WO2021189971A1

WO2021189971A1 - Medical plan recommendation system and method based on knowledge graph representation learning

Info

Publication number: WO2021189971A1
Application number: PCT/CN2020/136060
Authority: WO
Inventors: 颜泽龙; 王健宗; 吴天博; 程宁
Original assignee: 平安科技（深圳）有限公司
Priority date: 2020-10-26
Filing date: 2020-12-14
Publication date: 2021-09-30
Also published as: CN112242187A; CN112242187B

Abstract

Disclosed by the present application are a medical plan recommendation system and method based on knowledge graph representation learning, relating to the technical field of artificial intelligence, and capable of solving the problem that medical information recommended by existing medical recommendation systems is insufficiently accurate and such systems are prone to problems with potential risks. The system comprises: an extraction module, used for obtaining patient data of a target user and extracting a target entity in the patient data; a dividing module, used for dividing the medical knowledge graph into knowledge graph sub-graphs according to the target entity; a first determining module, used for determining, on the basis of representation learning, a low-dimensional vector corresponding to the knowledge graph sub-graph; an obtaining module, used for inputting the low-dimensional vector into a recommendation model which meets a preset training standard, and obtaining a medical recommendation result matching the patient data. The present application is suitable for the intelligent recommendation of medical solutions.

Description

Medical plan recommendation system and method based on knowledge graph representation learning

This application claims the priority of a Chinese patent application filed with the Chinese Patent Office on October 26, 2020, the application number is 202011153510.4, and the invention title is "Medical Solution Recommendation System and Method Based on Knowledge Graph Characterization Learning", the entire content of which is incorporated by reference Incorporated in this application.

Technical field

This application relates to the field of artificial intelligence technology, and in particular to a medical solution recommendation system and method based on knowledge graph representation learning.

Background technique

With the development of modern information technology, in order to build a smart city based on a new generation of information technology such as the Internet of Things, cloud computing, big data, and spatial geographic information integration, and to improve the digital experience of the people, various intelligent systems have emerged. In the medical field, by configuring a medical recommendation system, it can help patients and doctors greatly shorten the time for consultation and save manpower and material resources.

The inventor realizes that the current medical recommendation system often uses a fixed search method or simply uses the historical interactive information of doctors and patients as input to recommend relevant medical information. However, it does not pass the personal information well. Comprehensive analysis to better recommend medical information, resulting in the recommended medical plan is not accurate enough, and it is prone to potential risks.

technical problem

In view of this, this application provides a medical plan recommendation system and method based on knowledge graph representation learning, which mainly solves the problem that the medical information recommended by the existing medical recommendation system is not accurate enough and is prone to potential risks.

Technical solutions

According to one aspect of the present application, a medical solution recommendation system based on knowledge graph representation learning is provided, the system including:

The extraction module is used to obtain patient data of the target user and extract the target entity in the patient data;

The dividing module is used to divide the subgraph of the knowledge graph from the medical knowledge graph according to the target entity;

The first determining module is configured to determine the low-dimensional vector corresponding to the knowledge graph sub-graph based on characterization learning;

The obtaining module is used to input the low-dimensional vector into a recommendation model that meets the preset training standard, and obtain a medical recommendation result matching the patient data.

According to another aspect of the present application, a method for recommending medical solutions based on knowledge graph representation learning is provided, the method including:

Acquiring patient data of the target user, and extracting the target entity in the patient data;

Dividing the knowledge graph subgraph from the medical knowledge graph according to the target entity;

Determining the low-dimensional vector corresponding to the knowledge graph sub-graph based on representation learning;

The low-dimensional vector is input into a recommendation model that meets a preset training standard, and a medical recommendation result matching the patient data is obtained.

According to another aspect of the present application, there is provided a storage medium on which a computer program is stored, and when the program is executed by a processor, the above medical information recommendation method based on knowledge graph representation learning is implemented, including:

According to another aspect of the present application, a computer device is provided, including a storage medium, a processor, and a computer program that is stored on the storage medium and can run on the processor. When the processor executes the program, the above-mentioned The medical information recommendation method of knowledge graph representation learning includes:

Beneficial effect

With the above technical solutions, this application provides a medical solution recommendation system and method based on knowledge graph representation learning. Compared with the current medical recommendation system, this application can first use the extraction module to extract entities in the patient data, and use Divide modules and extract subgraphs from the knowledge graph based on entities. Then through the knowledge map representation learning, using the method of triple embedding, each entity (doctor, patient) and relationship (seeing a doctor, professional field, prescription, prescription, etc.) are embedded to obtain a low-dimensional vector, and the medical relationship map is maintained Semantic information. After that, the low-dimensional vector obtained by embedding is input into the recommendation model corresponding to the recommendation algorithm, and the recommendation model can classify the recommendation according to the low-dimensional vector of the patient, and further output the medical recommendation result for the patient's reference. In this application, obtaining low-dimensional vectors through characterization learning can improve the accuracy of the recommendation results output by the recommendation system, and provide higher support for subsequent personalized recommendations.

Description of the drawings

The drawings described here are used to provide a further understanding of the application and constitute a part of the application. The exemplary embodiments and descriptions of the application are used to explain the application, and do not constitute an improper limitation of the local application. In the attached picture:

Fig. 1 shows a schematic structural diagram of a medical scheme recommendation system based on knowledge graph representation learning provided by an embodiment of the present application;

FIG. 2 shows a schematic structural diagram of another medical scheme recommendation system based on knowledge graph representation learning provided by an embodiment of the present application;

FIG. 3 shows a schematic diagram of the principle structure of a knowledge graph representation learning provided by an embodiment of the present application;

FIG. 4 shows a schematic flowchart of a method for recommending medical solutions based on knowledge graph characterization learning provided by an embodiment of the present application.

The best mode of the present invention

Hereinafter, the present application will be described in detail with reference to the drawings and in conjunction with the embodiments. It should be noted that the embodiments in the application and the features in the embodiments can be combined with each other if there is no conflict.

Aiming at the problem that the medical information recommended by the existing medical recommendation system is not accurate enough and is prone to potential risks, an embodiment of the present application provides a medical solution recommendation system based on knowledge graph representation learning. As shown in FIG. 1, the system includes: An extraction module 31, a division module 32, a first determination module 33, and an acquisition module 34;

In a specific application scenario, the extraction module 31 can be used to obtain patient data of the target user and extract the target entity in the patient data. Among them, the patient data can be case consultation information manually uploaded by the target user in the recommendation system, or case information about the target user extracted based on the medical platform. Specifically, it can include various data forms such as text and images. After obtaining the patient information After that, firstly, based on the existing text conversion technology (such as OCR recognition technology, etc.), the patient data in each data form is uniformly converted into a text form for subsequent extraction of the target entity. The target entity refers to a word or phrase that has a descriptive meaning. It can usually be a person's name, place name, organization name, product name, or content with a certain meaning in a certain field, such as the name of a disease, drug, or organism in the medical field Wait. For this embodiment, it is necessary to first extract the target entity from the structured and unstructured information of the patient data, so as to further extract the subgraph of the knowledge map corresponding to the patient data by matching the target entity with the entity of the medical knowledge graph.

Correspondingly, since the dimensionality reduction expression method of the knowledge graph based on subgraph division is more reasonable than the traditional method for encoding entities, the dimensionality reduction expression based on the subgraph can fully consider the local characteristics of the knowledge graph, and the obtained entity encoding vector Can better reflect the essential characteristics of the entity. Therefore, in this application, it is necessary to divide the knowledge graph subgraph, and the larger the subgraph usually learns the better features, so the extraction range of the subgraph can be preset according to the required running time to ensure the medical plan The accuracy of the recommended results. Therefore, the system also includes a dividing module 32, which is used to divide the knowledge graph subgraph from the medical knowledge graph according to the target entity.

In a specific application scenario, the first determining module 33 may be used to determine the low-dimensional vector corresponding to the subgraph of the knowledge graph based on the representation learning. For this embodiment, when performing characterization learning, a new method can be used to encode triples on the basis of the traditional TRANS method, that is, by introducing position encoding and relational memory networks to mine the potential dependencies of triples, Further obtain the low-dimensional vector of the target entity. This process applies position coding and relational memory network coding to visit triples, which can solve the problem that the TRANS method cannot describe the potential dependency of the knowledge graph triples to a certain extent, thereby improving the accuracy of the triple embedding vector. Provide higher support for subsequent personalized recommendations.

Correspondingly, the obtaining module 34 can be used to input the low-dimensional vector determined by the first determining module 33 into a recommendation model that meets the preset training standard, and obtain a medical recommendation result matching the patient data. For this embodiment, in order to determine the medical recommendation scheme corresponding to the target user, specifically, the recommendation model can be pre-trained in the recommendation system based on preset classification rules, so that the recommendation model can be determined according to the entity low-dimensional vector corresponding to the patient Result of corresponding medical recommendation. Among them, the result of the medical recommendation may include the combination of medications, the treatment plan adopted, and the corresponding attending doctor candidates.

With the medical plan recommendation system based on knowledge graph representation learning in this embodiment, entities in patient data can be extracted first, and subgraphs can be extracted from the knowledge graph based on the entities. Then through the knowledge map representation learning, using the method of triple embedding, each entity (doctor, patient) and relationship (seeing a doctor, professional field, prescription, prescription, etc.) are embedded to obtain a low-dimensional vector, and the medical relationship map is maintained Semantic information. After that, the low-dimensional vector obtained by embedding is input into the recommendation model corresponding to the recommendation algorithm, and the recommendation model can classify the recommendation according to the low-dimensional vector of the patient, and further output the medical recommendation result for the patient's reference. In this application, obtaining low-dimensional vectors through representation learning can improve the accuracy of the recommendation results of the recommendation system and provide higher support for subsequent personalized recommendations.

Further, as a refinement and extension of the specific implementation of the foregoing embodiment, in order to fully illustrate the specific implementation process in this embodiment, as shown in FIG. 2, another medical solution recommendation system based on knowledge graph representation learning is provided. In the medical plan recommendation system, the extraction module 31 may further include: a first training unit 311 and an extraction unit 312.

In a specific application scenario, the first training unit 311 can be used to train an entity extraction model for extracting entity classes, where, when training an entity extraction model for extracting entity classes in patient data, the first training unit 311 can specifically Used to: tag the entity classes contained in the training set data; input the training set data after the annotation processing into the entity extraction model, and the training entity extraction model extracts the entity classes based on the Jieba natural language processing library; if the entity class is determined If the extraction error of is less than the preset threshold, it is determined that the entity extraction model has passed the training; if it is determined that the extraction error of the entity class is greater than or equal to the preset threshold, it is determined that the entity extraction model has not passed the training, and the training set data with pre-marked parts of speech is used to repeat the correction. The entity extraction model is trained so that the entity extraction model meets the first preset training standard.

For this embodiment, when performing part-of-speech tagging on the entity classes in the training set data, the part-of-speech tagging can be performed based on the ICTCLAS Chinese part-of-speech tagging set, so as to determine the part of speech of each entity class after word segmentation. In the specific training process, the data can be analyzed through the Jibba natural language processing library to classify all entity classes. Among them, the Jibba natural language processing database contains super large-scale corpus data, including 349,046 words, each line corresponds to a word, and contains three parts: word, word number, and part of speech. The preset threshold value should be a value from 0 to 1, indicating the maximum extraction error when the entity extraction model passes the training. The specific value can be set according to actual application requirements. The smaller the preset threshold value, it indicates the training of the entity extraction model. The higher the accuracy.

In a specific application scenario, the extraction unit 312 may be used to extract a target entity in the patient data using an entity extraction model that meets the first preset training standard. Among them, when extracting the target entity in the patient data according to the entity extraction model, the specific implementation process can be: loading the dictionary file, identifying each word segment in the patient data; constructing a directed acyclic graph based on each word segment; according to the directed acyclic The graph calculates the maximum path probability from each node to the ending position of the sentence, and determines the optimal ending position of the corresponding segment of the node when the probability is the largest; segmenting the patient data at the optimal ending position in order to obtain each target entity.

For example, if the input patient data is "some fever", when using the entity extraction model to extract the target entity, a directed acyclic graph DAG of the phrase will be constructed first. In the process of searching the dictionary for string matching, there may be several possible segmentation methods. These combinations can form a directed acyclic graph. For example, four paths can be formed: 1), ① one ③/some ④/fever; 2), ①One ③/some ⑤/fever; 3), ①have ②/some ④/fever; 4), ①have ②/some ⑤/fever, which can be determined according to the directed acyclic graph corresponding to the four paths Figure out the starting position and possible ending position of each word. Then, the probability of different ending positions corresponding to the same word is calculated, and the ending position with the highest probability is determined as the optimal ending position. Among them, the probability of each word = the number of words in the dictionary / The total number of words in the dictionary. If it is determined that the beginning position of the word segment in the text to be extracted is ①, two corresponding ending positions can be identified, namely: ② and ③, then the probability corresponding to the two ending positions can be calculated, and the probability corresponding to "Yes" If greater than "one", it can be determined that the position ② corresponds to the optimal ending position of position ①, and then the optimal ending position corresponding to other starting positions is determined based on the same method. For the beginning position ② of the word segment, the corresponding position can be identified There are two ending positions, namely: ④ and ⑤. If it is determined that the position ④ corresponds to the optimal ending position of the position ②, then the text to be extracted can be segmented at the optimal ending position ② and ④, and the target entity is obtained as " Yes, "some", "fever".

In a specific application scenario, in order to obtain the subgraphs of the knowledge graph, as shown in FIG. 2, in the medical plan recommendation system, the dividing module 32 may specifically include: a marking unit 321, a traversal unit 322, and a dividing unit 323; , The marking unit 321 can be used to mark the core object entity and the secondary object entity in the target entity; the traversal unit 322 can be used to traverse the medical knowledge graph with each core object entity as the starting point of the traversal, and traverse to the secondary object entity The traversal in this direction is stopped at time; the dividing unit 323 can be used to divide the knowledge graph sub-graphs according to the traversal results of each core object entity.

Among them, entity labeling is based on the importance and pivotality of this type of entity in the knowledge graph, marking it as a core object or a secondary object. Since the knowledge graphs in different fields have different entity types and association relationships, the task of labeling core objects and secondary objects can be completed manually. When traversing the subgraph, the breadth-first traversal rule can be used to traverse the subgraph according to the input core object entity. When the core object entity is traversed, the entity can be retained as the starting point of the subsequent traversal; when the secondary object entity is traversed Then stop traversing in that direction. The entity obtained in this step is actually the surrounding entity directly connected to the starting entity. Repeat the traversal steps until the entities obtained from a certain traversal, except for those already in the subgraph of the knowledge graph, the rest are all secondary object entities.

Correspondingly, in order to obtain the low-dimensional vector corresponding to the knowledge graph sub-graph, as shown in FIG. 2, in the medical plan recommendation system, the first determining module 33 may specifically include: an extraction unit 331, a configuration unit 332, and an encoding unit 333 2. The second training unit 334; the extraction unit 331, which can be used to extract each triplet in the knowledge graph sub-graph; the configuration unit 332, which can be used to position the entity vectors in the triples by encoding the triples Vector; encoding unit 333, which can be used to encode the triples after adding position vectors based on the relational network, to obtain the encoding vector; the second training unit 334 can be used to use the decoder to score the encoding vector, and use The adaptive moment estimation (Adam) optimizer performs iterative training to further obtain the low-dimensional vector corresponding to the subgraph of the knowledge map.

For this embodiment, the principle of knowledge graph representation learning can be seen in the medical triplet embedded coding structure shown in Figure 3. Specifically, when obtaining low-dimensional vectors through representation learning, the medical triplet can first be stored as ( The form of entity, relationship, entity), such as (patient, disease history, disease), (doctor, level, specialty) and other forms are used to construct triples. Afterwards, the position relationship can be embedded into the entity vector corresponding to the triplet, that is, by encoding the position of the triplet, the position vector can be configured for the entity vector in the embedding training. Then, the relational memory network can be used to encode the triples, and the specific encoding process can be realized based on the multi-head self-attention mechanism. In addition, in order to avoid that the entity vector obtained during the initialization process of the entity encoding is not accurate enough, it can be scored based on the decoder and iterated with the Adam optimizer. Through the positive and negative network training process, the entity can be further evaluated. The vector is optimized and adjusted so that the resulting low-dimensional vector meets the preset accuracy requirements. In order to finally get the low-dimensional vector of the medical entity, input it into the recommendation model to complete the framework of sequential learning.

In a specific application scenario, as shown in Figure 2, the medical plan recommendation system may specifically include: a labeling module 35, a training module 36, and a second determining module 37; the labeling module 35 can be used to determine sample patient data , And label the corresponding preset medical recommendation plan for the sample patient data; the training module 36 can be used to train the recommendation model using the low-dimensional vector corresponding to the sample patient data; the second determination module 37 can be used to determine the medical recommendation output by the recommendation model If the result meets the second preset training standard, it is determined that the recommended model has passed the training; the training module 36 can also be used to repeatedly train the recommended model with sample patient data if it is determined that the recommended model has not passed the training, so that the recommended model meets the second preset Training standards.

For this embodiment, the corresponding medical recommendation plan can be marked in advance based on different types of sample patient data. For example, the cancer patient data can be marked with an authoritative attending doctor in the field of cancer, and the corresponding treatment plan can also be marked. , Medication combination, etc. Furthermore, using the sample patient data labeled corresponding to the medical recommendation plan to carry out targeted training on the recommendation model can further strengthen the classification and recognition ability of the recommendation model, so that the output result of the recommendation model matches the labeling result.

Correspondingly, as shown in FIG. 2, in the medical plan recommendation system, the acquisition module 34 may specifically include: an input unit 341 and a determination unit 342.

In a specific application scenario, the input unit 341 may be used to input low-dimensional vectors into a recommendation model that meets the second preset training standard, and obtain recommendation scores corresponding to each preset medical recommendation plan. For this embodiment, after obtaining the low-dimensional vector corresponding to the target patient, the low-dimensional vector can be input into the recommendation model that meets the second preset training standard, and the recommendation model will output the recommendation score corresponding to each preset recommendation plan , The higher the recommended score, the higher the reference value.

Correspondingly, the determining unit 342 can be used to determine the preset medical recommendation plan with the highest recommended score as the medical recommendation result of the target user. For this embodiment, the preset medical recommendation scheme with the highest recommendation score can be determined as the medical recommendation result matching the target user, and then the recommendation system is output and displayed to the target user, so as to serve as a reference for the target user.

Through the above-mentioned medical plan recommendation system based on knowledge graph representation learning, entities in patient data can be extracted first, and subgraphs can be extracted from the knowledge graph based on the entities. Then through the knowledge map representation learning, using the method of triple embedding, each entity (doctor, patient) and relationship (seeing a doctor, professional field, prescription, prescription, etc.) are embedded to obtain a low-dimensional vector, and the medical relationship map is maintained Semantic information. After that, the low-dimensional vector obtained by embedding is input into the recommendation model corresponding to the recommendation algorithm, and the recommendation model can classify the recommendation according to the low-dimensional vector of the patient, and further output the medical recommendation result for the patient's reference. In this application, obtaining low-dimensional vectors through representation learning can improve the accuracy of the recommendation results of the recommendation system and provide higher support for subsequent personalized recommendations. In addition, for representation learning, on the basis of traditional methods, this application additionally introduces position coding and relational memory network to mine the potential dependency of the triples, and further obtain the low-dimensional vector of the target entity. This process applies position coding and relational memory network coding to visit triples, which can solve the problem that existing methods cannot describe the potential dependence of knowledge graph triples to a certain extent, and can improve the accuracy of triple embedding vectors. Provide higher support for subsequent personalized recommendations.

Further, as a specific embodiment of the method shown in FIG. 1 and FIG. 2, an embodiment of the present application provides a method for recommending a medical plan based on knowledge graph representation learning. As shown in FIG. 4, the method includes: acquiring patients of the target user Data, and extract the target entity in the patient data; divide the knowledge graph subgraph from the medical knowledge graph according to the target entity; determine the low-dimensional vector corresponding to the knowledge graph subgraph based on the representation learning; input the low-dimensional vector to meet the preset training standards In the recommendation model of, obtain the medical recommendation results that match the patient data.

In a specific application scenario, when extracting the target entity in the patient data, it may specifically include: training an entity extraction model for extracting the entity class; using an entity extraction model that meets the first preset training standard to extract the patient data Target entity. Among them, the specific method of training the entity extraction model for extracting entity classes can be: performing part-of-speech tagging on the entity classes contained in the training set data; inputting the labeled training set data into the entity extraction model to train the entity extraction The model extracts entity classes based on the Jieba natural language processing library; if it is determined that the extraction error of the entity class is less than the preset threshold, it is determined that the entity extraction model has passed the training; if it is determined that the extraction error of the entity class is greater than or equal to the preset threshold, the entity extraction model is determined If the training fails, the training entity extraction model is repeatedly modified using the training set data that is pre-marked with part of speech, so that the entity extraction model meets the first preset training standard.

In specific application scenarios, when dividing the subgraph of the knowledge graph from the medical knowledge graph according to the target entity, it can specifically include: marking the core object entity and the secondary object entity in the target entity; using each core object entity as the starting point for traversal The medical knowledge graph is traversed, and the traversal in this direction is stopped when the secondary object entity is traversed; the knowledge graph subgraph is divided according to the traversal results of each core object entity.

Correspondingly, when determining the low-dimensional vector corresponding to the knowledge graph sub-graph based on the representation learning, it may specifically include: extracting each triplet in the knowledge graph sub-graph; The entity vector configures the position vector; encodes the triples after adding the position vector based on the relational network to obtain the encoded vector; uses the decoder to evaluate the encoded vector, and uses the adaptive moment estimation (Adam) optimizer to perform Iterative training further obtains the low-dimensional vector corresponding to the subgraph of the knowledge graph.

In specific application scenarios, before inputting low-dimensional vectors into a recommendation model that meets the preset training standards and obtaining medical recommendation results matching the patient data, it also specifically includes: determining the sample patient data and setting it as the sample patient data Label the corresponding preset medical recommendation plan; use the low-dimensional vector corresponding to the sample patient data to train the recommendation model; if it is determined that the medical recommendation result output by the recommendation model meets the second preset training standard, the recommendation model is determined to pass the training; if the recommendation model is determined If the training fails, the recommended model is repeatedly trained using sample patient data so that the recommended model meets the second preset training standard.

Correspondingly, inputting the low-dimensional vector into the recommendation model that meets the preset training standard to obtain medical recommendation results matching the patient data may specifically include: inputting the low-dimensional vector into the recommendation model that meets the second preset training standard In, the recommended score corresponding to each preset medical recommendation plan is obtained; the preset medical recommendation plan with the highest recommended score is determined as the medical recommendation result of the target user.

It should be noted that, for other corresponding descriptions of the method for recommending medical solutions based on knowledge graph characterization learning provided in this embodiment, reference may be made to the corresponding descriptions in FIGS. 1 to 2, and details are not repeated here.

Based on the above-mentioned method shown in FIG. 4, correspondingly, an embodiment of the present application also provides a storage medium. The above-mentioned storage medium may be a volatile storage medium or a non-volatile storage medium; a computer program is stored thereon. When the program is executed by the processor, the above-mentioned method for recommending medical plans based on the knowledge graph representation learning as shown in FIG. 4 is realized.

Based on this understanding, the technical solution of this application can be embodied in the form of a software product. The software product can be stored in a non-volatile storage medium (which can be a CD-ROM, U disk, mobile hard disk, etc.), including several The instructions are used to make a computer device (which may be a personal computer, a server, or a network device, etc.) execute the methods in each implementation scenario of the present application.

Based on the above system shown in Figure 1 and Figure 2 and the method embodiment shown in Figure 4, in order to achieve the above objective, the embodiment of the present application also provides a computer device, which may be a personal computer, a server, or a network device. The physical device includes a storage medium and a processor; the storage medium is used to store a computer program; the processor is used to execute the computer program to implement the above-mentioned method for recommending medical solutions based on the knowledge graph representation learning as shown in FIG. 4.

Optionally, the computer device may also include a user interface, a network interface, a camera, a radio frequency (RF) circuit, a sensor, an audio circuit, a WI-FI module, and so on. The user interface may include a display screen (Display), an input unit such as a keyboard (Keyboard), etc., and the optional user interface may also include a USB interface, a card reader interface, and the like. The optional network interface can include standard wired interface, wireless interface (such as Bluetooth interface, WI-FI interface), etc.

Those skilled in the art can understand that the computer device structure provided in this embodiment does not constitute a limitation on the physical device, and may include more or fewer components, or combine certain components, or arrange different components.

The non-volatile readable storage medium may also include an operating system and a network communication module. The operating system is a program for the hardware and software resources of the data processing entity equipment based on the knowledge graph, and supports the operation of the information processing program and other software and/or programs. The network communication module is used to implement communication between various components in the non-volatile readable storage medium, and communication with other hardware and software in the physical device.

Through the description of the above embodiments, those skilled in the art can first extract entities in the patient data, and extract subgraphs from the knowledge graph based on the entities. Then through the knowledge map representation learning, using the method of triple embedding, each entity (doctor, patient) and relationship (seeing a doctor, professional field, prescription, prescription, etc.) are embedded to obtain a low-dimensional vector, and the medical relationship map is maintained Semantic information. After that, the low-dimensional vector obtained by embedding is input into the recommendation model corresponding to the recommendation algorithm, and the recommendation model can classify the recommendation according to the low-dimensional vector of the patient, and further output the medical recommendation result for the patient's reference. In this application, obtaining low-dimensional vectors through representation learning can improve the accuracy of the recommendation results of the recommendation system and provide higher support for subsequent personalized recommendations. In addition, for representation learning, on the basis of traditional methods, this application additionally introduces position coding and relational memory network to mine the potential dependency of the triples, and further obtain the low-dimensional vector of the target entity. This process applies position coding and relational memory network coding to visit triples, which can solve the problem that existing methods cannot describe the potential dependence of knowledge graph triples to a certain extent, and can improve the accuracy of triple embedding vectors. Provide higher support for subsequent personalized recommendations.

Those skilled in the art can understand that the accompanying drawings are only schematic diagrams of preferred implementation scenarios, and the modules or processes in the accompanying drawings are not necessarily necessary for implementing this application. Those skilled in the art can understand that the modules in the device in the implementation scenario can be distributed in the device in the implementation scenario according to the description of the implementation scenario, or can be changed to be located in one or more devices different from the implementation scenario. The modules of the above implementation scenarios can be combined into one module or further divided into multiple sub-modules.

The above serial number of this application is for description only, and does not represent the pros and cons of implementation scenarios. What has been disclosed above are only a few specific implementation scenarios of this application, but this application is not limited to these, and any changes that can be thought of by those skilled in the art should fall into the protection scope of this application.

Claims

A medical plan recommendation system based on knowledge graph representation learning, which includes:

The extraction module is used to obtain patient data of the target user and extract the target entity in the patient data;

The dividing module is used to divide the subgraph of the knowledge graph from the medical knowledge graph according to the target entity;

The first determining module is configured to determine the low-dimensional vector corresponding to the knowledge graph sub-graph based on characterization learning;

The obtaining module is used to input the low-dimensional vector into a recommendation model that meets the preset training standard, and obtain a medical recommendation result matching the patient data.
A medical scheme recommendation method based on knowledge graph representation learning, which includes:

Acquiring patient data of the target user, and extracting the target entity in the patient data;

Dividing the knowledge graph subgraph from the medical knowledge graph according to the target entity;

Determining the low-dimensional vector corresponding to the knowledge graph sub-graph based on representation learning;

The low-dimensional vector is input into a recommendation model that meets a preset training standard, and a medical recommendation result matching the patient data is obtained.
The method for recommending medical solutions based on knowledge graph representation learning according to claim 2, wherein said extracting the target entity in the patient data specifically includes:

Train the entity extraction model used to extract the entity classes;

The entity extraction model that meets the first preset training standard is used to extract the target entity in the patient data.
The medical solution recommendation method based on knowledge graph representation learning according to claim 3, wherein the specific method for training an entity extraction model for extracting entity classes is:

Perform part-of-speech tagging on the entity classes contained in the training set data;

Input the training set data after annotation processing into the entity extraction model, and the training entity extraction model extracts entity classes based on the Jieba natural language processing library;

If it is determined that the extraction error of the entity class is less than the preset threshold, it is determined that the entity extraction model has passed the training;

If it is determined that the extraction error of the entity class is greater than or equal to the preset threshold, it is determined that the entity extraction model has not passed the training, and the training set data with pre-marked parts of speech is used to repeatedly modify the training entity extraction model so that the entity extraction model meets the first prediction. Set training standards.
The medical solution recommendation method based on knowledge graph representation learning according to claim 2, wherein the dividing the knowledge graph subgraph from the medical knowledge graph according to the target entity specifically comprises:

Mark the core object entity and the secondary object entity in the target entity;

Use each core object entity as the starting point to traverse the medical knowledge graph, and stop the traversal in this direction when the secondary object entity is traversed;

According to the traversal results of each core object entity, the knowledge graph subgraph is divided.
The medical treatment plan recommendation method based on knowledge graph representation learning according to claim 2, wherein the determination of the low-dimensional vector corresponding to the knowledge graph sub-graph based on the representation learning specifically includes:

Extracting each triplet in the subgraph of the knowledge graph;

Configuring position vectors for the entity vectors in the triples by performing position encoding on the triples;

Performing encoding processing on the triplet after adding the position vector based on the relational network to obtain the encoding vector;

A decoder is used to evaluate the code vector, and an adaptive moment estimation optimizer is used to perform iterative training to obtain the low-dimensional vector corresponding to the knowledge map sub-graph.
The medical treatment plan recommendation method based on knowledge graph characterization learning according to claim 6, wherein said inputting said low-dimensional vector into a recommendation model meeting a preset training standard to obtain a medical treatment matching said patient data Before recommending the results, it also includes:

Determine the sample patient data, and mark the corresponding preset medical recommendation plan for the sample patient data;

Training a recommendation model by using low-dimensional vectors corresponding to the sample patient data;

If it is determined that the medical recommendation result output by the recommendation model meets the second preset training standard, determining that the recommendation model passes the training;

If it is determined that the recommendation model fails the training, the recommendation model is repeatedly trained using the sample patient data, so that the recommendation model meets the second preset training standard.
The method for recommending medical plans based on knowledge graph representation learning according to claim 7, wherein said inputting said low-dimensional vector into a recommendation model that meets a preset training standard to obtain a medical plan matching the patient data Recommended results, including:

Input the low-dimensional vector into a recommendation model that meets the second preset training standard, and obtain a recommendation score corresponding to each preset medical recommendation plan;

The preset medical recommendation scheme with the highest recommendation score is determined as the medical recommendation result of the target user.
A storage medium with a computer program stored thereon, wherein when the program is executed by a processor, a method for recommending medical information based on knowledge graph representation learning is realized: including:

Acquiring patient data of the target user, and extracting the target entity in the patient data;

Dividing the knowledge graph subgraph from the medical knowledge graph according to the target entity;

Determining the low-dimensional vector corresponding to the knowledge graph sub-graph based on representation learning;

The low-dimensional vector is input into a recommendation model that meets a preset training standard, and a medical recommendation result matching the patient data is obtained.
The storage medium according to claim 9, wherein the extracting the target entity in the patient data specifically includes:

Train the entity extraction model used to extract the entity classes;

The entity extraction model that meets the first preset training standard is used to extract the target entity in the patient data.
The storage medium according to claim 10, wherein the specific method for training an entity extraction model for extracting entity classes is:

Perform part-of-speech tagging on the entity classes contained in the training set data;

Input the training set data after annotation processing into the entity extraction model, and the training entity extraction model extracts entity classes based on the Jieba natural language processing library;

If it is determined that the extraction error of the entity class is less than the preset threshold, it is determined that the entity extraction model has passed the training;

If it is determined that the extraction error of the entity class is greater than or equal to the preset threshold, it is determined that the entity extraction model has not passed the training, and the training set data with pre-marked parts of speech is used to repeatedly modify the training entity extraction model so that the entity extraction model meets the first prediction. Set training standards.
The storage medium according to claim 9, wherein the dividing the knowledge graph subgraph from the medical knowledge graph according to the target entity specifically comprises:

Mark the core object entity and the secondary object entity in the target entity;

Use each core object entity as the starting point to traverse the medical knowledge graph, and stop the traversal in this direction when the secondary object entity is traversed;

According to the traversal results of each core object entity, the knowledge graph subgraph is divided.
The storage medium according to claim 9, wherein the determining the low-dimensional vector corresponding to the knowledge graph sub-graph based on the representation learning specifically comprises:

Extracting each triplet in the subgraph of the knowledge graph;

Configuring position vectors for the entity vectors in the triples by performing position encoding on the triples;

Performing encoding processing on the triplet after adding the position vector based on the relational network to obtain the encoding vector;

A decoder is used to evaluate the code vector, and an adaptive moment estimation optimizer is used to perform iterative training to obtain the low-dimensional vector corresponding to the knowledge map sub-graph.
The storage medium according to claim 13, wherein the inputting the low-dimensional vector into a recommendation model that meets a preset training standard, and before obtaining a medical recommendation result matching the patient data, further comprises:

Determine the sample patient data, and mark the corresponding preset medical recommendation plan for the sample patient data;

Training a recommendation model by using low-dimensional vectors corresponding to the sample patient data;

If it is determined that the medical recommendation result output by the recommendation model meets the second preset training standard, determining that the recommendation model passes the training;

If it is determined that the recommendation model fails the training, the recommendation model is repeatedly trained using the sample patient data, so that the recommendation model meets the second preset training standard.
The storage medium according to claim 14, wherein said inputting said low-dimensional vector into a recommendation model meeting a preset training standard to obtain a medical recommendation result matching said patient data specifically comprises:

Input the low-dimensional vector into a recommendation model that meets the second preset training standard, and obtain a recommendation score corresponding to each preset medical recommendation plan;

The preset medical recommendation scheme with the highest recommendation score is determined as the medical recommendation result of the target user.
A computer device, including a storage medium, a processor, and a computer program stored on the storage medium and running on the processor, wherein the processor implements a kind of medical information based on knowledge graph representation learning when the program is executed Recommended methods include:

Acquiring patient data of the target user, and extracting the target entity in the patient data;

Dividing the knowledge graph subgraph from the medical knowledge graph according to the target entity;

Determining the low-dimensional vector corresponding to the knowledge graph sub-graph based on representation learning;

The low-dimensional vector is input into a recommendation model that meets a preset training standard, and a medical recommendation result matching the patient data is obtained.
The computer device according to claim 16, wherein said extracting the target entity in the patient data specifically comprises:

Train the entity extraction model used to extract the entity classes;

The entity extraction model that meets the first preset training standard is used to extract the target entity in the patient data.
The computer device according to claim 17, wherein the specific method for training an entity extraction model for extracting entity classes is:

Perform part-of-speech tagging on the entity classes contained in the training set data;

Input the training set data after annotation processing into the entity extraction model, and the training entity extraction model extracts entity classes based on the Jieba natural language processing library;

If it is determined that the extraction error of the entity class is less than the preset threshold, it is determined that the entity extraction model has passed the training;

If it is determined that the extraction error of the entity class is greater than or equal to the preset threshold, it is determined that the entity extraction model has not passed the training, and the training set data with pre-marked parts of speech is used to repeatedly modify the training entity extraction model so that the entity extraction model meets the first prediction. Set training standards.
The computer device according to claim 16, wherein said dividing the subgraph of the knowledge graph from the medical knowledge graph according to the target entity specifically comprises:

Mark the core object entity and the secondary object entity in the target entity;

Use each core object entity as the starting point to traverse the medical knowledge graph, and stop the traversal in this direction when the secondary object entity is traversed;

According to the traversal results of each core object entity, the knowledge graph subgraph is divided.
The computer device according to claim 16, wherein the determining the low-dimensional vector corresponding to the subgraph of the knowledge graph based on the representation learning specifically comprises:

Extracting each triplet in the subgraph of the knowledge graph;

Configuring position vectors for the entity vectors in the triples by performing position encoding on the triples;

Performing encoding processing on the triplet after adding the position vector based on the relational network to obtain the encoding vector;

A decoder is used to evaluate the code vector, and an adaptive moment estimation optimizer is used to perform iterative training to obtain the low-dimensional vector corresponding to the knowledge map sub-graph.