Noninvasive blood glucose monitoring method and system based on PPG signal
Through intelligent wearable devices, collect and analyze PPG signals, combine user physiological information, build a personalized physiological characteristic model, and push customized blood sugar management strategies, solving the trauma and noise interference problems of traditional blood sugar monitoring, and realizing non-invasive and personalized blood sugar monitoring and health management.
Patent Information
- Application Number
- CN202411051006.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-01
- Publication Date
- 2024-11-15
AI Technical Summary
Traditional blood sugar monitoring methods are invasive and cumbersome, which can easily cause patient discomfort and infection risks. The PPG signal of the existing non-invasive methods is susceptible to noise interference and lacks comprehensive considerations for users' personalized physiological characteristics and health assessment.
By collecting PPG signals from smart wearable devices, filtering and analyzing, and building a personalized physiological characteristic model based on user basic physiological information, determining health assessment results and pushing customized blood sugar health management strategies.
It realizes non-invasive and continuous blood sugar monitoring, provides personalized health management solutions, improves user experience and health management levels, reduces noise interference, and enhances monitoring accuracy.
Smart Images

Figure CN118948271A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of intelligent health monitoring, and in particular to a non-invasive blood glucose monitoring method and system based on PPG signals. Background Art
[0002] Traditional blood glucose monitoring methods rely on invasive blood sampling, which is not only cumbersome but also prone to cause patient discomfort and infection risks. With the development of health monitoring technology, non-invasive blood glucose monitoring technology has become a research hotspot. A non-invasive blood glucose monitoring method based on photoplethysmography (PPG) signals has been proposed. However, PPG signals are susceptible to noise interference, which affects monitoring accuracy, and existing technologies mostly focus on monitoring a single blood glucose value, lacking comprehensive consideration of users' personalized physiological characteristics and health assessments. Summary of the invention
[0003] The object of the present invention is to provide a non-invasive blood glucose monitoring method and system based on PPG signals.
[0004] In a first aspect, an embodiment of the present invention provides a non-invasive blood glucose monitoring method based on a PPG signal, characterized in that it includes:
[0005] Collect the original PPG signal of the target user from the smart wearable device equipped with an optical sensor;
[0006] Filtering the original PPG signal to obtain a target PPG signal;
[0007] Analyzing the target PPG signal to obtain a blood glucose characterization result of the target PPG signal;
[0008] Acquire pre-stored basic physiological information of the user from the smart wearable device, and construct a physiological feature of the user based on the blood glucose characterization result and the basic physiological information of the user;
[0009] The user health assessment result and population category information of the target user are determined according to the user physiological characteristics, and a blood sugar health management strategy is pushed to the target user according to the user health assessment result and the population category information.
[0010] In a second aspect, an embodiment of the present invention provides a server system, including a server, wherein the server is used to execute the method described in the first aspect.
[0011] Compared with the prior art, the beneficial effects provided by the present invention include: the present invention discloses a non-invasive blood glucose monitoring method and system based on PPG signals, including: firstly collecting the original PPG signal in the smart wearable device, obtaining the target PPG signal through filtering, analyzing the blood glucose characterization results, and building a personalized physiological characteristic model in combination with the user's basic physiological information. Furthermore, based on the model, the user's health assessment results and population categories are determined, and a customized blood glucose health management strategy is pushed. Such a design not only realizes non-invasive and continuous blood glucose monitoring, but also provides a personalized health management plan by integrating the user's physiological characteristics and health assessment, thereby improving the user experience and health management level. BRIEF DESCRIPTION OF THE DRAWINGS
[0012] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only illustrate certain embodiments of the present invention and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can also be obtained based on these drawings without creative work.
[0013] Figure 1 A schematic diagram of the steps of a non-invasive blood glucose monitoring method based on a PPG signal provided in an embodiment of the present invention;
[0014] Figure 2 A schematic block diagram of the structure of a computer device provided in an embodiment of the present invention. DETAILED DESCRIPTION
[0015] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.
[0016] The specific implementation modes of the present invention are described in detail below in conjunction with the accompanying drawings.
[0017] In order to solve the technical problems in the aforementioned background technology, Figure 1 The present invention provides a flow chart of a non-invasive blood glucose monitoring method based on a PPG signal according to an embodiment of the present invention. The non-invasive blood glucose monitoring method based on a PPG signal is introduced in detail below.
[0018] Step S201, collecting the original PPG signal of the target user from a smart wearable device equipped with an optical sensor;
[0019] Step S202, filtering the original PPG signal to obtain a target PPG signal;
[0020] Step S203, analyzing the target PPG signal to obtain a blood glucose characterization result of the target PPG signal;
[0021] Step S204, obtaining pre-stored basic physiological information of the user from the smart wearable device, and constructing a physiological feature of the user based on the blood glucose characterization result and the basic physiological information of the user;
[0022] Step S205, determining the user health assessment result and population category information of the target user according to the user physiological characteristics, and pushing a blood sugar health management strategy to the target user according to the user health assessment result and the population category information.
[0023] In an embodiment of the present invention, exemplarily, the server receives data from a smart wearable device (such as a smart watch or a health monitoring bracelet) paired with a target user. The device has a built-in optical sensor that can continuously collect PPG signals at the user's wrist. For example, user Zhang San wears a smart watch with an integrated PPG sensor, which regularly sends the collected raw PPG signal data to the server. After receiving the raw PPG signal, the server applies a digital filtering algorithm (such as a low-pass filter) to remove noise and interference, and retains the effective signal part reflecting the change in blood volume. For example, the server uses a Butterworth low-pass filter to process the raw PPG signal sent by Zhang San's watch, removes high-frequency noise, and obtains a clear target PPG signal. The server extracts and analyzes the filtered target PPG signal, uses a machine learning model (such as a deep learning network) to identify features related to blood sugar levels, and outputs blood sugar characterization results. For example, the server uses a trained neural network model to analyze Zhang San's PPG signal and predicts a preliminary estimate of the blood sugar level. The server retrieves and obtains pre-stored basic physiological information of the user, such as age, gender, height, weight, etc., from the smart wearable device or the user's health record. The server then combines this information with the blood glucose characterization results to construct a complete set of user physiological characteristics. For example, the server obtains basic information such as Zhang San's age and gender, and combines it with the estimated blood glucose value to form Zhang San's physiological characteristic vector. Based on the user's physiological characteristics, the server uses a risk assessment model to assess the user's health status and classifies the user into a specific population category (such as low risk, medium risk, and high risk). For example, based on Zhang San's physiological characteristics, the server determines that he belongs to the "medium risk" population and gives a corresponding health assessment report. Finally, the server customizes personalized blood glucose health management strategies based on the user's health assessment results and population category information, and pushes them to the user through smart wearable devices or mobile applications. For example, the server generates a blood glucose health management plan for Zhang San that includes diet recommendations, exercise plans, and regular monitoring reminders, and displays it to Zhang San through the application interface of the smart watch.
[0024] In the embodiment of the present invention, the aforementioned step S203 can be implemented through the following example execution.
[0025] Performing spectrum feature extraction processing on the target PPG signal according to a first feature extraction network to obtain spectrum features;
[0026] Loading the spectrum features into a feature integration network for feature integration processing to obtain blood glucose features;
[0027] Performing feature reconstruction processing on the blood glucose feature according to the feature inference network to obtain a blood glucose characterization result of the target PPG signal;
[0028] Among them, the first feature extraction network, the feature integration network and the feature inference network are obtained according to the target error integrated learning, and the target error includes a first error and a second error; the first error is obtained according to the first blood glucose feature instance and the second blood glucose feature instance, the first blood glucose feature instance is a PPG signal instance that is subjected to spectral feature extraction processing by the first feature extraction network and then subjected to feature integration processing by the feature integration network, and the second blood glucose feature instance is obtained by performing blood glucose feature extraction processing on the blood glucose characterization instance matched with the PPG signal instance; the second error is obtained according to the target value of the relationship between the inferred blood glucose characterization result and the PPG signal, and the inferred blood glucose characterization result is obtained by performing feature reconstruction processing on the first blood glucose feature instance by the feature inference network.
[0029] In an embodiment of the present invention, exemplarily, the server receives a filtered target PPG signal from a smart wearable device. For example, the server receives filtered PPG signal data sent by Zhang San's smart watch. The server uses a trained first feature extraction network (such as a part of a convolutional neural network) to extract spectral features of the target PPG signal. This network can extract spectral features related to blood glucose from the PPG signal. For example, the server extracts the spectral features of Zhang San's PPG signal through operations such as convolutional layers and pooling layers. The server inputs the extracted spectral features into a feature integration network (such as a fully connected neural network) for further feature integration processing to obtain more abstract and advanced blood glucose features. For example, the server integrates the spectral features into a feature vector related to blood glucose through a fully connected layer. The server uses a feature inference network (such as another fully connected neural network or a regression model) to perform feature reconstruction processing on the integrated blood glucose features, and finally obtains the blood glucose characterization result of the target PPG signal. For example, the server outputs the blood glucose estimation value corresponding to Zhang San's PPG signal through a regression layer. In the training stage, the server uses a large number of PPG signal instances and corresponding blood glucose characterization instances to train the first feature extraction network, the feature integration network, and the feature inference network. During the training process, the server calculates two errors: the first error and the second error. First error: The server first extracts the spectrum features of the PPG signal instance through the first feature extraction network, and then integrates the features through the feature integration network to obtain the first blood glucose feature instance. At the same time, the server extracts the blood glucose features of the blood glucose characterization instance that matches the PPG signal instance to obtain the second blood glucose feature instance. The first error is the difference between the two. Second error: The server uses the feature inference network to reconstruct the features of the first blood glucose feature instance to obtain the inferred blood glucose characterization result. The second error is the difference between the inferred blood glucose characterization result and the target value of the PPG signal relationship (i.e., the actual blood glucose value). The server performs integrated learning based on these two errors and optimizes the network parameters to improve the accuracy of blood glucose characterization. Through the above steps, the server can accurately analyze the blood glucose characterization results from the target PPG signal and provide basic data support for subsequent blood glucose health management.
[0030] In an embodiment of the present invention, the integrated learning process of the first feature extraction network, the feature integration network and the feature inference network can be implemented through the following examples.
[0031] Acquire a sample instance data set, wherein the sample instance data set includes a plurality of PPG signal instances and a PPG signal relationship target value matched with each PPG signal instance;
[0032] Performing spectrum feature extraction processing on the PPG signal instance according to the first feature extraction network to obtain a spectrum feature instance, and performing feature integration processing on the spectrum feature instance according to the feature integration network to obtain the first blood glucose feature instance;
[0033] performing PPG signal analysis on the PPG signal instance to obtain a matching blood glucose characterization instance, and performing blood glucose feature extraction processing on the blood glucose characterization instance according to a second feature extraction network to obtain the second blood glucose feature instance, wherein the initial feature of the second feature extraction network has the same feature dimension as the result feature of the feature inference network;
[0034] Calculate a time series regression error according to the first blood sugar characteristic instance and the second blood sugar characteristic instance to obtain a first error;
[0035] performing feature reconstruction processing on the first blood glucose feature instance according to the feature inference network to obtain a first inferred blood glucose characterization result, and obtaining a second error according to a target value of a relationship between the first inferred blood glucose characterization result and the PPG signal;
[0036] The network parameters of the first feature extraction network, the feature integration network and the feature inference network are optimized according to the first error and the second error.
[0037] In an embodiment of the present invention, exemplarily, the server obtains a sample instance data set containing multiple PPG signal instances and their matching PPG signal relationship target values (i.e., real blood glucose values) from a database. For example, the server loads a data set containing PPG signal instances and corresponding blood glucose values of 1,000 different users. The server uses a first feature extraction network to extract spectrum features from each PPG signal instance to obtain a spectrum feature instance. For example, the server performs spectrum analysis on the PPG signal through a convolutional neural network to extract spectrum features. Then, the server uses a feature integration network to perform feature integration processing on the spectrum feature instance to obtain a first blood glucose feature instance. For example, the server integrates the spectrum features into a feature vector related to blood glucose through a fully connected layer. The server performs PPG signal analysis on each PPG signal instance to obtain a matching blood glucose representation instance (i.e., a feature representation corresponding to the real blood glucose value). For example, the server obtains the actual blood glucose value through an existing blood glucose monitoring device or method and converts it into a feature representation. The server uses a second feature extraction network to perform blood glucose feature extraction processing on the blood glucose representation instance to obtain a second blood glucose feature instance. Here, the initial features of the second feature extraction network have the same feature dimension as the result features of the feature inference network to ensure feature consistency. For example, the server uses another fully connected network to convert the blood glucose value into a feature vector of the same dimension as the feature inference network output. The server calculates the time series regression error between the first blood glucose feature instance and the second blood glucose feature instance to obtain a first error. This error reflects the accuracy of feature extraction and integration. For example, the server uses the mean square error (MSE) as a loss function to calculate the first error. The server uses the feature inference network to perform feature reconstruction processing on the first blood glucose feature instance to obtain a first inferred blood glucose characterization result. For example, the server outputs the inferred blood glucose value through the regression layer. The server calculates the error between the first inferred blood glucose characterization result and the PPG signal relationship target value (i.e., the true blood glucose value) to obtain a second error. This error reflects the accuracy of feature inference. For example, the server also uses the mean square error (MSE) as a loss function to calculate the second error. The server optimizes the network parameters of the first feature extraction network, the feature integration network, and the feature inference network according to the first error and the second error. For example, the server uses the back propagation algorithm and the gradient descent method to update the weights and biases of the network to minimize the total error. Through the above steps, the server can complete the integrated learning process of the first feature extraction network, the feature integration network and the feature inference network, thereby improving the accuracy of analyzing the blood glucose characterization results from the PPG signal.
[0038] In the embodiment of the present invention, the optimization of the network parameters of the first feature extraction network, the feature integration network and the feature inference network according to the first error and the second error can be implemented through the following examples.
[0039] Obtaining a first influencing factor of the first error matching, and obtaining a second influencing factor of the second error matching;
[0040] Performing a linear superposition operation on the first error and the second error according to the first influencing factor and the second influencing factor to obtain a first target error;
[0041] The network parameters of the first feature extraction network, the feature integration network and the feature inference network are optimized according to the first target error.
[0042] In an embodiment of the present invention, exemplarily, the server matches a first impact factor for each first error, and the impact factor reflects the importance of the error in the overall optimization. For example, the server can assign different weights to different first errors according to the source or type of the error. Similarly, the server matches a second impact factor for each second error, and the impact factor also reflects the importance of the error in the overall optimization. For example, the server can assign different weights to different second errors according to the accuracy requirements of the blood sugar value prediction. The server performs a linear superposition operation on the first error and the second error according to the first impact factor and the second impact factor to obtain the first target error. This operation process can be to multiply each error by its corresponding impact factor and then add all the results. For example, the server can calculate a weighted sum, where the weight of each error is its corresponding impact factor. The server uses the first target error as the optimization target to optimize the network parameters of the first feature extraction network, the feature integration network, and the feature inference network. This optimization process can be to update the weights and biases of the network by a gradient descent method or other optimization algorithms to minimize the first target error. For example, the server can use a back propagation algorithm to calculate the gradient and use a gradient descent method to update the network parameters. Through the above steps, the server can effectively optimize the network parameters of the first feature extraction network, the feature integration network and the feature inference network according to the importance of the first error and the second error, thereby improving the accuracy of analyzing the blood glucose characterization results from the PPG signal.
[0043] In an embodiment of the present invention, after performing PPG signal analysis on the PPG signal instance to obtain a matching blood glucose characterization instance, and performing blood glucose feature extraction processing on the blood glucose characterization instance according to a second feature extraction network to obtain the second blood glucose feature instance, the following implementation manner is also provided.
[0044] identifying a blood glucose-related component in the blood glucose characterization instance, and identifying a feature representation matching the blood glucose-related component in the second blood glucose feature instance;
[0045] performing feature integration processing on the feature representation in the second blood glucose feature instance according to the first blood glucose feature instance to obtain an integrated blood glucose feature;
[0046] Performing feature reconstruction processing on the integrated blood glucose feature according to the feature inference network to obtain a second inferred blood glucose characterization result, and performing feature reconstruction processing on the second blood glucose feature instance according to the feature inference network to obtain a third inferred blood glucose characterization result;
[0047] calculating a third error according to the second inferred blood glucose characterization result and the third inferred blood glucose characterization result;
[0048] The aforementioned optimization of the network parameters of the first feature extraction network, the feature integration network, and the feature inference network according to the first error and the second error can be implemented through the following examples.
[0049] The network parameters of the first feature extraction network, the feature integration network and the feature inference network are optimized according to the first error, the second error and the third error.
[0050] In an embodiment of the present invention, illustratively, the server identifies blood glucose related components in a blood glucose characterization instance. For example, the server identifies a portion directly related to changes in blood glucose levels from a blood glucose value time series obtained from a blood glucose monitoring device.
[0051] At the same time, the server identifies a feature representation that matches the blood sugar-related component in the second blood sugar feature instance, that is, the server finds a feature representation corresponding to the blood sugar-related component in the feature vector output by the second feature extraction network.
[0052] The server performs feature integration processing on the feature representation in the second blood glucose feature instance according to the first blood glucose feature instance to obtain an integrated blood glucose feature. In this process, the server uses a certain fusion strategy to combine the useful information in the first blood glucose feature instance with the blood glucose-related features in the second blood glucose feature instance.
[0053] The server uses the feature inference network to perform feature reconstruction processing on the integrated blood sugar feature to obtain a second inferred blood sugar representation result. This result is inferred based on the integrated blood sugar feature.
[0054] At the same time, the server also performs feature reconstruction processing on the second blood sugar feature instance to obtain a third inferred blood sugar characterization result. This result is directly inferred based on the second blood sugar feature instance.
[0055] The server calculates the error between the second inferred blood glucose characterization result and the third inferred blood glucose characterization result to obtain a third error. This error reflects the consistency and accuracy of the feature integration and feature inference process.
[0056] When optimizing network parameters, the server not only considers the first error and the second error, but also takes the third error into consideration. The server optimizes the network parameters of the first feature extraction network, the feature integration network, and the feature inference network based on the first error, the second error, and the third error. This optimization process can be implemented by a gradient descent algorithm, aiming to minimize the sum or weighted sum of the three errors.
[0057] Through the above steps, the server can more comprehensively consider the accuracy of blood glucose characterization and the consistency of feature extraction, integration, and inference processes, thereby more effectively optimizing the network parameters of the first feature extraction network, feature integration network, and feature inference network. This helps to improve the accuracy and reliability of blood glucose characterization results analyzed from PPG signals.
[0058] In the embodiment of the present invention, the feature integration processing is performed on the feature representation in the second blood glucose feature instance according to the first blood glucose feature instance to obtain the integrated blood glucose feature, which can be implemented through the following examples.
[0059] Performing operations on each feature representation and the first blood glucose feature instance according to a focusing mechanism to obtain blood glucose-related component spectrum features matching each blood glucose-related component;
[0060] The feature representations of the matches of each blood sugar related component in the second blood sugar feature instance are replaced with the matched blood sugar related component spectrum features to obtain an integrated blood sugar feature.
[0061] In the embodiment of the present invention, exemplarily, the server first determines a focusing mechanism, which defines how to emphasize or focus on a specific part of the second blood sugar feature instance based on the first blood sugar feature instance. For example, the focusing mechanism can be based on certain key features in the first blood sugar feature instance, such as blood sugar change trend or fluctuation amplitude, to weight the corresponding features in the second blood sugar feature instance. Then, the server performs a focusing mechanism operation on each feature representation in the second blood sugar feature instance and the first blood sugar feature instance. This operation process can be to perform a dot multiplication, weighted summation or other forms of operations on each feature representation in the second blood sugar feature instance and the relevant features in the first blood sugar feature instance to obtain the blood sugar related component spectral features matching each blood sugar related component. For example, if the first blood sugar feature instance indicates that blood sugar rises rapidly within a certain time period, the server will enhance the feature representation of the corresponding time period in the second blood sugar feature instance. After obtaining the blood sugar related component spectral features matching each blood sugar related component, the server replaces these spectral features into the second blood sugar feature instance to replace the original feature representation. This replacement process can be direct, that is, to cover the original feature representation with a new spectral feature; or it can be indirect, for example, by combining the new spectral feature with the original feature representation through a certain fusion strategy. By replacing the feature representation, the server obtains an integrated blood glucose feature. This integrated blood glucose feature contains both the original information in the second blood glucose feature instance and the key features in the first blood glucose feature instance, thereby improving the accuracy and reliability of blood glucose characterization. Through the above steps, the server can use the key information in the first blood glucose feature instance to enhance and optimize the feature representation in the second blood glucose feature instance, thereby obtaining a more accurate and reliable integrated blood glucose feature. This helps to improve the accuracy and reliability of blood glucose characterization results analyzed from PPG signals.
[0062] In an embodiment of the present invention, after performing feature reconstruction processing on the integrated blood glucose feature according to the feature inference network to obtain a second inferred blood glucose characterization result, and performing feature reconstruction processing on the second blood glucose feature instance according to the feature inference network to obtain a third inferred blood glucose characterization result, the following implementation manner may also be provided.
[0063] Obtaining a fourth error according to the target value of the relationship between the second inferred blood glucose characterization result and the PPG signal, and obtaining a fifth error according to the target value of the relationship between the third inferred blood glucose characterization result and the PPG signal;
[0064] The optimizing the network parameters of the first feature extraction network, the feature integration network and the feature inference network according to the first error, the second error and the third error comprises:
[0065] Calculate a second target error according to the first error, the second error, the third error, the fourth error and the fifth error;
[0066] The network parameters of the first feature extraction network, the feature integration network and the feature inference network are optimized according to the second target error.
[0067] In an embodiment of the present invention, exemplarily, the server first obtains a fourth error based on the second inferred blood glucose characterization result and the PPG signal relationship target value. The PPG signal relationship target value may be a blood glucose-related index directly extracted from the PPG signal, such as blood glucose concentration, blood glucose change trend, etc. The server calculates the difference between the second inferred blood glucose characterization result and these target values to obtain a fourth error. Similarly, the server also obtains a fifth error based on the third inferred blood glucose characterization result and the PPG signal relationship target value. This calculation process is similar to the calculation of the fourth error, except that the third inferred blood glucose characterization result is used. After obtaining the first error, the second error, the third error, the fourth error, and the fifth error, the server calculates the second target error based on these errors. This calculation process may perform a weighted summation of these errors, wherein the weight of each error reflects its importance in the overall optimization. The server uses the second target error as the optimization target to optimize the network parameters of the first feature extraction network, the feature integration network, and the feature inference network. This optimization process may be to update the weights and biases of the network by a gradient descent method or other optimization algorithms to minimize the second target error. For example, in practical applications, the server may extract blood glucose concentration from the PPG signal as the relationship target value. Then, the server calculates the mean square error between the second inferred blood glucose characterization result and the third inferred blood glucose characterization result and the target blood glucose concentration, and obtains the fourth error and the fifth error respectively. Then, the server performs weighted summation of these errors with the first error, the second error, and the third error to obtain the second target error. Finally, the server uses the gradient descent method to update the weights and biases of the network to minimize the second target error, thereby improving the accuracy and reliability of the blood glucose characterization results analyzed from the PPG signal.
[0068] In an embodiment of the present invention, after performing PPG signal analysis on the PPG signal instance to obtain a matching blood glucose characterization instance, and performing blood glucose feature extraction processing on the blood glucose characterization instance according to a second feature extraction network to obtain the second blood glucose feature instance, the following implementation manner may also be provided.
[0069] Loading the first blood sugar feature instance into a deep belief network for feature dimensionality reduction processing to obtain a first reduced-dimensional blood sugar feature, and loading the second blood sugar feature instance into the deep belief network for feature dimensionality reduction processing to obtain a second reduced-dimensional blood sugar feature;
[0070] Performing feature mapping processing on the first reduced-dimensionality blood sugar feature according to a feature mapping network to obtain a third reduced-dimensionality blood sugar feature, and performing feature mapping processing on the second reduced-dimensionality blood sugar feature according to the feature mapping network to obtain a fourth reduced-dimensionality blood sugar feature;
[0071] Calculating a sixth error according to the third reduced-dimensionality blood sugar feature and the fourth reduced-dimensionality blood sugar feature;
[0072] The calculating a second target error according to the first error, the second error, the third error, the fourth error and the fifth error comprises:
[0073] A second target error is calculated according to the first error, the second error, the third error, the fourth error, the fifth error, and the sixth error.
[0074] In an embodiment of the present invention, illustratively, the server first loads the first blood glucose feature instance into a deep belief network (DBN) for feature dimensionality reduction processing to obtain a first reduced-dimensional blood glucose feature. In this process, the deep belief network performs a nonlinear transformation on the first blood glucose feature instance through its multi-layer structure, extracts key information therein, and reduces the feature dimension. Similarly, the server also loads the second blood glucose feature instance into the deep belief network for feature dimensionality reduction processing to obtain a second reduced-dimensional blood glucose feature. This processing process is the same as the processing process of the first blood glucose feature instance, which is intended to extract key information in the second blood glucose feature instance and reduce the feature dimension. The server then uses the feature mapping network to perform feature mapping processing on the first reduced-dimensional blood glucose feature to obtain a third reduced-dimensional blood glucose feature. The feature mapping network can be a neural network or other machine learning model, which can further map the reduced-dimensional features to a new feature space for better subsequent processing. At the same time, the server also performs feature mapping processing on the second reduced-dimensional blood glucose feature to obtain a fourth reduced-dimensional blood glucose feature. This processing process is the same as the processing process of the first reduced-dimensional blood glucose feature, which is intended to map the second reduced-dimensional blood glucose feature to a new feature space. After obtaining the third reduced-dimensional blood sugar feature and the fourth reduced-dimensional blood sugar feature, the server calculates the error between the two features to obtain the sixth error. This error reflects the consistency of the feature mapping network after processing the two reduced-dimensional blood sugar features. When optimizing the network parameters, the server not only considers the first error, the second error, the third error, the fourth error and the fifth error, but also takes the sixth error into consideration. The server calculates the second target error based on these errors, and this calculation process can be a weighted summation or other combination of these errors. For example, in practical applications, the server can first use a deep belief network to reduce the dimension of the first blood sugar feature instance and the second blood sugar feature instance to extract key information. Then, the server uses the feature mapping network to map these reduced-dimensional features to a new feature space. Next, the server calculates the error between the mapped features (i.e., the sixth error) and uses it together with other errors to calculate the second target error. Finally, the server uses a gradient descent method or other optimization algorithm to update the weights and biases of the network to minimize the second target error, thereby improving the accuracy and reliability of the blood sugar characterization results analyzed from the PPG signal.
[0075] In an embodiment of the present invention, after the first reduced-dimensionality blood sugar feature is feature mapped according to the feature mapping network to obtain the third reduced-dimensionality blood sugar feature, and the second reduced-dimensionality blood sugar feature is feature mapped according to the feature mapping network to obtain the fourth reduced-dimensionality blood sugar feature, the following implementation manner may also be provided.
[0076] Performing feature reconstruction processing on the third reduced-dimensionality blood glucose feature according to the feature inference network to obtain a fourth inferred blood glucose characterization result, and performing feature reconstruction processing on the fourth reduced-dimensionality blood glucose feature according to the feature inference network to obtain a fifth inferred blood glucose characterization result;
[0077] calculating a seventh error according to the fourth inferred blood glucose characterization result and the fifth inferred blood glucose characterization result;
[0078] The calculating a second target error according to the first error, the second error, the third error, the fourth error, the fifth error and the sixth error comprises:
[0079] A second target error is calculated according to the first error, the second error, the third error, the fourth error, the fifth error, the sixth error, and the seventh error.
[0080] In an embodiment of the present invention, exemplarily, the server uses a feature inference network to perform feature reconstruction processing on the third reduced-dimensional blood glucose feature to obtain a fourth inferred blood glucose characterization result. In this process, the feature inference network reconstructs a blood glucose characterization result that is closer to the original blood glucose characterization based on the key information in the third reduced-dimensional blood glucose feature. At the same time, the server also performs feature reconstruction processing on the fourth reduced-dimensional blood glucose feature to obtain a fifth inferred blood glucose characterization result. This processing process is the same as the processing process of the third reduced-dimensional blood glucose feature, and is intended to reconstruct the blood glucose characterization result based on the key information in the fourth reduced-dimensional blood glucose feature. After obtaining the fourth inferred blood glucose characterization result and the fifth inferred blood glucose characterization result, the server calculates the error between the two results to obtain a seventh error. This error reflects the degree of consistency after the feature inference network reconstructs the two reduced-dimensional blood glucose features. When performing network parameter optimization, the server not only considers the first error, the second error, the third error, the fourth error, the fifth error and the sixth error, but also takes the seventh error into consideration. The server calculates the second target error based on these errors, and this calculation process can be a weighted summation or other combination of these errors to obtain a comprehensive error index. For example, in actual applications, the server may first use a feature mapping network to map the first reduced-dimensionality blood glucose feature and the second reduced-dimensionality blood glucose feature to obtain a third reduced-dimensionality blood glucose feature and a fourth reduced-dimensionality blood glucose feature. Then, the server uses a feature inference network to reconstruct these reduced-dimensionality features to obtain a fourth inferred blood glucose characterization result and a fifth inferred blood glucose characterization result. Next, the server calculates the error between the two inference results (i.e., the seventh error) and uses it together with other errors to calculate the second target error. Finally, the server uses a gradient descent method or other optimization algorithm to update the weights and biases of the network to minimize the second target error, thereby improving the accuracy and reliability of the blood glucose characterization results analyzed from the PPG signal.
[0081] In the embodiment of the present invention, the calculation of the second target error according to the first error, the second error, the third error, the fourth error, the fifth error, the sixth error and the seventh error can be implemented through the following example.
[0082] Get the learning cycle;
[0083] Calculate a third impact factor of each error match according to the learning cycle;
[0084] A linear superposition operation is performed on the first error, the second error, the third error, the fourth error, the fifth error, the sixth error and the seventh error according to the third influencing factor matched by each error to obtain a second target error.
[0085] In an embodiment of the present invention, illustratively, the server first obtains the current learning cycle. The learning cycle is a parameter in the training process, which is used to control the number of iterations or time of training. In different learning cycles, the server will pay different attention to the error in order to better adjust the network parameters. The server calculates the third impact factor matched by each error according to the learning cycle. The third impact factor is a weight value used to balance the importance of each error when calculating the second target error. As the learning cycle proceeds, the server gradually adjusts these weight values to more finely control the optimization process of the network parameters. The server uses the third impact factor matched by each error to perform a linear superposition operation on the first error, the second error, the third error, the fourth error, the fifth error, the sixth error, and the seventh error. This operation process is to multiply each error by its corresponding third impact factor, and then add all the results to obtain a comprehensive error indicator, that is, the second target error. For example, in actual applications, the server can first obtain the current learning cycle, such as the 10th cycle. Then, the server calculates the third impact factor of each error according to this learning cycle, such as the third impact factor of the first error is 0.1, the third impact factor of the second error is 0.2, and so on. Next, the server uses these third influencing factors to perform linear superposition operations on each error, such as the second target error = 0.1*first error + 0.2*second error + ... + 0.1*seventh error. Finally, the server uses this second target error to update the weights and biases of the network to further optimize the accuracy and reliability of the blood glucose characterization results analyzed from the PPG signal.
[0086] In an embodiment of the present invention, after performing PPG signal analysis on the PPG signal instance to obtain a matching blood glucose characterization instance, and performing blood glucose feature extraction processing on the blood glucose characterization instance according to a second feature extraction network to obtain the second blood glucose feature instance, the following implementation manner may also be provided.
[0087] Loading the first blood glucose feature instance into a deep belief network for feature dimensionality reduction processing, and loading the features obtained by dimensionality reduction into a feature mapping network for feature mapping processing to obtain a fifth dimensionality reduced blood glucose feature;
[0088] Loading the second blood glucose feature instance into a deep belief network for feature dimensionality reduction processing, and loading the features obtained by dimensionality reduction into a feature mapping network for feature mapping processing to obtain a sixth dimensionality reduced blood glucose feature;
[0089] Calculating an eighth error according to the fifth reduced-dimensionality blood sugar feature and the sixth reduced-dimensionality blood sugar feature;
[0090] The aforementioned optimization of the network parameters of the first feature extraction network, the feature integration network, and the feature inference network according to the first error and the second error can be implemented through the following examples.
[0091] The network parameters of the first feature extraction network, the feature integration network and the feature inference network are optimized according to the first error, the second error and the eighth error.
[0092] In an embodiment of the present invention, illustratively, the server loads the first blood glucose feature instance into the deep belief network for feature dimension reduction processing. The deep belief network performs nonlinear transformation on the first blood glucose feature instance through its multi-layer structure, extracts key information therein, and reduces the feature dimension, thereby obtaining the reduced dimension feature. Similarly, the server also loads the second blood glucose feature instance into the deep belief network for feature dimension reduction processing, extracts key information and reduces the feature dimension, and obtains another reduced dimension feature. The server then loads the features of the first blood glucose feature instance obtained by dimension reduction into the feature mapping network for feature mapping processing. The feature mapping network can be a neural network or other machine learning model, which can further map the reduced dimension features to a new feature space for better subsequent processing to obtain the fifth reduced dimension blood glucose feature. At the same time, the server also loads the features of the second blood glucose feature instance obtained by dimension reduction into the feature mapping network for feature mapping processing to obtain the sixth reduced dimension blood glucose feature. After obtaining the fifth reduced dimension blood glucose feature and the sixth reduced dimension blood glucose feature, the server calculates the error between the two features to obtain the eighth error. This error reflects the degree of consistency after the feature mapping network processes the two reduced dimension blood glucose features. When optimizing network parameters, the server not only considers the first error and the second error, but also takes the eighth error into consideration. The server optimizes the network parameters of the first feature extraction network, the feature integration network, and the feature inference network according to these errors. This optimization process can be to update the weights and biases of the network using a gradient descent method or other optimization algorithms to minimize the overall error. For example, in practical applications, the server can first use a deep belief network to reduce the dimension of the first blood glucose feature instance and the second blood glucose feature instance to extract the key information therein. Then, the server uses a feature mapping network to map these reduced-dimensional features to a new feature space to obtain the fifth reduced-dimensional blood glucose feature and the sixth reduced-dimensional blood glucose feature. Next, the server calculates the error (i.e., the eighth error) between the two mapped features, and uses it together with the first error and the second error to optimize the network parameters of the first feature extraction network, the feature integration network, and the feature inference network. Finally, the server uses the optimized network parameters to perform blood glucose characterization analysis on the new PPG signal instance to improve the accuracy and reliability of the analysis.
[0093] In the embodiment of the present invention, before performing feature reconstruction processing on the blood glucose feature according to the feature inference network to obtain the blood glucose characterization result of the target PPG signal, the following implementation manner may also be provided.
[0094] Performing quality analysis on the target PPG signal to obtain a PPG signal quality index;
[0095] Performing blood glucose feature extraction processing on the PPG signal quality indicator to obtain a blood glucose reliable feature;
[0096] fusing the blood glucose reliable feature with the blood glucose feature to obtain a target blood glucose feature;
[0097] The step of performing feature reconstruction processing on the blood glucose feature according to the feature inference network to obtain the blood glucose characterization result of the target PPG signal includes:
[0098] The target blood glucose feature is subjected to feature reconstruction processing according to the feature inference network to obtain the blood glucose characterization result of the target PPG signal.
[0099] In an embodiment of the present invention, exemplarily, the server first performs a quality analysis on the target PPG signal. This analysis process may include detecting quality problems such as noise, artifacts, signal loss, etc. in the signal, and evaluating the stability and reliability of the signal. Through quality analysis, the server obtains a PPG signal quality index, which quantifies the quality level of the target PPG signal. The server then performs a blood glucose feature extraction process on the PPG signal quality index. This process may involve extracting features related to blood glucose levels from the quality index using a machine learning algorithm or a statistical method. Through feature extraction processing, the server obtains blood glucose reliable features, which reflect the potential relationship between the quality of the PPG signal and the blood glucose level. The server fuses the blood glucose reliable features with the blood glucose features previously extracted from the PPG signal. The fusion process may be a combination of concatenating, weighted summing, or other forms of two feature vectors to obtain a comprehensive target blood glucose feature. This target blood glucose feature contains both blood glucose related information extracted from the original PPG signal and considers the impact of the quality of the PPG signal on the blood glucose characterization result. Finally, the server performs feature reconstruction processing on the target blood glucose feature according to the feature inference network. The feature inference network can be a neural network or other machine learning model, which can reconstruct a blood glucose characterization result that is closer to the original blood glucose characterization based on the key information in the target blood glucose feature. Through the feature reconstruction process, the server obtains the blood glucose characterization result of the target PPG signal, which reflects the blood glucose level information contained in the PPG signal. For example, in practical applications, the server can first receive a target PPG signal and perform a quality analysis on the signal to obtain a PPG signal quality index. Then, the server uses a machine learning algorithm to extract reliable blood glucose features from the quality index. Next, the server fuses these reliable blood glucose features with the blood glucose features extracted from the PPG signal to obtain a comprehensive target blood glucose feature. Finally, the server uses the feature inference network to reconstruct the target blood glucose feature to obtain the blood glucose characterization result of the target PPG signal. This result can be used to monitor changes in the patient's blood glucose level and provide valuable reference information for diabetes management.
[0100] In the embodiment of the present invention, the aforementioned step S205 can be implemented through the following example execution.
[0101] Determine a health management evaluation result according to the user physiological characteristics of the target user, the policy implementation characteristics and matching characteristics of the blood sugar health management policy, wherein the matching characteristics are used to characterize the degree of matching between the blood sugar health management policy and the target user, and the health management evaluation result is used to characterize the evaluation results of the health status of multiple users after the blood sugar health management policy is pushed to the target user;
[0102] Determining crowd category information according to the user's physiological characteristics, wherein the crowd category information is used to characterize the confidence that the target user is classified into a corresponding crowd category;
[0103] Determine a push coefficient of the blood sugar health management strategy according to the health management assessment result and the population category information, wherein the push coefficient is used to characterize the push priority of the blood sugar health management strategy;
[0104] In response to the push coefficient satisfying a push criterion, the blood sugar health management strategy is pushed to the target user.
[0105] In an embodiment of the present invention, exemplarily, the server first collects physiological characteristics of the target user, which may include age, gender, height, weight, BMI index, etc. The server also considers the policy implementation characteristics and matching characteristics of the blood glucose health management strategy. The policy implementation characteristics may include the difficulty of executing the strategy, cost, user acceptance, etc.; the matching characteristics are used to characterize the degree of match between the blood glucose health management strategy and the target user, such as whether the strategy is suitable for the user's daily living habits, dietary preferences, etc. Based on this information, the server uses a machine learning algorithm or a statistical model to perform a health management assessment on the user's health status and obtain a health management assessment result. This result predicts the improvement of the health status of multiple users after the blood glucose health management strategy is pushed to the target user. The server uses a classification algorithm or a clustering algorithm to classify the user into different population categories based on the physiological characteristics of the target user. The server calculates the confidence that the target user is classified into the corresponding population category, which reflects the possibility that the user belongs to the population category. The server uses a decision tree, logistic regression or other machine learning model to determine the push coefficient of the blood glucose health management strategy in combination with the health management assessment results and the population category information. The push coefficient is a quantitative indicator that characterizes the push priority of the blood glucose health management strategy. The higher the push coefficient, the more applicable the strategy is to the target user and the higher the priority. The server checks whether the push coefficient meets the preset push criteria. The push criteria can be a threshold, and only when the push coefficient exceeds this threshold is the strategy considered suitable for pushing to the user. If the push coefficient meets the push criteria, the server pushes the blood glucose health management strategy to the target user through a mobile application, SMS, email, or other communication methods. For example, in actual applications, the server can first collect physiological characteristics such as age, gender, BMI index, etc. of the target user, and consider the implementation difficulty and user acceptance of a blood glucose health management strategy. Then, the server uses a machine learning algorithm to predict the improvement of the user's blood glucose level after pushing the strategy to the user. Next, the server classifies the user into different population categories and calculates the confidence that the user belongs to each category. Based on this information and the preset push criteria, the server determines a push coefficient and decides whether to push the blood glucose health management strategy to the user. Finally, the user receives a text message or mobile application notification containing personalized blood glucose management suggestions.
[0106] In the embodiment of the present invention, the health management assessment result is determined based on the user physiological characteristics of the target user, the policy implementation characteristics and the matching characteristics of the blood glucose health management policy, and can be implemented through the following examples.
[0107] Performing combined feature combination processing on the user physiological features of the target user, the strategy implementation features of the blood sugar health management strategy, and the common features in the matching features to obtain combined features;
[0108] Determining a plurality of inferred features according to the user physiological features, the policy implementation features and the matching features, wherein the plurality of inferred features are used to characterize transition evaluation results of the health states of the plurality of users;
[0109] The health management assessment result is determined according to the combined feature and the multiple inferred features.
[0110] In an embodiment of the present invention, exemplarily, the server first identifies the common features in the user physiological characteristics of the target user, the policy implementation characteristics of the blood glucose health management policy, and the matching characteristics. These common features are features that span different dimensions (such as user physiology, policy implementation, and policy matching) but have similar or related meanings. The server performs a combined feature combination process on these common features, which may include methods such as feature splicing, feature weighted summation, or feature fusion to obtain a comprehensive combined feature. This combined feature integrates key information from different dimensions for subsequent health management evaluation. The server uses a machine learning algorithm or a statistical model to infer the transition evaluation results of the health status of multiple users based on the user physiological characteristics, policy implementation characteristics, and matching characteristics of the target user. These inferred features may include the expected changes in the user's blood glucose level, the degree of improvement in the user's health status, the user's potential response to the strategy, and the like. Each inferred feature reflects the possible changes in the user's health status in a specific aspect. The server combines the previously obtained combined features and multiple inferred features, and uses another machine learning model or decision algorithm to determine the health management evaluation result. This evaluation result is a comprehensive indicator that takes into account the potential impact of user physiological characteristics, policy implementation characteristics, and matching characteristics on user health status, and predicts the overall improvement of user health status after the blood glucose health management strategy is pushed to the user. For example, in actual applications, the server can first collect physiological characteristics such as age, gender, BMI index, etc. of the target user, and consider the implementation difficulty of a blood glucose health management strategy, user acceptance, and the degree of match between the strategy and the user's living habits. Then, the server identifies common features among these features, such as age and difficulty of policy implementation, and performs combined feature combination processing. Next, the server uses a machine learning algorithm to infer inferred features such as the expected change in the user's blood glucose level and the degree of improvement in health status. Finally, the server combines the combined features and inferred features to determine the health management evaluation results and decide whether to push the blood glucose health management strategy to the user. This process helps the server more accurately evaluate the potential impact of the strategy on the user's health status and make more personalized push decisions.
[0111] In the embodiment of the present invention, the user physiological feature of the target user, the policy implementation feature and the common feature in the matching feature are combined to obtain the combined feature, which can be implemented through the following example.
[0112] Performing linear superposition processing on multiple valid items in the user physiological characteristics of the target user, the policy implementation characteristics of the blood sugar health management policy, and the common characteristics in the matching characteristics to obtain primary features, wherein one valid item corresponds to one valid item weight;
[0113] Compressing the common features among the user physiological features, the policy implementation features and the matching features to obtain a plurality of first compressed features;
[0114] Determining a high-level combined feature based on the plurality of first compressed features;
[0115] The primary feature and the high-level combined feature are fused to obtain the combined feature.
[0116] In an embodiment of the present invention, exemplarily, the server first identifies the user physiological characteristics of the target user, the policy implementation characteristics of the blood sugar health management policy, and the common characteristics in the matching characteristics, and determines multiple valid items in these common characteristics. The server assigns a corresponding weight to each valid item, and these weights reflect the importance of the valid items in the combined features. The server performs linear superposition processing on multiple valid items, that is, multiplying each valid item with its corresponding weight, and adding the results to obtain a primary feature. This primary feature is a comprehensive feature that integrates information of multiple valid items. The server compresses the common features in the user physiological characteristics, policy implementation characteristics, and matching characteristics. The compression processing may include methods such as dimensionality reduction, feature selection, or feature extraction, which are intended to reduce the number of features and retain key information. Through the compression processing, the server obtains multiple first compressed features. Each first compressed feature is a compressed representation of the original common feature, which contains the key information of the original feature. The server uses a machine learning algorithm or a statistical model to determine a high-level combined feature based on multiple first compressed features. The high-level combined feature is a feature that integrates information of multiple first compressed features, which may have a higher level of abstraction and stronger representation capabilities. The server performs fusion processing on the primary features and the high-level combined features. The fusion process may include methods such as feature concatenation, feature weighted summation or feature fusion, aiming to integrate the key information in the primary features and advanced combined features into a unified feature representation. Through the fusion process, the server obtains the final combined feature. This combined feature integrates the key information of the common features from the user's physiological features, policy implementation features and matching features, and is used for subsequent health management evaluation. For example, in practical applications, the server can first identify the target user's physiological characteristics such as age, gender, BMI index, and common features such as the difficulty of implementing the blood sugar health management strategy, user acceptance, and the degree of matching between the strategy and the user's living habits. Then, the server performs linear superposition processing on the valid items in these common features to obtain the primary features. Next, the server compresses the common features to obtain multiple first compressed features, and determines the advanced combined features based on these first compressed features. Finally, the server fuses the primary features and the advanced combined features to obtain the final combined features, which are used for subsequent health management evaluation. This process helps the server to more effectively integrate and utilize feature information from different dimensions, and improve the accuracy and personalization of health management evaluation.
[0117] In the embodiment of the present invention, determining the high-level combined feature according to the multiple first compression features may be implemented through the following examples.
[0118] Determine a scalar product result of any two second compressed features in the plurality of first compressed features;
[0119] A superposition operation is performed on the determined multiple scalar product results to obtain the high-level combination feature.
[0120] In an embodiment of the present invention, exemplarily, the server first obtains a plurality of first compressed features previously obtained by compression processing. The server calculates the scalar product result of any two second compressed features in these first compressed features. The scalar product is the result of summing the corresponding elements of two vectors after multiplication, which reflects the similarity or correlation between the two vectors. By calculating the scalar product of all possible pairs of second compressed features, the server obtains a set of scalar product results. The server performs a superposition operation on the determined multiple scalar product results. The superposition operation can be a simple summation or a weighted summation, where the weight reflects the importance of different scalar product results. Through the superposition operation, the server obtains a high-level combined feature that integrates the similarity and correlation between all first compressed features. For example, in practical applications, the server can first obtain the first compressed features of the target user's age, gender, and BMI index through compression processing, as well as the first compressed features of the difficulty of implementing the blood sugar health management strategy, the user's acceptance, and the degree of matching between the strategy and the user's living habits. Then, the server calculates the scalar product results of any two of these first compressed features, such as the scalar product of age and implementation difficulty, the scalar product of gender and user acceptance, etc. Next, the server performs a superposition operation on these scalar product results to obtain a high-level combined feature that integrates the similarity and correlation between the user's physiological characteristics and the policy implementation characteristics. Finally, this high-level combined feature is used in subsequent health management assessments to improve the accuracy and personalization of the assessment.
[0121] In the embodiment of the present invention, the determination of multiple inferred features based on the user's physiological features, the policy implementation features, and the matching features may be implemented through the following examples.
[0122] fusing the key features of the user physiological features, the policy implementation features, and the matching features with a plurality of second compressed features to obtain a first user feature, wherein the plurality of second compressed features are obtained by compressing the common features of the user physiological features, the policy implementation features, and the matching features;
[0123] According to the multiple classroom networks corresponding to the multiple user health states, the first user characteristics are processed respectively to obtain multiple classroom network evaluation characteristics;
[0124] According to the multiple self-attention components corresponding to the multiple user health states, the first user feature and the multiple classroom network evaluation features are processed respectively to obtain multiple evaluation feature self-attention features;
[0125] According to the multiple nonlinear mapping components corresponding to the multiple user health states, the multiple evaluation feature self-attention features are processed respectively to obtain the multiple inference features.
[0126] In an embodiment of the present invention, exemplarily, the server first identifies key features in user physiological features, policy implementation features, and matching features. These key features are the most important features for evaluating the health status of the user. The server also compresses the common features in the user's physiological features, policy implementation features, and matching features to obtain multiple second compressed features. These second compressed features are a dimensionality reduction or compression representation of the original common features. The server fuses the key features and multiple second compressed features, and can obtain a comprehensive first user feature by methods such as feature splicing, feature weighted summation, or feature fusion. This first user feature integrates key information and compressed representations from different dimensions. The server processes the first user feature according to multiple classroom networks corresponding to multiple user health states. The classroom network can be a deep learning model for extracting features related to the user's health status. Through the processing of the classroom network, the server obtains multiple classroom network evaluation features. Each classroom network evaluation feature reflects a certain relationship or pattern between the first user feature and the user's health status. The server processes the first user feature and multiple classroom network evaluation features according to multiple self-attention components corresponding to multiple user health states. The self-attention component can be an attention mechanism for capturing dependencies between features. Through the processing of the self-attention component, the server obtains multiple evaluation feature self-attention features. These features not only contain information about the first user feature and the classroom network evaluation feature, but also reflect the dependency between them. The server processes the multiple evaluation feature self-attention features according to the multiple nonlinear mapping components corresponding to the health states of multiple users. The nonlinear mapping component can be a nonlinear function or a neural network layer for mapping features to a new space or representation. Through the processing of the nonlinear mapping component, the server obtains multiple inferred features. These inferred features are used to characterize the transition evaluation results of the health states of multiple users, that is, they reflect the possible changes or trends in specific aspects of the health states of users. For example, in practical applications, the server can first identify the key physiological characteristics of the target user, such as age, gender, BMI index, and key strategy features and matching features such as the difficulty of implementing the blood sugar health management strategy, user acceptance, and the degree of matching between the strategy and the user's living habits. Then, the server compresses the common features in these features to obtain multiple second compressed features. Then, the server fuses the key features and the second compressed features to obtain the first user feature. Subsequently, the server processes the first user feature according to the classroom network corresponding to the health states of multiple users to obtain multiple classroom network evaluation features. Then, the server processes the first user feature and the classroom network evaluation feature according to the self-attention components corresponding to the health states of the multiple users to obtain multiple evaluation feature self-attention features.Finally, the server processes the self-attention features of the evaluation features according to the nonlinear mapping components corresponding to the health states of multiple users to obtain multiple inferred features, which are used to characterize the transition evaluation results of the user's health state in specific aspects.
[0127] In an embodiment of the present invention, the first user feature and the multiple classroom network evaluation features are processed respectively according to the multiple self-attention components corresponding to the multiple user health states to obtain multiple evaluation feature self-attention features, which can be implemented through the following examples.
[0128] For any user health state, according to the self-attention component of the user health state, the first user feature is fully mapped and transformed to obtain multiple user weights, each of the user weights corresponds to each of the classroom network evaluation features;
[0129] According to the multiple user weights, the multiple classroom network evaluation features are linearly superimposed to obtain the evaluation feature self-attention feature corresponding to the user health status.
[0130] In an embodiment of the present invention, exemplarily, for any user health state, the server first obtains a self-attention component corresponding to the user health state. The server uses the self-attention component to perform a comprehensive mapping conversion on the first user feature. This mapping conversion can be a complex function or neural network layer, which can convert the first user feature into a set of user weights. Through the mapping conversion, the server obtains multiple user weights. Each user weight corresponds to each classroom network evaluation feature one by one, indicating the importance or contribution of the classroom network evaluation feature in the user health state evaluation. The server performs linear superposition processing on multiple classroom network evaluation features based on the obtained multiple user weights. Linear superposition processing is a process of multiplying each classroom network evaluation feature with its corresponding user weight and adding the results. In this way, the server can obtain an evaluation feature self-attention feature that integrates multiple classroom network evaluation feature information. This evaluation feature self-attention feature not only contains the information of the classroom network evaluation feature, but also reflects their relative importance or contribution in the user health state evaluation. For example, in practical applications, the server can first obtain a self-attention component corresponding to the target user health state "hyperglycemia". Then, the server uses this self-attention component to perform a comprehensive mapping transformation on the first user feature to obtain a set of user weights. These user weights correspond to different classroom network evaluation features, such as "age feature evaluation", "gender feature evaluation" and "BMI index feature evaluation". Then, the server performs linear superposition processing on the classroom network evaluation features based on these user weights to obtain an evaluation feature self-attention feature that integrates multiple classroom network evaluation feature information. This evaluation feature self-attention feature is used to characterize the self-attention evaluation results of the target user's health status "hyperglycemia" in specific aspects, that is, it reflects the relative importance or contribution of different classroom network evaluation features in the user's health status evaluation.
[0131] In the embodiment of the present invention, determining the health management assessment result according to the combined feature and the multiple inferred features may be implemented through the following example.
[0132] The combined feature and the multiple inferred features are respectively fused to obtain multiple second fused features;
[0133] The health management assessment result is determined according to the multiple second fusion features.
[0134] In an embodiment of the present invention, exemplarily, the server first obtains the previously determined combined feature and multiple inferred features. The server performs a fusion process on the combined feature and each inferred feature. The fusion process may include methods such as feature splicing, feature weighted summation or feature fusion, which is intended to integrate the information of the combined feature and the inferred feature. Through the fusion process, the server obtains multiple second fused features. Each second fused feature integrates the information of the combined feature and the corresponding inferred feature. The server performs further analysis and processing based on the obtained multiple second fused features. The server may use a machine learning model, a deep learning model or a statistical model to process the second fused feature to determine the health management assessment result of the user's health status. Finally, the server outputs the health management assessment result of the user's health status, which may be a value, a category or a probability distribution, used to represent the health management effect or risk of the user's health status. For example, in actual applications, the server may first obtain the previously determined combined feature and multiple inferred features, which respectively reflect the user's physiological condition, the implementation of the health management strategy, and the degree of matching between the strategy and the user's living habits. Then, the server performs a fusion process on the combined feature and each inferred feature to obtain multiple second fused features. Next, the server processes the second fusion feature using a machine learning model to determine the health management assessment result of the user's health status. Finally, the server outputs a numerical value indicating the degree of health improvement or risk reduction of the user under the current health management strategy. This assessment result can be used to guide the user to adjust the health management strategy to achieve better health results.
[0135] In the embodiment of the present invention, the determination of the crowd category information according to the physiological characteristics of the user may be implemented through the following examples.
[0136] Acquire a plurality of object compression features, wherein the object compression features are obtained by compressing common features in the user's physiological features;
[0137] Fusing the plurality of object compression features with key features in the user physiological features to obtain a second user feature;
[0138] The second user feature is subjected to multiple comprehensive mapping conversions to obtain the population category information.
[0139] In an embodiment of the present invention, exemplarily, the server first obtains common features in the physiological features of the user, which are the parts of the physiological features of the user that are ubiquitous and non-specific. The server compresses the common features and may obtain multiple object compressed features by methods such as dimensionality reduction, feature selection or feature encoding. These object compressed features are a compact representation of the original common features. The server obtains key features in the physiological features of the user, which are the most important features for determining the category of the crowd. The server fuses the multiple object compressed features and the key features and may obtain a comprehensive second user feature by methods such as feature concatenation, feature weighted summation or feature fusion. This second user feature integrates key information and compressed representations from different dimensions. The server performs multiple comprehensive mapping conversions on the second user feature. This mapping conversion may be a complex function or a neural network layer that can convert the second user feature into a vector or label representing a crowd category. Through multiple mapping conversions, the server gradually refines and abstracts the crowd category information in the second user feature. Finally, the server outputs crowd category information, which may be a value, a category or a probability distribution, used to represent the crowd category to which the user belongs. For example, in actual applications, the server can first obtain the common features of the user's physiological characteristics such as age, gender, BMI index, etc., and compress these common features to obtain multiple object compressed features. Then, the server obtains the key features of the user's physiological characteristics, such as blood pressure, blood sugar, etc. Next, the server fuses the multiple object compressed features and key features to obtain a comprehensive second user feature. Subsequently, the server performs multiple comprehensive mapping transformations on the second user feature, gradually refining and abstracting the population category information to which the user belongs. Finally, the server outputs a population category label, such as "young and middle-aged healthy population" or "hypertension risk population", to indicate the population category to which the user belongs. This population category information can be used to guide users to choose appropriate health management strategies or participate in corresponding health research.
[0140] In the embodiment of the present invention, the push criteria may be implemented through the following examples.
[0141] The position of the push priority represented by the push coefficient among the push priorities represented by the push coefficients of the plurality of blood sugar health management strategies does not exceed a preset position threshold;
[0142] The push priority represented by the push coefficient exceeds a preset priority threshold.
[0143] In an embodiment of the present invention, exemplarily, the server first obtains multiple blood glucose health management strategies and their corresponding push coefficients. The server sorts the multiple blood glucose health management strategies according to the push coefficient and determines the push priority of each strategy. The server checks the position of the push priority of the target blood glucose health management strategy in the sorted list to ensure that the position does not exceed the preset position threshold. For example, the preset position threshold can be set to the top 10, which means that only strategies with push priorities in the top 10 will be considered for push. The server obtains a preset priority threshold, which is used to determine whether the push priority of the blood glucose health management strategy is high enough to be worth pushing. The server compares the push coefficient of the target blood glucose health management strategy with the preset priority threshold. If the push priority represented by the push coefficient of the target strategy exceeds the preset priority threshold, then the strategy meets part of the conditions of the push criteria. In actual applications, the server can first obtain 100 blood glucose health management strategies and their corresponding push coefficients. Then, the server sorts these strategies according to the push coefficient and determines the push priority of each strategy. Next, the server checks the push priority of a specific blood glucose health management strategy (e.g., Strategy A) and finds that it is ranked 5th in the sorted list, which satisfies the condition that "the position of the push priority does not exceed the preset position threshold" (assuming that the preset position threshold is the top 10). In addition, the server also obtains a preset priority threshold (e.g., 0.8) and compares the push coefficient of Strategy A with this threshold. If the push priority represented by the push coefficient of Strategy A exceeds 0.8, then it also meets the condition that "the push priority exceeds the preset priority threshold." In this case, Strategy A is considered to meet the push criteria and can be pushed to the user.
[0144] The embodiment of the present invention provides a computer device 100, which includes a processor and a non-volatile memory storing computer instructions. When the computer instructions are executed by the processor, the computer device 100 executes the aforementioned non-invasive blood glucose monitoring method based on PPG signals. Figure 2 As shown, Figure 2 This is a structural block diagram of a computer device 100 provided in an embodiment of the present invention. The computer device 100 includes a memory 111 , a processor 112 , and a communication unit 113 .
[0145] To achieve data transmission or interaction, the memory 111, processor 112 and communication unit 113 are electrically connected to each other directly or indirectly. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines. For illustrative purposes, the foregoing description is made with reference to specific embodiments. However, the above illustrative discussion is not intended to be exhaustive or limit the present disclosure to the precise form disclosed. According to the above teachings, many modifications and variations are feasible. These embodiments are selected and described in order to best illustrate the principles of the present disclosure and its practical application, so that those skilled in the art can best utilize the present disclosure and utilize various embodiments with different modifications to be suitable for the expected specific application.
Claims
1. A non-invasive blood glucose monitoring method based on PPG signals, characterized in that: include: Collect the original PPG signal of the target user from the smart wearable device equipped with an optical sensor; Filtering the original PPG signal to obtain a target PPG signal; Analyzing the target PPG signal to obtain a blood glucose characterization result of the target PPG signal; Acquire pre-stored basic physiological information of the user from the smart wearable device, and construct a physiological feature of the user based on the blood glucose characterization result and the basic physiological information of the user; The user health assessment result and population category information of the target user are determined according to the user physiological characteristics, and a blood sugar health management strategy is pushed to the target user according to the user health assessment result and the population category information.
2. The method according to claim 1, characterized in that The analyzing the target PPG signal to obtain the blood glucose characterization result of the target PPG signal includes: acquiring the target PPG signal to be identified; Performing spectrum feature extraction processing on the target PPG signal according to a first feature extraction network to obtain spectrum features; Loading the spectrum features into a feature integration network for feature integration processing to obtain blood glucose features; Performing feature reconstruction processing on the blood glucose feature according to the feature inference network to obtain a blood glucose characterization result of the target PPG signal; Among them, the first feature extraction network, the feature integration network and the feature inference network are obtained according to the target error integrated learning, and the target error includes a first error and a second error; the first error is obtained according to the first blood glucose feature instance and the second blood glucose feature instance, the first blood glucose feature instance is a PPG signal instance that is subjected to spectral feature extraction processing by the first feature extraction network and then subjected to feature integration processing by the feature integration network, and the second blood glucose feature instance is obtained by performing blood glucose feature extraction processing on the blood glucose characterization instance matched with the PPG signal instance; the second error is obtained according to the target value of the relationship between the inferred blood glucose characterization result and the PPG signal, and the inferred blood glucose characterization result is obtained by performing feature reconstruction processing on the first blood glucose feature instance by the feature inference network.
3. The method according to claim 2, characterized in that The integrated learning process of the first feature extraction network, the feature integration network and the feature inference network includes: Acquire a sample instance data set, wherein the sample instance data set includes a plurality of PPG signal instances and a PPG signal relationship target value matched with each PPG signal instance; Performing spectrum feature extraction processing on the PPG signal instance according to the first feature extraction network to obtain a spectrum feature instance, and performing feature integration processing on the spectrum feature instance according to the feature integration network to obtain the first blood glucose feature instance; performing PPG signal analysis on the PPG signal instance to obtain a matching blood glucose characterization instance, and performing blood glucose feature extraction processing on the blood glucose characterization instance according to a second feature extraction network to obtain the second blood glucose feature instance, wherein the initial feature of the second feature extraction network has the same feature dimension as the result feature of the feature inference network; Calculate a time series regression error according to the first blood sugar characteristic instance and the second blood sugar characteristic instance to obtain a first error; performing feature reconstruction processing on the first blood glucose feature instance according to the feature inference network to obtain a first inferred blood glucose characterization result, and obtaining a second error according to a target value of a relationship between the first inferred blood glucose characterization result and the PPG signal; The network parameters of the first feature extraction network, the feature integration network and the feature inference network are optimized according to the first error and the second error.
4. The method according to claim 3, characterized in that: The optimizing the network parameters of the first feature extraction network, the feature integration network and the feature inference network according to the first error and the second error includes: Obtaining a first influencing factor of the first error matching, and obtaining a second influencing factor of the second error matching; Performing a linear superposition operation on the first error and the second error according to the first influencing factor and the second influencing factor to obtain a first target error; The network parameters of the first feature extraction network, the feature integration network and the feature inference network are optimized according to the first target error.
5. The method according to claim 3, characterized in that: After performing PPG signal analysis on the PPG signal instance to obtain a matching blood glucose characterization instance, and performing blood glucose feature extraction processing on the blood glucose characterization instance according to a second feature extraction network to obtain the second blood glucose feature instance, the method further includes: identifying a blood glucose-related component in the blood glucose characterization instance, and identifying a feature representation matching the blood glucose-related component in the second blood glucose feature instance; Performing operations on each feature representation and the first blood glucose feature instance according to a focusing mechanism to obtain blood glucose-related component spectrum features matching each blood glucose-related component; replacing the feature representations of the respective blood sugar related components in the second blood sugar feature instance with the matched blood sugar related component spectrum features to obtain an integrated blood sugar feature; Performing feature reconstruction processing on the integrated blood glucose feature according to the feature inference network to obtain a second inferred blood glucose characterization result, and performing feature reconstruction processing on the second blood glucose feature instance according to the feature inference network to obtain a third inferred blood glucose characterization result; Obtaining a fourth error according to the target value of the relationship between the second inferred blood glucose characterization result and the PPG signal, and obtaining a fifth error according to the target value of the relationship between the third inferred blood glucose characterization result and the PPG signal; calculating a third error according to the second inferred blood glucose characterization result and the third inferred blood glucose characterization result; The optimizing the network parameters of the first feature extraction network, the feature integration network and the feature inference network according to the first error and the second error includes: Calculate a second target error according to the first error, the second error, the third error, the fourth error and the fifth error; The network parameters of the first feature extraction network, the feature integration network and the feature inference network are optimized according to the second target error.
6. The method according to claim 5, characterized in that After performing PPG signal analysis on the PPG signal instance to obtain a matching blood glucose characterization instance, and performing blood glucose feature extraction processing on the blood glucose characterization instance according to a second feature extraction network to obtain the second blood glucose feature instance, the method further includes: Loading the first blood sugar feature instance into a deep belief network for feature dimensionality reduction processing to obtain a first reduced-dimensional blood sugar feature, and loading the second blood sugar feature instance into the deep belief network for feature dimensionality reduction processing to obtain a second reduced-dimensional blood sugar feature; Performing feature mapping processing on the first reduced-dimensionality blood sugar feature according to a feature mapping network to obtain a third reduced-dimensionality blood sugar feature, and performing feature mapping processing on the second reduced-dimensionality blood sugar feature according to the feature mapping network to obtain a fourth reduced-dimensionality blood sugar feature; Performing feature reconstruction processing on the third reduced-dimensionality blood glucose feature according to the feature inference network to obtain a fourth inferred blood glucose characterization result, and performing feature reconstruction processing on the fourth reduced-dimensionality blood glucose feature according to the feature inference network to obtain a fifth inferred blood glucose characterization result; calculating a seventh error according to the fourth inferred blood glucose characterization result and the fifth inferred blood glucose characterization result; Calculating a sixth error according to the third reduced-dimensionality blood sugar feature and the fourth reduced-dimensionality blood sugar feature; The calculating a second target error according to the first error, the second error, the third error, the fourth error and the fifth error comprises: Get the learning cycle; Calculate a third impact factor of each error match according to the learning cycle; A linear superposition operation is performed on the first error, the second error, the third error, the fourth error, the fifth error, the sixth error and the seventh error according to the third influencing factor matched by each error to obtain a second target error.
7. The method according to claim 3, characterized in that After performing PPG signal analysis on the PPG signal instance to obtain a matching blood glucose characterization instance, and performing blood glucose feature extraction processing on the blood glucose characterization instance according to a second feature extraction network to obtain the second blood glucose feature instance, the method further includes: Loading the first blood glucose feature instance into a deep belief network for feature dimensionality reduction processing, and loading the features obtained by dimensionality reduction into a feature mapping network for feature mapping processing to obtain a fifth dimensionality reduced blood glucose feature; Loading the second blood glucose feature instance into a deep belief network for feature dimensionality reduction processing, and loading the features obtained by dimensionality reduction into a feature mapping network for feature mapping processing to obtain a sixth dimensionality reduced blood glucose feature; Calculating an eighth error according to the fifth reduced-dimensionality blood sugar feature and the sixth reduced-dimensionality blood sugar feature; The optimizing the network parameters of the first feature extraction network, the feature integration network and the feature inference network according to the first error and the second error includes: The network parameters of the first feature extraction network, the feature integration network and the feature inference network are optimized according to the first error, the second error and the eighth error.
8. The method according to claim 2, characterized in that: Before performing feature reconstruction processing on the blood glucose feature according to the feature inference network to obtain the blood glucose characterization result of the target PPG signal, the method further includes: Performing quality analysis on the target PPG signal to obtain a PPG signal quality index; Performing blood glucose feature extraction processing on the PPG signal quality indicator to obtain a blood glucose reliable feature; fusing the blood glucose reliable feature with the blood glucose feature to obtain a target blood glucose feature; The step of performing feature reconstruction processing on the blood glucose feature according to the feature inference network to obtain the blood glucose characterization result of the target PPG signal includes: The target blood glucose feature is subjected to feature reconstruction processing according to the feature inference network to obtain a blood glucose characterization result of the target PPG signal.
9. The method according to claim 1, characterized in that: The determining the user health assessment result and population category information of the target user according to the user physiological characteristics, and pushing the blood sugar health management strategy to the target user according to the user health assessment result and the population category information, includes: Performing linear superposition processing on multiple valid items in the user physiological characteristics of the target user, the strategy implementation characteristics of the blood sugar health management strategy, and the common characteristics in the matching characteristics to obtain primary features, where one valid item corresponds to one valid item weight; Compressing the common features among the user physiological features, the policy implementation features and the matching features to obtain a plurality of first compressed features; Determine a scalar product result of any two second compressed features in the plurality of first compressed features; Performing superposition operation on the determined multiple scalar product results to obtain advanced combination features; fusing the primary feature and the high-level combined feature to obtain the combined feature; fusing the key features of the user physiological features, the policy implementation features, and the matching features with a plurality of second compressed features to obtain a first user feature, wherein the plurality of second compressed features are obtained by compressing the common features of the user physiological features, the policy implementation features, and the matching features; According to a plurality of classroom networks corresponding to a plurality of user health states, the first user characteristics are processed respectively to obtain a plurality of classroom network evaluation characteristics; For any user health state, according to the self-attention component of the user health state, the first user feature is fully mapped and transformed to obtain multiple user weights, each of the user weights corresponds to each of the classroom network evaluation features; According to the multiple user weights, the multiple classroom network evaluation features are linearly superimposed to obtain the evaluation feature self-attention feature corresponding to the user health status; According to the multiple nonlinear mapping components corresponding to the multiple user health states, the multiple evaluation feature self-attention features are processed respectively to obtain multiple inference features, and the multiple inference features are used to characterize the transition evaluation results of the multiple user health states; The combined feature and the multiple inferred features are respectively fused to obtain multiple second fused features; Determine the health management evaluation result according to the multiple second fusion features, the matching feature is used to characterize the matching degree between the blood sugar health management strategy and the target user, and the health management evaluation result is used to characterize the evaluation results of the health status of multiple users after the blood sugar health management strategy is pushed to the target user; Acquire a plurality of object compression features, wherein the object compression features are obtained by compressing common features in the user's physiological features; Fusing the plurality of object compression features with key features in the user physiological features to obtain a second user feature; Performing multiple comprehensive mapping conversions on the second user feature to obtain the crowd category information, where the crowd category information is used to characterize the confidence that the target user is classified into the corresponding crowd category; Determine a push coefficient of the blood sugar health management strategy according to the health management assessment result and the population category information, wherein the push coefficient is used to characterize the push priority of the blood sugar health management strategy; In response to the push coefficient satisfying the push criteria, the blood glucose health management strategy is pushed to the target user; the push criteria include: the push priority represented by the push coefficient does not exceed a preset position threshold in the push priorities represented by the push coefficients of multiple blood glucose health management strategies; or the push priority represented by the push coefficient exceeds a preset priority threshold.
10. A server system, characterized in that: The method comprises a server, wherein the server is used to execute the method described in any one of claims 1 to 9.
Citation Information
Patent Citations
Noninvasive blood glucose detection method based on time-frequency domain comprehensive analysis
CN113662538A
Noninvasive personalized dynamic blood glucose trend monitoring and early warning method and device
CN116250834A
Method, apparatus and device for obtaining blood glucose measurement result
WO2021238810A1
