Method and system for constructing clinical knowledge graph based on real world data

Abstract

The invention provides a method and a system for constructing a clinical knowledge graph based on real world data. The method comprises the following steps: acquiring electronic health archives, medical insurance data, medical literatures, medical equipment data, patient report outcome and public health data, and carrying out cleaning, desensitization, medical term code mapping and structured conversion on the data; schema adaptive to the clinical field is designed, and entities and core relations of patients, diseases, drugs, symptoms, inspection and treatment schemes and the like are defined; entity, relation and attribute extraction is carried out by adopting a rule method in combination with machine learning, and knowledge fusion is completed through entity alignment, confidence fusion and conflict resolution; and the fused knowledge is stored in a knowledge base, and hidden knowledge is reasoned and mined by adopting rules, embedding or a graph neural network, is finally used for clinical decision support, drug alert and disease research, and is continuously updated along with new data and user feedback, so that the quality and practicability of the graph are improved.

Description

Technical Field

[0001] This invention belongs to the field of medical informatics and medical data processing technology, and in particular relates to a method and system for constructing a clinical knowledge graph based on real-world data. Background Technology

[0002] Real-world data comes from a wide range of sources, including electronic health records, health insurance data, medical literature, and medical device data, containing a wealth of clinical information that can be used for diagnosis, research, and drug development. Clinical knowledge graphs can organize scattered information into a structured and reasonable knowledge network, thereby providing support for clinical decision-making, drug development, and disease research.

[0003] However, existing solutions for constructing clinical knowledge graphs based on real-world data still face several challenges: First, the data formats are complex (structured / semi-structured / unstructured data coexist), and the data quality varies, with issues such as duplication, missing data, anomalies, and input errors, leading to unstable extraction quality. Second, entity duplication or relationship conflicts may exist between multiple data sources, and the lack of effective fusion and conflict resolution mechanisms can result in decreased graph consistency. Third, the accuracy of entity and relationship extraction from unstructured text is insufficient, and erroneous knowledge used in clinical decision-making can affect treatment outcomes. Fourth, insufficient reasoning ability may lead to the omission of important knowledge such as potential disease associations and drug side effects. Fifth, real-world data is dynamically growing, and existing solutions still have shortcomings in terms of dynamic knowledge updates and iterations. Furthermore, real-world data contains a large amount of sensitive patient information, making desensitization and compliant use unavoidable engineering challenges.

[0004] Therefore, there is a need for a clinical knowledge graph construction method that can effectively preprocess and standardize multi-source real-world data, improve the accuracy of knowledge extraction, reduce redundancy and conflicts, support multiple storage and reasoning methods, and be continuously updated with new data and user feedback. Summary of the Invention

[0005] In view of this, the present invention aims to propose a method and system for constructing a clinical knowledge graph based on real-world data, so as to at least solve one of the problems in the background art.

[0006] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A method for constructing a clinical knowledge graph based on real-world data includes: S1. Acquire multi-source real-world data, wherein the multi-source real-world data includes at least electronic health records, medical insurance data, medical literature, medical device data, patient-reported outcomes, and public health data; S2. Perform preprocessing on the multi-source real-world data. The preprocessing includes deduplication, missing and anomaly handling, input error correction, desensitization of personal identity information fields, mapping medical terms to a preset coding database in a unified format, and converting unstructured data into semi-structured or structured data. S3. Perform schema design, defining entity types, relation types, and attribute types. The entity types must include at least patients, diseases, drugs, symptoms, laboratory tests, and treatment plans. S4. Under the schema constraints, entities, relations and attributes are extracted from the preprocessed data. Entity extraction and relation extraction adopt a combination of rule-based methods and machine learning methods. S5. Perform knowledge fusion on the extracted knowledge, including entity alignment and confidence-based relation fusion and conflict resolution; S6. Store the fused knowledge in a preset knowledge base and perform at least one knowledge reasoning to uncover implicit knowledge; S7. Apply the clinical knowledge graph to clinical decision support, pharmacovigilance, or disease research, and update and iterate the knowledge graph when receiving new data and / or user feedback.

[0007] Furthermore, the schema design in step S3 further defines at least the following relation types: patient has disease, disease accompanied by symptoms, disease requires testing, disease requires medication, and medication has side effects; and the schema design satisfies the principles of domain adaptability, scalability, and simplicity.

[0008] Furthermore, the mapping of medical terms to a preset coding database in step S2 includes mapping disease names to ICD-10 codes and mapping drug names to ATC codes or the RxNorm database.

[0009] Furthermore, the rule-based method for entity extraction in step S4 is based on a medical dictionary and regular expressions; the machine learning method for entity extraction includes traditional machine learning and deep learning, wherein traditional machine learning includes at least conditional random fields and support vector machines, and deep learning includes at least BERT or BioBERT pre-trained models.

[0010] Furthermore, the rule-based method for relation extraction in step S4 is based on trigger words and syntactic structure; the machine learning method for relation extraction includes supervised learning and remote supervision, wherein supervised learning transforms relation extraction into a classification problem and trains a CNN or LSTM model using labeled data, and remote supervision automatically labels training data using existing knowledge bases.

[0011] Furthermore, the attribute extraction in step S4 includes: directly extracting attributes from structured data, and extracting attributes from unstructured text and associating them with corresponding entities.

[0012] Furthermore, the entity alignment described in step S5 includes at least string matching and attribute matching, and further includes a machine learning alignment approach that treats entity alignment as a classification problem.

[0013] Furthermore, the machine learning alignment method uses entity names, attributes, and contextual relationships as features to train a classification model, and the classification model includes at least a logistic regression model or a graph neural network model.

[0014] Furthermore, the confidence-based relationship fusion in step S5 includes: assigning confidence levels to same-name relationships from different data sources and retaining high-confidence relationships; the conflict resolution includes determining the relationships to be retained by using conclusions from multiple sources or by combining clinical data verification.

[0015] Furthermore, the knowledge base storage method described in step S6 includes at least one of graph database, relational database, or triplet repository; wherein the graph database includes at least Neo4j, OrientDB, or JanusGraph, the relational database includes at least MySQL or PostgreSQL, and the triplet repository includes at least Virtuoso or Apache Jena.

[0016] Furthermore, the knowledge reasoning described in step S6 includes at least one of the following reasoning methods: rule-based reasoning, embedding-based reasoning, or graph neural network-based reasoning; wherein embedding-based reasoning uses the TransE or DistMult model, rule-based reasoning uses a rule engine to perform implicit relationship reasoning, and graph neural network-based reasoning uses a graph neural network to capture neighbor relationships and structural features for reasoning.

[0017] Furthermore, in step S4 or step S5, at least a metadata set is generated and bound for each extracted or fused piece of knowledge. The metadata set includes at least a source identifier, a generation timestamp, and a confidence level. The confidence level is calculated by weighting the confidence score output by the knowledge extraction and the data quality results of the preprocessing stage, and is used for the relationship fusion in step S5 or the update iteration decision in step S7.

[0018] Furthermore, the update iteration described in step S7 includes an incremental update mechanism: when new or changed data is detected, preprocessing, extraction, fusion, and storage are performed only on entities, relationships, and attributes related to the new or changed data; and the graph versions before and after the update are identified and archived traceably.

[0019] Furthermore, the continuous iterative optimization based on user feedback in step S7 includes: receiving manual correction information on knowledge extraction results, relation conflict resolution results, or reasoning conclusions; converting the correction information into training samples or rule revision items; and updating the entity extraction model, relation extraction model, and / or conflict resolution rules accordingly.

[0020] Furthermore, this solution discloses a system for constructing a clinical knowledge graph based on real-world data, including a data access module, a preprocessing module, a schema design module, a knowledge extraction module, a knowledge fusion module, a knowledge storage module, a knowledge reasoning module, and an application iteration module; The data access module is used to acquire multi-source real-world data. The preprocessing module is used to perform deduplication, missing and anomaly handling, data entry error correction, desensitization, terminology mapping, and structured transformation on the multi-source real-world data; The schema design module is used to define entity types, relationship types, and attribute types; The knowledge extraction module is used to extract entities, relations, and attributes under schema constraints by combining rule-based methods with machine learning methods. The knowledge fusion module is used to perform entity alignment and confidence-based relation fusion and conflict resolution. The knowledge storage module is used to store the merged knowledge in a preset knowledge base; The knowledge reasoning module is used to perform at least one knowledge reasoning based on the knowledge base to uncover implicit knowledge; The application iteration module is used to apply clinical knowledge graphs to clinical decision support, pharmacovigilance, or disease research, and to perform update iterations when new data and / or user feedback are received.

[0021] Furthermore, the knowledge fusion module is further configured to bind source identifiers, timestamps, and confidence levels to the fused knowledge, and to perform relationship fusion and conflict resolution based on the confidence levels; the application iteration module is further configured to perform incremental updates and graph version archiving.

[0022] Furthermore, this solution discloses an electronic device including a processor, a memory, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the method described herein.

[0023] Furthermore, this solution discloses a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method described.

[0024] Compared with existing technologies, the method and system for constructing a clinical knowledge graph based on real-world data described in this invention have the following advantages: (1) This invention cleans, standardizes and structures multi-source real-world data to improve data quality and lay the foundation for subsequent knowledge extraction; (2) The present invention designs a schema adapted to the clinical field, covering core entities, relationships and attribute types, and takes into account domain adaptability, scalability and simplicity, making knowledge organization clearer and more reasonable; (3) The knowledge extraction of this invention adopts a rule-based approach combined with machine learning, which can extract entities, relationships and attributes more accurately from different types of data, thereby improving the accuracy and efficiency of extraction. (4) This invention reduces redundancy and conflict and enhances knowledge consistency and accuracy by aligning entities, merging relationships and resolving conflicts; (5) This invention provides a variety of storage options, such as graph database, relational database or triplet repository, to facilitate selection and deployment according to scale and application requirements; (6) This invention supports multiple reasoning methods based on rules, embedding, and graph neural networks, which can uncover hidden knowledge and enhance the utilization value of graphs. Attached Figure Description

[0025] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram illustrating the logic of the method described in an embodiment of the present invention. Detailed Implementation

[0026] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0027] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0028] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0029] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0030] Example 1: Overall Process Example This embodiment provides a method for constructing a clinical knowledge graph based on real-world data, which is performed according to the following steps: S1 Data Acquisition: Acquire real-world data from multiple sources, including at least electronic health records, medical insurance data, medical literature, medical device data, patient-reported outcomes, and public health data. To ensure the universality of data access, a unified access interface and data dictionary can be established for data from different sources, recording basic information such as field source, collection time, and collection institution.

[0031] S2 Data Preprocessing: Performs cleaning, standardization, and structuring transformation on multi-source real-world data. Cleaning includes at least removing duplicate, missing, and outlier values, and correcting data entry errors. Considering that real-world data contains sensitive patient information, this step de-identifies personally identifiable information fields to meet compliance requirements. Standardization includes at least mapping medical terms to corresponding coding databases to unify data formats. Structuring transformation includes at least converting unstructured data into semi-structured or structured data for subsequent extraction and processing.

[0032] In an optional implementation, medical terminology mapping can specifically involve mapping disease names to ICD-10 codes and drug names to ATC codes or the RxNorm database; at the same time, mapping tables are established for synonyms, abbreviations, and aliases to reduce the differences in the expression of the same medical concept in different data sources.

[0033] S3 Schema Design: Defines the entity types, relation types, and attribute types of the clinical knowledge graph. Core entity types include at least patient, disease, drug, symptom, test / examination, and treatment plan; core relation types include at least patient-disease, disease-accompanying-symptom, disease-need-test, disease-use-drug, and drug-have-side effects; and adheres to the principles of domain adaptability, scalability, and simplicity. To support subsequent reasoning and tracing, the schema can further reserve attribute bits such as "source," "time," and "confidence" for entities and relations.

[0034] S4 Knowledge Extraction: Extract entities, relationships, and attributes from preprocessed data.

[0035] Entity extraction: A combination of rule-based and machine learning methods is used. Rule-based methods can be used for initial screening and localization based on medical dictionaries and regular expressions. Machine learning methods can include traditional machine learning and deep learning. Traditional machine learning can use conditional random fields and support vector machines, while deep learning can use BERT and BioBERT pre-trained models to improve the ability to recognize complex contexts.

[0036] Relation extraction: The same rule-based approach is used in combination with machine learning methods. The rule-based approach can identify relation candidates based on trigger words and syntactic structure. The machine learning methods can include supervised learning and remote supervision. Supervised learning can transform relation extraction into a classification problem and train CNN and LSTM models using labeled data. Remote supervision can automatically label data using existing knowledge bases to expand the training samples.

[0037] Attribute extraction: Attributes can be directly extracted from structured data; for attributes in unstructured text, attributes can be extracted after entity location and associated with the corresponding entity.

[0038] S5 Knowledge Fusion: Performs entity alignment, relation fusion, and conflict resolution on the extracted results.

[0039] Entity alignment: String matching, attribute matching and machine learning methods can be used; among them, machine learning methods can treat entity alignment as a classification problem, using features such as entity name, attributes, and contextual relationships to train logistic regression or graph neural network models to improve the accuracy of cross-source entity merging.

[0040] Relationship fusion and conflict resolution: Relationships with the same name from different sources are assigned confidence levels, retaining those with high confidence. When conflicts exist, the majority opinion or validation using clinical data can be used to determine which relationship to retain. For ease of maintenance, the source set and the basis for confidence level calculations can be recorded for each relationship.

[0041] S6 Knowledge Storage: The merged knowledge is written to a knowledge base. The knowledge base can be stored in a graph database, relational database, or triplet repository; for example, graph databases may include Neo4j, OrientDB, and JanusGraph; relational databases may include MySQL and PostgreSQL; and triplet repositories may include Virtuoso and Apache Jena. To meet auditing and traceability needs, the storage layer can simultaneously store metadata fields such as entity / relationship source identifiers, timestamps, and confidence levels.

[0042] S7 Knowledge Reasoning: Performs reasoning based on stored knowledge to uncover implicit knowledge. Reasoning methods may include: Rule-based reasoning: Automatically inferring implicit relationships through a rule engine; Embedded reasoning: Maps entities and relations to a low-dimensional vector space, and infers implicit relations through vector operations. TransE and DistMult models can be used. Reasoning based on graph neural networks: Using GNNs to capture the neighbor relationships and structural features of entities in the graph for reasoning.

[0043] S8 Application and Iteration: The constructed clinical knowledge graph will be applied to clinical decision support, pharmacovigilance, and disease research; and continuously iterated and optimized based on new data and user feedback. To adapt to the continuous growth of real-world data, an incremental update mechanism can be adopted in one optional implementation: when new or changed data is detected, preprocessing, extraction, fusion, and storage are performed only on related entities, relationships, and attributes, and the graph before and after the update is version-marked and traceably archived; at the same time, the corrective feedback from clinicians on the extraction / fusion / inference results is converted into training samples or rule revision items to update the extraction model and conflict handling rules.

[0044] Example 2: Demonstrating the construction process using diabetes-related knowledge as an example In this embodiment, electronic health records and medical insurance settlement records of a medical institution are used as the main structured data source, and medical literature abstracts are used as the unstructured data source, while some public health data are also accessed. First, cleaning (duplicate removal, missing / abnormal handling, and data entry correction) and desensitization are performed according to Embodiment 1; then, disease terms such as "type 2 diabetes" are mapped to corresponding ICD-10 codes, and drug terms such as "metformin" are mapped to ATC codes or RxNorm identifiers, thus unifying the representation of different data sources.

[0045] Subsequently, entities and relationships are defined according to the schema: patients, diseases, drugs, symptoms, tests / examinations, and treatment plans are used as core entities, and core relationships such as patient-disease-symptom, disease-need-test, disease-use-drug, and drug-have-side effects are constructed. For text data such as literature abstracts, candidate entities are first identified using a medical dictionary and regular expressions, and then the boundaries and types are confirmed using a BERT / BioBERT model; relation extraction simultaneously uses trigger word rules and supervised learning classification models (CNN / LSTM) to output relations and confidence scores.

[0046] For duplicate disease and drug entities from different sources, string similarity and encoding consistency are used for merging. For entities with partial aliases or homonyms, logistic regression or graph neural network alignment models are used to determine their identity, combining attribute and contextual relationship features. For conflicting relationships such as drug side effects, a comprehensive confidence level is assigned based on source reliability and model confidence, prioritizing the retention of high-confidence relationships. If necessary, the final retained relationship is determined by the conclusions of the majority of sources or by combining clinical data for verification.

[0047] Finally, the graph is written into a knowledge base (e.g., stored using a graph database), and implicit relationships such as "detection anomaly - possible disease" are inferred based on a rule engine, or embedded reasoning is performed using TransE / DistMult to fill in any possible missing "disease - medication" associations, thereby providing candidate suggestions for clinical decision support and pharmacovigilance; when new follow-up data or new literature evidence arrives, incremental updates are performed and version records are generated to ensure traceability.

[0048] Example 3: System Implementation of Modular Implementation Corresponding to the Method A system for constructing a clinical knowledge graph based on real-world data may include: a data access module, a preprocessing module, a schema design module, a knowledge extraction module, a knowledge fusion module, a knowledge storage module, a knowledge reasoning module, and an application iteration module.

[0049] The data access module is used to access electronic health records, medical insurance data, medical literature, medical device data, patient reported outcomes, and public health data. The preprocessing module is used to perform cleaning, desensitization, medical terminology mapping, and structured transformation; The schema design module is used to define core entities, relationships, and attributes; The knowledge extraction module is used to extract entities / relationships / attributes using rule-based methods combined with machine learning methods; The knowledge fusion module is used for entity alignment, confidence fusion, and conflict resolution; The knowledge storage module is used to write to graph databases / relational databases / triple databases; The knowledge reasoning module is used for rule-based reasoning / embedded reasoning / GNN reasoning; The application iteration module is used to apply the atlas to clinical decision support, pharmacovigilance, and disease research, and is continuously updated based on new data and user feedback. This modular division of labor corresponds one-to-one with the overall steps, facilitating engineering deployment and expansion.

[0050] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for constructing a clinical knowledge graph based on real-world data, characterized in that, include: S1. Acquire multi-source real-world data, wherein the multi-source real-world data includes at least electronic health records, medical insurance data, medical literature, medical device data, patient-reported outcomes, and public health data; S2. Perform preprocessing on the multi-source real-world data. The preprocessing includes deduplication, missing and anomaly handling, input error correction, desensitization of personal identity information fields, mapping medical terms to a preset coding database in a unified format, and converting unstructured data into semi-structured or structured data. S3. Perform schema design, defining entity types, relation types, and attribute types. The entity types must include at least patients, diseases, drugs, symptoms, laboratory tests, and treatment plans. S4. Under the schema constraints, entities, relations and attributes are extracted from the preprocessed data. Entity extraction and relation extraction adopt a combination of rule-based methods and machine learning methods. S5. Perform knowledge fusion on the extracted knowledge, including entity alignment and confidence-based relation fusion and conflict resolution; S6. Store the fused knowledge in a preset knowledge base and perform at least one knowledge reasoning to uncover implicit knowledge; S7. Apply the clinical knowledge graph to clinical decision support, pharmacovigilance, or disease research, and update and iterate the knowledge graph when receiving new data and / or user feedback.

2. The method according to claim 1, characterized in that, The step S2 of mapping medical terms to a preset coding database includes mapping disease names to ICD-10 codes and drug names to ATC codes or the RxNorm database.

3. The method according to claim 1, characterized in that, The rule-based method for entity extraction in step S4 is based on a medical dictionary and regular expressions; the machine learning method for entity extraction includes traditional machine learning and deep learning, wherein traditional machine learning includes at least conditional random fields and support vector machines, and deep learning includes at least BERT or BioBERT pre-trained models.

4. The method according to claim 1, characterized in that, The rule-based method for relation extraction in step S4 is based on trigger words and syntactic structure; the machine learning method for relation extraction includes supervised learning and remote supervision, wherein supervised learning transforms relation extraction into a classification problem and trains a CNN or LSTM model using labeled data, and remote supervision automatically labels training data using existing knowledge bases.

5. The method according to claim 1, characterized in that, The confidence-based relationship fusion in step S5 includes: assigning confidence levels to same-name relationships from different data sources and retaining high-confidence relationships; the conflict resolution includes determining the relationships to be retained by using the conclusions of the majority of sources or by combining clinical data verification.

6. The method according to claim 1, characterized in that, The knowledge reasoning described in step S6 includes at least one of the following reasoning methods: rule-based reasoning, embedding-based reasoning, or graph neural network-based reasoning; wherein embedding-based reasoning uses the TransE or DistMult model, rule-based reasoning uses a rule engine to perform implicit relation reasoning, and graph neural network-based reasoning uses a graph neural network to capture neighbor relationships and structural features for reasoning.

7. The method according to claim 1, characterized in that, In step S4 or step S5, at least a metadata set is generated and bound for each piece of extracted or fused knowledge. The metadata set includes at least the source identifier, generation timestamp, and confidence level. The confidence level is calculated by weighting the confidence score of the knowledge extraction output and the data quality results of the preprocessing stage, and is used for relation fusion in step S5 or update iteration decision in step S7.

8. The method according to claim 1, characterized in that, The update iteration described in step S7 includes an incremental update mechanism: when new or changed data is detected, preprocessing, extraction, fusion, and storage are performed only on entities, relationships, and attributes related to the new or changed data; The updated and unupdated map versions are identified and archived in a traceable manner.

9. The method according to claim 1, characterized in that, The continuous iterative optimization based on user feedback in step S7 includes: receiving manual correction information on knowledge extraction results, relation conflict resolution results, or reasoning conclusions; converting the correction information into training samples or rule revision items; and updating the entity extraction model, relation extraction model, and / or conflict resolution rules accordingly.

10. A system for constructing a clinical knowledge graph based on real-world data, characterized in that, It includes a data access module, a preprocessing module, a schema design module, a knowledge extraction module, a knowledge fusion module, a knowledge storage module, a knowledge reasoning module, and an application iteration module; The data access module is used to acquire multi-source real-world data. The preprocessing module is used to perform deduplication, missing and anomaly handling, data entry error correction, desensitization, terminology mapping, and structured transformation on the multi-source real-world data; The schema design module is used to define entity types, relationship types, and attribute types; The knowledge extraction module is used to extract entities, relations, and attributes under schema constraints by combining rule-based methods with machine learning methods. The knowledge fusion module is used to perform entity alignment and confidence-based relation fusion and conflict resolution. The knowledge storage module is used to store the merged knowledge in a preset knowledge base; The knowledge reasoning module is used to perform at least one knowledge reasoning based on the knowledge base to uncover implicit knowledge; The application iteration module is used to apply clinical knowledge graphs to clinical decision support, pharmacovigilance, or disease research, and to perform update iterations when new data and / or user feedback are received.

Images