An intelligent processing method for physical examination data based on a medical knowledge graph

Abstract

The present invention discloses an intelligent processing method for physical examination data based on a medical knowledge graph, comprising the steps of: S1, establishing a medical knowledge graph; S2, collecting the physical examination data of a patient, including user basic information, examination indicators, test indicators, imaging description texts, and general examination data; S3, intelligently cleaning the general examination data based on the medical knowledge graph; S4, intelligently extracting the imaging descriptions; S5, discretizing the indicators, establishing abnormal indicator labels, and performing unified structured processing. The present invention can not only improve the processing quality of physical examination data, but also provide more accurate and convenient data analysis tools for medical researchers, doctors, and health management personnel, thus playing an important role in the field of medical health.

Description

Technical Field

[0001] The present invention relates to the technical field of medical data analysis, and particularly to an intelligent processing method for physical examination data based on a medical knowledge graph. Background Art

[0002] In the field of medical health, physical examination is an important means of preventing diseases and detecting health problems at an early stage. With the development of medical informatization, the electronic processing of physical examination data has become increasingly common. However, physical examination data usually contains a large amount of unstructured text information, such as examination reports, test results, and imaging descriptions. The processing and analysis of these data face many challenges.

[0003] Firstly, the sources of physical examination data are diverse, and different medical institutions may use different data formats and terminology standards, resulting in problems with data consistency and comparability. In addition, the parsing of unstructured text data requires a high level of professional knowledge and manual participation, which not only increases the burden on medical staff but also limits the efficiency and accuracy of data processing.

[0004] Secondly, traditional data processing methods rely on manual review and manual input, which are not only time-consuming and laborious but also easily affected by subjective judgment, resulting in the quality and reliability of the data being affected. In addition, with the continuous growth of the volume of physical examination data, how to quickly and accurately extract valuable medical information from a large amount of data has become the key to improving the quality of medical services.

[0005] To solve the above problems, in recent years, artificial intelligence technologies, especially natural language processing (NLP) and machine learning, have begun to be applied to medical data analysis. These technologies can automatically identify and extract key information such as medical concepts, symptoms, and diseases in text, thereby assisting medical staff in making more accurate diagnoses and treatment decisions. However, existing AI solutions often require a large amount of labeled data for training, and may lack sufficient depth and accuracy when dealing with medical knowledge in specific fields.

[0006] The knowledge graph, officially named after it was proposed by Google in 2012, describes concepts, entities, and their relationships in the objective world in a structured form, expresses the information on the Internet in a form closer to the human cognitive world, and provides an ability to better organize, manage, and understand the vast amount of information on the Internet. Summary of the Invention

[0007] Aiming at the deficiencies of the prior art, the present invention aims to provide an intelligent processing method for physical examination data based on a medical knowledge graph.

[0008] To achieve the above object, the present invention adopts the following technical solutions:

[0009] An intelligent processing method for physical examination data based on a medical knowledge graph, comprising the following steps:

[0010] S1. Establish a medical knowledge graph;

[0011] S2. Collect the physical examination data of the patient, including user basic information, test indicators, imaging description text, and general inspection data;

[0012] S3. Intelligently clean the general inspection data based on the medical knowledge graph:

[0013] S3.1. Denote the set of general inspection data collected from the patient as S. Traverse the set S. For the i-th general inspection data S i , extract the detection method and general inspection description in the general inspection data according to the regular expression of the rule, and perform data preprocessing on S i to remove interference items;

[0014] S3.2. Use a pre-trained NLP model to extract the corresponding symptom entities, main parts, and secondary parts in the general inspection description;

[0015] S3.3. Obtain the relevant symptoms of the medical knowledge points from the medical knowledge graph KG tcm , convert them into the record form of <part, symptom KR sx , keyword KW s , diagnosis KR s >, and form a record set KR ssx ;

[0016] S3.4. Enter the symptom diagnosis process and output the diagnosis conclusion:

[0017] S3.4.1. Extract the records of <main part, symptom KR ssx , keyword KW sx , diagnosis KR s > from KR s according to the main part, and form a record set KR mssx ;

[0018] S3.4.2. Sort the record set KR mssx , where the main part takes precedence, and among those with the same main part, those with more keywords KW s take precedence;

[0019] S3.4.3. Match the symptom entities extracted in step S3.2 with the symptom KR mssx in the sorted record set KR sx to obtain a new ordered record set KR match , and calculate the matching degree score for each record in the S i and KR match sets: First, for Si and KR match the keyword set KW of each record in the set s Perform preprocessing, including part removal and symbol removal, and then S i Perform similarity matching with each keyword of each record in the set. Set the variable matching times count and the keyword similarity score score. If S i contains a certain keyword, the matching times of this record are counted as 1, otherwise the matching times are counted as 0, and the cosine similarity algorithm is used to calculate S i and the keyword set KW of each record s the similarity of each keyword in; finally, accumulate the similarities of each keyword and calculate the average value to obtain the final matching score score of the corresponding record. If count is equal to the number of keywords of the record, determine whether the matching score meets the set credibility threshold. If it meets, the record is judged to match successfully, otherwise the record does not match;

[0020] S3.4.4. If the number of main part matching results is greater than 1, select the diagnostic record with the highest matching score as the diagnostic conclusion; otherwise, select the secondary part, and then calculate according to steps S3.4.1 - S3.4.3 to obtain the diagnostic record with the highest matching degree as the diagnostic conclusion;

[0021] S4. Intelligent extraction of imaging descriptions:

[0022] S4.1. Based on the Sinohealth - AI annotation open platform, annotate the text data of imaging descriptions to form the NLP information extraction model M;

[0023] Before inputting the new imaging description text into the model M, it needs to be preprocessed, and the regular rule set is used to process the data to make its format consistent with the data during training;

[0024] S4.3. The model M extracts the attributes of the preprocessed imaging description text, including extracting the part, nodule type, ultrasonic characteristics, nodule diameter, and outputs them in a structured manner;

[0025] S5. Discretize the test indicators, establish abnormal indicator labels, and perform unified structured processing.

[0026] Furthermore, the specific process of step S1 is as follows:

[0027] S1.1. Knowledge modeling, determining the mutual association relationship between parts, diseases, symptoms, and indicators;

[0028] S1.2. Determine the knowledge source;

[0029] S1.3. Knowledge extraction:

[0030] (1) For structured data, map each field to a graph node and establish the association relationships between graph nodes;

[0031] (2) For unstructured data, according to the structural normativity of the text and the obviousness of sentence features, select the corresponding processing method for entity extraction; for texts with normative and fixed structures, use the method in 2.1) below for processing; for texts with obvious sentence features, use the method in 2.2) below for processing, and in other cases, use the method in 2.3) below for processing:

[0032] 2.1) Use a rule-based regular expression template to extract entities from the text;

[0033] 2.2) Apply statistical learning methods for sequence labeling, including using a labeled set for data labeling, setting a feature template, then using the CRF++ tool for model training, finally forming a CRF model, and finally inputting the preprocessed unstructured data into the CRF model to identify entities;

[0034] 2.3) Adopt deep learning methods for sequence labeling. The training stage includes data cleaning, labeling with the BIO labeled set, and the design of a BiLSTM model. The BiLSTM model includes an input layer, an embedding layer, a BiLSTM layer, a CRF layer, and an output layer, where the embedding layer is pre-trained using GloVe; after training, input the unstructured data to be processed into the model for entity extraction;

[0035] After entity extraction, according to the task requirements and data characteristics, select any of the following methods for relation extraction:

[0036] a) On the basis of an existing partial knowledge graph, adopt the remote supervision method for relation extraction;

[0037] b) For the case of a large amount of data with significant characteristics, use the template matching method to identify the relation expressions between entities and map them to relations;

[0038] c) In complex scenarios, manually label entity relations to ensure the accuracy and reliability of extraction;

[0039] (3) For semi-structured data, combine (1) and (2) to solve: First, use a parsing tool to identify the data format, parse the data to extract structured elements and attributes; extract the clearly identifiable structured information from the semi-structured data, and use the method in (1) for parsing and extraction. For the unstructured part in the extracted data, use the method in (2) for parsing and extraction;

[0040] Finally, construct and improve the medical knowledge graph manually;

[0041] S1.4, Knowledge Fusion: Based on the term entities in medical terms and medical dictionaries, entity alignment is achieved through the following methods:

[0042] (1.4.1) Term Synonym Mapping: Utilize the synonym information in medical terms and dictionaries to achieve the alignment of term entities;

[0043] (1.4.2) Entity Matching with Expert Intervention: When dealing with complex or ambiguous entity alignment problems, rely on the knowledge and experience of medical experts for accurate matching;

[0044] S1.5, Knowledge Storage: Store the constructed medical knowledge graph in a database system for easy retrieval and update.

[0045] The present invention also provides a computer-readable storage medium, in which a computer program is stored, and when the computer program is executed by a processor, the above method is implemented.

[0046] The present invention also provides a computer device, including a processor and a memory, where the memory is used to store a computer program; when the processor executes the computer program, the above method is implemented.

[0047] The beneficial effects of the present invention are as follows:

[0048] 1. By integrating and utilizing the medical knowledge graph, the present invention enables artificial intelligence to understand and process complex medical terms and concepts. The structured medical knowledge in the knowledge graph provides depth and accuracy for physical examination data cleaning and information extraction, ensuring the professionalism and authority of the processing results.

[0049] 2. By using advanced natural language processing technology, the present invention can automatically identify disease descriptions, test results, and imaging features in physical examination reports, reducing the dependence on manual input and improving the speed and accuracy of data processing.

[0050] 3. By optimizing the data processing process through machine learning algorithms, the present invention can quickly process large-scale physical examination data sets while maintaining high-accuracy medical information extraction, significantly improving the efficiency of medical data analysis.

[0051] 4. By converting unstructured physical examination data into standardized medical terms, the present invention helps to achieve data interoperability between different medical institutions and information systems, providing a solid data foundation for electronic medical records and telemedicine.

[0052] 5. The design of the present invention allows for easy integration of new medical knowledge, adapts to the continuously updated medical research and clinical practice, and ensures the long-term efficiency and accuracy of the system.

[0053] In summary, the present invention can not only improve the processing quality of physical examination data, but also provide more accurate and convenient data analysis tools for medical researchers, doctors, and health managers, thus playing an important role in the field of medical health. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is the overall flowchart of the method of the embodiment of the present invention;

[0055] Figure 2 is an example diagram of the medical knowledge graph constructed in the method of the embodiment of the present invention;

[0056] Figure 3 is the symptom diagnosis flowchart in the embodiment of the present invention;

[0057] Figure 4 is an example diagram of the imaging description annotation in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0058] The present invention will be further described below with reference to the accompanying drawings. It should be noted that this embodiment is based on the present technical solution and gives detailed implementation manners and specific operation processes, but the protection scope of the present invention is not limited to this embodiment.

[0059] This embodiment provides an intelligent processing method for physical examination data based on a medical knowledge graph, as Figure 1 shown, including the following steps:

[0060] S1. Establish a medical knowledge graph. The medical knowledge graph established in this embodiment is as Figure 2 shown. The establishment process of the medical knowledge graph mainly includes knowledge modeling, determining knowledge sources, knowledge extraction, knowledge fusion, knowledge storage, etc.

[0061] S1.1. Knowledge modeling, determining the mutual association relationships among parts, diseases, symptoms, and indicators;

[0062] S1.2. Determine knowledge sources, which may include medical literature, clinical practice data, medical knowledge bases, encyclopedic knowledge on the Internet, ancient medical records and medical books, etc.

[0063] S1.3. Knowledge extraction: This embodiment method adopts various strategies to process different types of data to construct a medical knowledge graph.

[0064] (1) For structured data, map each field to a graph node and establish the association relationships between the graph nodes. For example, the parts, symptoms, indicators, etc. in the physical examination report are used as graph nodes, and the relationships between them are expressed as the association between symptoms and diseases, the association between parts and parts, the association between parts and diseases, etc.;

[0065] (2) For unstructured data, first select the appropriate processing method for entity extraction based on the structural standardization of the text and the obviousness of the sentence features. For text with a standard and fixed structure, use the following 2.1) method for processing; for text with obvious sentence features, use the following 2.2) method for processing; otherwise, use the following 2.3) method for processing.

[0066] 2.1) Extract entities from text based on rule-based regular expression templates.

[0067] 2.2) Use statistical learning methods, such as the conditional random field (CRF) algorithm, for sequence labeling. This process includes using a labeling set such as BMEWO to label data, setting feature templates, and then using the CRF++ tool to train the model to eventually form a CRF model. Finally, the unstructured data is input into the CRF model to identify entities.

[0068] 2.3) Use deep learning methods, such as BiLSTM combined with CRF, for sequence labeling. The training phase includes data cleaning, labeling of the BIO label set, and BiLSTM model design. The BiLSTM model includes an input layer, an embedding layer, a BiLSTM layer, a CRF layer, and an output layer. The embedding layer is pre-trained using GloVe. After the training is completed, the unstructured data to be processed is input into the model for entity extraction.

[0069] After entity extraction, choose any of the following methods to extract relationships based on task requirements and data characteristics:

[0070] a) Based on the existing knowledge graph, a remote supervision method is used to extract relations;

[0071] b) For large amounts of data and significant features, use template matching methods to identify the relationship expressions between entities and map them to relationships;

[0072] c) In complex scenarios, entity relationships are manually labeled to ensure the accuracy and reliability of extraction.

[0073] (3) For semi-structured data, combine (1) and (2) to solve the problem: first use a parsing tool to identify the data format and parse the data to extract structured elements and attributes. Extract clearly identifiable structured information from semi-structured data, such as text content within XML or HTML tags, using method (1) for parsing and extraction. For the unstructured part of the extracted data (such as text description), use method (2) for parsing and extraction.

[0074] Finally, the medical knowledge graph is constructed and improved through manual methods such as expert method and crowdsourcing method;

[0075] S1.4, Knowledge Fusion: Based on the term entities in medical terms and medical dictionaries, entity alignment is achieved through the following methods:

[0076] (1.4.1) Term Synonym Mapping: Utilize the synonym information in medical terms and dictionaries to achieve entity alignment of term entities. For example, the synonymous terms of the term "swelling" include "edema", "distension", and "lump". By mapping these synonyms, their corresponding relationships with the main term "swelling" are established.

[0077] (1.4.2) Entity Matching with Expert Intervention: When dealing with complex or ambiguous entity alignment problems, rely on the knowledge and experience of medical experts for accurate matching. The intervention of experts ensures the accuracy of entity associations and compliance with medical professionalism.

[0078] S1.5, Knowledge Storage: Store the constructed medical knowledge graph in a database system for easy retrieval and update.

[0079] S2. Collect the physical examination data of patients, including basic user information, examination indicators, test indicators, imaging description texts, and general examination data;

[0080] S3. Intelligently clean the general examination data based on the medical knowledge graph, as Figure 2 shown. The specific process is as follows:

[0081] S3.1. Denote the set of general examination data collected from patients as S. Traverse the set S, and for the i-th general examination data S i , extract the detection method and general examination description in the general examination data according to the regular expression of the rule, and perform data preprocessing on S i to remove interference items, such as symbol standardization, digital noise elimination, special symbol processing, etc.;

[0082] S3.2. Use a pre-trained NLP model to extract the corresponding symptom entities, main parts, and secondary parts in the general examination description.

[0083] S3.3. Obtain the relevant symptoms of medical knowledge points from the medical knowledge graph KG tcm , and convert them into the record form of <part, symptom KR sx , keyword KW s , diagnosis KR s >, and form a record set KR ssx ;

[0084] S3.4. Enter the symptom diagnosis process, as Figure 3 shown, and output the diagnosis conclusion:

[0085] S3.4.1. According to the main part, from KR ssxExtract <main part, symptom KR sx , keyword KW s , diagnosis KR s > records to form a record set KR mssx ;

[0086] S3.4.2. Sort the record set KR mssx , where the main part has priority, and for those with the same main part, the one with more keywords KW s has priority;

[0087] S3.4.3. Match the symptom entities extracted in step S3.2 with the symptoms KR mssx in the sorted record set KR sx to obtain a new ordered record set KR match , and calculate the matching degree score for each record in the S i and KR match sets: First, preprocess the keyword sets KW i in each record of the S match and KR s sets, including part elimination (parts are not included in the score), symbol elimination, etc., and then perform similarity matching between S i and each keyword of each record in the set. Set variables matching times count and keyword similarity score score. If S i contains a certain keyword, the matching times of this record are counted as 1, otherwise the matching times are counted as 0, and use the Cosine Similarity algorithm to calculate the similarity between S i and each keyword in the keyword set KW s of each record. Finally, accumulate the similarities of each keyword and calculate the average value to obtain the final matching score score of the corresponding record. If count is equal to the number of keywords of the record, determine whether the matching score meets the set confidence threshold. If it meets, the record is judged to be successfully matched, otherwise the record is not matched;

[0088] S3.4.4. If the number of main part matching results is greater than 1, select the diagnostic record with the highest matching score; otherwise, select the secondary part, and then calculate according to steps S3.4.1 - S3.4.3 to obtain the diagnostic record with the highest matching degree;

[0089] S4. Intelligent extraction of imaging descriptions:

[0090] S4.1. Based on the Sinohealth - AI annotation open platform, annotate the text data of imaging descriptions to form an NLP information extraction model M. An example of imaging description annotation is as Figure 4 shown.

[0091] S4.2. Before inputting the new imaging description text into model M, it needs to be preprocessed, and the regular rule set is used to process the data to make its format consistent with that of the data during training.

[0092] S4.3. Model M extracts the attributes of the preprocessed imaging description text, extracts information such as the location, nodule type, ultrasonic characteristics, and nodule diameter, and outputs them in a structured manner.

[0093] S5. Discretize the test indicators, establish abnormal indicator labels, and perform unified structured processing.

[0094] For those skilled in the art, various corresponding changes and deformations can be given according to the above technical solutions and concepts, and all these changes and deformations should be included within the protection scope of the claims of the present invention.

Claims

1. A method for intelligent processing of physical examination data based on medical knowledge graph, characterized in that: The steps include: S1. Establish a medical knowledge graph; S2. Collect the patient's physical examination data, including basic user information, test indicators, imaging description text and general examination data; S3. Intelligent cleaning of general inspection data based on medical knowledge graph: S3.

1. Let the total examination data set of the collected patients be S. Traverse the set S. For the i-th total examination data S i , extract the detection method and general inspection description in the general inspection data according to the regular expression of the rule, and i Perform data preprocessing to remove interference items; S3.

2. Use the pre-trained NLP model to extract the corresponding symptom entities and primary and secondary parts in the general examination description; S3.

3. From the medical knowledge graph KG tcm Get the relevant symptoms of medical knowledge points and convert them into <parts, symptoms KR sx , keyword KW s , diagnosis KR s > record form, forming a record set KR ssx ; S3.

4. Enter the symptom diagnosis process and output the diagnosis conclusion: S3.4.1, according to the main part from KR ssx Extraction < Main part, symptoms KR sx , keyword KW s , diagnosis KR s > records, forming a record set KR mssx ; S3.4.

2. Record set KR mssx Sorting is performed, with the main part taking priority, and the same main part is sorted by keyword KW s The one with more number will be given priority; S3.4.

3. Compare the symptom entities extracted in step S3.2 with the sorted record set KR mssx Symptoms in KR sx Match to get a new ordered record set KR match , use S i and KR match Calculate the matching score for each record in the collection: First, i and KR match The keyword set KW for each record in the collection s Preprocessing is performed, including part removal and symbol removal, and then S i With KR match Perform similarity matching on each keyword of each record in the collection, set the variable matching count and keyword similarity score. If S i If a record contains a keyword, the number of matches for that record is counted as 1, otherwise the number of matches is counted as 0, and the cosine similarity is used to calculate S i The keyword set KW for each record s The similarity of each keyword in the record is calculated; finally, the similarity of each keyword is accumulated and the average value is calculated to obtain the final matching score of the corresponding record. If count is equal to the number of keywords in the record, it is determined whether the matching score meets the set credibility threshold. If it meets the threshold, the record is matched successfully, otherwise the record is not matched. S3.4.

4. If the number of matching results for the primary part is greater than 1, select the diagnostic record with the highest matching score as the diagnostic conclusion; otherwise, select the secondary part and follow steps S3.4.1- S3.4.3 performs calculations to obtain the diagnostic record with the highest matching degree as the diagnostic conclusion; S4. Intelligent extraction of imaging descriptions: S4.

1. Based on the Sinohealth-AI annotation open platform, the text data of the imaging description is annotated to form an NLP information extraction model M; S4.

2. Before inputting the new radiological description text into the model M, it needs to be preprocessed, and the data is processed using a regular rule set to make it consistent with the data format during training; S4.3, model M extracts the attributes of the preprocessed imaging description text, including extraction location, nodule type, ultrasound characteristics, nodule diameter, and outputs them in a structured manner; S5. Discretize the inspection indicators, establish abnormal indicator labels, and perform unified structured processing.

2. The intelligent processing method for physical examination data based on medical knowledge graph according to claim 1 is characterized in that: The specific process of step S1 is as follows: S1.1, knowledge modeling, determine the interrelationships among parts, diseases, symptoms, and indicators; S1.2, identify the sources of knowledge; S1.3, Knowledge Extraction: (1) For structured data, each field is mapped to a graph node, and the association relationship between the graph nodes is established; (2) For unstructured data, select the corresponding processing method for entity extraction based on the structural standardization of the text and the obviousness of the sentence features; for text with a standard and fixed structure, use the following 2.1) method for processing; for text with obvious sentence features, use the following 2.2) method for processing; otherwise, use the following 2.3) method for processing: 2.1) Extract entities from text based on regular expression templates; 2.2) Use statistical learning methods for sequence labeling, including using label sets to label data, setting feature templates, and then using CRF++ tools to train the model to eventually form a CRF model. Finally, the preprocessed unstructured data is input into the CRF model to identify entities. 2.3) Using deep learning methods for sequence labeling, the training phase includes data cleaning, labeling of BIO labeling sets, and BiLSTM model design. The BiLSTM model includes an input layer, an embedding layer, a BiLSTM layer, a CRF layer, and an output layer, wherein the embedding layer is pre-trained using GloVe; After training is completed, the unstructured data to be processed is input into the model for entity extraction; After entity extraction, choose any of the following methods to extract relationships based on task requirements and data characteristics: a) Based on the existing knowledge graph, a remote supervision method is used to extract relations; b) For large amounts of data and significant features, use template matching methods to identify the relationship expressions between entities and map them to relationships; c) In complex scenarios, manually annotate entity relationships to ensure the accuracy and reliability of extraction; (3) For semi-structured data, combine (1) and (2) to solve the problem: first use the parsing tool to identify the data format and parse the data to extract structured elements and attributes; extract clearly identifiable structured information from the semi-structured data using (1) and parse and extract the unstructured part of the extracted data using (2); Finally, the medical knowledge graph is constructed and improved manually; S1.4, Knowledge Fusion: Based on medical terms and term entities in medical dictionaries, entity alignment is achieved through the following methods: (1.4.1)Term synonym mapping: Use synonym information in medical terms and dictionaries to achieve term entity alignment; (1.4.2) Expert-mediated entity matching: When dealing with complex or ambiguous entity alignment problems, rely on the knowledge and experience of medical experts for accurate matching; S1.

5. Knowledge storage: The constructed medical knowledge graph is stored in the database system for easy retrieval and updating.

3. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 2 is implemented.

4. A computer device, characterized in that: The method comprises a processor and a memory, wherein the memory is used to store a computer program; when the processor is used to execute the computer program, the method described in any one of claims 1 to 2 is implemented.

Images