Patient examination item recommendation method based on medical knowledge graph

Abstract

The invention relates to a patient examination item recommendation method based on a medical knowledge graph. The method comprises the steps that firstly, historical data in the medical field is acquired, key information is extracted from the historical data to generate an entity relation triple, and then a medical knowledge graph is constructed; then, generating a patient state sub-graph based on patient input information, processing the medical knowledge graph by adopting an improved TransD model, and determining unique representation of an entity-relationship pair; mapping the patient state sub-map to a unique representation space, obtaining vector representation of a patient and an examination item related entity, and calculating a similarity value between vectors to obtain a similarity value matrix; and finally, performing priority ranking on the inspection items according to the matrix, and generating a final recommendation list through threshold screening and combination with graph neural network verification, optimization and adjustment of priorities. By adopting the method, examination items can be accurately recommended for patients, doctors are assisted to improve the treatment efficiency, the queuing time of the patients is shortened, and the medical experience of the patients is remarkably improved.

Description

Technical Field

[0001] The present invention belongs to the field of medical information technology, and in particular relates to a method for recommending patient examination items based on a medical knowledge graph. Background Art

[0002] With the development of medical information technology, technology for recommending patient examination items based on medical knowledge graphs has emerged. In the existing technology, hospitals usually rely on doctors' professional knowledge to prescribe examination items based on patients' complaints and medical history. Traditional recommendation methods such as content-based recommendation and collaborative filtering-based recommendation have obvious shortcomings. Content-based recommendation has high requirements for structured data extraction, while collaborative filtering-based recommendation has a low recommendation success rate when processing sparse data in historical data sets, faces cold start problems, and is difficult to expand recommended objects and has low accuracy in the initial stage of the system. For example, although patent CN112037345A proposes medical recommendations based on knowledge graphs, it does not combine the improved TransD model for entity relationship vectorization, making it difficult to effectively handle the diversity of medical data and the complexity of semantic associations. Summary of the Invention

[0003] Based on this, it is necessary to provide a patient examination item recommendation method based on medical knowledge graphs that can accurately recommend examination items to patients and assist doctors in improving medical treatment efficiency in response to the above technical problems.

[0004] In a first aspect, the present application provides a method for recommending patient examination items based on a medical knowledge graph, comprising:

[0005] Acquire historical data in the medical field, extract key information to generate entity relationship triples, and build a medical knowledge graph.

[0006] A patient status subgraph is generated based on patient input information, and an improved TransD model is used to process the medical knowledge graph to determine the unique representation of entity-relationship pairs.

[0007] The patient status subgraph is mapped to a unique representation space, the vector representation of entities related to the patient and examination items is obtained, and the similarity between the vectors is calculated to obtain a similarity value matrix.

[0008] The inspection items are prioritized according to the similarity value matrix, and the priorities are optimized and adjusted through threshold screening and combined with graph neural network verification to generate the final recommendation list.

[0009] In one embodiment, obtaining historical data in the medical field, extracting key information to generate entity relationship triples, and constructing a medical knowledge graph include:

[0010] Acquire historical medical data from multiple sources; historical data includes patient medical records and treatment records.

[0011] Named entity recognition algorithm and information extraction technology are used to extract key information such as patients, symptoms, diagnosis and examination items from preprocessed historical data.

[0012] The relationship between the extracted key information is extracted by combining rule matching with machine learning technology to generate entity relationship triples.

[0013] The entity-relationship triples are combined with the graph completion algorithm to optimize the graph structure and construct a complete medical knowledge graph.

[0014] The quality of the medical knowledge graph is evaluated, and erroneous information is corrected through redundancy detection and contradiction verification to obtain a queryable medical knowledge graph.

[0015] In one embodiment, a patient status subgraph is generated based on patient input information, and an improved TransD model is used to process the medical knowledge graph to determine a unique representation of entity-relationship pairs, including:

[0016] Based on the patient input information and the entities and relationships in the medical knowledge graph, a patient status subgraph is generated through graph query and screening algorithms.

[0017] The improved TransD model is used to vectorize the triples in the patient status subgraph to generate the meaning vectors and projection vectors of entities and relations respectively.

[0018] An adaptive learning mechanism is used to dynamically adjust the mapping matrix parameters, map the meaning vector and the projection vector into the same semantic space, and determine the unique representation of the entity-relationship pair.

[0019] In one embodiment, an improved TransD model is used to vectorize the triples in the patient status subgraph to generate meaning vectors and projection vectors for entities and relationships, respectively, including:

[0020] The triplet data in the patient status sub-graph is obtained, and the noise data and invalid records are removed from the data to obtain the preprocessed triplet.

[0021] Based on the improved TransD model structure, the initial model parameters are set, and the multi-head attention mechanism is used to enhance the ability to capture entity and relationship features in triples to obtain the optimized encoder.

[0022] The preprocessed triples are encoded using the optimized encoder to generate the initial meaning vectors and projection vectors of entities and relations, respectively.

[0023] According to the residual learning strategy, the initial meaning vector and projection vector are input into the residual network for optimization to obtain the final entity and relationship meaning vector and projection vector.

[0024] In one embodiment, the patient state sub-graph is mapped to a unique representation space, vector representations of entities related to the patient and the examination items are obtained, and similarity values ​​between the vectors are calculated to obtain a similarity value matrix, including:

[0025] Obtain a mapping matrix between the patient state sub-map and the unique representation space, and perform spatial conversion preprocessing on the patient state sub-map.

[0026] Based on the mapping matrix, the entity nodes in the patient state subgraph are projected into a unique representation space to extract the patient entity semantic vector.

[0027] Entity nodes related to examination items are retrieved from the medical knowledge graph, and the examination item entity vectors are obtained through the same mapping matrix transformation.

[0028] The optimized cosine similarity algorithm is used to calculate the cosine value of the angle between the semantic vectors of all patient entities and the entity vectors of examination items to generate a similarity value matrix.

[0029] In one embodiment, the similarity value matrix is ​​calculated using the following formula:

[0030]

[0031] Among them, S represents the similarity value matrix, S i,j represents the similarity between the symptom of the i-th patient and the j-th examination item, represents the mapping vector of the i-th patient entity in the unique representation space, represents the mapping vector of the j-th inspection item entity in the unique representation space, Λ i,j represents the prior association weight between the i-th patient symptom and the j-th examination item in the knowledge graph, m represents the number of patient entities, and n represents the number of examination item entities.

[0032] In one embodiment, the inspection items are prioritized according to the similarity value matrix, and the priorities are optimized and adjusted through threshold screening and combined with graph neural network verification to generate a final recommendation list, including:

[0033] The similarity value matrix is ​​sorted in descending order by column to obtain the similarity ranking vector of each examination item for all patients, and an initial examination item priority list is generated based on the vector.

[0034] An adaptive threshold screening algorithm is used to perform dynamic threshold filtering on the priority list, retaining inspection items with similarity values ​​higher than the dynamic threshold to obtain a preliminary recommendation candidate set.

[0035] Graph neural networks are used in combination with the semantic relationships between entities in the medical knowledge graph to verify the association strength of the examination items in the preliminary recommendation candidate set and generate a verification score matrix.

[0036] A multi-objective optimization model was constructed based on the similarity value and verification score, a genetic algorithm was used to perform secondary sorting of the inspection item priorities, and the Pareto optimal solution was used to obtain the optimized list.

[0037] The final screening strategy based on the attention mechanism adjusts the weights of the optimized list, extracts high-confidence examination items directly related to the patient's symptoms, and generates the final recommendation list.

[0038] In a second aspect, the present application further provides a computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the processor implements the above method when executing the computer program.

[0039] In a third aspect, the present application further provides a computer-readable storage medium having a computer program stored thereon, which implements the above method when executed by a processor.

[0040] The aforementioned method, computer device, and storage medium for recommending patient examination items based on a medical knowledge graph first acquires historical medical data, extracts key information from it, and generates entity-relationship triples, thereby constructing a medical knowledge graph. Next, a patient state subgraph is generated based on patient input information. The medical knowledge graph is processed using an improved TransD model to determine unique representations for entity-relationship pairs. The patient state subgraph is then mapped to a unique representation space, obtaining vector representations of entities related to the patient and examination items. A similarity matrix is ​​calculated by calculating the similarity between these vectors. Finally, examination items are prioritized based on this matrix, and the priorities are optimized and adjusted using threshold screening and graph neural network validation to generate a final recommendation list. By integrating knowledge graphs with recommendation algorithms, this method effectively addresses existing issues such as data sparsity, cold start, and low recommendation accuracy. Leveraging the semantic relationships and rich historical data of the medical knowledge graph, it can accurately recommend examination items to patients, assisting doctors in improving treatment efficiency, reducing patient waiting times, and preventing random billing and overcharging, significantly enhancing the patient experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or related technologies, the following briefly introduces the drawings required for use in the embodiments or related technical descriptions. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0042] Figure 1A flowchart of a method for recommending patient examination items based on a medical knowledge graph provided by an embodiment of the present invention;

[0043] Figure 2 A flowchart of an embodiment of the present invention for generating a patient status subgraph based on patient input information, processing a medical knowledge graph using an improved TransD model, and determining a unique representation of entity-relationship pairs;

[0044] Figure 3 The embodiment of the present invention provides a flowchart for prioritizing inspection items according to a similarity value matrix, optimizing and adjusting the priorities through threshold screening and combined with graph neural network verification to generate a final recommendation list. DETAILED DESCRIPTION

[0045] In order to make the purpose, technical solutions and advantages of this application more clear, the following further describes this application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain this application and are not intended to limit this application.

[0046] In one embodiment, Figure 1 As shown, this application provides a method for recommending patient examination items based on a medical knowledge graph, which may include the following steps:

[0047] Step S101: Acquire historical data in the medical field, extract key information to generate entity relationship triples, and construct a medical knowledge graph.

[0048] Specifically, multi-source, heterogeneous medical history data, including electronic medical records, diagnostic reports, and examination records, is first collected and pre-processed for cleaning, denoising, and standardization. Using named entity recognition algorithms and information extraction techniques, entities such as patients, symptoms, diagnoses, and examination items are accurately extracted from the pre-processed data. Relationships are then extracted using a combination of rule matching and machine learning, generating triples of the form "head entity-relationship-tail entity." These triples are stored in a graph database, and missing relationships are filled in using a graph completion algorithm. Quality assessment is performed through redundancy detection and contradiction verification, ultimately constructing a structurally complete and semantically accurate medical knowledge graph.

[0049] Step S102: Generate a patient status subgraph based on the patient input information, use the improved TransD model to process the medical knowledge graph, and determine the unique representation of the entity-relationship pair.

[0050] The system receives patient input, including symptom descriptions, age, and gender. Using semantic parsing and graph query algorithms, it filters entities and relationships related to the patient information from the medical knowledge graph, generating a patient status subgraph. An improved TransD model is used to vectorize the triples in this subgraph, defining meaning vectors and projection vectors for each entity and relationship. Using an adaptive learning mechanism to dynamically adjust the mapping matrix parameters, the entity and relationship vectors are mapped into a unified semantic space, ensuring that each entity-relationship pair receives a unique structured representation, capturing the diverse semantic associations of medical data.

[0051] Step S103: Map the patient status sub-graph to a unique representation space, obtain vector representations of entities related to the patient and the examination items, and calculate similarity values ​​between the vectors to obtain a similarity value matrix.

[0052] The mapping matrix generated by the TransD model is obtained, and the entity nodes in the patient state subgraph are projected into a unique semantic representation space to extract the semantic vector of the patient entity. Simultaneously, entities related to examination items are retrieved from the medical knowledge graph and converted into examination item entity vectors using the same mapping matrix. Using an optimized cosine similarity algorithm, dot products and modulus normalization are performed on the patient entity vectors and examination item entity vectors. The cosine values ​​of the angles between all vector pairs are calculated in batches to generate a similarity matrix with the dimension "number of patient entities × number of examination item entities" to quantify the semantic relevance between the two.

[0053] Step S104: Prioritize the inspection items according to the similarity value matrix, and optimize and adjust the priorities through threshold screening and combined with graph neural network verification to generate a final recommendation list.

[0054] Specifically, the similarity value matrix is ​​sorted in descending order by column, and a similarity ranking vector for each examination item for all patients is generated to form an initial priority list. An adaptive threshold algorithm is used to filter low-similarity items to obtain a preliminary candidate set. A graph neural network is used to traverse the medical knowledge graph to verify the semantic association strength between candidate examination items and patient symptoms, and generate a verification score matrix. A multi-objective optimization model is constructed based on the similarity values ​​and verification scores. A Pareto optimal solution is solved using a genetic algorithm, and the examination items are secondary sorted. Then, an attention mechanism is used to screen high-confidence items directly related to patient symptoms. Finally, a recommended list of examination items is generated that takes into account both recommendation accuracy and medical rationality.

[0055] In one embodiment, obtaining historical data in the medical field, extracting key information to generate entity relationship triples, and constructing a medical knowledge graph may include the following steps:

[0056] Step S201: Acquire multi-source medical field historical data; the historical data includes patient medical records and diagnosis and treatment records.

[0057] Step S202 , using named entity recognition algorithm and information extraction technology to extract key information such as patients, symptoms, diagnoses and examination items from the pre-processed historical data.

[0058] Step S203 , using a combination of rule matching and machine learning technology to perform relationship extraction on the extracted key information to generate entity relationship triples.

[0059] Preferably, a combination of rule matching and machine learning technology is used to extract relationships from the extracted key information and generate entity relationship triplets: first, a relationship extraction rule base for the medical field is established, such as structured patterns such as "symptoms + cause + diagnosis" and "patient + acceptance + examination items", and preliminary association matching is performed on the extracted entities. At the same time, machine learning models such as convolutional neural networks (CNN) or graph neural networks (GNN) are used to train relationship classifiers with entity context semantic features and syntactic structure features as input to identify implicit relationships between entities. The rule matching results are fused and verified with the machine learning prediction results, and conflicting relationships are corrected through manual annotation or secondary model training. Finally, standardized triplets in the form of "head entity-relationship-tail entity" are generated, such as (fever, related symptoms, blood routine test), to ensure the accuracy of relationship extraction and semantic consistency.

[0060] Step S204: Optimize the graph structure by combining the entity relationship triples with the graph completion algorithm to construct a complete medical knowledge graph.

[0061] Step S205: perform a quality assessment on the medical knowledge graph, correct erroneous information through redundancy detection and contradiction verification, and obtain a searchable medical knowledge graph.

[0062] Specifically, the algorithm first acquires historical medical data from multiple sources, including patient medical records and treatment notes. After preprocessing this data, named entity recognition algorithms and information extraction techniques are used to extract key information, such as patients, symptoms, diagnoses, and examination items. Next, a combination of rule matching and machine learning techniques is used to extract relationships from this extracted key information, generating entity-relationship triples. These entity-relationship triples are then combined with a graph completion algorithm to optimize the graph structure and construct a complete medical knowledge graph. Finally, the medical knowledge graph undergoes a quality assessment, correcting erroneous information through redundancy detection and contradiction checking, resulting in a searchable medical knowledge graph.

[0063] This implementation systematically integrates historical medical data, transforming scattered patient medical records and treatment notes into structured entity-relationship triples, effectively addressing the fragmented and difficult-to-associate nature of medical data. The optimized graph is structurally complete and accurate, helping to improve the accuracy and medical rationality of recommendations while also facilitating the query and application of medical data.

[0064] In one embodiment, Figure 2 As shown in the figure, a patient status subgraph is generated based on the patient input information, and the improved TransD model is used to process the medical knowledge graph to determine the unique representation of entity-relationship pairs, including:

[0065] Step S301: Based on the patient input information and the entities and relationships in the medical knowledge graph, a patient status subgraph is generated through a graph query and screening algorithm.

[0066] Step S302: vectorize the triples in the patient status subgraph using the improved TransD model to generate meaning vectors and projection vectors for entities and relationships, respectively.

[0067] Preferably, an improved TransD model is used to vectorize the triplets in the patient state subgraph to generate meaning vectors and projection vectors for entities and relationships respectively: first, the triple data in the patient state subgraph is obtained, duplicate records and invalid associations are removed, and a preprocessed triple set is formed. Based on the improved TransD model architecture, two sets of vectors are defined for each entity and relationship - the meaning vector is used to represent the semantic content (such as the symptom attribute of "fever"), and the projection vector is used to construct a mapping matrix (such as mapping the association between "fever" and "blood routine examination" to a unified space). The multi-head attention mechanism is used to enhance the model's ability to capture the entity relationship features in the triplet. For example, when processing the "patient-symptom-cough" triplet, the semantic association between the symptom entity and the patient entity is strengthened through the attention weight. The training process is optimized using batch normalization and residual connection, and the vector parameters are iteratively adjusted so that the head entity, relationship, and tail entity of each triplet satisfy the mapping relationship of "head entity meaning vector + relationship meaning vector ≈ tail entity meaning vector" in the vector space, and finally a structured meaning vector and projection vector are generated.

[0068] Step S303 : dynamically adjust the mapping matrix parameters using an adaptive learning mechanism, map the meaning vector and the projection vector to the same semantic space, and determine a unique representation of the entity-relationship pair.

[0069] Furthermore, based on the patient input information and the entities and relationships in the medical knowledge graph, a patient status subgraph is generated through a graph query and screening algorithm. Specifically, the symptoms, medical history and other information entered by the patient are semantically parsed and converted into queryable graph database retrieval conditions. Entities and relationships directly related to the patient information are screened from the medical knowledge graph to construct a subgraph containing the patient's personalized information. Next, an improved TransD model is used to vectorize the triples in the patient status subgraph, generating a meaning vector and a projection vector for each entity and relationship. The meaning vector represents its semantic connotation, and the projection vector is used to construct a mapping matrix. Finally, an adaptive learning mechanism is used to dynamically adjust the mapping matrix parameters, mapping the meaning vector and the projection vector to the same semantic space, thereby determining the unique representation of the entity-relationship pair.

[0070] This embodiment combines personalized patient information with the structured knowledge of the medical knowledge graph to generate a targeted patient status subgraph. It also implements vectorized representation of entities and relationships through an improved TransD model. This effectively captures the semantic connections between entities in medical data, addressing the data sparsity and inaccurate semantic representation issues of traditional methods. It provides a semantically consistent and well-structured vector representation foundation for subsequent similarity calculations and examination item recommendations, improving the accuracy and interpretability of the recommendation system.

[0071] In one embodiment, the improved TransD model is used to vectorize the triples in the patient state subgraph to generate the meaning vectors and projection vectors of the entities and relationships, respectively. The following steps may be included:

[0072] Step S401: Acquire triplet data in the patient status sub-graph, remove noise data and invalid records from the data, and obtain pre-processed triplet data.

[0073] In step S402, the initial model parameters are set based on the improved TransD model structure, and a multi-head attention mechanism is used to enhance the ability to capture entity and relationship features in triples to obtain an optimized encoder.

[0074] Step S403: Encode the preprocessed triples using the optimized encoder to generate initial meaning vectors and projection vectors of entities and relationships, respectively.

[0075] Step S404: input the initial meaning vector and the projection vector into the residual network for optimization according to the residual learning strategy to obtain the final entity and relationship meaning vector and projection vector.

[0076] Specifically, the triplet data in the patient status subgraph is first acquired and denoised to remove noisy data and invalid records, resulting in preprocessed triples. Initial model parameters are then set based on the improved TransD model structure. By employing a multi-head attention mechanism, the model's ability to capture entity and relationship features in the triplet is enhanced, resulting in an optimized encoder. This optimized encoder is then used to encode the preprocessed triplet, generating initial meaning vectors and projection vectors for the entities and relationships, respectively. Finally, based on a residual learning strategy, the initial meaning vectors and projection vectors are input into a residual network for optimization, yielding the final entity and relationship meaning vectors and projection vectors.

[0077] This example effectively processes triplet data in the patient status subgraph, removing noise and invalid information to ensure data quality. By combining an improved TransD model with a multi-head attention mechanism and residual learning strategy, it can more accurately capture the characteristics of entities and relationships in triples, generate more accurate meaning vectors and projection vectors, and map entities and relationships into a unified semantic space, improving the accuracy and effectiveness of the entire recommendation method.

[0078] In one embodiment, mapping the patient state subgraph to a unique representation space, obtaining vector representations of entities related to the patient and examination items, and calculating similarity values ​​between the vectors to obtain a similarity value matrix may include the following steps:

[0079] Step S501: Obtain a mapping matrix between the patient state sub-map and the unique representation space, and perform space conversion preprocessing on the patient state sub-map.

[0080] Step S502 : Project the entity nodes in the patient state subgraph to a unique representation space based on the mapping matrix, and extract the patient entity semantic vector.

[0081] Step S503: retrieve entity nodes related to the examination items from the medical knowledge graph, and obtain the examination item entity vectors through the same mapping matrix conversion.

[0082] Step S504 : Calculate the cosine values ​​of the angles between all patient entity semantic vectors and examination item entity vectors using an optimized cosine similarity algorithm to generate a similarity value matrix.

[0083] Specifically, the mapping matrix between the patient state subgraph obtained by training the improved TransD model and the unique representation space is obtained, and the entity and relationship data in the patient state subgraph are normalized, formatted, and other spatial transformation preprocessing is performed. Secondly, based on the mapping matrix, each entity node in the patient state subgraph is projected into the unique representation space to complete the semantic feature extraction and obtain the patient entity semantic vector. Then, the entity nodes related to the examination items are retrieved from the constructed medical knowledge graph, and the same mapping matrix is ​​used for spatial transformation to obtain the examination item entity vector. Finally, the optimized cosine similarity algorithm is used to perform dot product operations and modulus normalization on all patient entity semantic vectors and examination item entity vectors, calculate the cosine value of the angle between the vectors, and generate a similarity value matrix with the dimension of the number of patient entities multiplied by the number of examination item entities.

[0084] This embodiment uses a unified mapping matrix to map patient status subgraphs and examination item entities into the same semantic space, ensuring consistency in data representation and effectively addressing the heterogeneity of medical data. An optimized cosine similarity algorithm accurately quantifies the degree of semantic association between entities. This improves the recommendation system's ability to integrate personalized patient information and medical knowledge, enhancing the accuracy and medical rationality of recommendation results and providing more valuable reference for clinical decision support.

[0085] In one embodiment, the similarity value matrix can be calculated using the following formula:

[0086]

[0087] Among them, S represents the similarity value matrix, S i,j represents the similarity between the symptom of the i-th patient and the j-th examination item, represents the mapping vector of the i-th patient entity in the unique representation space, represents the mapping vector of the j-th inspection item entity in the unique representation space, Λ i,j represents the prior association weight between the i-th patient symptom and the j-th examination item in the knowledge graph, m represents the number of patient entities, and n represents the number of examination item entities.

[0088] This example quantifies the semantic relevance between patient and examination item entities by combining vector dot products with modulus-normalized cosine similarity. It also incorporates the prior association weights of the knowledge graph, incorporating inherent medical knowledge into the computational process. By leveraging data-driven vector representations and incorporating domain knowledge constraints, this approach effectively avoids recommendations that contradict medical common sense, improving both accuracy and medical rationality.

[0089] In one embodiment, Figure 3As shown, the inspection items are prioritized according to the similarity value matrix, and the priorities are optimized and adjusted through threshold screening and combined with graph neural network verification to generate the final recommendation list, which can include the following steps:

[0090] Step S601 : sorting the similarity value matrix in descending order by column to obtain a similarity ranking vector of each examination item for all patients, and generating an initial examination item priority list based on the vector.

[0091] Step S602 : Adopting an adaptive threshold screening algorithm to perform dynamic threshold filtering on the priority list, retaining inspection items with similarity values ​​higher than the dynamic threshold, and obtaining a preliminary recommendation candidate set.

[0092] Step S603: Utilize the graph neural network combined with the semantic relationship between entities in the medical knowledge graph to verify the association strength of the examination items in the preliminary recommendation candidate set and generate a verification score matrix.

[0093] Preferably, a graph neural network is used in combination with the semantic relationships between entities in the medical knowledge graph to verify the association strength of the examination items in the preliminary recommended candidate set and generate a verification score matrix: first, the examination items in the preliminary recommended candidate set are used as target nodes, and their multi-hop association paths with patient symptom entities, diagnosis entities, etc. are extracted from the medical knowledge graph, such as the relationship chain of "cough → respiratory infection → chest CT". The message passing mechanism of the graph neural network (GNN) is used to pass and aggregate the feature vectors of each node along the edge. For example, the semantic information of neighboring nodes is fused through the aggregation function of the GraphSAGE algorithm (such as mean aggregation and LSTM aggregation) to capture the indirect association between the examination items and patient symptoms. During model training, the triplets in the knowledge graph are used as supervisory signals, and the parameters are optimized by minimizing the cross-entropy loss function so that the generated verification score can reflect the actual medical association strength between the examination items and patient symptoms. Finally, a verification score between 0 and 1 is generated for each examination item, forming a verification score matrix with the same dimension as the similarity value matrix.

[0094] Step S604: construct a multi-objective optimization model based on the similarity value and the verification score, use a genetic algorithm to perform secondary sorting on the priorities of the inspection items, and use the Pareto optimal solution to obtain an optimized list.

[0095] Preferably, for medical examination item recommendation, the following objective function can be defined:

[0096] Maximize the similarity value: Among them, S i,j represents the similarity value between patient i and examination item j, x i,j ∈{0,1} indicates whether examination item j is recommended to patient i.

[0097] Maximize the association strength validation score: Among them, V i,j Represents the verification score of the association strength between examination item j and patient i's symptoms in the knowledge graph.

[0098] Minimize the number of inspection items:

[0099] Minimize consumption of medical resources: Among them, C j Represents the resource consumption cost (such as time, expenses, etc.) of inspecting project j.

[0100] Step S605: The final screening strategy based on the attention mechanism adjusts the weight of the optimized list, extracts high-confidence examination items directly related to the patient's symptoms, and generates a final recommendation list.

[0101] Specifically, the similarity value matrix is ​​sorted in descending order by column to obtain the similarity ranking vector of each examination item for all patients, and then an initial examination item priority list is generated. Next, an adaptive threshold screening algorithm is used to dynamically filter the priority list, retaining examination items with similarity values ​​higher than the dynamic threshold to obtain a preliminary recommendation candidate set. Then, a graph neural network is used in combination with the semantic relationship between entities in the medical knowledge graph to verify the association strength of the examination items in the candidate set and generate a verification score matrix. A multi-objective optimization model is then constructed based on the similarity value and verification score, and a genetic algorithm is used to perform a secondary sorting of the examination item priorities. The optimized list is obtained through the Pareto optimal solution. Finally, a final screening strategy based on the attention mechanism adjusts the weights of the optimized list, extracts high-confidence examination items directly related to the patient's symptoms, and generates a final recommendation list.

[0102] This embodiment combines data-driven similarity calculation with knowledge-driven semantic verification through a multi-stage optimization strategy. Adaptive threshold screening dynamically filters out low-relevance items, reducing redundant recommendations; graph neural network verification strengthens the medical logic support for recommendations; multi-objective optimization combined with genetic algorithms balances recommendation accuracy with constraints such as medical resource consumption; and an attention mechanism further focuses on high-confidence items. This effectively improves the accuracy and clinical practicality of recommendation results, helping doctors quickly identify key examination items, reducing unnecessary medical examinations, and optimizing patient care processes.

[0103] It should be understood that, although the various steps in the flowcharts involved in the various embodiments described above are displayed in sequence according to the instructions of the arrows, these steps are not necessarily executed in sequence in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order restriction on the execution of these steps, and these steps can be executed in other orders. Moreover, at least a portion of the steps in the flowcharts involved in the various embodiments described above can include multiple steps or multiple stages, and these steps or stages are not necessarily executed and completed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily to be carried out in sequence, but can be executed in turn or alternately with other steps or at least a portion of steps or stages in other steps.

[0104] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of the method for recommending patient examination items based on a medical knowledge graph as described above are implemented.

[0105] In one embodiment, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps in the above-mentioned method embodiments are implemented.

[0106] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to the partial description of the method embodiments. The device embodiments described above are merely illustrative, wherein the components described as separate parts may or may not be physically separated, and the parts displayed as units may or may not be physical units, that is, they may be located in one place, or they may be distributed on multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the disclosed solution. A person of ordinary skill in the art can understand and implement it without expending creative work.

[0107] The above-described embodiments merely represent several implementation methods of the embodiments of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that a person skilled in the art may make various modifications and improvements without departing from the concept of the embodiments of the present application, and these modifications and improvements fall within the scope of protection of the embodiments of the present application.

Claims

1. A method for recommending patient examination items based on medical knowledge graph, characterized in that: The method comprises: Obtain historical data in the medical field, extract key information to generate entity relationship triples, and build a medical knowledge graph; Generate a patient status subgraph based on patient input information, process the medical knowledge graph using an improved TransD model, and determine a unique representation of entity-relationship pairs; Mapping the patient state subgraph to the unique representation space, obtaining vector representations of entities related to the patient and the examination items, and calculating similarity values ​​between the vectors to obtain a similarity value matrix; The inspection items are prioritized according to the similarity value matrix, and the priorities are optimized and adjusted through threshold screening and combined with graph neural network verification to generate a final recommendation list.

2. The method according to claim 1, characterized in that The process of acquiring historical data in the medical field, extracting key information to generate entity relationship triples, and constructing a medical knowledge graph includes: Acquire multi-source historical medical data; the historical data includes patient medical records and diagnosis and treatment records; Using named entity recognition algorithms and information extraction technology to extract key information such as patients, symptoms, diagnoses, and examination items from the pre-processed historical data; Using a combination of rule matching and machine learning technology to extract relationships from the extracted key information and generate entity relationship triples; The entity relationship triples are combined with a graph completion algorithm to optimize the graph structure and construct a complete medical knowledge graph; The quality of the medical knowledge graph is evaluated, and erroneous information is corrected through redundancy detection and contradiction verification to obtain a queryable medical knowledge graph.

3. The method according to claim 1, characterized in that The patient status subgraph is generated based on the patient input information, and the improved TransD model is used to process the medical knowledge graph to determine the unique representation of the entity-relationship pair, including: Generate a patient status subgraph based on the patient input information combined with the entities and relationships in the medical knowledge graph through graph query and screening algorithms; Using an improved TransD model to vectorize the triples in the patient status subgraph, and respectively generate meaning vectors and projection vectors of the entities and relationships; An adaptive learning mechanism is used to dynamically adjust the mapping matrix parameters, map the meaning vector and the projection vector into the same semantic space, and determine a unique representation of the entity-relationship pair.

4. The method according to claim 3, characterized in that The improved TransD model is used to vectorize the triples in the patient status subgraph to generate the meaning vectors and projection vectors of the entities and relationships, respectively, including: Acquire triplet data in the patient state sub-graph, remove noise data and invalid records from the data, and obtain preprocessed triplet; The initial model parameters are set based on the improved TransD model structure, and a multi-head attention mechanism is used to enhance the ability to capture the entity and relationship features described in the triples, resulting in an optimized encoder. Encoding the preprocessed triples using the optimized encoder to generate initial meaning vectors and projection vectors of entities and relationships respectively; According to the residual learning strategy, the initial meaning vector and the projection vector are input into the residual network for optimization to obtain the final entity and relationship meaning vector and projection vector.

5. The method according to claim 1, wherein Mapping the patient state subgraph to the unique representation space, obtaining vector representations of entities related to the patient and the examination items, and calculating similarity values ​​between the vectors to obtain a similarity value matrix include: Obtaining a mapping matrix between the patient state sub-map and the unique representation space, and performing space conversion preprocessing on the patient state sub-map; Projecting the entity nodes in the patient state subgraph to the unique representation space based on the mapping matrix, and extracting the patient entity semantic vector; Retrieving entity nodes related to examination items from the medical knowledge graph, and obtaining examination item entity vectors through the same mapping matrix conversion; An optimized cosine similarity algorithm is used to calculate the cosine values ​​of the angles between all the patient entity semantic vectors and the examination item entity vectors to generate a similarity value matrix.

6. The method according to claim 5, characterized in that The similarity value matrix is ​​calculated by the following formula: Among them, S represents the similarity value matrix, S i,j represents the similarity between the symptom of the i-th patient and the j-th examination item, represents the mapping vector of the i-th patient entity in the unique representation space, represents the mapping vector of the j-th inspection item entity in the unique representation space, Λ i,j represents the prior association weight between the i-th patient symptom and the j-th examination item in the knowledge graph, m represents the number of patient entities, and n represents the number of examination item entities.

7. The method according to claim 1, characterized in that Prioritizing the inspection items according to the similarity value matrix, optimizing and adjusting the priorities through threshold screening and combined with graph neural network verification to generate a final recommendation list, including: Sorting the similarity value matrix in descending order by column to obtain a similarity ranking vector of each examination item for all patients, and generating an initial examination item priority list based on the vector; Adopting an adaptive threshold screening algorithm to perform dynamic threshold filtering on the priority list, retaining the inspection items with similarity values ​​higher than the dynamic threshold, and obtaining a preliminary recommendation candidate set; Utilizing a graph neural network in combination with semantic relationships between entities in the medical knowledge graph, the association strength of the examination items in the preliminary recommendation candidate set is verified to generate a verification score matrix; A multi-objective optimization model is constructed based on the similarity values ​​and the verification scores, a genetic algorithm is used to perform secondary sorting of the priorities of the inspection items, and an optimization list is obtained using a Pareto optimal solution; The final screening strategy based on the attention mechanism adjusts the weights of the optimized list, extracts high-confidence examination items directly related to the patient's symptoms, and generates a final recommendation list.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

Images