A health management method and device based on atlas reasoning and a computer readable storage medium

Abstract

This application belongs to the interdisciplinary field of graph databases, artificial intelligence, and reinforcement learning. It provides a health management method, device, and computer-readable storage medium based on graph reasoning. By jointly extracting entities and relations from unstructured medical text and eliminating conflicting edges based on confidence rules, it constructs a target evidence-based knowledge graph with a deterministic pathfinding structure. By quantifying features of heterogeneous multi-dimensional health data to construct a set of state features, it determines the starting query node for graph reasoning, realizing the generation and distribution of the initial plan. It uses the time-series interactive data asynchronously uploaded by the terminal device to calculate the task deviation as a reward signal for reinforcement learning, driving the model to adaptively adjust the output action instruction vector. While effectively updating the timestamp and execution frequency parameters of the management plan, the underlying algorithm architecture strictly constrains the adjustment range of strategy iteration, balancing the intelligent optimization efficiency of the health management plan with the security of execution.

Description

Technical Field

[0001] This invention relates to a health management method, device, and computer-readable storage medium based on graph reasoning, belonging to the interdisciplinary technical field of graph databases, artificial intelligence, and reinforcement learning, with main classification numbers G06N5 / 02, G06N20 / 00, or G06F40 / 20. Background Technology

[0002] With the advancement of the tiered medical system, health management systems for chronic disease patients during their discharge transition period are becoming increasingly widespread. Chronic diseases such as cardiovascular disease, diabetes, and chronic obstructive pulmonary disease are characterized by long disease courses and persistent symptoms, requiring long-term monitoring and intervention. During the discharge transition period, chronic disease patients are at high risk of relapse or even unexpected readmissions due to the absence of close 24-hour monitoring from in-hospital medical staff, coupled with insufficient understanding of their disease and poor adherence to treatment.

[0003] Currently, routine outpatient health management largely relies on paper-based medical orders or static health education manuals issued upon discharge, or on consultations between patients' families and hospitals via telephone or other remote means. This traditional management model has significant limitations. On the one hand, intervention plans are often generic templates, unable to accurately match the multi-dimensional conditions of individual patients, including complications, home environment, and subjective psychological state. On the other hand, health management is a dynamic process that changes over time, and static initial plans cannot be timely and safely adapted to the patient's feedback during home care. For example, if a patient experiences sudden fluctuations in vital signs or a decline in compliance, caregivers will find it difficult to make quick and accurate adjustments. Although some advanced medical information systems attempt to introduce clinical guidelines or expert consensus to assist decision-making, this evidence-based medicine knowledge is mostly in the form of unstructured natural language text, making it difficult for computer systems to automatically parse and execute it.

[0004] When transforming the aforementioned complex and evolving health management demands into automated computer software processing systems, existing data processing technologies and underlying algorithm architectures reveal the following fundamental technical deficiencies: First, in processing unstructured medical text, there is a lack of ability to accurately convert natural language with triggering conditions and logically mutually exclusive relationships into a traversable directed graph data structure; second, there is a lack of unified dimensional fusion of multimodal and heterogeneous objective physiological characteristics and subjective state scalars, especially the need to construct a quantitative mechanism for precise starting query nodes; finally, in the parameter adjustment stage, existing systems mostly rely on simple rule thresholds or conventional machine learning algorithms lacking safety boundary constraints, making it difficult to strictly limit the adjustment range of single intervention parameters to avoid medical safety risks while ensuring continuous optimization of the strategy. Summary of the Invention

[0005] To address the shortcomings of the prior art, the present invention aims to provide a health management method, device, and computer-readable storage medium based on graph reasoning.

[0006] According to an embodiment of the present invention, a first embodiment is provided: a health management method based on graph reasoning, applied to a health management system for the discharge transition period of patients with chronic diseases, the method comprising: Acquire medical text character sequences and multi-dimensional health data of the target object; Entity recognition and relation extraction are performed on the medical text character sequence to construct an initial graph data structure containing feature nodes, instruction nodes and directed association edges. Conflicting association edges that cause mutual exclusion are eliminated based on confidence rules, and a target evidence-based knowledge graph is generated. The multi-dimensional health data is subjected to feature quantization processing to construct a set of state features; The set of state features is mapped to the starting query node. A graph path search algorithm is executed in the target evidence-based knowledge graph to obtain the target instruction node. An initial health management plan containing timestamp and execution frequency parameters is generated based on the target instruction node and sent to the terminal device. The system receives time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, updates the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculates the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback. Using the current environmental state as input data and the execution feedback as a reward signal, an action instruction vector is output through a reinforcement learning model. The timestamp and execution frequency parameters of the initial health management plan are then corrected based on the action instruction vector to generate an updated health management plan.

[0007] Furthermore, the medical text character sequence includes unstructured natural language sequences extracted from domain protocols, system reviews, or expert consensus; the multidimensional health data includes static attribute vectors characterizing the baseline state of the target object, and subjective state scalars obtained through evaluation.

[0008] Further, the steps of performing entity recognition and relation extraction on the medical text character sequence, constructing an initial graph data structure containing feature nodes, instruction nodes, and directed association edges, and eliminating mutually exclusive conflicting association edges based on confidence rules to generate the target evidence-based knowledge graph include: extracting the feature nodes representing the baseline state of the object in the medical text character sequence through a pre-trained entity relation extraction model, and extracting the instruction nodes representing intervention actions and embedding the execution frequency parameter and the timestamp; establishing the directed association edges between the feature nodes and the instruction nodes with conditional triggering relationships based on semantic dependency logic, and aggregating them to form the initial graph data structure; performing graph path traversal in the initial graph data structure, and determining the existence of conflicting association edges when multiple directed association edges originating from the same feature node point to logically mutually exclusive instruction nodes; extracting the source attribute features of the corresponding medical text character sequence that generated the conflicting association edges and substituting them into the confidence rules for comprehensive evaluation, eliminating conflicting association edges with evaluation scores lower than the comparison threshold, and retaining uniquely connected association edges to generate the target evidence-based knowledge graph.

[0009] Further, the step of extracting the source attribute features of the corresponding medical text character sequences that generate the conflicting association edges, substituting them into the confidence rule for comprehensive evaluation, eliminating conflicting association edges with evaluation scores below the comparison threshold, and retaining uniquely connected association edges to generate the target evidence-based knowledge graph includes: obtaining the source attribute features of the corresponding medical text character sequences that generate each conflicting association edge, wherein the source attribute features include evidence rating quantification value and literature publication time; using the evidence rating quantification value and the literature publication time as input variables, substituting them into the confidence decay function included in the confidence rule, and calculating the comprehensive evaluation score of each conflicting association edge, wherein the confidence decay function is expressed as: C=α·f(E level )+β·e -λ(Tn-Tp) ; Where C is the comprehensive evaluation score, f(E) level T is the mapping function for the quantitative value of the evidence rating. p T represents the publication time of the aforementioned document. n Let λ be the current system time, α be the time decay constant, α be the first weight coefficient, and β be the second weight coefficient. Conflicting edges with a comprehensive evaluation score lower than the comparison threshold are removed. When there are multiple conflicting edges with scores not lower than the comparison threshold, the edge with the highest score is retained as the uniquely connected edge to generate the target evidence-based knowledge graph with a uniquely determined traversal path.

[0010] Further, the step of performing feature quantization processing on the multi-dimensional health data to construct a state feature set includes: obtaining the static attribute vector and the subjective state scalar in the multi-dimensional health data; performing dimensional unification and standardization mapping processing on the discrete and continuous feature values ​​in the static attribute vector and the subjective state scalar to generate a standardized feature matrix; constructing a feature evaluation judgment matrix based on the analytic hierarchy process; performing a consistency check on the standardized feature matrix; calculating and obtaining the weight factors of each dimension feature in the standardized feature matrix; using the weight factors to perform weighted fusion calculation on the standardized feature matrix to generate a quantified feature vector containing a discharge preparation metric score and a baseline risk label; and using the quantified feature vector as the state feature set.

[0011] Further, the steps of mapping the state feature set to a starting query node, executing a graph path search algorithm in the target evidence-based knowledge graph to obtain a target instruction node, generating an initial health management plan containing timestamps and execution frequency parameters based on the target instruction node, and sending it to the terminal device include: extracting the baseline risk label from the state feature set, using the baseline risk label as a graph retrieval entry point, and mapping it to the starting query node in the target evidence-based knowledge graph; starting from the starting query node, executing the graph path search algorithm in the target evidence-based knowledge graph, traversing along the associated edges with uniquely determined traversal paths, hitting and obtaining the logically connected target instruction node; parsing the target instruction node, extracting the embedded intervention task and the timestamp and execution frequency parameters bound to the intervention task; serializing the intervention task into discrete time nodes according to a preset discharge transition period timeline, encapsulating the arranged intervention task, timestamp, and execution frequency parameters into structured data, generating the initial health management plan, and establishing a communication link to send it to the terminal device.

[0012] Further, the steps of receiving the time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, updating the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculating the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback include: extracting the actual occurrence timestamp and actual interaction frequency, as well as dynamic physiological change features contained in the time-series interaction data; substituting the dynamic physiological change features into a preset feature evaluation judgment matrix for incremental calculation to refresh the quantified feature vector in the state feature set and obtain the current environmental state characterizing the target object; comparing the actual occurrence timestamp and actual interaction frequency with the timestamp and execution frequency parameters contained in the initial health management plan respectively to generate the multi-dimensional task deviation; and using a preset evaluation function to perform weighted summation and normalization processing on the multi-dimensional task deviation, and outputting a scalar value as the quantified execution feedback.

[0013] Further, the step of using the current environmental state as input data, the execution feedback as a reward signal, outputting an action instruction vector through a reinforcement learning model, and modifying the timestamp and execution frequency parameters of the initial health management plan based on the action instruction vector to generate an updated health management plan includes: constructing the reinforcement learning model based on a proximal policy optimization algorithm; setting the state space of the reinforcement learning model as a sequence tensor containing the current environmental state at multiple consecutive time steps; setting the action space of the reinforcement learning model as a continuous value matrix adjusting the timestamp offset and the increase / decrease of the execution frequency parameter; and constructing a composite reward function to obtain the reward signal, wherein the composite reward function includes an immediate reward value negatively correlated with the quantized execution feedback. The system generates a delayed reward discount value that is positively correlated with the gradient of risk reduction in the current environmental state; it inputs the sequence tensor into the reinforcement learning model, outputs an action probability distribution constrained by a preset truncation function boundary through the Actor network to constrain the magnitude of a single policy adjustment; it evaluates the value baseline of the current input data through the Critic network, and calculates the advantage function by combining the immediate reward value and the delayed reward discount value, using the advantage function to drive model gradient descent to update network parameters; it samples the action instruction vector from the converged action probability distribution, extracts the offset and the increase / decrease in the action instruction vector, and numerically overwrites the timestamp and execution frequency parameters in the initial health management plan to generate the updated health management plan.

[0014] Furthermore, the step of receiving the time-series interactive data asynchronously uploaded by the terminal device during the execution of the initial health management plan further includes: extracting real-time physiological values ​​from the time-series interactive data, comparing the real-time physiological values ​​with a preset set of safety baseline thresholds; when it is determined that the real-time physiological values ​​exceed the safety boundary in the set of safety baseline thresholds, suspending the input process to the reinforcement learning model and triggering a hierarchical routing early warning mechanism; calculating the deviation of the real-time physiological values ​​to match the corresponding early warning severity level, and executing a corresponding communication scheduling strategy based on the early warning severity level, wherein the communication scheduling strategy includes: generating autonomous intervention instructions and pushing them to the terminal device, issuing online collaborative prompts to the associated responsible business end, or directly establishing a multidisciplinary emergency dispatch communication link.

[0015] Furthermore, the method further includes the steps of: within a preset long-term time window, acquiring the quantified execution feedback generated after multiple target objects execute the same target instruction node; aggregating and calculating the multiple quantified execution feedbacks to obtain a group execution performance index; when it is determined that the group execution performance index is lower than a preset expected threshold and the duration reaches the evaluation period, generating a negative penalty factor corresponding to the target instruction node; asynchronously feeding the negative penalty factor back to the target evidence-based knowledge graph, adaptively adjusting the weight parameters of the associated edges pointing to the target instruction node in the confidence rule, so as to dynamically reconstruct the local topology of the target evidence-based knowledge graph.

[0016] According to an embodiment of the present invention, utilizing the graph-based reasoning-based health management method in the first solution provided by the present invention, a second solution is provided as follows: A graph-based reasoning-based health management device, used to apply any of the above-mentioned graph-based reasoning-based health management methods, the device comprising: The data acquisition module is used to acquire medical text character sequences and multi-dimensional health data of the target object; The graph construction module is used to perform entity recognition and relation extraction on the medical text character sequence, construct an initial graph data structure containing feature nodes, instruction nodes and directed association edges, eliminate conflicting association edges that cause mutual exclusion based on confidence rules, and generate the target evidence-based knowledge graph. The feature quantization module is used to perform feature quantization processing on the multi-dimensional health data to construct a set of state features; The graph reasoning module is used to map the set of state features to the starting query node, execute a graph path search algorithm in the target evidence-based knowledge graph to obtain the target instruction node, generate an initial health management plan containing timestamp and execution frequency parameters based on the target instruction node, and send it to the terminal device. The feedback update module is used to receive the time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, update the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculate the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback. The reinforcement learning module is used to take the current environmental state as input data, the execution feedback as a reward signal, output an action instruction vector through the reinforcement learning model, and modify the timestamp and execution frequency parameters of the initial health management plan according to the action instruction vector to generate an updated health management plan.

[0017] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: The process involves acquiring medical text character sequences and multi-dimensional health data of the target object; performing entity recognition and relation extraction on the medical text character sequences to construct an initial graph data structure containing feature nodes, instruction nodes, and directed association edges; eliminating mutually exclusive conflicting association edges based on confidence rules to generate a target evidence-based knowledge graph; performing feature quantization processing on the multi-dimensional health data to construct a state feature set; mapping the state feature set to starting query nodes; executing a graph path search algorithm in the target evidence-based knowledge graph to obtain target instruction nodes; and generating an initial health management plan containing timestamps and execution frequency parameters based on the target instruction nodes. The system plans and distributes the data to the terminal device; it receives the time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, updates the state feature set using the time-series interaction data to characterize the current environmental state of the target object, calculates the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback; it uses the current environmental state as input data, the execution feedback as a reward signal, outputs an action instruction vector through a reinforcement learning model, and corrects the timestamp and execution frequency parameters of the initial health management plan based on the action instruction vector to generate an updated health management plan.

[0018] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps: The process involves acquiring medical text character sequences and multi-dimensional health data of the target object; performing entity recognition and relation extraction on the medical text character sequences to construct an initial graph data structure containing feature nodes, instruction nodes, and directed association edges; eliminating mutually exclusive conflicting association edges based on confidence rules to generate a target evidence-based knowledge graph; performing feature quantization processing on the multi-dimensional health data to construct a state feature set; mapping the state feature set to starting query nodes; executing a graph path search algorithm in the target evidence-based knowledge graph to obtain target instruction nodes; and generating an initial health management plan containing timestamps and execution frequency parameters based on the target instruction nodes. The system plans and distributes the data to the terminal device; it receives the time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, updates the state feature set using the time-series interaction data to characterize the current environmental state of the target object, calculates the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback; it uses the current environmental state as input data, the execution feedback as a reward signal, outputs an action instruction vector through a reinforcement learning model, and corrects the timestamp and execution frequency parameters of the initial health management plan based on the action instruction vector to generate an updated health management plan.

[0019] Compared with the prior art, the unique advantages of the technical solution provided in this application are as follows: This invention successfully constructs a target evidence-based knowledge graph with a deterministic pathfinding structure by jointly extracting entities and relations from unstructured medical text and eliminating conflicting edges based on confidence rules. It also determines the starting query node for graph inference by quantifying features of heterogeneous, multi-dimensional health data to construct a set of state features, enabling highly personalized initial plan generation and distribution. Furthermore, this invention constructs a data-driven optimization closed-loop process, using asynchronously uploaded time-series interactive data from terminal devices to calculate task deviation as a reward signal for reinforcement learning. This drives the model to adaptively adjust the output action command vector, effectively updating the management plan timestamp and execution frequency parameters while strictly constraining the adjustment range of strategy iterations through the underlying algorithm architecture, thus balancing the intelligent optimization efficiency and execution security of health management solutions.

[0020] The solution of this invention overcomes the long-held technical bias in the field of health management system applications that the unbounded exploration characteristics of reinforcement learning are prone to causing loss of control risks, and therefore cannot be safely applied to the modification of underlying medical and health instructions. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] in: Figure 1 This is a flowchart of a health management method based on graph reasoning in one embodiment; Figure 2 This is an interactive flowchart of the initial health management plan generation stage of a graph-based health management method in one embodiment; Figure 3 This is an interactive flowchart of the health management plan update phase of a graph-based health management method in one embodiment; Figure 4 This is a flowchart illustrating the construction and conflict resolution of a target evidence-based knowledge graph in one embodiment. Figure 5 This is a flowchart illustrating the execution process of a graph path search algorithm in one embodiment. Figure 6 This is a schematic diagram of the architecture and data processing flow of a medical text entity and relation joint extraction model in one embodiment; Figure 7 This is a schematic diagram of the underlying architecture and data flow of a reinforcement learning model in one embodiment; Figure 8 This is a logical structure block diagram of a health management device based on graph reasoning in one embodiment; Figure 9 This is a structural block diagram of a computer device used to perform the graph-based reasoning health management method in one embodiment. Detailed Implementation

[0023] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] Example 1 With the advancement of the tiered medical system, health management systems for chronic disease patients during their discharge transition period are becoming increasingly widespread. Chronic diseases such as cardiovascular disease, diabetes, and chronic obstructive pulmonary disease are characterized by long disease courses and persistent symptoms, requiring long-term monitoring and intervention. During the discharge transition period, chronic disease patients are at high risk of relapse or even unexpected readmissions due to the absence of close 24-hour monitoring from in-hospital medical staff, coupled with insufficient understanding of their disease and poor adherence to treatment. Currently, routine outpatient health management largely relies on paper-based medical orders or static health education manuals issued upon discharge, or on consultations between patients' families and hospitals via telephone or other remote means. This traditional management model has significant limitations. On the one hand, intervention plans are often generic templates, unable to accurately match the patient's individual, multi-dimensional baseline status, including complications, home environment, and subjective psychological state. On the other hand, health management is a dynamic process that changes over time. Static initial plans cannot be timely and safely adaptively adjusted based on the patient's feedback during home care. For example, if a patient experiences sudden fluctuations in vital signs or a decline in compliance, caregivers find it difficult to make quick and accurate adjustments. Although some advanced medical information systems attempt to introduce clinical guidelines or expert consensus to assist decision-making, this evidence-based medicine knowledge often exists in unstructured natural language text, making it difficult for computer systems to automatically parse and execute it.

[0025] See Figure 1 , Figure 2 and Figure 3 As shown, this embodiment provides a health management method based on graph reasoning, applied to the health management system for the discharge transition period of patients with chronic diseases. The method includes: S1. Obtain medical text character sequences and multi-dimensional health data of the target object; In step S1, unstructured "Guidelines for the Diagnosis and Treatment of Chronic Diseases" and expert consensus are obtained from authoritative medical databases as medical text character sequences; at the same time, continuous physiological signs and discrete subjective assessments of patients are collected as multi-dimensional health data through IoT devices and patient terminals.

[0026] S2. Perform entity recognition and relation extraction on the medical text character sequence, construct an initial graph data structure containing feature nodes, instruction nodes and directed association edges, eliminate conflicting association edges that lead to mutual exclusion based on confidence rules, and generate a target evidence-based knowledge graph. In step S2, a deep learning natural language processing joint extraction model is used to convert the text into a graph data structure containing conditional triggering relationships. For mutually exclusive medical orders generated during multi-source guide fusion, a confidence evaluation model is introduced to compare and remove the weights of mutually exclusive related edges, retaining the legal execution paths.

[0027] S3. Perform feature quantization processing on the multi-dimensional health data to construct a set of state features; In step S3, feature engineering and dimensionality reduction mapping algorithms are used to map the heterogeneous health data of continuous vital sign values ​​and discrete questionnaire scores into scalars or vectors with unified dimensions, and package and integrate them into a set of state features that can be efficiently computed by a computer.

[0028] S4. Map the set of state features to the starting query node, execute the graph path search algorithm in the target evidence-based knowledge graph to obtain the target instruction node, generate an initial health management plan containing timestamp and execution frequency parameters based on the target instruction node, and send it to the terminal device. In step S4, the quantized set of state features is used as the retrieval entry point, and a graph traversal algorithm is executed in the pruned target evidence-based knowledge graph. After the target instruction node is hit, the specific time and frequency parameters are extracted by parsing the attribute slots embedded in the node, and a structured management plan is generated and distributed to the execution end.

[0029] S5. Receive the time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, update the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculate the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback. In step S5, the system establishes an asynchronous communication mechanism to continuously receive time-series data such as terminal check-in and vital sign fluctuations. This data is input into the state update module to reconstruct the current state space. Simultaneously, based on the difference between the actual execution time / frequency and the planned requirements, the task deviation is calculated as a reward / penalty factor for reinforcement learning.

[0030] S6. Using the current environmental state as input data and the execution feedback as a reward signal, an action instruction vector is output through a reinforcement learning model. The timestamp and execution frequency parameters of the initial health management plan are corrected based on the action instruction vector to generate an updated health management plan.

[0031] In step S6, the current state space and reward / penalty factor constructed in the previous steps are input into the reinforcement learning neural network. After policy gradient inference, the model outputs an action vector containing frequency increase / decrease and timestamp offset, thereby directly overwriting the execution parameters of the original plan in a purely data-driven manner.

[0032] The beneficial effects of the solution provided in this embodiment are as follows: By jointly extracting entities and relationships from unstructured medical text and eliminating conflicting edges based on confidence rules, a target evidence-based knowledge graph with a deterministic pathfinding structure is successfully constructed; by quantifying features of heterogeneous multi-dimensional health data to construct a set of state features, the starting query node for graph reasoning is determined, realizing highly personalized initial plan generation and distribution; this embodiment also constructs a data-driven optimization closed-loop step, using time-series interactive data asynchronously uploaded by terminal devices to calculate task deviation as a reward signal for reinforcement learning, driving the model to adaptively adjust the output action instruction vector. While effectively updating the timestamp and execution frequency parameters of the management plan, the underlying algorithm architecture strictly constrains the adjustment range of strategy iteration, balancing the intelligent optimization efficiency and execution security of the health management solution. The solution of this invention overcomes the long-held technical bias in the field that the unbounded exploration characteristics of reinforcement learning in health management system applications are prone to causing runaway risks and therefore cannot be safely applied to the modification of underlying medical and health instructions.

[0033] It should be noted that the acquisition, collection, storage, transmission, and use of all multi-dimensional health data concerning the target object involved in this application (including but not limited to continuous physiological sign values, discrete subjective assessment scalars, and time-series interaction data) are strictly conducted under legal and compliant conditions. Furthermore, strict anonymization and de-identification processing has been performed on all data before it enters the graph reasoning and reinforcement learning model of this invention for processing.

[0034] Example 2 This embodiment provides a further technical solution based on Embodiment 1.

[0035] In this embodiment, the medical text character sequence includes unstructured natural language sequences extracted from domain protocols, systematic reviews, or expert consensus. The multidimensional health data includes a static attribute vector characterizing the baseline state of the target object, and a subjective state scalar obtained through evaluation.

[0036] It should be noted that text parsing converts medical knowledge carriers with extremely high levels of clinical evidence into plain text format. These unstructured natural language sequences retain the original grammatical rules, contextual information, and terminology of medical knowledge, and are used as the raw input corpus for word segmentation and word embedding operations in the natural language processing encoder. A multimodal data fusion mechanism is employed when constructing the digital profile of the target subject. On the one hand, relatively fixed objective baseline information (such as age, gender, genetic polymorphism, and encoded history of chronic diseases) is constructed into a high-dimensional static attribute vector through one-hot encoding or numerical standardization; on the other hand, the patient's subjective feelings, which change dynamically with time or environment, are transformed into a normalized subjective state scalar.

[0037] See Figure 4 As shown, in this embodiment, the steps of performing entity recognition and relation extraction on the medical text character sequence, constructing an initial graph data structure containing feature nodes, instruction nodes, and directed association edges, eliminating mutually exclusive conflicting association edges based on confidence rules, and generating a target evidence-based knowledge graph include: S201. Extract the feature nodes representing the baseline state of the object in the medical text character sequence through a pre-trained entity relation extraction model, and extract the instruction nodes representing the intervention action and embedding the execution frequency parameter and the timestamp. S202. Based on semantic dependency logic, the directed association edges are established between the feature nodes and the instruction nodes that have conditional triggering relationships, and aggregated to form the initial graph data structure. S203. Perform graph path traversal in the initial graph data structure. When multiple directed associated edges originating from the same feature node point to logically mutually exclusive instruction nodes, it is determined that there are conflicting associated edges. S204. Extract the source attribute features of the corresponding medical text character sequences that generate the conflicting association edges, substitute them into the confidence rules for comprehensive evaluation, remove conflicting association edges with evaluation scores lower than the comparison threshold, and retain uniquely connected association edges to generate the target evidence-based knowledge graph.

[0038] It should be noted that, for reference Figure 6The diagram illustrates the architecture and data processing flow of the medical text entity and relation joint extraction model in this embodiment. The entity and relation extraction in this embodiment employs a dual-branch joint extraction neural network architecture with shared underlying semantic encoding to overcome the error cascading accumulation problem present in traditional pipeline extraction models and to accurately extract complex nested attributes from medical instructions. First, the system inputs unstructured medical text character sequences obtained from domain procedures or expert consensus into a pre-trained text semantic encoder, transforming them into high-dimensional word vectors containing rich medical contextual representations. Subsequently, the high-dimensional word vectors are synchronously input into parallel named entity recognition and semantic dependency extraction branches. In the named entity recognition branch, the model uses a decoder based on fragment or sequence labeling. Besides accurately identifying feature nodes representing the patient's baseline state and instruction nodes representing specific actions, this branch also incorporates an attribute slot filling mechanism to forcibly extract and structurally extract the execution frequency and timestamp attributes strongly bound to the action from the context vocabulary of the instruction entity. In the semantic dependency extraction branch, a multi-head attention mechanism is used to calculate the dependency weights between the identified entities, classifying and determining whether conditional triggering relationships exist between feature nodes and instruction nodes. Finally, the nodes output from both branches are graph-aligned and aggregated with directional conditional triggering relationships, constructing an initial graph data structure in memory that includes feature nodes, instruction nodes with embedded time / frequency parameters, and directed association edges. This provides a structured data foundation for subsequent graph resolution and pruning based on confidence rules.

[0039] In this embodiment, the step of extracting the source attribute features of the corresponding medical text character sequences that generate the conflicting association edges, substituting them into the confidence rule for comprehensive evaluation, eliminating conflicting association edges with evaluation scores below the comparison threshold, and retaining uniquely connected association edges to generate the target evidence-based knowledge graph includes: Obtain the source attribute features of the corresponding medical text character sequences that generate each of the conflict association edges, wherein the source attribute features include evidence rating quantification value and literature publication time; Using the evidence rating quantification value and the publication time of the literature as input variables, and substituting them into the confidence decay function included in the confidence rule, the comprehensive evaluation score of each conflicting association edge is calculated. The confidence decay function is expressed as follows: C=α·f(E level )+β·e -λ(Tn-Tp) ; Where C is the comprehensive evaluation score, f(E) level T is the mapping function for the quantitative value of the evidence rating. p T represents the publication time of the aforementioned document. n Let λ be the current system time, λ be the time decay constant, α be the first weighting coefficient, and β be the second weighting coefficient. Conflicting edges with a comprehensive evaluation score lower than the comparison threshold are removed. When there are multiple conflicting edges with scores not lower than the comparison threshold, the edge with the highest score is retained as the uniquely connected edge to generate the target evidence-based knowledge graph with a unique traversal path.

[0040] It should be noted that conflicting edges with a comprehensive evaluation score lower than the comparison threshold are removed. When there are multiple conflicting edges with scores not lower than the comparison threshold, the edge with the highest score is retained as the only connected edge, so as to generate a target evidence-based knowledge graph with a unique traversal path.

[0041] When the graph construction module traverses downwards along the same feature node and identifies mutually exclusive edges pointing to different instruction nodes—for example, for the same complication, Guideline A recommends increasing water intake while Guideline B recommends strictly limiting water—a conflict resolution mechanism is triggered. First, the source attribute features corresponding to the conflicting edges are traced and extracted, namely, the evidence rating quantification value and the publication time of the literature. These features are then substituted into the confidence decay function. In this core formula, the first part α·f(E level The baseline score for the level of evidence is represented by β, and the system uses a mapping function to convert unstructured evidence ratings into quantitative values; the second part is β·e. -λ(Tn-Tp) This is a penalty term based on the time difference. Where T... n -T p Representing the difference between the current system time and the publication time of the literature, a negative exponential decay model composed of a time decay constant and a natural constant is introduced to achieve a non-linear smooth reduction of the weight of old medical evidence. Deterministic graph pruning: After calculating the comprehensive evaluation score of each conflicting edge, the system implements a hard threshold filtering and maximum value retention strategy. Inferior edges below the comparison threshold are removed, and the edge with the highest C-value among those with acceptable scores is retained, thereby forcibly pruning other branches in the data structure.

[0042] In this embodiment, the step of performing feature quantization processing on the multi-dimensional health data to construct a state feature set includes: Obtain the static attribute vector and the subjective state scalar from the multi-dimensional health data, and perform dimensional unification and standardization mapping on the discrete and continuous feature values ​​in the static attribute vector and the subjective state scalar to generate a standardized feature matrix; Specifically, by allocating an independent data processing space in the underlying memory, the system receives continuous feature values ​​(such as absolute blood pressure and blood glucose concentration) from various sensors and discrete feature values ​​(such as gender and chronic disease history labels) from questionnaire forms. Since these data types are different and their value ranges vary greatly, the system uses a specific mapping function to eliminate differences in different physical units and magnitudes, mapping all data to [0,1] or a specific interval, and finally assembling them into a high-dimensional standardized feature matrix in memory.

[0043] A feature evaluation judgment matrix is ​​constructed based on the analytic hierarchy process (AHP). The consistency of the standardized feature matrix is ​​checked, and the weight factors of each dimension of the features in the standardized feature matrix are calculated. Specifically, the analytic hierarchy process (AHP) is introduced to construct a feature evaluation matrix that assesses the relative importance of each pair of features in each dimension. Subsequently, the maximum eigenvalue and the corresponding eigenvector of the matrix are solved, and a consistency check is strictly performed. After passing the consistency check, the normalized eigenvectors are extracted as weighting factors for each dimension.

[0044] The standardized feature matrix is ​​weighted and fused using the weighting factors to generate a quantitative feature vector containing a discharge preparation metric score and a baseline risk label. The quantitative feature vector is then used as the state feature set.

[0045] Specifically, using the underlying tensor computation unit, the obtained weight factor vector is fused with the standardized feature matrix through a dot product. Through this linear transformation, the system compresses the originally sparse, high-dimensional matrix into an extremely compact quantized feature vector. Two core parameters are specifically extracted from this vector: a continuous discharge preparation quantification score and a discrete baseline risk label.

[0046] It should be noted that the consistency check eliminates self-contradictory weight allocation logic caused by the complexity and interweaving of medical indicators. For example, setting A to be more important than B, B to be more important than C, but incorrectly setting C to be more important than A, thus giving the weight allocation rigor. Transforming the implicit medical experience of experts into computer-executable and verifiable linear algebra matrix operations greatly improves the computational robustness of the model during feature fusion.

[0047] See Figure 5 As shown, in this embodiment, the steps of mapping the state feature set to a starting query node, executing a graph path search algorithm in the target evidence-based knowledge graph to obtain a target instruction node, generating an initial health management plan containing timestamps and execution frequency parameters based on the target instruction node, and distributing it to the terminal device include: Extract the baseline risk label from the state feature set, and use the baseline risk label as a graph retrieval entry point to map it to the starting query node in the target evidence-based knowledge graph; Starting from the initial query node, the graph path search algorithm is executed in the target evidence-based knowledge graph to find the path along the associated edges with a unique traversal path, and to hit and obtain the logically connected target instruction node. The target instruction node is parsed to extract the embedded intervention task and the timestamp and execution frequency parameters bound to the intervention task; The intervention tasks are serialized and arranged according to the preset discharge transition period timeline. The arranged intervention tasks, timestamps, and execution frequency parameters are encapsulated into structured data to generate the initial health management plan, and a communication link is established to send it to the terminal device.

[0048] It should be noted that the baseline risk label extracted from the state feature set is used as the entry point for graph retrieval. This module calls the built-in indexing service of the graph database at its core, performing a global search across the entire feature node set of the target evidence-based knowledge graph using exact string matching or hash mapping algorithms. Once a match is successful, the unique resource identifier of that feature node in the graph database is obtained, and the node containing that unique resource identifier is locked as the starting query node for the graph traversal algorithm. In this way, dimensionality reduction and anchoring of high-dimensional state data to graph network nodes are achieved.

[0049] After locating the initial query node, the graph path reasoning and optimization module executes pathfinding logic based on associated edges. Since conflicting associated edges have been eliminated in the previous steps based on confidence rules, the target evidence-based knowledge graph now presents a directed acyclic graph structure. The graph path reasoning and optimization module uses the initial query node as the root node and calls breadth-first search or depth-first search algorithms from graph theory to perform a layer-by-layer pathfinding along associated edges with uniquely determined traversal paths. During this process, the algorithm filters out intermediate conditional flow nodes based on node attribute labels until it hits the target instruction node representing the specific action at the end.

[0050] After the target instruction node is hit, the instruction parsing and parameter extraction module performs the unpacking of the underlying data. In the graph database, the target instruction node is actually stored as an unstructured / semi-structured object in JSON or XML format containing multi-level nested key-value pairs. This module deserializes and parses the hit target instruction node, extracting fields to obtain the intervention task, and simultaneously extracting the timestamp and execution frequency fields strongly bound to the task to obtain the corresponding timestamp and execution frequency parameters. This module cleans the extracted heterogeneous parameters into a standardized variable format that can be called by a computer, facilitating subsequent encapsulation of the health management plan.

[0051] Example 3 This embodiment provides a further technical solution based on Embodiment 1 or Embodiment 2.

[0052] In this embodiment, the steps of receiving the time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, updating the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculating the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback include: Extract the actual timestamps and actual interaction frequencies, as well as dynamic physiological change characteristics, from the time-series interaction data; Specifically, an asynchronous data receiving service runs in the background, capturing data packets from smart terminals in real time. By deserializing and parsing data payloads in formats such as JSON or XML, the system extracts scalar data representing the patient's actions and tensor data representing changes in the patient's condition from the disordered physical world.

[0053] The dynamic physiological change features are substituted into a preset feature evaluation and judgment matrix for incremental calculation to refresh the quantized feature vector in the state feature set and obtain the current environmental state representing the target object. Specifically, an incremental computing architecture was adopted. The computer only extracts the dynamic physiological changes that cause fluctuations and uses them as a tiny update vector. With this extremely low computational complexity, a quantitative feature vector that can represent the patient's true physiological environment at the current time step is refreshed.

[0054] The actual occurrence timestamp and actual interaction frequency are compared with the timestamp and execution frequency parameters included in the initial health management plan to generate a multi-dimensional task deviation. Specifically, by performing vector subtraction or absolute error calculation logic at the underlying level, the extracted actual value is aligned and compared with the baseline value in the initial plan.

[0055] The multi-dimensional task deviation is weighted, summed, and normalized using a preset evaluation function, and a scalar value is output as the quantified execution feedback.

[0056] It should be noted that the underlying normalization process outputs a standard scalar, effectively avoiding the gradient explosion or vanishing problems that occur in reinforcement learning models when updating policy gradients during backpropagation, thus greatly improving the convergence stability of the model's optimization process. This quantifies complex medical intervention outcomes into a single scalar reward / penalty factor, establishing a clear logical chain from medical behavior assessment to the reward mechanism of computer neural networks.

[0057] In this embodiment, the steps of using the current environmental state as input data, the execution feedback as a reward signal, outputting an action instruction vector through a reinforcement learning model, and correcting the timestamp and execution frequency parameters of the initial health management plan based on the action instruction vector to generate an updated health management plan include: Construct the reinforcement learning model based on the proximal policy optimization algorithm, and set the state space of the reinforcement learning model as a sequence tensor containing the current environment state at multiple consecutive time steps; The action space of the reinforcement learning model is defined as a continuous value matrix that adjusts the timestamp offset and the increase or decrease of the execution frequency parameter. Construct a composite reward function to obtain the reward signal. The composite reward function includes an immediate reward value that is negatively correlated with the quantized execution feedback and a delayed reward discount value that is positively correlated with the quantized risk reduction gradient in the current environmental state. The sequence tensor is input into the reinforcement learning model, and the Actor network outputs an action probability distribution constrained by a preset truncation function boundary to constrain the magnitude of a single policy adjustment. The Critic network is used to evaluate the value baseline of the current input data, and the advantage function is calculated by combining the immediate reward value and the delayed reward discount value. The advantage function is then used to drive model gradient descent to update the network parameters. The action command vector is generated by sampling from the converged action probability distribution, and the offset and increment / decrement are extracted from the action command vector. The timestamp and execution frequency parameters in the initial health management plan are numerically overwritten to generate the updated health management plan.

[0058] It should be noted that, for reference Figure 7 As shown, in this embodiment, a proximal strategy with a truncation mechanism is used to optimize the dual-network architecture at the underlying level. Specifically, the model includes an Actor network and a Critic network. During the forward propagation of the model, a sequence tensor (state St) containing patient signs and interactions is simultaneously input to both networks. The Critic network is used to evaluate the baseline value of the current state, St; combined with the actual acquired composite reward signal Rt, the system calculates the generalized advantage estimate (GAE) to obtain the advantage function. The advantage function characterizes the effectiveness of the current intervention strategy adjustment relative to the average level. To prevent the model from outputting aggressive health management parameter modification instructions during the exploration process (such as significantly increasing medication or monitoring frequency, leading to medical risks), this invention sets a preset truncation function boundary in the update path of the Actor network parameter θ. Its core objective function is expressed as:

[0059] in, The ratio of the probabilities of the new and old strategies. To characterize the truncation hyperparameter of the safety constraint boundary, it is set to a small constant between 0.1 and 0.2 in this preferred embodiment. If the model attempts to make a large policy leap that may bring unknown risks, the clip function will forcibly reduce its gradient update magnitude. The converged Actor network outputs a safe action probability distribution, and the system extracts the action instruction vector At through random sampling or maximum a posteriori probability. The parsing module reduces the dimensionality of the action instruction vector and maps it to the timestamp offset and the increase or decrease of the execution frequency parameter relative to the initial health management plan, thereby completing the intelligent closed-loop dynamic optimization of the health management plan under the premise of clinical safety baseline.

[0060] In this embodiment, the step of receiving the time-series interactive data asynchronously uploaded by the terminal device during the execution of the initial health management plan further includes: Extract real-time physiological values ​​from the time-series interaction data and compare the real-time physiological values ​​with a preset set of safety baseline thresholds; When it is determined that the real-time physiological value exceeds the safety boundary in the set of safety baseline thresholds, the input process to the reinforcement learning model is suspended, and a hierarchical routing early warning mechanism is triggered. The deviation of the real-time physiological values ​​is calculated to match the corresponding severity level of the early warning, and the corresponding communication scheduling strategy is executed based on the severity level of the early warning. The communication scheduling strategy includes: generating autonomous intervention instructions and pushing them to the terminal device, issuing online collaborative prompts to the associated responsible business end, or directly establishing a multidisciplinary emergency scheduling communication link.

[0061] It should be noted that a front-end edge computing defense line was established before the complex and time-consuming feature quantization and reinforcement learning matrix operations. This effectively filtered out extreme and abnormal data, preventing the system from mistaking the patient's acute attack state for a normal environmental state and entering a lengthy inference pipeline. This greatly shortened the system latency of crisis response, prevented extreme outliers from polluting the reinforcement learning environment state pool, and avoided policy gradient calculation collapse caused by extreme feedback.

[0062] In this embodiment, the method further includes the step of: Within a preset long-term time window, the quantified execution feedback generated after multiple target objects execute the same target instruction node is obtained, and the multiple quantified execution feedbacks are aggregated and calculated to obtain the group execution performance index. Specifically, a long-period sliding time window is set in the underlying time series database. Using a distributed computing architecture similar to MapReduce, the system maps all users assigned to specific target instruction nodes at a macroscopic level and reduces the quantitative execution feedback generated during this period. By applying aggregation functions to remove extreme outliers, a statistically significant global scalar—the group execution performance index—is finally calculated.

[0063] When the group execution performance index is determined to be lower than the preset expected threshold and the duration reaches the evaluation cycle, a negative penalty factor corresponding to the target instruction node is generated. Specifically, the system introduces a time-constrained threshold trigger, which not only determines whether the current group performance indicator is lower than a preset expected threshold, but also starts a duration timer. Only when the indicator remains below the threshold level throughout the entire evaluation period will the system trigger a specific interrupt, generating a negative penalty factor in scalar or vector form based on the area of ​​the deviation.

[0064] The negative penalty factor is asynchronously fed back to the target evidence-based knowledge graph, and the weight parameters of the associated edges pointing to the target instruction node in the confidence rule are adaptively adjusted to dynamically reconstruct the local topology of the target evidence-based knowledge graph.

[0065] It should be noted that the system sends the aforementioned negative penalty factor back to the central graph database via an asynchronous message queue. Upon receiving this factor, the graph computation engine locates all associated edges pointing to the target instruction node and substitutes them into the aforementioned confidence rule formula, forcibly reducing the weight parameters of these edges in the matrix. As the weights are reduced, the confidence scores of some edges will fall below the comparison threshold, thus being automatically pruned by the system. Alternatively, in future graph path searches, their low weights will reduce the probability of being traversed and hit, thereby achieving directed reconstruction of the local graph topology at the physical memory level. Specifically, this overcomes the feedback bias caused by poor compliance or sudden situations in a single individual, feeding back the chaotic massive number of micro-nodes into smooth and convergent statistical characteristics. This effectively prevents neurotic misjudgments caused by short-term group data fluctuations, ensuring that penalty factors are only generated when a truly systemic and long-term intervention fails, greatly enhancing the robustness of the system's global judgment.

[0066] Example 4 In the health management scenario of chronic disease patients during the discharge transition period, due to the absence of close monitoring by professional medical staff in the hospital, changes in real physiological indicators, such as a substantial decrease in glycated hemoglobin and a gradual improvement in lung function, often have a time lag. When conventional reinforcement learning algorithms are used to dynamically adjust the timestamp and execution frequency parameters of the health management plan, the model faces a sparse reward dilemma in the early stages. This is because the lack of timely external environmental feedback makes the model prone to getting stuck in repeated trial and error in the state / action space. This not only leads to the ineffective consumption of underlying computing power and slows down the network convergence time, but may also output extreme intervention instructions that deviate significantly from the evidence-based medicine safety baseline during the exploration process.

[0067] To address the aforementioned technical issues, this application further introduces a composite reward generation and computational power circuit breaker algorithm based on graph topological semantic guidance in step S6 to strictly constrain the policy exploration direction of reinforcement learning. Specifically, the step of constructing the composite reward function to obtain the reward signal further includes: During the forward propagation exploration phase of the reinforcement learning model, the initial action instruction vector output by the Actor network is obtained; the initial action instruction vector is back-addressed to lock the target instruction node corresponding to the instruction vector in the target evidence-based knowledge graph; at the same time, the quantized feature vector representing the current environmental state is extracted, the corresponding starting feature node is anchored in the graph, and the system calls the underlying graph traversal algorithm to detect the topological connectivity between the starting feature node and the target instruction node.

[0068] When a connected traversal path exists, the logical hop count of the path and the comprehensive evaluation score of each associated edge along the path, which was retained during the pre-built graph, are extracted. Based on the logical rule that the logical hop count is negatively correlated and the comprehensive evaluation score is positively correlated, the system calculates and outputs a scalar value as an intrinsic reward signal to constrain the exploration direction of the model.

[0069] The system constructs a composite reward calculation pipeline that includes a dynamic balancing factor. In the early stage of model training, the intrinsic reward signal is assigned a high weight parameter, and the static topology of the graph dominates the direction of network gradient descent. As the model training time step increases and the time-series interaction data of the terminal device accumulates, the system adaptively decays the weight parameter of the intrinsic reward signal and simultaneously increases the weight parameter of the quantized execution feedback calculated using the task deviation.

[0070] When probing topological connectivity, if the graph traversal algorithm returns a disconnected state, or if the calculated intrinsic reward signal is lower than the preset medical logic safety threshold, the system forcibly generates a negative extreme value penalty signal and directly overwrites the composite reward function. At the same time, a truncation instruction is activated in the backpropagation computation graph to directly reject the gradient update operation of the current network parameter matrix.

[0071] By adopting the above technical solution, this application achieves the following significant beneficial effects: This approach overcomes the limitations of conventional reinforcement learning, which relies solely on external feedback. By introducing intrinsic reward signals, it transforms the static, evidence-based medical knowledge logic of human experts into an intrinsic motivation driving the optimization of neural network parameters. The adaptive dual-track reward fusion mechanism addresses the data scarcity problem during the model's cold start phase, reducing the ineffective exploration range of the model in the vast state space and significantly improving the convergence speed of the algorithm.

[0072] Unlike conventional technologies that set a single numerical alarm threshold at the business output end, this application deeply embeds medical safety rules into the underlying computer control logic. By determining the disconnection of the graph topology, the system performs computational power circuit breaking at the backpropagation source of the neural network to reject gradient updates caused by erroneous actions. This greatly reduces the risk of potential medical accidents caused by extreme command issuance and avoids the ineffective waste of computer computing power in meaningless solution spaces.

[0073] Existing technical solutions typically treat knowledge graphs and reinforcement learning as two independent physical black boxes, assembling them in a pipeline manner. This application, however, establishes a data link between the two, directly substituting the pathfinding results of the underlying graph database into the tensor gradient calculation link of the neural network model, achieving cross-coupling at the underlying architecture level.

[0074] Example 5 See Figure 8 As shown, this embodiment provides a health management device based on graph reasoning, used to apply the health management method based on graph reasoning described in any of the above embodiments. The device includes: The data acquisition module is used to acquire medical text character sequences and multi-dimensional health data of the target object; The graph construction module is used to perform entity recognition and relation extraction on the medical text character sequence, construct an initial graph data structure containing feature nodes, instruction nodes and directed association edges, eliminate conflicting association edges that cause mutual exclusion based on confidence rules, and generate the target evidence-based knowledge graph. The feature quantization module is used to perform feature quantization processing on the multi-dimensional health data to construct a set of state features; The graph reasoning module is used to map the set of state features to the starting query node, execute a graph path search algorithm in the target evidence-based knowledge graph to obtain the target instruction node, generate an initial health management plan containing timestamp and execution frequency parameters based on the target instruction node, and send it to the terminal device. The feedback update module is used to receive the time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, update the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculate the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback. The reinforcement learning module is used to take the current environmental state as input data, the execution feedback as a reward signal, output an action instruction vector through the reinforcement learning model, and modify the timestamp and execution frequency parameters of the initial health management plan according to the action instruction vector to generate an updated health management plan.

[0075] It should be noted that the beneficial effects of the graph-based reasoning health management device provided in this embodiment are as follows: By jointly extracting entities and relationships from unstructured medical text and eliminating conflicting edges based on confidence rules, a target evidence-based knowledge graph with a deterministic pathfinding structure is successfully constructed; by quantifying features of heterogeneous multi-dimensional health data to construct a set of state features, the starting query node for graph reasoning is determined, realizing highly personalized initial plan generation and distribution; the present invention also constructs a data-driven optimization closed-loop step, using the time-series interactive data asynchronously uploaded by the terminal device to calculate the task deviation as a reward signal for reinforcement learning, driving the model to adaptively adjust the output action command vector, effectively updating the timestamp and execution frequency parameters of the management plan, while strictly constraining the adjustment range of strategy iteration through the underlying algorithm architecture, thus balancing the intelligent optimization efficiency and execution security of the health management solution.

[0076] It should be noted that the underlying algorithm architecture based on graph reasoning and reinforcement learning provided in the embodiments of the present invention is not simply a stacking of general-purpose computer underlying computing power or the application of open-source code, but rather a deep integration of the evidence-based medical thinking and advanced mathematical logic mapping capabilities of experts in the field of clinical medicine.

[0077] Figure 9 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 9 As shown, the computer device includes a processor, memory, and a network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a health management method based on graph reasoning. The memory may also store a computer program, which, when executed by the processor, enables the processor to implement a health management method based on graph reasoning. Those skilled in the art will understand that... Figure 9The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0078] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: Acquire medical text character sequences and multi-dimensional health data of the target object; Entity recognition and relation extraction are performed on the medical text character sequence to construct an initial graph data structure containing feature nodes, instruction nodes and directed association edges. Conflicting association edges that cause mutual exclusion are eliminated based on confidence rules, and a target evidence-based knowledge graph is generated. The multi-dimensional health data is subjected to feature quantization processing to construct a set of state features; The set of state features is mapped to the starting query node. A graph path search algorithm is executed in the target evidence-based knowledge graph to obtain the target instruction node. An initial health management plan containing timestamp and execution frequency parameters is generated based on the target instruction node and sent to the terminal device. The system receives time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, updates the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculates the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback. Using the current environmental state as input data and the execution feedback as a reward signal, an action instruction vector is output through a reinforcement learning model. The timestamp and execution frequency parameters of the initial health management plan are then corrected based on the action instruction vector to generate an updated health management plan.

[0079] In one embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, causes the processor to perform the following steps: Acquire medical text character sequences and multi-dimensional health data of the target object; Entity recognition and relation extraction are performed on the medical text character sequence to construct an initial graph data structure containing feature nodes, instruction nodes and directed association edges. Conflicting association edges that cause mutual exclusion are eliminated based on confidence rules, and a target evidence-based knowledge graph is generated. The multi-dimensional health data is subjected to feature quantization processing to construct a set of state features; The set of state features is mapped to the starting query node. A graph path search algorithm is executed in the target evidence-based knowledge graph to obtain the target instruction node. An initial health management plan containing timestamp and execution frequency parameters is generated based on the target instruction node and sent to the terminal device. The system receives time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, updates the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculates the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback. Using the current environmental state as input data and the execution feedback as a reward signal, an action instruction vector is output through a reinforcement learning model. The timestamp and execution frequency parameters of the initial health management plan are then corrected based on the action instruction vector to generate an updated health management plan.

[0080] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0081] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0082] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A health management method based on graph reasoning, applied to a health management system for the discharge transition period of patients with chronic diseases, characterized in that, The method includes: Acquire medical text character sequences and multi-dimensional health data of the target object; Entity recognition and relation extraction are performed on the medical text character sequence to construct an initial graph data structure containing feature nodes, instruction nodes and directed association edges. Conflicting association edges that lead to mutual exclusion are eliminated based on confidence rules, and a target evidence-based knowledge graph is generated. The multi-dimensional health data is subjected to feature quantization processing to construct a set of state features; The set of state features is mapped to the starting query node. A graph path search algorithm is executed in the target evidence-based knowledge graph to obtain the target instruction node. An initial health management plan containing timestamp and execution frequency parameters is generated based on the target instruction node and sent to the terminal device. The system receives time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, updates the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculates the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback. Using the current environmental state as input data and the execution feedback as a reward signal, an action instruction vector is output through a reinforcement learning model. The timestamp and execution frequency parameters of the initial health management plan are then corrected based on the action instruction vector to generate an updated health management plan.

2. The health management method based on graph reasoning according to claim 1, characterized in that: The medical text character sequence includes unstructured natural language sequences extracted from domain protocols, systematic reviews, or expert consensus. The multidimensional health data includes a static attribute vector characterizing the baseline state of the target object, and a subjective state scalar obtained through evaluation.

3. The health management method based on graph reasoning according to claim 2, characterized in that, The steps of performing entity recognition and relation extraction on the medical text character sequence, constructing an initial graph data structure containing feature nodes, instruction nodes, and directed association edges, eliminating mutually exclusive conflicting association edges based on confidence rules, and generating a target evidence-based knowledge graph include: The feature nodes representing the baseline state of the object in the medical text character sequence are extracted by a pre-trained entity relation extraction model, and the instruction nodes representing the intervention action and embedded with the execution frequency parameter and the timestamp are extracted. Based on semantic dependency logic, directed association edges are established between the feature nodes and the instruction nodes that have conditional triggering relationships, and aggregated to form the initial graph data structure; In the initial graph data structure, a graph path traversal is performed. When multiple directed associated edges originating from the same feature node are identified and pointing to logically mutually exclusive instruction nodes, it is determined that there are conflicting associated edges. The source attribute features of the corresponding medical text character sequences that generate the conflicting association edges are extracted and substituted into the confidence rules for comprehensive evaluation. Conflicting association edges with evaluation scores lower than the comparison threshold are removed, and uniquely connected association edges are retained to generate the target evidence-based knowledge graph.

4. The health management method based on graph reasoning according to claim 3, characterized in that, The steps of extracting the source attribute features of the corresponding medical text character sequences that generate the conflicting association edges, substituting them into the confidence rules for comprehensive evaluation, removing conflicting association edges with evaluation scores below the comparison threshold, and retaining uniquely connected association edges to generate the target evidence-based knowledge graph include: Obtain the source attribute features of the corresponding medical text character sequences that generate each of the conflict association edges, wherein the source attribute features include evidence rating quantification value and literature publication time; Using the evidence rating quantification value and the publication time of the literature as input variables, and substituting them into the confidence decay function included in the confidence rule, the comprehensive evaluation score of each conflicting association edge is calculated. The confidence decay function is expressed as follows: C=α·f(E level )+β·e -λ(Tn-Tp) ; Where C is the comprehensive evaluation score, f(E) level T is the mapping function for the quantitative value of the evidence rating. p T represents the publication time of the aforementioned document. n Let λ be the current system time, λ be the time decay constant, α be the first weighting coefficient, and β be the second weighting coefficient. Conflicting edges with a comprehensive evaluation score lower than the comparison threshold are removed. When there are multiple conflicting edges with scores not lower than the comparison threshold, the edge with the highest score is retained as the uniquely connected edge to generate the target evidence-based knowledge graph with a unique traversal path.

5. The health management method based on graph reasoning according to claim 2, characterized in that, The step of performing feature quantization processing on the multi-dimensional health data to construct a set of state features includes: Obtain the static attribute vector and the subjective state scalar from the multi-dimensional health data, and perform dimensional unification and standardization mapping on the discrete and continuous feature values ​​in the static attribute vector and the subjective state scalar to generate a standardized feature matrix; A feature evaluation judgment matrix is ​​constructed based on the analytic hierarchy process (AHP). The consistency of the standardized feature matrix is ​​checked, and the weight factors of each dimension of the features in the standardized feature matrix are calculated. The standardized feature matrix is ​​weighted and fused using the weighting factors to generate a quantitative feature vector containing a discharge preparation metric score and a baseline risk label. The quantitative feature vector is then used as the state feature set.

6. The health management method based on graph reasoning according to claim 5, characterized in that, The steps of mapping the state feature set to a starting query node, executing a graph path search algorithm in the target evidence-based knowledge graph to obtain a target instruction node, generating an initial health management plan containing timestamps and execution frequency parameters based on the target instruction node, and distributing it to the terminal device include: Extract the baseline risk label from the state feature set, and use the baseline risk label as a graph retrieval entry point to map it to the starting query node in the target evidence-based knowledge graph; Starting from the initial query node, the graph path search algorithm is executed in the target evidence-based knowledge graph to find the path along the associated edges with a unique traversal path, and to hit and obtain the logically connected target instruction node. The target instruction node is parsed to extract the embedded intervention task and the timestamp and execution frequency parameters bound to the intervention task; The intervention tasks are serialized and arranged according to the preset discharge transition period timeline. The arranged intervention tasks, timestamps, and execution frequency parameters are encapsulated into structured data to generate the initial health management plan, and a communication link is established to send it to the terminal device.

7. The health management method based on graph reasoning according to claim 1, characterized in that, The steps of receiving the time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, updating the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculating the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback include: Extract the actual timestamps and actual interaction frequencies, as well as dynamic physiological change characteristics, from the time-series interaction data; The dynamic physiological change features are substituted into a preset feature evaluation and judgment matrix for incremental calculation to refresh the quantized feature vector in the state feature set and obtain the current environmental state representing the target object. The actual occurrence timestamp and actual interaction frequency are compared with the timestamp and execution frequency parameters included in the initial health management plan to generate a multi-dimensional task deviation. The multi-dimensional task deviation is weighted, summed, and normalized using a preset evaluation function, and a scalar value is output as the quantified execution feedback.

8. The health management method based on graph reasoning according to claim 1, characterized in that, The steps of using the current environmental state as input data, the execution feedback as a reward signal, outputting an action command vector through a reinforcement learning model, and correcting the timestamp and execution frequency parameters of the initial health management plan based on the action command vector to generate an updated health management plan include: Construct the reinforcement learning model based on the proximal policy optimization algorithm, and set the state space of the reinforcement learning model as a sequence tensor containing the current environment state at multiple consecutive time steps; The action space of the reinforcement learning model is defined as a continuous value matrix that adjusts the timestamp offset and the increase or decrease of the execution frequency parameter. Construct a composite reward function to obtain the reward signal. The composite reward function includes an immediate reward value that is negatively correlated with the quantized execution feedback and a delayed reward discount value that is positively correlated with the quantized risk reduction gradient in the current environmental state. The sequence tensor is input into the reinforcement learning model, and the Actor network outputs an action probability distribution constrained by a preset truncation function boundary to constrain the magnitude of a single policy adjustment. The Critic network is used to evaluate the value baseline of the current input data, and the advantage function is calculated by combining the immediate reward value and the delayed reward discount value. The advantage function is then used to drive model gradient descent to update the network parameters. The action command vector is generated by sampling from the converged action probability distribution, and the offset and increment / decrement are extracted from the action command vector. The timestamp and execution frequency parameters in the initial health management plan are numerically overwritten to generate the updated health management plan.

9. The health management method based on graph reasoning according to claim 1, characterized in that, The step of receiving the time-series interactive data asynchronously uploaded by the terminal device during the execution of the initial health management plan further includes: Extract real-time physiological values ​​from the time-series interaction data and compare the real-time physiological values ​​with a preset set of safety baseline thresholds; When it is determined that the real-time physiological value exceeds the safety boundary in the set of safety baseline thresholds, the input process to the reinforcement learning model is suspended, and a hierarchical routing early warning mechanism is triggered. The deviation of the real-time physiological values ​​is calculated to match the corresponding severity level of the early warning, and the corresponding communication scheduling strategy is executed based on the severity level of the early warning. The communication scheduling strategy includes: generating autonomous intervention instructions and pushing them to the terminal device, issuing online collaborative prompts to the associated responsible business end, or directly establishing a multidisciplinary emergency scheduling communication link.

10. The health management method based on graph reasoning according to claim 1, characterized in that, The method further includes the following steps: Within a preset long-term time window, the quantified execution feedback generated after multiple target objects execute the same target instruction node is obtained, and the multiple quantified execution feedbacks are aggregated and calculated to obtain the group execution performance index. When the group execution performance index is determined to be lower than the preset expected threshold and the duration reaches the evaluation cycle, a negative penalty factor corresponding to the target instruction node is generated. The negative penalty factor is asynchronously fed back to the target evidence-based knowledge graph, and the weight parameters of the associated edges pointing to the target instruction node in the confidence rule are adaptively adjusted to dynamically reconstruct the local topology of the target evidence-based knowledge graph.

11. A health management device based on graph reasoning, characterized in that, The apparatus for applying the graph-based reasoning health management method according to any one of claims 1 to 10, the apparatus comprising: The data acquisition module is used to acquire medical text character sequences and multi-dimensional health data of the target object; The graph construction module is used to perform entity recognition and relation extraction on the medical text character sequence, construct an initial graph data structure containing feature nodes, instruction nodes and directed association edges, eliminate conflicting association edges that cause mutual exclusion based on confidence rules, and generate the target evidence-based knowledge graph. The feature quantization module is used to perform feature quantization processing on the multi-dimensional health data to construct a set of state features; The graph reasoning module is used to map the set of state features to the starting query node, execute a graph path search algorithm in the target evidence-based knowledge graph to obtain the target instruction node, generate an initial health management plan containing timestamp and execution frequency parameters based on the target instruction node, and send it to the terminal device. The feedback update module is used to receive the time-series interaction data asynchronously uploaded by the terminal device during the execution of the initial health management plan, update the state feature set using the time-series interaction data to characterize the current environmental state of the target object, and calculate the task deviation of the time-series interaction data relative to the initial health management plan to obtain quantified execution feedback. The reinforcement learning module is used to take the current environmental state as input data, the execution feedback as a reward signal, output an action instruction vector through the reinforcement learning model, and modify the timestamp and execution frequency parameters of the initial health management plan according to the action instruction vector to generate an updated health management plan.

12. A computer-readable storage medium, characterized in that, The system stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 10.

Images