A non-invasive blood glucose monitoring method and system fusing multi-modal physiological parameters

Abstract

The application discloses a non-invasive blood glucose monitoring method and system fusing multi-modal physiological parameters, and relates to the technical field of blood glucose monitoring.The method comprises the following steps: collecting multi-modal physiological data of a target user, and obtaining a multi-modal physiological indication set; combining the multi-modal physiological indication set with the multi-modal physiological data, and analyzing state distribution data of the user; making a decision according to the state distribution data and a pre-trained meta-learning model, and obtaining an ensemble learning parameter; constructing a blood glucose fusion prediction model based on ensemble learning based on the ensemble learning parameter, inputting the state distribution data, the multi-modal physiological indication set and the multi-modal physiological data into the blood glucose fusion prediction model, and obtaining a blood glucose concentration monitoring value.The technical problem that existing non-invasive blood glucose monitoring technology is easily disturbed by external environmental factors, cannot meet the needs of clinical and daily monitoring, and leads to poor blood glucose monitoring effect is solved.

Description

Technical Field

[0001] This application relates to the field of blood glucose monitoring technology, specifically to a non-invasive blood glucose monitoring method and system that integrates multimodal physiological parameters. Background Technology

[0002] With increasing health awareness and a growing number of diabetic patients, blood glucose monitoring has become increasingly important. While traditional invasive blood glucose monitoring methods can measure blood glucose levels relatively accurately, they pose risks of pain and infection to patients, and frequent blood draws are inconvenient. Existing non-invasive blood glucose monitoring technologies still have shortcomings in accuracy, stability, and reliability. Some non-invasive monitoring technologies are susceptible to interference from external environmental factors, failing to meet the needs of clinical and routine monitoring, resulting in poor blood glucose monitoring outcomes. Summary of the Invention

[0003] This application provides a non-invasive blood glucose monitoring method and system that integrates multimodal physiological parameters, solving the technical problem that existing non-invasive blood glucose monitoring technologies are easily affected by external environmental factors, cannot meet the needs of clinical and daily monitoring, and result in poor blood glucose monitoring effects.

[0004] The technical solution to the above-mentioned technical problems in this application is as follows:

[0005] In a first aspect, this application provides a non-invasive blood glucose monitoring method that integrates multimodal physiological parameters, the method comprising:

[0006] Wearable devices are used to non-invasively and synchronously collect multimodal physiological data from target users, and the multimodal physiological data is analyzed to obtain multimodal physiological indicators.

[0007] By combining the multimodal physiological indicators and the multimodal physiological data, the user's state distribution data is analyzed and obtained;

[0008] Decisions are made based on the state distribution data and the pre-trained meta-learning model to obtain ensemble learning parameters, wherein the ensemble learning parameters include at least the number of ensemble models, the model grouping ratio, and the ensemble path parameters;

[0009] Based on the ensemble learning parameters, an ensemble learning-based blood glucose fusion prediction model is constructed, and the state distribution data, the multimodal physiological indicators, and the multimodal physiological data are input into the blood glucose fusion prediction model to obtain blood glucose concentration monitoring values.

[0010] Secondly, this application provides a non-invasive blood glucose monitoring system that integrates multimodal physiological parameters, comprising:

[0011] The data acquisition module is used to non-invasively and synchronously collect multimodal physiological data of the target user through wearable devices, and analyze the multimodal physiological data to obtain multimodal physiological indicators.

[0012] The data analysis module is used to combine the multimodal physiological indicators and the multimodal physiological data to analyze and obtain the user's state distribution data;

[0013] The parameter acquisition module is used to make decisions based on the state distribution data and the pre-trained meta-learning model to obtain ensemble learning parameters, wherein the ensemble learning parameters include at least the number of ensemble models, the model grouping ratio, and the ensemble path parameters.

[0014] The concentration monitoring module is used to construct a blood glucose fusion prediction model based on the ensemble learning parameters, and input the state distribution data, the multimodal physiological indicators and the multimodal physiological data into the blood glucose fusion prediction model to obtain blood glucose concentration monitoring values.

[0015] This application provides one or more technical solutions, which have at least the following technical effects or advantages:

[0016] This application provides a non-invasive blood glucose monitoring method and system that integrates multimodal physiological parameters. First, a wearable device non-invasively and synchronously collects multimodal physiological data from the target user, avoiding the pain and infection risks associated with traditional invasive monitoring. Second, the collected data is analyzed to obtain multimodal physiological indicators, which helps to understand the user's physiological state. Combining multimodal physiological indicators with multimodal physiological data to analyze and obtain the user's state distribution data allows for a more accurate grasp of the user's physiological characteristics in different states. Then, decisions are made based on the state distribution data and a pre-trained meta-learning model to obtain ensemble learning parameters, enabling the model to flexibly adjust according to the state distribution of different users, improving the model's adaptability and accuracy. A blood glucose fusion prediction model is constructed based on the ensemble learning parameters, and relevant data is input to obtain blood glucose concentration monitoring values. The ensemble learning method integrates the advantages of multiple models, further improving the accuracy and stability of blood glucose monitoring.

[0017] The above technical solution solves the problem that existing non-invasive blood glucose monitoring technologies are easily affected by external environmental factors and have poor monitoring effects. It provides a more reliable, accurate and comfortable solution for clinical and daily blood glucose monitoring. Patients only need to wear wearable devices to obtain blood glucose concentration monitoring values ​​in real time and non-invasively, which improves the convenience of life. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a non-invasive blood glucose monitoring method that integrates multimodal physiological parameters, as provided in an embodiment of this application.

[0020] Figure 2 This is a schematic diagram of the structure of a non-invasive blood glucose monitoring system that integrates multimodal physiological parameters, provided in an embodiment of this application.

[0021] The components represented by each number in the attached diagram are explained below:

[0022] Data acquisition module 11, data analysis module 12, parameter acquisition module 13, concentration monitoring module 14. Detailed Implementation

[0023] This application provides a non-invasive blood glucose monitoring method and system that integrates multimodal physiological parameters, addressing the technical problem that existing non-invasive blood glucose monitoring technologies are easily affected by external environmental factors, failing to meet the needs of clinical and daily monitoring, resulting in poor blood glucose monitoring performance.

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0025] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0026] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use this application. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that this application can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid unnecessarily obscuring the description of this application. Therefore, this application is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0027] Example 1, as Figure 1 As shown, this application provides a non-invasive blood glucose monitoring method that integrates multimodal physiological parameters, including:

[0028] S10: Non-invasively and synchronously collect multimodal physiological data of the target user through wearable devices, and analyze the multimodal physiological data to obtain multimodal physiological indicators;

[0029] In this embodiment, the wearable device first collects multimodal physiological data of the target user simultaneously. The wearable device is equipped with a variety of sensors and can collect multimodal physiological data such as heart rate, blood pressure, and blood oxygen saturation at the same time.

[0030] After collecting multimodal physiological data, data analysis algorithms are used for processing. For example, by analyzing heart rate data in the time and frequency domains, physiological indicators such as heart rate variability can be obtained; statistical analysis of blood pressure data can yield information such as the average blood pressure value and fluctuation range. The various physiological indicators obtained from the analysis are then integrated to obtain a multimodal physiological signature dataset.

[0031] Multimodal physiological indicators can describe the physiological characteristics of target users from multiple dimensions, providing a rich and accurate data foundation for subsequent blood glucose monitoring and analysis. Then, by combining multimodal physiological indicators with multimodal physiological data, further analysis can be conducted to obtain user state distribution data.

[0032] Specifically, step S10 in the method includes:

[0033] Wearable devices are used to synchronously collect multimodal physiological data of target users, including at least photoplethysmography (PPG) signals, electrocardiogram (ECG) signals, and behavioral signal data derived from accelerometers.

[0034] The photoplethysmography (PPG) signal and the electrocardiogram (ECG) signal are preprocessed, wherein the preprocessing includes filtering and noise reduction and baseline drift correction.

[0035] Based on the preprocessing results, joint feature analysis is performed to extract heart rate features, time-domain features of heart rate variability, and frequency-domain features of heart rate variability.

[0036] Based on the preprocessing results, the pulse wave conduction time is calculated and obtained, and then combined with the heart rate features, the heart rate variability time-domain features, and the heart rate variability frequency-domain features to output the multimodal physiological indicators.

[0037] In this embodiment, firstly, photoplethysmography (PPG) signals, electrocardiogram (ECG) signals, and behavioral signal data derived from an accelerometer are synchronously collected by a wearable device. The PPG signals contain cardiovascular information; the ECG signals reflect the electrical activity of the heart; and the behavioral signal data reflects the user's daily activity status.

[0038] Secondly, the photoplethysmography (PPG) and electrocardiogram (ECG) signals are preprocessed. Filtering and denoising can remove noise interference from the signals, making them purer. Baseline drift correction can ensure the accuracy and stability of the signals and avoid the impact of baseline fluctuations on subsequent feature extraction.

[0039] Next, after preprocessing, joint feature analysis is performed. Heart rate features can intuitively reflect the speed of the heartbeat; heart rate variability time-domain features and frequency-domain features further reveal the regulatory function of the cardiac autonomic nervous system; and the calculation of pulse wave conduction time is related to the elasticity and function of the cardiovascular system. This is combined with the extracted heart rate features, heart rate variability time-domain features, and heart rate variability frequency-domain features to output a multimodal physiological indicator set, which comprehensively describes the physiological state of the target user from multiple perspectives.

[0040] Specifically, different physiological characteristics manifest differently under different states. By comprehensively analyzing these characteristics, it is possible to more accurately determine whether a user is in deep sleep, light sleep, REM sleep, or awake, thus providing more precise state information for blood glucose monitoring. Simultaneously, multimodal physiological indicator collection also provides input data for ensemble learning-based blood glucose fusion prediction models, helping to improve the accuracy and reliability of blood glucose concentration monitoring values.

[0041] The above-mentioned methods for acquiring multimodal physiological indicators provide a foundation for subsequent analysis of user state distribution data by combining multimodal physiological data.

[0042] S20: Combine the multimodal physiological indicators and the multimodal physiological data to analyze and obtain the user's state distribution data;

[0043] In this embodiment, multimodal physiological indicators and multimodal physiological data are deeply fused and analyzed. Since multimodal physiological indicators describe the physiological characteristics of the target user from multiple dimensions, and multimodal physiological data contains the user's real-time physiological signals, the combination of the two can more comprehensively reflect the user's physiological state.

[0044] The specific analysis process can employ machine learning algorithms, such as cluster analysis. Clustering algorithms classify multimodal physiological indicators and data, grouping data points with similar physiological characteristics into the same category, with each category representing a user state.

[0045] For example, data points with low heart rate, stable blood pressure, and behavioral signal data indicating a resting state are classified as sleep state; data points with high heart rate, fluctuating blood pressure, and behavioral signal data indicating an active state are classified as exercise state.

[0046] Furthermore, to improve the accuracy of state distribution data, historical user data and prior knowledge can be incorporated. For example, based on users' past sleep habits and time patterns, the clustering results can be further adjusted and optimized. If users are typically asleep between 11 PM and 7 AM, then the static state clusters obtained during this time period can be more clearly identified as sleep states.

[0047] Blood sugar levels are affected by different factors in different states. For example, during exercise, the body's energy consumption increases, and more blood sugar is used, leading to a decrease in blood sugar levels. During sleep, the body's metabolic rate decreases, and blood sugar consumption decreases accordingly, but the liver releases a certain amount of glucose to maintain stable blood sugar levels.

[0048] Specifically, step S20 in the method includes:

[0049] Feature analysis is performed on the behavioral signal data and the photoplethysmography (PPG) signal to establish a baseline model of the target user's behavior.

[0050] The user behavior baseline model is compared with the standard behavior baseline model, and the behavior residuals are calculated.

[0051] The preset state partitioning rules are modified based on the behavioral residuals to obtain adaptive state partitioning rules;

[0052] Based on the adaptive state division rules, the behavioral signal data, and the photoplethysmography (PPG) signal, the target user's state is divided, and corresponding state distribution data is generated.

[0053] In this embodiment, since behavioral signals and photoplethysmography (PPG) signals can reflect user behavior and physiological state information, feature analysis is first performed on the behavioral signal data and PPG signals. By extracting and analyzing features such as motion amplitude and frequency in the behavioral signal data and features such as peaks and troughs in the PPG signals, a baseline model of the target user's behavior is constructed. This model can be regarded as a typical representation of the user's daily behavior and physiological state.

[0054] Secondly, the user behavior baseline model is compared with the standard behavior baseline model to calculate the behavioral residuals. The standard behavior baseline model is built upon a large amount of behavioral data and physiological signals from the same or similar groups and is generally representative. By comparing the differences between the two, the behavioral residuals are obtained, which reflect the degree of deviation between the target user and the standard group in terms of behavior and physiological state.

[0055] Next, the preset state classification rules are modified based on the behavioral residuals to obtain adaptive state classification rules. The preset state classification rules are general rules, but each user has unique behavioral and physiological characteristics, and using general rules may not accurately classify user states. Modifying them using behavioral residuals allows the state classification rules to better adapt to the characteristics of the target user.

[0056] Finally, based on adaptive state segmentation rules, behavioral signal data, and photoplethysmography (PPG) signals, the target user's state is segmented, and corresponding state distribution data is generated. Using the revised rules, combined with real-time acquired behavioral signal data and PPG signals, the user's state (e.g., exercise, rest, sleep) is determined. The proportion of data in each state is then statistically analyzed to generate state distribution data. Since blood glucose levels change differently in different states, accurate state distribution data provides more precise input for the blood glucose fusion prediction model, thereby improving the accuracy and reliability of blood glucose concentration monitoring values ​​and better meeting the needs of clinical and daily blood glucose monitoring.

[0057] Specifically, based on the adaptive state division rules, the behavioral signal data, and the photoplethysmography (PPG) signal, the target user's state is divided, and corresponding state distribution data is generated, including:

[0058] Acquire accelerometer data and the photoplethysmography signal within multiple preset prediction periods;

[0059] Using the prediction period as the analysis window, iterative state identification based on the adaptive state partitioning rule is performed, and the confidence screening of the iterative identification results is carried out to form the original state distribution dataset.

[0060] Based on the original state distribution dataset, regression analysis is used to model the state distribution and obtain the state distribution data.

[0061] In this embodiment, firstly, accelerometer data and photoplethysmography (PPG) signals are acquired within multiple preset prediction periods. Accelerometer data reflects the user's motion state, while PPG signals contain cardiovascular physiological information. The combination of the two can reflect the user's physiological and behavioral state from different perspectives.

[0062] Secondly, using the prediction period as the analysis window, iterative state recognition based on adaptive state partitioning rules is performed. Within each analysis window, accelerometer data and photoplethysmography (PPG) signals are analyzed according to the adaptive state partitioning rules to determine the user's state within that time period. Through multiple iterative recognitions, changes in the user's state are captured. Then, the iterative recognition results are subjected to confidence filtering, removing recognition results with low confidence and retaining only data with high confidence, thus forming the original state distribution dataset. During the iterative recognition process, data points with ambiguous state recognition results or significant differences from the states in preceding and following time periods are deemed to have low confidence and are discarded.

[0063] For example, a global distribution probability analysis is performed on the iterative identification results to obtain the distribution probability of the REM sleep stage in each time period. Multiple time periods with a distribution probability of less than 1% are identified, and periodic samples with REM sleep stage markers in the iterative segmentation results are removed accordingly.

[0064] Finally, based on the original state distribution dataset, regression analysis was used to model the state distribution. Regression analysis can identify the relationship between the state distribution and various physiological and behavioral data, thereby establishing a model describing the user's state distribution and obtaining state distribution data.

[0065] State distribution data clearly shows the probability and proportion of a user being in various states at different times, providing input information for subsequent blood glucose fusion prediction models. Through these methods, the accuracy and reliability of blood glucose concentration monitoring values ​​are improved, providing stronger support for clinical and daily blood glucose monitoring. This helps doctors better understand patients' blood glucose changes and develop more reasonable treatment plans, while also allowing patients to more conveniently and accurately grasp their blood glucose levels and improve their quality of life.

[0066] S30: Make a decision based on the state distribution data and the pre-trained meta-learning model to obtain ensemble learning parameters, wherein the ensemble learning parameters include at least the number of ensemble models, the model grouping ratio, and the ensemble path parameters;

[0067] In this embodiment, state distribution data is input into a pre-trained meta-learning model. The meta-learning model, based on its own learning mechanism and existing knowledge reserves, performs in-depth analysis and reasoning on the state distribution data. Through learning from a large amount of historical data, the meta-learning model identifies the optimal ensemble learning strategy corresponding to different state distribution data.

[0068] Furthermore, during the decision-making process, the meta-learning model first determines the appropriate number of ensemble models based on the characteristics of the state distribution data. If the state distribution data is complex, containing various types of state changes and feature information, then more ensemble models are needed to capture the information and improve the accuracy of blood glucose prediction; conversely, if the state distribution data is relatively simple, fewer ensemble models may be sufficient to complete the task, while also reducing computational costs and time.

[0069] Regarding the model grouping ratio, the meta-learning model considers the characteristics and importance of data under different states. For example, in certain states, certain physiological characteristics have a more significant impact on blood glucose changes. Therefore, when grouping, more resources will be allocated to the model group related to that physiological characteristic to ensure more accurate information capture.

[0070] Determining the ensemble path parameters involves effectively integrating the outputs of various ensemble models. The meta-learning model selects an appropriate ensemble path, such as weighted averaging or voting, based on the characteristics of the state distribution data, so that the final ensemble result can reflect the true blood glucose situation to the greatest extent.

[0071] By employing a decision-making process based on state distribution data and a pre-trained meta-learning model, accurate and reasonable ensemble learning parameters are provided for subsequent ensemble learning, thereby improving the performance of the ensemble learning model and the accuracy of blood glucose monitoring. Subsequently, based on these ensemble learning parameters, a specific ensemble learning model is constructed and trained to more accurately predict the user's blood glucose concentration.

[0072] Specifically, step S30 in the method includes:

[0073] Obtain sample state distribution data and sample blood glucose concentration from multiple sample users, and calculate the data entropy of the sample state distribution data;

[0074] When the data entropy is greater than a preset entropy threshold, multiple sample prediction training tasks based on ensemble learning are constructed according to the sample state distribution data and the sample blood glucose concentration, and model training is performed accordingly.

[0075] When the model training meets the preset prediction performance, the sample ensemble learning parameters corresponding to the multiple sample prediction training tasks are output.

[0076] Based on the sample state distribution data and the sample ensemble learning parameters, a meta-learning model is constructed and trained to obtain the ensemble learning parameters. The state distribution data is then input into the meta-learning model to obtain the ensemble learning parameters.

[0077] In this embodiment, firstly, sample state distribution data and sample blood glucose concentrations of multiple sample users are acquired, and the data entropy of the sample state distribution data is calculated. Data entropy is an indicator that measures the degree of uncertainty and disorder in data; calculating data entropy helps to understand the complexity of the sample state distribution data. When the data entropy is greater than a preset entropy threshold, it indicates that the sample state distribution data is relatively complex and has high uncertainty. At this point, based on the sample state distribution data and sample blood glucose concentrations, multiple sample prediction training tasks based on ensemble learning are constructed, and corresponding model training is performed. When constructing the sample prediction training tasks, the relationship between the physiological characteristics and blood glucose concentrations of different sample users in different states is considered, and a suitable training task is formulated for each sample user.

[0078] For example, the training tasks differ depending on the different sample state distributions, so the data of each individual user is used as a training task.

[0079] Furthermore, during model training, the model parameters are continuously adjusted to enable the model to better fit the sample data. When the model training meets the preset prediction performance, that is, when the model's prediction error on the sample data reaches an acceptable range, the ensemble learning parameters for multiple sample prediction training tasks are output. The ensemble learning parameters include the number of ensemble models, the model grouping ratio, and the ensemble path parameters, etc.

[0080] Secondly, based on the sample state distribution data and sample ensemble learning parameters, a meta-learning model is constructed and trained to obtain the model. The role of the meta-learning model is to learn how to select appropriate ensemble learning parameters according to different state distribution data. Through learning from a large amount of sample data, it identifies the optimal ensemble learning strategy corresponding to different state distribution data.

[0081] Finally, the state distribution data of the target user is input into the trained meta-learning model. The meta-learning model will output appropriate ensemble learning parameters for the target user based on its own learning mechanism and existing knowledge reserves.

[0082] For example, the specific steps for building and training a meta-learning model based on a neural network are as follows:

[0083] First, data acquisition involves collecting sample state distribution data and sample ensemble learning parameters.

[0084] Secondly, the model is built by taking the sample state distribution data and sample ensemble learning parameters as inputs, and the predicted sample ensemble learning parameters as outputs. The number of nodes in the input layer is equal to the dimension of the input features. For example, if there are two features (sample state distribution data and sample ensemble learning parameters), the input layer contains two nodes. One to three hidden layers are set, with the number of nodes in each layer adjusted experimentally, such as 64 or 32. The activation function is ReLU. The number of nodes in the output layer is equal to the number of generated learning parameters. For example, if prediction time requires one node, the output layer generally does not use an activation function and directly outputs continuous values.

[0085] Then, the model is trained. In each training iteration, the Adam optimizer and mean squared error (MSE) loss function are used to build the training framework. The batch size is set to 32 and the total number of training rounds is 50. An early stopping mechanism (patience=5) is introduced. When the validation set loss does not decrease for 5 consecutive rounds, the training process is automatically terminated, and the trained meta-learning model is obtained.

[0086] By utilizing the meta-learning model to learn and analyze sample data, specifically by predicting the parameters of the ensemble learning model that will be truly used for blood glucose prediction through state distribution, we can ensure a good starting point when training the ensemble learning model. This provides the target user with accurate and reasonable ensemble learning parameters, thereby improving the performance of the ensemble learning model and the accuracy of blood glucose monitoring, and providing stronger support for clinical and daily blood glucose monitoring.

[0087] S40: Based on the ensemble learning parameters, construct an ensemble learning-based blood glucose fusion prediction model, and input the state distribution data, the multimodal physiological indicators, and the multimodal physiological data into the blood glucose fusion prediction model to obtain blood glucose concentration monitoring values.

[0088] In this embodiment, an ensemble learning-based blood glucose fusion prediction model is constructed based on ensemble learning parameters. Ensemble learning combines multiple weak learners into a strong learner. By setting the number of ensemble models, the model grouping ratio, and the ensemble path parameters, the advantages of each model are leveraged to improve the overall performance of the model.

[0089] When constructing a blood glucose fusion prediction model, the number of ensemble models determines the number of sub-models involved in the fusion. More ensemble models can capture more data features and patterns, but also increase computational complexity and training time; fewer ensemble models may not be able to fully extract data information. The model grouping ratio allocates resources rationally among the model groups based on the characteristics and importance of data under different states, to better adapt to the patterns of blood glucose changes under different states. The ensemble path parameters determine how to integrate the outputs of the various ensemble models; common methods include weighted averaging and voting. Choosing an appropriate ensemble path can make the final prediction result closer to the actual blood glucose situation.

[0090] Furthermore, the state distribution data, multimodal physiological indicators, and multimodal physiological data are input into the constructed blood glucose fusion prediction model. The state distribution data clearly shows the probability and proportion of a user being in various states at different times; the multimodal physiological indicators include multiple feature indicators extracted from different physiological signals, such as heart rate, blood pressure, and blood oxygen saturation, reflecting the body's physiological state from multiple dimensions; and the multimodal physiological data are the specific measurements of these physiological indicators, serving as the direct basis for the model's predictions.

[0091] Secondly, the blood glucose fusion prediction model performs comprehensive analysis and processing of the input data. First, it adjusts the weights and processing methods for physiological data under different states based on the state distribution data.

[0092] For example, during exercise, more attention is paid to physiological indicators related to energy consumption and metabolism; during sleep, the focus is on indicators related to basal metabolism and blood glucose regulation.

[0093] Then, the model utilizes the correlations between multimodal physiological indicators and multimodal physiological data to uncover key factors and potential patterns affecting blood glucose concentration. Through learning and training on a large amount of historical data, a complex mapping relationship between key factors and blood glucose concentration is established, thereby accurately predicting the current blood glucose concentration and obtaining blood glucose concentration monitoring values.

[0094] By leveraging the advantages of state distribution data, multimodal physiological indicators, and multimodal physiological data, the blood glucose fusion prediction model based on ensemble learning can more comprehensively and accurately reflect the physiological and behavioral state of the human body, thereby improving the accuracy and reliability of blood glucose concentration monitoring values.

[0095] Specifically, step S40 in the method includes:

[0096] With the number of integrated models as a constraint, multiple shallow basic models are configured differently;

[0097] Based on the intrinsic characteristics of the target user and combined with a preset feature neighborhood, prior sample data is obtained, wherein the prior sample data includes prior sample physiological index collection, prior sample physiological data and prior sample blood glucose concentration.

[0098] The prior sample data is divided into sample groups based on the state distribution data, and multiple shallow basic models are divided into model groups based on the model grouping ratio. The correlation between the sample grouping results and the model grouping results is established.

[0099] Based on the aforementioned correlation, using the sample grouping results as training data, supervised training is performed on the model grouping results, and the trained model grouping results are integrated according to the integration path parameters to obtain the blood glucose fusion prediction model.

[0100] The integration path parameters include at least the integration weights.

[0101] In this embodiment, firstly, multiple shallow base models are configured differently, constrained by the number of integrated models. Different shallow base models can analyze and process data from different perspectives; for example, some models are good at processing time series data, while others have a better fitting ability for image data. Through differentiated configuration, the advantages of each model are fully utilized, improving the overall model performance.

[0102] Secondly, based on the intrinsic characteristics of the target users and combined with a pre-defined feature neighborhood, prior sample data is obtained. The intrinsic characteristics of the users include basic information such as age, gender, and physical condition. The pre-defined feature neighborhood is a range of features related to the intrinsic characteristics, statistically derived from a large amount of sample data. Prior sample data includes prior sample physiological indicators, prior sample physiological data, and prior sample blood glucose concentration.

[0103] Then, the prior sample data is grouped according to the state distribution data. The state distribution data reflects the user's state at different time periods, and the physiological data under different states may have different effects on blood glucose concentration. Therefore, grouping the prior sample data according to the state distribution allows the model to better adapt to the blood glucose change patterns under different states.

[0104] Meanwhile, multiple shallow basic models are grouped according to the model grouping ratio, and the correlation between the sample grouping results and the model grouping results is established to ensure that each model group has corresponding sample data, thereby improving the relevance of model training.

[0105] Subsequently, based on the association relationships, the sample grouping results were used as training data to conduct supervised training of the model corresponding to the grouping results. During training, prior sample blood glucose concentrations were used as labels, and the model parameters were continuously adjusted to make the model's output as close as possible to the true blood glucose concentration. During training, optimization algorithms such as stochastic gradient descent can be employed to improve training efficiency and the model's generalization ability.

[0106] Finally, based on the grouping results of the trained models using the ensemble path parameters, a blood glucose fusion prediction model is obtained. The ensemble path parameters include at least ensemble weights, which determine the contribution of each model group to the final prediction result. Optimal ensemble weights can be determined using methods such as cross-validation, ensuring that the final prediction result reflects the true blood glucose level to the greatest extent possible.

[0107] By constructing a blood glucose fusion prediction model using the above methods, and utilizing multimodal physiological data and state distribution information, the accuracy and reliability of blood glucose concentration monitoring values ​​can be improved, providing strong support for the health management of diabetic patients.

[0108] The state distribution data includes at least the discrete distribution of the duration of the user's deep sleep, light sleep, REM sleep, and wakefulness stages.

[0109] In this embodiment, the discrete distribution of the duration of deep sleep, light sleep, REM sleep, and wakefulness stages in the state distribution data can provide important reference information for blood glucose monitoring. Different sleep stages and wakefulness states result in differences in physiological activities and metabolic levels, which in turn have different effects on blood glucose concentration.

[0110] For example, during deep sleep, the body's metabolic rate decreases, glucose consumption is relatively reduced, and blood glucose concentration may be relatively stable. However, during REM sleep, brain activity is high, which may cause changes in hormone levels, thus affecting blood glucose concentration. During wakefulness, factors such as activity level and food intake can also cause fluctuations in blood glucose concentration.

[0111] By incorporating the discrete distribution of time duration into the blood glucose monitoring model, the relationship between physiological data and blood glucose concentration under different states can be analyzed. When grouping samples, prior sample data can be divided into different groups based on the state distribution data, enabling the model to better learn the patterns of physiological characteristics and blood glucose changes under different states. The determination of the model group ratio can also be based on the importance of different states and data characteristics, allocating resources to each model group to improve the model's accuracy in predicting blood glucose under different states.

[0112] Furthermore, a non-invasive blood glucose monitoring method integrating multimodal physiological parameters also includes:

[0113] Sweat data of the target user is acquired through an external sensor group in the target scene, and sweat data features are extracted;

[0114] Based on the sweat data characteristics, corresponding sample sweat feature data are collected, and the blood glucose fusion prediction model is subjected to dimensionality-incremental learning based on the sample sweat feature data.

[0115] Input the sweat data features, the state distribution data, the multimodal physiological indicators, and the multimodal physiological data into the blood glucose fusion prediction model after dimensionality-incremental learning to obtain blood glucose concentration monitoring values.

[0116] In this embodiment, firstly, sweat data of the target user is acquired through an external sensor array in the target scene. Sweat contains various chemical components related to human physiological state, such as glucose, lactic acid, and electrolytes. Changes in the content of these chemical components can reflect the body's metabolism and blood sugar levels. The sweat data features are extracted by screening out features related to blood sugar concentration from the chemical components, such as the concentration of glucose in sweat and the proportion of certain electrolytes.

[0117] Secondly, based on the characteristics of sweat data, corresponding sample sweat feature data are collected. Sample sweat feature data is a set of features extracted from sweat data collected from different individuals under different physiological conditions. It contains a wider range of sweat feature information, providing information for the learning of the blood glucose fusion prediction model.

[0118] Furthermore, based on the sample sweat feature data, the blood glucose fusion prediction model is subjected to dimensionality-incremental learning. Dimensionality-incremental learning refers to adding new dimensions related to sweat data features to the original model input feature dimensions, so that the model can take into account more physiological information. Incremental learning, on the other hand, updates and optimizes the model using new sample sweat feature data without destroying the original model knowledge, so that the model can adapt to new data patterns and changes.

[0119] Next, sweat data features, state distribution data, multimodal physiological indicators, and multimodal physiological data are input into the blood glucose fusion prediction model after dimensionality-incremental learning. At this point, the model can not only use the original state distribution data and multimodal physiological information to predict blood glucose, but also combine the newly added sweat data features to analyze the physiological state of the human body from more angles and dimensions.

[0120] Sweat data features provide additional information to the model, enabling it to more accurately capture changes in blood glucose concentration, especially in special circumstances such as after exercise or eating, where sweat data features may more sensitively reflect blood glucose fluctuations.

[0121] The dimensionality-incremental learning-based blood glucose fusion prediction model performs comprehensive and in-depth analysis and processing of all input data. The non-invasive blood glucose monitoring method, combined with sweat data characteristics, further improves the accuracy and reliability of blood glucose monitoring, providing more precise health management for diabetic patients and offering a more effective means for clinical blood glucose monitoring.

[0122] In summary, compared to existing technologies, this application provides suitable parameters for the ensemble learning-based blood glucose fusion prediction model through a meta-learning model. It fully considers the physiological characteristics and blood glucose variation patterns of the human body under different states by utilizing state distribution data, multimodal physiological indicator sets, and multimodal physiological data. Furthermore, by incorporating sweat data features into the monitoring system, and further enriching the model's input features through dimensionality-incremental learning, blood glucose monitoring becomes more comprehensive and accurate.

[0123] In summary, the embodiments of this application have at least the following technical effects:

[0124] This application provides a non-invasive blood glucose monitoring method that integrates multimodal physiological parameters. First, a wearable device non-invasively and synchronously collects multimodal physiological data from the target user, avoiding the pain and infection risks associated with traditional invasive monitoring. Second, the collected data is analyzed to obtain multimodal physiological indicators, which helps to understand the user's physiological state. Combining multimodal physiological indicators with multimodal physiological data to analyze and obtain the user's state distribution data allows for a more accurate grasp of the user's physiological characteristics in different states. Then, decisions are made based on the state distribution data and a pre-trained meta-learning model to obtain ensemble learning parameters. This allows the model to flexibly adjust according to the state distribution of different users, improving the model's adaptability and accuracy. A blood glucose fusion prediction model is constructed based on the ensemble learning parameters, and relevant data is input to obtain blood glucose concentration monitoring values. The ensemble learning method integrates the advantages of multiple models, further improving the accuracy and stability of blood glucose monitoring. The above technical solution solves the problem that existing non-invasive blood glucose monitoring technologies are easily affected by external environmental factors and have poor monitoring effects. It provides a more reliable, accurate and comfortable solution for clinical and daily blood glucose monitoring. Patients only need to wear wearable devices to obtain blood glucose concentration monitoring values ​​in real time and non-invasively, which improves the convenience of life.

[0125] Example 2, as Figure 2 As shown, based on the same inventive concept as the non-invasive blood glucose monitoring method integrating multimodal physiological parameters provided in Embodiment 1, this application also provides a non-invasive blood glucose monitoring system integrating multimodal physiological parameters, including:

[0126] The data acquisition module 11 is used to non-invasively and synchronously collect multimodal physiological data of the target user through a wearable device, and analyze the multimodal physiological data to obtain multimodal physiological indicators.

[0127] Data analysis module 12 is used to combine the multimodal physiological indicators and the multimodal physiological data to analyze and obtain the user's state distribution data;

[0128] The parameter acquisition module 13 is used to make decisions based on the state distribution data and the pre-trained meta-learning model to obtain ensemble learning parameters, wherein the ensemble learning parameters include at least the number of ensemble models, the model grouping ratio, and the ensemble path parameters.

[0129] The concentration monitoring module 14 is used to construct a blood glucose fusion prediction model based on the ensemble learning parameters, and input the state distribution data, the multimodal physiological indicators and the multimodal physiological data into the blood glucose fusion prediction model to obtain blood glucose concentration monitoring values.

[0130] In one embodiment, the data acquisition module 11 is specifically used for:

[0131] Wearable devices are used to synchronously collect multimodal physiological data of target users, including at least photoplethysmography (PPG) signals, electrocardiogram (ECG) signals, and behavioral signal data derived from accelerometers.

[0132] The photoplethysmography (PPG) signal and the electrocardiogram (ECG) signal are preprocessed, wherein the preprocessing includes filtering and noise reduction and baseline drift correction.

[0133] Based on the preprocessing results, joint feature analysis is performed to extract heart rate features, time-domain features of heart rate variability, and frequency-domain features of heart rate variability.

[0134] Based on the preprocessing results, the pulse wave conduction time is calculated and obtained, and then combined with the heart rate features, the heart rate variability time-domain features, and the heart rate variability frequency-domain features to output the multimodal physiological indicators.

[0135] In one embodiment, the data analysis module 12 is specifically used for:

[0136] Feature analysis is performed on the behavioral signal data and the photoplethysmography (PPG) signal to establish a baseline model of the target user's behavior.

[0137] The user behavior baseline model is compared with the standard behavior baseline model, and the behavior residuals are calculated.

[0138] The preset state partitioning rules are modified based on the behavioral residuals to obtain adaptive state partitioning rules;

[0139] Based on the adaptive state division rules, the behavioral signal data, and the photoplethysmography (PPG) signal, the target user's state is divided, and corresponding state distribution data is generated.

[0140] Furthermore, in one embodiment, the target user's state is divided based on the adaptive state division rule, the behavioral signal data, and the photoplethysmography (PPG) signal, and corresponding state distribution data is generated, including:

[0141] Acquire accelerometer data and the photoplethysmography signal within multiple preset prediction periods;

[0142] Using the prediction period as the analysis window, iterative state identification based on the adaptive state partitioning rule is performed, and the confidence screening of the iterative identification results is carried out to form the original state distribution dataset.

[0143] Based on the original state distribution dataset, regression analysis is used to model the state distribution and obtain the state distribution data.

[0144] In one embodiment, the parameter acquisition module 13 is specifically used for:

[0145] Obtain sample state distribution data and sample blood glucose concentration from multiple sample users, and calculate the data entropy of the sample state distribution data;

[0146] When the data entropy is greater than a preset entropy threshold, multiple sample prediction training tasks based on ensemble learning are constructed according to the sample state distribution data and the sample blood glucose concentration, and model training is performed accordingly.

[0147] When the model training meets the preset prediction performance, the sample ensemble learning parameters corresponding to the multiple sample prediction training tasks are output.

[0148] Based on the sample state distribution data and the sample ensemble learning parameters, a meta-learning model is constructed and trained to obtain the ensemble learning parameters. The state distribution data is then input into the meta-learning model to obtain the ensemble learning parameters.

[0149] In one embodiment, the concentration monitoring module 14 is specifically used for:

[0150] With the number of integrated models as a constraint, multiple shallow basic models are configured differently;

[0151] Based on the intrinsic characteristics of the target user and combined with a preset feature neighborhood, prior sample data is obtained, wherein the prior sample data includes prior sample physiological index collection, prior sample physiological data and prior sample blood glucose concentration.

[0152] The prior sample data is divided into sample groups based on the state distribution data, and multiple shallow basic models are divided into model groups based on the model grouping ratio. The correlation between the sample grouping results and the model grouping results is established.

[0153] Based on the aforementioned correlation, using the sample grouping results as training data, supervised training is performed on the model grouping results, and the trained model grouping results are integrated according to the integration path parameters to obtain the blood glucose fusion prediction model.

[0154] The integration path parameters include at least the integration weights.

[0155] Furthermore, the state distribution data includes at least the discrete distribution of the duration of the user's deep sleep, light sleep, REM sleep, and wakefulness stages.

[0156] Furthermore, a non-invasive blood glucose monitoring method integrating multimodal physiological parameters also includes:

[0157] Sweat data of the target user is acquired through an external sensor group in the target scene, and sweat data features are extracted;

[0158] Based on the sweat data characteristics, corresponding sample sweat feature data are collected, and the blood glucose fusion prediction model is subjected to dimensionality-incremental learning based on the sample sweat feature data.

[0159] Input the sweat data features, the state distribution data, the multimodal physiological indicators, and the multimodal physiological data into the blood glucose fusion prediction model after dimensionality-incremental learning to obtain blood glucose concentration monitoring values.

[0160] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0161] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0162] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications within the scope of this application.

[0163] Variations, combinations, or equivalents. Clearly, those skilled in the art can make various modifications and variations to this application without departing from its scope.

[0164] The scope of this application. Thus, if these modifications and variations of this application fall within the scope of this application and its equivalents,

[0165] Therefore, this application is intended to include these modifications and variations.

Claims

1. A non-invasive blood glucose monitoring method integrating multimodal physiological parameters, characterized in that, include: Wearable devices are used to non-invasively and synchronously collect multimodal physiological data from target users, and the multimodal physiological data is analyzed to obtain multimodal physiological indicators, including: Wearable devices are used to synchronously collect multimodal physiological data of target users, including at least photoplethysmography (PPG) signals, electrocardiogram (ECG) signals, and behavioral signal data derived from accelerometers. The photoplethysmography (PPG) signal and the electrocardiogram (ECG) signal are preprocessed, wherein the preprocessing includes filtering and noise reduction and baseline drift correction. Based on the preprocessing results, joint feature analysis is performed to extract heart rate features, time-domain features of heart rate variability, and frequency-domain features of heart rate variability. Based on the preprocessing results, the pulse wave conduction time is calculated and obtained, and then combined with the heart rate features, the heart rate variability time-domain features, and the heart rate variability frequency-domain features to output the multimodal physiological indicators. By combining the aforementioned multimodal physiological indicators and the aforementioned multimodal physiological data, user state distribution data is analyzed and obtained, including: Feature analysis is performed on the behavioral signal data and the photoplethysmography (PPG) signal to establish a baseline model of the target user's behavior. The user behavior baseline model is compared with the standard behavior baseline model, and the behavior residuals are calculated. The preset state partitioning rules are modified based on the behavioral residuals to obtain adaptive state partitioning rules; Based on the adaptive state division rules, the behavioral signal data, and the photoplethysmography (PPG) signal, the target user's state is divided, and corresponding state distribution data is generated. Specifically, based on the adaptive state division rules, the behavioral signal data, and the photoplethysmography (PPG) signal, the target user's state is divided, and corresponding state distribution data is generated, including: Acquire accelerometer data and the photoplethysmography signal within multiple preset prediction periods; Using the prediction period as the analysis window, iterative state identification based on the adaptive state partitioning rule is performed, and the confidence screening of the iterative identification results is carried out to form the original state distribution dataset. Based on the original state distribution dataset, a regression analysis method is used to model the state distribution to obtain the state distribution data. The state distribution data includes at least the discrete distribution of the duration of the user's deep sleep, light sleep, REM sleep, and wakefulness stages; Decisions are made based on the state distribution data and the pre-trained meta-learning model to obtain ensemble learning parameters, wherein the ensemble learning parameters include at least the number of ensemble models, the model grouping ratio, and the ensemble path parameters; Based on the ensemble learning parameters, an ensemble learning-based blood glucose fusion prediction model is constructed, and the state distribution data, the multimodal physiological indicators, and the multimodal physiological data are input into the blood glucose fusion prediction model to obtain blood glucose concentration monitoring values.

2. The non-invasive blood glucose monitoring method integrating multimodal physiological parameters as described in claim 1, characterized in that, Decisions are made based on the state distribution data and the pre-trained meta-learning model to obtain ensemble learning parameters, including: Obtain sample state distribution data and sample blood glucose concentration from multiple sample users, and calculate the data entropy of the sample state distribution data; When the data entropy is greater than a preset entropy threshold, multiple sample prediction training tasks based on ensemble learning are constructed according to the sample state distribution data and the sample blood glucose concentration, and model training is performed accordingly. When the model training meets the preset prediction performance, the sample ensemble learning parameters corresponding to the multiple sample prediction training tasks are output. Based on the sample state distribution data and the sample ensemble learning parameters, a meta-learning model is constructed and trained to obtain the ensemble learning parameters. The state distribution data is then input into the meta-learning model to obtain the ensemble learning parameters.

3. The non-invasive blood glucose monitoring method integrating multimodal physiological parameters as described in claim 1, characterized in that, Based on the ensemble learning parameters, an ensemble learning-based blood glucose fusion prediction model is constructed. The state distribution data, the multimodal physiological indicators, and the multimodal physiological data are input into the blood glucose fusion prediction model to obtain blood glucose concentration monitoring values, including: With the number of integrated models as a constraint, multiple shallow basic models are configured differently; Based on the intrinsic characteristics of the target user and combined with a preset feature neighborhood, prior sample data is obtained, wherein the prior sample data includes prior sample physiological index collection, prior sample physiological data and prior sample blood glucose concentration. The prior sample data is divided into sample groups based on the state distribution data, and multiple shallow basic models are divided into model groups based on the model grouping ratio. The correlation between the sample grouping results and the model grouping results is established. Based on the aforementioned correlation, using the sample grouping results as training data, supervised training is performed on the model grouping results, and the trained model grouping results are integrated according to the integration path parameters to obtain the blood glucose fusion prediction model. The integration path parameters include at least the integration weights.

4. The non-invasive blood glucose monitoring method integrating multimodal physiological parameters as described in claim 1, characterized in that, Also includes: Sweat data of the target user is acquired through an external sensor group in the target scene, and sweat data features are extracted; Based on the sweat data characteristics, corresponding sample sweat feature data are collected, and the blood glucose fusion prediction model is subjected to dimensionality-incremental learning based on the sample sweat feature data. Input the sweat data features, the state distribution data, the multimodal physiological indicators, and the multimodal physiological data into the blood glucose fusion prediction model after dimensionality-incremental learning to obtain blood glucose concentration monitoring values.

5. A non-invasive blood glucose monitoring system integrating multimodal physiological parameters, characterized in that, A method for performing a non-invasive blood glucose monitoring method integrating multimodal physiological parameters as described in any one of claims 1-4, comprising: The data acquisition module is used to non-invasively and synchronously collect multimodal physiological data of the target user through a wearable device, and analyze the multimodal physiological data to obtain multimodal physiological indicators, including: Wearable devices are used to synchronously collect multimodal physiological data of target users, including at least photoplethysmography (PPG) signals, electrocardiogram (ECG) signals, and behavioral signal data derived from accelerometers. The photoplethysmography (PPG) signal and the electrocardiogram (ECG) signal are preprocessed, wherein the preprocessing includes filtering and noise reduction and baseline drift correction. Based on the preprocessing results, joint feature analysis is performed to extract heart rate features, time-domain features of heart rate variability, and frequency-domain features of heart rate variability. Based on the preprocessing results, the pulse wave conduction time is calculated and obtained, and then combined with the heart rate features, the heart rate variability time-domain features, and the heart rate variability frequency-domain features to output the multimodal physiological indicators. The data analysis module is used to combine the multimodal physiological indicators and the multimodal physiological data to analyze and obtain the user's state distribution data, including: Feature analysis is performed on the behavioral signal data and the photoplethysmography (PPG) signal to establish a baseline model of the target user's behavior. The user behavior baseline model is compared with the standard behavior baseline model, and the behavior residuals are calculated. The preset state partitioning rules are modified based on the behavioral residuals to obtain adaptive state partitioning rules; Based on the adaptive state division rules, the behavioral signal data, and the photoplethysmography (PPG) signal, the target user's state is divided, and corresponding state distribution data is generated. Specifically, based on the adaptive state division rules, the behavioral signal data, and the photoplethysmography (PPG) signal, the target user's state is divided, and corresponding state distribution data is generated, including: Acquire accelerometer data and the photoplethysmography signal within multiple preset prediction periods; Using the prediction period as the analysis window, iterative state identification based on the adaptive state partitioning rule is performed, and the confidence screening of the iterative identification results is carried out to form the original state distribution dataset. Based on the original state distribution dataset, a regression analysis method is used to model the state distribution to obtain the state distribution data. The state distribution data includes at least the discrete distribution of the duration of the user's deep sleep, light sleep, REM sleep, and wakefulness stages; The parameter acquisition module is used to make decisions based on the state distribution data and the pre-trained meta-learning model to obtain ensemble learning parameters, wherein the ensemble learning parameters include at least the number of ensemble models, the model grouping ratio, and the ensemble path parameters. The concentration monitoring module is used to construct a blood glucose fusion prediction model based on the ensemble learning parameters, and input the state distribution data, the multimodal physiological indicators and the multimodal physiological data into the blood glucose fusion prediction model to obtain blood glucose concentration monitoring values.

Images