instructions or data that have been used or are frequently used by the processor 310. If the processor 310 needs to use the instruction or data, it may be directly called from the memory. This avoids repeated access, reduces the waiting time of the processor 310, and thus improves the efficiency of the system.

[0080] In some embodiments, the processor 310 may be used to perform calculations for blood pressure prediction. For example, the processor 310 may predict blood pressure based on the phase difference between the pressure waveform and the PPG waveform, the period of the PPG waveform, and the signal amplitude of the pressure waveform.

[0081] The display screen 320 is used to display images, videos, etc. The display screen 320 includes a display panel. The display panel can be a liquid crystal display (LCD), an organic light-emitting diode (OLED), an active-matrix organic light-emitting diode or an active-matrix organic light-emitting diode (AMOLED), a flexible light-emitting diode (FLED), Miniled, MicroLed, Micro-oLed, a quantum dot light-emitting diode (QLED), etc. In some embodiments, the wearable device may include one or more display screens 320.

[0082] In some embodiments, the display screen 320 may display the predicted blood pressure value and related controls for blood pressure measurement, such as a start control, an end control, etc.

[0083] The memory 360 can be used to store computer executable program codes, which include instructions. The memory 360 may include a program storage area and a data storage area. Among them, the program storage area may store an operating system, an application required for at least one function (such as a sound playback function, an image playback function), etc. The data storage area may store data created during the use of the wearable device (such as audio data, a phone book, etc.), etc. In addition, the memory 360 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, a universal flash storage (UFS), etc. The processor 310 executes various functional methods or data processing of the wearable device by running instructions stored in the memory 360 and / or instructions stored in a memory provided in the processor.

[0084] The wireless communication module 370 can support data exchange between electronic devices and other electronic devices through wireless communications including Bluetooth, global navigation satellite system (GNSS), wireless LAN, frequency modulation (FM), near field communication technology (NFC), infrared technology (IR), etc.

[0085] The PPG sensor 330 is used to emit light and detect reflected light signals, and convert the light signals into electrical signals. The AC component in the electrical signal can reflect the characteristics of blood flow, and thus can be used to calculate blood pressure.

[0086] The pressure sensor 340 is used to sense the pressure signal and convert the pressure signal into an electrical signal. There are many types of pressure sensors 340, such as resistive pressure sensors, inductive pressure sensors, capacitive pressure sensors, etc. A capacitive pressure sensor may include at least two parallel plates with conductive materials. When a force acts on the pressure sensor 340, the capacitance between the electrodes changes. The wearable device determines the intensity of the pressure based on the change in capacitance.

[0087] In some embodiments, the wearable device further includes an acceleration sensor 350. The acceleration sensor 350 can be used to detect the magnitude of the acceleration of the wearable device in various directions (generally three axes). This is used to determine the movement of the wearable device. For example, the wearable device can measure blood pressure when the wearable device is not moving, so as to avoid the movement of the wearable device affecting the accuracy of the measurement result.

[0088] At this point, it should be noted that in actual implementation, the hardware composition of wearable devices is not based on Fig. 9 Exemplarily, the wearable device may also include a power management module, a battery, a camera, a motor, a button, and other modules.

[0089] Combine the following Fig.10 This paper introduces the software structure of wearable devices.

[0090] Fig.10 : is a software structure diagram of a wearable device provided in an embodiment of the present application. The layered architecture of the wearable device divides the software into several layers, and each layer has a clear role and division of labor. The layers communicate with each other through software interfaces. In some embodiments, taking the wearable device as a smart watch as an example, Fig.10 As shown, the operating system of the smart watch may include five layers, namely, from top to bottom, UI (User Interface) application layer, system service layer, algorithm layer, hardware abstraction layer, and kernel layer.

[0091] The UI application layer may include a series of application packages, such as watch faces, sports records, calls, exercises, etc.

[0092] The system service layer can include a series of system services. Fig.10As shown, the system service layer may include a blood pressure service, which can provide physiological parameter information of the smart watch wearer, such as heart rate, blood pressure, blood oxygen, pulse rate (referred to as pulse), respiratory rate and body temperature, and can also detect physiological parameter change information of the smart watch wearer. The system service layer may also include a step counting service, a calorie service, and a heart health service.

[0093] The algorithm layer can include a series of algorithm models. Fig.10 As shown, the algorithm layer may include a blood pressure algorithm model and a multi-source data alignment algorithm model. The blood pressure algorithm model is used to calculate the physiological parameters of the smart watch wearer, and the multi-source data alignment algorithm model is used to align the data collected by multi-source sensors. The algorithm layer may also include a sleep algorithm model, a wearing algorithm model, a dimming algorithm model, etc.

[0094] The hardware abstraction layer (HAL) is an interface layer between the operating system kernel and the hardware circuit. Fig.10 As shown, the HAL layer includes storage HAL, display HAL, touch HAL, Bluetooth HAL, and global positioning system (GPS) HAL. Among them, the storage HAL can be used to store and process the data collected by the sensor. The display HAL can be used to display information. The touch HAL can be used to detect user trigger operations. The Bluetooth HAL can be used to implement the Bluetooth communication function. The GPS HAL can be used to implement the positioning function. The HAL layer can be implemented through a C++ library.

[0095] like Fig.10 As shown, the kernel layer includes at least the operating system kernel.

[0096] Understandably, Fig.10 The layers in the software structure shown and the components contained in each layer do not constitute a specific limitation on the smart watch. In other embodiments of the present application, the smart watch may include more or fewer layers than shown in the figure, and each layer may include more or fewer components, which is not limited in the present application.

[0097] Fig.11 A hardware structure block diagram of an example of a cloud device provided in an embodiment of the present application. The cloud in the embodiment of the present application may be one or more server clusters. The cluster may include one or more cloud devices. The cloud device may be a cloud server or other device, which is not limited in the present application. Fig.11As shown, the cloud device 400 includes a processor 401, which further includes one or more processors, and a memory resource represented by a memory 402, for storing instructions that can be executed by the processor 401, such as an application. The application stored in the memory 402 may include one or more modules, each corresponding to a set of instructions. In addition, the processor 401 is configured to execute instructions to perform the blood pressure measurement method provided in the embodiment of the present application.

[0098] The cloud device 400 may further include a power supply component 403 configured to perform power management of the cloud device 400 , a wired or wireless network interface 404 configured to connect the cloud device 400 to a network, and an input / output interface 405 . The cloud device 400 may operate based on an operating system stored in the memory 402 .

[0099] The methods in the following embodiments can all be implemented in a wearable device and a cloud device having the above hardware structure. Figure 6 Taking the wearable device and cloud shown as an example, the specific implementation method of the embodiment of the present application is described in detail. Specifically, the blood pressure measurement process in the embodiment of the present application can be divided into two parts. The first part is the model generation scenario, that is, the scenario in which the cloud performs blood pressure prediction model training based on a large amount of sample data. The second part is the blood pressure prediction scenario, that is, the scenario in which the wearable device performs blood pressure measurement based on the blood pressure prediction model.

[0100] First, for the blood pressure prediction scenario, refer to Fig.12 As shown, taking the wearable device as a smart watch and the cloud device as a cloud server as an example, the blood pressure measurement method provided in the embodiment of the present application may include S501-S505: S501. In response to a blood pressure measurement operation of a user, the smart watch collects a PPG signal of the user through a PPG sensor to obtain PPG signal data.

[0101] The blood pressure measurement operation may be an operation in which the user triggers a blood pressure measurement button in the smart watch. The blood pressure measurement button may be a button in a blood pressure measurement application, or a button in other applications in the smart watch that can implement a blood pressure measurement function.

[0102] When a user needs to measure his or her own blood pressure, the user can use a smart watch and a mobile phone to measure the blood pressure. For example, the user can wear a smart watch and first use the smart watch to collect the user's physical characteristics, so as to obtain the blood pressure value according to the user's physical characteristics. Among them, the user's physical characteristics include PPG signals.

[0103] In some examples, the process of a smart watch collecting PPG signals through a PPG sensor may include: first, the user starts the blood pressure measurement function of the smart watch. Then, after the blood pressure measurement function is turned on, the smart watch may prompt the user to perform a first preset behavior action and detect whether the user has performed the first preset behavior action. Exemplarily, the first preset behavior action may be to wear the smart watch on a specified part of the user's body, such as a wrist. When it is detected that the user has performed the first preset behavior action, the smart watch controls the PPG sensor to collect PPG signals. It is understandable that when it is detected that the user has triggered the first preset behavior action, the smart watch can also prompt the user through the voice module to perform the first preset behavior action more standardizedly, for example, prompting the user to stay still.

[0104] S502: The smart watch sends a blood pressure measurement request to the cloud server through the wireless communication module.

[0105] The blood pressure measurement request is used to request the cloud server to determine the user's blood pressure value.

[0106] As a possible implementation method, after obtaining the user's PPG signal data, the smart watch can send a blood pressure measurement request containing the PPG signal data to the cloud server through the wireless communication module.

[0107] As another possible implementation method, after obtaining the user's PPG signal data, the smart watch can upload the PPG signal data to the cloud server through the wireless communication module, and then send a blood pressure measurement request to the cloud server.

[0108] In some implementations, in order to ensure the accuracy of the blood pressure value determined based on the PPG signal, after the PPG signal is collected, the embodiment of the present application can also perform a signal quality test on the PPG signal. If the PPG signal passes the signal quality test, the PPG signal can be further processed. If the PPG signal does not pass the signal quality test, it is necessary to execute S501 to S502 again to re-collect the PPG signal until the collected PPG signal meets the requirements of the signal quality test. In this way, it can be ensured that the signal quality and signal stability of the collected PPG signal meet the requirements.

[0109] S503: The cloud server inputs the PPG signal data into the trained blood pressure prediction model to obtain the blood pressure measurement result.

[0110] Among them, the cloud server is deployed with a trained blood pressure prediction model, which is trained based on a target training sample set. The target training samples are obtained by sampling the original training sample set. The difference between any two sample PPG signal data in the target training sample set is greater than a preset threshold, and the target training sample set contains the target sample PPG signal data, while the original training sample set does not contain the target sample PPG signal data.

[0111] As a possible implementation method, after receiving a blood pressure measurement request containing PPG signal data sent by a smart watch, the cloud server inputs the PPG signal data into a trained blood pressure prediction model to obtain a blood pressure measurement result.

[0112] As another possible implementation, after receiving the blood pressure measurement request sent by the smart watch, the cloud server inputs the PPG signal data previously uploaded by the smart watch into the trained blood pressure prediction model to obtain the blood pressure measurement result.

[0113] S504: The cloud server sends the blood pressure measurement result to the wireless communication module of the smart watch.

[0114] As a possible implementation method, the cloud server sends the blood pressure measurement results to the smart watch. Correspondingly, the smart watch receives the blood pressure measurement results sent by the cloud server through the wireless communication module and transmits the blood pressure measurement results to the display screen.

[0115] S505. The smart watch displays the blood pressure measurement result on the display screen.

[0116] Among them, the blood pressure measurement results include diastolic pressure and systolic pressure. The smart watch can display the blood pressure measurement results to the user through the display screen, and can also save the test time of the blood pressure measurement results.

[0117] As a possible implementation, the smart watch can display the blood pressure measurement result on the interface through the display screen. After the wearable device displays the blood pressure measurement result, it can wait for the user's operation, or display an option to measure again to facilitate measuring the blood pressure again, or display an option to exit to end the blood pressure measurement.

[0118] Optionally, if the smart watch detects that the user triggers the re-measurement option, the above S501 to S505 may be executed.

[0119] Through the above scheme, after obtaining the photoplethysmography PPG signal data of the user to be tested, the present application inputs the PPG signal data into the trained blood pressure prediction model to obtain the blood pressure measurement result of the user to be tested. Since the blood pressure prediction model of the present application is trained based on the target training sample set, rather than the original training sample set. The target training sample set is obtained after sampling the original training samples. The present application effectively screens and augments the original training sample set, so that the target training sample set is richer and more representative than the original training sample set. Therefore, compared with the related art that directly uses the original training samples to train the blood pressure prediction model, the present application can suppress the overfitting phenomenon of the blood pressure prediction model by optimizing the training samples, enhance the generalization of the blood pressure prediction model, and thus improve the accuracy of the blood pressure prediction model.

[0120] For blood pressure prediction scenarios, refer to Fig.13 As shown, taking the wearable device as a smart watch and the cloud device as a cloud server as an example, the blood pressure measurement method provided in the embodiment of the present application may also include S601-S604: S601. The cloud server sends the trained blood pressure prediction model to the blood pressure algorithm module of the smart watch.

[0121] As a possible implementation method, after the cloud server has trained the blood pressure prediction model, it can send the trained blood pressure prediction model to the smart watch. Correspondingly, the smart watch can store and deploy the blood pressure prediction model so that the user can call the blood pressure prediction model when measuring blood pressure.

[0122] In some embodiments, the cloud server may send the model parameters of the blood pressure prediction model to each smart watch, and each smart watch may update the model by updating the model parameters of the local blood pressure prediction model.

[0123] For example, Fig.14 The figure shows a schematic diagram of the distribution process of the exemplary blood pressure prediction model. After training the blood pressure prediction model, the cloud 62 can send the model parameters of the blood pressure prediction model to the smart watch 61, so that the smart watch 61 can measure the blood pressure based on the blood pressure prediction model.

[0124] S602: In response to the user's blood pressure measurement operation, the smart watch collects the user's PPG signal through the PPG sensor to obtain PPG signal data.

[0125] The specific implementation method of this step can refer to the above S501 and will not be repeated here.

[0126] S603: The smart watch inputs the PPG signal data into a blood pressure prediction model to obtain a blood pressure measurement result of the user to be tested.

[0127] As a possible implementation method, the smart watch inputs the PPG signal data into a locally deployed blood pressure prediction model to obtain the blood pressure measurement results of the user to be tested.

[0128] S604. The smart watch displays the blood pressure measurement result on the display screen.

[0129] The specific implementation method of this step can refer to the above S505 and will not be repeated here.

[0130] Combine the following Fig.15 The method flow chart shown in FIG. 1 is a schematic diagram for explaining the training process of the blood pressure prediction model in detail. Fig.15 , including: S701. The cloud server obtains the original training sample set.

[0131] Among them, the original training sample set can be obtained by collecting multiple real sample users, and the original training sample set includes the blood pressure values ​​of multiple real sample users and the PPG signal data corresponding to each blood pressure value.

[0132] As a possible implementation method, the cloud server can obtain the data of these real sample users from the database and use these data as the original training sample set.

[0133] S702: The cloud server samples the original training sample set to obtain a target training sample set.

[0134] As a possible implementation, the cloud server may perform a first sampling on the original training sample set based on a first sampling strategy to obtain a first training sample set. The number of sample PPG signal data in the first training sample set is less than the number of sample PPG signal data in the original training sample, and the degree of difference between any two sample PPG signal data in the first training sample set is greater than a preset threshold. Further, the cloud server performs a second sampling on the first training sample set based on a second sampling strategy to obtain a second training sample set, wherein the second training sample set contains target sample PPG signal data, and the original training sample set and the first training sample set do not contain target sample PPG signal data. The cloud server may use the second training sample set as the target training sample set.

[0135] It can be understood that the cloud server first performs the first sampling of the original training sample set through the first sampling strategy. Selecting samples with large differences can reduce the redundant information between samples, avoid overfitting due to too much redundant information, and cover the original feature space as much as possible without losing personal information, which can fully improve the training efficiency. At the same time, when processing unbalanced data sets, it can help select minority class samples and improve the model's recognition ability for minority classes. Furthermore, the cloud server performs the second sampling of the first training sample set through the second sampling strategy to fill in some samples that are missing in the original training sample set, thereby enhancing the generalization of the model in subsequent training.

[0136] As another possible implementation, the cloud server may perform a first sampling on the original training sample set based on a first sampling strategy to obtain a first training sample set, wherein the number of sample PPG signal data in the first training sample set is less than the number of sample PPG signal data in the original training sample, and the difference between any two sample PPG signal data in the first training sample set is greater than a preset threshold. Further, the cloud server performs a second sampling on the original training sample set based on a second sampling strategy to obtain a second training sample set, wherein the second training sample set contains target sample PPG signal data, and the original training sample set does not contain target sample PPG signal data. The cloud server uses the first training sample set and the second training sample set as target training sample sets.

[0137] It is understandable that the cloud server can perform first sampling and second sampling on the original training samples. The first sampling can select samples with large differences to reduce redundant information between samples and avoid overfitting due to excessive redundant information. The first sampling can fill in some samples that are missing in the original training sample set. Furthermore, the cloud server uses the training sample set with two advantages as the target training sample set, which can better enhance the generalization ability of the model.

[0138] Optionally, the first sampling strategy can be any downsampling strategy, for example, the first sampling strategy includes diversity sampling. The second sampling strategy can be any upsampling strategy, for example, the second sampling strategy includes interpolation sampling. Diversity sampling aims to select samples with different characteristics or attributes from the data set to ensure that the machine learning model can learn the data distribution comprehensively and evenly. These selected samples may have features that are rare in the training data, or they may represent real statistical data that are currently underrepresented in the model. Through diversity sampling, the generalization ability and robustness of the model can be improved. Interpolation sampling refers to the process of estimating unknown data points through a certain algorithm or model based on known data points.

[0139] It can be understood that the embodiment of the present application aims at the technical defects of the original training sample set, and effectively screens and expands the original training sample set by designing a targeted sampling strategy and performing targeted sampling processing on the original training sample set, so that the target training sample set is richer and more representative than the original training sample set. Therefore, compared with the related art of directly training the blood pressure prediction model with the original training samples, the present application can suppress the overfitting phenomenon of the blood pressure prediction model by optimizing the training samples, enhance the generalization of the blood pressure prediction model, and thus improve the accuracy of the blood pressure prediction model.

[0140] The following takes the first sampling strategy of diversity sampling and the second sampling strategy of interpolation sampling as an example to introduce the specific first sampling strategy and the second sampling strategy in detail.

[0141] In some embodiments, for the diversity sampling strategy, the cloud server needs to classify the PPG signal data in the original training sample set before performing the first sampling on the original training sample set based on the diversity sampling strategy, and then perform diversity sampling on the PPG signal data in each classification result.

[0142] For example, the cloud server can classify the PPG signal data in the original training sample set according to the users corresponding to the PPG signal data, so that the PPG signal data of each sample user can be separated. Furthermore, the PPG signal data of each sample user is sampled in diversity.

[0143] As a possible implementation method, the cloud server can input the sample PPG signal data in the original training sample set into the trained classification model to obtain the classification results of the sample PPG signal data, wherein the classification model is used to extract the identity features of each sample PPG signal data, and classify each sample PPG signal data according to the identity features; the sample PPG signal data of the same category in the classification result corresponds to the same sample user. Further, the cloud server can perform the first sampling in the sample PPG signal data of each category based on the first sampling strategy to obtain the first training sample set.

[0144] The specific classification model is introduced below. In order to classify the PPG signal data in the original training sample set, the cloud server can train a classification model. Specifically, the cloud server can first divide the PPG signal data in a large-scale data set (such as a public data set) according to individuals and classify the individuals. Furthermore, the cloud server can use a feature extractor to extract individual identity-related features as an individual identity recognition classifier based on PPG signal data.

[0145] The feature extractor may be a small-scale simple feature extractor, such as a simple multi-layer stack of convolutional neural networks (CNNs).

[0146] During the training process, since it is assumed that the test set individuals have never been seen, and they are required to be classified (i.e., the zero-shot problem), when designing the loss function, the embodiment of the present application can use the cosine margin normalized loss function (Angular-margin Softmax) based on elasticity penalty to replace the normalized exponential function (Softmax) commonly used in classification tasks. Its purpose is to make the identity encoding in the latent space divisible, more compact within the class, more distinguishable between classes, and elastic and flexible through additive angular margin penalties. For example, the loss function formula can be designed as follows:

[0147] Among them, θ is the angle between the weight vector W of the fully connected layer in the classifier network and the feature vector x. The angle constraint generates a margin between the decision boundaries. The larger the margin, the stricter the classification boundary. S is a fixed scale factor used to enlarge the feature space, E is a random Gaussian elastic penalty added, m is the mean, and σ is the standard deviation.

[0148] It can be understood that Angular-margin Softmax strengthens the classification effect by strengthening the angle constraint between weights and feature vectors, and adds elastic penalties to assign different margins to each sample. The purpose is to derive the class center with a larger penalty for samples that are farther away, and give less penalty attention to samples that are closer.

[0149] Exemplary, reference Fig.16 The training set of the classification model can include the training set PPG data and the individual category labels of the training set. The feature extractor of the classification model is composed of multiple convolutional layers (Conv) stacked together. After the cloud server inputs the training set PPG data into the feature extractor, the feature extractor embeds the training set PPG data to extract the individual identity features corresponding to the training set PPG data, and uses the loss function Angular-margin Softmax as the evaluation basis to modify the parameters of each convolutional layer.

[0150] like Fig.17As shown in the figure, after the classification model training is completed, the cloud server can also test the classification effect of this classification model on the test set. The cloud server can use the dimension reduction visualization algorithm (such as the tSNE algorithm) to reduce the dimension of the identity coding features extracted by the feature extractor and visualize them, and then obtain the degree of clustering of the identity coding features from the same individual. In addition, the cloud server can use the random forest machine learning algorithm to perform classification tests on the identity coding features to complete the classification effect confirmation of the classifier on the test set. Among them, the random forest machine learning algorithm is an ensemble learning method that implements classification and regression tasks by constructing multiple decision trees.

[0151] After obtaining the classification results of the sample PPG signal data through the above classification model, the cloud server can perform diversity sampling on the sample PPG signal data of each category, that is, diversity sampling is performed in one classification result. Among them, diversity sampling can define the difference based on the cosine similarity distance, and select samples that cover the feature space as much as possible and have large differences through a greedy algorithm.

[0152] As a possible implementation, for any category of sample PPG signal data, the cloud server can collect a target number of sample PPG signal data from the sample PPG signal data of the current category to obtain a first sampling data set. The cosine similarity between the identity features of any two sample PPG signal data in the first sampling data set is greater than a first preset similarity. Further, the cloud server uses each first sampling data set as a first training sample set.

[0153] Specifically, for any sample PPG signal data that has been collected, the cloud server can respectively calculate the cosine similarity between the uncollected sample PPG signal data and the collected sample PPG signal data. Furthermore, the cloud server can add the sample PPG signal data with the largest cosine similarity among the uncollected sample PPG signal data to the first sampling data set.

[0154] It is understandable that the diversity sampling strategy can reduce the redundant information between samples by selecting samples with large differences, avoid overfitting due to too much redundant information, and cover the original feature space as much as possible without losing personal information, thereby fully improving the training efficiency. In addition, the diversity sampling strategy can also help select minority class samples when processing unbalanced data sets, improving the model's ability to recognize minority classes.

[0155] From the classification results of the above classification model, it can be seen that there are obvious gaps (GAPs) between the clustering results, and this GAP corresponds to individual samples that the blood pressure prediction model has not seen. Therefore, the existence of GAPs in the training stage will affect the generalization of the model. Therefore, the embodiment of the present application needs to fill the GAPs by interpolating between the clustering results through an interpolation sampling strategy to enhance the generalization of the blood pressure prediction model. The specific interpolation sampling strategy is introduced below.

[0156] In some embodiments, in order to fill the gap between clustering results, the cloud server may input the first training sample set obtained by the diversity sampling strategy into the trained sampling model to obtain the second training sample set.

[0157] like Fig.18 As shown, the sampling model includes an encoder, a sampler, and a decoder. The encoder is used to encode the sample PPG signal data in the first training sample set according to the sample user to which it belongs, to obtain multiple first codes. One first code corresponds to the sample PPG signal data of a sample user in the first training sample set. The sampler is used to perform interpolation sampling between the multiple first codes to obtain multiple second codes; the decoder is used to decode the multiple second codes to obtain the second training sample set.

[0158] The cloud server can perform interpolation sampling between multiple first codes through a sampler to obtain multiple second codes. Fig.19 As shown, the cloud server can construct a data structure diagram corresponding to multiple first codes by adopting a model. Wherein, the data structure diagram includes multiple nodes, one node corresponds to one first code, each node is connected by an edge, and the length of the edge between any two nodes is used to reflect the coding similarity between the two nodes. Further, the cloud server determines the candidate node pairs in the data structure diagram by adopting a model, wherein the candidate node pairs are any two nodes in the data structure diagram whose edge lengths are greater than the preset edge lengths. Then, the model can be used to determine the target node pairs based on the candidate node pairs, and insert new nodes between the target node pairs, wherein the target node pairs are candidate node pairs where no other candidate nodes exist between any two candidate nodes.

[0159] The interpolation process of the sampling model is as follows: First, select individual representative pairs according to the distance judgment (individual representatives can be generated by clustering all feature samples of the individual). For individual representative pairs that meet sufficient distance, each time interpolation is performed, a strategy is formulated to determine whether other individuals already exist between individual clusters. If so, it means that the GAP has been filled and no interpolation processing is required; if not, it means that there is a significant gap, and the interpolation formula is used for significant interpolation sampling. The interpolation formula can be: z=(1-α)*z1+α*z2, where z is the newly inserted value, z1 and z2 are a pair of individual representative pairs that meet sufficient distance, and α is a constant between 0 and 1.

[0160] As a possible implementation, the cloud server can determine the code corresponding to the node to be inserted according to the code corresponding to the first target node and the code corresponding to the second target node through a sampling model, and obtain the node to be inserted, wherein the first target node and the second target node are any two target node pairs. Further, the cloud server inserts the node to be inserted between the first target node and the second target node.

[0161] For example, the cloud server may perform weighted processing on the code corresponding to the first target node and the code corresponding to the second target node to obtain the node to be removed.

[0162] Fig.19 The line between any two points shown in the figure represents the candidate individual representative pairs selected based on the distance judgment. The solid line indicates that there are no other individual samples between the individual clusters, indicating that there is a significant gap, and the interpolation formula is used for significant interpolation and sampling. The dotted line indicates that there are other individual samples between the individuals (represented by the white ball individuals in this example), and no interpolation is required.

[0163] Fig. 20 The effect diagram after sampling by sampling model interpolation is shown. Fig. 20 a represents the training sample set without interpolation sampling. Fig. 20 The white nodes in b represent the target nodes that meet the interpolation conditions. Fig. 20 c represents the training sample set after inserting nodes between target nodes. Fig. 20 It can be seen that for the input first training sample set, the sampling model performs interpolation sampling on the target nodes therein and inserts new nodes between its target nodes, so that the number of nodes in the output second training sample set is greater than the number of nodes in the input first training sample set, thereby achieving the goal of supplementing the first training sample set.

[0164] It can be understood that after the sampling model collects all the representative pairs of individuals that meet the conditions, it obtains new feature vectors and individual label encodings through significant interpolation, and generates a large number of interpolated sampling samples through the decoder of the sampling model. Furthermore, the cloud server can use the samples sampled by the above two sampling strategies as training sample sets to train the blood pressure prediction model.

[0165] S703: The cloud server constructs a blood pressure prediction model and trains the model using a target training sample set.

[0166] The constructed blood pressure prediction model can be any neural network model, such as a Transformer model. The constructed blood pressure prediction model consists of multiple hidden layers, each of which includes linear and nonlinear processing units, which are connected in an end-to-end manner to achieve the task of blood pressure prediction.

[0167] As a possible implementation method, after constructing the network structure of the blood pressure prediction model, the cloud server takes the sample PPG signal data in the target training sample set as input, and the blood pressure value corresponding to each sample PPG signal data as a label, and trains the model until the number of training times reaches a preset number, and determines that the training is completed.

[0168] It can be understood that the blood pressure prediction model trained by the above embodiment is deployed on the cloud to predict the blood pressure of the wearable device user. Different from the traditional blood pressure prediction model, the blood pressure prediction model of the present application benefits from the overfitting suppression effect of the sampling strategy. Compared with the model of the related technology, the model has stronger generalization and higher prediction accuracy for unseen individual data.

[0169] In some embodiments, the training process of the blood pressure prediction model of the present application includes two parts: metamodel training and calibration. Among them, the metamodel is a general model trained based on a large batch of data (such as a public data set). Calibration is to perform personalized calibration of the metamodel based on a small batch of data of the individual to be tested. After the cloud server trains the metamodel using the above method, it can also calibrate the trained metamodel according to a specific user group, thereby improving the blood pressure prediction ability of the blood pressure prediction model for a specific user group.

[0170] For example, for some older user groups, their blood pressure is more likely to be higher than that of younger users. Therefore, after training the meta-model, the cloud server can fine-tune the output of the meta-model upward to achieve a better effect of adapting to the older user groups.

[0171] It is understandable that the embodiment of the present application improves the training samples so that the improved training samples are more representative and comprehensive, and thus the generalization of the metamodel can be optimized to improve the recognition effect of the metamodel. In the subsequent fine-tuning process of the metamodel, the difficulty of fine-tuning is reduced and the pressure of fine-tuning work is alleviated.

[0172] In addition, in order to realize the above functions, the wearable device includes a hardware structure or software module corresponding to each function, or a combination of the two. Those skilled in the art should easily realize that, in combination with the units and algorithm steps of each example described in the embodiments disclosed herein, the present application can be implemented in the form of hardware or a combination of hardware and computer software. Whether a function is executed in the form of hardware or computer software driving hardware depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0173] The embodiment of the present application can divide the functional modules of the wearable device according to the above method example. For example, each functional module can be divided corresponding to each function, or two or more functions can be integrated into one processing module. The above integrated modules can be implemented in the form of hardware or in the form of software functional modules. It should be noted that the division of modules in the embodiment of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation. The following is an example of dividing each functional module corresponding to each function.

[0174] Fig.21 This is a schematic diagram of the structure of a wearable device provided in an embodiment of the present application. Fig.21 As shown, the device may include a collection unit 801 and a processing unit 802 .

[0175] The acquisition unit 801 is used to acquire the user's photoplethysmography (PPG) signal in response to the user's blood pressure measurement operation to obtain PPG signal data.

[0176] The processing unit 802 is used to obtain the user's blood pressure measurement result based on the PPG signal data and the trained blood pressure prediction model, and display the blood pressure measurement result; the blood pressure prediction model is trained based on a target training sample set, the target training sample is obtained by sampling the original training sample set, the difference between any two sample PPG signal data in the target training sample set is greater than a preset threshold, and the target training sample set contains the target sample PPG signal data, and the original training sample set does not contain the target sample PPG signal data.

[0177] This embodiment also provides a computer storage medium, in which computer instructions are stored. When the computer instructions are executed on an electronic device, the electronic device executes the above-mentioned related method steps to implement the blood pressure measurement method in the above-mentioned embodiment.

[0178] This embodiment also provides a computer program product. When the computer program product is run on a computer, the computer is caused to execute the above-mentioned related steps to implement the blood pressure measurement method in the above-mentioned embodiment.

[0179] In addition, an embodiment of the present application also provides a device, which can specifically be a chip, component or module, and the device may include a connected processor and memory; wherein the memory is used to store computer-executable instructions, and when the device is running, the processor can execute the computer-executable instructions stored in the memory so that the chip executes the blood pressure measurement method in the above-mentioned method embodiments.

[0180] Among them, the electronic devices (such as wearable devices, cloud devices, etc.), computer storage media, computer program products or chips provided in this embodiment are all used to execute the corresponding methods provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding methods provided above, and will not be repeated here.

[0181] Through the description of the above implementation methods, technical personnel in the relevant field can understand that for the convenience and simplicity of description, only the division of the above-mentioned functional modules is used as an example. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

[0182] In the several embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0183] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A blood pressure measurement method, characterized in that: Applied to a wearable device, the method comprises: In response to the user's blood pressure measurement operation, collecting the user's photoplethysmography (PPG) signal to obtain PPG signal data; Based on the PPG signal data and the trained blood pressure prediction model, the user's blood pressure measurement result is obtained and displayed; the blood pressure prediction model is trained based on a target training sample set, the target training sample is obtained by sampling an original training sample set, the difference between any two sample PPG signal data in the target training sample set is greater than a preset threshold, and the target training sample set contains the target sample PPG signal data, while the original training sample set does not contain the target sample PPG signal data.

2. The method according to claim 1, characterized in that The wearable device is equipped with the blood pressure prediction model; and obtaining the blood pressure measurement result of the user based on the PPG signal data and the trained blood pressure prediction model includes: The PPG signal data is input into the blood pressure prediction model to obtain the user's blood pressure measurement result.

3. The method according to claim 1, characterized in that The obtaining of the user's blood pressure measurement result based on the PPG signal data and the trained blood pressure prediction model includes: Sending a blood pressure measurement request including the PPG signal data to a cloud; the blood pressure prediction model is deployed on the cloud; Receive the blood pressure measurement result sent by the cloud; the blood pressure measurement result is obtained by the cloud inputting the PPG signal data into the blood pressure prediction model.

4. A blood pressure measurement method, characterized in that: Applied to the cloud, the method includes: Receive a blood pressure measurement request including PPG signal data sent by a wearable device; The PPG signal data is input into a trained blood pressure prediction model to obtain a blood pressure measurement result, and the blood pressure measurement result is sent to the wearable device; the blood pressure prediction model is trained based on a target training sample set, the target training sample is obtained by sampling an original training sample set, the difference between any two sample PPG signal data in the target training sample set is greater than a preset threshold, and the target training sample set contains the target sample PPG signal data, and the original training sample set does not contain the target sample PPG signal data.

5. The method according to claim 4, characterized in that The method further comprises: Obtaining the original training sample set; Sampling the original training sample set to obtain the target training sample set; The sample PPG signal data in the target training sample set is used as input, and the blood pressure corresponding to the sample PPG signal data in the target training sample set is used as a label to train and obtain the blood pressure prediction model.

6. The method according to claim 5, characterized in that The sampling of the original training sample set to obtain the target training sample set includes: Performing a first sampling on the original training sample set based on a first sampling strategy to obtain a first training sample set; the number of sample PPG signal data in the first training sample set is less than the number of sample PPG signal data in the original training sample, and the difference between any two sample PPG signal data in the first training sample set is greater than the preset threshold; Performing a second sampling on the first training sample set based on a second sampling strategy to obtain a second training sample set; the second training sample set includes the target sample PPG signal data, and the original training sample set and the first training sample set do not include the target sample PPG signal data; The second training sample set is used as the target training sample set.

7. The method according to claim 5, characterized in that The sampling of the original training sample set to obtain the target training sample set includes: Performing a first sampling on the original training sample set based on a first sampling strategy to obtain a first training sample set; the number of sample PPG signal data in the first training sample set is less than the number of sample PPG signal data in the original training sample, and the difference between any two sample PPG signal data in the first training sample set is greater than the preset threshold; Performing a second sampling on the original training sample set based on a second sampling strategy to obtain a second training sample set; the second training sample set includes the target sample PPG signal data, and the original training sample set does not include the target sample PPG signal data; The first training sample set and the second training sample set are used as the target training sample sets.

8. The method according to claim 6 or 7, characterized in that: The first sampling strategy includes diversity sampling, and the second sampling strategy includes interpolation sampling.

9. The method according to claim 6, characterized in that The first sampling of the original training sample set based on the first sampling strategy to obtain the first training sample set includes: The sample PPG signal data in the original training sample set is input into the trained classification model to obtain the classification result of the sample PPG signal data; the classification model is used to extract the identity features of each sample PPG signal data, and classify each sample PPG signal data according to the identity features; the sample PPG signal data of the same category in the classification result corresponds to the same sample user; Based on the first sampling strategy, first sampling is performed in the sample PPG signal data of each category to obtain the first training sample set.

10. The method according to claim 9, characterized in that The first sampling is performed in each category of sample PPG signal data based on the first sampling strategy to obtain the first training sample set, including: For any category of sample PPG signal data, a target number of sample PPG signal data are collected from the sample PPG signal data of the current category to obtain a first sampling data set; the cosine similarity between the identity features of any two sample PPG signal data in the first sampling data set is greater than a first preset similarity; Each first sampling data set is used as the first training sample set.

11. The method according to claim 10, characterized in that The step of collecting a target number of sample PPG signal data from the sample PPG signal data of the current category to obtain a first sample data set includes: For any collected sample PPG signal data, respectively calculating the cosine similarity between the uncollected sample PPG signal data and the collected sample PPG signal data; The sample PPG signal data with the largest cosine similarity among the uncollected sample PPG signal data are added to the first sampling data set.

12. The method according to claim 9, characterized in that The method further comprises: Dividing the sample PPG signal data in the original training sample set according to the sample users to which they belong, to obtain sample PPG signal data of multiple sample users; The sample PPG signal data of the multiple sample users are taken as input, and the cosine residual normalization loss function Angular-margin Softmax is set to train the classification model; the Angular-margin Softmax is used to instruct the classification model to maximize the probability of the category to which it belongs and minimize the probability gap with the adjacent category.

13. The method according to claim 6, characterized in that The performing a second sampling on the first training sample set based on a second sampling strategy to obtain a second training sample set includes: The first training sample set is input into the trained sampling model to obtain the second training sample set.

14. The method according to claim 13, characterized in that The sampling model includes an encoder, a sampler, and a decoder; the first training sample set is input into the trained sampling model to obtain the second training sample set, including: The encoder encodes the sample PPG signal data in the first training sample set according to the sample users to which they belong, obtaining a plurality of first codes; one first code corresponds to the sample PPG signal data of a sample user in the first training sample set; Performing interpolation sampling between the plurality of first codes by the sampler to obtain a plurality of second codes; The plurality of second codes are decoded by the decoder to obtain the second training sample set.

15. The method according to claim 14, characterized in that The interpolation sampling between the plurality of first codes by the sampler to obtain the plurality of second codes comprises: Constructing a data structure graph corresponding to the multiple first codes; the data structure graph includes multiple nodes, one node corresponds to one first code, the nodes are connected by edges, and the length of the edge between any two nodes is used to reflect the coding similarity between the two nodes; Determine a candidate node pair in the data structure graph; the candidate node pair is any two nodes in the data structure graph whose side lengths are greater than a preset side length; According to the candidate node pairs, a target node pair is determined, and a new node is inserted between the target node pairs; the target node pair is a candidate node pair where no other candidate nodes exist between any two candidate nodes.

16. The method according to claim 15, characterized in that The inserting a new node between the target node pair comprises: According to the code corresponding to the first target node and the code corresponding to the second target node, the code corresponding to the node to be inserted is determined to obtain the node to be inserted; the first target node and the second target node are any two target node pairs; The node to be removed is inserted between the first target node and the second target node.

17. The method according to claim 16, characterized in that The step of determining the code corresponding to the node to be inserted according to the code corresponding to the first target node and the code corresponding to the second target node to obtain the node to be inserted includes: The code corresponding to the first target node and the code corresponding to the second target node are weighted to obtain the node to be removed.

18. A wearable device, characterized in that: include: one or more processors; Memory; And one or more computer programs, wherein the one or more computer programs are stored in the memory, and when the computer programs are executed by the one or more processors, the wearable device performs the blood pressure measurement method according to any one of claims 1 to 3.

19. A cloud device, characterized in that: include: one or more processors; Memory; and one or more computer programs, wherein the one or more computer programs are stored in the memory, and when the computer programs are executed by the one or more processors, the cloud device executes the blood pressure measurement method as described in any one of claims 4-17.

20. A computer-readable storage medium comprising a computer program, characterized in that: When the computer program is executed on an electronic device, the electronic device executes the blood pressure measurement method according to any one of claims 1 to 3 or the blood pressure measurement method according to any one of claims 4 to 17.

Images