Method for constructing relational database of clinical medicine

By performing multi-level classification and paragraph clauses on clinical evidence-based guide texts, combined with the joint extraction method of artificial intelligence and domain language models, a clinical medical relation database was constructed, solving the knowledge extraction of clinical evidence-based guide text files in the existing technology Inadequately, the automated joint extraction and fusion of pediatric disease knowledge graphs has been achieved, and the accuracy and efficiency of the knowledge graphs have been improved.

CN119066205APending Publication Date: 2024-12-03FUDAN UNIVERSITY +1

Patent Information

Application Number
CN202310643222.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-06-01
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The existing technology lacks effective knowledge extraction methods for clinical evidence-based guide text files in the construction of knowledge graphs in the medical field, and the lack of joint extraction mode, resulting in entity identification errors, redundancy and insufficient information utilization, and the inability to effectively process long texts and Multi-entity relationship across sentences.

Method used

After multi-level classification and paragraph classification of clinical evidence-based guide text files, an artificial intelligence model is used for joint extraction of relational recognition and entity extraction, and knowledge alignment and auditing are combined with domain language models to build a relational database of clinical medicine.

Benefits of technology

It realizes the automated joint extraction and integration of clinical medical knowledge, improves the accuracy and efficiency of the knowledge graph, solves the problems of difficulty in identifying English abbreviation and information redundancy, and establishes a unified pediatric disease relation database.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119066205A_ABST
    Figure CN119066205A_ABST
Patent Text Reader

Abstract

The invention discloses a method for constructing a relational database of clinical medicine, and the method comprises the following steps: S1, inputting a file to be subjected to knowledge extraction for constructing the relational database, S2, processing the file to be subjected to knowledge extraction, and converting the file to be subjected to knowledge extraction into a primary preprocessing text; s3, performing sentence segmentation on paragraphs of the primary pre-processed text, and then performing tiled title processing to obtain a secondary pre-processed text which comprises a plurality of long texts with tiled titles; s4, extracting the plurality of long texts in the step S3 through an artificial intelligence model so as to obtain a knowledge triple; and S5, constructing a database based on the knowledge triad and the secondary preprocessed text obtained in the previous step, thereby obtaining a clinical medicine relational database, and establishing a pediatric disease relational database for realizing automatic knowledge joint extraction and knowledge fusion through the method.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to the technical field of medical knowledge graphs, and in particular to a method for constructing a relational database (knowledge graph) for clinical medicine. Background Art

[0002] my country currently faces problems such as a shortage of high-quality pediatric resources, insufficient pediatric diagnosis and treatment capabilities among primary care doctors and difficulty in training them on a large scale and efficiently. Pediatric disease databases are mainly composed of some specialized disease databases, and pediatric clinical guidelines and consensus lack a unified high-standard evaluation system. A pediatric disease knowledge graph and construction system based on unstructured data, especially clinical evidence-based guideline documents, has not yet been found.

[0003] In addition, in the existing technology, the exploration of knowledge graphs in the medical industry is still in its initial development stage. The existing technical solutions are mainly aimed at knowledge extraction from structured data (electronic medical records) in medical systems, mainly based on logical rules. Most of them are based on deep learning methods, which extract entities and relationships in knowledge separately.

[0004] Specifically, the application of knowledge graphs in the prior art has the following shortcomings: (1) There is a lack of knowledge extraction methods for clinical evidence-based guideline text files. The main reason is that the formats of clinical evidence-based guideline documents are diverse, and the long professional text descriptions make it difficult to extract entity relationships with contextual connections, and there are a large number of English abbreviations; (2) The entity relationships in the knowledge graph are extracted separately, and there is a lack of joint extraction mode. The main reason is that entity extraction is mainly based on sequence models, and relationship extraction is mainly based on classification models, and the two cannot be unified and integrated. First, entity recognition is performed, and then the identified entities are paired for relationship classification, and finally the entity pairs with existing relationships are output as triples. However, this approach has the following disadvantages: (a) Error accumulation: errors in the entity recognition module will affect the performance of the following relationship classification; (b) Entity redundancy: unrelated entity pairs will bring redundant information, increase the error rate, and reduce the efficiency of the entire extraction process; (c) Insufficient information utilization: entity extraction and relationship classification are two relatively independent subtasks that cannot reflect the internal connection and dependency relationship; (3) Lack of extraction of multi-entity relationships in long texts and across sentences; (4) Lack of integrated knowledge alignment, review, and storage mechanism.

[0005] Therefore, there is an urgent need in this field to develop a method for constructing a relational database (knowledge graph) for clinical medicine that overcomes the above-mentioned shortcomings, so as to establish a pediatric disease relational database that realizes automated knowledge joint extraction and knowledge fusion. Summary of the invention

[0006] The purpose of this application is to provide a method for constructing a relational database for clinical medicine, thereby establishing a pediatric disease relational database that realizes automated knowledge joint extraction and knowledge fusion.

[0007] The present application provides a method for constructing a relational database of clinical medicine, comprising the following steps:

[0008] Step S1: inputting a file to be extracted for building the relational database, wherein the file contains clinical medical knowledge;

[0009] Step S2: Processing the file to be extracted with knowledge and converting it into a primary preprocessed text, wherein the primary preprocessed text includes multi-level classification items, and the multi-level classification items include multi-level titles from large to small, and each level of title is followed by all paragraphs under the title of that level;

[0010] Step S3: Segmenting the paragraphs of the primary preprocessed text and then performing title concatenation processing to obtain a secondary preprocessed text, wherein the secondary preprocessed text includes a plurality of long texts with concatenated titles, each concatenated title being formed by connecting titles arranged in hierarchical order in series;

[0011] Step S4: extracting the multiple long texts in step S3 through an artificial intelligence model to obtain knowledge triples, wherein the knowledge triples include a main entity, a relationship, and an object entity related to clinical medicine; and

[0012] Step S5: Based on the knowledge triples obtained in the above steps and the secondary preprocessed text, a database is constructed to obtain the clinical medicine relational database.

[0013] In another preferred example, the primary preprocessed text is a json format text.

[0014] In another preferred example, the primary preprocessed text is a json format text, in which titles of the same level are in the same level classification items and titles of the same level are arranged in the order of titles in the file to be extracted knowledge.

[0015] In another preferred example, the multi-level title in step S2 includes an article title, a chapter title, and a section title.

[0016] In another preferred embodiment, in step S3, the paragraphs of the primary preprocessed text are divided into sentences to obtain a plurality of long texts including sentences, and then a title is concatenated before each sentence to form a plurality of long texts with concatenated titles;

[0017] In another preferred example, the plurality of long texts including sentences may be one or more sentences ending with a period or all sentences ending with a period in a paragraph of the primary pre-processed text.

[0018] In another preferred example, the spliced ​​title can be a first-level title, a first-level title plus a second-level title, a first-level title plus a second-level title plus a third-level title, and so on.

[0019] In another preferred example, the spliced ​​title can be an article title, an article title plus a chapter title, or an article title plus a chapter title plus a subsection title.

[0020] In another preferred embodiment, if there is no paragraph between the upper-level title and the lower-level title in the document to be extracted knowledge, then in the primary preprocessed text obtained by processing in step S2, the upper-level title is in a separate line.

[0021] In another preferred example, if there is no paragraph between the upper-level title and the lower-level title in the file to be knowledge extracted, then in the primary preprocessed text obtained by step S2, the upper-level title is not followed by any paragraph, and in the json file in the primary preprocessed text, the upper-level title is in a separate line.

[0022] In another preferred embodiment, the artificial intelligence model in step S3 is configured to perform a joint extraction step, and the joint extraction step includes the following steps:

[0023] S4a: performing relationship recognition on the multiple long texts in step S3 to determine the existing relationships;

[0024] S4b: extract multiple entities from the multiple long texts in step S3;

[0025] S4c: Sequentially align the relationship determined in step S4a and the entities extracted in step S4b to obtain a knowledge triple in the form of a main entity, a relationship, and an object entity related to clinical medicine, wherein the relationship refers to the relationship between the main entity and the object entity.

[0026] In another preferred example, the entity includes a main entity and an object entity, wherein the main entity includes but is not limited to the description of the disease name, and the object entity includes but is not limited to the clinical symptoms and drug treatment associated with the disease.

[0027] In another preferred embodiment, the main entity is selected from the following group: childhood pneumonia, hand, foot and mouth disease, childhood diarrhea, eczema, allergy, bile duct dilatation, childhood Mycoplasma pneumonia, childhood tuberculous meningitis, bronchial asthma, and childhood monkeypox, and the objective entity includes: the name of symptomatic medicine, symptoms of the disease, causes of the disease, film examinations required for the disease, and other treatment methods.

[0028] In another preferred embodiment, the relationship is selected from the following group: drug treatment, clinical symptoms, complications, etiology, high-risk factors, biomedical examinations, and other treatments.

[0029] In another preferred example, step S6 is also included after step S5, in which the information in the clinical medicine relational database obtained in step S5 is reviewed through machine learning, and after each piece of information is reviewed, a reviewed clinical medicine relational database is constructed; conversely, if the information fails the review, the information is corrected and used again to construct the clinical medicine relational database.

[0030] In another preferred example, after step S4 and before step S5, step S50 is further included: performing knowledge alignment on the knowledge triples obtained in step S4 using a domain language model.

[0031] In another preferred example, the knowledge alignment is performed by the following method, including: converting the entity input domain language model obtained in step S4 into a high-dimensional vector, and then calculating the cosine similarity between the obtained high-dimensional vector of the entity and the standard knowledge vector, and when the calculated cosine similarity reaches a certain threshold, the compared vectors are the same knowledge.

[0032] In another preferred embodiment, the threshold is 0.9.

[0033] In another preferred embodiment, the long text is at least 256 characters.

[0034] In another preferred embodiment, step S3 further includes replacing the English abbreviations in the secondary preprocessed text with Chinese.

[0035] In another preferred example, the text file to be extracted with knowledge is a multi-format file, including but not limited to PDF format files, doc format files, and caj format files.

[0036] In another preferred example, the document to be extracted with knowledge contains a hierarchical relationship of titles, subtitles, paragraphs, and sentences.

[0037] In another preferred example, the file to be extracted knowledge is a text file related to pediatric diseases.

[0038] The present application also provides a device for constructing a relational database of clinical medicine, comprising:

[0039] An input module, the input module is configured to input a file to be extracted for building the relational database, the file containing clinical medical knowledge;

[0040] A primary preprocessing module, wherein the primary preprocessing module is configured to process the file to be extracted with knowledge and convert it into a primary preprocessing text, wherein the primary preprocessing text includes multi-level classification items, wherein the multi-level classification items include multi-level titles from large to small, and each level of title is followed by all paragraphs under the title of that level;

[0041] A sentence splitting and title splicing processing module, wherein the sentence splitting and title splicing module is configured to perform sentence splitting processing on the paragraphs of the primary preprocessed text and then perform title splicing processing to obtain a secondary preprocessed text, wherein the secondary preprocessed text includes a plurality of long texts with spliced ​​titles, and each spliced ​​title is formed by connecting titles arranged in a hierarchical order;

[0042] a joint extraction module, the joint extraction module being configured to extract the plurality of long texts to obtain knowledge triples, the knowledge triples comprising a main entity, a relationship, and an object entity related to clinical medicine; and

[0043] A storage module is configured to store the obtained knowledge triples and secondary preprocessed texts, wherein the obtained knowledge triples and secondary preprocessed texts constitute a relational database of clinical medicine.

[0044] It should be understood that within the scope of the present invention, the above-mentioned technical features of the present invention and the technical features specifically described below (such as embodiments) can be combined with each other to form a new or preferred technical solution. Due to space limitations, they will not be described one by one here. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. It should be understood that the drawings described below are only some implementation examples of the present invention, and ordinary technicians in this field can also obtain other implementation examples based on these drawings without paying creative work.

[0046] Figure 1 Schematic diagram of a method for constructing a relational database for clinical medicine according to an embodiment of the present application

[0047] Figure 2 Detailed schematic diagram of a method for constructing a clinical medical relational database according to an embodiment of the present application

[0048] Figure 3 It is a schematic diagram of the homepage of the clinical evidence-based guideline pdf file "Guidelines for the diagnosis and treatment of community-acquired pneumonia in children (2019 edition).pdf" according to the embodiment of the present application;

[0049] Figure 4 It is a schematic diagram of a json format file "Guidelines for the diagnosis and treatment of community-acquired pneumonia in children (2019 edition)_Community-acquired pneumonia in children.json" parsed into multi-level titles and paragraphs according to an embodiment of the present application;

[0050] Figure 5It is a long text schematic diagram of the Diagnosis and Treatment Guidelines for Community-Acquired Pneumonia in Children (2019 Edition) obtained according to the embodiment of the present application;

[0051] Figure 6 is a schematic diagram of knowledge triples extracted from a long text according to an embodiment of the present application;

[0052] Figure 7 is a schematic diagram of a relational database to be reviewed according to an implementation method of the present application;

[0053] Figure 8 It is a schematic diagram of a graph formed by a relational database to be reviewed according to an implementation method of the present application. DETAILED DESCRIPTION

[0054] After extensive and in-depth research, the inventors provide a method for constructing a pediatric disease relational database. The method establishes a hierarchical relationship between titles, subtitles, paragraphs, and sentences by parsing or processing multi-format files to be extracted for knowledge (such as clinical evidence-based guidelines pdf text files), and integrates these hierarchical relationship information into context information. In other words, a long text formed by one or more sentences in each paragraph in the text file is assigned a spliced ​​title, and the spliced ​​title includes all titles before the long text, so that the paragraphs and sentences in the text file are integrated into the hierarchical relationship of the title; in addition, the method also adopts a joint extraction model of relationship recognition, entity extraction and sequence alignment to realize the joint extraction of long texts and multi-entity relationships across sentences; and vectorizes knowledge triples through a domain language model, and performs knowledge alignment by calculating similarity; and also performs knowledge audit on the obtained relational data, and a process mechanism for knowledge storage in a non-relational database (audited relational database).

[0055] the term

[0056] As used herein, “non-relational database” and “audited relational database” are used interchangeably;

[0057] It should be noted that in the application documents of this patent, relational terms such as first and second, etc. are merely used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations.

[0058] The terms "include", "comprises" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one" do not exclude the existence of other identical elements in the process, method, article or device including the elements. In the application document of this patent, if it is mentioned that a certain action is performed according to a certain element, it means that the action is performed at least according to the element, which includes two situations: performing the action only according to the element, and performing the action according to the element and other elements. Expressions such as multiple, multiple, and multiple include 2, 2 times, 2 kinds, and more than 2, more than 2 times, and more than 2 kinds. It should be noted that in the application document of this patent, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations.

[0059] Relational Database

[0060] As used herein, the terms “relational database” and “knowledge graph” are used interchangeably;

[0061] Knowledge Graph

[0062] The knowledge graph is essentially a semantic network consisting of nodes and edges. The network nodes represent entities, concepts or attributes, and the edges represent the relationships between them.

[0063] Entity: refers to an independent thing with clear characteristics that can be distinguished from other things. In the fields of information extraction, natural language processing, and knowledge graphs, the information used to describe these things is called an entity. Entities can be abstract or concrete. In knowledge graphs, knowledge points are represented as entities; in graph theory, knowledge storage, or graph databases, entities are represented as vertices.

[0064] Relationship: describes the objective association between nodes, and is a directed, semantic representation between entities. In knowledge graphs, the associations and connections between knowledge are represented as relationships; in graph theory, knowledge storage, or graph databases, relationships are represented as edges.

[0065] The main advantages of the present invention are:

[0066] (a) The method of constructing a relational database for clinical medicine in the present application realizes the parsing and text extraction of multi-format text files (such as clinical evidence-based guidelines PDF files), the segmentation and sentence division of multi-level title texts, the replacement of the same English abbreviations with different Chinese meanings, and the integration of title information into context information. In other words, the method of the present application realizes format conversion after parsing multi-format text files (such as clinical evidence-based guidelines PDF files), establishes hierarchical relationships between titles, subtitles, paragraphs, and sentences, and integrates this information into context information;

[0067] (b) The method of constructing a clinical medicine relational database in this application uses a joint extraction model of relation recognition, entity extraction and sequence alignment to extract multi-entity relations in long texts and across sentences, thereby improving the prediction ability and the generalization and extraction efficiency;

[0068] (c) The method of constructing a relational database for clinical medicine in this application solves the problem of a large number of English abbreviations and the difficulty in identifying abbreviations in professional academic documents. In the absence of a synonym list, the domain language model is used to vectorize knowledge triples and calculate similarity to perform knowledge alignment;

[0069] (d) The method of constructing a relational database for clinical medicine in the present application establishes an integrated knowledge review and storage system that integrates relational databases (unaudited relational databases) and non-relational databases (audited relational databases).

[0070] In the following description, many technical details are provided to help readers better understand the present application. However, those skilled in the art can understand that the technical solution claimed in the present application can be implemented even without these technical details and various changes and modifications based on the following embodiments.

[0071] This application focuses on multi-source data knowledge in the field of pediatrics, condenses clinical knowledge, explores technologies such as large-scale, multimodal data integration processing, diagnosis and treatment rule knowledge mining, designs a pediatric disease knowledge graph construction system, and realizes automated knowledge joint extraction and fusion, providing a solid resource foundation for pediatrician talent training, related clinical research, and optimal allocation of resources in the pediatric field.

[0072] The purpose of this application is to parse the text files of clinical evidence-based guidelines, divide long texts according to title paragraphs, replace English abbreviations, and integrate contextual information; based on the sequence model, realize the joint extraction of multi-entity relationships in the knowledge graph in long texts; extract multi-entity relationships in long texts and across sentences; and design an integrated process mechanism for knowledge alignment, review, and storage.

[0073] See also Figure 1The present application provides a method for constructing a relational database of clinical medicine, comprising the following steps:

[0074] Step S1: inputting a file to be extracted for building the relational database, wherein the file contains clinical medical knowledge;

[0075] Step S2: Processing the file to be extracted with knowledge and converting it into a primary preprocessed text, wherein the primary preprocessed text includes multi-level classification items, and the multi-level classification items include multi-level titles from large to small, and each level of title is followed by all paragraphs under the title of that level;

[0076] Step S3: Segmenting the paragraphs of the primary preprocessed text and then performing title concatenation processing to obtain a secondary preprocessed text, wherein the secondary preprocessed text includes a plurality of long texts with concatenated titles, each concatenated title being formed by connecting titles arranged in hierarchical order in series;

[0077] Step S4: extracting the multiple long texts in step S3 through an artificial intelligence model to obtain knowledge triples, wherein the knowledge triples include a main entity, a relationship, and an object entity related to clinical medicine; and

[0078] Step S5: Based on the knowledge triples obtained in the above steps and the secondary preprocessed text, a database is constructed to obtain the clinical medicine relational database.

[0079] In one embodiment, the primary preprocessed text is a json format text, in which titles of the same level are in the same level classification items and are arranged in the order of titles in the file to be extracted. In one embodiment, the multi-level titles in step S2 include article titles, chapter titles, and section titles.

[0080] In one embodiment, if there is no paragraph between the upper-level title and the lower-level title in the file to be extracted, then in the primary preprocessed text obtained by step S2, the upper-level title is in a separate line. In one embodiment, if there is no paragraph content between the upper-level title and the lower-level title in the file to be extracted, then in the primary preprocessed text obtained by step S2, the upper-level title is not followed by any paragraph, and in the json file in the primary preprocessed text, the upper-level title is in a separate line.

[0081] In one embodiment, in step S3, the paragraphs of the primary preprocessed text are divided into sentences to obtain a plurality of long texts including sentences, and then a title is concatenated before each sentence to form a plurality of long texts with concatenated titles; in one embodiment, the plurality of long texts including sentences may be one or more sentences ending with a period or all sentences ending with a period in the paragraphs of the primary preprocessed text (see Figure 4 content).

[0082] In one embodiment, the spliced ​​title can be a first-level title, a first-level title plus a second-level title, a first-level title plus a second-level title plus a third-level title, and so on. In one embodiment, the spliced ​​title can be an article title, an article title plus a chapter title, or an article title plus a chapter title plus a subsection title.

[0083] In one embodiment, the artificial intelligence model in step S3 is configured to perform a joint extraction step, and the joint extraction step includes the following steps:

[0084] S4a: performing relationship recognition on the multiple long texts in step S3 to determine the existing relationships;

[0085] S4b: extract multiple entities from the multiple long texts in step S3;

[0086] S4c: Sequentially align the relationship determined in step S4a and the entities extracted in step S4b to obtain a knowledge triple in the form of a main entity, a relationship, and an object entity related to clinical medicine, wherein the relationship refers to the relationship between the main entity and the object entity.

[0087] In one embodiment, the entity includes a main entity and a guest entity, wherein the main entity includes but is not limited to a description of the disease name, and the guest entity includes but is not limited to clinical symptoms and drug treatment associated with the disease.

[0088] In one embodiment, the main entity is selected from the following group: childhood pneumonia, hand, foot and mouth disease, childhood diarrhea, eczema, allergy, bile duct dilatation, childhood Mycoplasma pneumonia, childhood tuberculous meningitis, bronchial asthma, and childhood monkeypox, and the objective entity includes: the name of symptomatic medicine, symptoms of the disease, causes of the disease, film examinations required for the disease, and other treatment methods.

[0089] In one embodiment, the relationship is selected from the group consisting of: drug treatment, clinical symptoms, complications, etiology, high-risk factors, biopsy, and other treatments.

[0090] In one embodiment, step S6 is further included after step S5, in which the information in the clinical medicine relational database obtained in step S5 is reviewed through machine learning, and after each piece of information is reviewed, a reviewed clinical medicine relational database is constructed; conversely, if the information fails the review, the information is corrected and used again to construct the clinical medicine relational database.

[0091] In one embodiment, after step S4 and before step S5, step S50 is further included: performing knowledge alignment on the knowledge triples obtained in step S4 by using a domain language model.

[0092] In one embodiment, the knowledge alignment is performed by the following method, including: converting the entity input domain language model obtained in step S4 into a high-dimensional vector, and then calculating the cosine similarity between the obtained high-dimensional vector of the entity and the standard knowledge vector. When the calculated cosine similarity reaches a certain threshold (0.9), the compared vectors are the same knowledge.

[0093] In one embodiment, the long text is at least 256 characters. In one embodiment, in step S3, it also includes replacing the English abbreviations in the secondary preprocessed text with Chinese. In one embodiment, the text file to be extracted for knowledge is a multi-format file, including but not limited to PDF format files, doc format files, and caj format files. In one embodiment, the file to be extracted for knowledge contains a hierarchical relationship of titles, subtitles, paragraphs, and sentences. In one embodiment, the file to be extracted for knowledge is a text file related to pediatric diseases.

[0094] The present application also provides a device for constructing a relational database of clinical medicine, comprising:

[0095] An input module, the input module is configured to input a file to be extracted for building the relational database, the file containing clinical medical knowledge;

[0096] A primary preprocessing module, wherein the primary preprocessing module is configured to process the file to be extracted with knowledge and convert it into a primary preprocessing text, wherein the primary preprocessing text includes multi-level classification items, wherein the multi-level classification items include multi-level titles from large to small, and each level of title is followed by all paragraphs under the title of that level;

[0097] A sentence splitting and title splicing processing module, wherein the sentence splitting and title splicing module is configured to perform sentence splitting processing on the paragraphs of the primary preprocessed text and then perform title splicing processing to obtain a secondary preprocessed text, wherein the secondary preprocessed text includes a plurality of long texts with spliced ​​titles, and each spliced ​​title is formed by connecting titles arranged in a hierarchical order;

[0098] A joint extraction module, wherein the joint extraction module is configured to extract the plurality of long texts to obtain knowledge triples, wherein the knowledge triples include a main entity, a relationship, and an object entity related to clinical medicine;

[0099] and a storage module, wherein the storage module is configured to store the obtained knowledge triples and the secondary preprocessed texts, wherein the obtained knowledge triples and the secondary preprocessed texts constitute a relational database of clinical medicine.

[0100] The present application provides a method for constructing a relational database of clinical medicine, comprising the following steps:

[0101] Step S1: input a multi-format file, wherein the multi-format file refers to a file containing hierarchical relationships such as titles, subtitles, paragraphs, and sentences. The present invention uses a clinical evidence-based guideline pdf file;

[0102] Step S2: Call the parser of the corresponding format to parse the PDF file in step 01 into a json format file with multi-level titles and paragraphs;

[0103] Step S3: After dividing the paragraphs into sentences, the title names are concatenated to form a long text (at least 256 characters);

[0104] Step S4: Use the joint extraction model to extract the long text. First, perform relationship recognition to determine which relationships are included, then extract multiple entities, and finally align the order of entities and relationships to obtain knowledge triples.

[0105] Step S5: Store the knowledge-aligned triples, long texts, and file information into a relational database to be reviewed; for example Figure 7 An example is given, Figure 7 is a schematic diagram of a relational database to be reviewed according to an embodiment of the present application;

[0106] Step S6: Multiple experts review the information stored in the relational database in step 05. After at least half of the experts have reviewed each piece of information, it is stored in the non-relational database (the reviewed relational database); the information that has not passed the review is added, deleted, or modified, and then stored again in the relational database to be reviewed. For example, Figure 8 A schematic diagram of the graph formed by the relational database to be reviewed is given

[0107] To make the purpose, technical solution and advantages of the present invention clearer, the embodiments of the present invention will be further described in detail below in conjunction with the accompanying drawings. It should be understood that these are only some examples that the present invention may take for the reader, but are not intended to limit the scope of the present invention.

[0108] Example

[0109] See also Figure 2-Figure 8 , describes a method for constructing a pediatric disease relational database based on the clinical evidence-based guideline pdf file "Guidelines for the diagnosis and treatment of community-acquired pneumonia in children (2019 edition).pdf". Among them, Figure 2 The detailed steps of the method are shown. Figure 3-Figure 7 Example pictures for each step are shown, and the method includes the following steps:

[0110] Step 01: Input the clinical evidence-based guideline pdf file "Guidelines for the diagnosis and treatment of community-acquired pneumonia in children (2019 edition).pdf", where Figure 3 The first page of the clinical evidence-based guideline pdf file "Guidelines for the diagnosis and treatment of community-acquired pneumonia in children (2019 edition).pdf" is displayed;

[0111] Step 02: Call the corresponding format parser to parse the pdf file into a json format file with multi-level titles and paragraphs "Guidelines for the diagnosis and treatment of community-acquired pneumonia in children (2019 edition)_Community-acquired pneumonia in children.json", see Figure 4 ;

[0112] The json format file is a primary preprocessed text, which includes multi-level classification items, including multi-level titles from large to small (for example, the first-level title "1 Introduction", "2 Main Pathogen Spectrum and Current Status of Drug Resistance"; the second-level title "2.1 Main Pathogen Spectrum", the third-level title "2.1.1 Respiratory Viruses", "2.1.2 Bacteria"), each level of title is followed by all paragraphs under the title; for example, the first-level title "1 Introduction" is followed by the paragraph content under the title. Of course, in this embodiment, the first-level title (2 Main Pathogen Spectrum and Current Status of Drug Resistance) is followed by the second-level title (2.1 Main Pathogen Spectrum), and "2 Main Pathogen Spectrum and Current Status of Drug Resistance" is not followed by paragraph content, and it is in a separate line;

[0113] In the json format file, the titles at the same level are in the same level classification items and are arranged in the order of the titles in the file to be extracted, such as the first-level titles "1 Introduction", "2 Main Pathogen Spectrum and Drug Resistance Status", "2.1.1 Respiratory Viruses", and "2.1.2 Bacteria";

[0114] Step 03: After dividing the paragraphs into sentences, concatenate the title names, replace the Chinese characters with English abbreviations, and form a long text; this step establishes the hierarchical relationship between titles, subtitles, paragraphs, and sentences, and integrates these hierarchical relationship information into the context information, see Figure 5 .

[0115] In this step, a secondary preprocessed text is obtained, which contains multiple long texts with concatenated titles, each of which is composed of titles arranged in hierarchical order, such as "Community-acquired pneumonia in children @ [Introduction]", "Community-acquired pneumonia in children @ [Main pathogen spectrum and drug resistance status] [Main pathogen spectrum] [Respiratory virus]";

[0116] Step 04: Use the joint extraction model to extract long texts. First, identify the relationships to determine which relationships are included, then extract multiple entities, and finally align the order of entities and relationships to obtain knowledge triples. Figure 6 ,In order to more clearly display the extracted knowledge triples, as shown below.

[0117]

[0118] Community-acquired pneumonia in children @

Main pathogen spectrum and drug resistance status

Main pathogen spectrum

Respiratory viruses

[0119]

[0120]

[0121] The method may also include a knowledge alignment step, and the present invention uses a machine-trained domain language model for knowledge alignment. The present invention first inputs the training set into the domain language model, converts the knowledge triples into high-dimensional vectors, calculates the cosine similarity with the standard knowledge vector, and obtains a threshold of 0.9; after verification by the validation set, when the cosine similarity between the vectors is above 0.9, it represents that the compared vectors are the same knowledge. This embodiment uses a language model in the field of pediatric diseases to align knowledge into the same knowledge. The knowledge described in this step includes but is not limited to disease names, disease descriptions, etc.

[0122] Step05: Store the knowledge triples, long texts, and file information extracted in step04 into a relational database to be reviewed;

[0123] Step 06: Multiple experts review the information stored in the relational database in step 05. After at least half of the experts have reviewed and approved each piece of information, it will be stored in the non-relational database (the reviewed relational database); after adding, deleting, and modifying the information that has not passed the review, it will be stored again in the relational database to be reviewed to form a graph.

[0124] All documents mentioned in this application are considered to be included in the disclosure of this application as a whole, so that they can be used as the basis for modification when necessary. In addition, it should be understood that after reading the above disclosure of this application, those skilled in the art can make various changes or modifications to this application, and these equivalent forms also fall within the scope of protection claimed in this application.

Claims

1. A method for constructing a relational database for clinical medicine, It is characterized in that The following steps are involved: Step S1: inputting a file to be extracted for building the relational database, wherein the file contains clinical medical knowledge; Step S2: Processing the file to be extracted with knowledge and converting it into a primary preprocessed text, wherein the primary preprocessed text includes multi-level classification items, and the multi-level classification items include multi-level titles from large to small, and each level of title is followed by all paragraphs under the title of that level; Step S3: Segmenting the paragraphs of the primary preprocessed text and then performing title concatenation processing to obtain a secondary preprocessed text, wherein the secondary preprocessed text includes a plurality of long texts with concatenated titles, each concatenated title being formed by connecting titles arranged in hierarchical order in series; Step S4: extracting the multiple long texts in step S3 through an artificial intelligence model to obtain knowledge triples, wherein the knowledge triples include a main entity, a relationship, and an object entity related to clinical medicine; and Step S5: Based on the knowledge triples obtained in the above steps and the secondary preprocessed text, a database is constructed to obtain the clinical medicine relational database.

2. The method according to claim 1, It is characterized in that The primary preprocessed text is in JSON format.

3. The method according to claim 1, It is characterized in that If there is no paragraph between the upper-level title and the lower-level title in the document to be extracted, then in the primary preprocessed text obtained by processing in step S2, the upper-level title is in a separate line.

4. The method according to claim 1, It is characterized in that The artificial intelligence model in step S3 is configured to perform a joint extraction step, and the joint extraction step includes the following steps: S4a: performing relationship recognition on the multiple long texts in step S3 to determine the existing relationships; S4b: extract multiple entities from the multiple long texts in step S3; S4c: Sequentially align the relationship determined in step S4a and the entities extracted in step S4b to obtain a knowledge triple in the form of a main entity, a relationship, and an object entity related to clinical medicine, wherein the relationship refers to the relationship between the main entity and the object entity.

5. The method according to claim 4, It is characterized in that The entity includes a main entity and an object entity. The main entity includes but is not limited to the description of the disease name, and the object entity includes but is not limited to clinical symptoms and drug treatment associated with the disease.

6. The method according to claim 5, It is characterized in that The relationship is selected from the following group: drug treatment, clinical symptoms, complications, etiology, high-risk factors, pathological examinations, and other treatments.

7. The method according to claim 1, It is characterized in that After step S5, step S6 is also included, in which the information in the clinical medicine relational database obtained in step S5 is reviewed through machine learning. After each piece of information is reviewed, a reviewed clinical medicine relational database is constructed; conversely, if the information fails the review, the information is corrected and used again to construct the clinical medicine relational database.

8. The method according to claim 1, It is characterized in that After step S4 and before step S5, the method further includes step S50: performing knowledge alignment on the knowledge triples obtained in step S4 by using a domain language model.

9. The method according to claim 1, It is characterized in that In step S3, the process also includes replacing the English abbreviations in the secondary preprocessed text with Chinese.

10. A device for constructing a relational database for clinical medicine, It is characterized in that include: An input module, the input module is configured to input a file to be extracted for building the relational database, the file containing clinical medical knowledge; A primary preprocessing module, wherein the primary preprocessing module is configured to process the file to be extracted with knowledge and convert it into a primary preprocessing text, wherein the primary preprocessing text includes multi-level classification items, wherein the multi-level classification items include multi-level titles from large to small, and each level of title is followed by all paragraphs under the title of that level; A sentence splitting and title splicing processing module, wherein the sentence splitting and title splicing module is configured to perform sentence splitting processing on the paragraphs of the primary preprocessed text and then perform title splicing processing to obtain a secondary preprocessed text, wherein the secondary preprocessed text includes a plurality of long texts with spliced ​​titles, and each spliced ​​title is formed by connecting titles arranged in a hierarchical order; A joint extraction module, wherein the joint extraction module is configured to extract the plurality of long texts to obtain knowledge triples, wherein the knowledge triples include a main entity, a relationship, and an object entity related to clinical medicine; and A storage module is configured to store the obtained knowledge triples and secondary preprocessed texts, wherein the obtained knowledge triples and secondary preprocessed texts constitute a relational database of clinical medicine.

Citation Information

Patent Citations

  • Long text classification method and device

    CN112100389A

  • Knowledge graph construction method, knowledge graph construction system and computing equipment

    CN114817553A

  • Intelligent archive construction method fusing artificial intelligence and knowledge graph technology

    CN115994230A

Cited By

  • Digital intelligence system for plasma cell disease and related big health management fused with AI model

    CN121094134A