Diabetic patient multi-source data real-time analysis method based on large language model

Abstract

The invention relates to the field of intelligent medical monitoring, in particular to a diabetic multi-source data real-time analysis method based on a large language model, and the method mainly comprises the steps: continuously collecting multi-source data of a patient from a CGM device, an intelligent watch and a mobile terminal, and enabling the multi-source data to at least comprise blood glucose data, heart rate and input data; constructing a deep learning model comprising four layers of attention mechanisms, inputting the fusion data into the deep learning model, predicting a blood glucose trend within a preset time by using a sliding time window to obtain a blood glucose prediction result, and pushing the blood glucose prediction result to the mobile terminal or the smart watch; and performing reinforcement learning on the large language model based on a deep Q network architecture, and when the blood glucose control time in the blood glucose data is lower than a preset threshold value, performing parameter updating on the large language model. According to the invention, personalized monitoring and intervention effects on the diabetic patient can be improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of intelligent medical monitoring, in particular to a diabetes patient multi-source data real-time analysis method based on a large language model. BACKGROUND

[0002] With the rapid development of wearable devices and Internet of Things technology, continuous glucose monitoring (CGM) systems have gradually become popular, enabling continuous and real-time collection of blood glucose data. However, current CGM-based systems still have significant shortcomings in data analysis and utilization. On the one hand, most systems only focus on blood glucose values and lack sufficient integration and analysis of multi-source heterogeneous data such as heart rate, exercise volume, sleep quality, and dietary composition, making it difficult to comprehensively and deeply understand the complex causes of blood glucose fluctuations. On the other hand, in terms of data analysis methods, traditional statistical analysis and simple machine learning algorithms are mainly relied upon. These methods have limited feature extraction capabilities when dealing with multi-source, high-dimensional, and dynamic diabetes data, making it difficult to uncover the underlying deep patterns in the data. As a result, the accuracy and timeliness of blood glucose trend prediction are subpar, and reliable evidence for clinical intervention cannot be provided.

[0003] In recent years, large language models have made breakthroughs in natural language processing, with strong semantic understanding, knowledge reasoning, and generation capabilities providing new ideas for solving real-time analysis of diabetes multi-source data. However, there are still many challenges in directly applying large language models to real-time analysis of diabetes data. First, how to convert multi-source heterogeneous diabetes data into a format that large language models can understand and process, effectively integrating and representing the data, is a key problem that needs to be addressed. Second, the knowledge in the diabetes field is highly specialized and rapidly evolving, and how to enable large language models to quickly and accurately learn and master this knowledge and apply it to blood glucose trend prediction and personalized intervention suggestion generation is a core difficulty in improving model practicality. In addition, existing technologies lack a mechanism for continuous optimization of models based on real-time patient feedback and the latest medical research findings, making it difficult for models to adapt to changes in patient conditions and updates in medical knowledge over the long term, and unable to provide continuous, precise, and personalized data analysis and intervention services for diabetes patients.

[0004] Therefore, there is an urgent need for a diabetes blood glucose data analysis method that can achieve real-time monitoring and personalized intervention. SUMMARY

[0005] The present application provides a diabetes patient multi-source data real-time analysis method based on a large language model, which can analyze multi-source data of diabetes patients in real time and accurately, provide personalized intervention plans through a hierarchical feedback mechanism, and update parameters of the large language model based on medical knowledge and feedback from different users to strengthen the large language model, achieving more personalized intervention for diabetes patients.

[0006] To achieve the above object, the present application adopts the following technical solutions:

[0007] In a first aspect, the present application provides a real-time analysis method for multi-source data of diabetic patients based on a large language model, comprising:

[0008] Continuously collecting multi-source data of patients from CGM devices, smart watches and mobile terminals, wherein the multi-source data at least includes blood glucose data, heart rate and input data, and the input data is behavior data manually input by patients in the mobile application interface of the mobile terminal;

[0009] Storing the multi-source data in the mobile terminal and sending the multi-source data to the cloud;

[0010] In the cloud, extracting the time sequence features of the multi-source data through convolutional neural networks and recurrent neural networks to obtain fusion data;

[0011] Constructing a deep learning model containing four layers of attention mechanism, inputting the fusion data into the deep learning model, and predicting the blood glucose trend in a preset time by using a sliding time window to obtain a blood glucose prediction result, and pushing the blood glucose prediction result to the mobile terminal or the smart watch;

[0012] Inputting the multi-source data into a pre-trained large language model to generate personalized suggestions;

[0013] Pushing the personalized suggestions to the mobile terminal or the smart watch based on a hierarchical feedback mechanism, wherein the hierarchical feedback mechanism includes three levels of immediate intervention, short-term adjustment and long-term management;

[0014] The immediate intervention is to trigger the vibration function of the smart watch and the mobile terminal when the blood glucose data fluctuates beyond a preset threshold, and the short-term adjustment and long-term management are to generate short-term personalized suggestions and long-term personalized suggestions according to the multi-source data in a preset time period;

[0015] Based on the deep Q network architecture, reinforcement learning is performed on the large language model, and when the blood glucose control time in the blood glucose data is lower than a preset threshold, the parameters of the large language model are updated.

[0016] In a preferred example of the present application, it can be further provided to further comprise:

[0017] inputting the blood glucose prediction result into an adaptive weight matrix to obtain a blood glucose weight value, the adaptive weight matrix automatically adjusting the influence weight of each index on blood glucose fluctuation in the blood glucose data according to individual differences of the patient, adopting an 8*8 two-dimensional matrix structure, each element in the adaptive weight matrix representing a correlation coefficient between a specific physiological index and blood glucose fluctuation;

[0018] constructing a personalized blood glucose response model, the extracted features of the model including carbohydrate intake, protein intake, fat intake, exercise intensity, exercise duration, insulin dose or oral hypoglycemic drug dosage, stress level and sleep quality;

[0019] inputting the blood glucose weight value into the personalized blood glucose response model to obtain the influence degree of a specific factor on blood glucose fluctuation of the patient;

[0020] pushing the influence degree into the mobile terminal or smart watch.

[0021] In a preferred example of the present application, the large language model can be further configured to be constructed based on a Transformer architecture, and trained using a multi-type medical knowledge database including clinical guidelines and common food blood glucose load index as training data to obtain the pre-trained large language model.

[0022] In a preferred example of the present application, the large language model can be further configured to be constructed based on a Transformer architecture, and trained using a multi-type medical knowledge database including clinical guidelines and common food blood glucose load index as training data to obtain the pre-trained large language model.

[0023] When the blood glucose control time in the blood glucose data is less than 70%, the parameters of the large language model are updated.

[0024] In a preferred example of the present application, the large language model can be further configured to be constructed based on a Transformer architecture, and trained using a multi-type medical knowledge database including clinical guidelines and common food blood glucose load index as training data to obtain the pre-trained large language model.

[0025] The multi-source data, the personalized suggestions and the blood glucose prediction result are transmitted using end-to-end encryption technology.

[0026] In a preferred example of the present application, the large language model can be further configured to be constructed based on a Transformer architecture, and trained using a multi-type medical knowledge database including clinical guidelines and common food blood glucose load index as training data to obtain the pre-trained large language model.

[0027] The multi-source data, the personalized suggestions and the blood glucose prediction result are transmitted using end-to-end encryption technology.

[0028] In a preferred example of the present application, the large language model can be further configured to be constructed based on a Transformer architecture, and trained using a multi-type medical knowledge database including clinical guidelines and common food blood glucose load index as training data to obtain the pre-trained large language model.

[0029] The data acquisition module is used to continuously acquire multi-source data from patients from CGM devices, smartwatches, and mobile terminals. The multi-source data includes at least blood glucose data, heart rate, and input data, wherein the input data is behavioral data manually entered by the patient in the mobile application interface of the mobile terminal; the multi-source data is stored in the mobile terminal and sent to the cloud.

[0030] The blood glucose analysis module is used to extract the temporal features of the multi-source data in the cloud through convolutional neural networks and recurrent neural networks to obtain fused data; construct a deep learning model containing a four-layer attention mechanism, input the fused data into the deep learning model, and use a sliding time window to predict the blood glucose trend within a preset time period to obtain blood glucose prediction results, and push the results to the mobile terminal or smartwatch; input the multi-source data into a pre-trained large language model to generate personalized suggestions;

[0031] An intervention module is used to input the multi-source data into a pre-trained large language model to generate personalized suggestions; based on a hierarchical feedback mechanism, the personalized suggestions are pushed to the mobile terminal or the smartwatch. The hierarchical feedback mechanism includes three levels: immediate intervention, short-term adjustment, and long-term management. The immediate intervention is to trigger the vibration function of the smartwatch and the mobile terminal when the blood glucose data fluctuation exceeds a preset threshold. The short-term adjustment and long-term management are to generate short-term personalized suggestions and long-term optimization suggestions based on the multi-source data within a preset time period.

[0032] The model update module is used to perform reinforcement learning on the large language model based on a deep Q-network architecture. When the blood glucose control time in the blood glucose data is lower than a preset threshold, the parameters of the large language model are updated.

[0033] Thirdly, this application provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the real-time analysis method for multi-source data of diabetic patients based on a large language model as described in any of the preceding claims.

[0034] Fourthly, this application provides a computer-readable storage medium storing a program, wherein when the program is executed by a processor, it implements the real-time analysis method for multi-source data of diabetic patients based on a large language model as described in any of the preceding claims.

[0035] Fifthly, this application provides a computer program product, including computer instructions that, when executed by a processor, implement the steps of the real-time analysis method for multi-source data of diabetic patients based on a large language model as described in any of the preceding claims.

[0036] In summary, compared with the prior art, the beneficial effects of the technical solution provided in this application include at least the following:

[0037] It achieves real-time 24 / 7 blood glucose monitoring and multi-source data fusion, providing a comprehensive data foundation for blood glucose management. High-precision blood glucose trend prediction is achieved through deep learning models, particularly improving the accuracy of early warnings for hypoglycemia and hyperglycemia events. Personalized intervention recommendations are generated based on a medical knowledge-enhanced large language model, ensuring the medical accuracy and operability of the recommendations. A closed-loop blood glucose management system has been established, forming a complete chain from data collection, analysis and prediction to intervention feedback. Intervention strategies are continuously optimized through reinforcement learning methods, enabling the large language model to self-evolve and personalize. A hierarchical feedback mechanism meets the blood glucose management needs at different time scales, forming a systematic solution from immediate intervention to long-term management. Attached Figure Description

[0038] Figure 1 This is a flowchart illustrating a real-time analysis method for multi-source data of diabetic patients based on a large language model, as provided in one embodiment of this application.

[0039] Figure 2 This is a system architecture diagram of a real-time analysis method for multi-source data of diabetic patients based on a large language model, provided as an embodiment of this application.

[0040] Figure 3 This is a block diagram of a device for real-time analysis of multi-source data of diabetic patients based on a large language model, provided as an embodiment of this application. Detailed Implementation

[0041] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0042] In one embodiment of this application, a method for real-time analysis of multi-source data of diabetic patients based on a large language model is provided. Please refer to [link / reference]. Figure 1 As shown, the method includes:

[0043] S100: Continuously collects multi-source data from the patient from the CGM device, smartwatch, and mobile terminal. The multi-source data includes at least blood glucose data, heart rate, and input data, wherein the input data is behavioral data manually entered by the patient in the mobile application interface of the mobile terminal.

[0044] Specifically, the multi-source data acquisition includes a continuous glucose monitoring (CGM) device and a smartwatch. The CGM device continuously monitors and records the patient's blood glucose level at 5-minute intervals with a sampling accuracy of ±0.3 mmol / L. The smartwatch simultaneously collects physiological parameters including heart rate, activity level, sleep quality, and stress level. Data transmission is achieved through Bluetooth 5.0 technology, and the device maintains a 72-hour local data cache capability in the event of a network interruption.

[0045] The mobile application interface allows patients to manually enter dietary components, medication information, and special events. This interface uses visual components to help patients accurately estimate carbohydrate, protein, and fat intake, and supports automatic nutritional analysis of food photos using image recognition technology. The mobile application interface allows patients to record food intake manually. It employs a hierarchical menu structure and includes a database of over 5,000 common foods and their nutritional components. Patients can directly select food types and enter the intake quantity, and the system will automatically calculate the intake of six major nutrient categories, including carbohydrates, protein, fat, dietary fiber, sugar alcohols, and sodium. The application interface also allows patients to modify and mark the entered data. Each modification is automatically recorded, including the modification time and content, and the modification history is retained in a cloud database. Modification thresholds are set for specific sensitive data such as blood glucose levels and medication dosages. When the modification exceeds the preset threshold, a confirmation prompt will be triggered to prevent accidental operation.

[0046] The mobile application interface is designed based on a cross-platform architecture, supporting both iOS and Android operating systems. It achieves adaptive display on devices with different screen sizes through a responsive layout. The application adopts a modular design and includes five core functional modules: data collection, data visualization, nutrition management, medication reminders, and special event recording.

[0047] The visualization component consists of three parts: a food portion estimation tool, a carbohydrate calculator, and a nutrient distribution display. The food portion estimation tool provides a comparison image of standard units of measurement (such as teaspoons, tablespoons, and cups) with the actual weight of the food. The carbohydrate calculator can calculate the total carbohydrate content in real time based on different combinations of ingredients. The nutrient distribution display uses a pie chart to visually present the percentage of each nutrient and labels the corresponding glycemic load index.

[0048] S200: Store the multi-source data in the mobile terminal and send the multi-source data to the cloud.

[0049] Specifically, the mobile application transmits the multi-source data to a cloud server via a wireless network for storage and advanced analysis. Local data caching mechanisms are implemented at three levels: the CGM device, the smartwatch, and the smartphone. The CGM device can locally store 7 days of raw data, the smartwatch can store 48 hours of complete physiological parameters, and the smartphone can retain 72 hours of all collected data and preliminary analysis results. When the network connection is interrupted, the system ensures data integrity through a timestamp mechanism and automatically uploads and synchronizes the data with the cloud after the network is restored, thereby guaranteeing the reliability and continuity of the system in various usage scenarios.

[0050] S300: In the cloud, the temporal features of the multi-source data are extracted through convolutional neural networks and recurrent neural networks to obtain fused data; a deep learning model containing a four-layer attention mechanism is constructed, the fused data is input into the deep learning model, and the blood glucose trend within a preset time period is predicted using a sliding time window to obtain blood glucose prediction results, and the results are pushed to the mobile terminal or smartwatch.

[0051] Specifically, a hybrid architecture integrating convolutional neural networks and recurrent neural networks is used to perform time-series analysis on collected blood glucose data and related physiological parameters. Furthermore, a deep learning model incorporating a four-layer attention mechanism is established, with 12 attention heads, and a sliding time window technique is used to predict blood glucose trends at 30-minute, 60-minute, and 120-minute intervals, respectively.

[0052] The number of attention heads was set to 12 based on clinical validation and model performance optimization. Each attention head is responsible for capturing specific types of data patterns and correlations. The 12 attention heads can focus on various blood glucose fluctuation factors, including postprandial blood glucose response, exercise effects, drug action cycles, and circadian rhythm changes, while maintaining a balance between computational efficiency and model complexity. Experiments show that when the number of attention heads is less than 8, the model's ability to capture complex blood glucose fluctuation patterns is insufficient, while when it is greater than 16, the computational resources are consumed too much and the improvement in prediction accuracy is limited.

[0053] The sliding time window technique is based on the principle of recursive prediction. First, it uses data from the past 6 hours as an input window to predict the blood glucose value for the next 30 minutes. Then, it updates the input window with the prediction result and the actual observation value, and then predicts the blood glucose value for the next 60 minutes. The blood glucose trend for 120 minutes is predicted in the same way. Compared with directly predicting a long time window, this method can effectively reduce the problem of error accumulation in long-term prediction.

[0054] S400: Input the multi-source data into a pre-trained large language model to generate personalized suggestions.

[0055] The large language model is based on a pre-trained Transformer architecture model with 175 billion parameters. It is fine-tuned for specific tasks through medical knowledge enhancement technology. This enhancement is mainly based on four types of medical knowledge: clinical guidelines published by international diabetes academic organizations, evidence-based medicine research evidence bases, knowledge system of metabolic endocrinology textbooks, and validated clinical nutrition and sports medicine knowledge. Through multi-stage instruction fine-tuning technology, the model can understand professional medical terminology and translate it into expressions that patients can understand.

[0056] The sources of the medical knowledge-enhancing data include: clinical guidelines and treatment standards published by authoritative organizations such as the American Diabetes Association (ADA), the International Diabetes Federation (IDF), the European Association for the Study of Diabetes (EASD), and the Chinese Diabetes Society (CDS); 150,000 diabetes-related research papers from PubMed; 1 million de-identified medical records of diabetes patients provided by 8 top-tier hospitals; 20 clinically validated diabetes diet and exercise intervention programs; and a database containing the glycemic load index of 5,000 common foods.

[0057] The knowledge graph construction method adopts a three-stage process: The first stage uses natural language processing technology to extract medical entities and relationships from text data, identifying entities in more than 50 categories, including disease subtypes, symptoms, complications, drugs, foods, and exercise methods, as well as 150 semantic relationships between these entities; the second stage uses ontology fusion technology to integrate knowledge from different sources into a unified framework, and adopts a standardization method based on SNOMED-CT medical terminology to ensure conceptual consistency; the third stage uses graph neural network technology to perform representation learning on the knowledge graph, mapping entities and relationships to a low-dimensional vector space, enabling the large language model to effectively retrieve and utilize structured medical knowledge;

[0058] The model considers the following factors when generating personalized recommendations: the patient's short-term (within 24 hours) and long-term (7-30 days) blood glucose fluctuation patterns, including average blood glucose level, standard deviation, time range, and frequency of low / high blood glucose occurrences; analysis of the patient's dietary records, including total carbohydrate intake, intensity of postprandial peak blood glucose response, food combination ratio, and regularity of meal times; assessment of the patient's exercise status, including exercise type, intensity, duration, and its impact on blood glucose; monitoring of the patient's medication use, including insulin or oral hypoglycemic agent dosage, duration, adherence, and drug response characteristics; and the patient's physiological and psychological parameters, including sleep quality, stress level, menstrual cycle (where applicable), and emotional state.

[0059] The personalized suggestion generation process is based on a combination of parametric templates and conditional generation. First, an appropriate suggestion category is selected based on the patient's blood glucose control status and the analysis results of influencing factors. Then, the core content of the suggestion is generated according to preset medical guideline rules. Finally, the natural language generation capability of the large language model is used to transform this content into personalized and easy-to-understand expressions. The suggestions specifically include: dietary adjustment suggestions down to the type and quantity of food, such as "Today's lunch suggestion is to replace white rice (75g) with whole grain rice (60g) and extra vegetables (100g) to slow down the absorption of carbohydrates"; precise exercise prescriptions based on heart rate zones and exercise tolerance, such as "It is recommended to do 20-30 minutes of moderate-intensity walking (heart rate controlled at 110-125 beats / minute) 90 minutes after dinner to help stabilize evening blood glucose"; and medication adjustment prompts based on real-time monitoring data, such as "Based on the trend of high morning blood glucose over the past 3 days, it is recommended to consider adjusting the basal insulin dose under the guidance of a doctor, increasing it from the current 14 units to 15 units".

[0060] S500: Based on a hierarchical feedback mechanism, the personalized suggestions are pushed to the mobile terminal or the smartwatch. The hierarchical feedback mechanism includes three levels: immediate intervention, short-term adjustment, and long-term management.

[0061] The immediate intervention involves triggering the vibration function of the smartwatch and the mobile terminal when blood glucose data fluctuations exceed a preset threshold. The short-term adjustment and long-term management involve generating short-term personalized suggestions and long-term optimization suggestions based on the multi-source data within a preset time period.

[0062] In practice, the push notification method is based on a multi-device collaboration strategy. The most suitable push method is selected according to the urgency, complexity, and current state of the suggestion: For urgent blood glucose abnormality warnings, a vibration reminder is sent via the smartwatch, and a brief handling suggestion is displayed on the watch face; for diet and exercise suggestions that require detailed reading, a push notification is sent via the mobile application, and the complete suggestion content is provided within the application; for lifestyle adjustment suggestions that require long-term adherence, comprehensive suggestions are sent regularly via email in the form of weekly reports; in addition, the smartwatch is also equipped with a voice prompt function, which can provide real-time voice guidance when it detects that the patient is engaged in a specific activity (such as exercise). The suggestion content will automatically adjust the complexity of expression according to the patient's reading habits and feedback history to ensure the effectiveness of information delivery.

[0063] The immediate intervention level is characterized by high priority, rapid response, and safety orientation. This level intervenes in the predicted risk of sharp fluctuations in blood glucose, including a low blood glucose risk warning (below 3.9 mmol / L), a high blood glucose risk warning (above 13.9 mmol / L), and a rapid fluctuation warning (a rate of change in blood glucose exceeding 0.2 mmol / L / minute). The necessary steps of immediate intervention include: alerting the patient via both vibration and sound from a smartwatch; simultaneously displaying the warning level and specific blood glucose value on both the smartwatch and mobile phone screens; providing standardized emergency treatment guidance, such as recommending the intake of 15-20g of fast-acting carbohydrates in case of hypoglycemia; and reminding the patient to retest blood glucose 15 minutes after taking action. Optional steps include: adjusting the warning intensity according to the patient's current activity status (e.g., exercising or driving); allowing the patient to notify a preset emergency contact with one click; providing the location information of nearby medical institutions; and triggering smart home devices (e.g., turning on lights or adjusting room temperature) under specific conditions (e.g., hypoglycemia occurring during sleep at night).

[0064] The short-term adjustment level is characterized by its preventative, personalized, and context-adaptive nature. This level addresses deviations from the target range in blood glucose management patterns detected within 24-72 hours, including patterns such as persistently high postprandial blood glucose peaks, increased nighttime blood glucose fluctuations, abnormal exercise-related blood glucose fluctuations, and unstable medication effects. Necessary steps for short-term adjustment include: conducting a comprehensive analysis of the patient's recent blood glucose data every 24 hours to identify problem patterns in blood glucose management; generating targeted dietary, exercise, or medication adjustment suggestions based on the problem type; and providing specific and actionable action plans, such as "recommending to reduce breakfast carbohydrate intake from the current 65g to 45g and increase protein intake by 10g to slow glucose absorption." Optional steps include: providing personalized alternative food recommendations; automatically adjusting exercise recommendations based on weather and the patient's schedule; generating a visual comparison chart of blood glucose trends and intervention effects; and initiating customized behavioral reminders (such as reminding to check blood glucose 30 minutes before meals).

[0065] The long-term management level is characterized by its comprehensiveness, educational nature, and adaptability. This level analyzes blood glucose management data from 7 to 30 days, focusing on long-term indicators such as the time-in-range of blood glucose control, estimated glycated hemoglobin, and coefficient of variation of blood glucose, as well as the patient's lifestyle patterns and medication adherence. Essential steps of long-term management include: generating a comprehensive assessment report every 7 days; analyzing the gap between the patient's blood glucose control status and the target; identifying key factors affecting long-term blood glucose control; providing personalized lifestyle adjustment plans, including dietary optimization, exercise habit cultivation, and stress management strategies; and creating phased blood glucose management goals. Optional steps include: providing diabetes-related educational content based on the patient's learning preferences; generating a 14-day dietary plan tailored to the patient's taste preferences; establishing personalized incentive mechanisms and achievements; supporting the participation of telemedicine teams in assessment and plan adjustments; and dynamically adjusting long-term management strategies according to seasonal changes, work stress cycles, and special physiological periods (such as the female menstrual cycle).

[0066] The three levels of intervention work together to form a closed-loop management system. The records of immediate intervention serve as input data for short-term adjustments. The effectiveness evaluation of short-term adjustments is integrated into the decision-making process of long-term management. The individual patient characteristics identified in long-term management are fed back into the parameter settings of immediate intervention and short-term adjustments. The trigger thresholds and content depth of the three levels of intervention are dynamically adjusted through automated rules and artificial intelligence algorithms to ensure the consistency and personalization of intervention recommendations.

[0067] S600: The large language model is subjected to reinforcement learning based on a deep Q-network architecture. When the blood glucose control time in the blood glucose data is lower than a preset threshold, the parameters of the large language model are updated.

[0068] Specifically, the reinforcement learning method is implemented based on the Deep Q-Network (DQN) architecture. This method models the blood glucose management process as a Markov decision process, where the state space includes the patient's blood glucose level, physiological parameters, and environmental factors; the action space includes intervention suggestions of different types and intensities; the reward function is constructed based on three dimensions: time in range (TIR), blood glucose coefficient of variation, and avoidance of extreme blood glucose events. The historical state-action-reward-new state quadruple is stored through experience replay technology, and a bi-objective network design is used to reduce the bias of Q-value estimation, thereby continuously optimizing the intervention strategy while maintaining model stability.

[0069] The intervention process is divided into two phases: exploration and utilization. In the exploration phase, an ε-greedy strategy is used to select random intervention suggestions with an initial probability of 0.3, which is gradually reduced to a stable exploration rate of 0.05 over time, and the blood glucose response pattern corresponding to each intervention suggestion is recorded. In the utilization phase, the optimal intervention strategy is selected based on historical effects, and samples with high temporal difference errors are learned first through a priority experience replay mechanism to achieve efficient learning of rare but important scenarios (such as acute hypoglycemic events).

[0070] The patient feedback and intervention effectiveness evaluation mechanism adopts a multi-dimensional indicator system, including two main categories: objective indicators and subjective indicators. Objective indicators cover clinically relevant parameters such as time interval (TIR) ​​of blood glucose, coefficient of variation (CV) of blood glucose, frequency of hypoglycemia and hyperglycemia events, and estimated glycated hemoglobin (eA1c). Subjective indicators include patient satisfaction ratings (1-5 points) of intervention suggestions provided through the application, evaluation of implementation difficulty, and text-based feedback. Natural language processing technology is used to analyze the emotional tendencies and key issues in the patient's text feedback.

[0071] The 70% threshold for time range of blood glucose control (TIR) ​​is based on the clinical goals of international diabetes management guidelines. This indicator reflects the percentage of time a patient's blood glucose is maintained within the target range of 3.9-10.0 mmol / L. Studies have shown that when TIR < 70%, the risk of diabetic complications increases significantly. The 80% threshold for adherence is based on behavioral medicine research results. Data shows that when treatment adherence is below 80%, the intervention effect is significantly reduced. The setting of these two thresholds takes into account both clinical safety and behavioral feasibility, ensuring that situations with poor blood glucose management can be identified and addressed in a timely manner.

[0072] The model parameter update process adopts a hierarchical adaptive adjustment strategy. When the update condition is triggered, the specific reasons for the performance decline are first analyzed. For the case of decreased blood glucose prediction accuracy, the weight parameters of the deep learning model are updated. For the case of low compliance with intervention suggestions, the expression mode and suggestion complexity of the natural language generation module are adjusted. For the case of poor blood glucose control, the reward function and state representation of the reinforcement learning model are recalibrated. The update process adopts a progressive strategy, first verifying the effectiveness of the new parameters in an offline environment, then testing them on a limited number of samples, and finally fully deploying the update.

[0073] In this method, the overall system architecture is as follows: Figure 2 As shown, Figure 2 The multi-source acquisition module corresponds to steps S100-S200, the pattern recognition and analysis module corresponds to steps S300-S400, and the intelligent intervention module corresponds to steps S500-S600.

[0074] In this embodiment, real-time 24 / 7 blood glucose monitoring and multi-source data fusion are achieved, providing a comprehensive data foundation for blood glucose management. High-precision blood glucose trend prediction is achieved through a deep learning model, particularly improving the accuracy of early warnings for hypoglycemia and hyperglycemia events. Personalized intervention suggestions are generated based on a medical knowledge-enhanced large language model, ensuring the medical accuracy and operability of the recommendations. A closed-loop blood glucose management system is established, forming a complete chain from data collection, analysis and prediction to intervention feedback. Intervention strategies are continuously optimized through reinforcement learning methods, enabling the large language model to self-evolve and personalize. A hierarchical feedback mechanism meets the blood glucose management needs at different time scales, forming a systematic solution from immediate intervention to long-term management.

[0075] In some embodiments, it also includes:

[0076] The blood glucose prediction results are input into an adaptive weight matrix to obtain blood glucose weight values. The adaptive weight matrix automatically adjusts the influence weights of various indicators on blood glucose fluctuations in the blood glucose data according to the individual differences of the patients. It adopts an 8×8 two-dimensional matrix structure. Each element in the adaptive weight matrix represents the correlation coefficient between a specific physiological indicator and blood glucose fluctuations.

[0077] A personalized blood glucose response model was constructed, and the extracted features of the model included carbohydrate intake, protein intake, fat intake, exercise intensity, exercise duration, insulin dose or oral hypoglycemic drug dosage, stress level, and sleep quality.

[0078] The blood glucose weight value is input into the personalized blood glucose response model to obtain the degree of influence of specific factors on the patient's blood glucose fluctuations.

[0079] The level of influence is pushed to the mobile terminal or smartwatch.

[0080] In practice, the adaptive weight matrix is ​​a dynamically adjusted multidimensional mathematical model. The purpose of introducing this matrix is ​​to solve the problem that different diabetic patients have significant individual differences in response to the same influencing factors. Traditional fixed weight models cannot adapt to individualized blood glucose response characteristics, while the adaptive weight matrix achieves dynamic balance of the importance of different factors by continuously learning the patient's blood glucose fluctuation patterns.

[0081] The adaptive weight matrix adopts an 8×8 two-dimensional matrix structure. Each element in the matrix represents the correlation coefficient between a specific physiological indicator and blood glucose fluctuation. The matrix element values ​​range from [-1, 1], where positive values ​​indicate a positive correlation, negative values ​​indicate a negative correlation, and zero values ​​indicate no significant correlation. The matrix is ​​compressed in dimension using singular value decomposition technology to extract the interaction relationship between the main influencing factors.

[0082] The dynamic trade-off mechanism continuously optimizes the values ​​in the weight matrix through the gradient descent algorithm. The system calculates the mean square error between the predicted blood glucose value and the actual blood glucose value every 24 hours and uses this as the loss function to backpropagate and update the weight coefficients. At the same time, an L1 regularization term is introduced to prevent overfitting and ensure that the model has sufficient generalization ability while maintaining accuracy.

[0083] The individual difference adjustment is achieved based on three levels of data analysis: First, the system establishes a baseline model for patients, acquiring their basic physiological parameters and blood glucose variation patterns through a 14-day data collection period; second, the system establishes a patient classification model based on patients' age, gender, diabetes type, disease duration, and clinical biochemical indicators, comparing patients with groups of similar characteristics; finally, the system continuously adjusts weights in the real-time data stream using reinforcement learning methods, enabling it to quickly adapt and readjust the weight matrix when patients' blood glucose response patterns change significantly (e.g., seasonal changes, lifestyle changes).

[0084] The personalized blood glucose response model includes at least eight characteristic variables, specifically: carbohydrate intake (g / meal), protein intake (g / meal), fat intake (g / meal), exercise intensity (measured in metabolic equivalents MET), exercise duration (minutes), insulin dose (units) or oral hypoglycemic agent dosage, stress level (1-10 scale), and sleep quality (percentage of sleep stages). These eight variables are the core variables with the most significant impact on blood glucose fluctuations, selected from more than 20 possible influencing factors through characteristic importance analysis.

[0085] The selection of the eight characteristic variables is based on the results of clinical research and data-driven analysis. Clinical research shows that these factors are the main variables affecting blood glucose fluctuations, while data-driven analysis confirms the significance of these variables through random forest feature importance ranking and analysis of variance. In certain special patient groups (such as patients with gestational diabetes or elderly patients with diabetes), the system can include additional characteristic variables as needed, such as changes in hormone levels or cognitive function status.

[0086] The quantitative assessment results are presented in the form of an "influence index," which ranges from 0 to 100 and represents the degree of influence of specific factors on the patient's blood glucose fluctuations. Based on this index, the system generates a personalized sensitivity map of influencing factors, which intuitively displays the patient's sensitivity to different factors. For example, a patient's sensitivity to carbohydrates is 85 (high sensitivity), to exercise is 45 (moderate sensitivity), and to stress is 65 (moderate to high sensitivity). These quantitative assessment results serve as an important basis for the intelligent intervention module to generate personalized suggestions, helping patients understand the main influencing factors of their blood glucose fluctuations and to adjust their lifestyle and medication regimens accordingly.

[0087] In this embodiment, the influence weights of eight core feature variables on blood glucose are quantified in real time using an adaptive weight matrix, which significantly improves the individualized fit of the model.

[0088] In some embodiments, the large language model is built on the Transformer architecture and trained using multiple medical knowledge databases, including clinical guidelines and glycemic load indices of common foods, to obtain the pre-trained large language model.

[0089] In practice, the large language model is based on a pre-trained Transformer architecture model with 175 billion parameters. It is fine-tuned for specific tasks through medical knowledge enhancement technology. This enhancement is mainly based on four types of medical knowledge: the clinical guidelines system published by the International Diabetes Association, the evidence base of evidence-based medicine research, the knowledge system of metabolic endocrinology textbooks, and the validated clinical nutrition and sports medicine knowledge. Through multi-stage instruction fine-tuning technology, the model can understand professional medical terms and translate them into expressions that patients can understand.

[0090] The sources of the medical knowledge-enhancing data include: clinical guidelines and treatment standards published by authoritative organizations such as the American Diabetes Association (ADA), the International Diabetes Federation (IDF), the European Association for the Study of Diabetes (EASD), and the Chinese Diabetes Society (CDS); 150,000 diabetes-related research papers from PubMed; 1 million de-identified medical records of diabetes patients provided by 8 top-tier hospitals; 20 clinically validated diabetes diet and exercise intervention programs; and a database containing the glycemic load index of 5,000 common foods.

[0091] The knowledge graph construction method adopts a three-stage process: The first stage uses natural language processing technology to extract medical entities and relationships from text data, identifying more than 50 categories of entities, including disease subtypes, symptoms, complications, drugs, foods, and exercise methods, as well as 150 semantic relationships between these entities; the second stage uses ontology fusion technology to integrate knowledge from different sources into a unified framework, and adopts a standardization method based on the SNOMED-CT medical terminology system to ensure conceptual consistency; the third stage uses graph neural network technology to perform representation learning on the knowledge graph, mapping entities and relationships to a low-dimensional vector space, enabling the large language model to effectively retrieve and utilize structured medical knowledge;

[0092] In this embodiment, the prediction accuracy of the large language model is improved by using a specific training dataset.

[0093] In some embodiments, it also includes:

[0094] The multi-source data, the personalized recommendations, and the blood glucose prediction results are transmitted using end-to-end encryption technology.

[0095] In practical implementation, the end-to-end encryption technology adopts a layered encryption architecture. At the device level, AES-256-bit hardware encryption is used to protect locally stored data. The transport layer uses the TLS 1.3 protocol combined with Elliptic Curve Cryptography (ECC) to achieve secure data transmission. The cloud storage layer uses homomorphic encryption technology to allow the system to perform specific computational operations on encrypted data without decryption. The implementation process of this encryption system includes: generating a unique device key pair during device initialization and storing the private key through a secure element; establishing a session key through key negotiation before data transmission; using attribute-based encryption (ABE) technology in the cloud to achieve fine-grained access control, ensuring that only authorized users and specific medical personnel can access the corresponding data; the system also implements a key rotation mechanism, automatically updating the encryption key every 90 days and retaining historical keys for decrypting earlier data.

[0096] The federated learning method is implemented based on a horizontal federated learning framework. This method distributes the model training process across multiple client devices without centralizing the original data. The specific implementation steps include: the system first deploys the initial model architecture and parameters on a central server; client devices train the model using local data, but only transmit gradient updates or model parameter differences (not the original data) to the central server; the central server integrates model updates from multiple clients using a secure aggregation algorithm, and performs global model updates using the federated averaging (FedAvg) algorithm; to further enhance privacy protection, the system applies differential privacy technology before parameter aggregation, preventing the model from back-deriving personal data by adding calibration noise to the gradient.

[0097] The federated learning method can simultaneously achieve model optimization and privacy protection because: the original patient data is always kept on the local device and does not need to be uploaded to the central server, fundamentally eliminating the risk of data leakage; model weight updates are aggregated through secure multi-party computation (SMC) technology, so even if the intermediate transmission data is intercepted, the original information cannot be recovered; data from different regions and populations contribute training signals through the federated learning framework, significantly improving the model's generalization ability and applicability; the system also sets differential privacy budget limits, which automatically pause a user's model contribution when a single user's privacy contribution exceeds a threshold, preventing the inference of personal data through multiple model updates;

[0098] In this embodiment, patient privacy is protected and the security of output transmission is improved.

[0099] In some embodiments, it also includes:

[0100] The system receives remote access requests from healthcare professionals via an application programming interface (API) and sends the multi-source data, personalized recommendations, and blood glucose prediction results to the terminal devices of healthcare professionals with access permissions.

[0101] In practice, remote access and intervention support for medical professionals are achieved through hierarchical permission management. The system provides a medical professional version application interface that supports access for three roles: attending physicians have full data viewing permissions and treatment plan adjustment permissions; diabetes educators can view patients' blood glucose trends and behavioral data and provide educational guidance; nutritionists and exercise therapists can access data in their respective professional fields and provide professional advice; medical professionals can view patients' blood glucose fluctuation trends, intervention recommendation history, and patient feedback through a secure console, set personalized treatment goals and warning parameters, send professional advice to patients through the system (distinguished from automatically generated advice), and initiate video consultations when necessary; the system also provides doctor-patient collaboration tools, such as a shared decision-making interface and treatment goal tracking table, allowing patients to selectively share specific data with the medical team and grant time-limited emergency data access permissions as needed, thereby achieving a closed-loop management model for doctor-patient collaboration.

[0102] In this embodiment, the accuracy of personalized intervention is improved by connecting with medical professionals.

[0103] This application also provides a real-time analysis device for multi-source data of diabetic patients based on a large language model. Please refer to [link / reference]. Figure 3 As shown, the device includes:

[0104] The data acquisition module 100 is used to continuously acquire multi-source data from patients from CGM devices, smartwatches, and mobile terminals. The multi-source data includes at least blood glucose data, heart rate, and input data, wherein the input data is behavioral data manually entered by the patient in the mobile application interface of the mobile terminal. The multi-source data is stored in the mobile terminal and sent to the cloud.

[0105] The blood glucose analysis module 200 is used to extract the temporal features of the multi-source data in the cloud through convolutional neural networks and recurrent neural networks to obtain fused data; construct a deep learning model containing a four-layer attention mechanism, input the fused data into the deep learning model, and use a sliding time window to predict the blood glucose trend within a preset time period to obtain blood glucose prediction results, and push the results to the mobile terminal or smartwatch; input the multi-source data into a pre-trained large language model to generate personalized suggestions;

[0106] The intervention module 300 is used to input the multi-source data into a pre-trained large language model to generate personalized suggestions; and to push the personalized suggestions to the mobile terminal or the smartwatch based on a hierarchical feedback mechanism, which includes three levels: immediate intervention, short-term adjustment, and long-term management; the immediate intervention is to trigger the vibration function of the smartwatch and the mobile terminal when the blood glucose data fluctuation exceeds a preset threshold; the short-term adjustment and long-term management are to generate short-term personalized suggestions and long-term optimization suggestions based on the multi-source data within a preset time period.

[0107] The model update module 400 is used to perform reinforcement learning on the large language model based on a deep Q-network architecture. When the blood glucose control time in the blood glucose data is lower than a preset threshold, the parameters of the large language model are updated.

[0108] The functional implementation of each module in the above-mentioned real-time analysis device for multi-source data of diabetic patients based on a large language model corresponds to the steps in the above-mentioned real-time analysis method embodiment for multi-source data of diabetic patients based on a large language model. Their functions and implementation processes will not be described in detail here.

[0109] This application also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the real-time analysis method for multi-source data of diabetic patients based on a large language model as described in any of the above embodiments.

[0110] This application also provides a computer-readable storage medium on which a program is stored. The computer-readable storage medium refers to a data storage medium, which may include, but is not limited to, floppy disks, optical disks, hard disks, flash memory, USB flash drives, and / or Memory Sticks. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The working process, details, and technical effects of the computer-readable storage medium provided in this embodiment can be found in the above embodiment regarding a real-time analysis method for multi-source data of diabetic patients based on a large language model, and will not be repeated here.

[0111] The application also provides a computer program product, including computer instructions that, when executed by a processor, implement the steps of the real-time analysis method for multi-source data of diabetic patients based on a large language model as described in any of the above embodiments.

[0112] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0113] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as the combination of these technical features does not contradict each other, it should be considered within the scope of this specification. The above embodiments only illustrate several implementation methods of this application, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for real-time analysis of multi-source data of diabetic patients based on a large language model, characterized in that, include: Multi-source data from patients are continuously collected from CGM devices, smartwatches, and mobile terminals. The multi-source data includes at least blood glucose data, heart rate, and input data, wherein the input data is behavioral data manually entered by the patient in the mobile application interface of the mobile terminal. The multi-source data is stored in the mobile terminal and then sent to the cloud. In the cloud, the temporal features of the multi-source data are extracted using convolutional neural networks and recurrent neural networks to obtain fused data; A deep learning model containing a four-layer attention mechanism is constructed. The fused data is input into the deep learning model, and the blood glucose trend within a preset time period is predicted using a sliding time window to obtain the blood glucose prediction result. The blood glucose prediction result is then pushed to the mobile terminal or smartwatch. The multi-source data is input into a pre-trained large language model to generate personalized suggestions. The personalized suggestions are pushed to the mobile terminal or the smartwatch based on a hierarchical feedback mechanism, which includes three levels: immediate intervention, short-term adjustment and long-term management. The immediate intervention is to trigger the vibration function of the smartwatch and the mobile terminal when the blood glucose data fluctuation exceeds the preset threshold. The short-term adjustment and long-term management are to generate short-term personalized suggestions and long-term optimization suggestions based on the multi-source data within the preset time period. The large language model is subjected to reinforcement learning based on a deep Q-network architecture. When the blood glucose control time in the blood glucose data is lower than a preset threshold, the parameters of the large language model are updated.

2. The method for real-time analysis of multi-source data of diabetic patients based on a large language model according to claim 1, characterized in that, Also includes: The blood glucose prediction results are input into an adaptive weight matrix to obtain blood glucose weight values. The adaptive weight matrix automatically adjusts the influence weights of various indicators on blood glucose fluctuations in the blood glucose data according to the individual differences of the patients. It adopts an 8×8 two-dimensional matrix structure. Each element in the adaptive weight matrix represents the correlation coefficient between a specific physiological indicator and blood glucose fluctuations. A personalized blood glucose response model was constructed, and the extracted features of the model included carbohydrate intake, protein intake, fat intake, exercise intensity, exercise duration, insulin dose or oral hypoglycemic drug dosage, stress level, and sleep quality. The blood glucose weight value is input into the personalized blood glucose response model to obtain the degree of influence of specific factors on the patient's blood glucose fluctuations. The level of influence is pushed to the mobile terminal or smartwatch.

3. The method for real-time analysis of multi-source data of diabetic patients based on a large language model according to claim 1, characterized in that, The large language model is built on the Transformer architecture and is trained using multiple medical knowledge databases, including clinical guidelines and glycemic load indices of common foods, to obtain the pre-trained large language model.

4. The method for real-time analysis of multi-source data of diabetic patients based on a large language model according to claim 1, characterized in that, When the blood glucose control time in the blood glucose data is lower than a preset threshold, the parameters of the large language model are updated, including: When the blood glucose control time in the blood glucose data is less than 70%, the parameters of the large language model are updated.

5. The method for real-time analysis of multi-source data of diabetic patients based on a large language model according to claim 1, characterized in that, Also includes: The multi-source data, the personalized recommendations, and the blood glucose prediction results are transmitted using end-to-end encryption technology.

6. The method for real-time analysis of multi-source data of diabetic patients based on a large language model according to claim 5, characterized in that, Also includes: The system receives remote access requests from healthcare professionals via an application programming interface (API) and sends the multi-source data, personalized recommendations, and blood glucose prediction results to the terminal devices of healthcare professionals with access permissions.

7. A real-time analysis device for multi-source data of diabetic patients based on a large language model, characterized in that, include: The data acquisition module is used to continuously acquire multi-source data from patients from CGM devices, smartwatches and mobile terminals. The multi-source data includes at least blood glucose data, heart rate and input data. The input data is behavioral data manually entered by the patient in the mobile application interface of the mobile terminal. The multi-source data is stored in the mobile terminal and then sent to the cloud. The blood glucose analysis module is used to extract the temporal features of the multi-source data in the cloud through convolutional neural networks and recurrent neural networks to obtain fused data; construct a deep learning model containing a four-layer attention mechanism, input the fused data into the deep learning model, and use a sliding time window to predict the blood glucose trend within a preset time period to obtain blood glucose prediction results, and push the results to the mobile terminal or smartwatch. The multi-source data is input into a pre-trained large language model to generate personalized suggestions. The intervention module is used to input the multi-source data into a pre-trained large language model to generate personalized suggestions; The personalized suggestions are pushed to the mobile terminal or the smartwatch based on a hierarchical feedback mechanism, which includes three levels: immediate intervention, short-term adjustment, and long-term management. The immediate intervention is to trigger the vibration function of the smartwatch and the mobile terminal when the blood glucose data fluctuates beyond a preset threshold. The short-term adjustment and long-term management are to generate short-term personalized suggestions and long-term optimization suggestions based on the multi-source data within a preset time period. The model update module is used to perform reinforcement learning on the large language model based on a deep Q-network architecture. When the blood glucose control time in the blood glucose data is lower than a preset threshold, the parameters of the large language model are updated.

8. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for real-time analysis of multi-source data of diabetic patients based on a large language model as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program, wherein when the program is executed by a processor, it implements the method for real-time analysis of multi-source data of diabetic patients based on a large language model as described in any one of claims 1 to 6.

10. A computer program product comprising computer instructions, characterized in that, When executed by a processor, the computer instructions implement the steps of the real-time analysis method for multi-source data of diabetic patients based on a large language model as described in any one of claims 1 to 6.

Images