Diabetes health management method and system based on AI big data

Abstract

The invention relates to the technical field of medical health, and discloses a diabetes health management method and system based on AI big data, and the method comprises the steps: collecting and preprocessing multi-dimensional data, collecting the biological state, behavior selection and environmental factor data of a patient through a multi-modal sensing network, and carrying out the standardization and feature extraction; modeling a triple helix dynamic system, constructing a coupling dynamic model of three dimensions, and realizing complex relationship expression through a heterogeneous graph neural network; performing cooperative resonance state identification and optimization, identifying a forward enhancement mode among three-dimensional factors, and searching an optimal combination relationship; multi-objective reinforcement learning intervention strategy generation: generating a personalized intervention strategy for guiding the patient to transfer to the optimal resonance state; closed-loop feedback and adaptive optimization are carried out, and system parameters and intervention strategies are continuously adjusted. Unification of accurate monitoring and comprehensive management is realized, the synergistic interaction among three dimensions is found, and the intervention complexity is reduced.

Description

Technical Field

[0001] This invention relates to the field of medical and health technology, and more specifically, to a method and system for diabetes health management based on AI big data. Background Technology

[0002] Diabetes mellitus is a common chronic metabolic disease with a continuously increasing global prevalence. Effective long-term health management is crucial for controlling disease progression and preventing complications. Currently, diabetes health management technologies mainly achieve this through monitoring biomarkers, providing behavioral interventions, or adjusting environmental support. However, existing technologies have the following shortcomings: First, traditional diabetes management systems typically optimize biological indicators (such as blood glucose control), behavioral interventions (such as dietary recommendations), or environmental modifications (such as social support) independently, neglecting the complex interactions between the three, resulting in unsatisfactory intervention effects and low patient compliance.

[0003] Secondly, existing technologies struggle to balance the precision and comprehensiveness of management. Pursuing precise control often results in too many monitoring indicators and complex intervention plans, while pursuing comprehensive management often leads to insufficient monitoring of key indicators.

[0004] Third, current management systems often focus on short-term blood sugar control and lack systematic consideration for cultivating long-term healthy behaviors, which may lead to good results for patients in the early stages, but make it difficult to maintain in the long term.

[0005] Fourth, existing technologies often employ single or limited-dimensional intervention models, failing to form a cohesive overall plan. This leads to conflicts or overlaps between multiple independent interventions, increasing the management burden and complexity for patients.

[0006] Fifth, most traditional systems adopt decision-making models based on fixed rules, which cannot adapt to the dynamic changes in patient status, behavior and environment, and lack personalization and adaptive capabilities.

[0007] Finally, although existing technologies can collect a large amount of patient data, the interaction relationships and potential value between the data have not been fully explored, especially the correlation patterns and synergistic effects between data from different dimensions have been overlooked. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a diabetes health management method and system based on AI big data.

[0009] A diabetes health management method based on AI and big data includes the following steps: Data on patients' biological status, behavioral choices, and environmental factors were collected and preprocessed to form a structured three-dimensional dataset. A coupled dynamic model of biological state, behavioral choice, and environmental factors is constructed based on the triple helix resonance theory, and the complex relationship between the three dimensions is expressed and calculated through a heterogeneous graph neural network. By using a synergistic resonance detector to identify positive enhancement patterns between three-dimensional factors, we can find resonance points that can produce extraordinary effects. By applying a multi-objective reinforcement learning algorithm, a personalized intervention strategy can be generated that guides the patient from the current state to the optimal resonance state. Based on feedback data from patients after implementing intervention strategies, a closed-loop adaptive optimization mechanism is constructed to continuously adjust and improve the triple helix resonance framework and its parameters. The heterogeneous graph neural network includes a node type mapping module, a relation-specific convolutional layer, and an attention fusion module, which are used to capture asymmetric interaction patterns between three types of nodes.

[0010] Preferably, the biological status data includes physiological indicators such as blood glucose curves, blood pressure values, weight, and heart rate variability; the behavioral selection data includes behavioral data such as diet records, exercise data, sleep patterns, and medication records; and the environmental factor data includes geographical location, weather conditions, participation in social activities, work status and stress levels, and family and social support.

[0011] Preferably, the steps of constructing a coupled dynamic model of biological state, behavioral choice, and environmental factors using the triple helix resonance theory include: Construct a system of differential equations describing the interaction of the three dimensions; Construct a heterogeneous graph containing three types of nodes and multiple types of edges to represent complex relationships in the system; Construct graph convolutional layers suitable for heterogeneous graphs, applying different transformation matrices to different types of nodes and edges; Heterogeneous graph neural networks are trained using time-series data to learn the parameters and evolutionary patterns in the dynamic equations.

[0012] Preferably, the synergistic resonance detector identifies positive enhancement patterns among three-dimensional factors by calculating a synergistic resonance index. The synergistic resonance index is calculated by dividing the overall health outcome score generated by the combination of the current biological state, behavioral choices, and environmental factors by the sum of the health outcome scores under the three single-dimensional optimization conditions.

[0013] Preferred: The multi-objective reinforcement learning algorithm includes: using a dynamic model as an environment simulator for reinforcement learning; constructing a multi-objective composite reward function that includes blood glucose control, intervention compliance, quality of life, and synergistic effects; training the intervention strategy network using a deep Q-network or a proximal policy optimization algorithm; performing personalized fine-tuning on the trained basic strategy model; and generating short-term, medium-term, and long-term intervention plans.

[0014] Preferably, the multi-objective composite reward function is: ; in, For the reward function of improved blood glucose control, To intervene in compliance reward functions, For quality of life reward function, It is a cooperative resonance exponential function. to These are the weighting coefficients for each objective.

[0015] Preferably, the closed-loop adaptive optimization mechanism includes: By collecting multidimensional feedback data on the intervention effect through real-time monitoring by smart devices, subjective evaluations by patients, and medical records, a comprehensive evaluation model is constructed to conduct multidimensional analysis of the effectiveness of the intervention strategy. Based on feedback data, key parameters in the triple helix dynamics model are automatically adjusted. Newly identified resonance patterns, effective intervention strategies, and feedback patterns are added to the system's knowledge base; Based on the latest model parameters and knowledge base, the intervention strategy is recalculated and optimized.

[0016] Preferably, the key parameters in the automatically adjusted triple helix dynamics model are obtained using a Bayesian optimization algorithm, which continuously updates the posterior distribution of the model parameters based on new data.

[0017] Preferably: the graph convolutional layer of the heterogeneous graph neural network pairs nodes. Representation vector The update process is as follows: ; in, Let v be the representation vector of node v at the (l+1)th layer. Let u be the representation vector of node u in the l-th layer. To pass through relationships With nodes The set of connected neighbor nodes, The size of the set, Let r be the weight matrix of relation r at level l. Let be the weight matrix of the self-loop at layer l. This is the activation function.

[0018] An AI-based big data-driven diabetes health management system, used to execute the above method, includes: The multi-dimensional data acquisition module is used to collect data on three dimensions: the patient's biological state, behavioral choices, and environmental factors, and to preprocess the data to form a structured three-dimensional dataset. The triple helix dynamics modeling module is used to construct a coupled dynamics model of biological state, behavioral choice and environmental factors based on triple helix resonance theory. It uses a heterogeneous graph neural network to express and calculate the complex relationships between the three dimensions. The co-resonance identification module is used to identify positive enhancement patterns between three-dimensional factors using a co-resonance detector, and to find resonance points that can produce extraordinary effects. The intervention strategy generation module is used to apply multi-objective reinforcement learning algorithms to generate personalized intervention strategies that can guide patients from their current state to the optimal resonance state. The closed-loop feedback optimization module is used to construct a closed-loop adaptive optimization mechanism based on feedback data from patients after implementing intervention strategies, continuously adjusting and improving the triple helix resonance framework and its parameters. The beneficial effects of this invention are as follows: The unity of precision and comprehensiveness: By constructing a three-dimensional health status mapping of biological-behavioral-environment space, the organic combination of precise monitoring and comprehensive management is achieved, so that health management can not only focus on the precise control of key physiological indicators, but also comprehensively consider the synergistic effects of behavioral patterns and environmental factors.

[0019] Synergistic effect: By identifying and optimizing synergistic resonance states, the system can discover positive reinforcing relationships among the three dimensions, generating a synergistic effect of "1+1+1>3". This significantly improves health management effectiveness, achieving better overall results while reducing the intensity of extreme interventions in a single dimension.

[0020] Dynamic adaptive capability: Through a closed-loop feedback mechanism, the system can continuously learn and optimize, adapting to the dynamic evolution of patients' health status, behavioral patterns and environmental changes, so that health management strategies always maintain the best effect.

[0021] Significantly reduced intervention complexity: Through the "triple helix resonance" framework, the system can identify key intervention points, obtain the greatest health benefits with the least intervention measures, integrate multiple fragmented interventions into a synergistic and consistent plan, reduce intervention complexity, and significantly improve patient compliance.

[0022] Enhanced long-term maintenance effects: The system optimizes the framework across multiple time scales, taking into account both short-term intervention effects and long-term behavioral development, thus creating a virtuous cycle between short-term intervention and long-term stability. This solves the problem of the disconnect between short-term effects and long-term maintenance in traditional health management.

[0023] Enhanced personalization: Based on big data and AI technology, the system can accurately capture the unique characteristics and needs of each patient, and provide highly personalized health management solutions according to a specific combination of individual biological status, behavioral patterns and environmental factors.

[0024] Multi-objective balance optimization: Through a multi-objective reinforcement learning framework, the system can simultaneously optimize multiple potentially conflicting objectives such as blood glucose control, quality of life, and adherence. It dynamically adjusts the weights of each objective based on the individual patient's preferences and needs to achieve optimal balance.

[0025] Maximizing data value: The system models the complex relationships between data of different dimensions through heterogeneous graph neural networks, fully explores the interaction relationships and potential value between data, realizes the deep utilization of big data, and significantly improves the efficiency of data use and value conversion rate. Attached Figure Description

[0026] Figure 1 This is a flowchart of the method of the present invention; Figure 2 It is a bar chart comparing the effects of single-dimensional optimization and three-dimensional collaborative optimization; Figure 3 It is a scatter plot of the co-resonance index distribution of patient state combinations; Figure 4 It is a radar chart showing the multi-objective reward weight configuration for different patient types; Figure 5 It is a line graph showing the changing trends of health indicators. Detailed Implementation

[0027] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0028] To aid in understanding the technical solution of this application, the terms used in this application will first be explained below: The triple helix resonance framework is a theoretical framework that models biological state, behavioral choice, and environmental factors as a coupled dynamic system, used to analyze and optimize the synergistic relationship among them.

[0029] Biological status: refers to the physiological indicators of diabetic patients, including but not limited to quantifiable physiological parameters such as blood glucose level, blood pressure, body mass index, and glycated hemoglobin.

[0030] Behavioral choices refer to the daily behavioral patterns of diabetic patients, including but not limited to observable behavioral data such as eating habits, exercise frequency, medication adherence, and sleep quality.

[0031] Environmental factors refer to external conditions that affect the health status of diabetic patients, including but not limited to situational factors such as social support networks, work stress, living environment, and accessibility of medical resources.

[0032] Synergistic resonance: refers to the phenomenon that when the three dimensions of biological state, behavioral choice and environmental factors reach a specific combination relationship, the overall performance of the system exceeds the sum of the independent optimization effects of each dimension.

[0033] Heterogeneous graph neural networks: a type of graph neural network capable of handling different types of nodes and edges, used to model the interactions between different types of entities in complex relational systems.

[0034] Multi-objective reinforcement learning: a reinforcement learning technique that simultaneously optimizes multiple potentially conflicting objectives by defining a composite reward function to achieve a balance between the multiple objectives.

[0035] The diabetes health management method based on AI big data proposed in this application is applicable to the following application scenarios: Personal health management platform: Patients install a diabetes health management application on their mobile devices (such as smartphones and wearable devices). This application integrates the triple helix resonance framework of this application and continuously collects patients' biological status data (such as blood glucose and blood pressure), behavioral data (such as dietary records and activity levels), and environmental data (such as location and social activities). Personalized health management recommendations are generated through cloud computing.

[0036] Diabetes Management Centers in Medical Institutions: Medical institutions deploy a diabetes management system based on the method described in this application, enabling healthcare professionals to remotely manage and intervene in patients' conditions. The system integrates patients' electronic health records, real-time monitoring data, and environmental information, providing healthcare professionals with a holistic view of patients' health and assisting in the development of more precise treatment and intervention plans.

[0037] Community health service stations: Community health service stations utilize the health management platform constructed using the method described in this application to manage diabetic patients within the community. The platform can identify the impact of community environmental factors on residents' health, and, in conjunction with individual patient characteristics, organize targeted health activities to form a community support network.

[0038] Insurance Company Health Management Program: The health management system developed by the insurance company based on the method described in this application provides personalized health management services to policyholders, continuously monitors and optimizes the policyholders' health status, reduces the risk of disease, and provides premium discounts based on participation and results.

[0039] Pharmaceutical Company Patient Management Project: Pharmaceutical companies utilize the method described in this application to build a patient management platform that provides comprehensive health management services to diabetic patients using their medications. By collecting real-world data, they can understand the actual effects of medications under different lifestyles and environmental conditions and optimize medication guidance.

[0040] In the aforementioned application scenarios, the method of this application can form a comprehensive understanding of the health status of diabetic patients by integrating data from three dimensions: biological state, behavioral choices, and environmental factors. This provides personalized and dynamically adaptive health management solutions, significantly improving management effectiveness and patients' quality of life.

[0041] Example A diabetes health management method based on AI and big data mainly includes the following steps: Step 1: Multi-dimensional data collection and preprocessing This step utilizes multimodal sensor networks and socio-physical fusion systems to collect data on patients' biological status, behavioral choices, and environmental factors. Through data cleaning, standardization, and feature extraction, a structured three-dimensional dataset is formed.

[0042] 1.1 Biological Status Data Acquisition: Medical-grade sensing devices such as continuous glucose monitors, smart blood pressure monitors, and smart scales are used to collect physiological indicators such as patients' blood glucose curves, blood pressure values, weight, and heart rate variability. The collected data undergoes time alignment, outlier detection, and noise filtering to extract features such as daily blood glucose fluctuation range, blood pressure trend, and weight change rate.

[0043] 1.2 Behavioral Selection Data Collection: Patient behavioral data, including dietary records (food type, intake, and eating time), exercise data (exercise type, intensity, and duration), sleep patterns, and medication records, were collected through smartphone applications, wearable devices, and electronic recording tools. Image recognition algorithms were used to automatically analyze food photos to extract nutritional components, and activity recognition algorithms were applied to classify exercise types. Combined with patients' manual records, a structured behavioral dataset was generated.

[0044] 1.3 Environmental Factor Data Collection: Information related to the patient's environment was obtained through location services, social media APIs, meteorological data interfaces, and patient questionnaires. This included geographical location, weather conditions, participation in social activities, work status and stress levels, and family and social support. The acquired environmental data was classified and quantified, transforming it into a computable environmental feature vector.

[0045] 1.4 Data Fusion and Preprocessing: The data from the three dimensions are aligned by timestamp, missing and outlier values ​​are handled, and feature selection algorithms are applied to filter key features, reducing dimensionality and increasing information density. A unified data representation format is constructed to provide high-quality input data for subsequent modeling.

[0046] Specific data preprocessing methods include: time series alignment using Dynamic Time Warping (DTW) to map data from different sampling frequencies to a unified time axis; missing value handling employing multiple interpolation and time-series-based interpolation algorithms, selecting the most suitable method based on data type and missing value pattern; outlier detection combining statistical methods (such as Z-scores and box plots) and machine learning algorithms (such as isolated forests and single-class SVMs) to identify and handle outliers; data standardization using min-max normalization or Z-score standardization to transform features at different scales to the same order of magnitude; and feature selection using algorithms such as L1-regularized Lasso regression, Recursive Feature Emission (RFE), and tree-based feature importance evaluation to select the most informative feature subsets. Simultaneously, dimensionality reduction techniques such as Principal Component Analysis (PCA) or t-SNE are employed to reduce the number of features while preserving data structure, thereby improving subsequent modeling efficiency.

[0047] It should be noted that in some embodiments, the data acquisition frequency can be dynamically adjusted according to the rate of change of different indicators. For example, blood glucose data can be collected every 5 minutes, while environmental data can be updated hourly or daily. Furthermore, data preprocessing may optionally include data augmentation techniques, such as synthesizing minority class samples or adding noise, to improve the model's robustness to anomalies.

[0048] Step 2: Modeling the Triple Helix Dynamics System This step utilizes the triple helix resonance theory to construct a coupled dynamic model of biological state, behavioral choice, and environmental factors. A heterogeneous graph neural network is used to express and calculate the complex relationships between the three dimensions, forming a mathematical description of the system's dynamic evolution.

[0049] 2.1 Construction of Dynamic Equations: Based on the collected multidimensional data, a system of differential equations describing the interaction of the three dimensions is constructed: in, Represents the biological state vector. This represents the behavior selection vector. Represents an environmental factor vector. , , These represent their rates of change over time. Functions , and The changing patterns of biological state, behavioral choices, and environmental factors are described respectively.

[0050] In the embodiments of this application, the dynamic changes of biological state, behavioral choice, and environmental factors are represented by a set of differential equations. First, the values ​​of the biological state vector, behavioral choice vector, and environmental factor vector at the current moment are obtained. Then, the rates of change of these three vectors over time are calculated respectively. The rate of change of biological state is a function of the biological state itself, behavioral choice, and environmental factors, reflecting how biological indicators change due to the combined influence of the biological state and the other two dimensions. Similarly, the rate of change of behavioral choice is also a function of the three vectors, representing how behavioral patterns are adjusted by biological state feedback and environmental conditions. The rate of change of environmental factors also depends on the current values ​​of the three dimensions, describing how environmental factors dynamically evolve with the patient's state and behavior. These three rate of change functions together constitute a coupled dynamic system capable of simulating the co-evolution of the three dimensions over time.

[0051] function , and The specific implementation involves parameterizing these functions using a neural network structure. First, feature extraction and nonlinear transformation are performed on the input vector. Then, complex interaction patterns are learned through a multi-layer fully connected network. For example, it includes an autoregressive term for the biological state itself, a term capturing behavioral influences, a term capturing environmental influences, and a term modeling the nonlinear interaction among the three. The weights of each term are optimized through backpropagation of time-series data, while incorporating domain knowledge constraints to ensure physiological plausibility. The function structure also includes a timescale parameter, enabling it to capture dynamic changes at different rates, such as rapid blood glucose fluctuations and slow habit formation.

[0052] 2.2 Construction of Heterogeneous Graph Neural Network: A heterogeneous graph containing three types of nodes (biological nodes, behavioral nodes, and environmental nodes) and multiple types of edges (representing interactions between different types of nodes) is constructed to represent complex relationships in the system. The graph is defined as follows: ,in This represents a set of nodes, containing three types of nodes; Represents the set of edges; It represents a set of relationship types, defining the interaction types between different nodes.

[0053] According to embodiments of this application, the heterogeneous graph neural network includes a node type mapping module, a relation-specific convolutional layer, and an attention fusion module. The node type mapping module first assigns dedicated feature transformation matrices to different types of nodes (biological nodes, behavioral nodes, and environmental nodes), mapping the original features to a unified hidden space. The relation-specific convolutional layer constructs specialized information transfer mechanisms for different types of node relationships (such as "behavior affects organisms," "environment affects behavior," etc.), achieving information exchange through weighted aggregation of neighboring nodes. The attention fusion module calculates the importance weights of different relationship types, achieving adaptive relationship importance assessment. This network structure can effectively capture asymmetric interaction patterns between the three types of nodes, learn time-dependent causal relationships, and adapt to individual differences among different patients, providing a foundation for subsequent dynamic system modeling.

[0054] 2.3 Graph Neural Network Layer Construction: Graph convolutional layers suitable for heterogeneous graphs are constructed, applying different transformation matrices to different types of nodes and edges to achieve efficient information transfer and aggregation. For nodes... Its representation vector The update process is as follows: in, Represents a node In the Layer representation vector, Indicates by relation type With nodes The set of connected neighbor nodes, and These are self-loop and relational types, respectively. The weight matrix, It is a non-linear activation function.

[0055] In the embodiments of this application, the update process of the node representation vector is as follows: First, for each node in the graph, consider all its neighboring nodes connected to it through different types of relationships; then, for each relationship type... This involves aggregating the representation vectors of all neighboring nodes under this relationship. Specifically, this is done by processing each neighboring node... Representation vector Apply the weight matrix specific to this relation type The transformation is performed, and then the average value is taken; next, the node's own representation vector is processed through the self-loop weight matrix. The transformation is performed; finally, the aggregated neighbor information of all relation types is added to the node's own information, and then a non-linear activation function is applied. Processing yields the updated representation vector of the node in the next layer. This process enables each node to acquire and integrate diverse information from its neighboring nodes based on different types of relationships, thereby learning richer feature representations.

[0056] According to embodiments of this application, the graph neural network layer adopts a multi-layer architecture, including an input layer, multiple heterogeneous graph convolutional layers, and an output layer. Each heterogeneous graph convolutional layer implements information transfer based on neighborhood aggregation, where the neighborhood aggregation operation considers the heterogeneity of node type and edge type. For each relation type r, it is first processed through a relation-specific weight matrix. The source node features are transformed; then, based on the type of the target node, an appropriate aggregation function (such as average aggregation, maximum aggregation, or weighted aggregation) is applied; finally, self-loop information is fused with neighbor information, and an updated node representation is generated using a non-linear activation function (such as ReLU or Tanh). Multi-layered stacked graph convolution operations enable the network to capture high-order interaction patterns, while residual connections and layer normalization techniques are used to improve training stability and model expressive power.

[0057] 2.4 Dynamic System Parameter Learning: A heterogeneous graphical neural network is trained using time-series data to learn the parameters and evolution laws in the dynamic equations. A time-series loss function is used to minimize the error between the model's predicted values ​​and the actual observed values, optimizing the model parameters so that they can accurately capture the dynamic evolution process of the system.

[0058] In addition, optionally, in some embodiments, an attention mechanism may be introduced to enhance the model’s ability to model long-term time-series dependencies, or variational inference methods may be used to handle uncertainty in the data and improve the model’s generalization ability.

[0059] Step 3: Co-resonance state identification and optimization This step utilizes a synergistic resonance detector to identify positive enhancement patterns between three-dimensional factors, discover and analyze the optimal combination of biological state, behavioral choices, and environmental factors, and identify resonance points that can produce extraordinary effects.

[0060] 3.1 State Space Construction: Based on the dynamic model established in step 2, a three-dimensional state space of biology, behavior, and environment is constructed. Each point in the space represents a specific system state configuration, namely a specific combination of biological state, behavioral choice, and environmental factors. By using state transition sequences from historical data, the patient's trajectory in the state space is plotted, identifying frequently visited areas and stable state areas.

[0061] 3.2 Synergistic Resonance Pattern Extraction: A synergistic resonance detection algorithm is constructed to identify, from historical trajectory data, situations where, when the three dimensions achieve a specific configuration relationship, the improvement in health indicators significantly exceeds the sum of the effects achievable by optimizing any single dimension individually. The synergistic resonance index is defined as follows: in, Indicates the synergistic resonance index. It is a health outcome assessment function. , , These represent the current biological state, behavioral choices, and environmental factors, respectively. , , Indicates the baseline status.

[0062] In the embodiments of this application, the synergistic resonance index The calculation method is as follows: First, considering the current biological state... Behavioral choices and environmental factors Combined, the overall health outcome scores produced are evaluated. Then, the health outcome scores were calculated for the three single-dimensional optimization scenarios, i.e., optimizing only the biological state while keeping the other two dimensions at the baseline level. Optimize only the behavioral choice while keeping the other two dimensions at the baseline level. Optimize only the environmental factors while keeping the other two dimensions at the baseline level. Next, the health outcome scores of these three single-dimensional optimizations are summed to obtain the sum of the effects of each individual optimization dimension; finally, the overall optimization effect is divided by the sum of the effects of each single dimension to obtain the synergistic resonance index. When the index is greater than 1, it indicates that the synergistic optimization of the three dimensions has produced an enhancement effect that goes beyond simple superposition. The higher the index, the stronger the resonance effect.

[0063] function The specific implementation involves constructing a weighted comprehensive score of multiple health indicators, encompassing three levels: short-term physiological indicators (such as blood glucose stability and blood pressure control), medium-term behavioral indicators (such as adherence and frequency of healthy behaviors), and long-term outcome indicators (such as quality of life and risk of complications). Each indicator is first normalized and mapped to the 0-1 range. Then, the weight coefficients are determined using indicator weights defined by medical experts or a data-driven importance learning algorithm. The function also includes a personalized adjustment mechanism that dynamically adjusts the relative importance of each indicator based on the patient's treatment stage, disease severity, and personal preferences, ensuring that the scoring results meet individual health needs.

[0064] The preprocessing methods for different types of data are as follows: Numerical indicators (such as blood glucose, blood pressure, and body mass index) are mapped to the 0-1 interval using min-max normalization or segmented normalization based on clinical thresholds; Categorical data (such as social support status and accessibility to medical resources) employs multiple coding strategies, including ordinal coding (for ordered categorical variables), one-hot coding (for unordered categorical variables), and target coding (mapping categorical variables to the historical average results of the corresponding health indicators), with all coding results undergoing a uniform scaling transformation; Frequency indicators (such as behavior execution frequency and social activity participation) are converted to saturation scores in the 0-1 interval using a sigmoid mapping function. For time-series indicators, their changing trends and stability are further considered, with the rate of change and fluctuation coefficient calculated using a sliding window, and these derived features are also incorporated into the scoring system. All preprocessed indicator combinations are weighted hierarchically to ensure that different types of indicators maintain an appropriate proportion in the overall score.

[0065] According to embodiments of this application, the synergistic resonance detector comprises three main functional modules: a baseline comparison module, a synergistic effect calculation module, and a threshold dynamic adjustment module. The baseline comparison module first constructs the patient's personalized baseline state. This is obtained through stable period characteristics in historical data or the average characteristics of demographically similar patients. The synergistic effect calculation module incorporates the actually observed health outcomes. The synergistic resonance index is calculated by comparing it with the sum of the expected results under the three single-dimensional optimization scenarios. Health outcome score The function comprehensively considers a weighted combination of multiple indicators such as glycemic stability, blood pressure control, and quality of life score, with weighting coefficients determined through sensitivity analysis. The threshold dynamic adjustment module dynamically sets values ​​based on historical patient data and population distribution characteristics. The system adjusts the judgment threshold to accommodate individual differences among patients. It also includes a pattern clustering component, which uses unsupervised learning algorithms (such as DBSCAN or hierarchical clustering) to cluster high-resolution patterns. Cluster analysis is performed on the state configuration of the values ​​to extract the common features of the resonance patterns.

[0066] Figure 2 The bar charts compare the health management effects of single-dimensional optimization and three-dimensional collaborative optimization, visually verifying the core claim of the triple helix resonance theory—that the effect of collaborative optimization exceeds the sum of the effects of single-dimensional optimization.

[0067] Figure 3The distribution of the co-resonance index under different combinations of patient states is presented in scatter plot form, verifying that the system can effectively identify high-resonance state points. The clustered regions in the scatter plot represent state combination patterns that may produce strong synergistic effects. These high CI value regions are the resonance points identified by the system through the co-resonance detection algorithm, providing target areas for resonance state optimization in step 3.4. The system will guide the patient state to transfer to these high-resonance regions to achieve optimization of the health state.

[0068] 3.3 Resonance Condition Modeling: Based on the extracted resonance patterns, the characteristics of the condition combinations leading to the resonance phenomenon are analyzed, and a resonance condition prediction model is constructed. Decision trees and rule extraction algorithms are applied to extract interpretable condition rules from state combinations with high co-resonance index values, forming a co-resonance knowledge base to provide a basis for subsequent intervention strategy generation.

[0069] 3.4 Resonance State Optimization: Using genetic algorithms and simulated annealing algorithms, the optimal resonance point is searched in the state space, i.e., the biological-behavioral-environment combination that produces the strongest synergistic effect under individual patient constraints. The optimization objective function is set as the synergistic resonance index, while considering state reachability and stability constraints to ensure that the recommended state combinations are both effective and feasible.

[0070] It should be noted that, in some embodiments, a multi-objective optimization method can be employed, simultaneously considering the intensity of the resonance effect and the complexity of state transitions, to generate personalized optimal resonance state transition paths for different patients. Alternatively, the system can also establish a classification system for resonance states, predefining several resonance patterns based on different clinical goals and patient characteristics to simplify the optimization process.

[0071] Step 4: Generation of intervention strategies for multi-objective reinforcement learning This step applies a multi-objective reinforcement learning algorithm, based on the triple helix dynamics model and synergistic resonance knowledge, to generate personalized intervention strategies that can guide patients from their current state to the optimal resonance state, thereby achieving multi-objective balance optimization.

[0072] 4.1 Construction of reinforcement learning environment: The dynamic model established in step 2 is used as the environment simulator for reinforcement learning. The state space is defined as a three-dimensional feature vector of organism-behavior-environment, the action space is a set of implementable intervention measures (such as dietary adjustment suggestions, exercise prescriptions, social activity arrangements, etc.), and the reward function is a weighted combination of the degree of improvement of health indicators and the synergistic resonance index.

[0073] 4.2 Construction of Multi-Objective Reward Function: Construct a composite reward function that includes multiple health management objectives: in, Represents the total reward function. and These represent the current state and the next state, respectively. Indicates the action taken. , , Let represent the sub-reward functions for blood glucose control, intervention adherence, and quality of life, respectively. Indicates the synergistic resonance index. arrive These are the weighting coefficients for each objective.

[0074] In the embodiments of this application, a multi-objective reward function The construction method is as follows: First, define corresponding sub-reward functions for different health management goals, including functions for evaluating the degree of improvement in blood glucose control. Functions for assessing intervention protocol adherence Functions for assessing quality of life And a function for evaluating the strength of the synergistic resonance effect. Then, weight coefficients are assigned to these sub-reward functions. arrive The first step is to reflect the relative importance of each objective. Next, the output of each sub-reward function is multiplied by its corresponding weighting coefficient. Finally, all weighted sub-reward results are summed to form a comprehensive reward value. The weighting coefficients can be dynamically adjusted based on individual patient preferences, clinical needs, and treatment stages to achieve a personalized balance of multiple objectives.

[0075] The specific implementations of each sub-reward function are as follows: The reward function for improved blood glucose control evaluates blood glucose indicators from the state... to state The degree of improvement is considered in terms of changes in multiple indicators, including average blood glucose, standard deviation, and the proportion of time within the target interval; the intervention compliance reward function is used to calculate the action. The degree to which the difficulty of implementation matches the patient's ability, as well as the personalization and ease of implementation of intervention recommendations; the quality of life reward function assesses quality of life indicators in the new state. The performance is assessed across dimensions including mental health, energy levels, and social activities; the synergistic resonance exponential function captures the strength of the synergistic effect of the three-dimensional factors. Weight coefficients are determined through a multi-level architecture: the bottom layer represents the minimum weight threshold set by medical safety constraints, the middle layer represents weight adjustments based on patient personal preferences, and the top layer represents weight fine-tuning by the meta-controller based on dynamic optimization of historical effects.

[0076] To ensure that the sub-reward functions can be reasonably combined in the multi-objective reward function, the outputs of all sub-reward functions are standardized. The specific standardization methods are as follows: Blood glucose control rewards use a clinically meaningful piecewise function mapping, mapping different degrees of blood glucose improvement to the [-1,1] interval, where positive values ​​represent improvement, negative values ​​represent deterioration, and zero values ​​represent no change; adherence rewards use a relative scoring mechanism based on the patient's historical adherence behavior, generating standardized scores in the [-0.5,1] range by comparing with the individual's baseline; quality of life rewards are based on validated quality of life assessment scales (such as SF-36 or WHOQOL-BREF), mapping the raw scores to the [0,1] interval through linear transformation; the synergistic resonance index, as a ratio, already has clear interpretive significance, and is calculated using the Sigmoid function. The output is transformed to fit within the range of [-0.5, 0.5], facilitating combination with other reward terms. Furthermore, all sub-reward functions include an adaptive normalization component, which dynamically adjusts the mapping parameters based on the characteristic distribution of different patient groups, ensuring the statistical stability of the reward distribution and thus supporting the effective convergence of the reinforcement learning algorithm.

[0077] According to embodiments of this application, the multi-objective reinforcement learning system adopts a value decomposition-based architecture, comprising three core components: an objective-specific value network, a value integration network, and a policy generation network. Each health management objective (such as blood glucose control, compliance, etc.) has an independent value network responsible for evaluating the value of state-action pairs under a specific objective. The value integration network receives the value evaluation results of each objective and performs weighted fusion according to the current weight configuration to generate a comprehensive value evaluation. The weight configuration is dynamically adjusted through a meta-policy controller, which adaptively adjusts the relative importance of each objective based on patient feedback, physician suggestions, and historical execution, while meeting clinical safety constraints. The policy generation network, based on the integrated value evaluation, generates a balanced multi-objective intervention strategy using a temperature-parameter-controlled soft Q-learning algorithm. The system also includes an experience replay buffer, which, through a priority sampling mechanism, focuses on learning experiences with high temporal difference errors or rare state transitions, improving learning efficiency.

[0078] Figure 4 A radar chart was used to illustrate the multi-objective reward weight configurations for different patient types, verifying that the system can personalize weight adjustments based on individual patient characteristics. This chart visually presents the multi-objective reward function... arrive The differences in the allocation of each weight coefficient across different patient types demonstrate the system's ability to dynamically adjust target weights based on patient characteristics (such as disease stage, personal preferences, and adherence history), thereby enabling personalized customization of intervention strategies and improving patient acceptance and adherence.

[0079] 4.3 Deep Reinforcement Learning Model Training: Based on the dynamic environment and reward function, the intervention policy network is trained using Deep Q-Network (DQN) or Proximal Policy Optimization (PPO) algorithms. Experience replay and target networks are employed to improve training stability, and priority sampling is used to focus on high-value transitions. The policy network structure includes a state encoding layer, an action value evaluation layer, and a multi-objective balancing module, outputting the optimal intervention action for the current state.

[0080] 4.4 Personalized Strategy Adaptation: The trained basic strategy model is fine-tuned in a personalized manner. Based on the patient's individual characteristics, feedback history, and compliance patterns, the reward weights and exploration parameters of the reinforcement learning model are adjusted. Thompson sampling technology is employed to strike a balance between exploration and utilization, gradually optimizing the individualized intervention strategy.

[0081] 4.5 Intervention Strategy Generation and Scheduling: Based on a reinforcement learning model, short-term, medium-term, and long-term intervention plans are generated, including daily behavior suggestions, environmental adjustment schemes, and phased goal setting. Through a multi-step planning algorithm, the optimal transition path from the current state to the target resonance state is constructed, and the path is decomposed into executable intervention sequences to form a structured health management plan.

[0082] In addition, optionally, in some embodiments, the system can introduce a human feedback reinforcement learning (RLHF) mechanism to integrate feedback from doctors and patients into the strategy optimization process, improving the clinical feasibility and patient acceptance of intervention strategies. Furthermore, a hierarchical reinforcement learning architecture can be employed to decompose complex health management tasks into multiple levels of sub-tasks, simplifying the strategy learning process.

[0083] Step 5: Closed-loop feedback and adaptive optimization This step, based on feedback data from patients after implementing intervention strategies, constructs a closed-loop adaptive optimization mechanism to continuously adjust and improve the triple helix resonance framework and its parameters, thereby achieving continuous optimization and evolution of the system.

[0084] 5.1 Multi-source feedback data collection: Multi-dimensional feedback data on intervention effectiveness is collected through real-time monitoring by smart devices, patient subjective evaluations, and medical records. This includes changes in objective biological indicators (such as blood glucose trends and blood pressure stability), behavioral performance (such as protocol adherence rate and evaluation of implementation difficulty), and subjective experience data (such as quality of life scores and satisfaction feedback).

[0085] 5.2 Intervention Effectiveness Evaluation: A comprehensive evaluation model was constructed to conduct multi-dimensional analysis of the effectiveness of the intervention strategy. The difference between the actual improvement in health indicators and the expected goals was calculated, factors influencing compliance were analyzed, the actual manifestation of the synergistic effect was assessed, and a structured effectiveness evaluation report was generated.

[0086] 5.3 Adaptive Adjustment of Model Parameters: Based on feedback data, key parameters in the triple helix dynamics model are automatically adjusted. A Bayesian optimization algorithm is applied to continuously update the posterior distribution of model parameters based on new data, enabling the model predictions to more accurately reflect individual characteristics and environmental changes. Simultaneously, the threshold parameters of the co-resonance detector are updated, allowing it to more precisely identify patient-specific resonance patterns.

[0087] 5.4 Incremental Knowledge Base Updates: Newly identified resonance patterns, effective intervention strategies, and feedback rules are added to the system's knowledge base to achieve continuous knowledge accumulation and optimization. An incremental learning algorithm is employed to integrate new knowledge while retaining existing knowledge, avoiding catastrophic forgetting and ensuring the knowledge base remains up-to-date.

[0088] 5.5 Dynamic Adjustment of Intervention Strategies: Based on the latest model parameters and knowledge base, intervention strategies are recalculated and optimized. Through online reinforcement learning algorithms, the content, intensity, and frequency of interventions are dynamically adjusted according to real-time patient feedback and status changes to ensure optimal effectiveness. A dynamic priority queue is constructed to rank intervention measures based on urgency and importance, optimizing resource allocation.

[0089] Figure 5 The changes in patients' health indicators before and after using the system are presented in the form of a line graph, verifying the long-term health management effectiveness of the system. The graph shows a trend of stable improvement in the health indicators of patients managed using this system over time. In particular, while fluctuations may occur in the initial stages of use, the system maintains a good state in the long term, demonstrating the value of closed-loop feedback and adaptive optimization mechanisms in ensuring long-term management effectiveness.

[0090] It should be noted that, in some embodiments, the system may employ an active learning strategy to collect feedback data under specific conditions, thereby improving the efficiency of model updates. Additionally, optionally, the system may construct a group knowledge-sharing mechanism to enable experience sharing and knowledge transfer among different patients while protecting privacy, thus accelerating the optimization process of individual models.

[0091] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for diabetes health management based on AI big data, characterized in that, Includes the following steps: Data on patients' biological status, behavioral choices, and environmental factors were collected and preprocessed to form a structured three-dimensional dataset. A coupled dynamic model of biological state, behavioral choice, and environmental factors is constructed based on the triple helix resonance theory, and the complex relationship between the three dimensions is expressed and calculated through a heterogeneous graph neural network. By using a synergistic resonance detector to identify positive enhancement patterns between three-dimensional factors, we can find resonance points that can produce extraordinary effects. By applying a multi-objective reinforcement learning algorithm, a personalized intervention strategy can be generated that guides the patient from the current state to the optimal resonance state. Based on feedback data from patients after implementing intervention strategies, a closed-loop adaptive optimization mechanism is constructed to continuously adjust and improve the triple helix resonance framework and its parameters. The heterogeneous graph neural network includes a node type mapping module, a relation-specific convolutional layer, and an attention fusion module, which are used to capture asymmetric interaction patterns between three types of nodes. 2.The AI big data-based diabetes health management method of claim 1, wherein, The biological status data includes physiological indicators such as blood glucose curves, blood pressure values, weight, and heart rate variability; the behavioral choice data includes behavioral data such as diet records, exercise data, sleep patterns, and medication records; and the environmental factor data includes geographical location, weather conditions, participation in social activities, work status and stress levels, and family and social support. 3.The AI big data-based diabetes health management method of claim 1, characterized in that, The steps of constructing a coupled dynamic model of biological state, behavioral choice, and environmental factors using the triple helix resonance theory include: Construct a system of differential equations describing the interaction of the three dimensions; Construct a heterogeneous graph containing three types of nodes and multiple types of edges to represent complex relationships in the system; Construct graph convolutional layers suitable for heterogeneous graphs, applying different transformation matrices to different types of nodes and edges; Heterogeneous graph neural networks are trained using time-series data to learn the parameters and evolutionary patterns in the dynamic equations.

4. The method for diabetes health management based on AI big data according to claim 1, characterized in that, The co-resonance detector identifies positive enhancement patterns among three-dimensional factors by calculating a co-resonance index. The co-resonance index is calculated by dividing the overall health outcome score generated by the combination of the current biological state, behavioral choices, and environmental factors by the sum of the health outcome scores under the three single-dimensional optimization conditions.

5. The method for diabetes health management based on AI big data according to claim 1, characterized in that, The multi-objective reinforcement learning algorithm includes: The dynamic model is used as an environment simulator for reinforcement learning; A multi-objective composite reward function was constructed, incorporating glycemic control, intervention adherence, quality of life, and synergistic effects; the intervention strategy network was trained using a deep Q-network or proximal strategy optimization algorithm. Personalized fine-tuning of the trained basic policy model; Generate short-term, medium-term, and long-term intervention plans.

6. The method for diabetes health management based on AI big data according to claim 5, characterized in that, The multi-objective composite reward function is as follows: ； in, For the reward function of improved blood glucose control, To intervene in compliance reward functions, For quality of life reward function, It is a cooperative resonance exponential function. to These are the weighting coefficients for each objective.

7. The method for diabetes health management based on AI big data according to claim 1, characterized in that, The closed-loop adaptive optimization mechanism includes: By collecting multidimensional feedback data on the intervention effect through real-time monitoring by smart devices, subjective evaluations by patients, and medical records, a comprehensive evaluation model is constructed to conduct multidimensional analysis of the effectiveness of the intervention strategy. Based on feedback data, key parameters in the triple helix dynamics model are automatically adjusted. Newly identified resonance patterns, effective intervention strategies, and feedback patterns are added to the system's knowledge base; Based on the latest model parameters and knowledge base, the intervention strategy is recalculated and optimized.

8. The method for diabetes health management based on AI big data according to claim 7, characterized in that, The key parameters in the automatically adjusted triple helix dynamics model are optimized using a Bayesian optimization algorithm, which continuously updates the posterior distribution of the model parameters based on new data.

9. The method for diabetes health management based on AI big data according to claim 1, characterized in that, The graph convolutional layer of the heterogeneous graph neural network pairs nodes. The representation vector The update process is as follows: ； in, Let v be the representation vector of node v at the (l+1)th layer. Let u be the representation vector of node u at layer l. To pass through relationships With nodes The set of connected neighbor nodes, The size of the set, Let r be the weight matrix of relation r at level l. Let be the weight matrix of the self-loop at layer l. This is the activation function.

10. A diabetes health management system based on AI big data, used to execute the method according to any one of claims 1-9, characterized in that, include: The multi-dimensional data acquisition module is used to collect data on three dimensions: the patient's biological state, behavioral choices, and environmental factors, and to preprocess the data to form a structured three-dimensional dataset. The triple helix dynamics modeling module is used to construct a coupled dynamics model of biological state, behavioral choice and environmental factors based on triple helix resonance theory. It uses a heterogeneous graph neural network to express and calculate the complex relationships between the three dimensions. The co-resonance identification module is used to identify positive enhancement patterns between three-dimensional factors using a co-resonance detector, and to find resonance points that can produce extraordinary effects. The intervention strategy generation module is used to apply multi-objective reinforcement learning algorithms to generate personalized intervention strategies that can guide patients from their current state to the optimal resonance state. The closed-loop feedback optimization module is used to construct a closed-loop adaptive optimization mechanism based on feedback data after patients implement intervention strategies, and to continuously adjust and improve the triple helix resonance framework and its parameters.

Images