Intelligent monitoring decision-making method based on knowledge graph and federal learning

Abstract

The invention discloses an intelligent monitoring decision-making method based on a knowledge graph and federal learning, and the method comprises the following steps: S1, collecting and preprocessing multi-source health data of a user, and generating a health data set; s2, constructing a local medical knowledge graph and performing knowledge embedding modeling to generate a knowledge representation vector; s3, constructing a health risk assessment model, and performing modeling in combination with knowledge representation and health data; s4, initializing a federated learning architecture, setting a client and an aggregation end, and distributing a model structure and parameters; s5, locally training the model by each federated client, and uploading parameters to an aggregation end to complete parameter aggregation; s6, combining the updated model with the real-time health data and a knowledge graph reasoning result to generate a personalized monitoring decision; and S7, collecting user feedback and newly added data, updating the knowledge graph and the model, and entering a new round of optimization. The method is used for realizing personalized health risk assessment and intelligent monitoring fusing the knowledge graph and federal learning while ensuring privacy.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical information processing technology, and in particular to an intelligent monitoring decision-making method based on knowledge graphs and federated learning. Background Technology

[0002] In the context of the development of contemporary intelligent medical systems, with the widespread deployment of wearable devices, edge computing nodes, and remote medical terminals, medical monitoring data exhibits characteristics of multi-source heterogeneity, high-dimensional temporal sequence, and strong individual privacy. Traditional centralized health monitoring systems generally rely on central servers for centralized collection and unified modeling of patient health data. However, due to tightening privacy protection policies, increased cross-institutional data barriers, and enhanced user sensitivity to personal health data, medical data exists as "data silos" across different terminals and institutions, severely restricting the improvement of cross-regional and cross-platform health risk prediction capabilities and intelligent intervention levels.

[0003] In existing technologies, most mainstream health risk prediction methods employ deep neural networks to train and infer multidimensional health characteristics. These models are typically built on large-scale centralized data, modeling an individual's historical vital signs, physiological indicators, and behavioral habits, and outputting disease risk levels, health scores, or suggested interventions. However, this approach suffers from three core technical limitations: First, its strong reliance on centralized data patterns results in insufficient model generalization ability, making it difficult to adapt to the diverse needs of different populations, regions, and device environments; second, due to the inability to integrate medical semantic knowledge and reliance on pure data-driven approaches, it lacks interpretable modeling of the health evolution logic, resulting in insufficient persuasiveness of the output for clinicians and users; and third, under isolated data conditions, it cannot leverage data from other terminals for collaborative optimization, leading to slow model updates and poor long-term robustness.

[0004] To address privacy concerns, some research has introduced federated learning, enabling terminal devices to train models locally and only upload model parameters to a central server for aggregation. This allows for multi-terminal collaborative modeling without exchanging raw data. This approach has already seen initial applications in tasks such as medical image recognition and diabetes risk screening. However, the practical application of federated learning in health monitoring still faces several key challenges: First, most methods are still based on traditional neural network structures and do not fully incorporate medical knowledge for semantic enhancement; second, the parameter aggregation process ignores individual semantic differences, making it difficult to balance personalization and universality in model updates; and third, the inconsistent training results across terminals under the federated mechanism may lead to "drifting aggregation," affecting the overall model performance.

[0005] On the other hand, knowledge graphs, as a structured knowledge representation method that can represent semantic relationships between entities, have been applied in medical question-answering systems and auxiliary diagnostic tools. Medical knowledge graphs are typically composed of entity nodes such as diseases, symptoms, indicators, drugs, and treatment plans, and are connected through "accompaniment," "causation," and "intervention" relationships to form a multi-layered semantic network. By introducing knowledge graphs, the reasoning ability and interpretability of models can be improved, making them particularly suitable for understanding and predicting complex health states in health monitoring systems. However, most existing systems deploy knowledge graph reasoning modules on the server side, making it difficult to support local real-time reasoning and dynamic updates; at the same time, under the federated training framework, the knowledge graph construction mechanism lacks standards and fusion strategies, making it impossible to achieve effective collaboration and semantic sharing between distributed graphs.

[0006] More seriously, existing methods for generating intelligent monitoring decision-making results often employ a single model-driven approach, failing to incorporate causal chains and multi-hop reasoning paths from the knowledge graph into the decision-making logic. This results in risk predictions and intervention recommendations lacking logical support, making it difficult to achieve structured, traceable, and explanatory reasoning. Furthermore, existing intervention recommendations are mostly static template matching or manually set rules, lacking adaptive association with the user's health history and knowledge graph context, thus failing to meet the personalized needs of "treating the same disease differently" and "dynamic intervention."

[0007] Therefore, existing health risk monitoring technologies have the following main shortcomings: First, they lack a mechanism to integrate medical knowledge graphs with federated learning architectures, making it impossible to achieve distributed modeling with structured semantic enhancement while protecting privacy. Second, they fail to design heterogeneous perception and adaptive collaborative model optimization strategies to address the differences in knowledge structures and data capabilities of different terminal devices. Third, the generation of intelligent monitoring results lacks inference chain support, and intervention suggestions lack interpretability and personalized relevance, making it difficult to form a reliable and dynamic monitoring loop. Fourth, traditional models focus on static parameter aggregation, ignoring the regulatory role of the knowledge graph on aggregation weights and path selection, and failing to form a semantically driven parameter update mechanism.

[0008] Therefore, how to provide an intelligent guardianship decision-making method based on knowledge graphs and federated learning is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0009] One objective of this invention is to propose an intelligent monitoring decision-making method based on knowledge graphs and federated learning. This invention fully integrates the semantic reasoning capabilities of medical knowledge graphs with the distributed privacy modeling mechanism of federated learning. It describes in detail how to construct a local medical knowledge graph, train an individual health risk assessment model, and achieve global model optimization through multi-terminal parameter aggregation without exposing the user's original health data. It has the advantages of strong data privacy protection, accurate health status identification, interpretable reasoning results, and strong ability to generate personalized intervention suggestions.

[0010] An intelligent guardianship decision-making method based on knowledge graphs and federated learning according to an embodiment of the present invention includes the following steps:

[0011] S1. Collect users' multi-source health data, preprocess the multi-source health data, and generate a health dataset;

[0012] S2. Based on the health dataset, construct a local medical knowledge graph, establish the relationship between entities through entity recognition and relation extraction, and perform local medical knowledge graph embedding modeling to form a knowledge representation vector;

[0013] S3. Construct a health risk assessment model, use graph neural networks to perform graph structure information propagation and feature extraction on knowledge representation vectors, and combine the health dataset to train and construct a health risk assessment model to complete the modeling of individual health status.

[0014] S4. Initialize the federated learning architecture, using multiple monitoring terminal devices as federated clients and the central coordination server as the federated aggregation terminal, distribute health risk assessment model parameters to each federated client, and configure communication and aggregation rules.

[0015] S5. Each federated client uses the health dataset and the embedded features of the local medical knowledge graph to train the health risk assessment model, and uploads the trained health risk assessment model parameters to the federated aggregation terminal, which performs parameter aggregation operation to generate and synchronously update the health risk assessment model.

[0016] S6. Based on the updated health risk assessment model, the system combines real-time health data with the reasoning results of the local medical knowledge graph to output the user's health risk level, possible disease categories, and suggested intervention measures, forming a personalized intelligent monitoring decision result.

[0017] S7. Incorporate user feedback data and newly added health data into the local medical knowledge graph and model for iterative updates, and participate in the joint optimization of the new round of health risk assessment model.

[0018] Optionally, the multi-source health data specifically includes physiological signals, medical records, examination and test results, medication records, and behavioral data, which are used to support the construction of a local medical knowledge graph and the training of a health risk assessment model.

[0019] Optionally, the preprocessing of multi-source health data specifically includes missing value imputation, standardization, structural transformation, and time alignment, which are used to improve the quality of multi-source health data and generate a health dataset that can be used for modeling.

[0020] Optionally, S2 specifically includes:

[0021] S21. Extract structured fields and unstructured text from the health dataset to construct an initial entity set E0 = {e1, e2, ..., e...}. n}, where e i Let i represent the i-th medical entity, where i = 1, 2, ..., n, and n is the total number of entities;

[0022] S22. Based on medical ontology and statistical models, construct a semantic relational database R0 = {r1, r2, ..., r...} m}, where r j Let j represent the j-th type of medical semantic relation, where j = 1, 2, ..., m, and m is the total number of medical semantic relation types. Through dependency parsing and co-occurrence network, entity pair relations are mined from E0 to obtain the local relation set R.

[0023] S23. Calculate the importance weight for each entity and filter out entities with high weight:

[0024] f(e i )=w1×freq(e i )+w2×coFreq(e i );

[0025] Where, f(e) i ) represents entity e i Importance weight, freq(e i ) is entity e i The number of times it appears in the health dataset, coFreq(e i ) is entity e i The number of times entities co-occur with diagnostic labels, w1 and w2 are empirical weighting coefficients; a threshold Θ is defined, and a set of high-weight entities E1 is constructed:

[0026] E1={e i ∈E0∣f(e i )>Θ};

[0027] S24. Generate a set of triples T = {(h, r, t) | h ∈ E1, r ∈ R, t ∈ E1} based on E1 and R, where h is the head entity, r is the edge type in the triple, and t is the tail entity. Remove low-confidence edges based on the confidence threshold δ to obtain the local medical knowledge graph G. local = (E1, R, T),

[0028] S25. For any entity e i Define the neighbor set N(e) i )={e j ∣(e i ,r,e j )∈T}, and incorporate the context into its own embedding:

[0029]

[0030] Among them, e j For the j-th entity, emb(e i ) is entity e i Initial embedding vector, emb(e j ) is entity e j The initial embedding vector, λ is the fusion coefficient, v(e i ) is entity e i Context-enhanced final embedding vector;

[0031] S26. Apply the TransE algorithm to the local medical knowledge graph, satisfying the condition that for any (h, r, t) ∈ T,

[0032] emb(h) + emb(r) ≈ emb(t);

[0033] Where emb(r) is the embedding vector of relation r, emb(h) is the embedding vector of head entity h, and emb(t) is the embedding vector of tail entity t;

[0034] S27, All entity vectors {v(e i )∣e i The knowledge representation vector K = {v(e ∈ E1} and the relation vector {emb(r) | r ∈ R} constitute the knowledge representation vector K = {v(e ∈ E1}}. i )}∪{emb(r)} is used as input for the health risk assessment model.

[0035] Optionally, S3 specifically includes:

[0036] S31. Obtain the formed knowledge representation vector set, construct a graph structure input according to the relationship between all entities using the embedding vectors of all entities, and construct a graph neural network input graph data;

[0037] S32. Construct a graph neural network structure containing multiple layers of graph neural computing units, and use a layer-by-layer propagation mechanism to diffuse and aggregate graph structure information on knowledge representation vectors, and extract a high-order representation that integrates structural features and semantic information.

[0038] S33. Extract time-series health data of individual users from the health dataset, including changes in physiological indicators, behavioral monitoring data and historical records, and construct data input samples for individual status modeling;

[0039] S34. The knowledge graph features extracted by the graph neural network are fused with the individual's health time series data to form a unified representation of the individual's health status.

[0040] S35. Construct a health risk assessment model based on fusion features, complete the learning of health risk assessment model parameters through a supervised training mechanism, and establish a mapping relationship between individual user health status and risk category;

[0041] S36. Complete the training process of the health risk assessment model to obtain the model structure and parameters that can be used for health status prediction, providing a basis for federal optimization and personalized monitoring decisions.

[0042] Optionally, S4 specifically includes:

[0043] S41. Establish a multi-terminal federated learning system architecture, set multiple terminal devices with health monitoring functions as federated clients, set the central coordination server as the federated aggregation terminal, and define the federated communication protocol and task scheduling process.

[0044] S42. Construct a unified structural framework for the health risk assessment model in the central coordination server, and set a health risk assessment model structure mapping table according to the computing resource level of each client. Load health risk assessment model substructures of corresponding complexity for devices with different performance to form a heterogeneous model adaptation mechanism.

[0045] S43. Distribute the initial health risk assessment model parameters, structural mapping relationships, and local training configuration files to each federated client to ensure the consistency constraints of the health risk assessment model structure and the coordination of personalized computing capabilities.

[0046] S44. Configure a local privacy protection policy set for each federated client. The client dynamically selects the differential privacy noise intensity from the policy set based on the sensitivity level of the collected data. The aggregation end coordinates and executes the privacy policy collaboration mechanism.

[0047] S45. Establish a communication channel between the client and the aggregation terminal to support the periodic synchronization of multiple types of data, including health risk assessment model parameters, privacy gradients, and structural recommendations, while configuring encryption strategies and fault-tolerant recovery mechanisms for the communication process.

[0048] S46. During each round of federated training, clients are allowed to propose suggestions for fine-tuning the health risk assessment model structure based on local training performance feedback. The aggregation end summarizes the structure proposals and conducts collaborative voting evaluation among multiple clients to determine whether to perform a global incremental update of the model structure.

[0049] Optionally, S5 specifically includes:

[0050] S51. Each federal client loads the initial structure and parameters of the health risk assessment model from the federal aggregation terminal and activates the local training task initialization process.

[0051] S52. Each federated client constructs model training input based on the local health dataset and the embedded features of the local medical knowledge graph. Based on the number of entities, connection density and semantic type coverage in the local medical knowledge graph, it dynamically adjusts the amount of data and training rounds used in this round of local training, forming a semantic-driven training scheduling mechanism for the local medical knowledge graph.

[0052] S53. Each federated client trains the health risk assessment model locally, and after training, identifies the key entities in the local medical knowledge graph and the feature channels associated with their upstream and downstream paths during the training process, extracts the corresponding health risk assessment model parameters, and uses them as the upload candidate subset of the gradient filtering mechanism guided by the local medical knowledge graph retention factor.

[0053] S54. Each federal client uploads the filtered health risk assessment model parameter set to the federal aggregation terminal through an encrypted transmission protocol, forming a semantically valid parameter submission unit with compressed structure.

[0054] S55. Before performing the aggregation of health risk assessment model parameters, the federated aggregation end analyzes the distribution characteristics of the local medical knowledge graph embedding features corresponding to the parameters uploaded by each client, obtains a semantic consistency score, and adjusts the weight of each client parameter in the aggregation process based on the score results, thus constructing a personalized knowledge embedding weight adjustment mechanism.

[0055] S56. After completing the aggregation operation of the health risk assessment model parameters, the federated aggregation end generates an updated health risk assessment model and simultaneously distributes the updated health risk assessment model to each federated client for the next round of personalized training and collaborative optimization.

[0056] Optionally, S6 specifically includes:

[0057] S61. Based on the updated health risk assessment model, load the complete parameters and structure configuration locally and initialize the fusion inference engine;

[0058] S62. Collect current real-time health data locally, including continuous multidimensional physiological indicator streams, interactive behavior data and dynamic environmental variables, and perform short-cycle time window processing to form a set of current health status features.

[0059] S63. Call the local medical knowledge graph reasoning engine, activate the entity path backtracking module based on key entity nodes in the current health data, construct a causal chain reasoning structure for decision support, and generate reasoning results with semantic interpretation capabilities.

[0060] S64. Input the current real-time health data and the reasoning results of the local medical knowledge graph into the updated health risk assessment model, introduce a health status shift analysis mechanism, and dynamically identify the risk of health mutations compared with historical status.

[0061] S65. Based on the output risk results, label the corresponding knowledge graph path and output the user's health risk level, possible disease category and two-layer intervention suggestions. The standard suggestions are generated based on the local medical knowledge graph rule base, and the personalized response suggestions are automatically generated by combining the model prediction results and the user's historical behavior preferences.

[0062] S66. Based on the health risk level, possible disease categories and intervention recommendations, a personalized intelligent monitoring decision is generated locally. At the same time, user behavior responses are received through a feedback interface for fine-tuning of the health risk assessment model structure and dynamic correction of the local medical knowledge graph.

[0063] The beneficial effects of this invention are:

[0064] The intelligent monitoring decision-making method based on knowledge graphs and federated learning proposed in this invention overcomes multiple technical bottlenecks in existing health monitoring systems in terms of data silos, privacy protection, model interpretability, and personalized intervention. It achieves efficient, reliable, and semantically enhanced health risk modeling and decision output for multi-terminal environments.

[0065] This invention fully integrates knowledge graph reasoning and federated modeling mechanisms in its methodological structure. First, it constructs individualized local medical knowledge graphs on each client. Through entity recognition, semantic relationship extraction, and embedding modeling, semantic vector representations are formed, which makes up for the shortcomings of traditional models in expressing complex causal relationships and diagnostic semantic paths between medical entities. Second, by introducing graph semantic-driven training scheduling, graph retention factor screening of uploaded parameters, and personalized aggregation weight mechanisms into the federated learning architecture, this invention enables the model to not only achieve privacy-preserving modeling but also to have cross-user semantic perception capabilities, significantly improving the generalization of the global model and the adaptability of the local model.

[0066] Furthermore, this invention constructs a semantically enhanced reasoning and decision-making generation mechanism, introducing a graph causal path chain backtracking module, a temporal state offset analysis unit, and a two-layer structure of intervention suggestions (i.e., graph rule-driven standard suggestions and model prediction-driven personalized response suggestions) into the health risk prediction process. This achieves an intelligent closed loop from data perception and semantic understanding to structured reasoning and personalized output. Compared with traditional static template matching or black-box model prediction, this method significantly improves the interpretability, responsiveness, and user trust of monitoring suggestions while maintaining a high degree of automation.

[0067] In summary, this invention offers the following advantages: First, it enables efficient collaborative modeling of multi-source health data without data leakage, meeting the industry's urgent need for data privacy protection; second, by integrating knowledge graphs with the entire model training process, the model possesses medical semantic understanding capabilities, outputting risk assessment results with more clinical logic; and third, through dynamic reasoning and personalized intervention mechanisms, it enhances the responsiveness and accuracy of interventions to sudden changes in health status, improving the overall intelligence, adaptability, and user acceptability of the system, making it suitable for various application scenarios such as smart healthcare, remote chronic disease management, and elderly health monitoring. Attached Figure Description

[0068] 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:

[0069] Figure 1 This is a flowchart of an intelligent guardianship decision-making method based on knowledge graphs and federated learning proposed in this invention;

[0070] Figure 2 This is a schematic diagram of the local medical knowledge graph construction and embedding modeling of the intelligent monitoring decision-making method based on knowledge graph and federated learning proposed in this invention. Detailed Implementation

[0071] 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.

[0072] refer to Figure 1 and Figure 2 A method for intelligent guardianship decision-making based on knowledge graphs and federated learning includes the following steps:

[0073] S1. Collect users' multi-source health data, preprocess the multi-source health data, and generate a health dataset;

[0074] S2. Based on the health dataset, construct a local medical knowledge graph, establish the relationship between entities through entity recognition and relation extraction, and perform local medical knowledge graph embedding modeling to form a knowledge representation vector;

[0075] S3. Construct a health risk assessment model, use graph neural networks to perform graph structure information propagation and feature extraction on knowledge representation vectors, and combine the health dataset to train and construct a health risk assessment model to complete the modeling of individual health status.

[0076] S4. Initialize the federated learning architecture, using multiple monitoring terminal devices as federated clients and the central coordination server as the federated aggregation terminal, distribute health risk assessment model parameters to each federated client, and configure communication and aggregation rules.

[0077] S5. Each federated client uses the health dataset and the embedded features of the local medical knowledge graph to train the health risk assessment model, and uploads the trained health risk assessment model parameters to the federated aggregation terminal, which performs parameter aggregation operation to generate and synchronously update the health risk assessment model.

[0078] S6. Based on the updated health risk assessment model, the system combines real-time health data with the reasoning results of the local medical knowledge graph to output the user's health risk level, possible disease categories, and suggested intervention measures, forming a personalized intelligent monitoring decision result.

[0079] S7. Incorporate user feedback data and newly added health data into the local medical knowledge graph and model for iterative updates, and participate in the joint optimization of the new round of health risk assessment model.

[0080] In this embodiment, the multi-source health data specifically includes physiological signals, medical records, examination and test results, medication records, and behavioral data, which are used to support the construction of a local medical knowledge graph and the training of a health risk assessment model.

[0081] In this embodiment, the preprocessing of multi-source health data specifically includes missing value imputation, standardization, structure transformation, and time alignment, which are used to improve the quality of multi-source health data and generate a health dataset that can be used for modeling.

[0082] In this embodiment, S2 specifically includes:

[0083] S21. Extract structured fields and unstructured text from the health dataset to construct an initial entity set E0 = {e1, e2, ..., e...}. n}, where e i Let i represent the i-th medical entity, where i = 1, 2, ..., n, and n is the total number of entities;

[0084] S22. Based on medical ontology and statistical models, construct a semantic relational database R0 = {r1, r2, ..., r...} m}, where r j Let j represent the j-th type of medical semantic relation, where j = 1, 2, ..., m, and m is the total number of medical semantic relation types. Through dependency parsing and co-occurrence network, entity pair relations are mined from E0 to obtain the local relation set R.

[0085] S23. Calculate the importance weight for each entity and filter out entities with high weight:

[0086] f(e i )=w1×freq(e i )+w2×coFreq(e i );

[0087] Where, f(e) i ) represents entity e i Importance weight, freq(e i ) is entity e i The number of times it appears in the health dataset, coFreq(e i ) is entity e i The number of times entities co-occur with diagnostic labels, w1 and w2 are empirical weighting coefficients; a threshold Θ is defined, and a set of high-weight entities E1 is constructed:

[0088] E1={e i ∈E0∣f(e i )>Θ};

[0089] S24. Generate a set of triples T = {(h, r, t) | h ∈ E1, r ∈ R, t ∈ E1} based on E1 and R, where h is the head entity, r is the edge type in the triple, and t is the tail entity. Remove low-confidence edges based on the confidence threshold δ to obtain the local medical knowledge graph G. local = (E1, R, T),

[0090] S25. For any entity e i Define the neighbor set N(e) i )={e j ∣(e i ,r,e j )∈T}, and incorporate the context into its own embedding:

[0091]

[0092] Among them, e j For the j-th entity, emb(e i ) is entity e i Initial embedding vector, emb(e j ) is entity ej The initial embedding vector, λ is the fusion coefficient, v(e i ) is entity e i Context-enhanced final embedding vector;

[0093] S26. Apply the TransE algorithm to the local medical knowledge graph, satisfying the condition that for any (h, r, t) ∈ T,

[0094] emb(h) + emb(r) ≈ emb(t);

[0095] Where emb(r) is the embedding vector of relation r, emb(h) is the embedding vector of head entity h, and emb(t) is the embedding vector of tail entity t;

[0096] S27, All entity vectors {v(e i )∣e i The knowledge representation vector K = {v(e ∈ E1} and the relation vector {emb(r) | r ∈ R} constitute the knowledge representation vector K = {v(e ∈ E1}}. i )}∪{emb(r)} is used as input for the health risk assessment model.

[0097] In this embodiment, S3 specifically includes:

[0098] S31. Obtain the formed knowledge representation vector set, construct a graph structure input according to the relationship between all entities using the embedding vectors of all entities, and construct a graph neural network input graph data;

[0099] S32. Construct a graph neural network structure containing multiple layers of graph neural computing units, and use a layer-by-layer propagation mechanism to diffuse and aggregate graph structure information on knowledge representation vectors, and extract a high-order representation that integrates structural features and semantic information.

[0100] S33. Extract time-series health data of individual users from the health dataset, including changes in physiological indicators, behavioral monitoring data and historical records, and construct data input samples for individual status modeling;

[0101] S34. The knowledge graph features extracted by the graph neural network are fused with the individual's health time series data to form a unified representation of the individual's health status.

[0102] S35. Construct a health risk assessment model based on fusion features, complete the learning of health risk assessment model parameters through a supervised training mechanism, and establish a mapping relationship between individual user health status and risk category;

[0103] S36. Complete the training process of the health risk assessment model to obtain the model structure and parameters that can be used for health status prediction, providing a basis for federal optimization and personalized monitoring decisions.

[0104] In this embodiment, S4 specifically includes:

[0105] S41. Establish a multi-terminal federated learning system architecture, set multiple terminal devices with health monitoring functions as federated clients, set the central coordination server as the federated aggregation terminal, and define the federated communication protocol and task scheduling process.

[0106] S42. Construct a unified structural framework for the health risk assessment model in the central coordination server, and set a health risk assessment model structure mapping table according to the computing resource level of each client. Load health risk assessment model substructures of corresponding complexity for devices with different performance to form a heterogeneous model adaptation mechanism.

[0107] S43. Distribute the initial health risk assessment model parameters, structural mapping relationships, and local training configuration files to each federated client to ensure the consistency constraints of the health risk assessment model structure and the coordination of personalized computing capabilities.

[0108] S44. Configure a local privacy protection policy set for each federated client. The client dynamically selects the differential privacy noise intensity from the policy set based on the sensitivity level of the collected data. The aggregation end coordinates and executes the privacy policy collaboration mechanism.

[0109] S45. Establish a communication channel between the client and the aggregation terminal to support the periodic synchronization of multiple types of data, including health risk assessment model parameters, privacy gradients, and structural recommendations, while configuring encryption strategies and fault-tolerant recovery mechanisms for the communication process.

[0110] S46. During each round of federated training, clients are allowed to propose suggestions for fine-tuning the health risk assessment model structure based on local training performance feedback. The aggregation end summarizes the structure proposals and conducts collaborative voting evaluation among multiple clients to determine whether to perform a global incremental update of the model structure.

[0111] In this embodiment, S5 specifically includes:

[0112] S51. Each federal client loads the initial structure and parameters of the health risk assessment model from the federal aggregation terminal and activates the local training task initialization process.

[0113] S52. Each federated client constructs model training input based on the local health dataset and the embedded features of the local medical knowledge graph. Based on the number of entities, connection density and semantic type coverage in the local medical knowledge graph, it dynamically adjusts the amount of data and training rounds used in this round of local training, forming a semantic-driven training scheduling mechanism for the local medical knowledge graph.

[0114] S53. Each federated client trains the health risk assessment model locally, and after training, identifies the key entities in the local medical knowledge graph and the feature channels associated with their upstream and downstream paths during the training process, extracts the corresponding health risk assessment model parameters, and uses them as the upload candidate subset of the gradient filtering mechanism guided by the local medical knowledge graph retention factor.

[0115] S54. Each federal client uploads the filtered health risk assessment model parameter set to the federal aggregation terminal through an encrypted transmission protocol, forming a semantically valid parameter submission unit with compressed structure.

[0116] S55. Before performing the aggregation of health risk assessment model parameters, the federated aggregation end analyzes the distribution characteristics of the local medical knowledge graph embedding features corresponding to the parameters uploaded by each client, obtains a semantic consistency score, and adjusts the weight of each client parameter in the aggregation process based on the score results, thus constructing a personalized knowledge embedding weight adjustment mechanism.

[0117] S56. After completing the aggregation operation of the health risk assessment model parameters, the federated aggregation end generates an updated health risk assessment model and simultaneously distributes the updated health risk assessment model to each federated client for the next round of personalized training and collaborative optimization.

[0118] In this embodiment, S6 specifically includes:

[0119] S61. Based on the updated health risk assessment model, load the complete parameters and structure configuration locally and initialize the fusion inference engine;

[0120] S62. Collect current real-time health data locally, including continuous multidimensional physiological indicator streams, interactive behavior data and dynamic environmental variables, and perform short-cycle time window processing to form a set of current health status features.

[0121] S63. Call the local medical knowledge graph reasoning engine, activate the entity path backtracking module based on key entity nodes in the current health data, construct a causal chain reasoning structure for decision support, and generate reasoning results with semantic interpretation capabilities.

[0122] S64. Input the current real-time health data and the reasoning results of the local medical knowledge graph into the updated health risk assessment model, introduce a health status shift analysis mechanism, and dynamically identify the risk of health mutations compared with historical status.

[0123] S65. Based on the output risk results, label the corresponding knowledge graph path and output the user's health risk level, possible disease category and two-layer intervention suggestions. The standard suggestions are generated based on the local medical knowledge graph rule base, and the personalized response suggestions are automatically generated by combining the model prediction results and the user's historical behavior preferences.

[0124] S66. Based on the health risk level, possible disease categories and intervention recommendations, a personalized intelligent monitoring decision is generated locally. At the same time, user behavior responses are received through a feedback interface for fine-tuning of the health risk assessment model structure and dynamic correction of the local medical knowledge graph.

[0125] Example 1:

[0126] To verify the feasibility of this invention in practice, it was applied to a pilot deployment of a remote elderly health management system in a demonstration community integrating medical and elderly care for six months. The method of this invention was integrated into a community family doctor platform, wearable health terminals, smart mattress monitoring devices, and a mobile terminal app for the elderly, realizing distributed health data collection, graph reasoning, and risk prediction decision output throughout the entire process. Each pilot participant's local terminal independently constructed a medical knowledge graph, including entities such as diseases, symptoms, physiological indicators, and behavioral habits, as well as the relationships between these entities. For example, some knowledge graph entity nodes are "hyperglycemia," "blurred vision," "diabetic retinopathy," and "insulin intervention," with relationships such as "trigger," "symptom indication," and "treatment plan." The graph structure is dynamically updated based on the chronic disease standards and guidelines issued by the National Health Commission and individual physical characteristics.

[0127] During implementation, each terminal trains a health risk assessment model based on its local knowledge graph and health data (such as blood pressure, heart rate, steps, and sleep quality). Through federated learning, the local model parameters are synchronized to the aggregation server without uploading the original data. The system automatically completes federated aggregation every 7 days and adaptively adjusts the training frequency and uploaded content of each terminal based on the complexity of the device knowledge graph, achieving differentiated optimization and model self-learning evolution. The updated global model is then sent back to all terminals and participates in personalized inference and intervention generation.

[0128] For example, after a terminal device detects a user's elevated nighttime heart rate and morning blood pressure for three consecutive days, it triggers a knowledge graph reasoning module to identify potential pathways for hypertension worsening. Without disturbing the user, it outputs non-emergency intervention suggestions, such as "suggest getting up slowly in the morning," "adjusting the timing of nighttime medication," and "scheduling a doctor's follow-up appointment this week." Compared to traditional monitoring systems that only issue fixed warnings when thresholds are exceeded, the suggestions provided by this invention are more logically interpretable and adaptable.

[0129] After six months of application, the system has covered 126 users with an average age of 69.3 years, recording over 860,000 pieces of valid health data. The system issued 121 remote health alerts, provided 853 intelligent intervention suggestions, and users proactively adopted the suggestions 629 times, representing 73.8% of the total. More importantly, the system significantly improved the early identification rate of health risks, reduced the frequency of emergencies, and demonstrated significant advantages in improving system response speed and intervention accuracy.

[0130] Table 1 Comparison of monitoring effects between the method of the present invention and traditional methods

[0131]

[0132] According to the data in Table 1, it can be seen that the "intelligent monitoring decision-making method based on knowledge graph and federated learning" proposed in this invention has significant advantages over the traditional centralized monitoring system in several key performance indicators, fully demonstrating its practicality and technical superiority in actual health management scenarios.

[0133] First, in terms of "real-time early warning accuracy," the system of this invention achieved 93.1%, significantly higher than the 78.6% of traditional systems. This demonstrates that by introducing knowledge graphs for semantic modeling and risk path reasoning, the system's accuracy in identifying complex health events can be effectively improved, especially playing a crucial role in the early identification of high-risk states for chronic diseases such as cardiovascular diseases.

[0134] Secondly, in terms of the "personalized matching degree of intervention recommendations," an indicator reflecting the system's intelligence level, the system of this invention achieved 89.3%, which is 27.9 percentage points higher than traditional systems. This improvement is attributed to the fusion and inference capabilities of knowledge graph reasoning and individual user health status, ensuring that the output intervention recommendations are not only closely related to the disease but also possess personalized targeting and contextual rationality, greatly enhancing user adoption willingness and intervention effectiveness.

[0135] The percentage of users adopting intervention suggestions increased from 58.2% in traditional methods to 73.8% in the system of this invention, indicating a significant improvement in users' understanding and trust in the suggestions. This also demonstrates that the invention has achieved a breakthrough in terms of "explainability" and "user participation." This improvement is of great significance for the long-term management of chronic diseases in the community and helps to promote the transformation from a "passive medical care" model to a "proactive health" model.

[0136] In terms of early warning response speed, the "average risk identification advance time" has been shortened from 64 minutes using traditional methods to 18 minutes, significantly reducing the time required to handle emergencies. This not only helps reduce the probability of acute health risks but also demonstrates the synergistic effect of the "knowledge graph + model prediction + state shift analysis" triple mechanism in this invention, achieving dynamic and continuous assessment of health status.

[0137] In addition, the "conversion rate triggered by doctor follow-up visits" increased from 33.5% to 47.2%, indicating that the system can more accurately identify cases that require further clinical intervention, enhance the intervention efficiency of family doctors and community doctors, and is a good supplement to the existing chronic disease service system.

[0138] Finally, regarding the "data privacy leakage risk" metric, the system of this invention, through federated learning technology, ensures that the original data does not leave the local machine, only exchanging parameters, effectively reducing the risk of privacy leakage. In today's increasingly stringent data security and compliance environment, this has extremely high practical value and promising prospects for widespread application.

[0139] In summary, this comparison table comprehensively verifies the leading advantages of the method of the present invention in terms of accuracy, personalization, response speed, user adoption rate, and privacy protection, demonstrating clear engineering feasibility and broad practical significance.

[0140] 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 guardianship decision-making method based on knowledge graphs and federated learning, characterized in that, Includes the following steps: S1. Collect users' multi-source health data, preprocess the multi-source health data, and generate a health dataset; S2. Based on the health dataset, construct a local medical knowledge graph, establish the relationship between entities through entity recognition and relation extraction, and perform local medical knowledge graph embedding modeling to form a knowledge representation vector; S3. Construct a health risk assessment model, use graph neural networks to perform graph structure information propagation and feature extraction on knowledge representation vectors, and combine the health dataset to train and construct a health risk assessment model to complete the modeling of individual health status. S4. Initialize the federated learning architecture, using multiple monitoring terminal devices as federated clients and the central coordination server as the federated aggregation terminal, distribute health risk assessment model parameters to each federated client, and configure communication and aggregation rules. S5. Each federated client uses the health dataset and the embedded features of the local medical knowledge graph to train the health risk assessment model, and uploads the trained health risk assessment model parameters to the federated aggregation terminal, which performs parameter aggregation operation to generate and synchronously update the health risk assessment model. S6. Based on the updated health risk assessment model, the system combines real-time health data with the reasoning results of the local medical knowledge graph to output the user's health risk level, possible disease categories, and suggested intervention measures, forming a personalized intelligent monitoring decision result. S7. Incorporate user feedback data and newly added health data into the local medical knowledge graph and model for iterative updates, and participate in the joint optimization of the new round of health risk assessment model.

2. The intelligent guardianship decision-making method based on knowledge graphs and federated learning according to claim 1, characterized in that, The multi-source health data specifically includes physiological signals, medical records, examination and test results, medication records, and behavioral data, which are used to support the construction of a local medical knowledge graph and the training of a health risk assessment model.

3. The intelligent guardianship decision-making method based on knowledge graphs and federated learning according to claim 1, characterized in that, The preprocessing of multi-source health data specifically includes missing value imputation, standardization, structural transformation, and time alignment, which are used to improve the quality of multi-source health data and generate health datasets that can be used for modeling.

4. The intelligent guardianship decision-making method based on knowledge graphs and federated learning according to claim 1, characterized in that, S2 specifically includes: S21. Extract structured fields and unstructured text from the health dataset to construct an initial entity set E0 = {e1, e2, ..., e...}. n }, where e i Let i represent the i-th medical entity, where i = 1, 2, ..., n, and n is the total number of entities; S22. Based on medical ontology and statistical models, construct a semantic relational database R0 = {r1, r2, ..., r...} m }, where r j Let j represent the j-th type of medical semantic relation, where j = 1, 2, ..., m, and m is the total number of medical semantic relation types. Through dependency parsing and co-occurrence network, entity pair relations are mined from E0 to obtain the local relation set R. S23. Calculate the importance weight for each entity and filter out entities with high weight: f(e i )=w1×freq(e i )+w2×coFreq(e i ); Where, f(e) i ) represents entity e i Importance weight, freq(e i ) is entity e i The number of times it appears in the health dataset, coFreq(e i ) is entity e i The number of times entities co-occur with diagnostic labels, w1 and w2 are empirical weighting coefficients; a threshold Θ is defined, and a set of high-weight entities E1 is constructed: E1 = {e i ∈E0∣f(e i )>Θ}; S24. Generate a set of triples T = {(h, r, t) | h ∈ E1, r ∈ R, t ∈ E1} based on E1 and R, where h is the head entity, r is the edge type in the triple, and t is the tail entity. Remove low-confidence edges based on the confidence threshold δ to obtain the local medical knowledge graph G. local = (E1, R, T), S25. For any entity e i Define the neighbor set N(e) i )={e j ∣(e i ,r,e j )∈T}, and incorporate the context into its own embedding: Among them, e j For the j-th entity, emb(e i ) is entity e i Initial embedding vector, emb(e j ) is entity e j The initial embedding vector, λ is the fusion coefficient, v(e i ) is entity e i Context-enhanced final embedding vector; S26. Apply the TransE algorithm to the local medical knowledge graph, satisfying the condition that for any (h, r, t) ∈ T, emb(h) + emb(r) ≈ emb(t); Where emb(r) is the embedding vector of relation r, emb(h) is the embedding vector of head entity h, and emb(t) is the embedding vector of tail entity t; S27, All entity vectors {v(e i )∣e i The knowledge representation vector K = {v(e ∈ E1} and the relation vector {emb(r) | r ∈ R} constitute the knowledge representation vector K = {v(e ∈ E1}}. i )}∪{emb(r)} is used as input for the health risk assessment model.

5. The intelligent guardianship decision-making method based on knowledge graphs and federated learning according to claim 1, characterized in that, S3 specifically includes: S31. Obtain the formed knowledge representation vector set, construct a graph structure input according to the relationship between all entities using the embedding vectors of all entities, and construct a graph neural network input graph data; S32. Construct a graph neural network structure containing multiple layers of graph neural computing units, and use a layer-by-layer propagation mechanism to diffuse and aggregate graph structure information on knowledge representation vectors, and extract a high-order representation that integrates structural features and semantic information. S33. Extract time-series health data of individual users from the health dataset, including changes in physiological indicators, behavioral monitoring data and historical records, and construct data input samples for individual status modeling; S34. The knowledge graph features extracted by the graph neural network are fused with the individual's health time series data to form a unified representation of the individual's health status. S35. Construct a health risk assessment model based on fusion features, complete the learning of health risk assessment model parameters through a supervised training mechanism, and establish a mapping relationship between individual user health status and risk category; S36. Complete the training process of the health risk assessment model to obtain the model structure and parameters that can be used for health status prediction, providing a basis for federal optimization and personalized monitoring decisions.

6. The intelligent guardianship decision-making method based on knowledge graphs and federated learning according to claim 1, characterized in that, S4 specifically includes: S41. Establish a multi-terminal federated learning system architecture, set multiple terminal devices with health monitoring functions as federated clients, set the central coordination server as the federated aggregation terminal, and define the federated communication protocol and task scheduling process. S42. Construct a unified structural framework for the health risk assessment model in the central coordination server, and set a health risk assessment model structure mapping table according to the computing resource level of each client. Load health risk assessment model substructures of corresponding complexity for devices with different performance to form a heterogeneous model adaptation mechanism. S43. Distribute the initial health risk assessment model parameters, structural mapping relationships, and local training configuration files to each federated client to ensure the consistency constraints of the health risk assessment model structure and the coordination of personalized computing capabilities. S44. Configure a local privacy protection policy set for each federated client. The client dynamically selects the differential privacy noise intensity from the policy set based on the sensitivity level of the collected data. The aggregation end coordinates and executes the privacy policy collaboration mechanism. S45. Establish a communication channel between the client and the aggregation terminal to support the periodic synchronization of multiple types of data, including health risk assessment model parameters, privacy gradients, and structural recommendations, while configuring encryption strategies and fault-tolerant recovery mechanisms for the communication process. S46. During each round of federated training, clients are allowed to propose suggestions for fine-tuning the health risk assessment model structure based on local training performance feedback. The aggregation end summarizes the structure proposals and conducts collaborative voting evaluation among multiple clients to determine whether to perform a global incremental update of the model structure.

7. The intelligent guardianship decision-making method based on knowledge graphs and federated learning according to claim 1, characterized in that, S5 specifically includes: S51. Each federal client loads the initial structure and parameters of the health risk assessment model from the federal aggregation terminal and activates the local training task initialization process. S52. Each federated client constructs model training input based on the local health dataset and the embedded features of the local medical knowledge graph. Based on the number of entities, connection density and semantic type coverage in the local medical knowledge graph, it dynamically adjusts the amount of data and training rounds used in this round of local training, forming a semantic-driven training scheduling mechanism for the local medical knowledge graph. S53. Each federated client trains the health risk assessment model locally, and after training, identifies the key entities in the local medical knowledge graph and the feature channels associated with their upstream and downstream paths during the training process, extracts the corresponding health risk assessment model parameters, and uses them as the upload candidate subset of the gradient filtering mechanism guided by the local medical knowledge graph retention factor. S54. Each federal client uploads the filtered health risk assessment model parameter set to the federal aggregation terminal through an encrypted transmission protocol, forming a semantically valid parameter submission unit with compressed structure. S55. Before performing the aggregation of health risk assessment model parameters, the federated aggregation end analyzes the distribution characteristics of the local medical knowledge graph embedding features corresponding to the parameters uploaded by each client, obtains a semantic consistency score, and adjusts the weight of each client parameter in the aggregation process based on the score results, thus constructing a personalized knowledge embedding weight adjustment mechanism. S56. After completing the aggregation operation of the health risk assessment model parameters, the federated aggregation end generates an updated health risk assessment model and simultaneously distributes the updated health risk assessment model to each federated client for the next round of personalized training and collaborative optimization.

8. The intelligent guardianship decision-making method based on knowledge graphs and federated learning according to claim 1, characterized in that, S6 specifically includes: S61. Based on the updated health risk assessment model, load the complete parameters and structure configuration locally and initialize the fusion inference engine; S62. Collect current real-time health data locally, including continuous multidimensional physiological indicator streams, interactive behavior data and dynamic environmental variables, and perform short-cycle time window processing to form a set of current health status features. S63. Call the local medical knowledge graph reasoning engine, activate the entity path backtracking module based on key entity nodes in the current health data, construct a causal chain reasoning structure for decision support, and generate reasoning results with semantic interpretation capabilities. S64. Input the current real-time health data and the reasoning results of the local medical knowledge graph into the updated health risk assessment model, introduce a health status shift analysis mechanism, and dynamically identify the risk of health mutations compared with historical status. S65. Based on the output risk results, label the corresponding knowledge graph path and output the user's health risk level, possible disease category and two-layer intervention suggestions. The standard suggestions are generated based on the local medical knowledge graph rule base, and the personalized response suggestions are automatically generated by combining the model prediction results and the user's historical behavior preferences. S66. Based on the health risk level, possible disease categories and intervention recommendations, a personalized intelligent monitoring decision is generated locally. At the same time, user behavior responses are received through a feedback interface for fine-tuning of the health risk assessment model structure and dynamic correction of the local medical knowledge graph.

Images