Noninvasive blood glucose monitoring optimization method and system

Abstract

The invention discloses a noninvasive blood glucose monitoring optimization method and system. The method comprises the steps that CGM data and blood glucose data of an optical / electrochemical hybrid sensor of wearable equipment are collected; basic numerical value features, time coding features and blood glucose dynamic features are extracted through feature engineering; performing two-stage training by adopting an ODE-LSTM model, learning a real blood glucose rule in a pre-training stage, newly adding a projection layer in a calibration stage, and dynamically optimizing an error by utilizing a dual-threshold MSE loss function; and finally, outputting the calibrated blood glucose value, and generating a blood glucose dynamic map to support personalized health management. The system comprises a data acquisition module, a preprocessing module, a model training module, a visual analysis module and the like, and real-time monitoring, error analysis and clinical decision support are realized through large-screen interaction. The problems that existing noninvasive equipment is insufficient in precision and poor in timeliness are solved, and the span from consumption level monitoring to medical auxiliary diagnosis level is achieved.

Description

Technical Field

[0001] This invention relates to the field of medical and health care and artificial intelligence interdisciplinary technology, and in particular to a non-invasive blood glucose monitoring optimization method and system. Background Technology

[0002] Blood glucose monitoring is a crucial aspect of diabetes management, and the accuracy of the data directly impacts treatment planning and patient prognosis. Traditional invasive testing methods primarily rely on finger-prick blood sampling, which is an intrusive procedure. Furthermore, due to cost and inconvenience, the testing frequency is limited, making it difficult to achieve continuous monitoring around the clock and failing to fully reflect the patterns of blood glucose fluctuations.

[0003] In recent years, while non-invasive / minimally invasive blood glucose monitoring devices (such as smart bracelets and watches) can achieve continuous data acquisition, their accuracy is limited by factors such as the 10-30 minute lag in the equilibrium of biofluid and blood glucose concentration, individual physiological differences (skin condition, blood flow velocity), and environmental interference (temperature, humidity). This results in insufficient accuracy of the raw data, with root mean square error (RMSE) generally exceeding 1.5 mmol / L, making it difficult to meet the requirements for clinical auxiliary diagnosis. For example, the commonly used evaluation criteria for clinical continuous glucose monitoring (CGM) systems typically require a mean absolute relative error (MARD) of less than 10%. Current non-invasive devices still fall short of this requirement in terms of accuracy.

[0004] For example, some non-invasive monitoring technologies based on indirect biomarkers (such as glucose concentration in sweat and tears) have a time delay compared to the actual blood glucose value because there is an equilibrium process between the glucose concentration in body fluids and blood. While solutions based on indirect physiological parameters (such as the vascular elasticity model of Huawei Watch4) can achieve risk warning, they cannot directly capture glucose molecule signals and lack data traceability.

[0005] Existing calibration methods (such as regression models and Kalman filtering) mostly rely on single-modal data and fail to effectively integrate time dynamics and physiological patterns of blood glucose, resulting in poor prediction timeliness and lack of personalized recommendations. Summary of the Invention

[0006] The purpose of this section is to outline some aspects of embodiments of the present invention and to briefly describe some preferred embodiments. Simplifications or omissions may be made in this section, as well as in the abstract and title of this application, to avoid obscuring the purpose of these documents; however, such simplifications or omissions should not be construed as limiting the scope of the invention.

[0007] Therefore, to solve the above-mentioned technical problems, the present invention provides the following technical solution: an optimized method for non-invasive blood glucose monitoring, comprising the following steps: S1: Data acquisition steps: Collect subcutaneous CGM blood glucose data obtained from professional medical testing equipment and blood glucose data from an optical / electrochemical hybrid sensor obtained from wearable non-invasive blood glucose testing equipment; S2: Data preprocessing steps: Clean the collected CGM data and wearable device blood glucose data to remove outliers and missing values; S3: Feature engineering steps: Extract multi-dimensional blood glucose-related features from the preprocessed data; S4: Model training steps: Two-stage training is performed using the ODE-LSTM model, including a pre-training stage and a calibration training stage. The pre-training stage uses CGM data to train the model, and the calibration training stage uses wearable device data to calibrate the model. S5: Blood glucose prediction step: Input the blood glucose data collected in real time by the wearable device into the trained model to obtain the blood glucose prediction value; S6: Management Steps: Generate a glycemic dynamics profile based on predicted blood glucose values ​​and provide personalized health management recommendations.

[0008] As a preferred embodiment of the non-invasive blood glucose monitoring optimization method described in this invention, the CGM data collected by the professional medical testing equipment meets the accuracy standards for medical devices, and its collection frequency is not less than once every 10 minutes.

[0009] As a preferred embodiment of the non-invasive blood glucose monitoring optimization method described in this invention, the multi-dimensional blood glucose-related features include basic numerical features, time features, and glycemic dynamics features.

[0010] As a preferred embodiment of the non-invasive blood glucose monitoring optimization method described in this invention, the basic numerical characteristics are obtained by performing maximum and minimum normalization on the CGM value and the blood glucose value from the wearable device. The normalization formula is as follows: ; in, This is the original blood glucose level. , These are the minimum and maximum values ​​in the dataset, respectively. The inverse normalization formula is: .

[0011] As a preferred embodiment of the non-invasive blood glucose monitoring optimization method described in this invention, the time features include sine coding of hours, cosine coding of hours, sine coding of minutes, cosine coding of minutes, and Boolean features for meal time markers and sleep time markers, specifically: Hourly sine code: ; Hourly cosine encoding: ; Minute sine code: ; Minute cosine encoding: ; Meal markers ( Meal times are marked using 0 / 1 Boolean characteristics; Sleep duration markers ( ): Sleep periods are marked using 0 / 1 Boolean characteristics.

[0012] As a preferred embodiment of the non-invasive blood glucose monitoring optimization method of the present invention, the glycemic dynamics features include: The rate of change in blood glucose: calculated based on the first derivative, using the following formula: , ( Unit: hour); The acceleration of blood glucose changes is calculated based on the second derivative, using the following formula: ; in: This represents the rate of change from time point i to j, used to reflect the intensity of blood glucose changes during a specific time period; Indicates time interval (unit: hours); Blood glucose stability indicators include the coefficient of variation (CV), calculated using the following formula: CV = (Standard deviation / Mean) × 100%; And through the mapping function Map CV values ​​to interval; TIR (Time Percentage Within Target Area): Its calculation formula is as follows: .

[0013] As a preferred embodiment of the non-invasive blood glucose monitoring optimization method described in this invention, the pre-training phase includes: A pre-trained model is constructed by selecting 7-dimensional input features. The model is an ODE-LSTM network, which includes an input layer (7 nodes), a hidden layer (1-2 layers of LSTM units, 64 or 128 nodes per layer, with ReLU activation function), and an output layer (1-dimensional predicted value, with linear activation function). The MSE loss function was used for 50 training epochs, and an early stopping mechanism was set up to stop training when the loss did not decrease for 5 consecutive epochs. Using the Adam optimizer, the learning rate is 0.001.

[0014] As a preferred embodiment of the non-invasive blood glucose monitoring optimization method described in this invention, the calibration training phase includes: Load the pre-trained model weights and build a calibration model, wherein the calibration model adds a projection layer, which is a 32-node fully connected layer with the Sigmoid activation function. The training is performed using a dual-threshold MSE loss function, which dynamically increases the loss weight based on the different threshold intervals in which the absolute value of the prediction error falls. If the prediction error is ≤5%, the weight is 1.0; If 5% < prediction error ≤ 10%, weight = 1.5; If the prediction error is greater than 10%, the weight is 2.0; The training rounds are set to 500, and an early stop mechanism is set up to stop training when the loss does not decrease for 30 consecutive rounds. Using the Adam optimizer, the learning rate is 0.0001.

[0015] A non-invasive blood glucose monitoring optimization system, the system being used to implement the non-invasive blood glucose monitoring optimization method described in any one of the above claims, the system comprising: The data acquisition module is used to collect CGM blood glucose data and wearable device blood glucose data; The data preprocessing module is used to clean the data to remove outliers and missing values; The feature engineering module is used to extract multi-dimensional blood glucose-related features; The model training module is used to perform two-stage training of ODE-LSTM; The blood glucose prediction module is used to output predicted blood glucose values; The visualization and analysis module is used to display blood glucose data, trends, and error analysis results on a large visual screen. The health management module is used to generate personalized recommendations based on predicted blood glucose values ​​and glycemic dynamics profiles. The clinical support module is used to assist medical decision-making and support the recording and adjustment of treatment plans.

[0016] As a preferred embodiment of the non-invasive blood glucose monitoring optimization system described in this invention, the wearable non-invasive blood glucose testing device includes a smart bracelet, a smartwatch, a belt-type blood glucose meter, and smart glasses.

[0017] As a preferred embodiment of the non-invasive blood glucose monitoring optimization system of the present invention, the visualization analysis module includes the following sub-modules: Real-time monitoring area: Displays the current blood glucose level, blood glucose status label, and a trend micro-chart of the last 20 historical blood glucose levels; Historical Trend Zone: A line chart is used to compare the actual CGM value, the original wearable device value, and the model calibration value, and statistical indicators are displayed. Error Analysis Area: Displays MARD values, error distribution histograms, and Clark error grid plots; Clinical Support Area: Displays glycemic dynamics profiles, clinical recommendations, and historical treatment protocols.

[0018] As a preferred embodiment of the non-invasive blood glucose monitoring optimization system described in this invention, the visualization analysis module supports interactive functions, including patient ID search, trend chart zooming, data point details pop-up, error distribution details viewing, and chart export.

[0019] The beneficial effects of this invention are: 1. This invention employs a two-stage training strategy using ODE-LSTM. In the pre-training stage, a baseline for blood glucose variation patterns is established based on CGM data. In the calibration stage, a proprietary learning error mapping for wearable devices is introduced in the projection layer. Combined with a dual-threshold MSE loss function, the penalty for large errors is dynamically optimized, thereby stabilizing the mean absolute relative error (MARD) of blood glucose detection below 9.5%. This indicator has reached the accuracy threshold of professional clinical-grade dynamic blood glucose monitoring systems, achieving a leap from consumer-grade to clinical auxiliary diagnostic level.

[0020] 2. This invention introduces 6-dimensional time coding (sine / cosine periodic features) and glycemic dynamics features (rate of change, acceleration, CV stability index, and TIR target range percentage) into feature engineering, effectively solving the problem of abrupt time changes and quantifying the pattern of blood glucose fluctuations. By integrating the continuous system modeling of ordinary differential equations with the time-dependent capture capability of LSTM through the ODE-LSTM network, the prediction accuracy of physiological dynamics such as postprandial fluctuations and nocturnal stability is significantly improved, avoiding the trend bias caused by the neglect of time periodicity in traditional models.

[0021] 3. The system of this invention displays calibrated blood glucose values, error analysis (Clark grid, MARD trend) and glycemic dynamics graphs in real time on a large visual screen, supporting medical staff to dynamically adjust insulin dosage; the health management module automatically generates diet and exercise suggestions based on graph features, realizing closed-loop management from monitoring to intervention; compared with the existing technology that only provides raw data or simple warnings, this invention realizes data-driven personalized clinical decision support. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a flowchart of the process of the present invention.

[0023] Figure 2 This is a system architecture diagram of the present invention.

[0024] Figure 3 This is a Clarke error grid analysis result diagram of the original wristband data of this invention.

[0025] Figure 4 This is a graph showing the Clarke error grid analysis results of the calibrated data from this invention. Detailed Implementation

[0026] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0027] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.

[0028] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.

[0029] Example 1 Reference Figures 1-2 This is the first embodiment of the present invention, which provides an optimized method for non-invasive blood glucose monitoring, applicable to, for example... Figure 2 In the system environment shown, wearable non-invasive blood glucose monitoring devices (such as smart bracelets and smartwatches) collect users' blood glucose data through optical / electrochemical hybrid sensors and transmit it to a server or local processing terminal via a network. The server integrates data preprocessing, feature extraction, model training, and prediction modules, and finally feeds back the calibrated blood glucose values, trend analysis, and health suggestions to the user terminal or medical platform.

[0030] CGM (Continuous Glucose Monitoring): A technology that monitors interstitial fluid glucose concentration in real time using a subcutaneous sensor. It meets the ISO15197 certification standard and collects data every 10 minutes.

[0031] ODE-LSTM (Ordinary Differential Equation-Long Short-Term Memory): A hybrid neural network structure that combines the ability of ordinary differential equations to model continuous-time dynamic systems with the ability of LSTM to capture long-term dependencies in time-series data. It is suitable for the characteristic of blood glucose values ​​changing continuously over time.

[0032] MARD (Mean Absolute Relative Difference): The core indicator of blood glucose detection accuracy, calculated as the average absolute value of the relative error between the predicted value and the actual value at each point; the goal of this invention is to control MARD below 9.5%.

[0033] In one exemplary embodiment, such as Figure 1 As shown, an optimized method for non-invasive blood glucose monitoring is provided, including the following steps: Step S201: Data acquisition and preprocessing; Subcutaneous CGM blood glucose data is collected from professional medical testing equipment that meets ISO15197 certification (collection frequency: once every 10 minutes), and blood glucose data from an optical / electrochemical hybrid sensor obtained from a wearable non-invasive blood glucose testing device is collected simultaneously to ensure time synchronization; The collected data is preprocessed using... The principle is to identify and remove outliers; for example, a patient's CGM data sequence showed a value of 20 mmol / L (far exceeding the normal range of 3.9-10.0 mmol / L). Values ​​detected as outliers will be removed. Missing values ​​are filled using linear interpolation (for example, missing periods in a patient's CGM data are filled by linear interpolation of preceding and following values ​​to ensure data integrity).

[0034] Step S202: Feature engineering; Extract multi-dimensional features from the preprocessed data, including: 1. Basic numerical characteristics: Maximum and minimum normalization were performed on CGM values ​​and wearable device blood glucose values. The normalization formula is: ; in, This is the original blood glucose level. , These are the minimum and maximum values ​​in the dataset, respectively. Example: A patient's CGM value ranges from 3.2 to 15.6 mmol / L. The normalized result for the original value of 8.5 mmol / L is: ; The inverse normalization formula is: ; 2. Time characteristics: The timestamp is periodically encoded in 6 dimensions (sine / cosine encoding ensures time continuity): Hourly sine code: ; Hourly cosine encoding: ; Minute sine code: ; Minute cosine encoding: ; Example: The data collection time is 8:30 AM (hour=8, minute=30), and its encoding is: , ; , ; Meal markers ( (Breakfast 7:00-8:00, Lunch 12:00-13:00, Dinner 18:00-19:00 are marked as 1, otherwise 0;) Sleep duration markers ( ): 1 is marked for 22:00-6:00 the next day, otherwise 0; 3. Glucose dynamics characteristics: Rate of change (first derivative): ; in, The blood glucose value at time point t (unit: mmol / L). Indicates the previous time point Blood sugar levels; Indicates time interval (unit: hours); Example: When =5.2mmol / L, =5.5mmol / L, time interval When =5 minutes = 5 / 60 hours, the rate of change is: ; Change in acceleration (second derivative): based on three consecutive time points ( , , The rate of change of blood glucose levels is calculated to identify accelerating or decelerating trends in blood glucose changes; the formula is: ; in, : Indicates the rate of change from time point i to j, used to reflect the intensity of blood glucose changes during a specific time period; Indicates time interval (unit: hours); Example: When =5.2mmol / L, =5.5mmol / L, =5.7mmol / L, at a time interval of 5 minutes: arrive The rate of change is: mmol / L; ; arrive The rate of change is (calculated in the same way as above): ; Therefore, the acceleration is: A negative value indicates that the rate of blood sugar rise is slower.

[0035] Stability metrics: The coefficient of variation (CV) was used to assess the stability of blood glucose fluctuations, and the results were converted into standardized metrics using a mapping function. The formula for calculating the coefficient of variation (CV) is: CV = (standard deviation / mean) × 100%. The smaller the CV value, the more stable the blood glucose. The coefficient of variation (CV) eliminates the influence of dimensions and facilitates the comparison of stability between different individuals or devices. The data were evaluated using the TIR evaluation criteria, and the coefficient of variation was transformed from... Mapped to This interval transforms CVs into model-friendly normalized values, enhancing feature interpretability and facilitating model processing. Mapping function ; like ,but , The closer the value is to 1, the higher the stability. TIR (Time in Range): This is used to assess the proportion of time blood glucose is within the target range of 3.9-10.0 mmol / L (i.e., the ideal range). It is a key indicator in clinical management, and its calculation formula is as follows: ; The time unit is either minutes or hours, and a uniform conversion is required; the higher the TIR value, the better the blood sugar control.

[0036] The aforementioned glucocorticoid features, together with the basic numerical features and temporal features, constitute a multi-dimensional feature input, providing rich temporal dynamic information for the ODE-LSTM model and significantly improving calibration accuracy.

[0037] Step S203: ODE-LSTM two-stage training; A two-stage training strategy using an ODE-LSTM hybrid network is employed. Through the collaborative optimization of pre-training and calibration training, accurate mapping of wearable device blood glucose data to clinical-grade precision is achieved. This strategy effectively addresses the core issues of large initial model bias and low calibration accuracy in existing technologies. The specific steps are as follows: 1. First stage: Pre-training; This phase uses subcutaneous CGM blood glucose data collected by professional medical testing equipment (meeting ISO15197 accuracy standards, collected every 10 minutes) as the training data source. The core objective is to build a basic understanding of the model's real-world blood glucose variation patterns, as detailed below: Data processing (strictly following feature engineering specifications): The raw CGM blood glucose values ​​are subjected to maximum and minimum normalization, and glycemic dynamics features (including rate of change, acceleration of change, and stability indicators) are extracted simultaneously. Finally, 7 core input features are selected (the selection criteria are the correlation between the features and the actual blood glucose values ​​and their predictive contribution, including normalized CGM values, hourly / minute sine / cosine coding, mealtime marking, etc.). The pre-training dataset is divided into training set: validation set: test set in a ratio of 8:1:1 to ensure the scientific nature of model training and validation.

[0038] Model architecture design: The ODE-LSTM hybrid network structure is adopted, which combines the ability of ordinary differential equations (ODE) to model continuous dynamic systems with the ability of long short-term memory networks (LSTM) to capture time-series dependencies, and adapts to the characteristics of blood glucose values ​​changing continuously over time (such as postprandial rise and nighttime stabilization). The specific structure is as follows: the input layer receives 7-dimensional feature data, the hidden layer is configured with 1-2 layers of LSTM units (64 or 128 nodes per layer, ReLU activation function), and the output layer generates a 1-dimensional normalized blood glucose prediction value (linear activation function).

[0039] The training optimization strategy is as follows: Loss function: Mean squared error (MSE) is used, and the calculation formula is as follows: Its significant penalty for large errors aligns with the clinical need for blood glucose monitoring to "prioritize reducing serious biases"; Training control: Fixed training for 50 rounds, using the Adam optimizer (initial learning rate 0.001), and setting an early stopping mechanism (training is terminated if the MSE on the validation set does not decrease for 5 consecutive rounds) to effectively avoid overfitting.

[0040] Core output: Saves the optimal model weights at the time of early cessation triggering. These weights have learned the temporal patterns of real blood glucose (such as time periodicity, meal / sleep fluctuation patterns), providing benchmark parameters for the calibration phase and solving the problem of model cognitive bias caused by initial data deviation.

[0041] Phase Two: Calibration Training; This stage, based on a pre-trained model, incorporates blood glucose data from wearable devices (such as smart bracelets) and achieves clinical-grade accuracy through an error correction mechanism, as detailed below: Data processing specifications: The feature engineering process is completely consistent with the pre-training stage: the blood glucose values ​​of wearable devices are normalized using the MIN / MAX values ​​of the CGM data in the pre-training stage, and time features and glycemic dynamics features are extracted simultaneously to ensure the uniformity of input dimensions; the dataset is split into training set and validation set in the same proportion to form paired samples of "wearable device blood glucose value - corresponding CGM true value".

[0042] Model architecture optimization: An "inheritance-extension" strategy is adopted: pre-trained model weights are directly loaded (retaining parameters from the input layer to the hidden layer), and a 32-node fully connected projection layer (Sigmoid activation function) is added after the output layer to specifically learn the error characteristics of wearable devices; this projection layer maps the output of the pre-trained model to the calibrated blood glucose value, effectively offsetting error sources such as sensor drift and environmental interference.

[0043] The training optimization strategy is as follows: Loss function: A dual-threshold MSE loss function is used, with dynamic adjustment of error penalty weights. like If so, then the weight = 1.0; like If so, then the weight = 1.5; like If so, then the weight = 2.0; This design prioritizes guiding the model to reduce large errors (such as predicted values ​​deviating from the true value by ≥3 mmol / L), meeting the high accuracy requirements of medical auxiliary diagnosis.

[0044] Training control: Fixed training for 500 rounds (to adapt to complex error mapping requirements), Adam optimizer learning rate reduced to 0.0001 (to avoid destroying pre-trained weights), early stopping mechanism set (termination if the loss on the validation set does not decrease after 30 consecutive rounds), balancing model convergence and overfitting risks.

[0045] Core output: Saves the final calibration model containing inherited weights and projection layer fine-tuning weights. This model can receive real-time data from wearable devices and output accurate blood glucose prediction values. This stage solves the core problem of wearable device error correction and achieves accurate mapping of device data to clinical-grade CGM data.

[0046] Step S204: Blood glucose detection and analysis; Real-time blood glucose data collected by wearable devices is processed through feature engineering and then input into a calibration model to output accurate blood glucose predictions. Combined with the rate of change of blood glucose (first derivative), acceleration of change (second derivative), and stability indicators (coefficient of variation (CV) mapping value, TIR percentage), a "glucose dynamics map" (including daily variation curves, mealtime fluctuation amplitude, and sleep stability intervals) is generated to provide personalized suggestions for patients (example: if the patient's postprandial blood glucose rise rate is >3...). The system recommends "reducing high-carbohydrate intake and increasing post-meal exercise" and provides clinical decision support for healthcare professionals.

[0047] The above-mentioned non-invasive blood glucose monitoring optimization method, through multi-dimensional feature engineering and ODE-LSTM two-stage training, achieves accurate mapping of wearable device data to clinical-grade precision, solves the problems of low calibration accuracy and poor prediction timeliness in existing technologies, and provides a reliable tool for diabetes management. By employing a progressive learning mechanism of "pre-training-calibration training," and the collaborative design of multi-dimensional feature engineering and dynamic loss functions, the nonlinear mapping problem between wearable device signals and gold standard data has been solved. Key steps such as projection layer augmentation and dual threshold weighting mechanisms are core technologies for overcoming accuracy bottlenecks, ensuring the irreplaceability of the technical solution.

[0048] Example 2 Reference Figure 2 This is the second embodiment of the present invention, which differs from the first embodiment in that it provides a non-invasive blood glucose monitoring optimization system. This system, through a modular design combined with a visual large-screen interactive interface, realizes a complete workflow from data acquisition to clinical decision support. The following, in conjunction with the accompanying drawings, provides a complete description of the system's constituent modules, functional implementation, and the detailed structure of the visual large screen.

[0049] The non-invasive blood glucose monitoring optimization system includes the following core modules: Data acquisition module: Used to collect subcutaneous CGM blood glucose data from professional medical testing equipment (compliant with ISO15197 certification), and blood glucose data from optical / electrochemical hybrid sensors from wearable non-invasive blood glucose testing devices (such as smart bracelets, smartwatches, belt-type blood glucose meters, and smart glasses). CGM data is collected every 10 minutes to ensure synchronization with wearable device data.

[0050] Data preprocessing module: Connected to the data acquisition module, it cleans the acquired blood glucose data, identifies and removes outliers using the 3σ principle, and completes missing values ​​using linear interpolation to ensure data integrity and validity.

[0051] Feature engineering module: Connected to the data preprocessing module, it performs multi-dimensional feature extraction operations, including: Basic numerical characteristics: CGM values ​​and wearable device blood glucose values ​​were normalized to maximum and minimum values; Time characteristics: Perform 6-dimensional time encoding (hour sine / cosine, minute sine / cosine, meal time markers, sleep period markers); Glucose dynamics characteristics: extracting the rate of change of blood glucose, acceleration of change, coefficient of variation (CV), and time interval (TIR).

[0052] Model training module: Connected to the feature engineering module, it performs two-stage training of ODE-LSTM. Pre-training phase: The model is trained using CGM data, with 7-dimensional input features selected. The MSE loss function is used for 50 training rounds, and the early stopping mechanism is to verify that the loss does not decrease for 5 consecutive rounds. Calibration training phase: Load pre-trained weights, add a projection layer (32-node fully connected layer), train for 500 rounds using the dual-threshold MSE loss function, and use the early stopping mechanism to verify that the loss does not decrease for 30 consecutive rounds.

[0053] Blood glucose prediction module: It is connected to the data acquisition module and the model training module respectively, inputting real-time data from the wearable device into the calibration model and outputting the calibrated blood glucose prediction value.

[0054] Visualization and Analysis Module: Connects to the blood glucose prediction module and displays blood glucose data, trends, and error analysis results on a large visual screen.

[0055] Health Management Module: Connects to the blood glucose prediction module and generates personalized recommendations based on blood glucose prediction results and "glucose dynamics map".

[0056] Clinical Support Module: Connects with the Visualization Analysis Module and Health Management Module to provide decision support data for medical staff and support the recording and adjustment of treatment plans.

[0057] Example 3 Reference Figures 3-4This is the second embodiment of the present invention. The difference between this embodiment and the first embodiment is that this embodiment provides a visual blood glucose screen based on a non-invasive blood glucose monitoring optimization system for displaying and analyzing blood glucose data for medical institutions or individual users. The screen adopts a modular design and is connected to the optimization system described in Embodiment 2 through a data interface to receive calibrated blood glucose data and related analysis results in real time.

[0058] The visualized blood glucose screen includes the following four functional modules: Module 1: Patient Basic Information and Real-time Blood Glucose Monitoring Area; Functionality: Provides healthcare professionals with an instant overview of patients' blood glucose status and supports quick switching between single and multiple patients.

[0059] Content to be displayed: Real-time blood glucose card: Precisely displays the current time (minutes) and calibrated blood glucose value (font enlarged to 24pt, unit). Blood glucose status label (“Normal”: 3.9-10.0) Green; "High blood sugar": >10.0 Red; "Hypoglycemia": <3.9 (yellow), the label color coding strictly follows the ISO15197 medical standard; Real-time trend mini-chart: The right side displays a mini-line chart of nearly 20 historical blood glucose values ​​(X-axis is time, Y-axis is blood glucose value), with the current time point marked in a prominent mark to help quickly identify the trend of blood glucose changes (rising / falling / stable).

[0060] Interactive features: Supports switching patient data via the "Patient ID search box"; clicking on the trend chart will automatically zoom in to the complete historical trend analysis area of ​​Module 2.

[0061] Module 2: Historical Blood Glucose Trend Analysis Area; Functional positioning: Quantitatively compare the differences between the original data and the calibrated data of wearable devices, and intuitively demonstrate the calibration effect of the ODE-LSTM model; Content to be displayed: Multi-curve comparison chart: Using ECharts line chart (X-axis is time, Y-axis is blood glucose value, range 3.0-21.0 mmol / L), three curves are displayed simultaneously: “CGM True Value” (solid green line, baseline data); "Raw values ​​for wearable devices" (solid red line, uncalibrated data); “Model calibration value” (blue solid line, output result of this system); Statistical indicator card: The bottom displays the core indicators for the specified time range, including average blood glucose level, highest / lowest blood glucose level, and TIR (time interval of the target range). Interactive features: Supports dragging the timeline to zoom and view specific time periods; clicking on any data point on the curve will bring up a detailed information box (including time, CGM value, original value, and calibration value).

[0062] Module 3: Error Analysis and Accuracy Evaluation Area; Functional positioning: To verify the calibration accuracy of the ODE-LSTM model through quantitative indicators, and to provide a reliable basis for clinical decision-making.

[0063] Content to be displayed: MARD value real-time monitoring: The left-hand number card displays the "MARD value in the last 24 hours" (retaining 1 decimal place, font enlarged to 20pt), and the right-hand side displays a line graph of the MARD value change in the last 7 days (X-axis is date, Y-axis is MARD value, marked with the 9.5% target threshold line). When MARD > 9.5%, the card automatically turns red as a warning. Error distribution bar chart: The X-axis represents the error range (0-2%, 2-4%, 4-6%, 6-8%, 8-10%, >10%), and the Y-axis represents the corresponding sample size. It is filled with a gradient of blue (0-2% dark blue → >10% light blue) to visually reflect the concentration of the error distribution. Clark error grid plot: drawn according to medical industry standards (X-axis is the true CGM value, Y-axis is the calibrated value), divided into five regions: A, B, C, D, and E. Zone A (error < 10%): Green, accounting for 92.5%; Zone B (10-20%): Yellow, accounting for 7.5%; Zones C / D / E (>20%): Red, accounting for 0%; It also displays the sample percentage for each region (e.g., "Area A 92.5%, Area B 7.5%").

[0064] Interactive features: Clicking on the error distribution bar chart allows you to view detailed error sample information; supports exporting Clark grid charts as PDF / image formats; The results of Clarke error grid analysis on the original wristband data are as follows: Figure 3 As shown, the Clarke error grid analysis results of the calibrated data are as follows: Figure 4 .

[0065] Module 4: Clinical Support and Health Advice Section; Functional positioning: Generate clinical decision support based on glycemic dynamics characteristics to achieve closed-loop management of treatment; Content to be displayed: Glucose dynamics graph: The left side shows a heatmap of "rate of change of blood glucose over time" over the past 7 days (X-axis is time, Y-axis is rate of change, units...). (Color depth indicates the magnitude of the rate of increase), marking the peak period of postprandial blood glucose rise (e.g., "Rise rate 3.2 mmol / L 1 hour after breakfast"). (”); Clinical recommendation cards: Automatically generate recommendations based on glycemic dynamics characteristics (e.g., "Postprandial 2-hour blood glucose rate has consistently been >3.0 mmol / L for the past 3 days"). It is recommended to adjust the insulin dose and increase postprandial exercise by 15 minutes. The logic of the recommendation is strictly based on the characteristic thresholds of rate of change, stability index (CV>15%), and TIR<70%. Historical treatment regimen comparison: The right side shows the treatment regimen adjustment record for the past 3 months (such as "2024-03-01: insulin dose increased from 10U to 12U") and the blood glucose trend curve after the adjustment, supporting efficacy assessment; Interactive functions: Healthcare staff can manually edit clinical recommendations and automatically save them to the Hospital Information System (HIS); clicking on the peak area of ​​the heat map will bring up detailed characteristic data for that period (including acceleration of change, CV value, and TIR).

[0066] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. An optimized method for non-invasive blood glucose monitoring, characterized in that: Includes the following steps: S1: Collect subcutaneous CGM blood glucose data obtained from professional medical testing equipment and blood glucose data from optical / electrochemical hybrid sensors obtained from wearable non-invasive blood glucose testing devices; S2: Clean the collected CGM data and wearable device blood glucose data to remove outliers and missing values; S3: Extract multi-dimensional glucose-related features from the preprocessed data; S4: The ODE-LSTM model is used for two-stage training, including a pre-training stage and a calibration training stage. The pre-training stage uses CGM data to train the model, and the calibration training stage uses wearable device data to calibrate the model. S5: Input the blood glucose data collected in real time by the wearable device into the trained model to obtain the blood glucose prediction value; S6: Generates glycemic dynamics profiles based on predicted blood glucose values ​​and provides personalized health management recommendations.

2. The optimized method for non-invasive blood glucose monitoring as described in claim 1, characterized in that: The CGM data collected by the professional medical testing equipment meets the accuracy standards for medical devices, and its collection frequency is no less than once every 10 minutes.

3. The optimized method for non-invasive blood glucose monitoring as described in claim 2, characterized in that: The multidimensional glucose-related features include baseline numerical features, temporal features, and glycemic dynamics features.

4. The optimized method for non-invasive blood glucose monitoring as described in claim 3, characterized in that: The basic numerical features are obtained by performing maximum and minimum normalization on the CGM value and the blood glucose value from the wearable device.

5. The optimized method for non-invasive blood glucose monitoring as described in claim 4, characterized in that: The time features include sine coding for hours, cosine coding for hours, sine coding for minutes, cosine coding for minutes, and Boolean features for mealtime markers and sleep time markers.

6. The optimized method for non-invasive blood glucose monitoring as described in claim 5, characterized in that: The glucocortic dynamics features include the rate of change of blood glucose calculated based on the first derivative, the acceleration of change of blood glucose calculated based on the second derivative, and the glucocortic stability index based on the coefficient of variation and / or time within the target range.

7. The optimized method for non-invasive blood glucose monitoring as described in claim 6, characterized in that: The pre-training phase includes selecting a subset of core features from the multi-dimensional blood glucose-related features to construct a pre-trained model, using the MSE loss function for training, setting the training rounds to 50, and setting an early stop mechanism to stop training when the loss does not decrease for 5 consecutive rounds.

8. The non-invasive blood glucose monitoring optimization method as described in claim 7, characterized in that: The calibration training phase includes: Load the pre-trained model weights and build a calibration model, wherein the calibration model adds a projection layer, which is a 32-node fully connected layer with the Sigmoid activation function. The training is performed using a dual-threshold MSE loss function, which dynamically increases the loss weight based on the different threshold intervals in which the absolute value of the prediction error falls. The training rounds are set to 500, and an early stop mechanism is set up to stop training when the loss does not decrease for 30 consecutive rounds. The dual-threshold MSE loss function dynamically adjusts the loss weights based on the relationship between the prediction error and the first and second thresholds.

9. A non-invasive blood glucose monitoring optimization system, the system being used to implement the non-invasive blood glucose monitoring optimization method according to any one of claims 1-8, characterized in that: The system includes: The data acquisition module is used to collect CGM blood glucose data and wearable device blood glucose data; The data preprocessing module is used to clean the data to remove outliers and missing values; The feature engineering module is used to extract multi-dimensional blood glucose-related features; The model training module is used to perform two-stage training of ODE-LSTM; The blood glucose prediction module is used to output predicted blood glucose values; The visualization and analysis module is used to display blood glucose data, trends, and error analysis results on a large visual screen. The health management module is used to generate personalized recommendations based on predicted blood glucose values ​​and glycemic dynamics profiles. The clinical support module is used to assist medical decision-making and support the recording and adjustment of treatment plans.

10. The non-invasive blood glucose monitoring optimization system as described in claim 9, characterized in that: The wearable non-invasive blood glucose testing devices include smart bracelets, smartwatches, belt-type blood glucose meters, and smart glasses.

Images