Medical clinical decision support method and system based on knowledge graph

Abstract

The invention discloses a medical clinical decision support method and system based on a knowledge graph, and the method comprises the following steps: S1, collecting structured and unstructured medical data, and constructing an initial medical knowledge graph; s2, performing term standardization and semantic alignment on the graph to generate a fusion knowledge graph; s3, constructing a time-labeled medical record graph structure based on the medical record data, and aligning the time-labeled medical record graph structure with the fusion graph; s4, inputting the fusion atlas and the medical record graph into the hypersphere graph neural model, and generating semantic representation; s5, calculating a gravitation vector by using a path traction module, and guiding the propagation direction of the reasoning path; s6, generating a diagnosis and treatment candidate set and a corresponding recommended path according to the node state; s7, optimizing a model structure and initial parameters through a black widow spider optimization algorithm; and S8, outputting diagnosis and treatment suggestions and reasoning paths. According to the method, intelligent organization of medical knowledge and accurate diagnosis and treatment path recommendation are realized, and the auxiliary decision making efficiency and reliability are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical information processing, and in particular to a medical clinical decision support method and system based on a knowledge graph. BACKGROUND

[0002] Existing medical auxiliary decision systems are mostly based on rule engines or statistical learning-based classification models, and mainly rely on fixed templates, expert knowledge bases or annotated data-driven shallow reasoning mechanisms. Such systems have certain response capabilities when dealing with structured data, but when faced with complex semantic unstructured text, time sequence information in medical records and multi-source medical data fusion scenarios, the reasoning accuracy and coverage range are obviously limited. Since traditional models usually perform feature learning in Euclidean space, they lack the ability to express the directionality and semantic hierarchical structure between entities, making it difficult to effectively capture the multi-hop, weak connection logical path between clinical symptoms and potential diseases, resulting in a lack of depth semantic support for auxiliary diagnosis results.

[0003] Some systems attempt to introduce knowledge graphs into the medical field to construct a relationship network composed of entities such as diseases, symptoms, examination items, and drugs to enhance semantic expression and information retrieval efficiency. However, existing knowledge graph modeling methods are rough in handling term standardization, consistency maintenance, and semantic disambiguation. The confusion of synonyms and concept superiors and inferiors often causes node redundancy and edge weight disorder in the graph, reducing the structural effectiveness of subsequent reasoning models. In terms of medical record information analysis, many systems cannot model the patient's disease evolution process in combination with the time dimension, nor can they structure the medical record information into a graph model to participate in downstream graph neural network reasoning, limiting the model's personalized expression ability and explainability output.

[0004] Graph neural networks, as a modeling tool widely used in knowledge reasoning tasks in recent years, have development prospects in medical graph reasoning. However, traditional graph neural network models mostly construct attention weights based on dot product or Euclidean distance, making it difficult to handle medical relationship structures with significant directional semantics. For symptom-disease reasoning paths with complex angle dependencies, the lack of directional expression ability will directly affect the accuracy of diagnosis. In addition, the existing graph reasoning process lacks a modeling mechanism for counterfactual paths, and the model cannot actively capture the implicit causal relationship between the patient's state and the potential diagnosis and treatment path in structure, resulting in a lack of reasoning basis and explanation ability for recommended results in addition to structural connectivity.

[0005] Some studies introduce hyperparameter optimization algorithms to enhance the stability of model performance, but the mainstream optimization methods such as genetic algorithm and particle swarm algorithm have the problem of premature convergence in high-dimensional structure search, which is not conducive to the joint regulation of multivariate and multi-objective weights in graph neural structure. At present, there is still a lack of a medical knowledge graph reasoning method that integrates hyper-spherical semantic reasoning structure, path guiding mechanism and dynamic optimization strategy, which is difficult to meet the requirements of structure accuracy, reasoning directionality and explainability of diagnosis and treatment recommendations at the same time.

[0006] Therefore, how to provide a medical clinical decision support method and system based on a knowledge graph is a problem that those skilled in the art need to solve. SUMMARY

[0007] One object of the present application is to provide a medical clinical decision support method and system based on a knowledge graph. The present application integrates medical entity standardization, medical record graph modeling, hyper-spherical graph neural network reasoning and path gravity regulation mechanism, and describes in detail the whole process of constructing a fusion knowledge graph from multi-source medical data, reasoning diagnosis and treatment nodes and realizing explainable diagnosis and treatment recommendations. It has the advantages of strong semantic consistency, high reasoning accuracy and good explainability of diagnosis and treatment path.

[0008] The medical clinical decision support method based on a knowledge graph according to an embodiment of the present application comprises the following steps: S1, collecting structured and unstructured medical data to construct an initial knowledge graph composed of medical entities and semantic relationships; S2, performing term standardization processing on the entities in the initial knowledge graph based on medical terminology ontology, including synonym normalization, hierarchical mapping and concept disambiguation, to generate a semantic-consistent fusion knowledge graph; S3, collecting patient electronic medical record data, extracting clinical entity nodes containing time labels, and mapping and embedding them into the fusion knowledge graph; S4, constructing a hyper-spherical attention graph neural reasoning model based on the fusion knowledge graph, performing hyper-spherical space embedding and multi-head attention propagation on the graph nodes; S5, introducing a medical path traction module in the graph neural reasoning model, identifying a set of counterfactual paths related to the target diagnosis node, generating a path gravity vector and integrating it into the attention mechanism to guide semantic propagation; S6, generating a diagnosis and treatment node set containing candidate diseases, examination items and treatment recommendations based on the gravity-regulated graph neural reasoning model; S7, optimizing the structure parameters and initial weights of the graph neural reasoning model using the black widow spider optimization algorithm, taking diagnosis accuracy and reasoning path consistency as fitness indicators; S8, outputting diagnosis and treatment recommendations and reasoning paths.

[0009] Optionally, the structured and unstructured medical data collected in S1 includes structured data such as diagnosis codes, examination records, prescription information and surgery records in the electronic medical record system, and unstructured data such as clinical course records, discharge summaries, medical literature abstracts and guideline texts.

[0010] Optionally, the process of constructing an initial knowledge graph composed of medical entities and semantic relationships in S1 comprises: extracting medical entities of diseases, symptoms, drugs, examination items and treatment methods from structured and unstructured data by using a medical named entity recognition model, identifying semantic relationships of diagnosis-symptom, treatment-disease and examination-index between entities based on a syntactic dependency analysis and a context semantic matching method, and constructing a graph structure dataset composed of medical entity nodes and semantic relationship edges as an initial knowledge graph.

[0011] Optionally, S2 specifically comprises: S21, extracting all medical entity nodes in the initial knowledge graph, obtaining the original name, entity type and source data label of each entity node, and constituting a set of entities to be standardized; S22, performing semantic comparison of each entity name in the set of entities to be standardized with the term entries in the standard medical terminology ontology library, determining the corresponding standard term node by using character matching, context word embedding similarity analysis and concept label comparison method, and completing entity normalization mapping operation; S23, for the entity nodes with hierarchical nested relationship after normalization mapping, judging the concept inclusion relationship based on the superior-inferior structure rules in the ontology library, preferentially retaining the superior concept node, merging the attribute values of the subordinate nodes, and updating to a single main node representation; S24, for the entity nodes with ambiguity, analyzing the connection relationship type, upstream and downstream node categories and edge semantic information in the graph, determining the standard term node that best fits the context semantics by semantic consistency judgment method, and replacing the original ambiguous expression; S25, updating the entity nodes processed by S22 to S24 together with the edge relationships in the graph, replacing the original entity nodes, and reconstructing the target node pointing of the connection edge, to form a standardized fusion knowledge graph.

[0012] Optionally, S3 specifically comprises: S31, collecting electronic medical record text data corresponding to the target patient, and extracting chief complaint description, history of present illness record, physical examination result, laboratory test data and treatment record text paragraphs; S32, performing medical entity recognition processing on the text paragraphs, and extracting a structured entity set containing disease name, symptom description, physiological signs, examination index and drug name, and adding a time label corresponding to the position where each entity first appears in the original medical record text; S33, performing one-to-one semantic mapping of the identified structured medical entities and the standardized entity nodes in the fusion knowledge graph, and retaining the mapping relationship between the entities and the standard terms; S34. Organize the medical entities that have completed semantic mapping into a time-ordered sequence of entity nodes according to the time tag order, and establish graph structure connection relationships between entities. Among them, "temporal connection edges" are established between entities that appear consecutively at adjacent time points, and "semantic co-occurrence edges" are established between entities that co-occur in the same paragraph. S35. Define the structured graph structure composed of the constructed nodes and edges as a medical record graph structure, and the medical record graph structure is consistent with the fused knowledge graph in the node semantic space.

[0013] Optionally, S4 specifically includes: S41. Unify the mapping of the knowledge graph and medical record graph structure to the input graph structure for graph neural modeling, maintain standardized entity representation for all nodes, maintain semantic consistency for edge relationship types, and generate a joint graph structure for graph neural modeling. S42. Initialize the hypersphere space vector for each entity node in the joint graph structure, constraining the node representation vector to the unit hypersphere, so that each node representation vector... Satisfy constraints ,in , indicating the first The initial representation of each node; S43. Construct a multi-head hyperspherical attention mechanism module. In each attention head, calculate the attention weight based on the angle between the representation vectors of the node pairs. The attention weights are calculated using the following formula: ; in, Represents a node With adjacent nodes The angle between them A learnable scaling factor. For nodes The set of adjacent nodes; S44. Perform weighted aggregation on the representation vectors of all adjacent nodes under each attention head to generate updated node state representations. The outputs of multiple attention heads are concatenated to form the final node embedding representation. S45. Input the updated node representation as the output of the graph neural reasoning model into the medical path traction module, as the basic semantic space representation for path reasoning.

[0014] Optionally, S5 specifically includes: S51. Receive the generated node state representation, take the starting node corresponding to the patient as the source node, take each node in the candidate diagnosis and treatment node set as the target node, and retrieve all reachable paths from the source node to each target node based on the fused knowledge graph to construct a path set; S52, calculate a path score value for each reachable path, wherein the path score value is calculated by the following formula: ; wherein, denotes the path score value, is the path length, is the semantic confidence of each edge in the path, is the frequency of occurrence of the path in historical cases, , , is a preset weight coefficient; S53, normalize and weight the path score values of each target node corresponding to multiple paths to generate a path attraction value associated with the node, arrange all target nodes to form a path attraction vector, which is used to express the potential guidance intensity of the current patient state to different diagnosis and treatment nodes; S54, fuse the path attraction vector into the attention calculation of the multi-head hyperspherical attention mechanism, and the specific fusion manner is to weight and adjust the original attention weight: ; wherein, is the attention weight after fusing the path attraction, is the original attention weight, is the corresponding component value of the target node in the path attraction vector.

[0015] Optionally, the S7 specifically comprises: S71, determine the graph neural inference model parameters to be optimized, including the number of attention heads, the embedding dimension size, the initial weight matrix distribution mean, the hyperspherical attention scaling factor γ, and the path attraction fusion regulation coefficient, to form an optimization parameter vector θ; S72, set an initial population size, randomly initialize multiple groups of parameter vectors θ, construct a group of graph neural inference model instances based on each group of parameter configurations, and respectively input the fusion knowledge graph and the medical record graph structure to perform a diagnosis and treatment node prediction task; S73, verify each group of model instances, record the diagnosis and treatment node prediction accuracy and the path reasoning consistency score, and calculate the comprehensive evaluation value according to the following fitness function: ; wherein, is the fitness value of the current group of parameters, denotes the matching accuracy of the model output candidate diagnosis and treatment node set and the standard answer, denotes the consistency score of the corresponding reasoning path and the historical true path, , is a weight coefficient; S74, using the black widow spider optimization algorithm to perform individual selection, cross pairing and local mutation operation on the current population, generate a new generation of parameter vector set, repeat the model construction and fitness calculation process; S75, after meeting the termination condition, select the parameter vector with the highest fitness value as the optimal structure parameter and initialization weight configuration of the graph neural inference model.

[0016] The medical clinical decision support system based on a knowledge graph according to the embodiment of the application comprises the following modules: A knowledge graph construction module is configured to collect medical data and construct a standardized fusion knowledge graph, and provide a unified semantic representation space. A medical record analysis module is configured to extract clinical entities in an electronic medical record and construct a time-labeled medical record graph structure. A graph neural inference module is configured to perform hyper-spherical attention propagation and generate node state representation, and support multi-hop semantic reasoning. A path traction regulation module is configured to calculate a path gravity vector and regulate the direction of attention propagation in the graph neural network. An optimization execution module is configured to iteratively update the structure parameters and initial weights of the inference model based on the black widow spider optimization algorithm. A diagnosis and treatment output module is configured to output candidate diagnosis and treatment suggestions and corresponding inference paths, and realize the display of the explainability of the results.

[0017] The beneficial effects of the application are: The application realizes unified semantic expression of heterogeneous medical data: by constructing a fusion knowledge graph, and introducing term normalization, upper and lower concept mapping and semantic disambiguation processing in the graph construction process, the problems of inconsistent expression of multi-source medical entities, semantic overlap and confusion of entity attribution are solved, and the structural accuracy and semantic integrity of the knowledge graph are improved.

[0018] The application improves the individualized reasoning ability of the model for clinical states: by embedding the medical record graph structure into the fusion knowledge graph, and using the hyper-spherical attention network for node embedding and directional semantic propagation, the expression ability of the model for the causal relationship between time-series medical entities is enhanced, and precise diagnosis and treatment node reasoning for a single patient state is realized.

[0019] The application enhances the explainability and optimization robustness of diagnosis and treatment recommendations: by introducing a path gravity vector mechanism, the attention distribution is guided to concentrate on potential medical causal paths during multi-hop reasoning, and the control ability of the reasoning direction is improved; at the same time, the black widow spider optimization algorithm is used to iteratively optimize the model structure parameters, and the adaptability of the system to complex medical graph structures and the output stability are improved. BRIEF DESCRIPTION OF DRAWINGS

[0020] The accompanying drawings are included to provide a further understanding of the application, and are incorporated in and constitute a part of the specification, illustrate embodiments of the application, and are used to explain the present application, but are not intended to limit the present application. In the drawings:

[0021] Fig. 1 The overall flowchart of the medical clinical decision support method based on the knowledge graph proposed in the application; Fig. 2 The medical history graph structure and fusion graph mapping relationship of the medical clinical decision support method based on the knowledge graph proposed in the application; Fig. 3 The path gravity regulation attention mechanism schematic diagram of the medical clinical decision support method and system based on the knowledge graph proposed in the application. DETAILED DESCRIPTION

[0022] The application will now be described in further detail with reference to the drawings. These drawings are simplified schematic diagrams, and only illustrate the basic structure of the application in a schematic manner, and thus only show the components related to the application.

[0023] Reference Figs. 1-3 The medical clinical decision support method and system based on the knowledge graph, comprising the following steps: S1, collecting structured and unstructured medical data, and constructing an initial knowledge graph composed of medical entities and semantic relationships; S2, performing term standardization processing on the entities in the initial knowledge graph based on a medical terminology ontology, including synonym normalization, hierarchical mapping and concept disambiguation, to generate a semantic-consistent fusion knowledge graph; S3, collecting patient electronic medical record data, extracting clinical entity nodes containing time labels, and mapping and embedding the fusion knowledge graph; S4, constructing a hyper-spherical attention graph neural inference model based on the fusion knowledge graph, performing hyper-spherical space embedding and multi-head attention propagation on the graph nodes; S5, introducing a medical path traction module in the graph neural inference model, identifying a set of counterfactual paths related to the target diagnosis node, generating a path gravity vector and fusing it into the attention mechanism to guide semantic propagation; S6, generating a diagnosis and treatment node set containing candidate diseases, examination items and treatment suggestions based on the gravity-regulated graph neural inference model; S7, optimizing the structure parameters and initialization weights of the graph neural inference model using the black widow spider optimization algorithm, and taking diagnosis accuracy and reasoning path consistency as fitness indicators; S8, outputting diagnosis and treatment suggestions and reasoning paths.

[0024] The application establishes a medical auxiliary decision-making method with complete structure, consistent semantics and strong directionality by constructing a process combining a knowledge graph, a medical record graph structure and a hyperspherical graph neural reasoning model. Compared with the traditional diagnosis recommendation method based on rule matching or shallow model, the application can realize multi-hop semantic propagation and explainable diagnosis and treatment path reasoning, improve the diagnosis accuracy and model reliability in complex scenarios, and is suitable for clinical auxiliary analysis tasks in multi-symptom and multi-time medical scenarios.

[0025] In the embodiment, the structured and unstructured medical data collected in S1 includes structured data such as diagnosis codes, inspection records, prescription information and operation records in the electronic medical record system, and unstructured data such as clinical course records, discharge summaries, medical literature abstracts and guideline texts.

[0026] The application ensures the richness and context consistency of the basic data in the knowledge graph construction process by explicitly introducing structured and unstructured medical data sources, including medical record fields, guideline literature, inspection records and the like. Compared with the prior art which only relies on structured diagnosis codes or de-identified text fragments, the application can realize a wider coverage of medical information, and provides higher quality input basis for subsequent entity extraction and knowledge modeling.

[0027] In the embodiment, the process of constructing an initial knowledge graph composed of medical entities and semantic relationships in S1 includes: extracting medical entities such as diseases, symptoms, drugs, inspection items and treatment methods from structured and unstructured data using a medical named entity recognition model, identifying semantic relationships such as diagnosis-symptom, treatment-disease and inspection-index between entities based on syntactic dependency analysis and context semantic matching method, constructing a graph structure dataset composed of medical entity nodes and semantic relationship edges as an initial knowledge graph.

[0028] The application introduces entity-level multi-class medical concept recognition and relationship extraction strategy based on syntactic semantics in the process of constructing the initial knowledge graph, and constitutes a medical semantic relationship network. Compared with the existing graph construction method which only supports disease classification or upper-lower static structure, the application can support more rich diagnosis and treatment relationship modeling between entities, which helps to enhance the semantic propagation ability and structural explainability of the graph in the graph neural model.

[0029] In the embodiment, S2 specifically includes: S21, extracting all medical entity nodes in the initial knowledge graph, obtaining the original name, entity type and source data label of each entity node, and constituting a set of entities to be standardized; S22, each entity name in the to-be-standardized entity set is semantically compared with a term entry in the standard medical terminology ontology, character matching, context word embedding similarity analysis and concept label comparison methods are used to determine the corresponding standard term node, and an entity mapping operation is completed; S23, for the entity nodes with hierarchical nested relationship after mapping, based on the upper and lower structure rules in the ontology library, the concept inclusion relationship is judged, the upper concept node is preferentially reserved, the attribute values of the subordinate nodes are merged, and the single main node is updated; S24, for the entity nodes with ambiguity, the connection relationship type, upstream and downstream node category and edge semantic information in the graph are analyzed, the standard term node most consistent with the context semantics is determined through semantic consistency judgment, and the original ambiguous expression is replaced; S25, the entity nodes processed through S22 to S24 are updated together with the edge relationship in the graph, replace the original entity nodes, and reconstruct the target node pointing of the connection edge, form a standardized fusion knowledge graph.

[0030] The application solves the problems of inconsistent entity expression and semantic confusion in the process of multi-source data fusion by unifying the term standardization process, including synonym normalization, upper and lower mapping and semantic ambiguity resolution. Unlike the simple keyword mapping in the prior art, the application enhances the consistency and uniqueness of the graph node through the fine semantic alignment mechanism, ensuring the accuracy of the information propagation path in the graph neural inference.

[0031] In the embodiment, the S3 specifically comprises: S31, collecting electronic medical record text data corresponding to the target patient, extracting chief complaint description, present illness record, physical examination result, laboratory test data and treatment record text paragraphs; S32, performing medical entity recognition processing on the text paragraphs, extracting structured entity set containing disease name, symptom description, physiological sign, examination index and drug name, and adding time label corresponding to the position of the first appearance in the original medical record text to each entity; S33, the identified structured medical entity is semantically mapped one by one with the standardized entity node in the fusion knowledge graph, and the mapping relationship between the entity and the standard term is retained; S34, the medical entities that have completed semantic mapping are organized into a time-ordered entity node sequence according to the time label sequence, and the graph structure connection relationship between entities is established, wherein the "time sequence connection edge" is established between entities that appear continuously at adjacent time points, and the "semantic co-occurrence edge" is established between entities that coexist in the same paragraph; S35. Define the structured graph structure composed of the constructed nodes and edges as a medical record graph structure, and the medical record graph structure is consistent with the fused knowledge graph in the node semantic space.

[0032] This invention constructs a medical record graph structure by structuring electronic medical records, combining time tags and contextual entity relationships, and completing semantic mapping with a fused knowledge graph. This provides structural input guarantees for the embedding and reasoning of individual patient states within the knowledge graph. Unlike traditional field-based indexing queries, this invention achieves dynamic patient state modeling based on a graph structure, enhancing the ability for personalized diagnosis and treatment reasoning.

[0033] In this embodiment, S4 specifically includes: S41. Unify the mapping of the knowledge graph and medical record graph structure to the input graph structure for graph neural modeling, maintain standardized entity representation for all nodes, maintain semantic consistency for edge relationship types, and generate a joint graph structure for graph neural modeling. S42. Initialize the hypersphere space vector for each entity node in the joint graph structure, constraining the node representation vector to the unit hypersphere, so that each node representation vector... Satisfy constraints ,in , indicating the first The initial representation of each node; S43. Construct a multi-head hyperspherical attention mechanism module. In each attention head, calculate the attention weight based on the angle between the representation vectors of the node pairs. The attention weights are calculated using the following formula: ; in, Represents a node With adjacent nodes The angle between them A learnable scaling factor. For nodes The set of adjacent nodes; S44. Perform weighted aggregation on the representation vectors of all adjacent nodes under each attention head to generate updated node state representations. The outputs of multiple attention heads are concatenated to form the final node embedding representation. S45. Input the updated node representation as the output of the graph neural reasoning model into the medical path traction module, as the basic semantic space representation for path reasoning.

[0034] The application is based on constructing a hyperspherical attention graph neural inference model based on a fusion graph and a medical record graph, and adopts unit hyperspherical embedding and directional attention mechanism to improve the multi-hop relationship representation capability. Unlike the traditional graph neural network which only considers Euclidean feature propagation, the application introduces an angle similarity driven attention structure to enhance the inference capability of the model in the directional semantic graph, and is suitable for complex clinical path modeling tasks.

[0035] In the embodiment, the S5 specifically includes: S51, receiving the generated node state representation, taking the starting node corresponding to the patient as the source node, taking each node in the candidate diagnosis and treatment node set as the target node, searching all reachable paths from the source node to each target node based on the fusion knowledge graph, and constructing a path set; S52, calculating the path score value of each reachable path, wherein the path score value is calculated by the following formula: ; Wherein, represents the path score value, is the path length, is the semantic confidence of each edge in the path, is the frequency of the path in the historical cases, 、 、 is a preset weight coefficient; S53, normalizing and weighting the multiple path score values corresponding to each target node to generate a path attraction value associated with the node, arranging all target nodes to form a path attraction vector, which is used to express the potential guiding strength of the current patient state to different diagnosis and treatment nodes; S54, the path attraction vector is fused into the attention calculation of the multi-head hyperspherical attention mechanism, and the specific fusion method is to adjust the original attention weight by weighting: ; Wherein, is the attention weight after fusing the path attraction, is the original attention weight, is the target node corresponding component value in the path attraction vector.

[0036] The application introduces a medical path traction module to construct a path attraction vector between the patient state and the diagnosis and treatment node, and fuses it into the attention mechanism to regulate the information propagation path. This mechanism makes up for the defects of traditional graph neural networks in weak connection semantic path recognition, realizes the "cause and effect traction" on the inference structure, and significantly improves the rationality, directionality and explanation ability of diagnosis and treatment recommendation.

[0037] In this embodiment, S7 specifically includes: S71, determining the graph neural inference model parameters to be optimized, including the number of attention heads, the embedding dimension size, the initial weight matrix distribution mean, the hypersphere attention scaling factor gamma, and the path gravity fusion regulation coefficient, to constitute an optimization parameter vector theta; S72, setting an initial population size, randomly initializing multiple sets of parameter vectors theta, constructing a set of graph neural inference model instances based on each set of parameter configurations, and respectively inputting the fusion knowledge graph and the medical record graph structure to perform diagnosis and treatment node prediction tasks; S73, performing verification on each set of model instances, recording the diagnosis and treatment node prediction accuracy and the path reasoning consistency score, and calculating the comprehensive evaluation value according to the following fitness function: Wherein, is the fitness value of the current set of parameters, represents the matching accuracy of the model output candidate diagnosis and treatment node set and the standard answer, represents the consistency score of the corresponding reasoning path and the historical true path, , is a weight coefficient; S74, using the black widow spider optimization algorithm to perform individual selection, cross pairing and local mutation operations on the current population to generate a new generation of parameter vector set, and repeatedly performing the model construction and fitness calculation process; S75, after meeting the termination condition, selecting the parameter vector with the highest fitness value as the optimal structure parameter and initialization weight configuration of the graph neural inference model.

[0038] The present application adopts the black widow spider optimization algorithm to jointly optimize the structure parameters and initialization weights of the graph neural model, and combines a multi-objective fitness function to evaluate the diagnosis accuracy and path consistency. Compared with genetic algorithm and particle swarm optimization algorithm, the optimization method of the present application has stronger global exploration and local search ability, improves the stability and convergence efficiency of the inference model, and is suitable for high-dimensional structure optimization tasks.

[0039] The medical clinical decision support system based on a knowledge graph according to an embodiment of the present application includes the following modules: A knowledge graph construction module is used to collect medical data and construct a standardized fusion knowledge graph, and provide a unified semantic representation space; A medical record analysis module is used to extract clinical entities in an electronic medical record and construct a time-labeled medical record graph structure; A graph neural inference module is used to perform hypersphere attention propagation, generate node state representation, and support multi-hop semantic reasoning; ​A path traction regulation module is configured to calculate a path traction vector and regulate a direction of attention propagation in the graph neural network. An optimization execution module is configured to iteratively update structure parameters and initial weights of the inference model based on a spider optimization algorithm. A diagnosis and treatment output module is configured to output candidate diagnosis and treatment suggestions and corresponding inference paths to realize an explainable display of results.

[0040] The medical clinical decision support system designed by the application has clear module division, covers functions such as atlas construction, medical record structure modeling, semantic reasoning, path traction regulation and optimization execution, and has a closed loop of data flow logic and unified terminology between modules. Compared with the existing system structure mainly based on a single model or nested function calling, the application has a highly modular, reconfigurable and traceable system architecture, which is convenient for deployment and upgrading.

[0041] Embodiment 1 In order to verify the feasibility of the application in implementation, the application is applied to a respiratory medicine intelligent auxiliary diagnosis and treatment platform construction project of a certain third-grade comprehensive hospital. The department has a large number of patients, complex diseases, especially in the fields of chronic obstructive pulmonary disease, pulmonary infection and asthma, which have characteristics such as diverse clinical symptom expression, time change sensitivity, and complex auxiliary examination information. Artificial diagnosis has a high load, and there are problems such as unclear preliminary diagnosis, redundant examination items, and low matching degree of treatment schemes. The project goal is to establish a knowledge graph reasoning auxiliary system for real medical record data, which is used to provide diagnosis suggestions and auxiliary decision paths for inpatient and outpatient doctors.

[0042] In the early stage of system construction, structured data fields such as inspection results, discharge diagnosis codes and prescription details were extracted from the hospital information system, and unstructured texts such as admission records, medical records and discharge summaries were extracted, a total of 11247 medical record documents were processed, covering about 980000 text paragraphs. Using the entity recognition method proposed in the application, diseases, symptoms, test indicators, drugs and other medical entities were extracted from the original text. There are 31542 standardized entity nodes in the initial recognition result, an average of 28.1 standardized entities per medical record, and the node redundancy rate of the initial entity relationship graph is 18.4%. After the term standardization processing, the redundancy rate is reduced to 4.2%, and the graph connectivity is improved by 37.5%.

[0043] The standardized fusion knowledge graph is then connected with real patient medical record data, the time information in each medical record is parsed by using the medical record graph construction process of the application to generate a clinical state graph with a time axis. By aligning with the standard entities of the fusion graph, a semantic consistent medical record graph structure is established to ensure the consistency of the input data of the graph neural network model. The system has completed the graph structure conversion of 9725 medical records, automatically extracted 219574 graph structure nodes, with an average of 22.6 graph nodes per medical record and an average of 45.1 edges.

[0044] After the construction is completed, the hyperspherical graph neural network model proposed in the application is used for reasoning modeling, and a path traction module is introduced to automatically identify the multi-hop path from the medical record state node to the recommended diagnosis node, and generate a path gravity vector to guide the attention propagation direction. In the diagnosis and treatment recommendation task, the Top3 diagnosis accuracy of the model of the application reaches 86.4%, which is improved by 9.8% compared with the ordinary GAT model. Especially in the multi-symptom cross case, such as the sample of asthma combined with respiratory tract infection, the proportion of the three-hop link of the reasoning path of the model of the application contains symptoms-examination-disease is 82.6%, while that of the comparative model is only 46.1%, which shows that the structure of the application has stronger ability to capture complex semantic relationships.

[0045] In order to further test the explainability of the model output, doctors conduct blind review evaluation on 300 cases of system recommendation results, of which 258 cases are consistent with the clinical thinking of the reasoning path, accounting for 86.0%; 29 cases are considered to have medical value but secondary reasoning path, accounting for 9.7%. After introducing the Latrodectus Mactans optimization algorithm in the model optimization process, the average convergence round of the model training is reduced from 48 rounds to 31 rounds, and the average validation accuracy fluctuation range is converged from ±3.2% to ±1.1%, which significantly improves the training stability and result consistency.

[0046] The embodiment shows that the application can complete large-scale knowledge graph construction, medical record graph structure generation and diagnosis and treatment suggestion reasoning work in a real clinical data environment. Through the introduction of the path gravity regulation mechanism and the structure optimization algorithm, the system has strong reasoning accuracy, semantic explanation ability and clinical adaptability.

[0047] Table 1: Key performance verification data table of the auxiliary diagnosis and treatment system driven by the fusion knowledge graph

[0048] The above-mentioned "key performance verification data table of the auxiliary diagnosis and treatment system driven by the fusion knowledge graph" fully shows the application performance and key performance indicators of the application in a real clinical environment. In terms of data size, the system processes 11247 real medical records, containing about 980000 unstructured text segments, reflecting its processing capacity under large-scale medical corpus. Through the entity recognition and term standardization process, the initial number of entity nodes is 31542, and the redundancy rate is reduced from 18.4% to 4.2%, which shows that the application has a significant advantage in medical term normalization and semantic consistency processing, and directly drives the connectivity of the fusion knowledge graph to increase by 37.5%, improving the integrity of the graph reasoning structure.

[0049] In the medical record graph construction layer, a total of 9725 medical record graph structures are generated, with an average of 22.6 nodes and 45.1 edges per medical record graph, showing the high efficiency of the system in structuring the patient's state. This structure lays a semantic foundation for subsequent reasoning tasks. In terms of diagnosis and treatment recommendation, the Top3 accuracy rate of the model is 86.4%, significantly better than the 76.6% of the traditional graph attention network model, indicating that the application has stronger reasoning effectiveness in the diagnosis task. Especially in complex conditions such as multiple symptom cross cases, the effective capture rate of the multi-hop path of the system is as high as 82.6%, much higher than the 46.1% of the comparative model, reflecting the key role of the path traction mechanism in semantic reasoning.

[0050] The artificial blind review evaluation results show that the consistency of the reasoning path with the clinical thinking is as high as 86.0%, indicating that the system output has good explainability and medical reliability. After introducing the Latrodectus mactans optimization algorithm, the average convergence round of the model is reduced from 48 rounds to 31 rounds, and the verification accuracy fluctuation range is converged from ±3.2% to ±1.1%, indicating that the application not only has high optimization efficiency, but also has a more stable training process and faster convergence process, improving the engineering practicability of model deployment.

[0051] Overall, the above data verifies the comprehensive technical advantages of the application in data fusion, semantic reasoning, clinical decision support and structure optimization, with strong landing ability and promotion prospects.

[0052] The above is only the preferred specific embodiment of the application, but the protection scope of the application is not limited thereto, and any person skilled in the art can make equivalent substitutions or changes within the technical scope disclosed by the application and according to the technical solutions and inventive concepts of the application, which should be covered within the protection scope of the application.

Claims

1. A medical clinical decision support method based on knowledge graphs, characterized in that, Includes the following steps: S1. Collect structured and unstructured medical data to construct an initial knowledge graph composed of medical entities and semantic relationships; S2. Perform terminology standardization processing on the entities in the initial knowledge graph based on the medical terminology ontology, including synonym unification, hyponym mapping, and concept disambiguation, to generate a semantically consistent fused knowledge graph; S3. Collect patient electronic medical record data, extract clinical entity nodes containing time tags, and map and embed them into the fused knowledge graph; S4. Construct a hyperspherical attention graph neural reasoning model based on a fused knowledge graph, and perform hyperspherical spatial embedding and multi-head attention propagation on the graph nodes; S5. Introduce a medical path traction module into the graph neural reasoning model to identify the set of counterfactual paths related to the target diagnostic node, generate path gravity vectors, and integrate them into the attention mechanism to guide semantic propagation; S6. Generate a set of diagnostic nodes containing candidate diseases, examination items, and treatment suggestions based on the gravity-controlled graph neural reasoning model; S7. The Black Widow Spider optimization algorithm is used to optimize the structural parameters and initial weights of the graph neural network inference model, with the diagnostic accuracy and inference path consistency as fitness indicators; S8. Output diagnostic suggestions and inference paths.

2. The medical clinical decision support method based on knowledge graphs according to claim 1, characterized in that, The structured and unstructured medical data collected in S1 include: structured data such as diagnostic codes, laboratory test records, prescription information and surgical records in the electronic medical record system, as well as unstructured data such as clinical progress records, discharge summaries, medical literature abstracts and guideline texts.

3. The medical clinical decision support method and system based on knowledge graphs according to claim 2, characterized in that, The process of constructing the initial knowledge graph composed of medical entities and semantic relationships in S1 includes: extracting medical entities such as diseases, symptoms, drugs, examination items, and treatment methods from structured and unstructured data using a medical named entity recognition model; identifying semantic relationships between entities such as diagnosis-symptom, treatment-disease, and examination-indicator based on syntactic dependency analysis and context semantic matching methods; and constructing a graph structure dataset composed of medical entity nodes and semantic relationship edges as the initial knowledge graph.

4. The medical clinical decision support method based on knowledge graphs according to claim 3, characterized in that, S2 specifically includes: S21. Extract all medical entity nodes from the initial knowledge graph, obtain the original name, entity type and source data label of each entity node, and form a set of entities to be standardized. S22. Perform semantic comparison between each entity name in the set of entities to be standardized and the term entries in the standard medical terminology ontology. Use character matching, context word embedding similarity analysis and concept tag comparison methods to determine the corresponding standard term nodes and complete the entity normalization mapping operation. S23. For entity nodes that have a hierarchical nesting relationship after normalization mapping, based on the hierarchical structure rules in the ontology library, determine the concept inclusion relationship, prioritize the retention of the superior concept node, merge the attribute values ​​of the subordinate node, and update it to a single master node representation. S24. For ambiguous entity nodes, analyze the connection relationship type, upstream and downstream node categories and edge semantic information in the graph, and determine the standard term node that best fits the context semantics through semantic consistency judgment, and replace the original ambiguous expression. S25. Update the entity nodes processed by S22 to S24, along with their edge relationships in the graph, replace the original entity nodes, and reconstruct the target node pointers of the connecting edges to form a standardized fusion knowledge graph.

5. The medical clinical decision support method based on knowledge graphs according to claim 4, characterized in that, S3 specifically includes: S31. Collect electronic medical record text data corresponding to the target patient, and extract the chief complaint description, present medical history record, physical examination results, laboratory test data and treatment record text paragraphs; S32. Perform medical entity recognition processing on the text paragraph to extract a set of structured entities containing disease name, symptom description, physiological signs, examination indicators and drug names, and attach a time tag to the position corresponding to the first appearance of each entity in the original medical record text. S33. Perform a one-to-one semantic mapping between the identified structured medical entities and the standardized entity nodes in the fusion knowledge graph, and retain the mapping relationship between entities and standard terms. S34. Organize the medical entities that have completed semantic mapping into a time-ordered sequence of entity nodes according to the time tag order, and establish graph structure connection relationships between entities. Among them, "temporal connection edges" are established between entities that appear consecutively at adjacent time points, and "semantic co-occurrence edges" are established between entities that co-occur in the same paragraph. S35. Define the structured graph structure composed of the constructed nodes and edges as a medical record graph structure, and the medical record graph structure is consistent with the fused knowledge graph in the node semantic space.

6. The medical clinical decision support method based on knowledge graphs according to claim 5, characterized in that, S4 specifically includes: S41. Unify the mapping of the knowledge graph and medical record graph structure to the input graph structure for graph neural modeling, maintain standardized entity representation for all nodes, maintain semantic consistency for edge relationship types, and generate a joint graph structure for graph neural modeling. S42. Initialize the hypersphere space vector for each entity node in the joint graph structure, constraining the node representation vector to the unit hypersphere, so that each node representation vector... Satisfy constraints ,in , indicating the first The initial representation of each node; S43. Construct a multi-head hyperspherical attention mechanism module. In each attention head, calculate the attention weight based on the angle between the representation vectors of the node pairs. The attention weights are calculated using the following formula: ； in, Represents a node With adjacent nodes The angle between them A learnable scaling factor. For nodes The set of adjacent nodes; S44. Perform weighted aggregation on the representation vectors of all adjacent nodes under each attention head to generate updated node state representations. The outputs of multiple attention heads are concatenated to form the final node embedding representation. S45. Input the updated node representation as the output of the graph neural reasoning model into the medical path traction module, as the basic semantic space representation for path reasoning.

7. The medical clinical decision support method based on knowledge graphs according to claim 6, characterized in that, S5 specifically includes: S51. Receive the generated node state representation, take the starting node corresponding to the patient as the source node, take each node in the candidate diagnosis and treatment node set as the target node, and retrieve all reachable paths from the source node to each target node based on the fused knowledge graph to construct a path set; S52. Calculate a path score for each reachable path, where the path score is calculated using the following formula: ； in, Indicates the path score. For path length, The semantic confidence score for each edge in the path. The frequency of the path in historical cases, , , These are preset weighting coefficients; S53. Normalize and weight the multiple path scores corresponding to each target node to generate a path gravity value associated with the node. Arrange all target nodes to form a path gravity vector, which is used to express the potential guidance strength of the current patient status on different diagnosis and treatment nodes. S54. The path gravity vector is fused into the attention calculation of the multi-head hyperspherical attention mechanism. Specifically, the fusion method is to adjust the original attention weights by weighting: ； in, The attention weights after merging path gravity. For the original attention weights, For the target node The corresponding component value in the path gravity vector.

8. The medical clinical decision support method based on knowledge graphs according to claim 7, characterized in that, Specifically, S7 includes: S71. Determine the parameters of the graph neural inference model to be optimized, including the number of attention heads, the size of the embedding dimension, the mean of the initial weight matrix distribution, the hypersphere attention scaling factor γ, and the path gravity fusion control coefficient, to form the optimization parameter vector θ. S72. Set the initial population size, randomly initialize multiple sets of parameter vectors θ, construct a set of graph neural reasoning model instances based on each set of parameter configuration, and input the fused knowledge graph and medical record graph structure respectively to perform the diagnosis and treatment node prediction task; S73. Perform validation on each set of model instances, record the prediction accuracy of diagnosis and treatment nodes and the consistency score of path inference, and calculate the comprehensive evaluation value according to the following fitness function: ； in, This represents the fitness value of the parameters in this set. This represents the accuracy of matching the model's output set of candidate diagnosis and treatment nodes with the standard answer. This represents the consistency score between the corresponding reasoning path and the historical actual path. , These are the weighting coefficients; S74. Use the black widow spider optimization algorithm to perform individual selection, crossover pairing and local mutation operations on the current population to generate a new generation of parameter vector set, and repeat the model construction and fitness calculation process. S75. After the termination condition is met, select the parameter vector with the highest fitness value as the optimal structural parameters and initial weight configuration of the graph neural inference model.

9. A deep learning-based knowledge graph-based medical clinical decision support system, applied to the knowledge graph-based medical clinical decision support method according to any one of claims 1 to 8, characterized in that, Includes the following modules: The knowledge graph construction module is used to collect medical data and build a standardized fusion knowledge graph, providing a unified semantic representation space; The medical record parsing module is used to extract clinical entities from electronic medical records and construct a time-tagged medical record graph structure. The graph neural reasoning module is used to perform hyperspherical attention propagation, generate node state representations, and support multi-hop semantic reasoning. The path traction control module is used to calculate the path gravity vector and control the direction of attention propagation in the graph neural network; The optimization execution module is used to iteratively update the inference model's structural parameters and initial weights based on the Black Widow Spider optimization algorithm. The diagnosis and treatment output module is used to output candidate diagnosis and treatment suggestions and corresponding reasoning paths, so as to achieve interpretable display of results.

Images