T1DM children blood glucose monitoring and early warning system

Abstract

The invention relates to the technical field of child health monitoring and artificial intelligence, and provides a T1DM child blood glucose monitoring and early warning system. An existing blood glucose monitoring system is generally only based on single blood glucose data acquisition, and blood glucose fluctuation of children in different situations is difficult to accurately capture in real time. Through a multi-modal data fusion technology, multi-modal physiological features are integrated, and deep feature extraction and fusion analysis of data are realized. The system further combines a Q-learning algorithm to generate a personalized blood glucose early warning strategy, and utilizes a hidden Markov model (HMM) to identify the behavior situation of the child and optimize the early warning threshold setting. Meanwhile, data privacy is protected in a data distributed processing process through a federated learning framework and a differential privacy protection technology. The real-time performance and accuracy of blood glucose monitoring are improved, and intelligent and personalized health management and early warning services are provided for children and guardians of the children on the premise of protecting the privacy of the children.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent health monitoring systems, and in particular to a blood sugar monitoring and early warning system for children with T1DM (type 1 diabetes). Background Art

[0002] With the rapid development of medical technology and information technology, blood glucose monitoring and health management for children with type 1 diabetes (T1DM) have become increasingly important topics. T1DM is a chronic disease that requires long-term insulin injections and close monitoring of blood glucose levels. However, traditional blood glucose monitoring methods mainly rely on fingertip blood sampling for instant testing, which has limitations such as low testing frequency and discomfort during the blood sampling process. Especially for pediatric patients, their compliance and tolerance are significantly insufficient. This traditional method makes it difficult to achieve continuous dynamic monitoring of blood glucose, and thus cannot promptly identify and warn of the risk of abnormal blood glucose. In addition, the physiological and behavioral characteristics of children determine that their blood glucose fluctuates frequently, requiring a more comprehensive and intelligent monitoring solution.

[0003] In recent years, the development of smart wearable devices and multimodal sensing technology has provided new possibilities for blood glucose management in children with T1DM. Multimodal data acquisition, fusion processing, behavioral analysis and other technologies can monitor a variety of physiological data including blood glucose, heart rate, body temperature and activity status in real time, providing a technical basis for continuous blood glucose monitoring. However, the existing system still has shortcomings in multimodal data fusion, personalized analysis and dynamic adjustment of warning thresholds. On the one hand, it is difficult for traditional algorithms to take into account the feature differences between data when facing different modal data, and it is difficult to build a stable and effective fusion model. On the other hand, the existing system lacks the ability to adapt to individual characteristics and behavioral situations, and cannot automatically optimize the warning strategy according to the health status of different individuals, which affects the monitoring accuracy and the accuracy of the warning. Blood glucose monitoring in children with T1DM also faces the problem of data privacy and security. Children's health data involves highly sensitive information, so a strict data privacy protection mechanism is required during data transmission and model training. The current blood glucose monitoring system lacks sufficient privacy protection measures, and it is difficult to effectively realize remote monitoring and analysis while ensuring data security.

[0004] In summary, there is an urgent need for a T1DM children's blood glucose monitoring and early warning system that integrates multimodal data collection, personalized analysis, privacy protection and remote monitoring functions, which can provide real-time and accurate blood glucose monitoring and abnormal early warning, thereby ensuring the health and safety of children. Summary of the invention

[0005] The purpose of the present invention is to provide a T1DM children's blood glucose monitoring and early warning system, which aims to use multimodal data collection and intelligent analysis to achieve real-time monitoring of children's blood glucose dynamic changes and abnormal early warning, generate personalized health intervention strategies, and solve the problems of discontinuous monitoring, insufficient personalized analysis, and weak data privacy protection in the prior art. To achieve the above purpose, the T1DM children's blood glucose monitoring and early warning system provided by the present invention includes:

[0006] Data acquisition module: The core of this module is to accurately and frequently obtain various physiological data of children (blood sugar concentration, heart rate, body temperature and activity status). The acquisition equipment uses portable wearable technology to realize real-time data recording. The system uses short-time Fourier transform (STFT) to perform frequency domain analysis on dynamic physiological data. The STFT window length is set to 25ms, and the 50% overlap rate design ensures a smooth transition effect, which is suitable for physiological data collection at different time points. At the same time, the frequency domain processing within the module can capture subtle physiological fluctuations and then extract short-term change characteristics, such as sudden changes in children's heart rate after intense activities or blood sugar fluctuations due to postprandial insulin response. For different data types, the data acquisition module uses specific sampling frequency and resolution optimization. Blood sugar data is recorded every 5 minutes, and heart rate data is sampled once a minute to ensure that the recording of different physiological parameters meets the real-time and accuracy requirements.

[0007] Data preprocessing module: responsible for cleaning, smoothing and standardizing the original collected data. This module uses a Kalman filter to smooth dynamic data. During the filtering process, the prediction step of the Kalman filter calculates the optimal estimate of the subsequent data based on the state transition matrix (A), and uses the covariance matrix (Q, R) to adjust the noise suppression effect to ensure that abnormal spikes do not interfere with the overall trend of the data stream. The Kalman filter uses dynamic adjustment of the prediction variance and measurement variance to adapt to the changes in different data modes. In addition, the module applies the Z-score algorithm, sets a threshold range of ±3σ, and removes outliers. After preprocessing, the normalization process uses linear normalization to standardize the data of different modes to the [0,1] interval to improve the adaptability of the model and the uniformity of calculation. Using this multi-layer data preprocessing method, the system eliminates interference factors in the data, making the subsequent algorithm processing more accurate and reliable.

[0008] Multimodal data fusion module: The core purpose of the multimodal data fusion module is to use the associated information of different physiological characteristics to improve the accurate judgment of children's health status. This module applies convolutional neural network (CNN) to extract features from blood glucose data. In order to enhance the sensitivity to blood glucose fluctuations, the number of convolutional layers is set to 5 layers and the convolution kernel size is set to 3×3. This structure has high effectiveness in extracting local features. For long-term dependent signals such as body temperature and activity status, the module introduces a long short-term memory network (LSTM). The recursive structure of the LSTM network enables the system to capture the change of heart rate over time and the continuous fluctuation of body temperature. The multimodal data fusion module uses the Bayesian optimization algorithm to dynamically adjust the modal weights. Under the Gaussian process model, the expected improvement function (EI) is used to determine the data contribution weights of different modalities, so that the system has good adaptability to different environments and individual differences of children. This optimized fusion ensures that accurate monitoring effects can still be achieved in specific scenarios (such as exercise and high temperature environments).

[0009] Adaptive learning module: This module is based on the Q-learning algorithm and uses reinforcement learning to adaptively adjust the differences in individual data. The learning rate α is set to 0.05, the discount factor γ is 0.85, and the system uses real-time feedback to continuously correct the Q value to optimize the individualized warning threshold. After collecting individual characteristic data, the reinforcement learning mechanism of Q-learning uses a reward and punishment feedback mechanism to dynamically adjust the warning value and algorithm parameters. During the training process, the system uses long-term data of individual children to iterate and optimize the Q value multiple times, so that the model can form personalized threshold settings and warning trigger conditions based on the physiological response characteristics of specific children.

[0010] Context analysis module: The main technology of this module is based on the hidden Markov model (HMM), which uses hidden state modeling and probability calculation of observation sequences to identify children's daily behavioral contexts. The context analysis module sets 5 hidden states (such as motion state, resting state, sleeping state, etc.), constructs the parameter space of HMM according to the actual changes in physiological data, and uses the Baum-Welch algorithm to achieve self-learning and updating of the model. The system uses forward and backward algorithms to calculate the probability of the observation sequence and predict the context of the next time step, thereby identifying physiological change patterns in specific scenarios, such as the dynamic decrease in blood sugar during exercise or the trend of blood sugar stabilization in the resting state. In different contexts, the system can adaptively adjust the warning threshold to make the warning strategy more context-adaptive and personalized.

[0011] Privacy protection module: The privacy protection module uses differential privacy mechanism and federated learning framework to achieve privacy protection for distributed learning. When the model weights are aggregated, the privacy budget ε is set to 0.5, and a noise introduction mechanism is used to ensure that children's personal health data cannot be leaked. At the same time, during data transmission, the system uses homomorphic encryption technology to encrypt gradient data. Before each transmission process, the model uses encrypt(gradient) to encrypt the gradient to ensure that the model weights are protected during transmission. The system uses differential privacy mechanism to evaluate the consumption of privacy budget after each data processing, and uses noise regulation to control the use of privacy budget to ensure data security.

[0012] Early warning module: This module generates personalized health early warning signals and determines the early warning level based on Q-learning strategies and situational analysis results. The early warning levels include normal, caution and danger. The system determines whether to trigger an early warning based on the changing trends of blood sugar, heart rate and body temperature; if a high-risk blood sugar value is detected, the system uses a remote terminal to send an emergency alarm and notify the guardian in real time.

[0013] The data acquisition module further includes an adaptive sampling module for dynamically managing the data acquisition frequency, which can adjust the data acquisition frequency according to different activity states and blood sugar fluctuations. The adaptive sampling module monitors the activity data collected by the acceleration sensor in real time. When a sudden change in activity state is detected, the heart rate and blood sugar data collection frequency is increased to once a minute to capture rapidly changing physiological parameters. The module uses a moving average filter with a window size of 10s to smooth the sampled data stream to reduce noise accumulation caused by oversampling.

[0014] The data preprocessing module performs multi-stage data cleaning on physiological data such as blood sugar, heart rate and body temperature. Specifically, it includes: noise suppression processing, using the Kalman filter algorithm, dynamically updating the state transfer matrix A based on historical data, and adaptively adjusting the covariance matrix R; data repair operation, filling in the missing data of intermittent collection based on the Lagrange interpolation method; normalization stage, ensuring the consistency of processing of different modal data to improve the accuracy of subsequent data fusion.

[0015] The multimodal data fusion module adopts a weighted fusion algorithm and uses Bayesian optimization to iteratively adjust the fusion weights to maximize the accuracy of multimodal data fusion. The module further includes an error control submodule, which calculates the mean square error (MSE) for each eigenvalue of heart rate, blood sugar and body temperature, and sets an error threshold to automatically exclude abnormal eigenvalues. The fusion process sets a dynamic weighting factor W for the output features of the convolutional neural network (CNN) and the long short-term memory network (LSTM) to achieve the accuracy of feature fusion.

[0016] The adaptive learning module uses the Q-learning algorithm to adjust the health warning threshold based on the child's individual physiological data. The module further includes a reward feedback mechanism. In each training cycle, immediate rewards are set for hypoglycemia and hyperglycemia scenarios, and the definition is to prioritize accurate warnings of blood sugar critical points. The update formula of Q-learning is to ensure the optimal convergence of Q value under different health conditions.

[0017] The hidden Markov model (HMM) of the scenario analysis module uses the Baum-Welch algorithm to adaptively adjust the model parameters, and learns the transition probabilities of different hidden states based on the blood sugar and heart rate data in the context of children's daily activities. This module includes a prediction algorithm module, which uses forward and backward algorithms to calculate the probability of observation sequences in different contexts. The system forms predictions for changes in activity contexts and changes in physiological data trends based on probability distribution, and optimizes personalized health management plans.

[0018] The privacy protection module uses differential privacy mechanism and federated learning framework to achieve distributed learning with privacy protection among devices. Specifically, it includes: privacy budget control submodule, where the value of privacy budget $\epsilon$ is 0.5, which is used as noise intensity control for each data transmission to ensure user privacy protection; encrypted transmission module, which uses homomorphic encryption algorithm encrypt(gradient) to process gradient data during transmission, and decrypts it when the aggregation model is updated to complete the update of the privacy-protected federated learning model.

[0019] The early warning module generates dynamic health warning signals based on real-time data. The module has a risk level determination algorithm that triggers three warning levels based on dynamic data flow to calculate the current risk value, and combines situational analysis and Q-learning feedback to determine whether a high-risk warning is triggered and send a notification to parents, prompting them to take appropriate intervention measures. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] In order to more clearly illustrate the technical solutions in the embodiments or exemplary embodiments of the present invention, the drawings required for use in the embodiments or exemplary descriptions will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained according to the drawings without paying creative work.

[0021] Figure 1 Flowchart of blood glucose monitoring and early warning system for children with T1DM

[0022] Figure 2 Class diagram for T1DM children's blood sugar monitoring and early warning system

[0023] Figure 3Data flow diagram of blood glucose monitoring and early warning system for T1DM children

[0024] The realization of the purpose, functional features and advantages of the present invention will be further explained in conjunction with embodiments and with reference to the accompanying drawings. DETAILED DESCRIPTION

[0025] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0026] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present application. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present application belongs.

[0027] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present application. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, it indicates the presence of features, steps, operations, devices, components and / or combinations thereof.

[0028] Figure 1 This is a flowchart provided for an embodiment of the present invention, and the specific implementation process includes the following steps:

[0029] System initialization and data input preparation; acquisition of real-time physiological data and data processing; the system uses the following steps to achieve multi-modal physiological data acquisition and processing:

[0030] Blood glucose data is processed using short-time Fourier transform (STFT), where the window length is set to 25ms and the overlap rate is 50%. STFT converts blood glucose signals from the time domain to the frequency domain in order to identify characteristic changes in short-term fluctuations. Next, the system uses wavelet analysis to generate spectral feature vectors as input for physiological features, and further combines Kalman filtering for smoothing to remove noise and abnormal fluctuations, thereby improving the accuracy of blood glucose data.

[0031] Behavior data is collected using an acceleration sensor and combined with a hidden Markov model (HMM) for behavior recognition. The system first uses the collected acceleration signal to calculate the activity intensity and uses a time series analysis method to identify different motion states. The HMM uses forward and backward algorithms to calculate the observation probability under different states and determine the child's activity state, including exercise, rest, or sleep.

[0032] The temperature data is processed by Fourier transform enhanced time series analysis to capture the short-term fluctuation characteristics of body temperature. After wavelet denoising and trend smoothing, the temperature data is used to identify the possibility of fever or hypothermia using time series analysis and anomaly detection methods, thereby assisting in judging the health status of children.

[0033] In the process of integrating multimodal physiological data, different modalities (such as blood sugar, heart rate, body temperature, and behavioral data) have different effects on health risk assessment. To achieve the optimal integration of multimodal features, the system dynamically adjusts the weight of each modal feature through the Bayesian optimization method to maximize the accuracy of the final health assessment. This optimization problem constructs a multi-objective function and regards weight adjustment as an optimization task to minimize errors.

[0034] The optimization of weight distribution is completed through Bayesian optimization, in which the Bayesian optimization algorithm estimates the evaluation error based on the Gaussian process (GP) model and gradually selects the optimal weight combination to improve the recognition accuracy of the system.

[0035] The optimization of weight distribution is completed through Bayesian optimization. The Bayesian optimization algorithm approximates the objective function by constructing a proxy model. Assuming that the classification loss function L(α) is complex and cannot be directly optimized, Bayesian optimization approximates the objective function through the Gaussian process (GP) model G(α). By selecting the optimal next sampling point, the expected improvement function is used to determine the next most likely point to be improved. The principle is:

[0036] EI(α)=E[max(L(α*)-L(α),0)]

[0037] In the code, the optimization process repeatedly calls the sampler in the Bayesian optimization library, updates the Gaussian process proxy model each time, and dynamically adjusts the weights based on the current results.

[0038] The weight of each modality is adjusted through the loss function feedback during the training process. After each sentiment analysis, the system calculates the actual loss based on the prediction results, and then updates the weight through gradient descent or other optimization algorithms. The implementation steps are as follows:

[0039] The optimizer repeatedly calculates gradients and updates weights until convergence or a set number of iterations is reached.

[0040] The modal features are passed to the optimizer through the feature extractor, and the loss value of each modality is returned to the main model to guide the adjustment of weights.

[0041] In the data flow, the features of each modality are first normalized after feature extraction and fusion to ensure that the features of different modalities are in the same numerical range. The system performs weighted combination of each feature data input through the weight optimizer module. In the code, the data flows from the input module to the feature extraction module, and after feature processing, it is input to the weight calculation module. The weight calculation module calls the Bayesian optimization API and finally outputs the optimized modality weights.

[0042] Weight calculation and multi-modal physiological feature fusion; the system evaluates the importance of each modal feature through a random forest classifier and uses the Gini Index to determine the feature splitting node. Then, the system dynamically adjusts the weights in combination with the Bayesian optimization algorithm to achieve an optimal balance in the contribution of blood sugar, heart rate, body temperature and behavioral data in health assessment.

[0043] The system determines the weight of each modal feature by calculating its mutual information with the target health status (such as hypoglycemia, hyperglycemia, and fever), ensuring that the system can accurately capture the influence of key features. Ultimately, the fused multimodal features are weighted summed to generate a unified health feature vector for subsequent risk assessment and early warning model analysis.

[0044] Personalized health warning strategy generation; Personalized health warning strategy is implemented using Q-learning algorithm, which processes health status and intervention behavior and optimizes intervention strategy. t and possible intervention behaviorsa t The system determines the optimal intervention strategy by updating the Q value. The update rule is as follows:

[0045]

[0046] The system is based on real-time health feedback t Dynamically adjust the Q value and gradually approach the optimal strategy. Through continuous learning, the system can adjust the next intervention strategy after each intervention, so that health fluctuations can be controlled to the greatest extent. During the intervention process, the health data flow flows from the multimodal analysis module to the reinforcement learning module, and the system evaluates the patient's status in real time. After each intervention, the system adjusts the Q value in the Q-learning model based on feedback, and the updated Q value determines the optimal strategy for the next round of intervention.

[0047] The epsilon-greedy strategy balances exploration and utilization of the dynamic adaptability of the optimized intervention strategy; the system adopts the epsilon-greedy strategy to achieve a dynamic balance between exploring new intervention methods and utilizing the existing optimal strategy. In the specific implementation, the system dynamically adjusts the learning rate α and the discount factor γ to adapt to the fluctuation characteristics of children's health data. The adjustment of the learning rate α is based on the fluctuation amplitude and frequency of the health data monitored in real time. When the health parameters show significant fluctuations, the system increases the learning rate to quickly respond to environmental changes; when the health status is stable, the system appropriately reduces the learning rate to ensure the robustness of the strategy optimization. The adjustment of the discount factor γ focuses on long-term return optimization. Especially in the scenario of continuous health abnormalities, the system will increase the discount factor to enhance the ability to predict future health status, thereby generating a more forward-looking intervention plan. Through the adaptive parameter adjustment function, the system tunes α and γ in real time to ensure that the intervention strategy can accurately match the dynamic changes of children's health status. This module realizes the collaborative operation of data privacy protection and model optimization under the federated learning framework. In the federated learning architecture, each smart device worn by the child independently runs a local version of the blood glucose monitoring model, and the model is optimized through a distributed training method. Specifically, the local device executes the back-propagation algorithm on the collected health data, and uses the chain derivation method to calculate the gradient, thereby updating the model parameters. During this process, the device optimizes the efficiency of gradient calculation through batching, and uses the cross-entropy loss function as the objective function to evaluate the classification accuracy of the model. For the distribution characteristics of non-independent and identically distributed (Non-IID) data, the system adopts a distributed optimization strategy to dynamically adjust the parameter updates of the global model through asynchronous aggregation.

[0048] To protect data privacy, the system uses homomorphic encryption technology to encrypt and transmit gradient data. The core principle of the Paillier encryption algorithm is to perform linear operations on ciphertext and implement gradient aggregation without plaintext access on the server side. The gradient data generated by each device is encrypted with a random key, which is dynamically generated and protected by the security module inside the device. During the aggregation process, the server uses a verifiable encryption protocol to verify the integrity of the gradient data to prevent tampering or forgery. To improve the efficiency of encryption calculations, the system relies on the AES-NI instruction set of modern hardware to achieve fast processing of gradient encryption, thereby significantly reducing the computational overhead of encryption operations.

[0049] The differential privacy protection mechanism protects individual data by introducing Laplace noise during the gradient update process. The system defines a privacy budget ∈ to control the intensity of injected noise to ensure privacy protection during the gradient update process. The system equips each device with a privacy budget tracker to monitor the consumption of the privacy budget in real time. The noise injection of differential privacy is supported by a hardware-level random number generator (such as the Intel Secure Key module), which ensures the uniform distribution of noise and the reliability of privacy protection by quickly generating high-quality random numbers.

[0050] After each round of model update, the server calculates the global model weight based on the asynchronous aggregation strategy, and the optimized global model parameters are encrypted and transmitted back to the local device to achieve continuous optimization of the distributed model. At the same time, the system's privacy budget scheduling mechanism dynamically adjusts the privacy budget allocation based on the frequency and scale of each data operation to ensure a balance between model accuracy and data protection.

[0051] The recognition model is optimized through a genetic algorithm. The parameters of the parent model are reorganized using a crossover operation with a crossover probability of 0.7. The model parameters are then perturbed through a mutation operation with a mutation rate of 0.01. The fitness function is called to evaluate the performance of the model in the recognition task, and models with higher fitness are retained for subsequent optimization to improve the accuracy and robustness of the recognition model. The system generates risk warnings and adjusts the next round of intervention strategies based on real-time data and personalized intervention strategies, forming a closed loop in the decision-making process.

[0052] Figure 2 A class diagram provided for an embodiment of the present invention;

[0053] like Figure 2 As shown, the present invention provides a class diagram, including:

[0054] The system includes the following key classes:

[0055] Data Collection Class (DataCollector): This class is mainly used to collect children's blood sugar concentration, heart rate, body temperature and activity status data in real time through sensor devices. The class attributes include sensor type (such as blood sugar sensor, heart rate sensor, acceleration sensor, etc.), sampling frequency and timestamp. Methods include:

[0056] collectGlucoseData(): Collects blood glucose data through the sensor, calls the short-time Fourier transform (STFT) function, converts the time domain signal into the frequency domain signal and extracts the feature changes.

[0057] collectHeartRate(): records heart rate data in real time and smoothes the noise signal through interpolation algorithm.

[0058] collectActivityData(): Collects acceleration data and calculates the child’s motion status by synthesizing the three-axis data.

[0059] Data Preprocessor: This class is responsible for standardizing, denoising, and detecting outliers on the collected data to ensure data quality and consistency. Properties include outlier thresholds (such as ±3σ), normalization ranges, etc. Methods include:

[0060] normalizeData(): Normalizes data to the [0,1] interval to improve model adaptability.

[0061] detectOutliers(): Calls the Z-score algorithm to detect outliers and remove them.

[0062] applyKalmanFilter(): Performs Kalman filter smoothing and optimizes data noise suppression by dynamically adjusting the covariance matrix.

[0063] Multimodal fusion class (DataFusion): This class is used to implement feature fusion of multimodal physiological data and generate comprehensive feature vectors for subsequent analysis. Attributes include weight distribution parameters and feature vector dimensions. Methods include:

[0064] extractFeaturesCNN(): Extracts local features of blood glucose data based on convolutional neural network (CNN).

[0065] processWithLSTM(): Uses a long short-term memory network (LSTM) to process time series features and capture long-term changing patterns in heart rate and activity data.

[0066] optimizeWeights(): Dynamically adjusts the feature weights of each modal data through the Bayesian optimization algorithm.

[0067] Personalized Learning: This class is used to implement adaptive learning based on individual differences. Its core is to continuously optimize the personalized strategy through reinforcement learning. Attributes include Q-learning parameters (such as learning rate α, discount factor γ) and reward function. Methods include:

[0068] updateQValues(): Call the Q-learning update formula to optimize the individual strategy.

[0069] adjustThresholds(): Dynamically adjust the warning thresholds of blood sugar and heart rate to meet personalized needs.

[0070] calculateRewards(): Calculates the immediate reward value based on the health status of individual data.

[0071] ContextAnalyzer: This class recognizes children's behavior contexts based on Hidden Markov Model (HMM) and provides accurate context-adaptive early warning support. Properties include the number of HMM states, state transition probability matrix, and observation sequence. Methods include:

[0072] trainHMM(): Calls the Baum-Welch algorithm to train the hidden Markov model.

[0073] predictContext(): Predicts the current context state through the forward algorithm.

[0074] adjustModelParameters(): Dynamically updates model parameters to adapt to new context data.

[0075] PrivacyProtector: This class is used to protect the privacy of data in the system. Properties include privacy budget parameters, noise distribution type, and encryption key. Methods include:

[0076] addNoise(): Injects Laplace noise into the data through the differential privacy mechanism.

[0077] encryptGradient(): Encrypts the local gradient using the Paillier homomorphic encryption algorithm.

[0078] trackPrivacyBudget(): monitors the consumption of the privacy budget in real time.

[0079] AlertGenerator: This class is responsible for generating health alerts based on real-time data. Properties include alert levels and policy libraries. Methods include:

[0080] generateAlerts(): Generates warning signals based on dynamic changes in blood sugar and heart rate.

[0081] sendAlerts(): Sends warning signals to the guardian via the remote terminal.

[0082] updateStrategies(): Optimize the existing warning plan according to the Q-learning strategy.

[0083] Figure 3The data flow process of the T1DM children's blood sugar monitoring and early warning system in actual usage scenarios is demonstrated. Taking the situation after strenuous exercise as an example, the system collects, processes and analyzes the child's blood sugar, heart rate, body temperature, activity status and other data, and finally generates a real-time health warning signal.

[0084] In this scenario, the specific implementation process of data flow and logical relationship is as follows:

[0085] First, the data acquisition module starts working, and various physiological data are recorded in real time through the sensor device worn on the child. Activity status data is collected through a 10Hz acceleration sensor, blood glucose concentration is sampled every 5 minutes, heart rate is sampled once a minute, and body temperature is updated to the system in real time. During this process, the collectActivityData() function records high-frequency activity status data and transmits the data stream to the data preprocessing module. At the same time, the collectGlucoseData() function performs a short-time Fourier transform (STFT) on the blood glucose data to extract features and add the frequency domain information of blood glucose fluctuations to the data stream.

[0086] In the data preprocessing module, the collected data is first normalized by the normalizeData() function to normalize the different modal data to the [0,1] interval so that different physiological characteristics can be compared on the same scale in subsequent analysis. Then, the applyKalmanFilter() function smoothes the data and suppresses high-frequency noise to eliminate short-term fluctuation anomalies caused by exercise. The prediction and update steps of the Kalman filter dynamically adjust the noise covariance matrix based on the real-time collected activity data to ensure stable blood sugar and heart rate trends in this intense exercise situation. The normalized and smoothed physiological signal data generated by this process is pushed to the data fusion module.

[0087] In the multimodal data fusion module, first, the blood glucose data is passed to the extractFeaturesCNN() function, which uses a CNN structure with 5 convolutional layers to extract local features. The convolutional layer kernel size is 3×3 to ensure that the detailed features of blood glucose fluctuations are extracted under high activity. For heart rate, body temperature and exercise status data, they are processed by the processWithLSTM() function. This function uses a long short-term memory network (LSTM) for time series processing to capture the dynamic change pattern of children's physiological data under intense exercise situations. In the optimizeWeights() function, the system dynamically assigns the importance weight of each modal feature in combination with the Bayesian optimization algorithm. This optimization process is based on data features and situational analysis results, so that multimodal features can reflect higher accuracy and real-time performance under intense exercise, and the fused comprehensive feature vector is passed to the next module.

[0088] The context analysis module uses the predictContext() function to identify contexts based on the child's current intense exercise state. The HMM model identifies that the child is in a high-activity context and automatically updates the model parameters to dynamically adapt to the physiological data characteristics of this high-activity state. Based on the identification, the adjustModelParameters() function calls the forward algorithm of the hidden Markov model and inputs the context state corresponding to the multimodal data into the adaptive learning module.

[0089] The personalized learning module updates the individual Q value through the updateQValues() function of the Q-learning algorithm based on the currently identified high activity state. This function dynamically adjusts the Q value in combination with real-time feedback, and continuously optimizes the warning threshold based on the changes in the child's physical signs in the context of intense exercise. In the adjustThresholds() function, the learning rate α is set to 0.05 to ensure that the system responds quickly in this intense exercise situation. After reinforcement learning optimization, the individualized warning threshold is adjusted, and the system passes the comprehensive individual context threshold and physiological data stream to the privacy protection module.

[0090] In the privacy protection module, first, differential privacy noise is added to the data through the addNoise() function to prevent individual data from being leaked. The system then encrypts the gradient generated during the Q-learning model training process through encryptGradient() to ensure that the child's health data cannot be directly accessed on the system server. The encrypted data is then stored locally for aggregation by the federated learning model.

[0091] Finally, the data enters the warning generation module. The system generates real-time warning signals based on individual thresholds. In the generateAlerts() function, the system determines whether the blood sugar and heart rate fluctuations in the high activity state exceed the set personalized thresholds. If they exceed the set safety range, an emergency notification is sent to the guardian's terminal device through the sendAlerts() function, suggesting that they take necessary intervention measures, such as checking blood sugar or adjusting insulin dosage.

[0092] The overall data flow starts from the data acquisition module, and gradually passes through the data preprocessing, multimodal fusion, situational analysis and personalized learning modules, and then realizes data encryption through the privacy protection module, and finally completes the generation and notification of personalized warning signals in the warning generation module. The system ensures the real-time, accuracy and privacy protection of monitoring through strict data flow design and modular functions.

Claims

1. A blood glucose monitoring and early warning system for children with T1DM, characterized in that: The system comprises: Data acquisition module: The module acquires various physiological data of children at a high frequency. The acquisition equipment uses portable wearable technology to achieve real-time data recording. The system uses short-time Fourier transform (STFT) to perform frequency domain analysis on dynamic physiological data. The frequency domain processing in the module captures subtle physiological fluctuations and extracts short-term change characteristics. For different data types, the data acquisition module uses specific sampling frequency and resolution optimization. Data preprocessing module: The module is responsible for cleaning, smoothing and standardizing the original collected data; the module uses a Kalman filter to smooth dynamic data. During the filtering process, the prediction step of the Kalman filter calculates the optimal estimate of subsequent data based on the state transfer matrix (A), and uses the covariance matrix (Q, R) to adjust the noise suppression to ensure that abnormal spikes interfere with the overall trend of the data stream; the Kalman filter uses dynamic adjustment of the prediction variance and measurement variance to adapt to the changes in different data modes; the module applies the Z score algorithm, sets the threshold range of ±3σ, and removes outliers; after preprocessing, the normalization process uses linear normalization, and the data of different modes are standardized to the [0,1] interval; Multimodal data fusion module: The core purpose of the multimodal data fusion module is to use the correlation information of different physiological characteristics; the module applies convolutional neural network (CNN) to extract features from blood glucose data, and the convolution kernel size is set to 3×3. For long-term dependent signals such as body temperature and activity status, the module uses long short-term memory network (LSTM). The recursive structure of the LSTM network captures the change of heart rate over time and the continuous fluctuation of body temperature; the multimodal data fusion module of the module uses Bayesian optimization algorithm to dynamically adjust the modal weights. Under the Gaussian process model, the expected improvement function (EI) is used to determine the data contribution weights of different modalities. The system is adaptive to different environments and individual differences of children; Adaptive learning module: This module is based on the Q-learning algorithm and uses reinforcement learning to adaptively adjust the differences in individual data. The system uses real-time feedback to continuously correct the Q value and optimize the individual warning threshold. After the reinforcement learning mechanism of Q-learning collects individual characteristic data, it uses the reward and punishment feedback mechanism to dynamically adjust the warning value and algorithm parameters. During the training process, the system uses the long-term data of individual children to iteratively optimize the Q value. The model forms personalized threshold settings and warning trigger conditions based on the physiological response characteristics of specific children. Situational analysis module: The main technology of the module is based on the Hidden Markov Model (HMM), which uses the modeling of hidden states and the probability calculation of observation sequences to identify children's daily behavioral situations; the situational analysis module sets hidden states, including: motion state, resting state, and sleep state, constructs the parameter space of HMM according to the actual physiological data changes, and uses the Baum-Welch algorithm to achieve self-learning and updating of the model; the system uses forward and backward algorithms to calculate the probability of the observation sequence, predict the situation of the next time step, identify the physiological change pattern in a specific scenario, the dynamic decline of blood sugar during exercise or the trend of blood sugar stabilization in the resting state; the system can adaptively adjust the warning threshold in different situations; Privacy protection module: The privacy protection module uses differential privacy mechanism and federated learning framework to jointly realize privacy protection of distributed learning; when the model weights are aggregated, the privacy budget ε is set to 0.5, and a noise introduction mechanism is used to ensure that the personal health data of children cannot be leaked; for data transmission, the system uses homomorphic encryption technology to encrypt gradient data; during each transmission process, the model uses encrypt(gradient) to encrypt the gradient to ensure that the model weights are protected during transmission; the system uses differential privacy mechanism to evaluate the consumption of privacy budget after each data processing, and uses noise regulation to control the use of privacy budget; Early warning module: The module generates personalized health early warning signals and determines the early warning level based on Q-learning strategies and situational analysis results; the early warning levels include normal, caution and danger; the system makes judgments based on the changing trends of blood sugar, heart rate and body temperature; high-risk blood sugar value scenarios trigger emergency alarms, which are sent remotely from terminal to terminal to notify relevant personnel in real time.

2. According to the T1DM children's blood glucose monitoring and early warning system according to claim 1, the data acquisition module further includes an adaptive sampling module for dynamically managing the data acquisition frequency, which can adjust the data acquisition frequency according to different activity states and blood glucose fluctuation amplitudes; the adaptive sampling module monitors the activity data collected by the acceleration sensor in real time, and when a sudden change in the activity state is detected, the heart rate and blood glucose data acquisition frequency is increased to capture physiological parameters; the module uses a moving average filter to smooth the sampled data stream.

3. The T1DM children's blood glucose monitoring and early warning system according to claim 1, wherein the data preprocessing module performs multi-stage data cleaning on physiological data such as blood glucose, heart rate and body temperature; specifically comprising: Noise suppression processing, using the Kalman filter algorithm, the state transfer matrix A is dynamically updated based on historical data, and the covariance moment R is adaptively adjusted; Data repair operation, based on Lagrange interpolation method to fill in the missing data of intermittent collection; In the normalization stage, the consistency of processing of data of different modalities is ensured to improve the accuracy of subsequent data fusion.

4. According to claim 1, the T1DM children's blood glucose monitoring and early warning system, wherein the multimodal data fusion module adopts a weighted fusion algorithm, uses Bayesian optimization to iteratively adjust the fusion weight, and maximizes the accuracy of multimodal data fusion; the module further includes an error control submodule, which calculates the mean square error (MSE) of each eigenvalue of heart rate, blood glucose and body temperature, and sets an error threshold to automatically exclude abnormal eigenvalues; the fusion process sets a dynamic weighting factor W for the output features of the convolutional neural network (CNN) and the long short-term memory network (LSTM) to achieve feature fusion.

5. The T1DM children's blood glucose monitoring and early warning system according to claim 1, wherein the adaptive learning module uses a Q-learning algorithm to adjust the health warning threshold based on the child's individual physiological data; the module further includes a reward feedback mechanism, which sets instant rewards for hypoglycemia and hyperglycemia scenarios in the training cycle, defines accurate early warnings that prioritize blood glucose critical points, and determines the optimal convergence of the Q value in Q-learning under different health conditions.

6. According to claim 1, the T1DM children's blood glucose monitoring and early warning system, wherein the hidden Markov model (HMM) of the context analysis module uses the Baum-Welch algorithm to adaptively adjust the model parameters, and learns the transition probabilities of different hidden states based on the blood glucose and heart rate data in the children's daily activity context; the module includes a prediction algorithm module, which uses forward and backward algorithms to calculate the observation sequence probabilities in different contexts respectively. The system forms predictions on changes in activity contexts and changes in physiological data trends based on probability distribution, and provides feedback to optimize personalized health management plans.

7. The T1DM children's blood glucose monitoring and early warning system according to claim 1, wherein the privacy protection module adopts a differential privacy mechanism and a federated learning framework to achieve distributed learning with privacy protection among various devices; Specifically include: Privacy budget control submodule, the privacy budget epsilon value is 0.5, which is used as the noise intensity control for each data transmission to ensure user privacy protection; The encrypted transmission module uses the homomorphic encryption algorithm encrypt(gradient) to process the gradient data during transmission, and decrypts it when the aggregation model is updated to complete the privacy-preserving federated learning model update.

8. According to claim 1, the T1DM children's blood glucose monitoring and early warning system, wherein the early warning module generates a dynamic health early warning signal based on real-time data; a risk level determination algorithm is provided inside the module, which calculates the current risk value according to the three early warning levels triggered by the dynamic data flow, and combines situational analysis and Q-learning feedback to determine whether to trigger a high-risk early warning and send a notification, suggesting corresponding intervention measures.

Images