Dynamic noninvasive blood glucose monitoring system based on intelligent ring

Abstract

The invention discloses a dynamic noninvasive blood glucose monitoring system based on an intelligent ring, and relates to the technical field of dynamic blood glucose monitoring. A system architecture is composed of an intelligent hardware layer, a software and data layer, an AI model layer and an application layer, the intelligent hardware layer comprises an intelligent ring, a PPG sensor, a Bluetooth module and a low-power-consumption battery are installed on the intelligent ring, and a painless needle-shaped sensor and a signal transmitter are installed on CGM equipment; the software and data layer comprises a data acquisition and synchronization module and a data storage and management module; the AI model adopts a combined architecture of a private model and a third-party open source model; the application layer comprises a real-time blood glucose prediction and monitoring module, a historical data analysis and personalized suggestion module and a model optimization and continuous learning module. A motion sensor is added to the hardware end of the intelligent ring, the intelligent ring is compared with CGM equipment in the system for learning in the learning and training stages, a user only needs to wear the intelligent ring, and a sustainable dynamic noninvasive blood glucose monitoring system is formed.

Description

Technical Field

[0001] The present invention relates to the technical field of dynamic non-invasive blood sugar monitoring, and in particular to a dynamic non-invasive blood sugar monitoring system based on a smart ring. Background Art

[0002] The dynamic blood glucose monitoring system is a medical device used to continuously monitor changes in human blood glucose levels. In the prior art, CGM is a minimally invasive dynamic blood glucose meter, which requires minimally invasive sampling of the human body, and the PPG signal of the dynamic blood glucose monitoring device is easily affected by motion, ambient light interference, etc., and the relationship between PPG and blood glucose varies between different users. The learning model inside the CGM device responds slowly, and when the learning model is deployed on a mobile device, it cannot take into account both efficiency and accuracy. Therefore, in response to the above problems, a combined system that can dynamically and continuously monitor and completely non-invasively monitor the human body is proposed. Summary of the invention

[0003] The purpose of the present invention is to provide a dynamic non-invasive blood glucose monitoring system based on a smart ring in response to the defects and shortcomings of the prior art. The smart ring uses CGM device support during development and training. The user only needs to wear the smart ring, which is easy to use and can infer blood glucose in real time through PPG signals. The smart ring reduces dependence on CGM devices after comparing and learning with the data in the CGM device in the system during the learning and training stages, realizes non-invasive blood glucose monitoring, and uses a more convenient health management method to provide real-time reminders of abnormal blood glucose status, improve user experience, integrate PPG sensors, mobile phones, and cloud hardware, realize AI control and cloud services, and form a sustainable ecological blood glucose monitoring system.

[0004] To achieve the above-mentioned purpose, the present invention adopts the following technical scheme: it is composed of a system architecture consisting of an intelligent hardware layer, a software and data layer, an AI model layer, and an application layer. The intelligent hardware layer includes a smart ring, on which a PPG sensor, a Bluetooth module, and a low-power battery are installed, and a painless needle sensor and a signal transmitter are installed on the CGM device; the software and data layer includes a data acquisition and synchronization module and a data storage and management module. The data acquisition and synchronization module uses the smart ring and the CGM device to simultaneously collect the signal emitted by the PPG sensor. The cloud storage of the data storage and management module uses a secure database to store data, and the local cache temporarily stores unuploaded data through a mobile phone App; the AI ​​model adopts a combination architecture of a private model + a third-party open source model, and adopts an adaptive training method, wherein the private model is an architecture that combines CNN and LSTM to extract the timing characteristics of the PPG signal; the application layer includes a real-time blood glucose prediction and monitoring module, a historical data analysis and personalized recommendation module, and a model optimization and continuous learning module.

[0005] The working steps of the dynamic non-invasive blood glucose monitoring system based on smart rings include the following points: 1. The smart ring in the intelligent hardware layer collects PPG signals in real time and optimizes the collection and preliminary processing of PPG signals. A local processing unit is added to the device end. The Bluetooth module transmits data to the mobile phone or the cloud. The CGM device collects blood glucose data in real time, and the blood glucose data is synchronized to the mobile phone through the Bluetooth module, which is aligned with the time of the PPG signal. The blood glucose data collected by CGM in real time can be used to train and verify the AI ​​model; 2. The smart ring and CGM device collect PPG signals and blood glucose data at the same time, continuously sample and upload them in real time, and record dynamic data every 3 minutes. Dynamic blood glucose level, and the cloud-based secure database encrypts the collected data to provide user privacy protection. At the same time, the mobile app temporarily stores unuploaded data, which is suitable for network interruption scenarios; 3. The private model combines CNN and LSTM to extract the time series features of the PPG signal, maps the PPG signal to dynamic blood glucose data, and uses the open source pre-trained model to process the basic features and generate auxiliary representations; when the AI ​​model inputs the PPG signal in real time, the private model extracts the local and full features of the PPG signal, and the third-party open source model provides high-dimensional features or prior knowledge, and uses the fusion model to output the dynamic blood glucose value as a blood glucose prediction; 4. The training method of the AI ​​model is to collect synchronous PPG-CGM data of multiple users for at least 1,000 hours. At the beginning of the training, the preprocessed PPG signal is input, and the blood glucose value at the corresponding time point is used everywhere. The loss function of the mean square error and the Adam or RMSProp optimization algorithm are used to establish a learning model, and the model is tested using unused user data. Its measurement indicators are MSE, MAE and clinical related indicators. The time window feature fusion, data enhancement technology, alignment and interpolation strategies are used to solve the problem of sparse labels relative to PPG signals and CGM signals; Fifth, the smart ring transmits the PPG data to the mobile phone App or the cloud via Bluetooth. The system calls the learned AI model, generates and displays the blood glucose value in real time, and uses a dynamic curve to represent the real-time blood glucose changes in the user interface, and provides high or low blood glucose alarm reminders; Sixth, after the user's blood glucose data is stored, the system can provide blood glucose trend analysis based on the learned AI model to provide users with personalized health advice; Seventh, the AI ​​model layer can use the newly collected data to regularly retrain the private model to improve the prediction accuracy, and use federated learning to update the model while protecting the user's privacy, so that the system is in a state of continuous learning to meet the user's long-term use experience.

[0006] Furthermore, the sampling rate of the smart ring for collecting PPG signals in real time is 50HZ and 100HZ, and the CGM has one label every 3 minutes.

[0007] Furthermore, the security database may be MySQL, Redis, or Mongodb.

[0008] Furthermore, the data encryption and user privacy protection comply with GDPR or HIPAA standards.

[0009] Furthermore, the open source pre-trained model is Pytorch or a deep learning model based on health signals.

[0010] Furthermore, the prior knowledge provided by the third-party open source model is the relationship between light signals and metabolism.

[0011] Furthermore, the clinically relevant indicator is the Clarke error grid.

[0012] The working principle of the present invention is as follows: a combination of intelligent hardware and CGM equipment is adopted, deep learning model optimization and reasoning acceleration are added to the server, mobile device or embedded device, the PPG signal is synchronized, compared and integrated with the CGM data, and a dynamic blood glucose prediction method based on machine learning and deep learning model is adopted to realize data preprocessing, real-time data processing and cleaning technology; personalized blood glucose prediction training and fine-tuning are realized through personalized training and transfer learning, and the robustness of the model in complex environments is improved through data enhancement, and the accuracy evaluation and dynamic calibration of blood glucose prediction are improved according to various data, so as to realize personalized health management for different users, and the integrated architecture can provide real-time feedback of user blood glucose prediction data.

[0013] After adopting the above technical scheme, the beneficial effects of the present invention are as follows: a PPG sensor is added to the hardware end of the smart ring, and the smart ring compares and learns with the CGM device in the system during the learning and training stage, so as to realize the process of training and learning through the smart ring combined with the CGM device. The end user only needs to wear the smart ring, and the filtering technology is combined to enhance the denoising process and anomaly detection. When the user uses it for the first time, the transfer learning method is used to collect 3 days of synchronous data for fine-tuning to generate a personalized model, and the model is lightweight. The Pytorch matching model or ONNX optimization model is used to realize the combination of cloud and local reasoning to ensure low latency; blood sugar is inferred in real time through PPG signals, reducing dependence on CGM devices, realizing non-invasive blood sugar monitoring, and in a more convenient health management method, reminding of abnormal blood sugar status in real time, improving user experience, integrating PPG sensors, mobile phones, and cloud hardware, realizing AI control and cloud services, forming a sustainable dynamic non-invasive blood sugar monitoring system, which has very high commercial value and high inventiveness. DETAILED DESCRIPTION

[0014] The technical solution adopted in this specific implementation is:

[0015] Collecting user data:

[0016] a. Single user data scale, CGM sampling frequency: record a blood glucose value every 3 minutes (about 480 data points per day);

[0017] PPG sampling frequency: Assuming the sampling rate is 100Hz, there are 100 data points per second;

[0018] Single-day PPG data: 24 hours × 3600 seconds × 100 Hz = 8,640,000 data points;

[0019] Data alignment: After aligning the PPG and CGM data, approximately 288 blood glucose values ​​are generated in one day, and each blood glucose value corresponds to the PPG data within a time period (e.g., a 3-minute window).

[0020] b. Total data requirements. In order to train the model, the following scale of data is usually required:

[0021] Target scale: at least 10,000 hours of PPG and CGM paired data (a conservative estimate);

[0022] Total data points: 10,000 hours × 100 Hz ≈ 3.6 billion (3,600,000,000) PPG data points;

[0023] Number of blood glucose labels: 10,000 hours / 3 minutes ≈ 200,000 labels.

[0024] c. Estimation of the number of users. Ideally, the data contributed by a single user is:

[0025] 8 hours / day x 10 days = 80 hours of PPG data. Calculation of number of users

[0026] Total data required: 10,000 hours, single user data volume: 80 hours;

[0027] The number of users required = total data volume / single user data volume = 10,000 hours / 80 hours ≈ 125 users.

[0028] d. Considering the data diversity, in order to ensure the generalization ability of the model, the number of users may need to be increased. The specific reasons are as follows:

[0029] Individual differences, such as age, gender, weight, skin type (such as pigment concentration), and basal metabolic rate may affect the relationship between PPG and blood sugar;

[0030] Users cover a variety of populations, especially those with diabetes (type I, type II) and healthy individuals;

[0031] Data diversity: collected data needs to cover a variety of scenarios: different activity states (stationary, moving); whether medication is taken, medication history information; different temperature or humidity environments; increasing scenario diversity requires more users.

[0032] Data redundancy and cleaning. During the collection process, some data may need to be discarded due to equipment problems or noise. For safety reasons, the actual number of users required should be higher than the theoretical value.

[0033] e. Recommended number of users. Considering the above factors, the recommended number of users is:

[0034] Basic coverage requirements: 150-200 users;

[0035] Diversity requirements: 300-500 users, covering a variety of populations, statuses and environments.

[0036] f. Notes on data collection

[0037] Gradually expand:

[0038] You can start with a small scale (e.g. 10-100 users) for preliminary model training and verification. Then gradually expand the number of users based on model performance and problems found.

[0039] User distribution:

[0040] Healthy individuals: 50%-70%;

[0041] Diabetic patients: 30%-50%;

[0042] Covering different age groups (such as children, teenagers, middle-aged and elderly people);

[0043] Data collection cycle,The collection cycle can last for several months, gradually enriching the data set and optimizing model performance.

[0044] g. Initial training of 150-200 users, full coverage: 300-500 users;

[0045] Data collection can be carried out in stages, with the goal of validating the model in the early stage and gradually expanding in the later stage to improve the generalization ability of the model.

[0046] Data preparation:

[0047] a. Data acquisition, PPG signal sampling frequency: 100 Hz (100 data points per second); acquisition window length is 30 seconds to 1 minute;

[0048] For example

[0049] Each 30-second window corresponds to 30×100=300030\times 100=300030×100=3000 data points;

[0050] Each 1-minute window corresponds to 60×100=600060\times 100=600060×100=6000 data points;

[0051] Blood glucose data: CGM devices record blood glucose values ​​every 3 minutes, which is strictly time-aligned with the PPG signal.

[0052] b. Data preprocessing and synchronous processing: every 3 minutes, the corresponding PPG signal is intercepted as input and the blood glucose value is used as the label;

[0053] Noise reduction and filtering: Filter adjustment: Bandpass filter for 100Hz sampling frequency: frequency range 0.5Hz~3Hz, remove motion artifacts and environmental interference;

[0054] Data normalization: PPG signal standardization (zero mean, unit variance) to avoid the influence of numerical range on training;

[0055] Feature extraction (optional): time domain features: heart rate, waveform interval, etc.;

[0056] Frequency domain features: frequency components are extracted through fast Fourier transform (FFT).

[0057] c. Data slicing, the input of each sample is the PPG signal within 3 minutes: input dimension: 300 seconds × 100 Hz = 30,000 data points, corresponding to the CGM blood glucose value at the time point; to adapt to model training, the PPG signal is divided into multiple small windows (such as 30 seconds), and the number of samples is expanded through sliding window technology (overlapping 10 seconds).

[0058] Model design:

[0059] a. Input adjustment. As the sampling frequency decreases, the temporal resolution of the PPG signal decreases. The model design requires the following adjustments:

[0060] Input length reduction: PPG signal of each input sample is 30 seconds (3000 data points) or 60 seconds (6000 data points);

[0061] Downsampling: In PPG signals, many variation patterns exist in lower frequency bands (<3Hz). Downsampling (e.g., 100Hz to 50Hz) can further reduce computational complexity.

[0062] b. Model architecture,

[0063] 1. Model architecture: lightweight CNN-LSTM model

[0064] 2.CNN part:

[0065] The convolution kernel captures short-term patterns in PPG signals, such as waveform shape and local features.

[0066] It is recommended to use a small convolution kernel (such as a size of 5 to 10).

[0067] 3.LSTM part:

[0068] Process the time series characteristics of PPG signals and capture long-term dynamic changes.

[0069] When the input data length is reduced, the depth of the LSTM layer can be increased, improving the time-dependent modeling capability.

[0070] Adjusted model structure:

[0071]

[0072]

[0073] c. Loss Function and Optimization

[0074] Regression task: use mean square error (MSE);

[0075] Optimizer: Adam. The learning rate is recommended to start from 10-310^{-3}10-3.

[0076] d. Training strategy:

[0077] 1. Batch training, batch size: 16 to 32, adjusted according to the capacity of video memory. The PPG signal samples input each time are [batch_size, time_steps(3000), channels(1)].

[0078] 2. Learning rate scheduling: Use the learning rate scheduler to automatically reduce the learning rate when the model performance improves slowly. The learning rate is reduced by 50% every 5 rounds (when the validation set loss does not decrease).

[0079] 3. Model evaluation, performance indicators: mean absolute error (MAE): <10mg / dL<10\,\text{mg / dL}<10mg / dL.

[0080] Clarke Error Grid Analysis: The proportion of predicted blood sugar values ​​in zone A (safe range) should reach more than 90%.

[0081] e. Deployment and optimization:

[0082] 1. Model lightweight, use Pytorch matching model or ONNX for model quantization:

[0083] Weights are quantized to 8-bit to reduce storage and reasoning overhead;

[0084] When deployed to mobile phones or embedded devices, ensure that the inference time is within 100ms;

[0085] 2. Real-time inference

[0086] Process a complete PPG window every 3 minutes and output blood glucose values ​​in real time; combine it with a sliding window during deployment to reduce latency and improve the real-time performance of dynamic blood glucose prediction;

[0087] f. Continuous optimization:

[0088] Expand the size of the data set, especially for data in scenarios with low sampling frequency; use transfer learning to quickly adapt to new users based on existing models.

[0089] Code sample section:

[0090] a. Data processing and loading

[0091] PyTorch is used to manage PPG data and corresponding blood glucose labels using the Dataset and DataLoader modules.

[0092] Example Code: Defining a Dataset

[0093]

[0094]

[0095] b. Model design

[0096] In PyTorch, you can directly define a hybrid CNN-LSTM model. Example code: CNN-LSTM model

[0097]

[0098]

[0099]

[0100] c. Training and evaluation

[0101] The training loop uses standard PyTorch training logic.

[0102] Example code: Training loop

[0103]

[0104]

[0105] d. Model evaluation Evaluate the model performance on the validation set or test set.

[0106] Example Code: Evaluate

[0107]

[0108]

[0109] e. Reasoning and deployment

[0110] Inference: Switch the model to evaluation mode model.eval(), processing single or batch samples.

[0111] Model saving: Save the model as a file for deployment:

[0112] torch.save(model.state_dict(),'ppg_to_glucose_model.pth')

[0113] Deployment optimization:

[0114] Convert models to a format optimized for mobile or embedded devices using TorchScript or ONNX.

[0115] Sample code (exported as ONNX):

[0116] dummy_input = torch.randn(1,3000).to('cuda') #Assume a single input sample

[0117] torch.onnx.export(model,dummy_input,"ppg_to_glucose.onnx").

[0118] The fusion methods of PPG signal and CGM data in the present invention are mainly as follows:

[0119] Data layer-based fusion method

[0120] Direct splicing and fusion: PPG signals and CGM data are directly spliced ​​on the time axis to form a long sequence containing both data information. For example, in a certain period of time, a section of PPG signal data is recorded first, followed by the CGM data of the same period of time, and so on. This method is simple and direct, but it does not fully consider the inherent relationship and feature differences of the two data, and the fusion effect may be limited. In subsequent analysis, different segments of data need to be processed separately.

[0121] Data interpolation fusion: Since the sampling frequencies of PPG signals and CGM data may be different, data interpolation is used to adjust the two to the same sampling frequency and then fuse them. For example, if the sampling frequency of the PPG signal is high and the sampling frequency of the CGM data is low, the CGM data can be interpolated to align it with the PPG signal in time, and then the data at the corresponding time points can be fused. This can make the two data more matched in the time dimension, which is convenient for subsequent joint analysis, but the interpolation process may introduce certain errors.

[0122] Feature layer-based fusion method

[0123] Feature extraction and selection: Extract representative features from PPG signals and CGM data, such as heart rate, heart rate variability, blood oxygen saturation, etc. from PPG signals, and extract blood sugar average, blood sugar fluctuation range, number of hypoglycemia and hyperglycemia events, etc. from CGM data. Then, based on specific application requirements and data characteristics, select highly relevant and complementary features for fusion.

[0124] Principal component analysis fusion: Dimensionality reduction methods such as principal component analysis (PCA) are used to map the high-dimensional features of PPG signals and CGM data to low-dimensional space to obtain their respective principal components. Then, according to the contribution rate and correlation of the principal components, appropriate principal components are selected for combination and fusion, thereby reducing the data dimension while retaining the main information and feature relationship of the two data and removing noise and redundant information.

[0125] Model-layer based fusion method

[0126] Bayesian network fusion: Construct a Bayesian network model, take PPG signals and CGM data as input nodes, and take physiological states or disease diagnosis results related to blood sugar as output nodes. By learning a large amount of sample data, determine the conditional probability distribution between nodes in the network, and thus infer the probability of physiological states or disease risks related to blood sugar based on PPG signals and CGM data. This method can handle uncertainty and complex relationships between variables well, but model construction and parameter learning require a lot of data support and have high computational complexity.

[0127] Neural network fusion: Using neural network models, such as multi-layer perceptron (MLP), convolutional neural network (CNN), recurrent neural network (RNN) and its variants (such as LSTM, GRU), etc., PPG signals and CGM data are used as neurons in the input layer, and after feature transformation and learning in the hidden layer, the fused results are finally obtained in the output layer, such as predicted blood sugar values, blood sugar change trends, etc. Neural networks have strong nonlinear fitting and feature learning capabilities, and can automatically mine the potential relationship between the two data, but model training requires a long time and a large amount of training data, and is prone to overfitting problems.

[0128] Fusion method based on decision layer

[0129] Voting fusion: Use PPG signals and CGM data for independent analysis and decision making. For example, use PPG signals to determine whether there are cardiovascular abnormalities, and use CGM data to determine whether blood sugar is within the normal range. Then, the two decision results are fused by voting, such as majority voting, weighted voting, etc., to obtain the final comprehensive decision result. This method is simple and intuitive, but it requires high accuracy in the preliminary processing of the two data and independent decision-making.

[0130] Expert system fusion: Establish an expert system to convert the knowledge and experience of medical experts into rules and logic, and combine PPG signals and CGM data to make comprehensive judgments and decisions. For example, when the PPG signal shows abnormal heart rate and the CGM data shows large fluctuations in blood sugar, the expert system infers possible health problems based on preset rules and gives corresponding suggestions and diagnosis results. Expert system fusion can make full use of domain knowledge, but the formulation and maintenance of rules require a lot of investment from professionals, and it is difficult to handle complex and changing situations.

[0131] The steps of combining a convolutional neural network (CNN) with a long short-term memory network (LSTM) to process PPG signal time series data and output blood glucose prediction values ​​in the present invention are as follows: 1. Design a convolutional neural network (CNN) module to extract local features from PPG signal time series data.

[0132] 2. The feature sequence extracted by CNN is input into the Long Short-Term Memory Network (LSTM) module to capture the long-term dependencies in time series data.

[0133] 3. The output of LSTM is fed into the fully connected layer for further feature fusion and dimensionality reduction.

[0134] 4. Finally, an output layer is used to convert the fused features into blood glucose prediction values.

[0135] In order to effectively integrate the output of the private model with the output of the third-party open source model in the present invention, the following steps can be considered:

[0136] 1. Data preprocessing: Ensure that the input data formats of the two models are consistent. If the data sources are different, normalization or standardization may be required to ensure data consistency.

[0137] 2. Model output standardization: If the output ranges or units of two models are different, they need to be converted to a unified scale for easy fusion.

[0138] 3. Feature fusion strategy: The following strategies can be used for feature fusion:

[0139] -Early fusion: The output features of the two models are merged at the input layer or intermediate layer, and then sent to the subsequent network layer for processing. This requires that the output features of the two models have the same dimension or can be converted to the same dimension.

[0140] - Mid-term fusion: Feature fusion is performed at the intermediate layers of the model, such as between CNN and LSTM, or between multiple LSTM layers.

[0141] -Late fusion: The final outputs of the two models are weighted averaged or combined in other forms to obtain the final prediction result. This usually requires determining the weight of each model output, which can be determined by methods such as cross-validation.

[0142] 4. Model training and optimization: The fused model needs to be retrained to optimize the model parameters. The validation set can be used to adjust hyperparameters such as learning rate, weight decay, etc. to achieve the best fusion effect.

[0143] 5. Evaluate the fusion effect: Use the test set to evaluate the performance of the fused model, and compare the model indicators before and after fusion, such as accuracy, recall, etc., to verify the effectiveness of the fusion strategy.

[0144] Through the above steps, the private model and the third-party open source model output can be effectively integrated. It should be noted that the selection of integration strategy and the adjustment of the model need to be determined according to the specific application scenario and data characteristics.

[0145] The above description is only used to illustrate the technical solution of the present invention rather than to limit it. Other modifications or equivalent substitutions made to the technical solution of the present invention by ordinary technicians in this field should be included in the scope of the claims of the present invention as long as they do not depart from the spirit and scope of the technical solution of the present invention.

Claims

1. Dynamic non-invasive blood glucose monitoring system based on smart ring, characterized by: It consists of a system architecture consisting of an intelligent hardware layer, a software and data layer, an AI model layer, and an application layer. The intelligent hardware layer includes a smart ring, which is equipped with a PPG sensor, a Bluetooth module, and a low-power battery. The CGM device is equipped with a painless needle sensor and a signal transmitter. The software and data layer includes a data acquisition and synchronization module and a data storage and management module. The data acquisition and synchronization module uses the smart ring and the CGM device to simultaneously collect the signals emitted by the PPG sensor. The cloud storage of the data storage and management module uses a secure database to store data, and the local cache temporarily stores unuploaded data through a mobile phone app. The AI ​​model adopts a combination architecture of private models + third-party open source models, and adopts an adaptive training method. The private model is an architecture that combines CNN and LSTM to extract the timing features of PPG signals. The application layer includes a real-time blood glucose prediction and monitoring module, a historical data analysis and personalized recommendation module, and a model optimization and continuous learning module.

2. The working steps of the dynamic non-invasive blood glucose monitoring system based on the smart ring are characterized by: It includes the following steps:

1. The smart ring in the smart hardware layer collects PPG signals in real time and optimizes the collection and preliminary processing of PPG signals. A local processing unit is added to the device side. The Bluetooth module transmits data to the mobile phone or the cloud. The CGM device collects blood glucose data in real time, and the blood glucose data is synchronized to the mobile phone through the Bluetooth module, and is aligned with the time of the PPG signal. The blood glucose data collected by the CGM in real time can be used to train and verify the AI ​​model; 2. The smart ring and CGM device collect PPG signals and blood glucose data at the same time, continuously sample and upload in real time, record dynamic blood glucose levels every 3 minutes, and the cloud-based secure database encrypts the collected data to provide user privacy protection. At the same time, the mobile phone App temporarily stores unuploaded data, which is suitable for network interruption scenarios; 3. The private model combines CNN and LSTM to extract the time series features of the PPG signal, maps the PPG signal to the dynamic blood glucose data, and uses the open source pre-trained model to process the basic features and generate auxiliary representations; when the AI ​​model inputs the PPG signal in real time, the private model extracts the local and full features of the PPG signal, and the third-party open source model provides high-dimensional features or prior knowledge, and uses the fusion model to output the dynamic blood glucose value as a blood glucose prediction; 4. The training method of the AI ​​model is to collect at least 1,000 hours of synchronized PPG-CGM data from multiple users. At the beginning of the training, the preprocessed PPG signal is input, and the blood glucose value at the corresponding time point is used everywhere. The learning model is established using the mean square error loss function and the Adam or RMSProp optimization algorithm, and the model is tested using unused user data. The measurement indicators are MSE, MAE and clinical related indicators. The time window feature fusion, data enhancement technology, alignment and interpolation strategies are used to solve the problem of sparse labels relative to PPG signals and CGM signals; 5. The smart ring transmits PPG data to the mobile phone App or the cloud via Bluetooth. The system calls the learned AI model, generates and displays the blood glucose value in real time, and uses a dynamic curve to represent the real-time blood glucose changes in the user interface, and provides high or low blood glucose alarm reminders; 6. After the user's blood sugar data is stored, the system can provide blood sugar trend analysis based on the learned AI model and provide users with personalized health advice; 7. The AI ​​model layer can use newly collected data to regularly retrain private models to improve prediction accuracy, and use federated learning to update the model while protecting user privacy, so that the system is in a state of continuous learning and meets users' long-term usage experience.

3. The working steps of the dynamic non-invasive blood glucose monitoring system based on the smart ring according to claim 2 are characterized by: The sampling rate of the smart ring for real-time collection of PPG signals is 50HZ and 100HZ, and the CGM has one label every 3 minutes.

4. The working steps of the dynamic non-invasive blood glucose monitoring system based on the smart ring according to claim 2 are characterized by: The security database may be MySQL, Redis, Mongodb, etc.

5. The working steps of the dynamic non-invasive blood glucose monitoring system based on the smart ring according to claim 2 are characterized by: The data encryption and user privacy protection described herein comply with GDPR or HIPAA standards.

6. The working steps of the dynamic non-invasive blood glucose monitoring system based on the smart ring according to claim 2 are characterized by: The open source pre-trained model is Pytorch or a deep learning model based on health signals.

7. The working steps of the dynamic non-invasive blood glucose monitoring system based on the smart ring according to claim 2 are characterized by: The prior knowledge provided by the third-party open source model is the relationship between light signals and metabolism.

8. The working steps of the dynamic non-invasive blood glucose monitoring system based on the smart ring according to claim 2 are characterized by: The clinically relevant indicator is the Clarke error grid.

Images