Noninvasive blood glucose prediction method and system based on personalized time sequence model

Abstract

The invention discloses a noninvasive blood glucose prediction method and system based on a personalized time sequence model. The method comprises the following steps: acquiring a noninvasive related parameter sample, a real blood glucose data sample and a continuous blood glucose data sample, and training a general time sequence model to obtain a trained general time sequence model; the general time sequence model is used for predicting blood glucose based on historical features and continuous features of blood glucose; and obtaining an individual blood sugar data sample, performing migration training on the general time sequence model according to the individual blood sugar data to obtain a personalized time sequence model, and performing individual blood sugar prediction based on the personalized time sequence model. According to the embodiment of the invention, the personalized time sequence model is obtained through migration training to predict the individual blood glucose, so that the accuracy of blood glucose prediction is improved. The method can be widely applied to the technical field of noninvasive detection.

Description

Technical Field

[0001] The present invention relates to the field of non-invasive detection technology, and in particular to a non-invasive blood glucose prediction method and system based on a personalized time series model. Background Art

[0002] Blood glucose (BG) monitoring technology is a core tool for diabetes management. Human blood glucose monitoring technology encompasses a variety of methods, including invasive, minimally invasive, and non-invasive monitoring techniques. Currently, continuous prediction of non-invasive blood glucose (BG) concentrations relies primarily on data analysis methods and machine learning techniques. Non-invasive monitoring technologies aim to avoid blood collection by inferring blood glucose concentrations using principles such as spectroscopy (such as near-infrared or mid-infrared spectroscopy), Raman spectroscopy, or human metabolic heat integration (MHC). These technologies are currently undergoing continuous exploration and optimization.

[0003] In the prior art, reference BG values are typically obtained through discrete finger-prick blood samples. These samples are then correlated with non-invasive parameters to form a training dataset. A general artificial intelligence model is then trained to calculate BG values from these non-invasive parameters. However, this general calculation method is susceptible to the complex physiological characteristics of diabetic patients and often becomes inaccurate due to individual physiological differences. Summary of the Invention

[0004] The purpose of the present invention is to solve one of the technical problems existing in the prior art to at least a certain extent.

[0005] To this end, the purpose of the present invention is to provide a high-precision non-invasive blood glucose prediction method and system based on a personalized time series model.

[0006] In order to achieve the above technical objectives, one aspect of an embodiment of the present invention provides a non-invasive blood glucose prediction method based on a personalized time series model, comprising the following steps: obtaining non-invasive related parameter samples, real blood glucose data samples, and continuous blood glucose data samples, training a universal time series model, and obtaining a trained universal time series model; the universal time series model is used to predict blood glucose based on historical and continuous features of blood glucose; obtaining individual blood glucose data samples, performing migration training on the universal time series model according to the individual blood glucose data, obtaining a personalized time series model, and performing individual blood glucose prediction based on the personalized time series model. The embodiment of the present application predicts individual blood glucose by obtaining a personalized time series model through migration training, which is conducive to improving the accuracy of blood glucose prediction.

[0007] In some embodiments, the universal time series model includes a first model component, a second model component, and a third model component, and the universal time series model is trained by the following steps:

[0008] Inputting historical non-invasive related parameter samples into the first model component to obtain historical parameter feature samples;

[0009] inputting the continuous blood glucose data sample into the second model component to obtain a blood glucose trend feature sample;

[0010] The current non-invasive related parameter samples, the historical parameter feature samples and the blood glucose trend feature samples are input into the third model component to obtain predicted values, and the universal time series model is trained according to the obtained predicted values to obtain a trained universal time series model.

[0011] In some embodiments, inputting historical non-invasive related parameter samples into the first model component to obtain historical parameter feature samples includes:

[0012] Constructing the first model component based on a single-layer bidirectional GRU network;

[0013] Performing sine-cosine embedding transformation on the historical non-invasive related parameter samples to obtain corresponding time period codes;

[0014] Based on the time period code and the historical non-invasive related parameter samples, a historical parameter feature sample is obtained.

[0015] In some embodiments, the third model component includes a plurality of connection layers; inputting the current non-invasive related parameter sample, the historical parameter feature sample, and the blood glucose trend feature sample into the third model component to obtain a predicted value includes:

[0016] Based on the connection layer, the current non-invasive related parameter samples, the historical parameter feature samples and the blood glucose trend feature samples are weightedly fused to obtain a predicted value.

[0017] In some embodiments, the method further comprises:

[0018] Embedding the personalized time series model into a wearable device to collect individual blood glucose data and corresponding time data;

[0019] An individual continuous blood glucose prediction result is obtained based on the individual blood glucose data and the time data.

[0020] In some embodiments, inputting the continuous blood glucose data sample into the second model component to obtain a blood glucose trend feature sample includes:

[0021] Constructing the second model component based on a single-layer bidirectional GRU network;

[0022] Obtaining discrete blood glucose data samples based on the continuous blood glucose data samples, and determining blood glucose fluctuation status in combination with meal times;

[0023] Performing sine-cosine embedding transformation on the continuous blood glucose data samples to obtain corresponding time period codes;

[0024] The second model component is trained based on the time period code, the historical non-invasive related parameters and the blood glucose fluctuation state to obtain historical parameter feature samples.

[0025] In some embodiments, obtaining non-invasive relevant parameter samples includes:

[0026] Obtain metabolic thermal parameter samples and cardiovascular parameter samples of the group.

[0027] On the other hand, an embodiment of the present invention provides a non-invasive blood glucose prediction system based on a personalized time series model, comprising:

[0028] The first module is used to obtain non-invasive relevant parameter samples, real blood glucose data samples, and continuous blood glucose data samples, and train a universal time series model to obtain a trained universal time series model; the universal time series model is used to predict blood glucose based on historical and continuous features of blood glucose;

[0029] The second module is used to obtain individual blood glucose data samples, perform migration training on the universal time series model according to the individual blood glucose data to obtain a personalized time series model, and perform individual blood glucose prediction based on the personalized time series model.

[0030] On the other hand, an embodiment of the present invention provides a non-invasive blood glucose prediction device based on a personalized time series model, comprising:

[0031] at least one processor;

[0032] at least one memory for storing at least one program;

[0033] When the at least one program is executed by the at least one processor, the at least one processor implements the above-mentioned non-invasive blood glucose prediction method based on the personalized time series model.

[0034] On the other hand, an embodiment of the present invention provides a storage medium storing a program executable by a processor. When the program is executed by the processor, it is used to implement the above-mentioned non-invasive blood glucose prediction method based on the personalized time series model.

[0035] The embodiments of the present application include at least the following beneficial effects: the method provided by the embodiments of the present invention includes: obtaining non-invasive relevant parameter samples, real blood glucose data samples, and continuous blood glucose data samples, training a universal time series model, and obtaining a trained universal time series model; the universal time series model is used to predict blood glucose based on historical and continuous features of blood glucose; obtaining individual blood glucose data samples, performing migration training on the universal time series model based on the individual blood glucose data, obtaining a personalized time series model, and performing individual blood glucose prediction based on the personalized time series model. The embodiments of the present application predict individual blood glucose by obtaining a personalized time series model through migration training, which is conducive to improving the accuracy of blood glucose prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following introduction is made to the drawings of the embodiments of the present invention or the related technical solutions in the prior art. It should be understood that the drawings introduced below are only for the convenience of clearly describing some embodiments of the technical solutions of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without any creative work.

[0037] Figure 1 A flow chart of an embodiment of a non-invasive blood glucose prediction method based on a personalized time series model provided by the present invention;

[0038] Figure 2 A flowchart of an embodiment of the training process of the universal time series model provided by the present invention;

[0039] Figure 3 A flowchart of an embodiment of the process of building a universal timing model provided by the present invention;

[0040] Figure 4 A flowchart of an embodiment of the process of building a personalized time series model provided by the present invention;

[0041] Figure 5 A schematic diagram illustrating an embodiment of the parameter correlation analysis of different test subjects provided by the present invention;

[0042] Figure 6 A schematic diagram showing an embodiment of the pre-training dataset provided by the present invention;

[0043] Figure 7 A schematic diagram showing an embodiment of a personalized patient dataset provided by the present invention;

[0044] Figure 8 A schematic diagram showing an embodiment of the time period coding provided by the present invention;

[0045] Figure 9 A schematic diagram showing the accuracy of blood glucose prediction according to an embodiment of the present invention;

[0046] Figure 10 This is a schematic diagram showing the accuracy of another embodiment of blood glucose prediction provided by the present invention. DETAILED DESCRIPTION

[0047] The embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention and are not to be construed as limiting the present invention. The step numbers in the following embodiments are provided for ease of explanation only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted according to the understanding of those skilled in the art.

[0048] Blood glucose (BG) monitoring is a core tool in diabetes management. Human BG monitoring encompasses a variety of methods, including invasive, minimally invasive, and non-invasive. Currently, continuous prediction of non-invasive BG concentrations relies primarily on data analysis and machine learning techniques. These non-invasive monitoring technologies aim to avoid blood collection and infer BG concentrations using spectroscopy (such as near-infrared or mid-infrared spectroscopy), Raman spectroscopy, or the human metabolic heat integration (MHC) method. These technologies are currently undergoing continuous exploration and optimization. These technologies collect various parameters to estimate the current BG level. In recent years, AI has developed rapidly, with a wide range of prediction algorithms emerging, from shallow learning models based on statistical methods to deep neural networks implemented using large-scale computing resources. These algorithms have been widely applied in various fields. Within the research area of BG prediction, related models can be broadly divided into two categories: one directly estimates BG concentrations through regression analysis; the other focuses on categorical judgment of disease status or risk level. Direct prediction of BG concentrations involves the precise estimation of continuous variables, thus posing greater complexity and challenges in model design and algorithm implementation. Continuous glucose monitoring (CGM) based on physical detection still does not meet clinical standards, showing low accuracy and insufficient specific recognition of glucose. The complex physiological characteristics of the human body, such as tissue properties and blood composition, have large individual differences, which hinders the detection signal. Traditional non-invasive BG sensors are usually manufactured through standardized processes and use common data sets from the population to build general algorithms or computational models. Sensors developed for general applications fail to take into account physiological variations between individuals, resulting in potential inaccuracies between people in actual applications. In addition, most existing non-invasive glucose sensing technologies are limited to detecting discrete blood glucose values, making it difficult to fully reflect complex blood glucose fluctuation trends.

[0049] In traditional calculation methods, reference BG values are typically obtained through discrete finger-prick blood samples, which are then correlated with non-invasive parameters to form a training dataset. A general artificial intelligence model is then trained to calculate the non-invasive parameters as BG values. However, this general calculation method is susceptible to the complex physiological characteristics of diabetic patients and often becomes inaccurate due to physiological differences between individuals.

[0050] When performing non-invasive BG concentration predictions, conventional AI models are essentially transferred from pre-trained AI models to predict patient blood glucose levels. The goal of general AI models built under this traditional model is to adapt to the large number of patients in the market and clinical settings. However, there are natural differences in physiological characteristics between different individuals, resulting in significant differences in the correlation between non-invasive physiological signals and true BG values. Therefore, relying on a single algorithm or universal model to derive BG values for all patients ignores individual heterogeneity, limits accuracy and practicality in practical applications, and fails to address the issue of individual diversity. Different physiological parameters for each specific individual may correspond to different changes in blood glucose trends and should be considered comprehensively.

[0051] In the non-invasive prediction of BG concentration, conventional AI models only consider the different non-invasive physiological parameters at the current moment, and do not consider the parameter characteristics of historical data. This may cause inaccurate model prediction results. The parameter characteristics of historical and current moments should be comprehensively considered.

[0052] In addition to minimizing the average error in predicting absolute blood glucose values, this application also aims to accurately reflect and monitor the overall fluctuation trend of blood glucose through non-invasive sensing. Therefore, we propose a personalized non-invasive blood glucose model (adoptive non-invasive BG model), which is also the personalized time series model in this application.

[0053] The model built in this application collects personalized CGM data sets for specific individuals, and then performs periodic time encoding (including but not limited to sine and cosine time encoding, vector time encoding, etc.) on non-invasive related parameters and BG trend characteristics. This is more conducive to the model learning the periodic fluctuations of the data set, so that the blood glucose data set has the characteristics of a dynamic time series.

[0054] The following describes in detail the non-invasive blood glucose prediction method and system based on the personalized time series model proposed in an embodiment of the present invention with reference to the accompanying drawings. First, the non-invasive blood glucose prediction method based on the personalized time series model proposed in an embodiment of the present invention will be described with reference to the accompanying drawings.

[0055] Reference Figure 1In an embodiment of the present invention, a non-invasive blood glucose prediction method based on a personalized time series model is provided. The non-invasive blood glucose prediction method based on a personalized time series model in the embodiment of the present invention can be applied to a terminal, can be applied to a server, or can be software running in a terminal or a server. The terminal can be a tablet computer, a laptop computer, a desktop computer, etc., but is not limited thereto. The server can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The non-invasive blood glucose prediction method based on a personalized time series model in the embodiment of the present invention mainly includes the following steps:

[0056] S100: Acquire non-invasive relevant parameter samples, real blood glucose data samples, and continuous blood glucose data samples, and train a universal time series model to obtain a trained universal time series model; the universal time series model is used to predict blood glucose based on historical features and continuous features of blood glucose;

[0057] S200: Acquire individual blood glucose data samples, perform migration training on the universal time series model according to the individual blood glucose data to obtain a personalized time series model, and perform individual blood glucose prediction based on the personalized time series model.

[0058] In some possible implementations, non-invasive relevant parameter samples, real blood glucose data samples, and continuous blood glucose data samples may be relevant data of a group.

[0059] In some embodiments, the universal time series model includes a first model component, a second model component, and a third model component, and the universal time series model is trained by the following steps:

[0060] Inputting historical non-invasive related parameter samples into the first model component to obtain historical parameter feature samples;

[0061] inputting the continuous blood glucose data sample into the second model component to obtain a blood glucose trend feature sample;

[0062] The current non-invasive related parameter samples, the historical parameter feature samples and the blood glucose trend feature samples are input into the third model component to obtain predicted values, and the universal time series model is trained according to the obtained predicted values to obtain a trained universal time series model.

[0063] In some embodiments, inputting historical non-invasive related parameter samples into the first model component to obtain historical parameter feature samples includes:

[0064] Constructing the first model component based on a single-layer bidirectional GRU network;

[0065] Performing sine-cosine embedding transformation on the historical non-invasive related parameter samples to obtain corresponding time period codes;

[0066] Based on the time period code and the historical non-invasive related parameter samples, a historical parameter feature sample is obtained.

[0067] In some embodiments, the third model component includes a plurality of connection layers; inputting the current non-invasive related parameter sample, the historical parameter feature sample, and the blood glucose trend feature sample into the third model component to obtain a predicted value includes:

[0068] Based on the connection layer, the current non-invasive related parameter samples, the historical parameter feature samples and the blood glucose trend feature samples are weightedly fused to obtain a predicted value.

[0069] In some embodiments, the method further comprises:

[0070] Embedding the personalized time series model into a wearable device to collect individual blood glucose data and corresponding time data;

[0071] An individual continuous blood glucose prediction result is obtained based on the individual blood glucose data and the time data.

[0072] In some embodiments, inputting the continuous blood glucose data sample into the second model component to obtain a blood glucose trend feature sample includes:

[0073] Constructing the second model component based on a single-layer bidirectional GRU network;

[0074] Obtaining discrete blood glucose data samples based on the continuous blood glucose data samples, and determining blood glucose fluctuation status in combination with meal times;

[0075] Performing sine-cosine embedding transformation on the continuous blood glucose data samples to obtain corresponding time period codes;

[0076] The second model component is trained based on the time period code, the historical non-invasive related parameters and the blood glucose fluctuation state to obtain historical parameter feature samples.

[0077] In some embodiments, obtaining non-invasive relevant parameter samples includes:

[0078] Obtain metabolic thermal parameter samples and cardiovascular parameter samples of the group.

[0079] The prediction method in this application is described in detail below with a specific embodiment:

[0080] The model constructed in this application includes a model component #1 for learning historical non-invasive multi-parameter time series features (including but not limited to GRU network, LSTM, Seq2Seq, Transformer, Bayesian time series network model, etc.), which can learn the non-invasive multi-parameter time series features before the current moment, extract multi-parameter information related to the current blood glucose value from historical time series data, and output it in the form of "historical parameter features".

[0081] The model constructed in this application includes model component #2 for capturing the blood sugar fluctuation trend law (including but not limited to GRU network, LSTM, Seq2Seq, Transformer, Bayesian time series network model, etc.). The input features and output features more directly map the blood sugar fluctuation trend information. Based on the continuous blood sugar fluctuation time series data, the blood sugar fluctuation trend information of the previous time period can be analyzed and output in the form of "blood sugar trend features".

[0082] The model constructed in this application includes feature fusion model component #3 (including but not limited to fully connected neural network, CNN model, GRU model, LSTM model, GAN generative adversarial network model, Fusion model, etc.), which can weightedly fuse the current non-invasive multi-parameter features, "historical parameter features" and "blood sugar trend features", and input them into the hidden layer structure together to finally obtain the correlation between non-invasive multi-parameters and blood sugar values.

[0083] When the model constructed in this application derives blood glucose values, the model will combine the non-invasive multi-parameter indicators at the current moment with historical parameter information and blood glucose trend characteristics to jointly realize the derivation of non-invasive metabolic thermal integration method parameters to blood glucose values.

[0084] The model built in this application is used as a personalized blood glucose sensing method. We only use the non-invasive blood glucose data of different subjects as a personalized transfer training set, instead of using a universal training set containing data from all subject groups. By using the personalized transfer training set to perform transfer enhancement on the pre-trained initialization model, a personalized Adoptive non-invasive BG model specific to different subjects can be quickly obtained. During the transfer enhancement process, the model can efficiently learn the subject's own physiological state and achieve accurate derivation of non-invasive metabolic heat multi-parameters to blood glucose. Finally, the personalized Adoptive non-invasive BG model is reconfigured and embedded into a wearable device, and a personalized non-invasive blood glucose monitoring device specific to different subjects can be obtained. By inputting the current non-invasive related parameters and timeline, the personalized AI model can derive the current BG level. That is, this application also includes: embedding a personalized time series model into a wearable device to continuously measure the individual's blood glucose.

[0085] When the model built in this application derives blood glucose values, the model will combine historical parameter information and blood glucose trend characteristics based on the non-invasive multi-parameter indicators at the current moment to achieve the derivation of non-invasive multi-parameters to blood glucose values. The model built in this application is a personalized blood glucose sensing method. We only use the non-invasive blood glucose data of different subjects as a personalized migration training set, instead of using a universal training set containing data from all subject groups. By using the personalized migration training set to perform migration enhancement on the pre-trained initialization model, a personalized Adoptive non-invasive BG model specific to different subjects can be quickly obtained. During the migration enhancement process, the model can efficiently learn the subject's own physiological state and achieve accurate derivation of non-invasive metabolic heat multi-parameters to blood glucose. Finally, by reconfiguring and embedding the personalized Adoptive non-invasive BG model into a wearable device, a personalized non-invasive blood glucose monitoring device specific to different subjects can be obtained. By inputting the current non-invasive related parameters and timeline, the personalized AI model can derive the current BG level.

[0086] This application compares the results of a universal model with those of a personalized model. The universal dataset is a reorganized MI / NI-KG dataset of different patients based on the time of day of testing. These datasets are identical to conventional discrete datasets. Key information such as age, gender, and diabetes history of each patient is also incorporated into the universal dataset for routine individual calibration. This dataset can be used as a control group.

[0087] The BG error in the Adoptive non-invasive BG model of this application is 12.5±10.3%. Clinically, the standard for BG measurement approved by FDA, MDR or NMPA is MARD<15%, and MARD<10% is considered to be very good accuracy. The MARD calculated value of the non-invasive CGM model developed in this study is <15.3%, which meets the clinical accuracy standard. In comparison, the MARD of ordinary non-invasive CGM is ≈49.1%, and the error of the universal CGM model in the control group of this study is 47%. It is far from the standard for clinical application. The method of the AI model of this application can provide unique opportunities for personalized diagnosis and treatment of diabetic patients.

[0088] Figure 2Schematic diagram of universal non-invasive CGM model training provided in this embodiment. After collecting a universal data set through sensors, the universal AI model is retrained from a conventional AI model. BG monitoring is then performed using a universal non-invasive CGM (the drawings in this application use MHC as an example). In non-invasive CGM, sensors are typically used to collect multi-channel physiological signals from a large number of patients, and then machine learning is used to learn the correlation between non-invasive physiological signals and BG data, so that the non-invasive signals in the test data set can be inferred as BG levels. The purpose of this method is to put a large number of patient data sets into the model for pre-training to obtain a universal AI model.

[0089] Figure 3 The universal AI model construction method provided in this embodiment collects MI / NI-CGM datasets from the population and uses non-invasive related parameters and timelines to train a universal AI model.

[0090] Figure 4 This embodiment provides a personalized AI model building method. After collecting a personalized MI / NI-CGM dataset for a specific individual, the non-invasive related parameters and BG trend characteristics are periodically encoded. The personalized AI model consists of three main parts, where Model#1 (i.e., the first model component) represents the historical related time series feature model component, Model#2 (i.e., the second model component) represents the BG fluctuation trend model component, and Model#3 (i.e., the third model component) represents the non-invasive related parameter model component. By inputting the current non-invasive related parameters and timeline, the personalized AI model can derive the current BG level.

[0091] Figure 5 Correlation analysis was performed on the BG levels and MHC non-invasive parameters of the six participants in the dataset provided in the examples of this application. The results showed that there were significant differences in the datasets between different participants. Building a personalized AI model is very necessary.

[0092] Figure 6 Schematic diagram of the patient pre-training dataset provided in the embodiments of this application. The universal dataset consists of fingerstick blood glucose (BG) test data, universal non-invasive parameter data, and test time. Discrete BG data from fingerstick blood is used as reference BG values for training and transfer learning of universal computing AI models. Non-invasive parameter data collected from sensors is used for non-invasive BG derivation to construct the AI model used for pre-training.

[0093] Figure 7Schematic diagram of a personalized patient dataset provided for an embodiment of the present application. The dataset consists of minimally invasive CGM data, non-invasive CGM data, and time period encoding. Among them, minimally invasive CGM data is collected from the MI-CGM module, which is used to calculate the reference BG value for AI model training and transfer learning. Non-invasive CGM data is collected from the NI-CGM module for non-invasive BG derivation, which contains patient-specific parameters. Based on the personalized MI / NI-BG dataset, a personalized AI model tailored for the individual can be migrated from the pre-trained AI model, thereby achieving accurate non-invasive BG monitoring.

[0094] Figure 8 This is a schematic diagram of the time period encoding of the MI / NI-BG dataset provided in this embodiment. Using the sine-cosine embedding transform, the BG test time can be cyclically encoded on the hour and minute scales, effectively representing the cyclical characteristics of BG values. By integrating the encoded time features, BG values, meal times, and noninvasive parameters, the dataset exhibits dynamic time series characteristics.

[0095] This application uses a personalized AI model for non-invasive CGM. It uses a GRU neural network as model component GRU#1 for historical non-invasive multi-parameter time series features, and uses the GRU neural network to construct a model component GRU#2 for capturing blood sugar fluctuation trends. It also uses a DNN neural network as feature fusion model component #3, and uses the MHC principle as an example for non-invasive multi-parameter collection.

[0096] GRU#1 learns noninvasive multi-parameter time series features prior to the current moment, extracts multi-parameter information related to the current blood glucose level from historical time series data, and outputs this information as "historical parameter features." GRU#2 analyzes blood glucose fluctuation trends over previous periods based on continuous blood glucose fluctuation time series data and outputs this information as "blood glucose trend features." The DNN performs a weighted fusion of the current noninvasive multi-parameter features, "historical parameter features," and "blood glucose trend features," and inputs these features into the hidden layer of the FCNN structure. Ultimately, the correlation between the noninvasive MHC multi-parameters and blood glucose levels is derived. GRU#1 is constructed using a single-layer bidirectional GRU network that uses historical MHC-related time series parameters prior to the current moment as input features. MHC-related and noninvasive related parameters include metabolic thermal parameters (skin surface temperature, skin radiant temperature, ambient temperature, and ambient humidity) and cardiovascular parameters (heart rate, pulse time difference, blood oxygen saturation, systolic blood pressure, diastolic blood pressure, and blood flow velocity). Before being input into GRU#1, MHC-related parameters undergo a sine-cosine embedding transformation at the hourly and minutely scales to encode temporal periods. GRU#1 learns continuous MHC-related parameters from the previous period and extracts time series features related to the current blood glucose value, outputting them as historical blood glucose features. Simultaneously, GRU#2 is constructed using a single-layer bidirectional GRU network, taking continuous blood glucose data from the previous period as input features. Continuous blood glucose data includes reference blood glucose values for individual participants collected by the MI-CGM module. Combined with the participant's meal times, a personalized postprandial blood glucose fluctuation profile can be generated to reflect blood glucose fluctuation patterns. Before being input into GRU#2, continuous blood glucose data undergoes a sine-cosine embedding transformation at the hourly and minutely scales to encode temporal periods. GRU#2 analyzes trends in blood glucose fluctuations and outputs these patterns as blood glucose trend features. The DNN network consists of two to three fully connected layers, taking historical blood glucose features, blood glucose trend features, and current MHC-related parameters as input features. By fusing these features, the DNN can derive correlations between MHC-related parameters and BG values. The input features of the DNN network are also encoded with temporal periods on the hourly and minute scales, which is more conducive to the model learning the periodic fluctuations of the input data. A pre-trained model is pre-built, and the training efficiency of the personalized AI model is improved by pre-training the initialization parameters. The pre-trained model adopts a conventional GRU network architecture and uses only pre-trained discrete non-invasive parameters as input. This pre-training method can accelerate the training and convergence of the personalized AI model and reduce the amount of individual personalized data required. After the personalized AI model is trained, test data from the MI / NI-CGM dataset is input into the model, and BG values are derived from the MHC-related parameters.

[0097] For example, a five-day dataset, including real-world blood glucose data collected by a minimally invasive CGM module (MI-CGM), personalized data collected by a non-invasive CGM module (NI-CGM), continuous blood glucose data, and continuous non-invasive multi-parameter data, was used to build and train an AI model. In this work, we used five days of non-invasive blood glucose data from subjects to train and evaluate the Adoptive non-invasive BG model. Specifically, we used the first 1.5 days of data as the training set for individualized training of the Adoptive non-invasive BG model, while the remaining 3.5 days of data served as the test set to verify the model's accuracy. To improve model training efficiency, we pre-trained the model's initialization parameters using a non-invasive blood glucose pre-training set. This corresponds to the initialization model pre-embedded in the wearable device during its development phase. Based on this pre-trained initialization model, subjects wore the wearable device to collect non-invasive blood glucose data for 1.5 days, which served as the transfer training set for the Adoptive non-invasive BG model and served as the basis for building a personalized AI model. We only use the non-invasive blood glucose data of different subjects as a personalized transfer training set, rather than a universal training set containing data from all subject groups. By using the personalized transfer training set to perform transfer enhancement on the pre-trained initialization model, we can quickly obtain personalized Adoptive non-invasive BG models specific to different subjects. During the transfer enhancement process, the model can efficiently learn the subject's own physiological state and achieve accurate derivation of non-invasive metabolic heat multi-parameters to blood glucose. Finally, by reconfiguring and embedding the personalized Adoptive non-invasive BG model into a wearable device, we can obtain a personalized non-invasive blood glucose monitoring device specific to each subject.

[0098] As a control group, we constructed an AI model for universal sensors, referencing a conventional universal noninvasive blood glucose sensing method. Conventional methods typically use noninvasive blood glucose monitoring devices to piecemeally collect noninvasive parameter blood glucose data from a patient population. The correlation between blood glucose and noninvasive parameters is then constructed using discrete datasets. This conventional universal dataset lacks individual patient data and the temporal characteristics of blood glucose. Therefore, the universal AI model used the first 1.5 days of noninvasive blood glucose data from all subjects as a universal training set. The data from each subject was reorganized and sorted based on the test time of the day and fed into a pretrained GRU model for secondary training. As a standard individual calibration step, basic information such as age and medical history of each patient was also input as individual parameters into the model training. Finally, the last 3.5 days of data from six subjects was used as a test set and fed into the universal AI model to evaluate the accuracy of the universal noninvasive blood glucose monitoring method.

[0099] The BG error (MARD) in the Adoptive non-invasive BG model of this application is 12.5±10.3%. In clinical practice, the standard for BG measurement approved by the FDA, MDR or NMPA is MARD <15%, and MARD <10% is considered to be very good accuracy. The non-invasive CGM model developed in this study has a calculated MARD value of <15.3%, which meets the clinical accuracy standard. In comparison, the MARD of ordinary non-invasive CGM is ≈49.1%, and the error of the universal CGM model in the control group of this study is 47%. It is far from meeting the standard for clinical application.

[0100] Reference Figure 9 Clark error grid analysis was performed on the BG values of all participants in the initial 1.5-day training set. The BG values obtained from the training of (a) the personalized CGM model and (b) the universal CGM model were compared with the reference BG. For the training results of the personalized CGM model, most of the BG value data in the MI / NI-BG training data set of all participants were distributed in area A, meeting the clinical requirement of MARD <15%. For the universal CGM model, all BG value data in the MI / NI-BG training data set of all participants were distributed in areas A, B, and D. The effect of the adaptive non-invasive BG model is significantly better than that of the general non-invasive BG model.

[0101] Figure 10Clark error grid analysis of BG values for all participants in the post-test 3.5 days. Red dots represent the effect of the adaptive non-invasive BG model, with most of the effect occurring in area A and a small amount in area B. Blue dots represent the effect of the universal model, with most of the effect occurring in area B and a small amount in areas A and D.

[0102] In summary, the method provided in the embodiment of the present application includes: obtaining non-invasive relevant parameter samples, real blood glucose data samples, and continuous blood glucose data samples, training a universal time series model, and obtaining a trained universal time series model; the universal time series model is used to predict blood glucose based on historical and continuous features of blood glucose; obtaining individual blood glucose data samples, performing migration training on the universal time series model based on the individual blood glucose data, obtaining a personalized time series model, and performing individual blood glucose prediction based on the personalized time series model. The embodiment of the present application predicts individual blood glucose by obtaining a personalized time series model through migration training, which is conducive to improving the accuracy of blood glucose prediction.

[0103] Secondly, an embodiment of the present invention proposes a non-invasive blood glucose prediction system based on a personalized time series model, which specifically includes:

[0104] The first module is used to obtain non-invasive relevant parameter samples, real blood glucose data samples, and continuous blood glucose data samples, and train a universal time series model to obtain a trained universal time series model; the universal time series model is used to predict blood glucose based on historical and continuous features of blood glucose;

[0105] The second module is used to obtain individual blood glucose data samples, perform migration training on the universal time series model according to the individual blood glucose data to obtain a personalized time series model, and perform individual blood glucose prediction based on the personalized time series model.

[0106] It can be seen that the contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0107] On the other hand, an embodiment of the present invention provides a non-invasive blood glucose prediction device based on a personalized time series model, comprising:

[0108] at least one processor;

[0109] at least one memory for storing at least one program;

[0110] When the at least one program is executed by the at least one processor, the at least one processor implements the non-invasive blood glucose prediction method based on the personalized time series model.

[0111] Similarly, the contents of the above method embodiments are applicable to the present device embodiments. The functions specifically implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0112] An embodiment of the present invention further provides a computer-readable storage medium storing a program executable by a processor. When the program is executed by the processor, it is used to execute the above-mentioned non-invasive blood glucose prediction method based on the personalized time series model.

[0113] Similarly, the contents of the above method embodiments are applicable to the present storage medium embodiment. The functions specifically implemented by the present storage medium embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0114] In some optional embodiments, the function / operation mentioned in the block diagram may not occur in the order mentioned in the operation diagram. For example, depending on the function / operation involved, the two boxes shown in succession can actually be executed substantially simultaneously or the boxes can sometimes be executed in reverse order. In addition, the embodiment presented and described in the flow chart of the present invention is provided in an exemplary manner for the purpose of providing a more comprehensive understanding of the technology. The disclosed method is not limited to the operation and logic flow presented herein. Optional embodiments are contemplated in which the order of the various operations is changed and the sub-operations described as a part of a larger operation are performed independently.

[0115] In addition, although the present invention is described in the context of functional modules, it should be understood that, unless otherwise stated, one or more of the functions and / or features may be integrated into a single physical device and / or software module, or one or more functions and / or features may be implemented in separate physical devices or software modules. It is also understood that a detailed discussion of the actual implementation of each module is not necessary for understanding the present invention. More specifically, given the properties, functions, and internal relationships of the various functional modules in the devices disclosed herein, the actual implementation of the module will be understood within the ordinary skill of an engineer. Therefore, a person skilled in the art will be able to implement the present invention set forth in the claims using ordinary skill without undue experimentation. It is also understood that the specific concepts disclosed are merely illustrative and are not intended to limit the scope of the present invention, which is determined by the full scope of the appended claims and their equivalents.

[0116] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several programs for enabling a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0117] The logic and / or steps represented in a flowchart or otherwise described herein, for example, may be considered as an ordered list of executable programs for implementing the logical functions, and may be embodied in any computer-readable medium for use by, or in conjunction with, a program execution system, apparatus, or device (e.g., a computer-based system, a system including a processor, or other system that can retrieve and execute a program from a program execution system, apparatus, or device). For purposes of this specification, a "computer-readable medium" may be any device that can contain, store, communicate, propagate, or transport a program for use by, or in conjunction with, a program execution system, apparatus, or device.

[0118] More specific examples (a non-exhaustive list) of computer-readable media include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk cartridge (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), a fiber optic device, and a portable compact disc read-only memory (CDROM). In addition, the computer-readable medium may even be paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering, or processing in another suitable manner as necessary, and then stored in a computer memory.

[0119] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above-described embodiments, multiple steps or methods can be implemented using software or firmware stored in a memory and executed by a suitable program execution system. For example, if implemented using hardware, as in another embodiment, any one of the following technologies known in the art or a combination thereof can be used: a discrete logic circuit having a logic gate circuit for implementing a logic function on a data signal, an application-specific integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc.

[0120] In the above description of this specification, reference to the terms "one embodiment / example," "another embodiment / example," or "certain embodiments / examples" means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in any one or more embodiments or examples.

[0121] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to the embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the claims and their equivalents.

[0122] The above is a specific description of the preferred implementation of the present invention, but the present invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of the present invention.

Claims

1. A non-invasive blood glucose prediction method based on a personalized time series model, characterized in that: The following steps are involved: Obtaining non-invasive relevant parameter samples, real blood glucose data samples, and continuous blood glucose data samples, and training a universal time series model to obtain a trained universal time series model; the universal time series model is used to predict blood glucose based on historical features and continuous features of blood glucose; Obtain individual blood glucose data samples, perform migration training on the universal time series model according to the individual blood glucose data to obtain a personalized time series model, and perform individual blood glucose prediction based on the personalized time series model.

2. The non-invasive blood glucose prediction method based on the personalized time series model according to claim 1, characterized in that: The universal time series model includes a first model component, a second model component, and a third model component. The universal time series model is trained by the following steps: Inputting historical non-invasive related parameter samples into the first model component to obtain historical parameter feature samples; inputting the continuous blood glucose data sample into the second model component to obtain a blood glucose trend feature sample; The current non-invasive related parameter samples, the historical parameter feature samples and the blood glucose trend feature samples are input into the third model component to obtain predicted values, and the universal time series model is trained according to the obtained predicted values to obtain a trained universal time series model.

3. The non-invasive blood glucose prediction method based on the personalized time series model according to claim 2, characterized in that: The step of inputting historical non-invasive related parameter samples into the first model component to obtain historical parameter feature samples includes: Constructing the first model component based on a single-layer bidirectional GRU network; Performing sine-cosine embedding transformation on the historical non-invasive related parameter samples to obtain corresponding time period codes; Based on the time period code and the historical non-invasive related parameter samples, a historical parameter feature sample is obtained.

4. The non-invasive blood glucose prediction method based on the personalized time series model according to claim 2, characterized in that: The third model component includes a plurality of connection layers; the current non-invasive related parameter sample, the historical parameter feature sample, and the blood glucose trend feature sample are input into the third model component to obtain a predicted value, including: Based on the connection layer, the current non-invasive related parameter samples, the historical parameter feature samples and the blood glucose trend feature samples are weightedly fused to obtain a predicted value.

5. The non-invasive blood glucose prediction method based on personalized time series model according to claim 1, characterized in that: The method further comprises: Embedding the personalized time series model into a wearable device to collect individual blood glucose data and corresponding time data; An individual continuous blood glucose prediction result is obtained based on the individual blood glucose data and the time data.

6. The non-invasive blood glucose prediction method based on personalized time series model according to claim 2, characterized in that: Inputting the continuous blood glucose data sample into the second model component to obtain a blood glucose trend feature sample includes: Constructing the second model component based on a single-layer bidirectional GRU network; Obtaining discrete blood glucose data samples based on the continuous blood glucose data samples, and determining blood glucose fluctuation status in combination with meal times; Performing sine-cosine embedding transformation on the continuous blood glucose data samples to obtain corresponding time period codes; The second model component is trained based on the time period code, the historical non-invasive related parameters and the blood glucose fluctuation state to obtain historical parameter feature samples.

7. The non-invasive blood glucose prediction method based on personalized time series model according to claim 1, characterized in that: The obtaining of non-invasive relevant parameter samples includes: Obtain metabolic thermal parameter samples and cardiovascular parameter samples of the group.

8. A non-invasive blood glucose prediction system based on a personalized time series model, characterized in that: include: The first module is used to obtain non-invasive relevant parameter samples, real blood glucose data samples, and continuous blood glucose data samples, and train a universal time series model to obtain a trained universal time series model; the universal time series model is used to predict blood glucose based on historical and continuous features of blood glucose; The second module is used to obtain individual blood glucose data samples, perform migration training on the universal time series model according to the individual blood glucose data to obtain a personalized time series model, and perform individual blood glucose prediction based on the personalized time series model.

9. A non-invasive blood glucose prediction device based on a personalized time series model, characterized in that: include: at least one processor; at least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the non-invasive blood glucose prediction method based on the personalized time series model as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a program executable by a processor, characterized in that: The processor-executable program is used to implement the non-invasive blood glucose prediction method based on the personalized time series model as claimed in any one of claims 1 to 7 when executed by the processor.

Images