Health management service system and method based on AI optimization

Abstract

The invention discloses a health management service system and method based on AI optimization. The method comprises the following steps: S1, collecting multi-source health data of a user and constructing a structured health data set; s2, constructing a health state graph containing node attributes, edge connection weights and time indexes; s3, inputting the health state atlas into a linear graph neural network to generate a health state embedded vector; s4, collecting context information of a user, and fusing to generate personalized state perception representation and a health target embedding vector; s5, inputting the improved general value layering model to generate a structured intervention action candidate set; s6, screening and outputting a personalized health intervention scheme based on the matching score; s7, constructing a state evolution sequence, and inputting the state evolution sequence into a semi-supervised anomaly detection algorithm for monitoring and recognition; and S8, when the risk threshold is exceeded, triggering early warning and dynamically adjusting the intervention scheme. According to the method, collaborative optimization of personalized modeling and intelligent intervention strategies is realized, and the method is suitable for health management service scenes.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and health management technology, and in particular to a health management service system and method based on AI optimization. Background Technology

[0002] With the development of artificial intelligence technology and the rise of personalized health needs, intelligent health management services based on multi-source data are gradually becoming an important direction for medical assistance and resident health intervention. Existing health management systems typically rely on simple statistical models or rule bases to analyze health status and recommend intervention strategies, but in practical applications, they generally suffer from the following problems:

[0003] First, most existing systems are based on single or static health data sources, making it difficult to integrate multi-source health data with strong temporal dynamics and complex structures, resulting in partial and one-sided health status assessment results. Second, there is a lack of effective data structure modeling methods, which cannot reveal the multi-dimensional correlation between users' physiological, behavioral, and diagnostic information, limiting the contextual adaptability of intervention plans. Third, the generation mechanism of intervention strategies mostly relies on fixed rules or static target matching, making it difficult to achieve dynamic response and strategy transfer in different target switching scenarios, lacking flexibility and intelligence. Fourth, in terms of health status monitoring and anomaly identification, most existing methods rely on threshold judgment or fully supervised models, which are difficult to cope with the evolutionary characteristics of users' health status and the problem of capturing early abnormal fluctuations, resulting in a delayed response of early warning mechanisms.

[0004] Therefore, how to provide a health management service system and method based on AI optimization is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an AI-optimized health management service system and method. This invention employs a linear graph neural network, an improved general value hierarchical model, and a semi-supervised anomaly detection algorithm, combined with personalized state perception and contextual information, to construct a health status graph, enabling hierarchical generation, priority selection, and dynamic adjustment of intervention strategies. This invention effectively integrates the structural characteristics and goal-driven information of health data, possessing advantages such as accurate modeling, real-time response, personalized strategies, and strong risk warning capabilities.

[0006] A health management service method based on AI optimization according to an embodiment of the present invention includes the following steps:

[0007] S1. Collect users' multi-source health data and perform preprocessing operations to build a structured health dataset;

[0008] S2. Construct a health status graph containing node attributes, edge connection weights, and time indexes based on the structured health dataset;

[0009] S3. Input the health status map into the linear graph neural network to generate the user's current health status embedding vector;

[0010] S4. Collect the user's contextual information, including the current time period, activity scenario, behavioral preferences and historical response records, and fuse the health status embedding vector with the contextual information to form a personalized status perception representation. At the same time, based on the user's medium- and long-term health management goals or the goals set by the doctor, generate the corresponding health goal embedding vector.

[0011] S5. Input the personalized state perception representation and health goal embedding vector into the improved general value hierarchical model. The high-level strategy module, which introduces a semantic perception goal embedding mechanism, generates sub-goal intention embedding vectors. The low-level strategy module, under the control of the dynamic goal switching mechanism, combines the personalized state perception representation and sub-goal intention embedding vectors to output a set of structured intervention action candidates.

[0012] S6. Label the candidate set of structured intervention actions, match the labels with user behavior preferences and historical response records to obtain a matching score, construct a priority list of intervention actions based on the matching score, and select the strategy with the highest score to output a personalized health intervention plan.

[0013] S7. Based on the user's behavioral response and changes in health indicators after implementing a personalized health intervention plan, construct a state evolution sequence and input it into a semi-supervised anomaly detection algorithm to monitor health status and identify anomalies in real time. When the detection result exceeds the preset risk threshold, trigger the intervention early warning mechanism and dynamically adjust the intervention plan.

[0014] Optionally, S1 specifically includes:

[0015] S11. Collect multi-source health data of users, including physiological indicator data, behavioral activity data, historical medical record information and time series data recorded by wearable devices;

[0016] S12. Perform format conversion, outlier removal and field reconstruction operations on the collected multi-source health data, and perform missing value imputation and multi-source alignment based on a unified time index to generate a structured health dataset with unified dimensions and continuous time sequence.

[0017] Optionally, S2 specifically includes:

[0018] S21. Based on the physiological indicator data, behavioral activity data and historical medical record information in the structured health dataset, construct physiological nodes, behavioral nodes and diagnosis and treatment nodes, and each node contains a related numerical or categorical node attribute vector.

[0019] S22. Based on the timestamp information of the nodes, perform node pairing within the time window and construct co-occurrence edges. Within the preset time window, if physiological nodes, behavioral nodes, or diagnostic nodes appear in the same time period, it is considered that the node pair has co-occurred, and the number of occurrences is accumulated. Finally, the co-occurrence frequency is calculated based on the number of co-occurrences of the node pair and the time span within the window, and then the co-occurrence frequency is normalized to the initial weight of the co-occurrence edge.

[0020] S23. Based on the sequential relationship of nodes in the time series, construct sequential edges and set a time index as an attribute for each sequential edge. The time index represents the time difference between nodes and describes the temporal dependency between nodes.

[0021] S24. Based on the definition of causal relationship in the health knowledge base, establish causal edges between eligible node pairs, and calculate the credibility of the causal relationship between node pairs through correlation analysis, which serves as the corresponding weight for each causal edge.

[0022] S25. Combine the node attribute vectors of physiological nodes, behavioral nodes, and diagnostic nodes with the information of co-occurrence edges, sequential edges, and causal edges, and use a weighted graph structure representation method to generate a health status map.

[0023] Optionally, S3 specifically includes:

[0024] S31. Use the health status graph as the input structure of the graph neural network. Each node in the health status graph is input with its attribute vector, and each edge is input with its corresponding edge attribute.

[0025] S32. The node attribute vectors are linearly aggregated and transformed by a linear graph neural network. The attribute vectors of each node are updated by combining the weights of adjacent edges and time indices, and the state representation vectors of each node in the health status graph are generated.

[0026] S33. Perform a global aggregation operation on the state representation vectors of all nodes in the health status graph to obtain the user's current health status embedding vector.

[0027] Optionally, S4 specifically includes:

[0028] S41. Collect the user's current context information, including activity scenarios, behavioral preferences and historical response records, and perform format conversion and numerical normalization processing.

[0029] S42. Perform dimension alignment processing on the health status embedding vector and the context information respectively, and combine them into a joint vector using feature concatenation method as a personalized state perception representation.

[0030] S43. Standardize and encode the user-defined medium- and long-term health management goals or doctor-defined goals, and input them into the embedding mapping network. Generate health goal embedding vectors through lookup-based embedding tables or fully connected mapping functions.

[0031] Optionally, S5 specifically includes:

[0032] S51. The personalized state perception representation and the health target embedding vector are concatenated to form a joint input representation, which is then input into the high-level policy processing module of the improved general value hierarchical model.

[0033] S52. The high-level strategy processing module introduces a semantically aware target embedding mechanism to expand the semantic features of the health target embedding vector, and combines the context information in the joint input representation with the health status embedding vector to perform multi-head attention calculation and generate multiple sub-target intent embedding vectors.

[0034] S53. The above-mentioned sub-target intent embedding vector and personalized state-aware representation are concatenated and input into the low-level policy module to construct a sub-target candidate set;

[0035] S54. Under the control of the dynamic target switching mechanism, the low-level strategy module combines the user's behavioral preferences, historical response records and personalized state perception representation to calculate the matching score of the intent embedding vector of each sub-target, and selects the one with the highest score as the currently active sub-target intent embedding vector.

[0036] S55. The currently activated sub-target intent embedding vector is concatenated with the personalized state perception representation and input into the multi-layer perception structure to generate multiple intervention action vectors. The health target label is marked according to the target type information contained in the health target embedding vector. The intervention method label is determined according to the behavior channel in the intervention action vector generation process. The execution difficulty is estimated by combining the user's historical response record of similar intervention actions. The execution difficulty label is marked.

[0037] S56. Combine the intervention action vectors and their corresponding labels according to a preset format to form a structured intervention action candidate set.

[0038] Optionally, S6 specifically includes:

[0039] S61. For each intervention action in the structured intervention action candidate set, compare its attached health goal label, intervention method label and execution difficulty label with the corresponding executed intervention records in the user behavior preferences and historical response records item by item.

[0040] S62. During the comparison process, each label type is assigned a matching weight, the category consistency is determined for the health target label, the behavioral channel similarity is calculated for the intervention method label, and the historical acceptance is estimated for the implementation difficulty label.

[0041] S63. Multiply the three matching results by their respective set weights and then perform a weighted sum to obtain the matching score for each intervention action;

[0042] S64. Sort all intervention actions from highest to lowest matching score, construct an intervention action priority list, and select the intervention action with the highest score as the current output personalized health intervention plan.

[0043] Optionally, S7 specifically includes:

[0044] S71. Based on the user's behavioral response and health indicator changes after implementing a personalized health intervention plan, construct a state evolution sequence: extract behavioral response and health indicator change data within each time window from the user's multi-source health data in chronological order, and combine them into an ordered state sequence. The state at each time point includes the user's health indicators, behavioral data, and user feedback information on the previous intervention plan, forming a complete state evolution sequence.

[0045] S72. Input the state evolution sequence into the semi-supervised anomaly detection algorithm. The semi-supervised anomaly detection algorithm is based on the state evolution sequence and combines historical health data to perform real-time monitoring and anomaly identification, and identify abnormal fluctuations or signs of health risks.

[0046] S73. When the test result exceeds the preset risk threshold, the intervention early warning mechanism is triggered, and new health intervention plan adjustment suggestions are output.

[0047] According to an embodiment of the present invention, an AI-optimized health management service system includes the following modules:

[0048] The data processing module is used to collect users' multi-source health data, perform preprocessing operations, and build a structured health dataset.

[0049] The graph construction module is used to construct a health status graph containing node attributes, edge connection weights, and time indexes based on user health-related information contained in the structured health dataset.

[0050] The state generation module is used to input the health state map into the linear graph neural network and generate the user's current health state embedding vector.

[0051] The information fusion module is used to collect user context information and fuse the health status embedding vector with the context information to construct a personalized status perception representation, while generating a health target embedding vector.

[0052] The strategy generation module is used to embed personalized state-aware representations and health goals into an improved general value hierarchical model of vector input to generate a set of structured intervention action candidates.

[0053] The strategy screening module is used to perform label adaptation and matching score calculation on the structured intervention action candidate set, construct an intervention action priority list, and output personalized health intervention plans.

[0054] The anomaly detection module is used to construct a state evolution sequence based on the user's behavioral response and changes in health indicators after implementing the intervention plan, and input it into the semi-supervised anomaly detection algorithm for real-time monitoring of health status and anomaly identification.

[0055] The dynamic adjustment module is used to trigger an intervention warning mechanism and dynamically adjust the current health intervention plan when the test results exceed the preset risk threshold.

[0056] The beneficial effects of this invention are:

[0057] (1) This invention constructs a health status graph that includes node attributes, edge connection weights and time indexes, and introduces a linear graph neural network to generate user health status embedding vectors. This enables structured modeling and time-series dynamic feature extraction of multi-source health data, overcoming the shortcomings of insufficient data correlation mining and one-sided status assessment in existing health management methods.

[0058] (2) This invention proposes an improved general value hierarchical model, which introduces a semantic perception target embedding mechanism and a dynamic target switching mechanism. It can generate sub-target intentions based on the user's personalized state perception representation and health target embedding vector, and dynamically filter intervention strategies, thus solving the problems of inflexible intervention generation and poor adaptability under traditional rule or single target matching methods.

[0059] (3) The present invention uses a semi-supervised anomaly detection algorithm to monitor and identify anomalies in the state evolution sequence in real time, and automatically adjusts the intervention plan when triggered by the risk threshold. It has stronger health risk warning capabilities and strategy adaptability, and makes up for the shortcomings of the existing system in terms of lagging early anomaly identification and single response mechanism. Attached Figure Description

[0060] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0061] Figure 1This is a flowchart of an AI-optimized health management service method proposed in this invention;

[0062] Figure 2 This is a schematic diagram of the structure of a health management service system based on AI optimization proposed in this invention. Detailed Implementation

[0063] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0064] refer to Figure 1 A health management service method based on AI optimization includes the following steps:

[0065] S1. Collect users' multi-source health data and perform preprocessing operations to build a structured health dataset;

[0066] S2. Construct a health status graph containing node attributes, edge connection weights, and time indexes based on the structured health dataset;

[0067] S3. Input the health status map into the linear graph neural network to generate the user's current health status embedding vector;

[0068] S4. Collect the user's contextual information, including the current time period, activity scenario, behavioral preferences and historical response records, and fuse the health status embedding vector with the contextual information to form a personalized status perception representation. At the same time, based on the user's medium- and long-term health management goals or the goals set by the doctor, generate the corresponding health goal embedding vector.

[0069] S5. Input the personalized state perception representation and health goal embedding vector into the improved general value hierarchical model. The high-level strategy module, which introduces a semantic perception goal embedding mechanism, generates sub-goal intention embedding vectors. The low-level strategy module, under the control of the dynamic goal switching mechanism, combines the personalized state perception representation and sub-goal intention embedding vectors to output a set of structured intervention action candidates.

[0070] S6. Label the candidate set of structured intervention actions, match the labels with user behavior preferences and historical response records to obtain a matching score, construct a priority list of intervention actions based on the matching score, and select the strategy with the highest score to output a personalized health intervention plan.

[0071] S7. Based on the user's behavioral response and changes in health indicators after implementing a personalized health intervention plan, construct a state evolution sequence and input it into a semi-supervised anomaly detection algorithm to monitor health status and identify anomalies in real time. When the detection result exceeds the preset risk threshold, trigger the intervention early warning mechanism and dynamically adjust the intervention plan.

[0072] In this embodiment, S1 specifically includes:

[0073] S11. Collect multi-source health data of users, including physiological indicator data, behavioral activity data, historical medical record information and time series data recorded by wearable devices;

[0074] S12. Perform format conversion, outlier removal and field reconstruction operations on the collected multi-source health data, and perform missing value imputation and multi-source alignment based on a unified time index to generate a structured health dataset with unified dimensions and continuous time sequence.

[0075] This implementation method significantly improves the integrity and temporal consistency of the collected multi-source health data by preprocessing it, including data cleaning, normalization, and time alignment, and effectively reduces noise interference and information redundancy.

[0076] In this embodiment, S2 specifically includes:

[0077] S21. Based on the physiological indicator data, behavioral activity data and historical medical record information in the structured health dataset, construct physiological nodes, behavioral nodes and diagnosis and treatment nodes, and each node contains a related numerical or categorical node attribute vector.

[0078] S22. Based on the timestamp information of the nodes, perform node pairing within the time window and construct co-occurrence edges. Within the preset time window, if physiological nodes, behavioral nodes, or diagnostic nodes appear in the same time period, it is considered that the node pair has co-occurred, and the number of occurrences is accumulated. Finally, the co-occurrence frequency is calculated based on the number of co-occurrences of the node pair and the time span within the window, and then the co-occurrence frequency is normalized to the initial weight of the co-occurrence edge.

[0079] S23. Based on the sequential relationship of nodes in the time series, construct sequential edges and set a time index as an attribute for each sequential edge. The time index represents the time difference between nodes and describes the temporal dependency between nodes.

[0080] S24. Based on the definition of causal relationship in the health knowledge base, establish causal edges between eligible node pairs, and calculate the credibility of the causal relationship between node pairs through correlation analysis, which serves as the corresponding weight for each causal edge.

[0081] S25. Combine the node attribute vectors of physiological nodes, behavioral nodes, and diagnostic nodes with the information of co-occurrence edges, sequential edges, and causal edges, and use a weighted graph structure representation method to generate a health status map.

[0082] This implementation constructs a health status graph that includes node attributes, edge connection weights, and time indexes, effectively expressing the temporal correlations and causal relationships among multidimensional elements such as physiological, behavioral, and medical treatment aspects of user health data. The graph employs a weighted graph structure representation, unifying and integrating co-occurrence, sequence, and causal edge types and their attributes, thereby enhancing the structural expressive power of health information.

[0083] In this embodiment, S3 specifically includes:

[0084] S31. Use the health status graph as the input structure of the graph neural network. Each node in the health status graph is input with its attribute vector, and each edge is input with its corresponding edge attribute.

[0085] S32. The node attribute vectors are linearly aggregated and transformed by a linear graph neural network. The attribute vectors of each node are updated by combining the weights of adjacent edges and time indices, and the state representation vectors of each node in the health status graph are generated.

[0086] S33. Perform a global aggregation operation on the state representation vectors of all nodes in the health status graph to obtain the user's current health status embedding vector.

[0087] This implementation utilizes the connectivity and edge weight information between nodes in the graph structure to perform linear propagation and aggregation operations on health features, generating a health status embedding vector that reflects the user's current overall health status. This approach ensures computational efficiency and interpretability while fully preserving key structural and temporal information within the graph, thus improving the accuracy and generalization ability of representing individual health status.

[0088] In this embodiment, S4 specifically includes:

[0089] S41. Collect the user's current context information, including activity scenarios, behavioral preferences and historical response records, and perform format conversion and numerical normalization processing.

[0090] S42. Perform dimension alignment processing on the health status embedding vector and the context information respectively, and combine them into a joint vector using feature concatenation method as a personalized state perception representation.

[0091] S43. Standardize and encode the user-defined medium- and long-term health management goals or doctor-defined goals, and input them into the embedding mapping network. Generate health goal embedding vectors through lookup-based embedding tables or fully connected mapping functions.

[0092] This implementation method collects the user's contextual information and performs vector-level fusion with the health status embedding vector to construct a personalized state-aware representation. Simultaneously, it generates a corresponding health goal embedding vector by combining the user's medium- to long-term health management goals or the goals set by the doctor. This approach achieves a deep integration of health status characteristics and individual behavioral context, making the generated state representation more individualized and goal-oriented.

[0093] In this embodiment, S5 specifically includes:

[0094] S51. The personalized state perception representation and the health target embedding vector are concatenated to form a joint input representation, which is then input into the high-level policy processing module of the improved general value hierarchical model.

[0095] S52. The high-level strategy processing module introduces a semantically aware target embedding mechanism to expand the semantic features of the health target embedding vector, and combines the context information in the joint input representation with the health status embedding vector to perform multi-head attention calculation and generate multiple sub-target intent embedding vectors.

[0096] S53. The above-mentioned sub-target intent embedding vector and personalized state-aware representation are concatenated and input into the low-level policy module to construct a sub-target candidate set;

[0097] S54. Under the control of the dynamic target switching mechanism, the low-level strategy module combines the user's behavioral preferences, historical response records and personalized state perception representation to calculate the matching score of the intent embedding vector of each sub-target, and selects the one with the highest score as the currently active sub-target intent embedding vector.

[0098] S55. The currently activated sub-target intent embedding vector is concatenated with the personalized state perception representation and input into the multi-layer perception structure to generate multiple intervention action vectors. The health target label is marked according to the target type information contained in the health target embedding vector. The intervention method label is determined according to the behavior channel in the intervention action vector generation process. The execution difficulty is estimated by combining the user's historical response record of similar intervention actions. The execution difficulty label is marked.

[0099] S56. Combine the intervention action vectors and their corresponding labels according to a preset format to form a structured intervention action candidate set.

[0100] This implementation improves the general value hierarchy model by inputting personalized state-aware representations and health goal embedding vectors. It uses a semantically aware goal embedding mechanism to generate multiple sub-goal intent embedding vectors. Under the control of a dynamic goal switching mechanism, and combined with user behavior preferences and historical response records, it selects the current optimal sub-goal intent embedding vector to further generate a structured set of intervention action candidates. This approach achieves goal-oriented multi-strategy generation and dynamic adaptation, possessing advantages such as clear task hierarchy, strong response flexibility, and a well-defined intervention strategy structure, providing a foundation for more intelligent and efficient intervention solutions for health management.

[0101] In this embodiment, S6 specifically includes:

[0102] S61. For each intervention action in the structured intervention action candidate set, compare its attached health goal label, intervention method label and execution difficulty label with the corresponding executed intervention records in the user behavior preferences and historical response records item by item.

[0103] S62. During the comparison process, each label type is assigned a matching weight, the category consistency is determined for the health target label, the behavioral channel similarity is calculated for the intervention method label, and the historical acceptance is estimated for the implementation difficulty label.

[0104] S63. Multiply the three matching results by their respective set weights and then perform a weighted sum to obtain the matching score for each intervention action;

[0105] S64. Sort all intervention actions from highest to lowest matching score, construct an intervention action priority list, and select the intervention action with the highest score as the current output personalized health intervention plan.

[0106] This implementation method calculates a matching score, constructs a priority list of intervention actions, and selects the one with the highest score as the personalized health intervention plan. Based on a comprehensive consideration of individual user characteristics and behavioral response patterns, this approach achieves precise matching and prioritization of intervention strategies, effectively improving the relevance, acceptability, and actual implementation effectiveness of the intervention plan.

[0107] In this embodiment, S7 specifically includes:

[0108] S71. Based on the user's behavioral response and health indicator changes after implementing a personalized health intervention plan, construct a state evolution sequence: extract behavioral response and health indicator change data within each time window from the user's multi-source health data in chronological order, and combine them into an ordered state sequence. The state at each time point includes the user's health indicators, behavioral data, and user feedback information on the previous intervention plan, forming a complete state evolution sequence.

[0109] S72. Input the state evolution sequence into the semi-supervised anomaly detection algorithm. The semi-supervised anomaly detection algorithm is based on the state evolution sequence and combines historical health data to perform real-time monitoring and anomaly identification, and identify abnormal fluctuations or signs of health risks.

[0110] S73. When the test result exceeds the preset risk threshold, the intervention early warning mechanism is triggered, and new health intervention plan adjustment suggestions are output.

[0111] This implementation method uses a semi-supervised anomaly detection algorithm to monitor changes in users' health status in real time and identify potential anomalies. This approach enables highly sensitive early risk identification in real-world scenarios lacking complete labeled data, effectively improving the dynamic response capability and risk control level of the health management system, and providing a scientific and timely basis for individual health interventions.

[0112] refer to Figure 2 A health management service system based on AI optimization includes the following modules:

[0113] The data processing module is used to collect users' multi-source health data, perform preprocessing operations, and build a structured health dataset.

[0114] The graph construction module is used to construct a health status graph containing node attributes, edge connection weights, and time indexes based on user health-related information contained in the structured health dataset.

[0115] The state generation module is used to input the health state map into the linear graph neural network and generate the user's current health state embedding vector.

[0116] The information fusion module is used to collect user context information and fuse the health status embedding vector with the context information to construct a personalized status perception representation, while generating a health target embedding vector.

[0117] The strategy generation module is used to embed personalized state-aware representations and health goals into an improved general value hierarchical model of vector input to generate a set of structured intervention action candidates.

[0118] The strategy screening module is used to perform label adaptation and matching score calculation on the structured intervention action candidate set, construct an intervention action priority list, and output personalized health intervention plans.

[0119] The anomaly detection module is used to construct a state evolution sequence based on the user's behavioral response and changes in health indicators after implementing the intervention plan, and input it into the semi-supervised anomaly detection algorithm for real-time monitoring of health status and anomaly identification.

[0120] The dynamic adjustment module is used to trigger an intervention warning mechanism and dynamically adjust the current health intervention plan when the test results exceed the preset risk threshold.

[0121] This implementation method constructs a health management service system comprising a structured data acquisition module, a health status map construction module, a graph neural network modeling module, a personalized status fusion module, a hierarchical strategy generation module, an intervention strategy screening module, a health status monitoring module, and an early warning feedback module. This system achieves a complete closed loop from multi-source data modeling, health status identification, intervention strategy formulation, to dynamic feedback adjustment. The system's modules are clearly defined and work in synergy, providing real-time, accurate, and personalized intelligent health management services in complex and ever-changing individual health scenarios, significantly improving the scientific rigor of intervention decisions and the level of intelligence in health services.

[0122] Example 1:

[0123] To verify the feasibility of this invention in practice, it was applied to a large-scale enterprise internal health management project targeting 324 frontline employees in a manufacturing workshop. The goal was to improve the overall health level of employees and enhance the ability to predict abnormal events through automated intervention, tracking, and optimization over a three-month period. This project relied on the company's internal health record system and supporting wearable devices, integrating the health management service methods proposed in this invention into the enterprise health platform, taking into account employees' daily work rhythms and health consultation services.

[0124] During project implementation, the system first collects multi-source health data from employees, primarily including heart rate, sleep, step count, and body temperature data collected by smart bracelets, regular blood pressure and blood sugar test data, as well as information such as health questionnaires, dietary preferences, and medical history filled out by employees. The data is uploaded in real-time through a device gateway, updated at least 24 times daily, forming a structured health dataset. After normalization using a unified time step, the platform constructs a health status map for each employee. The node attributes in the map cover the current health indicator level and historical trends, while the edge weights reflect the degree of correlation and sequence between events, such as the correlation between lunch content and afternoon heart rate fluctuations.

[0125] The system further employs a linear graph neural network to extract state embedding vectors, which represent the employee's current overall health status. Next, the platform combines the employee's contextual information, such as day / night shift status, workload score, recent dietary and sleep behaviors, and personalized health goals set by the health management department (such as improving sleep quality and controlling blood sugar fluctuations), to construct a state-aware representation and a health goal embedding vector, and inputs them into the improved general value hierarchical model.

[0126] The high-level model uses a semantically aware goal embedding mechanism to analyze the current goal intent and generate multiple feasible sub-goal intents (such as going to bed 30 minutes earlier on a certain day, or reducing high-intensity steps). The low-level model uses a dynamic goal switching mechanism, combined with the user's past preferences (some employees prefer outdoor activities, while others prefer meditation or diet control), to filter the optimal strategy intent, ultimately forming a set of labeled intervention action candidates. After matching the candidate set with the employee's behavioral preferences, the system pushes the optimal personalized health intervention plan, such as pushing a low-intensity stretching video two hours before the end of the night shift and recommending a soothing drink to accompany a low-fat dinner.

[0127] The system continuously records changes in health indicators after intervention, such as heart rate stability, body temperature fluctuations, and sleep efficiency, as well as employee feedback scores (submitted via the platform's mini-program). This data constitutes a state evolution sequence, which is input into an improved semi-supervised anomaly detection algorithm. The system can identify risk trends 2-3 hours before an abnormal state occurs; for example, it can issue an early warning 3 hours in advance if abnormal blood pressure fluctuations are detected, push adjustment suggestions, and trigger the management intervention approval mechanism.

[0128] During the three-month testing period, the platform pushed a total of 15,684 personalized intervention suggestions to 324 employees, issued 278 warnings of abnormal trends, and dynamically adjusted intervention strategies 945 times. Employees gave the platform an average satisfaction rating of 4.7 out of 5, with over 90% of employees saying that the system's push notifications were "timely, personalized, and helpful."

[0129] The table below compares the performance of the system of this invention with the enterprise's original health recommendation system on five core indicators:

[0130] Table 1 Performance Comparison of Health Management Systems

[0131] Evaluation indicators Traditional system This invention system Intervention matching accuracy rate (%) 72.3 89.6 Status recognition accuracy (%) 78.5 91.2 Anomaly detection lead time (hours) 1.8 3.4 Average user satisfaction rating (out of 5) 3.9 4.7 The strategy dynamically adjusts the response time (in seconds). 8.6 2.1

[0132] As can be seen from the table, the system of the present invention has improved the accuracy of intervention matching and health status identification by 17.3% and 12.7% respectively, and the average advance time for anomaly identification has been improved from 1.8 hours to 3.4 hours; user satisfaction scores have been improved by more than 20%, and the dynamic response time of the strategy has been shortened from 8.6 seconds to 2.1 seconds, which fully demonstrates the advantages of the present invention in terms of practicality, intelligence and efficiency.

[0133] The following beneficial effects can be obtained by combining the above embodiments: By introducing graph structure modeling and linear graph neural networks, the correlation information in multi-source health data is effectively integrated, significantly improving the accuracy of state modeling; through the semantic perception and dynamic switching mechanism in the hierarchical value strategy model, the intervention generation process has the characteristics of target perception and strong adaptability, improving the rationality and execution rate of intervention recommendations; through the improved semi-supervised anomaly detection algorithm, the early perception of health risks and dynamic intervention updates are realized without the need for complete annotation, significantly enhancing the responsiveness and risk prevention and control level of the health management system.

[0134] 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 health management service method based on AI optimization, characterized in that, Includes the following steps: S1. Collect users' multi-source health data and perform preprocessing operations to build a structured health dataset; S2. Construct a health status graph containing node attributes, edge connection weights, and time indexes based on the structured health dataset; S3. Input the health status map into the linear graph neural network to generate the user's current health status embedding vector; S4. Collect the user's contextual information, including the current time period, activity scenario, behavioral preferences and historical response records, and fuse the health status embedding vector with the contextual information to form a personalized status perception representation. At the same time, based on the user's medium- and long-term health management goals or the goals set by the doctor, generate the corresponding health goal embedding vector. S5. Input the personalized state perception representation and health goal embedding vector into the improved general value hierarchical model. The high-level strategy module, which introduces a semantic perception goal embedding mechanism, generates sub-goal intention embedding vectors. The low-level strategy module, under the control of the dynamic goal switching mechanism, combines the personalized state perception representation and sub-goal intention embedding vectors to output a set of structured intervention action candidates. S6. Label the candidate set of structured intervention actions, match the labels with user behavior preferences and historical response records to obtain a matching score, construct a priority list of intervention actions based on the matching score, and select the strategy with the highest score to output a personalized health intervention plan. S7. Based on the user's behavioral response and changes in health indicators after implementing a personalized health intervention plan, construct a state evolution sequence and input it into a semi-supervised anomaly detection algorithm to monitor health status and identify anomalies in real time. When the detection result exceeds the preset risk threshold, trigger the intervention early warning mechanism and dynamically adjust the intervention plan.

2. The AI-optimized health management service method according to claim 1, characterized in that, S1 specifically includes: S11. Collect multi-source health data of users, including physiological indicator data, behavioral activity data, historical medical record information and time series data recorded by wearable devices; S12. Perform format conversion, outlier removal and field reconstruction operations on the collected multi-source health data, and perform missing value imputation and multi-source alignment based on a unified time index to generate a structured health dataset with unified dimensions and continuous time sequence.

3. The AI-optimized health management service method according to claim 1, characterized in that, S2 specifically includes: S21. Based on the physiological indicator data, behavioral activity data and historical medical record information in the structured health dataset, construct physiological nodes, behavioral nodes and diagnosis and treatment nodes, and each node contains a related numerical or categorical node attribute vector. S22. Based on the timestamp information of the nodes, perform node pairing within the time window and construct co-occurrence edges. Within the preset time window, if physiological nodes, behavioral nodes, or diagnostic nodes appear in the same time period, it is considered that the node pair has co-occurred, and the number of occurrences is accumulated. Finally, the co-occurrence frequency is calculated based on the number of co-occurrences of the node pair and the time span within the window, and then the co-occurrence frequency is normalized to the initial weight of the co-occurrence edge. S23. Based on the sequential relationship of nodes in the time series, construct sequential edges and set a time index as an attribute for each sequential edge. The time index represents the time difference between nodes and describes the temporal dependency between nodes. S24. Based on the definition of causal relationship in the health knowledge base, establish causal edges between eligible node pairs, and calculate the credibility of the causal relationship between node pairs through correlation analysis, which serves as the corresponding weight for each causal edge. S25. Combine the node attribute vectors of physiological nodes, behavioral nodes, and diagnostic nodes with the information of co-occurrence edges, sequential edges, and causal edges, and use a weighted graph structure representation method to generate a health status map.

4. The AI-optimized health management service method according to claim 1, characterized in that, S3 specifically includes: S31. Use the health status graph as the input structure of the graph neural network. Each node in the health status graph is input with its attribute vector, and each edge is input with its corresponding edge attribute. S32. The node attribute vectors are linearly aggregated and transformed by a linear graph neural network. The attribute vectors of each node are updated by combining the weights of adjacent edges and time indices, and the state representation vectors of each node in the health status graph are generated. S33. Perform a global aggregation operation on the state representation vectors of all nodes in the health status graph to obtain the user's current health status embedding vector.

5. The AI-optimized health management service method according to claim 1, characterized in that, S4 specifically includes: S41. Collect the user's current context information, including activity scenarios, behavioral preferences and historical response records, and perform format conversion and numerical normalization processing. S42. Perform dimension alignment processing on the health status embedding vector and the context information respectively, and combine them into a joint vector using feature concatenation method as a personalized state perception representation. S43. Standardize and encode the user-defined medium- and long-term health management goals or doctor-defined goals, and input them into the embedding mapping network. Generate health goal embedding vectors through lookup-based embedding tables or fully connected mapping functions.

6. The AI-optimized health management service method according to claim 1, characterized in that, S5 specifically includes: S51. The personalized state perception representation and the health target embedding vector are concatenated to form a joint input representation, which is then input into the high-level policy processing module of the improved general value hierarchical model. S52. The high-level strategy processing module introduces a semantically aware target embedding mechanism to expand the semantic features of the health target embedding vector, and combines the context information in the joint input representation with the health status embedding vector to perform multi-head attention calculation and generate multiple sub-target intent embedding vectors. S53. The above-mentioned sub-target intent embedding vector and personalized state-aware representation are concatenated and input into the low-level policy module to construct a sub-target candidate set; S54. Under the control of the dynamic target switching mechanism, the low-level strategy module combines the user's behavioral preferences, historical response records and personalized state perception representation to calculate the matching score of the intent embedding vector of each sub-target, and selects the one with the highest score as the currently active sub-target intent embedding vector. S55. The currently activated sub-target intent embedding vector is concatenated with the personalized state perception representation and input into the multi-layer perception structure to generate multiple intervention action vectors. The health target label is marked according to the target type information contained in the health target embedding vector. The intervention method label is determined according to the behavior channel in the intervention action vector generation process. The execution difficulty is estimated by combining the user's historical response record of similar intervention actions. The execution difficulty label is marked. S56. Combine the intervention action vectors and their corresponding labels according to a preset format to form a structured intervention action candidate set.

7. The AI-optimized health management service method according to claim 1, characterized in that, S6 specifically includes: S61. For each intervention action in the structured intervention action candidate set, compare its attached health goal label, intervention method label and execution difficulty label with the corresponding executed intervention records in the user behavior preferences and historical response records item by item. S62. During the comparison process, each label type is assigned a matching weight, the category consistency is determined for the health target label, the behavioral channel similarity is calculated for the intervention method label, and the historical acceptance is estimated for the implementation difficulty label. S63. Multiply the three matching results by their respective set weights and then perform a weighted sum to obtain the matching score for each intervention action; S64. Sort all intervention actions from highest to lowest matching score, construct an intervention action priority list, and select the intervention action with the highest score as the current output personalized health intervention plan.

8. The AI-optimized health management service method according to claim 1, characterized in that, Specifically, S7 includes: S71. Based on the user's behavioral response and health indicator changes after implementing a personalized health intervention plan, construct a state evolution sequence: extract behavioral response and health indicator change data within each time window from the user's multi-source health data in chronological order, and combine them into an ordered state sequence. The state at each time point includes the user's health indicators, behavioral data, and user feedback information on the previous intervention plan, forming a complete state evolution sequence. S72. Input the state evolution sequence into the semi-supervised anomaly detection algorithm. The semi-supervised anomaly detection algorithm is based on the state evolution sequence and combines historical health data to perform real-time monitoring and anomaly identification, and identify abnormal fluctuations or signs of health risks. S73. When the test result exceeds the preset risk threshold, the intervention early warning mechanism is triggered, and new health intervention plan adjustment suggestions are output.

9. An AI-optimized health management service system, executing the AI-optimized health management service method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data processing module is used to collect users' multi-source health data, perform preprocessing operations, and build a structured health dataset. The graph construction module is used to construct a health status graph containing node attributes, edge connection weights, and time indexes based on user health-related information contained in the structured health dataset. The state generation module is used to input the health state map into the linear graph neural network and generate the user's current health state embedding vector. The information fusion module is used to collect user context information and fuse the health status embedding vector with the context information to construct a personalized status perception representation, while generating a health target embedding vector. The strategy generation module is used to embed personalized state-aware representations and health goals into an improved general value hierarchical model of vector input to generate a set of structured intervention action candidates. The strategy screening module is used to perform label adaptation and matching score calculation on the structured intervention action candidate set, construct an intervention action priority list, and output personalized health intervention plans. The anomaly detection module is used to construct a state evolution sequence based on the user's behavioral response and changes in health indicators after implementing the intervention plan, and input it into the semi-supervised anomaly detection algorithm for real-time monitoring of health status and anomaly identification. The dynamic adjustment module is used to trigger an intervention warning mechanism and dynamically adjust the current health intervention plan when the test results exceed the preset risk threshold.

Images