Personalized treatment decision system for depression based on knowledge graph

Abstract

The present application discloses a personalized treatment decision system for depression based on a knowledge graph, which belongs to the field of medical auxiliary decision technology. The patient-related data is collected by designing a collection module, and then a knowledge graph is constructed by a construction module. Then, the knowledge graph is optimized by an optimization module, and finally, a personalized auxiliary treatment plan is output through a personalized output module to provide auxiliary decision-making for the patient's treatment. The advantage of this application is that it integrates multi-source data and constructs a high-quality knowledge graph, which provides new ideas and methods for diagnosis and treatment in the medical field. This application also combines artificial intelligence technology to propose a personalized treatment decision system for depression based on a knowledge graph, and then optimizes the knowledge graph through operations such as entity extraction, relationship extraction, attribute extraction, entity alignment, entity linking, and relationship deduction. Finally, through multimodal data analysis and knowledge graph reasoning, intelligent support is provided for the diagnosis and treatment of depression.

Description

Technical Field

[0001] The present application relates to the field of medical decision-making assistance technology, and in particular to a knowledge graph-based personalized depression treatment decision-making system. Background Art

[0002] In current medical practice, the diagnosis and treatment of depression faces multiple challenges. Traditional diagnosis and treatment methods rely on the doctor's clinical experience and the patient's subjective description, which not only limits the accuracy of diagnosis, but also increases the risk of misdiagnosis and missed diagnosis. Existing treatment plans are often based on the results of large-scale clinical trials. Although they are universal, they are difficult to personalize for the specific conditions of different patients, thus affecting the treatment effect and patient satisfaction.

[0003] With the rapid development of big data and artificial intelligence technology, applying them to the medical field, especially the diagnosis and treatment of depression, has become an important research direction. By integrating and analyzing massive amounts of medical data and combining advanced algorithm models, more accurate and personalized diagnosis and treatment plans can be provided for patients with depression, improving the efficiency of diagnosis and treatment and the quality of life of patients. However, there is currently a lack of a system that can effectively integrate multi-source data, build a knowledge graph, and make personalized treatment decisions based on the knowledge graph. Summary of the invention

[0004] In response to the shortcomings of the existing technology, this application provides a personalized depression treatment decision system based on knowledge graphs to solve technical problems such as limited accuracy of traditional diagnosis and treatment, uneven distribution of medical resources, lack of personalization of treatment plans, and lack of ability to integrate multi-source data.

[0005] To achieve the above objectives, this application is implemented through the following technical solutions:

[0006] A knowledge graph-based personalized depression treatment decision system, including:

[0007] The collection module is used to collect and process the patient's relevant data;

[0008] A building module for integrating and systematizing medical knowledge based on a preset knowledge graph embedding model to construct a knowledge graph;

[0009] The optimization module is used to optimize the knowledge graph to improve the performance of the knowledge graph and thus derive medical auxiliary decision-making;

[0010] The personalized medical assistance module is used to generate personalized auxiliary treatment plan recommendations based on a preset personalized treatment decision model and the derived medical assistance decision, so as to provide auxiliary decision-making for patients.

[0011] Preferably, the patient-related data includes at least: biological marker data, clinical evaluation data, patient feedback data, and treatment response data.

[0012] Preferably, the building block further comprises:

[0013] The framework unit is used to build the knowledge graph framework, and the knowledge graph framework is expressed by the following formula:

[0014] ;

[0015] in, represents the knowledge graph framework, h represents the entity set, r represents the relationship set, and t represents the edge set;

[0016] The embedding unit is used to capture the complex characteristics of relations through two-dimensional deep convolution operations, map entities and relations to low-dimensional vector space to build a knowledge graph embedding model, and express the knowledge graph embedding model through the following formula:

[0017] ;

[0018] ;

[0019] in, Representing a collection of entities and relationship set The splicing result is expressed as , represents a two-dimensional convolution operation, is the convolution kernel, is the number of convolution kernels. The convolution operation captures the concatenated vector through a sliding window. The local features in DWConv() represent the deep two-dimensional convolution operation, which uses multiple convolution kernels W to extract the complex characteristics between relations and entities, and takes the sum of the output results of each convolution kernel. represents a nonlinear activation function, Represents the flattening operation, which converts the two-dimensional feature matrix generated by the convolution operation into a one-dimensional feature vector to facilitate subsequent matrix operations. Indicates that the rationality score of the triple is calculated by dot product , by combining the edge set The dot product of is calculated, which is measured by and The feature of represents whether the generated vector can well predict the edge set ;

[0020] The loss function unit is used to set the loss function of the knowledge graph embedding model. and and the relationship between them When embedded into the vector space, the loss function of the knowledge graph embedding model is expressed as follows:

[0021] ;

[0022] in, is a set of training triples, representing entities and The relationship between , is the relation score function obtained through the knowledge graph embedding model, representing the triple The score, is a preset hyperparameter used to distinguish positive samples from negative samples. are negative entity pairs used to train the model to distinguish correct relation triplets from incorrect negative samples.

[0023] Preferably, the optimization module further comprises:

[0024] A response unit, including a retrieval subunit and a generation subunit, wherein the response unit is used to extract and organize medical knowledge from relevant data of the patient, and jointly optimize the retrieval subunit and the generation subunit so that the dialogue output of the knowledge graph can generate a response that is more in line with the context and has higher accuracy;

[0025] A precision unit, which is used to optimize and process the medical knowledge in the knowledge graph to make it more precise. The precision unit includes a knowledge extraction subunit and a knowledge fusion subunit.

[0026] The inference unit is used to automatically learn complex relationships in the knowledge graph and perform reasoning to derive new knowledge.

[0027] Preferably, the operation process of the retrieval subunit is expressed by the following formula:

[0028] ;

[0029] ;

[0030] ;

[0031] in, It is the knowledge graph reasoning information, which contains structured entity and relationship information; is a document collection, which is the target collection for retrieval and contains all documents to be matched; It represents the retrieval result set, which is a document collection Inquiry Most relevant A collection of documents; It means to select the first Documents; is the similarity term, used for query Documents to be searched The similarity score of the query and documents to be searched The direct semantic similarity of and is the vector representation of the query and the document to be searched. and Representation vector and of the norm, which is the length or magnitude of a vector; Is a query in the knowledge graph and documents to be searched Entities in The relationship score of Representation vector of norm, which is the length of the vector; is an entity in the knowledge graph; h is an entity Embedding representation of ; is the embedding representation of the relation, is the embedding representation of the edge set;

[0032] And the operation process of the generated subunit is expressed by the following formula:

[0033] ;

[0034] in, It is the relevant information inferred from the knowledge graph, including entities, relationships, and attributes; query is the user's input question or request; the retrieval result set It is the context document selected by the retrieval subunit, which is retrieved from the document collection by the retrieval subunit. Selected from the query The most relevant collection of documents; is the generated natural language response, It consists of words; is the first The word is the model in The output generated by time steps; Indicates generation to time step All words before , as the next word conditions; In a given query , Search result set and knowledge graph reasoning information And the current generated Words In the case of Words probability;

[0035] The process of joint optimization is expressed as follows:

[0036] ;

[0037] ;

[0038] ;

[0039] in, is the probability that the document is selected; is the conditional probability of generating a subunit; Indicates that it is used for query With Documentation Similarity score of Represents the most relevant documents after retrieval; is the total length of the generated response, Is generated The conditional probability of a word.

[0040] Preferably, the knowledge extraction subunit includes:

[0041] A medical entity recognition sub-subunit, for identifying medically related entities from patient-related data;

[0042] The attribute extraction sub-unit is used to identify the attribute information of the entity;

[0043] The relation extraction sub-unit is used to identify the relations between entities in the text.

[0044] Preferably, the knowledge fusion subunit includes:

[0045] The entity alignment sub-unit is used to calculate the similarity of the same entity in different data sources or knowledge bases;

[0046] The entity linking sub-unit is used to correctly link entities in the text with entities in the knowledge base and perform entity reference disambiguation;

[0047] The relation inference sub-unit is used to expand known entity relations from the text into the knowledge base.

[0048] Preferably, the derivation unit comprises:

[0049] The construction unit is used to process the input entity and relation embedding through multiple hidden layers and perform reasoning to construct a neural network structure. and relationship The model output is the embedding vector of the tail entity This process is expressed by the following formula:

[0050] ;

[0051] in, Is Entity and relationship Embedded splicing; is the weight matrix of the neural network; is the bias term; is a nonlinear activation function; the output is the predicted embedding vector of the tail entity;

[0052] The entity relationship generation unit is used to learn complex patterns between entities and relationships through multiple hidden layers to form new entity relationships and infer unobserved triples. This process is expressed by the following formula:

[0053] ;

[0054] in, is the inference probability, outputting a probability value, indicating the entity pair Existence relationship credibility.

[0055] Preferably, the personalized medical assistance module includes a multimodal feature extraction unit and a personalized auxiliary treatment plan recommendation unit.

[0056] Preferably, the multimodal feature extraction unit comprises:

[0057] A user modeling subunit, used to establish a personalized feature vector based on the multimodal data of the patient;

[0058] A feature extraction subunit, which uses a long short-term memory network model to extract meaningful information from clinical assessment reports and patient feedback, and uses a recurrent neural network to process time-varying data;

[0059] The personalized auxiliary treatment plan recommendation unit includes:

[0060] The program summary subunit is used to summarize programs from aspects such as drug therapy, cognitive behavioral therapy, emotional support, and brain-computer interface therapy, and provide timely suggestions and resource links;

[0061] The weighted subunit is used to calculate the weighted score of the solution, which is expressed by the following formula:

[0062] ;

[0063] in, It is a treatment plan Total rating, is the number of features, It is The weight of a feature indicates the contribution of the feature to the score. It is characteristic function, reflecting the characteristics Treatment options The impact of

[0064] The recommendation subunit is used to recommend the best adjuvant therapy plan. This process is expressed by the following formula:

[0065] ;

[0066] in, represents the set of all preferred adjuvant treatment options, It is the best adjuvant treatment option finally recommended to patients. Represents a collection of preferred adjuvant treatment options A specific adjuvant therapy regimen is an index variable, which can range from 1 to , through different Values ​​are used to refer to different real-world treatment options in the collection.

[0067] Compared with the prior art, the knowledge graph-based personalized depression treatment decision system provided by this application has achieved the following beneficial effects:

[0068] (1) Improving the accuracy of diagnosis and treatment: By collecting multimodal data from patients, such as biomarker data, clinical evaluation data, and patient feedback data, and combining it with the medical knowledge and experience in the knowledge graph, the system can provide patients with more accurate and comprehensive diagnosis and treatment recommendations, reducing the risk of misdiagnosis and missed diagnosis.

[0069] (2) Personalized treatment: The system can recommend personalized treatment plans based on the patient's specific conditions, such as genotype, medical history, symptoms, etc., which can better meet the patient's needs and improve treatment outcomes and patient satisfaction.

[0070] (3) Optimizing the allocation of medical resources: Through the automation and intelligent diagnosis and treatment decision-making process, the system can reduce the workload of doctors and improve the utilization efficiency of medical resources. At the same time, the system can also provide remote diagnosis and treatment services for patients in remote areas, alleviating the problem of uneven distribution of medical resources.

[0071] (4) Promote medical knowledge sharing: The construction and optimization process of the knowledge graph promotes the integration and sharing of medical knowledge. Doctors can access the latest medical research results and clinical experience through the system to improve their diagnosis and treatment level. At the same time, the system can also provide strong support for medical education and training.

[0072] (5) Improve patient engagement: The system enhances patient engagement and trust by providing personalized treatment plans and timely feedback mechanisms. Patients can better understand their condition and treatment progress, work with doctors to develop treatment plans, and improve treatment compliance and effectiveness.

[0073] In summary, the knowledge graph-based personalized depression treatment decision system provided by this application has achieved remarkable beneficial effects in improving diagnosis and treatment accuracy, realizing personalized treatment, optimizing medical resource allocation, promoting medical knowledge sharing, and enhancing patient participation. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0075] Figure 1 It is a schematic diagram of system modules of a knowledge graph-based personalized depression treatment decision system provided in an embodiment of the present application;

[0076] Figure 2 It is a knowledge graph construction flowchart provided in an embodiment of the present application. DETAILED DESCRIPTION

[0077] The embodiments of the present application provide a method and device for optimizing the detection and extraction process of tea polyphenols based on near-infrared spectroscopy. The terms "first", "second", "third", "fourth", etc. (if any) in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchanged where appropriate, so that the embodiments described here can be implemented in an order other than that illustrated or described here. In addition, the terms "including" or "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units that are clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0078] As mentioned above, the current medical field faces many challenges in the diagnosis and treatment of patients. For example, depression is a serious mental disorder. Patients often show persistent low mood, loss of interest, cognitive dysfunction and even self-harm tendencies, which bring tremendous pressure to society, family and medical systems, and have a great impact on patients' quality of life and health. The diagnosis and treatment of depression currently faces many challenges. Traditional diagnostic methods rely on doctors' experience and patients' subjective descriptions, and their accuracy is limited; medical resources are unevenly distributed, and patients find it difficult to obtain timely and effective treatment; and existing treatment plans lack personalization and are difficult to meet the needs of different patients; combined with questionnaire scales and patients' subjective feedback, they have defects such as strong subjectivity, low compliance and small coverage.

[0079] In view of the above problems, the following Figure 1~Figure 2 To explain this application in detail.

[0080] The present application provides a personalized medical decision-making support system based on knowledge graphs. The system is mainly aimed at various current diseases. Through the entire multi-source data, a knowledge graph is constructed for data processing, and finally a personalized treatment plan is obtained to assist doctors in treating patients. Specifically, the present application provides a personalized treatment decision-making system for depression based on knowledge graphs, including:

[0081] The collection module is used to collect and process the patient's relevant data;

[0082] A building module for integrating and systematizing medical knowledge based on a preset knowledge graph embedding model to construct a knowledge graph;

[0083] The optimization module is used to optimize the knowledge graph to improve the performance of the knowledge graph and thus derive medical auxiliary decision-making;

[0084] The personalized medical assistance module is used to generate personalized auxiliary treatment plan recommendations based on a preset personalized treatment decision model and the derived medical assistance decision, so as to provide auxiliary decision-making for patients.

[0085] The knowledge graph-based personalized depression treatment decision system provided in this application collects patient-related data by designing a collection module, then constructs a knowledge graph through a construction module, optimizes the knowledge graph through an optimization module, and finally outputs a personalized auxiliary treatment plan through a personalized output module to provide auxiliary decision-making for the patient's treatment. The advantage of this application is that the optimal auxiliary treatment plan can be selected in combination with the knowledge graph to meet the patient's personalized needs.

[0086] In order to further elaborate and explain the above embodiments, based on the above embodiments, in another embodiment of the present application, the patient's relevant data at least includes: biological marker data (such as genomic data, blood samples, brain imaging data, body surface information, facial expressions, voice, gait, tongue, pulse and exhaled gas, etc.), clinical evaluation data (such as using standardized scales to evaluate symptom severity, disease course characteristics, diagnosis and treatment questionnaires, pathological information, disease severity, diagnostic criteria, treatment plans, etc.), patient feedback data (such as collecting emotional diaries, sleep quality, daily activities and other information through smartphone applications or wearable devices), and treatment response data (such as recording the patient's response to different treatment methods).

[0087] Then, the data are preprocessed as follows:

[0088] The collected data are subjected to preprocessing operations such as data standardization (converting the data into a distribution with zero mean and unit variance through Z-score standardization), missing value processing (for numerical data, missing values ​​can be filled by the mean of the data), and information gain calculation to ensure the quality and availability of the data and provide a reliable basis for subsequent analysis and decision-making.

[0089] For classification problems (such as determining the type or severity of depression based on multi-source data), information gain can be used to evaluate the importance of a feature for classification. Assuming that the dataset have categories, features have Different values Representation characteristics The value is A subset of samples, Representation dataset The total number of samples, Representation Subset The number of samples is, then the feature Information gain The calculation formula is:

[0090] ;

[0091] in, Is a collection of documents Information excerpt, Yes Category In the document collection The probability of occurrence in the information gain is larger, indicating that the feature The greater the help in classification.

[0092] In another embodiment of the present application, the construction module further includes: a framework unit, an embedding unit, and a loss function unit, wherein the framework unit is used to build a framework of the knowledge graph, as shown below:

[0093] Knowledge graph reasoning information include:

[0094] Entity Collection :Such as patient symptoms (depression, anxiety), physical characteristics (facial expressions, pulse), treatment plans (drugs, psychotherapy), relationship sets : such as "initiate", "improve", etc., edge set : Describe the relationship between entities. Each relationship can be expressed as a triple. Represents an entity and The relationship between , for example: (depression, triggers, insomnia).

[0095] The embedding unit is used to build a knowledge graph embedding model as follows:

[0096] The knowledge graph embedding model is operated through two-dimensional deep convolution Capture the complex characteristics of relationships, map entities and relationships into low-dimensional vector space, and use vectors to represent entity sets , Relationship Set and edge set ,in for -dimensional vector, the goal is to make the score of the correct triple Minimum:

[0097] ;

[0098] ;

[0099] in, Representing a collection of entities and relationship set The splicing result is expressed as , represents a two-dimensional convolution operation, where is the convolution kernel, is the number of convolution kernels. The convolution operation captures the concatenated vector through a sliding window. The local features in DWConv() represent the deep two-dimensional convolution operation, which uses multiple convolution kernels W to extract the complex characteristics between relations and entities, and takes the sum of the output results of each convolution kernel. represents a nonlinear activation function, Represents the flattening operation, which converts the two-dimensional feature matrix generated by the convolution operation into a one-dimensional feature vector to facilitate subsequent matrix operations. Indicates that the rationality score of the triple is calculated by dot product , by combining the edge set The dot product of is calculated, which is measured by and The feature of represents whether the generated vector can well predict the edge set .

[0100] The loss function unit is used to set the loss function of the knowledge graph embedding model, as shown below:

[0101] Assume Entity and and the relationship between them Embedded into the vector space, the goal of the knowledge graph embedding model is to minimize the following loss function:

[0102] ;

[0103] in: is a set of training triples (including entity pairs and relations), such as , representing the entity and The relationship between , is the relation score function obtained through the knowledge graph embedding model, representing the triple The score, is a preset hyperparameter used to distinguish positive samples from negative samples. It is a negative sample entity pair used to train the model to distinguish correct relation triplets from incorrect negative samples. The loss function optimizes the model embedding by comparing the scores of positive and negative samples, making similar entities and relations closer in the vector space.

[0104] In another embodiment of the present application, the optimization module further includes:

[0105] The response unit includes a retrieval subunit and a generation subunit. The response unit is used to extract and organize medical knowledge from the patient's relevant data, and jointly optimize the retrieval subunit and the generation subunit so that the dialogue output of the knowledge graph can generate a response that is more contextual and more accurate.

[0106] The operation of the retrieval subunit is as follows: the retrieval subunit combines the structural information in the knowledge graph to enhance the understanding of the query, and its purpose is to and document collections Find the most relevant set of documents to the query , by combining the query and Documentation Similarity, and structured information in the knowledge graph , sort the documents and select the most relevant Documents:

[0107] ;

[0108] ;

[0109] ;

[0110] in, It is the knowledge graph reasoning information, which contains structured entity and relationship information. is a document collection, the target collection for retrieval, containing all documents to be matched. It represents the retrieval result set, which is a document collection Inquiry Most relevant A collection of documents. It means to select the first documents, is the similarity term, used for query Documents to be searched The similarity score of the query and documents to be searched The direct semantic similarity of and is the vector representation of the query and the document to be searched. and Representation vector and of norm, which is the length or magnitude of a vector, Is a query in the knowledge graph and documents to be searched Entities in The relationship score of Representation vector of norm, which is the length of the vector, is an entity in the knowledge graph that may appear in the query or Document In the example, h is the entity The embedding representation of is the embedding representation of the relation, : An embedded representation of a set of edges that may appear in a document.

[0111] The operation mode of the generating subunit is as follows:

[0112] The generation subunit combines the context of the retrieval with the reasoning information in the knowledge graph. The goal of the generation subunit is to , the retrieved document collection and knowledge graph reasoning information , generating the response , the generation probability is decomposed into conditional probabilities at each time step , that is, at each time step , based on existing generated content and input context to generate the next word :

[0113] ;

[0114] It is the relevant information inferred from the knowledge graph, including entities, relationships, attributes, etc. It is the user's input question or request, the main basis for generating a response, and the retrieval result set It is the context document selected by the retrieval subunit, which is retrieved from the document collection by the retrieval subunit. Selected from the query The most relevant collection of documents, is the generated natural language response, The word composition, is the first The word is the model in The output generated by time steps is It means generating to time step All words before , as the next word conditions, In a given query , Search result set and knowledge graph reasoning information And the current generated Words In the case of Words probability.

[0115] Joint optimization is performed based on the retrieval subunit and the generation subunit.

[0116] The final joint optimization goal can combine the outputs of the knowledge graph reasoning and retrieval subunits to generate a response Probability The output document probability of the retrieval subunit is jointly optimized by the two parts and retrieve documents based on and knowledge graph reasoning The probability of generating a subunit , based on the retrieved documents and inference information, generate responses relevant to the query , which is mainly expressed by the following formula:

[0117] ;

[0118] ;

[0119] ;

[0120] Combining the reasoning and retrieval process, the dialogue output can generate responses that are more contextual and accurate, and the probability of the document being selected is increased. The document is to be checked With query The relevance score indicates that The probability of being selected in the search results, the conditional probability of generating subunits In a given query , To be checked and knowledge graph reasoning information Under the condition of The probability of Indicates that it is used for query With Documentation The similarity score of Represents the most relevant document after retrieval. is the total length of the generated response, Is generated The conditional probability of words, considering the historical generated content , Query , To be checked and inferential information.

[0121] The optimization module also includes: a precision unit, which is used to optimize and process the medical knowledge in the knowledge graph to make it more precise. The precision unit includes a knowledge extraction subunit and a knowledge fusion subunit; the knowledge extraction subunit includes a medical entity recognition subunit, which is expressed by the following process:

[0122] The medical entity recognition sub-unit is used to identify medical-related entities such as diseases, drugs, symptoms, etc. from multimodal data. This task is completed using a bidirectional long short-term memory network-conditional random field model.

[0123] Assume the input text , where each The text is words, and calculate the label of each word through the deep learning model , the goal of entity recognition is to predict the label of each word :

[0124] ;

[0125] It is a conditional probability based on a bidirectional long short-term memory network, used to predict the label of each word (such as "disease", "drug", "symptom", etc.),

[0126] For the bidirectional LSTM-CRF model, the conditional probability of the label sequence is It can be calculated by the conditional random field formula:

[0127] ;

[0128] From the tag arrive The transfer score,

[0129] is a given input and tags The launch score,

[0130] is the normalization factor.

[0131] The attribute extraction subunit is expressed through the following process: The goal of the attribute extraction task is to identify the attribute information of an entity (such as the severity of a disease, the dosage of a drug, etc.). and text input , predict the attributes of the entity The conditional probability of choosing The biggest attribute As the final prediction result, the attributes of each entity is extracted from the text, assuming that the entity The attribute set is The goal of attribute extraction is to predict attributes for each entity and select the attribute with the highest probability. As an entity The final properties:

[0132] ;

[0133] is a characteristic function that describes the entity and properties With input text For example, if It's aspirin. Is the dose ", then the characteristic function Is there a number '50' near 'aspirin').

[0134] : Calculate candidate attributes The score of Enter and text Mapped to a specific score, Indicates the number of eigenfunctions used in the model.

[0135] : Normalize all possible attributes The score makes the probability value range , candidate attribute set { }: Indicates an entity The set of all possible attribute candidates is used to normalize the denominator.

[0136] Conditional Probability : Indicates attributes Is Entity The probability of the attribute, given the text context ,if =0.95, indicating yes The confidence level of the attribute is 95%.

[0137] The relation extraction sub-unit is expressed through the following process: the goal of the relation extraction task is to identify the relationship between entities in the text. The distant supervision method is used to automatically annotate the training data through the existing knowledge base labels to reduce the dependence on manual annotation. The goal of distant supervision is to generate annotations by matching entity pairs in the text with the relationship patterns in the knowledge base.

[0138] Assumptions are two entities in the text, is the relationship between them. The knowledge base provides the true relationship between entity pairs. , the goal of relation extraction is through the feature function Combine entity pairs ,relation and input text Calculate the probability of the relationship between entity pairs and choose the probability The largest As the prediction result:

[0139] ;

[0140] Among them, the entity Represents text There are two entities of interest, and the goal is to predict the relationship between them , Represents the characteristic function, describing the entity pair and relationship With input text If the relationship between (Einstein, Germany), place of birth, "Einstein was born in Germany", then the characteristic function Yes: Whether "born" or an entity exists in the text Is it a geographic location? It is a computational relationship The feature score of The weighted sum of represents the correlation between entity pairs and relations. represents the total number of eigenfunctions used in the model, :For the candidate relationship set The scores are normalized to obtain the conditional probability distribution , the candidate relationship set : Entity pair The set of all possible relationships, including the target relationship and other candidate relationships , Example: For (Einstein, Germany), the candidate relationship set is Place of birth, place of residence, place of death , conditional probability : Entity pair In the context of the text There is a relationship probability.

[0141] The knowledge fusion subunit includes the entity alignment subunit, which is expressed through the following process: The goal of entity alignment is to calculate the similarity of the same entity in different data sources or knowledge bases. Assume that the entity and From different knowledge bases and , they may refer to the same concept, the similarity of entity alignment Calculated by feature matching:

[0142] ;

[0143] Is Entity and The vector representation of and Representation vector and of The norm, which is the length or magnitude of a vector.

[0144] The entity linking subunit is expressed through the following process: The task of entity linking is to correctly link the entities in the text with the entities in the knowledge base, and to perform entity reference disambiguation. and text context , calculate candidate entities Probability in the knowledge base , that is, the entity in the text Which knowledge base entity is most likely to correspond to :

[0145] ;

[0146] ;

[0147] is based on the similarity between context and entity embeddings, and Representation vector and of norm, which is the length or magnitude of a vector, is a contextual feature, which represents the context of the entity in the text. Is a computational entity Entities in the text The matching degree of For all candidate entities Normalize so that the sum of the probability distribution is 1; context features Represents the context feature function, describing the entity In the context of the text The relevance in the text represents the contextual information of the entity in the text. For example, the entity "apple" refers to "Apple Company" in some contexts and "fruit" in other contexts. Help the model determine the specific reference of the entity and the candidate entity set Represents the set of all candidate entities, including the target entity and other possible entities , conditional probability Represents a given text entity and text context ,entity is the probability of the best matching candidate entity.

[0148] The relation inference subunit is expressed through the following process: The goal of relation inference is to expand known entity relations from the text to the knowledge base. Given two entities in the relation inference and known relationships , calculate the relationship Conditional Probability in a Knowledge Base , the relationship Infer entity pairs from text into knowledge base:

[0149] ;

[0150] in, It is a characteristic function used to extend the relationship in the reasoning process, which represents the relationship between entity pairs. Is to calculate candidate relations The score of Indicates the relationship between entity pairs, feature number represents the number of eigenfunctions used in the model, For all candidate relations The scores are normalized to obtain the conditional probability distribution, characteristic function is the characteristic function Used to describe entity pairs and relationship and provide support for reasoning, such as the characteristic function Is the entity mentioned in the text? and entities , and there are relational terms like "born in" between them, the candidate relation set It is an entity pair The set of all possible relationships, including the target relationship and other candidate relationships For example, for the entity pair (Einstein, Germany), the candidate relationship set is Place of birth, place of residence, place of death}, conditional probability Represents a given entity pair , and infer that there is a relationship between them The probability that , then it represents the relationship between the entity pair (Einstein, Germany) The probability of birthplace is .

[0151] Then there is a derivation unit, which is used to automatically learn complex relationships in the knowledge graph and perform reasoning to derive new knowledge. In another embodiment of the present application, the derivation unit includes a construction unit and an entity relationship generation unit, which are respectively expressed by the following processes:

[0152] Construction unit: The deep neural network processes the input entity and relation embedding through multiple hidden layers and performs reasoning, assuming that the input is an entity and relationship The model output is the embedding vector of the tail entity :

[0153] ;

[0154] in, Representing Entities and relationship Embedded splicing, represents the weight matrix of the neural network, represents the bias term, Represents the activation function (such as ReLU, Sigmoid), output Represents the predicted embedding vector of the tail entity.

[0155] Entity Relationship Generation Unit: The neural network learns complex patterns between entities and relationships through multiple hidden layers. During the reasoning process, the goal of the neural network is to generate new entity relationships and infer unobserved triples. Given an entity and relationship , the network outputs the prediction of the tail entity (i.e., predictive relationship Reasoning entities between):

[0156] ;

[0157] Inference Probability : Output a probability value, indicating the entity pair Existence relationship credibility.

[0158] In another embodiment of the present application, the personalized medical assistance module includes a multimodal feature extraction unit and a personalized auxiliary treatment plan recommendation unit, and the multimodal feature extraction unit includes a user modeling subunit and a feature extraction subunit, which is expressed by the following process:

[0159] User modeling: based on multimodal patient data , build a personalized feature vector:

[0160] ;

[0161] .

[0162] .

[0163] .

[0164] Feature extraction: Use convolutional neural networks to extract features from brain imaging data; use long short-term memory network models to extract meaningful information from clinical assessment reports and patient feedback; use recurrent neural networks to process data that changes over time (such as EEG signals, pulse wave signals, etc.).

[0165] The optimizer updates the parameters The formula (including parameters such as network weights and biases) is as follows, which avoids the possible inconsistency between weight decay and learning rate adjustment:

[0166] ;

[0167] in, is the weight decay rate, which controls the regularization strength (i.e., the coefficient of L2 regularization), is the learning rate, is a small constant (usually ), used to avoid division by zero errors, is the bias-corrected first-order moment estimate, is the bias-corrected second-order moment estimate, is the current number of training steps,

[0168] in:

[0169] ;

[0170] ;

[0171] is the first moment estimate (the average of the gradients), Second-order moment estimate (the average of the squared gradients), and are the decay rates of the first and second order moments, respectively, and are usually set close to 1 (e.g. ).

[0172] in:

[0173] ;

[0174] ;

[0175] is the loss function over parameters The gradient of

[0176] The personalized adjuvant therapy plan recommendation unit includes a plan summary sub-unit, a weighted sub-unit, and a recommendation sub-unit, which are expressed through the following processes:

[0177] Summary of the plan: The treatment plan can be considered from the following four aspects:

[0178] 1) Drug treatment: Based on the patient's genotype and past treatment response, recommend antidepressant drugs and dosages, recommend one or more antidepressants, and give preliminary dosage recommendations.

[0179] 2) Cognitive behavioral therapy: Based on the patient's symptom characteristics, a series of customized cognitive behavioral training tasks are recommended, such as emotion management exercises in virtual reality scenarios.

[0180] 3) Emotional support: Provide emotional support through AI chatbots, and provide timely advice and resource links based on the patient's mood swings.

[0181] 4) Brain-computer interface therapy: Use transcranial magnetic stimulation as an auxiliary treatment method, and use feedback training to help patients improve their self-control ability.

[0182] Weighted subunit: The score of the treatment plan is obtained by weighted combination of multiple factors (such as patient characteristics, nature of the disease, effect of the treatment plan, etc.). The corresponding score is , is calculated using the following weighted sum model:

[0183] ;

[0184] in, Indicates treatment options Total rating, represents the number of features, Indicates The weight of a feature indicates the contribution of the feature to the score. Indicates characteristic function, reflecting the characteristics Treatment options The recommended subunit represents the assumption that all candidate treatment options are , we can make personalized recommendations by selecting the treatment with the highest score:

[0185] ;

[0186] Among them, among them, represents the set of all preferred adjuvant treatment options, It is the best adjuvant treatment option finally recommended to patients. Represents a collection of preferred adjuvant treatment options A specific adjuvant treatment plan. is an index variable, which can range from 1 to , through different Values ​​are used to refer to different real-world treatment options in the collection.

[0187] In many (That is arrive Among these different adjuvant treatment options, find the one whose score The largest solution, and the one that finally finds the largest score that is (That is, the optimal auxiliary treatment plan recommended to the patient in the end). Argmax is a mathematical operator (or a form of representation) used to find the value of the independent variable that makes a function reach its maximum value. In simple terms, it is essentially a way of expressing the value of the independent variable when the value of the related function reaches the maximum within a specific set. It is often used in mathematical operations and logical processing related to decision-making and optimization.

[0188] The application scenarios of the above knowledge graphs may include: clinical decision-making, medical question and answer, medical intelligent semantic search, disease management, etc.

[0189] Based on the same general inventive concept, the present application also protects a control device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the personalized treatment decision system described above.

[0190] Based on the same general inventive concept, the present application also protects a machine-readable storage medium, on which instructions are stored, and the instructions enable the machine to execute the above-mentioned personalized treatment decision system.

[0191] It should be understood that in the various embodiments of the present application, the size of the serial numbers of the above-mentioned processes does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present application.

[0192] In addition, the terms "system" and "network" are often used interchangeably in this article. The term "and / or" in this article is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the previously associated objects are in an "or" relationship.

[0193] It should be understood that in the embodiments of the present application, "B corresponding to A" means that B is associated with A, and B can be determined based on A, but it should also be understood that determining B based on A does not mean determining B only based on A, and B can also be determined based on A and / or other information.

[0194] A person of ordinary skill in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, computer software, or a combination of the two. In order to clearly illustrate the interchangeability of hardware and software, the composition and steps of each example have been generally described in terms of function in the above description. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel may use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0195] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0196] In the several embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. There may be other division methods in actual implementation. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. In addition, the mutual coupling or direct coupling or communication connection shown or discussed may be an indirect coupling or communication connection through some interfaces, devices or units, or may be electrical, mechanical or other forms of connection.

[0197] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they may be located in one place, or may be distributed on multiple network units, and some or all of the units may be selected according to actual needs to achieve the purpose of the embodiment of the present application.

[0198] In addition, the functional units in the various embodiments of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware or in the form of software functional units.

[0199] Through the description of the above implementation methods, those skilled in the art can clearly understand that the present application can be implemented by hardware, firmware, or a combination thereof. When implemented by software, the above functions can be stored in a computer-readable medium or transmitted as one or more instructions or codes on a computer-readable medium. Computer-readable media include computer storage media and communication media, wherein communication media include any media that facilitates the transmission of computer programs from one place to another. Storage media can be any available media that can be accessed by a computer. For example, but not limited to: computer-readable media can include RAM, ROM, EEPROM, CD-ROM or other optical disk storage, disk storage media or other magnetic storage devices, or can be used to carry or store desired data in the form of instructions or data structures. Any other medium in which program code can be stored and accessed by a computer. In addition, any connection can appropriately become a computer-readable medium. For example, if the software is transmitted from a website, server or other remote source using a coaxial cable, fiber optic cable, twisted pair, digital subscriber line (DSL), or wireless technologies such as infrared, radio and microwave, then the coaxial cable, fiber optic cable, twisted pair, DSL, or wireless technologies such as infrared, wireless and microwave are included in the fixation of the medium to which it belongs. As used in this application, disk and disc include compact disc (CD), laser disc, optical disc, digital versatile disc (DVD), floppy disk and Blu-ray disc, where disks usually reproduce data magnetically, while discs use lasers to optically reproduce data. The above combinations should also be included in the protection scope of computer-readable media.

[0200] In short, the above is only a preferred embodiment of the technical solution of the present application and is not intended to limit the scope of protection of the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the scope of protection of the present application.

Claims

1. A personalized depression treatment decision system based on knowledge graph, characterized in that: include: The collection module is used to collect and process the patient's relevant data; A building module for integrating and systematizing medical knowledge based on a preset knowledge graph embedding model to construct a knowledge graph; The optimization module is used to optimize the knowledge graph to improve the performance of the knowledge graph and thus derive medical auxiliary decision-making; The optimization module further comprises: A response unit, including a retrieval subunit and a generation subunit, wherein the response unit is used to extract and organize medical knowledge from relevant data of the patient, and jointly optimize the retrieval subunit and the generation subunit so that the dialogue output of the knowledge graph can generate a response that is more in line with the context and has higher accuracy; A precision unit, which is used to optimize and process the medical knowledge in the knowledge graph to make it more precise. The precision unit includes a knowledge extraction subunit and a knowledge fusion subunit. The inference unit is used to automatically learn complex relationships in the knowledge graph and perform reasoning to derive new knowledge; The operation process of the retrieval subunit is expressed by the following formula: ； ； ； in, It is the knowledge graph reasoning information, which contains structured entity and relationship information; is a document collection, which is the target collection for retrieval and contains all documents to be matched; It represents the retrieval result set, which is a document collection Inquiry Most relevant A collection of documents; It means to select the first Documents; is the similarity term, used for query Documents to be searched The similarity score of the query and documents to be searched The direct semantic similarity of and is the vector representation of the query and the document to be searched. and Representation vector and of the norm, which is the length or magnitude of a vector; Is a query in the knowledge graph and documents to be searched Entities in The relationship score of Representation vector of norm, which is the length of the vector; It is an entity in the knowledge graph; Is Entity Embedding representation of ; is the embedding representation of the relation, is the embedding representation of the edge set; And the operation process of the generated subunit is expressed by the following formula: ； in, It is the relevant information inferred from the knowledge graph, including entities, relationships, and attributes; query is the user's input question or request; the retrieval result set It is the context document selected by the retrieval subunit, which is retrieved from the document collection by the retrieval subunit. Selected from the query The most relevant collection of documents; is the generated natural language response, It consists of words; is the first The word is the model in The output generated by time steps; Represents generation to time step All words before , as the next word conditions; In a given query , Search result set and knowledge graph reasoning information and the currently generated Words In the case of Words probability; The process of joint optimization is expressed as follows: ； ； ； in, is the probability that the document is selected; is the conditional probability of generating a subunit; Indicates that it is used for query With Documentation Similarity score of ; Represents the most relevant documents after retrieval; Is generated The conditional probability of a word; Wherein, the derivation unit comprises: The construction unit is used to process the input entity and relation embedding through multiple hidden layers and perform reasoning to construct a neural network structure. and relationship The model output is the embedding vector of the tail entity This process is expressed by the following formula: ； in, Is Entity and relationship Embedded splicing; is the weight matrix of the neural network; is the bias term; is a nonlinear activation function; the output is the predicted embedding vector of the tail entity; The entity relationship generation unit is used to learn complex patterns between entities and relationships through multiple hidden layers to form new entity relationships and infer unobserved triples. This process is expressed by the following formula: ； in, is the inference probability, outputting a probability value, indicating the entity pair Existence relationship Credibility; The personalized medical assistance module is used to generate personalized auxiliary treatment plan recommendations based on a preset personalized treatment decision model and the derived medical assistance decision, so as to provide auxiliary decision-making for patients.

2. The personalized treatment decision system according to claim 1, characterized in that: The patient-related data includes at least: biological marker data, clinical evaluation data, patient feedback data, and treatment response data.

3. The personalized treatment decision system according to claim 1, characterized in that: The building blocks further include: The framework unit is used to build the knowledge graph framework, and the knowledge graph framework is expressed by the following formula: ； in, represents the knowledge graph framework, h represents the entity set, r represents the relationship set, and t' represents the edge set; The embedding unit is used to capture the complex characteristics of relations through two-dimensional deep convolution operations, map entities and relations to low-dimensional vector space to build a knowledge graph embedding model, and express the knowledge graph embedding model through the following formula: ； ； in, Representing a collection of entities and relationship set The splicing result is expressed as , represents a two-dimensional convolution operation, is the convolution kernel, is the number of convolution kernels. The convolution operation captures the concatenated vector through a sliding window. The local features in DWConv() represent the deep two-dimensional convolution operation, which uses multiple convolution kernels W to extract the complex characteristics between relations and entities, and takes the sum of the output results of each convolution kernel. represents a nonlinear activation function, Represents the flattening operation, which converts the two-dimensional feature matrix generated by the convolution operation into a one-dimensional feature vector to facilitate subsequent matrix operations. Indicates that the rationality score of the triple is calculated by dot product , by combining the edge set The dot product of is calculated, which is measured by and The feature of represents whether the generated vector can well predict the edge set ; The loss function unit is used to set the loss function of the knowledge graph embedding model. and and the relationship between them When embedded into the vector space, the loss function of the knowledge graph embedding model is expressed as follows: ； in, is a set of training triples, representing entities and The relationship between , is the relation score function obtained through the knowledge graph embedding model, representing the triple The score, is a preset hyperparameter used to distinguish positive samples from negative samples. are negative entity pairs used to train the model to distinguish correct relation triplets from incorrect negative samples.

4. The personalized treatment decision system according to claim 1, characterized in that: The knowledge extraction subunit comprises: A medical entity recognition sub-subunit, for identifying medically related entities from patient-related data; The attribute extraction sub-unit is used to identify the attribute information of the entity; The relation extraction sub-unit is used to identify the relations between entities in the text.

5. The personalized treatment decision system according to claim 1, characterized in that: The knowledge fusion subunit includes: The entity alignment sub-unit is used to calculate the similarity of the same entity in different data sources or knowledge bases; The entity linking sub-unit is used to correctly link entities in the text with entities in the knowledge base and perform entity reference disambiguation; The relation inference sub-unit is used to expand known entity relations from the text into the knowledge base.

6. The personalized treatment decision system according to claim 1, characterized in that: The personalized medical assistance module includes a multimodal feature extraction unit and a personalized auxiliary treatment plan recommendation unit.

7. The personalized treatment decision system according to claim 6, characterized in that: The multimodal feature extraction unit comprises: A user modeling subunit, used to establish a personalized feature vector based on the multimodal data of the patient; A feature extraction subunit, which uses a long short-term memory network model to extract meaningful information from clinical assessment reports and patient feedback, and uses a recurrent neural network to process time-varying data; The personalized auxiliary treatment plan recommendation unit includes: The program summary subunit is used to summarize programs from drug therapy, cognitive behavioral therapy, emotional support, and brain-computer interface therapy, and provide timely suggestions and resource links; The weighted subunit is used to calculate the weighted score of the solution, which is expressed by the following formula: ； in, It is a treatment plan Total rating, is the number of features, It is The weight of a feature indicates the contribution of the feature to the score. It is characteristic function, reflecting the characteristics Treatment options The impact of The recommendation subunit is used to recommend the best adjuvant therapy plan. This process is expressed by the following formula: ； in, represents the set of all preferred adjuvant treatment options, It is the best adjuvant treatment option finally recommended to patients. Represents a collection of preferred adjuvant treatment options A specific adjuvant therapy regimen is an index variable, ranging from 1 to , through different Values ​​are used to refer to different specific treatment options in the collection.

Images