Recommendation method for diagnosis and treatment of septic shock based on knowledge graph

Abstract

The present invention discloses a method for recommending the diagnosis and treatment of septic shock based on a knowledge graph. The method comprises the following steps: S1: defining relevant information in a knowledge graph for septic shock; S2: introducing symmetric, inverse, and combined relationships, and using the RotatE model to represent the knowledge graph; S3: using word2vec and the RotatE model to recommend treatment methods based on patient symptoms. The present invention constructs a knowledge graph based on five entities and seven relationships, introduces symmetric, inverse, and combined relationships, and uses the RotatE model to represent the knowledge graph for septic shock. Considering that some patient symptoms that may be queried in reality do not exist in the knowledge graph, the word2vec model is retrained to recommend treatment methods for symptoms that are not in the knowledge graph, assisting doctors in making clinical diagnosis and treatment decisions to achieve the best treatment results.

Description

Technical Field

[0001] The present invention belongs to the field of smart medicine and relates to an intelligent diagnosis and treatment decision-making assistance method for septic shock based on a knowledge graph, and specifically to a septic shock diagnosis and treatment recommendation method that combines a word2vec model based on a knowledge graph with a RotatE model. Background Art

[0002] Septic shock is one of the most serious complications leading to death in intensive care units (ICUs). Globally, over 30 million people are admitted to ICUs each year, and nearly 6 million do not survive. Septic shock can be caused by a variety of pathogens, including bacteria, viruses, and fungi. It's difficult to diagnose a patient's infection quickly, let alone determine a treatment. Therefore, research into the timely development of appropriate treatment strategies is crucial. It can help doctors detect and control the source of infection early, regulate the body's immune response, reduce organ damage, and ultimately improve patient survival.

[0003] A knowledge graph is a semantic network that reveals the relationships between entities and can effectively organize and utilize massive amounts of knowledge data. In the medical field, knowledge graphs can help doctors and patients obtain more accurate, comprehensive, and accessible medical knowledge, improving the efficiency and quality of diagnosis and treatment. In recent years, medical diagnostic systems based on knowledge graphs have received widespread attention and research. These systems leverage the reasoning capabilities of knowledge graphs, combined with technologies such as natural language processing and machine learning, to provide intelligent assistance and analysis of patient symptoms, diseases, and treatment plans. Currently, the commonly used knowledge graph construction method generally involves first collecting relevant text data, then performing entity recognition and relationship extraction on the text data, extracting triples from the text data, and completing the construction of the entire knowledge graph. However, current medical knowledge graphs do not specifically target septic shock. Summary of the Invention

[0004] In order to solve the problems existing in the technical background, the present invention provides a method for recommending the diagnosis and treatment of infectious shock based on a knowledge graph. The present invention uses a medical corpus related to infectious shock to retrain a word2vec model. The patient's symptoms are mapped into an embedding vector through word2vec. After a linear layer and an activation function, the symptom vector is mapped to the semantic space of the RotatE model. The embedding vector of the predicted tail entity is calculated based on the specific relationship vector. Then, based on the relationship in the knowledge graph, the vector of the tail entity is calculated. By performing similarity matching with the entity vector of the corresponding category in the infectious shock knowledge graph, the corresponding treatment method or symptom-related disease is recommended to assist doctors in clinical diagnosis.

[0005] The purpose of the present invention is achieved through the following technical solutions:

[0006] A method for recommending diagnosis and treatment of septic shock based on a knowledge graph includes the following steps:

[0007] Step S1: defining the relevant information of the septic shock knowledge graph;

[0008] Step S2: Introduce symmetric relations, inverse relations, and combination relations, and use the RotatE model to represent the septic shock knowledge graph;

[0009] Step S3: Use word2vec and RotatE models to recommend diagnosis and treatment methods based on the patient's symptoms.

[0010] Compared with the prior art, the present invention has the following advantages:

[0011] This paper collects data related to septic shock from Baidu Medical Dictionary, septic shock guidelines, and medical literature, and constructs a knowledge graph dedicated to septic shock based on five entities and seven relationships. It introduces symmetric relationships, inverse relationships, and combination relationships, and uses the RotatE model to represent the septic shock knowledge graph. Considering that in reality, some patient symptoms that may be queried do not exist in the knowledge graph, the word2vec model is retrained to recommend diagnosis and treatment methods for symptoms that do not exist in the knowledge graph, assisting doctors in making clinical diagnosis and treatment decisions to achieve the best treatment effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 It is the overall structure and workflow diagram of the present invention;

[0013] Figure 2 Recommended results. DETAILED DESCRIPTION

[0014] The technical solution of the present invention is further described below with reference to the accompanying drawings, but is not limited thereto. Any modification or equivalent replacement of the technical solution of the present invention that does not depart from the spirit and scope of the technical solution of the present invention should be included in the scope of protection of the present invention.

[0015] The present invention provides a method for recommending the diagnosis and treatment of infectious shock based on a knowledge graph. The method introduces symmetric relationships, inverse relationships, and combination relationships for a constructed infectious shock knowledge graph, and uses the RotatE model to represent and learn the knowledge graph. Taking into account the application situation where the patient's symptoms do not exist in the knowledge graph, the word2vec model is retrained with the relevant corpus of infectious shock, and the patient's symptoms are mapped into vectors. The vectors are mapped to the semantic space of RotatE through linear layers and activation functions, and then the vectors of the tail entities are calculated based on the relationships in the knowledge graph. By performing similarity matching with the entity vectors of the corresponding categories in the infectious shock knowledge graph, the top K entities similar to the tail entity are obtained to complete the recommendation of treatment methods or related diseases. Figure 1 As shown, the following steps are included:

[0016] Step S1: Define the information related to the septic shock knowledge graph. The specific steps are as follows:

[0017] Step S101: crawling data related to septic shock from Baidu Medical Dictionary, septic shock treatment guidelines, and related medical literature;

[0018] Step S102: Define the entity category of the knowledge graph of septic shock. Based on the characteristics of septic shock, septic shock is associated with diseases of many organ systems, so it is necessary to define disease-related entities. There are many different treatment methods for each disease. In order to further refine the treatment plan, the treatment is divided into physical therapy and drug therapy. In addition, the pathogens of septic shock are also diverse, so an entity category of pathogens and their causes is added. Therefore, the entities are divided into five categories: (1) disease, (2) symptoms, (3) physical therapy, (4) drug therapy, and (5) pathogens and their causes.

[0019] Step S103: Define the relationship categories of the septic shock knowledge graph, that is, divide the relationships into the following seven categories: clinical_manifestation: the relationship between disease and symptoms; physical: the relationship between disease and physical treatment; medical: the relationship between disease and drug treatment; relevant: the correlation between diseases; cause: the relationship between pathogens and their causes and diseases; same: synonyms; treatment: the treatment method corresponding to the symptoms.

[0020] Step S2: Introduce symmetric relationships, inverse relationships, and combination relationships, and use the RotatE model to represent the septic shock knowledge graph. The specific steps are as follows:

[0021] Step S201: Introducing a symmetric relationship: The constructed septic shock knowledge graph has a relationship of (entity 1, same, entity 2). Synonyms can exchange positions with each other, so a triple (entity 2, same, entity 1) is added to the knowledge graph, thus introducing a symmetric relationship into the knowledge graph.

[0022] Step S202: Introducing an inverse relationship: There is a (disease, clinical_manifestation, symptom) triple in the knowledge graph. In order to facilitate subsequent prediction based on symptoms, a relationship is added from symptoms to diseases: clinical_manifestation_inverse relationship, and the corresponding triple (symptom, clinical_manifestation_inverse, disease) is added. This relationship is the inverse relationship of clinical_manifestation, thus introducing an inverse relationship in the knowledge graph.

[0023] Step S203: Introducing a combination relationship: After adding the clinical_manifestation_inverse relationship, (disease, physical, treatment method) still exists in the knowledge graph. Combined with (symptoms, clinical_manifestation_inverse, disease), we can obtain the triple (symptoms, treatment, treatment method). It can be observed that treatment is composed of physical and clinical_manifestation_inverse, that is, after the inverse relationship is introduced, a combination relationship exists in the knowledge graph.

[0024] Step S204: Use the RotatE model to represent the knowledge graph with symmetric relationships, inverse relationships, and combination relationships. The specific steps are as follows:

[0025] Step S20401: Map the head entity h and the tail entity t to complex embeddings, and then define the function mapping of each relation r as performing element-wise rotation on the head entity h to obtain the tail entity t. That is, for a triple, we have:

[0026]

[0027] in, is the Hadmard product, r i |=1 means the modulus of the constraint relationship is 1.

[0028] Step S20402: Define the distance function d of RotatE r (h,t):

[0029]

[0030] Step S20403: Use the loss function of negative sampling loss to effectively optimize the distance-based model, and use the adaptive adversarial negative sampling method to extract negative triplets based on the current embedding model. Specifically, the formula for extracting the distribution of negative triplets is:

[0031]

[0032] Among them, p(h′ j ,r,t′ j |{h i ,r i ,t i}) is the probability of negative sample distribution, α is the sampling temperature, f r (h′ j ,t′ j ) represents the score function of the relationship between the currently extracted head entity and tail entity, f r (h′ i ,t′ i ) refers to the score function of the relationship between any head entity and tail entity in all the extracted negative triples, h′ j ,t′ j and h′ i ,t′ i are the head and tail of the negative sample.

[0033] In addition, in order to reduce the cost of the sampling process, p(h′ j ,r,t′ j |{h i ,r i ,t i}) as the weight of the negative sample. Therefore, the final negative sampling loss formula for adaptive adversarial training is as follows:

[0034]

[0035] Among them, γ is a fixed margin, in order to prevent the model from overfitting, σ is a sigmoid function, (h′ i ,r,t′ i ) is the i-th negative triplet.

[0036] After RotatE represents the knowledge graph, the entities and relationships in the knowledge graph are mapped into vectors.

[0037] Step S3: Use word2vec and RotatE models to recommend diagnosis and treatment methods based on the patient's symptoms. The specific steps are as follows:

[0038] Step S301: Because the RotatE model is a translation-based model that learns vector representations of entities and relationships based on existing knowledge graph data, if certain symptoms of the patient do not exist in the knowledge graph, that is, when encountering a new entity, it cannot give a reasonable vector representation. Therefore, a word2vec model is retrained using a corpus related to septic shock.

[0039] Step S302: Through the linear layer and activation function, the embedding vector of the entity in the knowledge graph obtained by the word2vec model is mapped to the semantic space where the RotatE model is located. The formula is as follows:

[0040] h RoataE =σ(Wh w2v +b)

[0041] Among them, h RoataE is the vector of the RotatE model vector space, h w2v is the embedding vector obtained by the word2vec model, and W and b are learnable parameters.

[0042] Step S303: First, the patient's symptoms are mapped into embedded vectors using the trained word2vec. After passing through a linear layer and an activation function, the symptom vectors are mapped into the semantic space of the RotatE model, and the embedded vectors of the relationships are obtained from the septic shock knowledge graph learned through the RotatE representation.

[0043] Step S304: Based on the specific relations treatment and clinical_manifestation_inverse in the knowledge graph, the embedding vector of the predicted tail entity is obtained. Then, similarity matching is performed on the embedding vectors of entities such as physical treatment methods, drug treatment methods, and related diseases in the knowledge graph to obtain the top K matching entities. These top K entities are fed back to the doctor to assist the doctor in making treatment decisions. The specific steps are as follows:

[0044] S30401: Get the patient's symptom embedding and the embedding of the relationship treatment, based on Obtain the embedding vector of the corresponding tail entity, obtain the set of physical therapy methods and drug therapy methods in the knowledge graph, perform vector similarity matching, and obtain the top K matching physical therapy methods and drug therapy methods to assist doctors in diagnosis and treatment. The K value is 5, and different K values can also be selected to control the number of related diseases, physical therapy methods, or drug therapy methods recommended to doctors.

[0045] S30402: Get the patient's symptom embedding and the embedding of the relation clinical_manifestation_inverse, based on Obtain the embedding vector of the corresponding tail entity, obtain the set of diseases in the knowledge graph, perform vector similarity matching, and obtain the top K matching diseases to assist doctors in diagnosis and treatment.

[0046] For example, the current patient has clinical symptoms of "hypoxic respiratory failure", but the symptoms of "hypoxic respiratory failure" do not exist in the knowledge graph. First, the trained word2vec model of "hypoxic respiratory failure" is mapped into an embedding vector, and then through the linear layer and activation function, the patient's symptoms are mapped to the semantic space of the RotatE model. According to the treatment relationship, the tail vector of the treatment method is obtained, and the similarity is matched with the embedding vectors of the existing physical treatment methods and drug treatment methods in the knowledge graph to obtain the top K corresponding physical treatment methods. Recommending diseases that may be related to symptoms is based on the clinical_manifestation_inverse relationship, and the rest is similar to the above process. At this time, K is 5, and the recommendation results are as follows: Figure 2 shown.

Claims

1. A method for recommending diagnosis and treatment of septic shock based on knowledge graph, characterized by The method comprises the following steps: Step S1: defining the relevant information of the septic shock knowledge graph; Step S2: Introduce symmetric relations, inverse relations, and combination relations, and use the RotatE model to represent the septic shock knowledge graph; Step S3: Use word2vec and RotatE models to recommend diagnosis and treatment methods based on the patient's symptoms. The specific steps are as follows: Step S301: retraining a word2vec model using a corpus related to septic shock; Step S302: Through the linear layer and activation function, the embedding vector of the entity in the knowledge graph obtained by the word2vec model is mapped to the semantic space where the RotatE model is located. The formula is as follows: h RoataE =σ(W×h w2v +b) Among them, h RoataE is the vector of the RotatE model vector space, h w2v is the embedding vector obtained by the word2vec model, W and b are learnable parameters, and σ is the sigmoid function; Step S303: Map the patient's symptoms into an embedded vector using the trained word2vec. After passing through a linear layer and an activation function, the symptom vector is mapped into the semantic space of the RotatE model, and the embedded vector of the relationship is obtained from the septic shock knowledge graph learned through the RotatE representation. Step S304: Based on the specific relations treatment and clinical_manifestation_inverse in the knowledge graph, the embedding vector of the predicted tail entity is obtained. Then, similarity matching is performed on the embedding vectors of the physical treatment methods, drug treatment methods, and related disease entities in the knowledge graph to obtain the top K matching entities. These top K entities are fed back to the doctor to assist the doctor in making treatment decisions. The specific steps are as follows: S30401: Get the patient's symptom embedding and the embedding of the relation treatment. Define the function mapping of each relation r as performing element-wise rotation on the head entity h to obtain the tail entity t. Get the embedding vector of the corresponding tail entity, It is the Hadmard product, which obtains the set of physical therapy methods and drug therapy methods in the knowledge graph, performs vector similarity matching, and obtains the top K matching physical therapy methods and drug therapy methods to assist doctors in diagnosis and treatment; S30402: Get the patient's symptom embedding and the embedding of the relation clinical_manifestation_inverse, based on Obtain the embedding vector of the corresponding tail entity, obtain the set of diseases in the knowledge graph, perform vector similarity matching, and obtain the top K matching diseases to assist doctors in diagnosis and treatment.

2. The method for recommending diagnosis and treatment of septic shock based on knowledge graph according to claim 1 is characterized in that The specific steps of step S1 are as follows: Step S101: crawling data related to septic shock from Baidu Medical Dictionary, septic shock treatment guidelines, and related medical literature; Step S102: define the entity categories of the septic shock knowledge graph and divide the entities into five categories: (1) disease, (2) symptoms, (3) physical treatment methods, (4) drug treatment methods, and (5) pathogens and their causes; Step S103: Define the relationship categories of the septic shock knowledge graph, that is, divide the relationships into the following seven categories: clinical_manifestation: the relationship between disease and symptoms; physical: the relationship between disease and physical treatment; medical: the relationship between disease and drug treatment; relevant: the correlation between diseases; cause: the relationship between pathogens and their causes and diseases; same: synonyms; treatment: the treatment method corresponding to the symptoms.

3. The method for recommending diagnosis and treatment of septic shock based on knowledge graph according to claim 1 is characterized in that The specific steps of step S2 are as follows: Step S201: Introducing a symmetric relationship: The constructed septic shock knowledge graph has a relationship of (entity 1, same, entity 2). Synonyms can be interchanged, so a triple (entity 2, same, entity 1) is added to the knowledge graph. Step S202: Introducing an inverse relationship: There is a (disease, clinical_manifestation, symptom) triple in the knowledge graph. To facilitate subsequent prediction based on symptoms, a relationship is added from symptom to disease: clinical_manifestation_inverse relationship. The corresponding triple (symptom, clinical_manifestation_inverse, disease) is also added. This relationship is the inverse relationship of clinical_manifestation. Step S203: Introduce a combination relationship: combine (disease, physical, treatment method) with (symptom, clinical_manifestation_inverse, disease) in the knowledge graph to obtain the triple (symptom, treatment, treatment method); Step S204: Use the RotatE model to represent the knowledge graph that adds symmetric relationships, inverse relationships, and combination relationships.

4. The method for recommending diagnosis and treatment of septic shock based on knowledge graph according to claim 3 is characterized in that The specific steps of step S204 are as follows: Step S20401: Map the head entity h and the tail entity t to complex embeddings, and then define the function mapping of each relation r as performing element-wise rotation on the head entity h to obtain the tail entity t. That is, for the triple, we have: in, is the Hadmard product, r i |=1 means the modulus of the constraint relationship is 1; Step S20402: Define the distance function d of RotatE r (h,t): Step S20403: Use the loss function of negative sampling loss to optimize the distance-based model, and use the adaptive adversarial negative sampling method to extract negative triplets according to the current embedding model.

5. The method for recommending diagnosis and treatment of septic shock based on knowledge graph according to claim 4 is characterized in that In step S20403, the formula for extracting the negative triplet distribution is: Among them, p(h′ j ,r,t′ j |{h i ,r i ,t i }) is the probability of negative sample distribution, α is the sampling temperature, f r (h′ j ,t′ j ) represents the score function of the relationship between the currently extracted head entity and tail entity, f r (h′ i ,t′ i ) refers to the scoring function of the relationship between any head entity and tail entity in all the extracted negative triples.

6. The method for recommending diagnosis and treatment of septic shock based on knowledge graph according to claim 4 is characterized in that In step S20403, the final negative sampling loss formula of the adaptive adversarial training is as follows: Among them, γ is the fixed margin, σ is the sigmoid function, (h′ i ,r,t′ i ) is the i-th negative triplet.

Images