Intelligent health system and data processing method thereof

Abstract

This invention relates to the field of data processing and analysis technology, and in particular to a smart health system and its data processing method. The system includes: real-time acquisition of data from multiple heterogeneous data sources to comprehensively capture multi-dimensional health data of users; dynamic construction and optimization of a multimodal health knowledge graph through efficient data preprocessing and feature alignment techniques to enhance the integration value and utilization efficiency of the data; accurate assessment of users' health status and timely prediction of future risks using graph neural networks and time-series analysis models to effectively improve the predictability and targeting of health management; generation of highly personalized intervention sequences based on individual user circumstances using reinforcement learning algorithms to enhance the accuracy of health interventions and promote positive changes in users' health behaviors; and execution of intervention commands through intelligent devices and real-time monitoring of user feedback to dynamically adjust intervention strategies, ensuring the flexibility and adaptability of intervention measures.

Description

Technical Field

[0001] This invention relates to the field of data processing and analysis technology, and in particular to a smart health system and its data processing method. Background Technology

[0002] Despite the rapid development of smart health management technology, traditional health management methods still face numerous challenges. The following problems persist in the current health management field: In traditional methods, user physiological parameters, behavioral activities, environmental perceptions, and medical records are often stored independently by different systems or devices, forming data silos and hindering data integration and comprehensive utilization; traditional health assessment methods rely heavily on manual examinations and user self-reporting, which are easily influenced by subjective factors and fail to comprehensively capture users' real-time health status, resulting in subjectivity and lag in assessments and potential biases; traditional health risk prediction models are typically based on limited historical data and static indicators, making it difficult to dynamically reflect real-time changes in user health status and predict potential health risks in a timely manner; traditional health intervention strategies often employ generic solutions, ignoring individual differences among users and dynamic changes in health status, failing to effectively meet specific user needs, impacting intervention effectiveness, and limiting the continuous optimization and improvement of health management services. Therefore, this invention proposes a smart big health system and its data processing method. Summary of the Invention

[0003] The purpose of this invention is to solve the problems in the background art by proposing a smart health system and its data processing method.

[0004] To achieve the above objectives, the present invention adopts the following technical solution:

[0005] A smart health system includes: a multi-source data acquisition module, a data fusion and knowledge graph construction module, a health status assessment and risk prediction module, a personalized intervention strategy generation module, and an intervention execution and feedback adaptation module;

[0006] The multi-source data acquisition module is used to collect users' physiological parameter data, behavioral activity data, environmental perception data, and medical record data from multiple heterogeneous data sources in real time or near real time.

[0007] The data fusion and knowledge graph construction module is used to preprocess, extract and align features from the collected multi-source data, and build and dynamically update a multimodal health knowledge graph based on this. Nodes represent entities including users, physiological indicators, symptoms, behaviors, environmental factors and diseases, while edges represent the semantic and statistical relationship strength between entities.

[0008] The health status assessment and risk prediction module is used to generate real-time health status assessment results and short-term health risk prediction results for individual users based on a multimodal health knowledge graph, using a graph neural network model to learn the embedded representation of nodes and relationships, and combining it with a time series analysis model.

[0009] The personalized intervention strategy generation module is used to infer and generate a personalized intervention sequence for the target user, which includes a series of ordered intervention actions, based on the health status assessment results and health risk prediction results, through a dynamic strategy optimization algorithm based on reinforcement learning.

[0010] The intervention execution and feedback adaptation module is used to convert personalized intervention sequences into execution instructions that match the interface protocol of smart devices, drive the connected smart devices to execute the intervention, and monitor the target user's response data to the intervention in real time. Based on the response data, it dynamically adjusts the relationship strength between the intervention strategy and the multimodal health knowledge graph.

[0011] Furthermore, the data fusion and knowledge graph construction module preprocesses, extracts, and aligns the collected multi-source data, and constructs and dynamically updates a multimodal health knowledge graph. This process includes:

[0012] Preprocessing of multi-source data includes data cleaning, noise reduction, outlier detection and imputation, and format standardization.

[0013] Key features are extracted from the preprocessed data, including time and frequency domain features of physiological parameters, pattern features of behavioral activities, periodic features of environmental data, and structured features of medical records.

[0014] The extracted key features are aligned under a unified timestamp and mapped to a preset entity space;

[0015] Based on the aligned feature data, entity linking and relation extraction techniques are used to construct a multimodal health knowledge graph with users as the core nodes. The node types include user nodes, physiological indicator nodes, symptom nodes, behavior nodes, environment nodes, and disease nodes. The edge relation types include user-have-indicator, indicator-may cause-symptom, behavior-influence-indicator, environment-exacerbate-risk, and disease-associate-symptom.

[0016] A graph database is used to store and index a multimodal health knowledge graph. An incremental update algorithm is used to dynamically update the node attributes, edge relationship strength, and newly added entity relationships in the multimodal health knowledge graph.

[0017] Furthermore, the health status assessment and risk prediction module, based on a multimodal health knowledge graph, uses a graph neural network model to learn the embedded representations of nodes and relationships, and combines this with a time series analysis model to generate real-time health status assessment results and short-term health risk prediction results for individual users. The process includes:

[0018] Based on a multimodal health knowledge graph, a graph embedding algorithm is used to learn the low-dimensional vector representation of each node and relationship in the graph, and obtain the graph embedding vector.

[0019] By combining the time series of user historical data, time series analysis is performed on the key features extracted from the preprocessed data. The time series prediction model is used to analyze the changing trends of each key feature and obtain the time series trend features of the key features in the future preset time period.

[0020] The graph embedding vector is fused with time-series trend features and input into a preset health score prediction model to calculate the user's current health status score.

[0021] Based on the extrapolation results of the association paths and temporal trend features in the multimodal health knowledge graph, the probability of a user's health risk event occurring within a preset time period is predicted, and key risk influencing factors are identified.

[0022] Furthermore, the process by which the target user filtering module identifies target users from all users includes:

[0023] Based on health status scores and health risk prediction probabilities, a comprehensive health risk value is calculated for each user; the comprehensive health risk value is compared with a preset comprehensive health risk threshold; users whose comprehensive health risk values ​​exceed the comprehensive health risk threshold are identified as target users who require priority intervention.

[0024] Furthermore, the personalized intervention strategy generation module, based on the health status assessment results and health risk prediction results, uses a reinforcement learning-based dynamic strategy optimization algorithm to infer and generate a personalized intervention sequence for the target user, containing a series of ordered intervention actions. This process includes:

[0025] The user's health status assessment results are integrated with the health risk prediction results to construct the state of the reinforcement learning agent.

[0026] The action space of an agent is defined as the various executable intervention methods and their combinations.

[0027] Construct a reward function that uses the expectation of health improvement as a positive reward and the probability of risk occurrence and user discomfort as negative rewards to guide the learning process of the agent;

[0028] In a simulated environment consisting of a multimodal health knowledge graph and user historical data, an agent is trained to learn a strategy that maps the optimal sequence of actions, i.e., a personalized intervention sequence, based on the current state.

[0029] Furthermore, the intervention execution and feedback adaptation module converts personalized intervention sequences into execution instructions that match the interface protocol of smart devices, driving the connected smart devices to execute the intervention. It also monitors the target user's response data to the intervention in real time, and dynamically adjusts the relationship strength between the intervention strategy and the multimodal health knowledge graph based on the response data. This process includes:

[0030] Collect data on changes in users' physiological parameters and behavioral activities after intervention as response data;

[0031] Calculate the deviation between the response data and the expected improvement target to evaluate the effectiveness of the intervention strategy;

[0032] Based on the evaluation results, the parameters of the decision algorithm in the personalized intervention strategy generation module were adjusted using the Bayesian optimization method.

[0033] Based on the intervention effect and user response, update the relationship strength between relevant entities in the multimodal health knowledge graph.

[0034] Furthermore, the process of extracting key features from the preprocessed data includes:

[0035] For physiological parameter data, extract its time-domain features, including mean, variance, maximum and minimum values, and extract its frequency-domain features through Fourier transform;

[0036] For behavioral activity data, clustering algorithms are used to identify behavioral patterns, and the frequency and duration of different patterns are extracted as pattern features.

[0037] For environmental data, daily and weekly periodic characteristics are extracted as periodic features through periodic graph analysis.

[0038] For medical record data, natural language processing technology is used to extract its structured entities and relationships as structured features.

[0039] Furthermore, based on the extrapolation results of the association paths and temporal trend characteristics in the multimodal health knowledge graph, the process of predicting the probability of a user experiencing a target health risk event within a preset time period and identifying key risk influencing factors includes:

[0040] In a multimodal health knowledge graph, user nodes corresponding to target users are located, and one or more physiological indicator nodes that are currently abnormal and mapped by key features of the target user are associated with them through user-have-indicator relationships, serving as the starting set of paths; at the same time, disease nodes corresponding to predicted health risk events are located, serving as the target nodes of paths; and all associated paths from any physiological indicator node in the starting set of paths to the target disease node of the path are extracted.

[0041] Calculate the weight of each associated path, where the path weight is determined by the product of the relation strength of all edges on the path and the risk coefficient of the nodes traversed by the path;

[0042] For each associated path, extract one or more physiological indicator nodes on the path; directly obtain the time-series trend features generated by the time-series prediction model for the key features corresponding to these physiological indicator nodes; integrate the weight of the associated path with the corresponding future change values ​​in the time-series trend features, and calculate the final health risk prediction probability through the probability prediction model.

[0043] Among all associated paths, the top-N critical paths that contribute the most to the probability of the final health risk prediction are selected; the nodes connected by the edges with the strongest relationship on these critical paths are extracted, and the entities corresponding to the nodes are the key risk factors. The key risk factors are used to prioritize the treatment in the generated personalized intervention sequence.

[0044] Furthermore, the process of training an agent to learn a policy includes:

[0045] The agent selects an action based on its current state and the agent's policy network.

[0046] The simulation environment updates to a new state based on the selected action and calculates an immediate reward according to the reward function;

[0047] The agent uses the policy gradient method to update the agent's policy network parameters based on the state, action, and reward sequence.

[0048] After multiple rounds of iterative training, a stable and optimized strategy is finally obtained. Applying this strategy to the user's current health status will output a corresponding personalized intervention sequence.

[0049] A data processing method for a smart healthcare system includes:

[0050] S1. Collect users' physiological parameter data, behavioral activity data, environmental perception data, and medical record data from multiple heterogeneous data sources in real time or near real time;

[0051] S2. Preprocess, extract features and align the collected multi-source data in a spatiotemporal manner, and build and dynamically update a multimodal health knowledge graph based on the processed data;

[0052] S3. Based on a multimodal health knowledge graph, a graph neural network model is used to learn the embedded representation of nodes and relationships, and combined with a time series analysis model, to generate real-time health status assessment results and short-term health risk prediction results for individual users.

[0053] S4. Based on the health status assessment results and health risk prediction results, a personalized intervention sequence for the target user is inferred and generated through a dynamic policy optimization algorithm based on reinforcement learning.

[0054] S5. Convert the personalized intervention sequence into execution instructions that match the interface protocol of the smart device, drive the connected smart device to perform the intervention, and monitor the target user's response data to the intervention in real time.

[0055] S6. Based on user response data, dynamically adjust the relationship strength between intervention strategies and the multimodal health knowledge graph.

[0056] Compared with existing technologies, the beneficial effects of this invention are as follows: By collecting data in real time from multiple heterogeneous data sources such as wearable devices, smart mobile terminals, environmental sensors, and hospital information systems, the comprehensiveness and real-time nature of health data are ensured; by preprocessing, extracting features, and aligning the collected multi-source data, a multimodal health knowledge graph is dynamically constructed and updated, effectively integrating scattered data resources, improving data utilization efficiency, and providing data support for accurate health assessment; based on the multimodal health knowledge graph, graph neural networks and time series analysis models are used to achieve real-time assessment of users' health status and accurate prediction of future health risks, so as to detect health problems in advance and take preventive measures; based on the health assessment results, personalized intervention sequences are generated through reinforcement learning algorithms to improve the pertinence and effectiveness of interventions and promote positive changes in users' health behaviors; by transforming intervention strategies into execution instructions, driving smart devices to execute, and monitoring user responses in real time, the intervention strategies are dynamically adjusted to ensure the effective implementation and timely adjustment of intervention measures, thereby improving the overall effect of health management. Attached Figure Description

[0057] Figure 1 This is a block diagram of a smart health system proposed in this invention.

[0058] Figure 2 This is a flowchart of a data processing method for a smart health system proposed in this invention. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Reference Figure 1 A smart health system, comprising a multi-source data acquisition module, a data fusion and knowledge graph construction module, a health status assessment and risk prediction module, a personalized intervention strategy generation module, and an intervention execution and feedback adaptation module;

[0061] The multi-source data acquisition module is used to collect users' physiological parameter data, behavioral activity data, environmental perception data, and medical record data from multiple heterogeneous data sources in real time or near real time. The heterogeneous data sources in the multi-source data acquisition module include wearable devices, smart mobile terminals, environmental sensors, and hospital information systems. Physiological parameter data is collected through wearable devices and includes at least three of the following: heart rate, blood pressure, blood glucose, blood oxygen saturation, body temperature, and electrocardiogram waveform. Behavioral activity data is collected through smart mobile terminals and includes sleep duration and quality, steps, exercise type and intensity, sedentary time, and calorie consumption. Environmental perception data is collected through environmental sensors and includes ambient temperature, humidity, air quality index, noise level, and light intensity. Medical record data is obtained through the hospital information system interface and includes diagnostic results, medication records, allergy history, and family medical history from electronic medical records.

[0062] The data fusion and knowledge graph construction module is used to preprocess, extract and align features from the collected multi-source data, and build and dynamically update a multimodal health knowledge graph based on this. Nodes represent entities including users, physiological indicators, symptoms, behaviors, environmental factors and diseases, while edges represent the semantic and statistical relationship strength between entities.

[0063] The health status assessment and risk prediction module is used to generate real-time health status assessment results and short-term health risk prediction results for individual users based on a multimodal health knowledge graph, using a graph neural network model to learn the embedded representation of nodes and relationships, and combining it with a time series analysis model.

[0064] The personalized intervention strategy generation module is used to infer and generate a personalized intervention sequence for the target user, which includes a series of ordered intervention actions, based on the health status assessment results and health risk prediction results, through a dynamic strategy optimization algorithm based on reinforcement learning.

[0065] The intervention execution and feedback adaptation module is used to convert personalized intervention sequences into execution instructions that match the interface protocol of smart devices, drive the connected smart devices to execute the intervention, and monitor the target user's response data to the intervention in real time. Based on the response data, it dynamically adjusts the relationship strength between the intervention strategy and the multimodal health knowledge graph.

[0066] It should be further explained that, in the specific implementation process, the data fusion and knowledge graph construction module preprocesses, extracts and aligns features from the collected multi-source data, and constructs and dynamically updates a multimodal health knowledge graph. This process includes:

[0067] Preprocessing of multi-source data includes data cleaning, noise reduction, outlier detection and imputation, and format standardization.

[0068] Key features are extracted from the preprocessed data, including time-domain and frequency-domain features of physiological parameters, pattern features of behavioral activities, periodic features of environmental data, and structured features of medical records. Specifically, for physiological parameter data, time-domain features such as mean, variance, maximum, and minimum values ​​are extracted, and frequency-domain features are extracted using Fourier transform. For behavioral activity data, clustering algorithms are used to identify behavioral patterns, and the frequency and duration of different patterns are extracted as pattern features. For environmental data, daily and weekly periodic features are extracted using periodic graph analysis. For medical record data, structured entities and relationships are extracted using natural language processing techniques as structured features.

[0069] The extracted key features are aligned under a unified timestamp and mapped to a preset entity space;

[0070] Based on the aligned feature data, entity linking and relation extraction techniques are used to construct a multimodal health knowledge graph with users as the core nodes. The node types include user nodes, physiological indicator nodes, symptom nodes, behavior nodes, environment nodes, and disease nodes. The edge relation types include user-have-indicator, indicator-may cause-symptom, behavior-influence-indicator, environment-exacerbate-risk, and disease-associate-symptom.

[0071] A graph database is used to store and index a multimodal health knowledge graph. As new data continuously flows in, an incremental update algorithm is employed to dynamically update node attributes, edge relationship strength, and newly added entity relationships within the multimodal health knowledge graph. Specifically, the incremental update algorithm is an efficient algorithm for processing streaming data. Its core principle is to update only the affected parts of the multimodal health knowledge graph as new data continuously flows in, thereby reducing computational overhead and improving real-time performance. The specific process includes:

[0072] A1. When new data flows in through the multi-source data acquisition module, the system compares the differences between the new data and the existing data in the multimodal health knowledge graph, and identifies newly added entities, updated node attributes, or changed relationships. For example, newly acquired heart rate data may deviate from existing data, triggering node attribute updates.

[0073] A2. In response to detected changes, update the relevant nodes and edges: For node attributes, dynamically adjust attribute values ​​using methods such as sliding window averaging or exponentially weighted moving average; for edge relationship strength, recalculate the strength value based on statistical correlation (such as Pearson correlation coefficient) or semantic similarity in the new data; for newly added entity relationships, embed the new entities into the graph using entity linking technology and establish edge connections.

[0074] A3. After the update, the graph structure is re-optimized using the indexing mechanism of a graph database (such as Neo4j) to ensure query efficiency. At the same time, consistency checks are performed regularly to avoid loops or contradictions in the graph caused by incremental updates. The incremental update algorithm ensures that the multimodal health knowledge graph always reflects the latest health status, providing an accurate data foundation for downstream modules.

[0075] It should be further explained that, in the specific implementation process, the health status assessment and risk prediction module, based on a multimodal health knowledge graph, uses a graph neural network model to learn the embedded representations of nodes and relationships, and combines this with a time series analysis model to generate real-time health status assessment results and short-term health risk prediction results for individual users. The process includes:

[0076] Based on a multimodal health knowledge graph, a graph embedding algorithm is used to learn the low-dimensional vector representation of each node and relationship in the graph, and obtain the graph embedding vector.

[0077] Specifically, B1 is a graph structure based on a multimodal health knowledge graph, where node types include user nodes, physiological indicator nodes, symptom nodes, behavior nodes, environment nodes, and disease nodes, and edge relationship types include user-possess-indicator, indicator-may-cause-symptom, behavior-influence-indicator, environment-exacerbate-risk, and disease-associate-symptom. Each node is associated with key features extracted from preprocessed data, such as: physiological indicator nodes are associated with time-domain features (e.g., mean, variance) and frequency-domain features (e.g., Fourier transform coefficients); behavior nodes are associated with pattern features (e.g., frequency of behavior patterns generated by clustering); environment nodes are associated with periodic features (e.g., daily cycle, weekly cycle); and medical record nodes are associated with structured features (e.g., entities and relationships). These features serve as the initial feature vectors for nodes, used for subsequent embedding learning.

[0078] B2. A graph neural network model (such as GraphSAGE or Graph Attention Network GAT) is used as the core of the graph embedding algorithm. This model contains a multi-layer message passing mechanism. In each layer, each node updates its own representation by aggregating information from its neighboring nodes. For relation embedding, the edge relation type is used as an important parameter in the message passing process. Relation-specific weight matrices (e.g., in R-GCN) are used to distinguish the influence of different relation types, thereby learning low-dimensional vector representations of nodes and relations simultaneously.

[0079] B3. Train the graph neural network model by defining a graph embedding training objective function, which includes node classification loss and link prediction loss. The node classification loss ensures that similar nodes (such as nodes with the same disease type) are close in the embedding space, and the link prediction loss ensures that node pairs with strong relationships (such as user-have-indicator) have high inner product scores. During training, the gradient descent algorithm (such as the Adam optimizer) is used to minimize the loss function, and an incremental update mechanism is combined to dynamically adjust the embedding representation as new data flows in to reflect the real-time changes of the knowledge graph.

[0080] B4. After training, output low-dimensional vector representations of all nodes and relationships (i.e., graph embedding vectors). These vectors capture the semantic relationships and statistical association strengths between entities in the graph. The graph embedding vectors are then used in the health status assessment module, fused with time-series trend features, and input into the health score prediction model to generate user health status scores and health risk prediction results.

[0081] By combining the time series of user historical data, time series analysis is performed on the key features extracted from the preprocessed data. The time series prediction model is used to analyze the changing trends of each key feature and obtain the time series trend features of the key features in the future preset time period.

[0082] The graph embedding vector is fused with time-series trend features and input into a preset health score prediction model to calculate the user's current health status score; wherein the health score prediction model is one of a linear regression model, a gradient boosting decision tree model, or a fully connected neural network model;

[0083] Based on the extrapolation results of the association paths and temporal trend features in the multimodal health knowledge graph, the probability of a user's health risk event occurring within a preset time period is predicted, and key risk influencing factors are identified. Specifically, in the multimodal health knowledge graph, user nodes corresponding to the target user are located, and one or more physiological indicator nodes mapped by key features of the target user are associated with the user-possess-indicator relationship, serving as the starting set of paths. At the same time, disease nodes corresponding to the predicted health risk event are located, serving as the target nodes of paths. All association paths from any physiological indicator node in the starting set of paths to the target disease node of the path are extracted.

[0084] Calculate the weight of each associated path, where the path weight is determined by the product of the relation strength of all edges on the path and the risk coefficient of the nodes traversed by the path;

[0085] For each associated path, extract one or more physiological indicator nodes on the path; directly obtain the time-series trend features generated by the time-series prediction model for the key features corresponding to these physiological indicator nodes; integrate the weight of the associated path with the corresponding future change value in the time-series trend feature, and calculate the final health risk prediction probability through a probability prediction model (e.g., logistic regression model); where the future change value is the specific value of the key feature predicted by the time-series trend feature at a specific point in the future.

[0086] Among all associated paths, the Top-N key paths that contribute the most to the final health risk prediction probability are selected. The nodes connected by the edges with the strongest relationships on these key paths are extracted, and the entities corresponding to these nodes are the key risk influencing factors. These key risk influencing factors are prioritized in the generated personalized intervention sequence. Extrapolation results refer to using time-series analysis models to predict trends in historical key feature data, thereby inferring changes in feature values ​​over future time periods. Specifically, time-series trend features are generated by analyzing preprocessed key features (such as heart rate and steps) using time-series prediction models (such as ARIMA, LSTM, or Prophet models) to generate a sequence of predicted values ​​for a preset future time period (e.g., the next 24 hours). Extrapolation results include point predictions (specific values) and interval predictions (confidence intervals) used to assess changes in health risk. In risk prediction, the extrapolation results are combined with the associated paths of the multimodal health knowledge graph, and the risk probability is quantified by calculating path weights and trend changes.

[0087] It should be further explained that, in the specific implementation process, the target user screening module's process for identifying target users from all users includes:

[0088] Based on health status scores and health risk prediction probabilities, a comprehensive health risk value is calculated for each user. This comprehensive health risk value integrates the health status score and the health risk prediction probability into a single indicator, facilitating user filtering and prioritization. The calculation process is as follows: Define the health status score HS as a value from 0 to 100, based on the output of the health score prediction model; a lower value indicates a worse health status. Define the health risk prediction probability RP as a value from 0 to 1, representing the probability of a target health risk event (such as a heart attack) occurring in the future. Normalize the health status score HS and the health risk prediction probability RP; for example, convert the health status score HS into a relative score (HS' = 1 - HS / 1). 00), so that the higher the normalized HS' value, the worse the condition; the comprehensive health risk value HR is calculated using a linear weighted formula: HR=α1×HS'+α2×RP, where α1 and α2 are weighting coefficients; the comprehensive health risk value is compared with the preset comprehensive health risk threshold; users whose comprehensive health risk value exceeds the comprehensive health risk threshold are identified as target users who need priority intervention; it can be understood that the health status assessment and risk prediction module performs health status assessment and health risk prediction for all users, generating real-time health status assessment results and short-term health risk prediction results for individual users. Here, individual users refer to any individual being assessed in the system, which is all users before screening.

[0089] It should be further explained that, in the specific implementation process, the personalized intervention strategy generation module, based on the health status assessment results and health risk prediction results, uses a reinforcement learning-based dynamic strategy optimization algorithm to infer and generate a personalized intervention sequence for the target user, containing a series of ordered intervention actions. This process includes:

[0090] The user's health status assessment results are integrated with the health risk prediction results to construct the state of the reinforcement learning agent.

[0091] The action space of an agent is defined as the various executable intervention methods and their combinations.

[0092] Construct a reward function that uses the expectation of health improvement as a positive reward and the probability of risk occurrence and user discomfort as negative rewards to guide the learning process of the agent;

[0093] In a simulated environment comprised of a multimodal health knowledge graph and user historical data, an agent is trained to learn a policy that maps the optimal sequence of actions—i.e., a personalized intervention sequence—based on the current state. The process of training the agent to learn this policy includes:

[0094] The agent selects an action based on its current state and the agent's policy network.

[0095] The simulated environment updates to a new state based on the selected action and calculates an immediate reward according to the reward function. In reinforcement learning training, the specific process of updating the simulated environment and calculating the reward includes: the current state... Including user health status score Health risk prediction probability And graph embedding vector ,Right now ( Indicates the current time step. Indicates at time step The calculated health status score, Indicates at time step The probability of predicting health risks Indicates at time step (Graph embedding vectors); agent selects actions. (e.g., sending health reminders or adjusting medication dosages), the simulated environment is action-based. And the built-in user model, update the state to The user model is built using historical data to simulate the user's real response to interventions; for example, after increasing exercise, heart rate may decrease and health scores may improve; reward function for: In the formula, This indicates the change in health status score from the current state to the next state; a positive value (health improvement) provides a positive reward. This represents the reduction in the probability of health risks; a positive value (reduced risk) provides a positive reward. Indicates the performance of an action The user discomfort or implementation cost caused by the intervention is quantified through user feedback questionnaires or a pre-set cost table (for example, the cost of high-intensity intervention is higher than that of reminder-type intervention), and this item is treated as a negative reward. The predefined weighting coefficients are used to balance the relative importance of health improvement, risk reduction, and user comfort. This process is repeated in a simulation environment, where the agent learns to maximize the cumulative reward through the policy gradient method and finally outputs an optimized policy.

[0096] The agent uses the policy gradient method to update the agent's policy network parameters based on the state, action, and reward sequence.

[0097] After multiple rounds of iterative training, a stable and optimized strategy is finally obtained. Applying this strategy to the user's current health status will output a corresponding personalized intervention sequence.

[0098] It should be further explained that, in the specific implementation process, the intervention execution and feedback adaptation module converts the personalized intervention sequence into execution instructions that match the interface protocol of the smart device, drives the connected smart device to execute the intervention, and monitors the target user's response data to the intervention in real time. The process of dynamically adjusting the relationship strength between the intervention strategy and the multimodal health knowledge graph based on the response data includes:

[0099] Collect data on changes in users' physiological parameters and behavioral activities after intervention as response data;

[0100] Based on the expected direction of health status improvement according to the personalized intervention sequence, specific and quantifiable expected improvement goals are set; the deviation between the response data and the expected improvement goals is calculated to evaluate the effectiveness of the intervention strategy;

[0101] Based on the evaluation results, the parameters of the decision-making algorithm in the personalized intervention strategy generation module were adjusted using a Bayesian optimization method. Specifically, the Bayesian optimization method was used to adjust the parameters (such as learning rate and discount factor) of the reinforcement learning algorithm in the personalized intervention strategy generation module to improve the intervention effect. The specific process included:

[0102] F1. Define the set of parameters to be optimized, mainly including the weight coefficients of each item in the reward function of the reinforcement learning agent, such as the weights of parameters like expected health improvement, probability of risk occurrence, and user discomfort; construct a Bayesian optimization objective function with the intervention effect as the measurement standard, where the function is calculated based on the deviation between user response data (such as actual changes in physiological parameters and behavioral activities) and the preset quantitative expected improvement target.

[0103] F2. Bayesian Optimization Iteration: In the initialization phase, a set of initial parameter values ​​is randomly selected from the parameter space; a Gaussian process is used as a surrogate model to probabilistically model the objective function; acquisition functions such as expectation improvement are used to balance exploration and utilization, and the next set of parameters to be evaluated is selected; in each iteration, the decision algorithm is run using the current parameters to generate and execute interventions, user response data is collected to calculate the objective function value, and the surrogate model is updated accordingly; when the objective function value converges or reaches the maximum number of iterations, the optimization is terminated, and the optimal parameter combination found is output; finally, the optimal parameters obtained by optimization are applied to the decision algorithm (i.e., reinforcement learning agent) in the personalized intervention strategy generation module to generate more effective subsequent intervention strategies;

[0104] Based on the intervention effects and user responses, update the relationship strength between relevant entities in the multimodal health knowledge graph, strengthen the association between effective interventions and positive outcomes, and weaken the association between ineffective interventions.

[0105] Reference Figure 2 A data processing method for a smart health system, comprising:

[0106] S1. Collect users' physiological parameter data, behavioral activity data, environmental perception data, and medical record data from multiple heterogeneous data sources in real time or near real time;

[0107] S2. Preprocess, extract features and align the collected multi-source data in a spatiotemporal manner, and build and dynamically update a multimodal health knowledge graph based on the processed data;

[0108] S3. Based on a multimodal health knowledge graph, a graph neural network model is used to learn the embedded representation of nodes and relationships, and combined with a time series analysis model, to generate real-time health status assessment results and short-term health risk prediction results for individual users.

[0109] S4. Based on the health status assessment results and health risk prediction results, a personalized intervention sequence for the target user is inferred and generated through a dynamic policy optimization algorithm based on reinforcement learning.

[0110] S5. Convert the personalized intervention sequence into execution instructions that match the interface protocol of the smart device, drive the connected smart device to perform the intervention, and monitor the target user's response data to the intervention in real time.

[0111] S6. Based on user response data, dynamically adjust the relationship strength between intervention strategies and the multimodal health knowledge graph.

[0112] Furthermore, the formulas mentioned above are all dimensionless calculations, derived from software simulation using a large amount of collected data to approximate the real situation. The weighting coefficients and preset thresholds in the formulas are set by those skilled in the art based on actual conditions or obtained through large-scale data simulation. The magnitude of the weighting coefficient is a specific value obtained by quantifying each parameter to facilitate subsequent comparison. The magnitude of the weighting coefficient depends on the amount of sample data and the processing coefficients initially set by those skilled in the art for each set of sample data. As long as it does not affect the proportional relationship between the parameter and the quantified value, it is acceptable.

[0113] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. The focus of each embodiment is on its differences from other embodiments. In particular, the apparatus embodiments are described simply because they are fundamentally based on the method embodiments; relevant parts can be referred to the descriptions of the method embodiments.

[0114] For ease of description, the above devices are described separately by function as various units. Of course, in implementing this application, the functions of each unit can be implemented in one or more software and / or hardware.

[0115] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can be implemented in one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs) containing computer-usable program code. The form of a computer program product implemented on ROM, optical memory, etc.

[0116] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0117] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0118] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0119] Secondly: The accompanying drawings of the embodiments disclosed in this invention only involve the structures involved in the embodiments disclosed in this invention. Other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of this invention can be combined with each other.

[0120] In conclusion, the above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart health system, characterized in that, It includes a multi-source data acquisition module, a data fusion and knowledge graph construction module, a health status assessment and risk prediction module, a personalized intervention strategy generation module, and an intervention execution and feedback adaptation module; The multi-source data acquisition module is used to collect users' physiological parameter data, behavioral activity data, environmental perception data, and medical record data from multiple heterogeneous data sources in real time or near real time. The data fusion and knowledge graph construction module is used to preprocess, extract and align features from the collected multi-source data, and build and dynamically update a multimodal health knowledge graph based on this. Nodes represent entities including users, physiological indicators, symptoms, behaviors, environmental factors and diseases, while edges represent the semantic and statistical relationship strength between entities. The health status assessment and risk prediction module is used to generate real-time health status assessment results and short-term health risk prediction results for individual users based on a multimodal health knowledge graph. It employs a graph neural network model to learn the embedded representations of nodes and relationships, and combines this with a time-series analysis model. These results include: Based on a multimodal health knowledge graph, a graph embedding algorithm is used to learn the low-dimensional vector representation of each node and relationship in the graph, and obtain the graph embedding vector. By combining the time series of user historical data, time series analysis is performed on the key features extracted from the preprocessed data. The time series prediction model is used to analyze the changing trends of each key feature and obtain the time series trend features of the key features in the future preset time period. The graph embedding vector is fused with time-series trend features and input into a preset health score prediction model to calculate the user's current health status score. Based on the extrapolation results of the association paths and temporal trend features in the multimodal health knowledge graph, the probability of a user's health risk event occurring within a preset time period is predicted, and key risk influencing factors are identified. The personalized intervention strategy generation module is used to infer and generate a personalized intervention sequence for the target user, which includes a series of ordered intervention actions, based on the health status assessment results and health risk prediction results, through a dynamic strategy optimization algorithm based on reinforcement learning. The intervention execution and feedback adaptation module is used to convert personalized intervention sequences into execution instructions that match the interface protocol of smart devices, drive the connected smart devices to execute the intervention, and monitor the target user's response data to the intervention in real time. Based on the response data, it dynamically adjusts the relationship strength between the intervention strategy and the multimodal health knowledge graph.

2. The smart health system according to claim 1, characterized in that, The data fusion and knowledge graph construction module preprocesses, extracts, and aligns collected multi-source data, and constructs and dynamically updates a multimodal health knowledge graph. The process includes: Preprocessing of multi-source data includes data cleaning, noise reduction, outlier detection and imputation, and format standardization. Key features are extracted from the preprocessed data, including time and frequency domain features of physiological parameters, pattern features of behavioral activities, periodic features of environmental data, and structured features of medical records. The extracted key features are aligned under a unified timestamp and mapped to a preset entity space; Based on the aligned feature data, entity linking and relation extraction techniques are used to construct a multimodal health knowledge graph with users as the core nodes. The node types include user nodes, physiological indicator nodes, symptom nodes, behavior nodes, environment nodes, and disease nodes. The edge relation types include user-have-indicator, indicator-may cause-symptom, behavior-influence-indicator, environment-exacerbate-risk, and disease-associate-symptom. A graph database is used to store and index a multimodal health knowledge graph. An incremental update algorithm is used to dynamically update the node attributes, edge relationship strength, and newly added entity relationships in the multimodal health knowledge graph.

3. The intelligent health system according to claim 1, characterized in that, The system further includes: a target user screening module for determining target users from all users, the process of which includes: Based on health status scores and health risk prediction probabilities, a comprehensive health risk value is calculated for each user; the comprehensive health risk value is compared with a preset comprehensive health risk threshold; users whose comprehensive health risk values ​​exceed the comprehensive health risk threshold are identified as target users who require priority intervention.

4. A smart health system according to claim 1 or 3, characterized in that, The personalized intervention strategy generation module, based on health status assessment results and health risk prediction results, uses a reinforcement learning-based dynamic strategy optimization algorithm to infer and generate a personalized intervention sequence for the target user, containing a series of ordered intervention actions. This process includes: The user's health status assessment results are integrated with the health risk prediction results to construct the state of the reinforcement learning agent. The action space of an agent is defined as the various executable intervention methods and their combinations. Construct a reward function that uses the expectation of health improvement as a positive reward and the probability of risk occurrence and user discomfort as negative rewards to guide the learning process of the agent; In a simulated environment consisting of a multimodal health knowledge graph and user historical data, an agent is trained to learn a strategy that maps the optimal sequence of actions, i.e., a personalized intervention sequence, based on the current state.

5. The intelligent health system according to claim 1, characterized in that, The intervention execution and feedback adaptation module converts personalized intervention sequences into execution instructions that match the interface protocol of smart devices, drives the connected smart devices to execute the intervention, and monitors the target user's response data to the intervention in real time. The process of dynamically adjusting the relationship strength between the intervention strategy and the multimodal health knowledge graph based on the response data includes: Collect data on changes in users' physiological parameters and behavioral activities after intervention as response data; Calculate the deviation between the response data and the expected improvement target to evaluate the effectiveness of the intervention strategy; Based on the evaluation results, the parameters of the decision algorithm in the personalized intervention strategy generation module were adjusted using the Bayesian optimization method. Based on the intervention effect and user response, update the relationship strength between relevant entities in the multimodal health knowledge graph.

6. The intelligent health system according to claim 2, characterized in that, The process of extracting key features from preprocessed data includes: For physiological parameter data, extract its time-domain features, including mean, variance, maximum and minimum values, and extract its frequency-domain features through Fourier transform; For behavioral activity data, clustering algorithms are used to identify behavioral patterns, and the frequency and duration of different patterns are extracted as pattern features. For environmental data, daily and weekly periodic characteristics are extracted as periodic features through periodic graph analysis. For medical record data, natural language processing technology is used to extract its structured entities and relationships as structured features.

7. The intelligent health system according to claim 1, characterized in that, The process of predicting the probability of a user experiencing a target health risk event within a preset time period based on extrapolation results of association paths and temporal trend features in a multimodal health knowledge graph, and identifying key risk influencing factors, includes: In a multimodal health knowledge graph, user nodes corresponding to target users are located, and one or more physiological indicator nodes that are currently abnormal and mapped by key features of the target user are associated with them through user-have-indicator relationships, serving as the starting set of paths; at the same time, disease nodes corresponding to predicted health risk events are located, serving as the target nodes of paths; and all associated paths from any physiological indicator node in the starting set of paths to the target disease node of the path are extracted. Calculate the weight of each associated path, where the path weight is determined by the product of the relation strength of all edges on the path and the risk coefficient of the nodes traversed by the path; For each associated path, extract one or more physiological indicator nodes on the path; directly obtain the time-series trend features generated by the time-series prediction model for the key features corresponding to these physiological indicator nodes; integrate the weight of the associated path with the corresponding future change values ​​in the time-series trend features, and calculate the final health risk prediction probability through the probability prediction model. Among all associated paths, the top-N critical paths that contribute the most to the probability of the final health risk prediction are selected; the nodes connected by the edges with the strongest relationship on these critical paths are extracted, and the entities corresponding to the nodes are the key risk factors. The key risk factors are used to prioritize the treatment in the generated personalized intervention sequence.

8. The smart health system according to claim 4, characterized in that, The process of training an agent to learn a policy includes: The agent selects an action based on its current state and the agent's policy network. The simulation environment updates to a new state based on the selected action and calculates an immediate reward according to the reward function; The agent uses the policy gradient method to update the agent's policy network parameters based on the state, action, and reward sequence. After multiple rounds of iterative training, a stable and optimized strategy is finally obtained. Applying this strategy to the user's current health status will output a corresponding personalized intervention sequence.

9. A data processing method for a smart health system, characterized in that, The method, applied to a smart health system as described in any one of claims 1-8, includes: S1. Collect users' physiological parameter data, behavioral activity data, environmental perception data, and medical record data from multiple heterogeneous data sources in real time or near real time; S2. Preprocess, extract features and align the collected multi-source data in a spatiotemporal manner, and build and dynamically update a multimodal health knowledge graph based on the processed data; S3. Based on a multimodal health knowledge graph, a graph neural network model is used to learn the embedded representation of nodes and relationships, and combined with a time series analysis model, to generate real-time health status assessment results and short-term health risk prediction results for individual users. S4. Based on the health status assessment results and health risk prediction results, a personalized intervention sequence for the target user is inferred and generated through a dynamic policy optimization algorithm based on reinforcement learning. S5. Convert the personalized intervention sequence into execution instructions that match the interface protocol of the smart device, drive the connected smart device to perform the intervention, and monitor the target user's response data to the intervention in real time. S6. Based on user response data, dynamically adjust the relationship strength between intervention strategies and the multimodal health knowledge graph.

Images