Medical term atlas reasoning construction method and device, equipment and storage medium

Abstract

The invention provides a medical term atlas reasoning construction method and device, equipment and a storage medium, and the method comprises the steps: carrying out the attribute pre-extraction of an input original medical term through employing a large language model based on an assembly rule system, and outputting an attribute group of the original medical term; the attribute group comprises each medical entity type of the original medical term, and an attribute category and an attribute value corresponding to each medical entity type; based on each medical entity type of the original medical terms and the attribute category and the attribute value corresponding to each medical entity type, verifying and correcting the attribute group; and based on the verified and corrected attribute group, performing reasoning retrieval in the medical term atlas, and / or constructing or updating the medical term atlas. Therefore, lossless analysis and standardization of medical term deep semantics are realized, information loss in a traditional normalization method is reduced, and term processing accuracy and medical term graph expandability are improved.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to a method, apparatus, device, and storage medium for constructing a medical terminology atlas. Background Technology

[0002] In the fields of healthcare, health insurance, and pharmaceuticals, a large amount of multi-source, heterogeneous data exists. Medical terminology, as a key carrier of information transmission and knowledge expression, plays a crucial pivotal role in the integration of these three areas. However, there is a serious problem of inconsistent expression of medical terminology across these three fields, causing significant difficulties in the collection, organization, and analysis of medical data, and severely hindering the effective sharing of medical information and the in-depth exploration of medical knowledge. To address this challenge, normalization has become the mainstream approach in current medical terminology processing.

[0003] Normalization aims to unify various medical terms into a standardized form, such as the Unified Medical Language System (UMLS), the Medical Subject Headings (MeSH), and the Systematic Medical Nomenclature – Clinical Terminology (SNOMED CT). The process involves two main steps: First, using rule-based regular expression matching, statistical methods based on manually labeled data, deep learning methods, and vectorization and generative methods from pre-trained models such as large models, to generate preliminary candidate terms or directly obtain normalization results. Second, using rule-based methods, online search enhancement, or deep learning algorithms distinct from the first step, the candidate terms are ranked.

[0004] All of the above normalization methods require a complete and comprehensive mapping standard to be found beforehand. If the input medical concept does not have a completely matching term in the standard, an approximate concept will be found from the standard based on similarity.

[0005] This method cannot guarantee the lossless nature of the original input concepts' connotations. When precise normalization is not possible, the input medical terms must lose some connotations and be mapped to a broader medical concept. The comprehensiveness of the mapping standard directly affects the usability of the mapping results. When the mapping standard is not comprehensive enough, it lacks good extensibility, and all extended terms require manual verification. In using non-precisely normalized terms, errors may occur in conditional judgments due to information loss. For example, if a drug is explicitly specified for use only in patients with advanced triple-negative breast cancer who have not undergone surgery, and the closest concept in the mapping standard (e.g., ICD-10) is "breast cancer," mapping the indication upwards to "breast cancer" will incorrectly expand the drug's indications. Summary of the Invention

[0006] This invention provides a method, apparatus, device, and storage medium for constructing a medical terminology atlas, in order to solve the aforementioned problems in the prior art.

[0007] This invention provides a method for constructing a medical terminology atlas, comprising the following steps: Based on the combination rule system, a large-scale language model is used to pre-extract attributes from the input original medical terms and output the attribute group of the original medical terms; the attribute group includes each medical entity type of the original medical terms, as well as the attribute category and attribute value corresponding to each medical entity type; Based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the attribute group is verified and corrected; Based on the verified and corrected attribute groups, perform reasoning retrieval in the medical terminology graph; and / or, based on the verified and corrected attribute groups, construct or update the medical terminology graph; The combination rule system classifies various medical entity types into simple entities, separable combination entities, and indivisible combination entities. The separable combination entities are defined with attribute categories and value ranges, and the value range of each attribute category is anchored to the defined medical concept and the sub-concepts of the defined medical concept. In the medical terminology graph, the nodes include concept nodes and terminology nodes; each concept node represents a medical concept, each terminology node represents a string with medical meaning, each concept node is connected to at least one terminology node, and the concept nodes are connected to each other through parent-child relationships and attribute relationships.

[0008] According to a medical terminology graph reasoning construction method provided by the present invention, the attribute group is validated based on each medical entity type of the original medical terminology and the attribute category and attribute value corresponding to each medical entity type, including: Based on each medical entity type of the original medical terminology, the attribute category and attribute value corresponding to each medical entity type, and the verification rules, the attribute group is verified. The verification rules include at least one of the following: parent concept type verification rules, entity-attribute matching rules, attribute-value range matching rules, and logical consistency rules; The parent concept type validation rules include: If the medical entity type is the parent concept and the medical entity type is a detachable and combinable entity, the medical entity type validation fails. The entity-attribute matching rules include: If the medical entity type and the attribute category do not belong to a pre-defined association relationship, the verification of the medical entity type and the attribute category fails. The attribute-range matching rules include: If the attribute value does not belong to the value range of the attribute category, the attribute category and the attribute value verification fails. The logical consistency rules include: If the relationship between any two of the medical entity type, the attribute category, and the attribute value does not satisfy medical logic, the verification of any two items will fail.

[0009] According to a medical terminology graph reasoning construction method provided by the present invention, based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the attribute group is modified, including: The attribute groups that failed the validation will be pushed to the medical terminology expert platform for manual correction. The manually modified attribute groups are added to the fine-tuning dataset, and the large language model is incrementally trained based on the fine-tuning dataset.

[0010] According to the present invention, a method for constructing a medical terminology graph for reasoning, based on a verified and corrected set of attribute groups, performs reasoning retrieval in a medical terminology graph, including: Based on the method of comparing the attribute groups after verification and correction and the parent-child relationship, the parent concept of the original medical term is searched in a breadth-first manner, starting from the root node of the medical terminology graph. If there exists any concept that is the parent concept of the original medical term, and none of the child concepts of the original medical term are the parent concepts of the original medical term, then the arbitrary concept shall be regarded as the nearest parent concept of the original medical term. When the nearest parent concept of the original medical term is a child concept of the original medical term, the nearest parent concept is used as the accurate normalized result obtained by retrieving the original medical term.

[0011] According to a medical terminology atlas reasoning construction method provided by the present invention, the method of comparing parent-child relationships includes: If all parent concepts of the first concept have a parent-child relationship or are the same as a parent concept of the second concept, and all attribute groups of the first concept have a parent-child relationship with a certain attribute group of the second concept, then the first concept is the parent concept of the second concept.

[0012] According to the present invention, a method for constructing a medical terminology graph based on a verified and corrected set of attribute data is provided to construct or update a medical terminology graph, including: When the medical entity type of the original medical term is a simple entity or an indivisible and combinable entity, the original medical term is normalized to the concept nodes in the medical terminology graph by means of string similarity or retrieval and rearrangement methods of large language models. If the medical entity type of the original medical term is a separable and combinable entity, and the original medical term does not have a precise normalization result found in the medical terminology graph, a new concept node is created based on the nearest parent concept of the original medical term and the verified and corrected attribute group.

[0013] The present invention also provides a medical terminology atlas reasoning and construction device, comprising the following modules: The attribute extraction module is used to pre-extract attributes from the input raw medical terms based on the combination rule system and a large language model, and output the attribute group of the raw medical terms; the attribute group includes each medical entity type of the raw medical terms, as well as the attribute category and attribute value corresponding to each medical entity type; The verification and correction module is used to verify and correct the attribute group based on each medical entity type of the original medical term and the attribute category and attribute value corresponding to each medical entity type. The reasoning construction module is used to perform reasoning retrieval in the medical terminology graph based on the verified and corrected attribute groups; and / or to construct or update the medical terminology graph based on the verified and corrected attribute groups. The combination rule system classifies various medical entity types into simple entities, separable combination entities, and indivisible combination entities. The separable combination entities are defined with attribute categories and value ranges, and the value range of each attribute category is anchored to the defined medical concept and the sub-concepts of the defined medical concept. In the medical terminology graph, the nodes include concept nodes and terminology nodes; each concept node represents a medical concept, each terminology node represents a string with medical meaning, each concept node is connected to at least one terminology node, and the concept nodes are connected to each other through parent-child relationships and attribute relationships.

[0014] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the medical terminology atlas reasoning construction method as described above.

[0015] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the medical terminology atlas reasoning construction method as described above.

[0016] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the medical terminology graph reasoning construction method as described above.

[0017] The medical terminology graph reasoning construction method, apparatus, device, and storage medium provided by this invention construct a set of combination rules that classify medical entities into simple entities, separable and combinable entities, and indivisible and combinable entities. It also utilizes a large-scale language model to pre-extract attributes from the original medical terms, and after verification and correction, forms standard format attribute groups. These groups are then used for reasoning retrieval and construction updates in the medical terminology graph. This achieves lossless parsing and standardization of the deep semantics of complex medical terms, significantly reduces information loss in traditional normalization methods, and improves the accuracy of terminology processing and the scalability of the medical terminology graph. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating the medical terminology atlas reasoning construction method provided by the present invention.

[0020] Figure 2 This is a schematic diagram of the medical entity type classification provided by the present invention.

[0021] Figure 3 This is a terminology graph reasoning flowchart provided by the present invention.

[0022] Figure 4 This is a schematic diagram illustrating an example of determining the parent-child relationship of entities provided by the present invention.

[0023] Figure 5 This is a schematic diagram of the medical terminology atlas reasoning and construction device provided by the present invention.

[0024] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0026] Figure 1 This is a flowchart illustrating the medical terminology atlas reasoning construction method provided by the present invention, such as... Figure 1 As shown, the method includes the following steps: Step 100: Based on the combination rule system, a large-scale language model is used to pre-extract attributes from the input original medical terms and output the attribute group of the original medical terms. The attribute group includes each medical entity type of the original medical terms, as well as the attribute category and attribute value corresponding to each medical entity type.

[0027] Step 101: Based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, verify and correct the attribute group.

[0028] Step 102: Based on the verified and corrected attribute groups, perform reasoning retrieval in the medical terminology atlas; and / or, based on the verified and corrected attribute groups, construct or update the medical terminology atlas.

[0029] Among them, the combination rule system classifies various medical entity types into simple entities, separable combination entities, and indivisible combination entities; separable combination entities are defined with attribute categories and value ranges, and the value range of each attribute category is anchored to the defined medical concept and the sub-concepts of the defined medical concept; In a medical terminology graph, nodes include concept nodes and terminology nodes. Each concept node represents a medical concept, and each terminology node represents a string with medical meaning. Each concept node is connected to at least one terminology node, and concept nodes are connected to each other through parent-child relationships and attribute relationships.

[0030] Specifically, in this embodiment of the invention, a combination rule system is first determined. The combination rule system predefines a classification framework for medical entity types, distinguishing medical entity types into simple entities, divisible combination entities, and indivisible combination entities. It also clarifies the attribute categories and the value range corresponding to each attribute category for divisible combination entities. This value range is anchored to the defined medical concepts and their sub-concepts.

[0031] In some implementations, Figure 2 This is a schematic diagram of the medical entity type classification provided by the present invention, such as... Figure 2As shown, 11 types of medical entity types can be defined based on the terminology characteristics of the medical, health insurance, and pharmaceutical fields. Among them, medical entities in the types of body structure, event, organism, physical force, qualifier, matter, and physical object are simple entities, characterized by clear semantics and fixed expression; medical entities in the types of disease / symptom, observable indicator, operation, and drug / preparation are composite entities.

[0032] Composite entities can be categorized into two types: indivisible and indivisible entities. Indivisible entities are typically fixed morpheme phrases (such as "jaundice" and "dysentery") or multi-layered modifier-head structures in the medical field (such as "diabetes" and "hypertension"). Indivisible entities are simple modifier-head structures centered around a core attribute (such as "lung cancer" and "type 2 diabetes"). For simple entities and indivisible composite entities, their phrase stability is extremely high, and there are very few synonyms with completely identical meanings. They can be directly normalized using similarity matching algorithms or large-scale model retrieval and reordering techniques. However, for indivisible composite entities, attributes need to be extracted first, and then normalized and assembled based on the attributes.

[0033] For detachable assembly entities, specific attribute categories and value ranges are defined for each concept.

[0034] In some implementations, regarding the definition of attribute categories, core attribute categories covering eight semantic dimensions are identified around the semantic composition dimensions of the separable and combinable entity. These dimensions include temporal association (e.g., after / before / during), spatial association (e.g., location of discovery), feature description (e.g., related morphology, clinical manifestations, frequency of onset), evaluation dimension (e.g., evaluation object / outcome), developmental state (e.g., disease stage, clinical course, pathological process), causal association (e.g., pathogenic factors, causes), constraints (e.g., treatment conditions, organismal state, physiological and age stage at the time of occurrence), and conceptual affiliation (e.g., indivisible parent concept), ensuring that all semantic information of the entity can be completely decomposed and expressed.

[0035] Each attribute category has a clearly defined value range. This value range limitation follows a specific anchoring logic, which anchors the value range of each attribute to a defined medical entity category and its sub-concepts, forming a strong association constraint between the attribute and the entity category. For example, "related morphology" is bound to the sub-concept "abnormal structure" under "body structure", and "after" is bound to the sub-concept "event / operation / disease / symptom", avoiding the problem of invalid attribute values ​​caused by value range generalization.

[0036] Based on the combination rule system, a large language model can be used to pre-extract attributes from the input original medical terms. That is, by utilizing the powerful natural language understanding and generation capabilities of the large language model, the original medical terms can be preliminarily structured and the attribute groups of the original medical terms can be output.

[0037] The original medical terminology can be a relatively complex medical term, such as "stage III PD-L1 positive non-small cell lung cancer after targeted therapy".

[0038] In some implementations, carefully designed cue word engineering ensures that large language models can accurately understand task requirements and extract attributes from the demonstrated medical entities according to specific output format constraints. For example, large language models can organize their output according to a standard structure of "entity type: {attribute category: attribute value, ...}, ...".

[0039] The attribute group contains each medical entity type identified in the original medical terminology, as well as the attribute category and attribute value corresponding to each medical entity type.

[0040] After obtaining the attribute groups of the original medical terms, each extracted medical entity type and its corresponding attribute category and attribute value can be comprehensively verified according to the combination rule system and other logical rules. During the verification process, any data that fails verification can be automatically marked, and medical experts with professional knowledge can then adjust the marked content appropriately to ensure it conforms to medical common sense and logic.

[0041] Based on the validated and corrected attribute groups, a query can be performed in the medical knowledge graph to find concept nodes that semantically match the original input medical terms. Alternatively, the original medical terms can be integrated into the medical knowledge graph as new medical concepts. These two functions can be executed individually or in combination, depending on actual needs.

[0042] In a medical terminology graph, nodes can be categorized into two basic types: concept nodes and terminology nodes. Concept nodes represent abstract medical concepts, each possessing a unique identifier, essentially a digital ID for that concept within the graph. Terminology nodes, on the other hand, represent specific medical vocabulary strings, the concrete expression of the concept in natural language. Each concept node establishes connections with one or more terminology nodes. This design enables unified management of the same medical concept across different expressions; for example, two different terminology nodes, "lung cancer" and "lung cancer," can both point to the same concept node representing the medical concept of lung cancer.

[0043] Any two concept nodes can be connected through parent-child relationships and attribute relationships. Parent-child relationships express the inclusion and hierarchical structure between concepts, reflecting their hierarchical relationship. Attribute relationships are specifically designed for concepts that can be split and combined in the combination rule system. These concepts establish labeled attribute relationship edges with the concept nodes corresponding to their attribute values. These edges explicitly label the attribute category and grouping information, thus forming a rich semantic network structure.

[0044] When performing reasoning retrieval based on the verified and corrected attribute groups, possible matching concepts can be found in the existing atlas structure based on the relevant information provided in the attribute groups. By comparing the semantic features of the attribute groups with the existing concepts in the medical terminology atlas, it can be determined whether there is a perfect match or the closest parent concept relationship can be identified.

[0045] When updating the graph based on the validated and corrected attribute groups, for medical concepts that need to be added after validation, a new concept node can be created and assigned a unique identifier. Simultaneously, a connection can be established between this node and related term nodes. Furthermore, based on the semantic relationships between concepts, a parent-child relationship can be established between the new concept node and its parent concept node, and labeled attribute relationship edges can be created between the new concept node and the concept nodes corresponding to each attribute value, thus achieving dynamic expansion and improvement of the graph structure.

[0046] The medical terminology graph reasoning construction method provided by this invention constructs a set of combination rules that classify medical entities into simple entities, separable and combinable entities, and indivisible and combinable entities. It uses a large-scale language model to pre-extract attributes from the original medical terms, and after verification and correction, forms standard format attribute groups. Then, it performs reasoning retrieval and construction updates in the medical terminology graph, thereby achieving lossless parsing and standardization of the deep semantics of complex medical terms. This significantly reduces the information loss in traditional normalization methods and improves the accuracy of terminology processing and the scalability of the medical terminology graph.

[0047] According to a medical terminology graph reasoning construction method provided by the present invention, based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the attribute group is validated, including: The attribute group is validated based on each medical entity type of the original medical terminology, the attribute category and attribute value corresponding to each medical entity type, and the validation rules.

[0048] Specifically, during the validation of attribute groups, predefined validation rules can be used to ensure the correctness of the extracted attribute groups in terms of structure, semantics, and logic.

[0049] In this embodiment of the invention, the verification rules may include at least one of the following: parent concept type verification rules, entity-attribute matching rules, attribute-value range matching rules, and logical consistency rules.

[0050] The parent concept type validation rules include: When the medical entity type is the parent concept and the medical entity type is a detachable and combinable entity, the medical entity type validation fails.

[0051] This rule requires that the medical entity type appearing as the parent concept must be a simple concept or a concept of an indivisible, combinable entity. When the detected parent concept is an indivisible, combinable entity type, the validation fails, and the parent concept needs to be further broken down into more basic semantic units.

[0052] Entity-attribute matching rules include: If the medical entity type and attribute category do not belong to a pre-defined association, the medical entity type and attribute category validation fails.

[0053] This rule checks whether each medical entity type and its associated attribute categories conform to a predefined correspondence. For example, it verifies whether the "Disease / Symptom (Separable)" entity type is associated only with the 18 allowed attribute categories, and not incorrectly associated with irrelevant attributes such as "Drug Ingredients" or "Operating Instruments". Any entity type and attribute category combination that does not conform to the preset association relationship will be marked as a validation failure.

[0054] Attribute-value range matching rules include: If the attribute value does not belong to the value range of the attribute category, the attribute category and attribute value validation will fail.

[0055] This rule ensures that the attribute values ​​corresponding to each attribute category are limited to the preset value range. For example, it verifies whether the value of the "disease staging" attribute is limited to standard values ​​such as "stage I, stage II, stage III, stage IV, early stage, late stage", and excludes meaningless values ​​such as "high" and "low" that do not conform to the value range specification, thereby ensuring the validity and standardization of attribute values.

[0056] Logical consistency rules include: If the relationship between any two of the medical entity type, attribute category, and attribute value does not meet the medical logic, the validation of any two items will fail.

[0057] This rule verifies the logical rationality between entity types, attribute categories, and attribute values ​​from a medical perspective. The system detects and flags combinations with conflicting medical logic, such as contradictory situations like "treatment condition: no surgery" and "after: surgery," or attribute combinations that do not conform to common clinical sense, such as "disease stage: late stage" and "clinical course: acute stage," ensuring the rationality and usability of the final attribute group in medical practice.

[0058] According to a medical terminology graph reasoning construction method provided by the present invention, based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the attribute group is modified, including: The attribute groups that failed the validation will be pushed to the medical terminology expert platform for manual correction. The manually modified attribute groups were added to the fine-tuning dataset, and the large language model was incrementally trained based on the fine-tuning dataset.

[0059] Specifically, upon discovering attribute groups that do not conform to the rules, the failed attribute groups can be pushed to the medical terminology expert platform for manual correction. For example, when encountering an attribute combination such as "acute advanced lung cancer" that does not conform to clinical common sense, the expert will correct it to a reasonable expression such as "chronic advanced lung cancer" based on their professional knowledge, thereby completing the standardization process of the attribute group by directly adjusting the attribute category or attribute value.

[0060] In some implementations, during the manual correction phase, the medical terminology expert at the receiving end should have more than five years of clinical or medical informatics experience to ensure that they possess sufficient professional judgment.

[0061] In some implementations, experts can review violation reports generated during the verification process. These reports can identify the specific rule types and locations violated by the attribute group and provide corrective suggestions based on medical common sense for expert reference.

[0062] After manual correction, the corrected attribute sets can be added to the fine-tuning dataset. By collecting high-quality data validated by experts, reliable training materials can be provided for subsequent optimization of large-scale language models.

[0063] Large language models can be incrementally trained according to predetermined strategies based on an ever-expanding fine-tuning dataset. For example, an incremental fine-tuning round can be performed every 10,000 corrected data points. In this way, large language models can gradually learn the patterns and rules of expert corrections, thereby continuously optimizing the accuracy of the attribute extraction process and gradually mastering more precise medical semantic understanding capabilities.

[0064] Through this feedback loop, the proportion of manual corrections can be gradually reduced. In the initial stage, since the model is not yet fully trained, the manual correction rate is about 30%; as the model undergoes multiple rounds of incremental training and the accuracy of autonomous processing improves, the proportion of manual corrections can be reduced to below 5% after three rounds of training.

[0065] According to the present invention, a method for constructing a medical terminology graph for reasoning, based on a verified and corrected set of attribute groups, performs reasoning retrieval in a medical terminology graph, including: Based on the method of comparing the attribute groups after verification and correction and the parent-child relationship, the parent concept of the original medical term is searched in a breadth-first manner, starting from the root node of the medical terminology graph. If there exists an arbitrary concept that is the parent concept of the original medical term, and none of the child concepts of the arbitrary concept in the medical terminology map are the parent concepts of the original medical term, then the arbitrary concept is taken as the nearest parent concept of the original medical term. When the nearest parent concept of the original medical term is a child concept of the original medical term, the nearest parent concept is used as the accurate normalized result obtained from retrieving the original medical term.

[0066] Specifically, in the semantic retrieval process of the medical terminology graph, the parent concepts of the original medical terms can be searched in a breadth-first manner, starting from the root node of the medical terminology graph, based on the verified and corrected attribute groups and the comparison of parent-child relationships.

[0067] In some implementations, efficient pruning mechanisms can be applied to optimize search efficiency. That is, when a concept node is not the parent concept of the original medical term, further exploration of all child branches of that node can be immediately terminated. This pruning strategy significantly reduces unnecessary computational overhead, thereby enabling rapid location of relevant regions in the medical terminology graph.

[0068] The search process requires finding the nearest parent concept of the original medical term. A concept node that serves as the nearest parent concept of the original medical term should meet two key requirements: first, the concept node must be the parent concept of the original medical term; second, none of the child concepts of this concept node in the graph are parent concepts of the original medical term. In other words, the nearest parent concept of the original medical term should be the closest direct superior node in the concept hierarchy to the original medical term.

[0069] After determining the nearest parent concept, it is necessary to further determine whether the nearest parent concept is the precise normalization result of the original medical term. This step is achieved through bidirectional semantic equivalence verification. If the nearest parent concept can also be a child concept of the original medical term, then this bidirectional parent-child relationship verification ensures that the two concepts are completely semantically equivalent, thereby accurately mapping the input original medical term to the existing concept nodes in the medical terminology graph.

[0070] According to the medical terminology atlas reasoning construction method provided by the present invention, the parent-child relationship comparison method includes: If all parent concepts of the first concept have a parent-child relationship or are the same as a parent concept of the second concept, and all attribute groups of the first concept have a parent-child relationship with a certain attribute group of the second concept, then the first concept is the parent concept of the second concept.

[0071] Specifically, before starting the comparison of parent-child relationships, both comparison concepts need to be converted into standard combination structure representations.

[0072] For concepts already existing in the medical terminology graph, their parent concept list and attribute groups can be directly extracted from the node relationships recorded in the graph to form a standard format. For newly input original medical terms, the attribute groups obtained by entity category judgment and attribute extraction through the large model are the standard structure of the original medical terms, ensuring that the representations of the two concepts are comparable.

[0073] First, a comparison of the parent concept sets can be performed. During the comparison, each parent concept of the first concept must have at least one corresponding parent concept in the parent concept set of the second concept, and the former and the latter must have a parent-child relationship or be completely identical. If no matching parent concept is found, the comparison is immediately terminated, and the parent-child relationship between the first and second concepts is determined to be invalid.

[0074] After the comparison of the parent concept set is successful, the comparative analysis of attribute groups continues. Each attribute group is a semantic unit composed of multiple key-value pairs. During the comparison, it is necessary to ensure that each attribute group of the first concept can be found in the attribute group set of the second concept. The criterion for determining corresponding attribute groups is that each attribute key-value pair of the former satisfies a parent-child relationship with some attribute key-value pair of the latter. That is, there is a parent-child relationship between the attribute categories and the attribute values. Only in this way can it be determined that there is an inclusion relationship between the two attribute groups, that is, the first concept is the parent concept of the second concept.

[0075] In some implementations, the parent-child relationship between two concepts can be determined by comparing attributes using the following formula.

[0076] (i) Determine whether A is the parent concept of B (A≥B). First, organize A and B into standard formats.

[0077] A has m parent concepts in its standard format. The standard format of B has n parent concepts. ; A's standard format has p groups of attributes. The standard format of B has q sets of attributes. .

[0078] (ii) For each Determine whether it is in In, there exists ,satisfy If this condition is true, proceed to the next step to continue reasoning; if this step is false, then A≥B is not true.

[0079] (iii) For each group Determine whether it is in In, there exists ,satisfy If this condition is true, then A≥B is true; otherwise, if even one condition is false, then A≥B is false.

[0080] In some implementations, the parent concept is a simple concept or an indivisible combination concept, which can be used to directly determine whether there is a direct or indirect parent-child relationship in the graph.

[0081] In the process of determining the parent-child relationship of an attribute group, each It is a list of dictionaries, that is , .

[0082] judge It is necessary to determine each pair All can be in Find a pair , making , making .

[0083] judge , equivalent to judgment .

[0084] According to the present invention, a method for constructing a medical terminology graph based on a verified and corrected set of attribute data is provided to construct or update a medical terminology graph, including: When the medical entity type of the original medical term is a simple entity or an indivisible and combinable entity, the original medical term is normalized to the concept nodes in the medical terminology graph by means of string similarity or retrieval and ranking methods of large language models. If the medical entity type of the original medical term is a separable and combinable entity, and the original medical term does not have a precise normalization result found in the medical terminology graph, a new concept node is created based on the nearest parent concept of the original medical term and the verified and corrected attribute group.

[0085] Specifically, for medical terminology types with stable structures and fixed semantics, including simple entities such as body structures and events, as well as entities that have been solidified into fixed phrases and cannot be separated, since these terms have few synonym variants and relatively uniform expressions, string similarity matching algorithms or retrieval and reordering methods using large language models can be used to directly find the most matching corresponding item among the existing concept nodes in the medical terminology graph, thereby achieving rapid concept mapping.

[0086] When the medical entity type of the original medical term is a divisible and combinable entity, such entities have a decomposable semantic structure. First, a complete reasoning retrieval process needs to be performed. By performing hierarchical search and parent-child relationship comparison in the medical terminology graph, it is necessary to find whether there is a precise normalized result that is completely consistent with the semantics of the input original medical term.

[0087] If a comprehensive search of the graph confirms that no precise normalization result exists, the original medical term can be incorporated into the knowledge system according to application requirements. A unique concept identifier can be assigned to the original medical term, and a corresponding concept node can be created.

[0088] At this point, the nearest parent concept information of the original medical term in the knowledge graph has been determined through the reasoning retrieval process, and the attribute group data has been fully verified and corrected. Therefore, based on the determined nearest parent concept, a parent-child relationship can be established between the new concept node and its nearest parent concept node, ensuring that the new concept is correctly placed in the appropriate position in the knowledge hierarchy. Simultaneously, based on the verified and corrected attribute group, attribute relationships can be established between the new concept node and the concept nodes corresponding to each attribute value, fully preserving the semantic composition information of the original medical term and achieving the orderly expansion and improvement of the knowledge system.

[0089] The following examples from specific application scenarios further illustrate the medical terminology atlas reasoning construction method provided by the present invention.

[0090] Figure 3 This is a terminology graph reasoning flowchart provided by the present invention, such as... Figure 3 As shown, this diagram illustrates how to find the normalized result of a real medical entity in the terminology graph. When an accurate normalized result cannot be found, a new concept is constructed and inserted into the terminology graph as necessary, or its semantic attributes are preserved losslessly during use.

[0091] The following are examples from some of the processes.

[0092] (1) Example of defining the attribute categories and value ranges of the assembled entity Disease / Symptom (Divisible Entities): Core attribute categories are: After…, related morphology, location of discovery, assessment result, assessment object, during…, frequency of onset, organismal state, physiological and age stage at occurrence, pathological process, before…, causative factors, clinical course, due to, disease stage, comprehensive clinical manifestation classification, clinical manifestations, treatment conditions, and indivisible parent concepts. The value range of each attribute category is certain entities and their sub-concepts. For example, the value range of "related morphology" is "abnormal structure" under "body structure" and its sub-concepts; the value range of "after…" is "event," "operation," "disease / symptom," and their sub-concepts; the value range of "parent concept" is "disease / symptom" and its sub-concepts.

[0093] (2) Use large models to pre-extract attributes of medical entities Enter a medical entity similar to "Stage III PD-L1 positive non-small cell lung cancer after targeted therapy".

[0094] First, use the large model to output labeled entity types: clearly define the separable entity categories to which the terms belong after splitting (e.g., "disease / symptom (separable)"), providing a basis for subsequent attribute-entity matching verification. Continue using the large model to output the underlying layer: organize attribute groups: using "{attribute category: attribute value}" as the basic unit, group attributes with the same semantic dimension into one group (e.g., in the example, "assessment object: PD-L1" and "assessment result: positive" are grouped together because they both belong to the "assessment dimension"), and separate multiple groups of attributes with commas to ensure a clear structure.

[0095] The large model, based on its built-in medical terminology semantic understanding capabilities, performs multi-dimensional semantic parsing of the original terms, accurately matches key information in the terms to corresponding "attribute categories," generates attribute values, and ultimately forms attribute groups that conform to the output format. { Following: targeted therapy, Location of discovery: Lung. Relevant morphology: Non-small cell carcinoma, Disease staging: Stage III }, { Assessment subject: PD-L1, Assessment result: Positive } (3) Compare the parent-child relationship between two medical entities (concepts). Figure 4 This is a schematic diagram illustrating an example of determining the parent-child relationship of entities provided by the present invention, such as... Figure 4 As shown, by comparing attributes, it is determined that "AIDS with candidiasis" and "AIDS with oropharyngeal candidiasis" satisfy the parent-child relationship: since the latter has two more attributes than the former, and all other attributes are the same, the two satisfy the parent-child relationship.

[0096] "Immune deficiency with candidiasis" and "AIDS with oropharyngeal candidiasis" satisfy a parent-child relationship: since all attribute groups of the former can be found in some attribute group of the latter, whose attribute value is a sub-concept of the former, the two satisfy a parent-child relationship. However, the relevant morphology in "congenital immunodeficiency with candidiasis" is "congenital defect," and no sub-concept of "congenital defect" can be found in "AIDS with oropharyngeal candidiasis." Therefore, "congenital immunodeficiency with candidiasis" and "AIDS with oropharyngeal candidiasis" do not satisfy a parent-child relationship.

[0097] The medical terminology graph reasoning and construction device provided by the present invention is described below. The medical terminology graph reasoning and construction device described below and the medical terminology graph reasoning and construction method described above can be referred to in correspondence with each other.

[0098] Figure 5 This is a schematic diagram of the medical terminology atlas reasoning and construction device provided by the present invention, such as... Figure 5 As shown, the device includes the following modules: The attribute extraction module 500 is used to pre-extract attributes from the input raw medical terms based on the combination rule system and a large language model, and output the attribute group of the raw medical terms. The attribute group includes each medical entity type of the raw medical terms, as well as the attribute category and attribute value corresponding to each medical entity type. The verification and correction module 510 is used to verify and correct the attribute group based on each medical entity type of the original medical terminology, as well as the attribute category and attribute value corresponding to each medical entity type. The reasoning construction module 520 is used to perform reasoning retrieval in the medical terminology graph based on the verified and corrected attribute groups; and / or to construct or update the medical terminology graph based on the verified and corrected attribute groups. Among them, the combination rule system classifies various medical entity types into simple entities, separable combination entities, and indivisible combination entities; separable combination entities are defined with attribute categories and value ranges, and the value range of each attribute category is anchored to the defined medical concept and the sub-concepts of the defined medical concept; In a medical terminology graph, nodes include concept nodes and terminology nodes. Each concept node represents a medical concept, and each terminology node represents a string with medical meaning. Each concept node is connected to at least one terminology node, and concept nodes are connected to each other through parent-child relationships and attribute relationships.

[0099] According to a medical terminology graph reasoning and construction device provided by the present invention, based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the device verifies the attribute group, including: Based on each medical entity type of the original medical terminology, the attribute category and attribute value corresponding to each medical entity type, and the validation rules, the attribute group is validated. The validation rules include at least one of the following: parent concept type validation rules, entity-attribute matching rules, attribute-value range matching rules, and logical consistency rules; The parent concept type validation rules include: When the medical entity type is the parent concept and the medical entity type is a detachable and combinable entity, the medical entity type validation fails. Entity-attribute matching rules include: If the medical entity type and attribute category do not belong to a pre-defined association, the validation of the medical entity type and attribute category will fail. Attribute-value range matching rules include: If the attribute value does not belong to the value range of the attribute category, the attribute category and attribute value validation will fail. Logical consistency rules include: If the relationship between any two of the medical entity type, attribute category, and attribute value does not meet the medical logic, the validation of any two items will fail.

[0100] According to the medical terminology graph reasoning construction apparatus provided by the present invention, based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the attribute group is modified, including: The attribute groups that failed the validation will be pushed to the medical terminology expert platform for manual correction. The manually modified attribute groups were added to the fine-tuning dataset, and the large language model was incrementally trained based on the fine-tuning dataset.

[0101] According to the present invention, a medical terminology graph reasoning construction device performs reasoning retrieval in a medical terminology graph based on a verified and corrected attribute group, including: Based on the method of comparing the attribute groups after verification and correction and the parent-child relationship, the parent concept of the original medical term is searched in a breadth-first manner, starting from the root node of the medical terminology graph. If there exists an arbitrary concept that is the parent concept of the original medical term, and none of the child concepts of the arbitrary concept in the medical terminology map are the parent concepts of the original medical term, then the arbitrary concept is taken as the nearest parent concept of the original medical term. When the nearest parent concept of the original medical term is a child concept of the original medical term, the nearest parent concept is used as the accurate normalized result obtained from retrieving the original medical term.

[0102] According to the medical terminology atlas reasoning construction device provided by the present invention, the method of comparing parent-child relationships includes: If all parent concepts of the first concept have a parent-child relationship or are the same as a parent concept of the second concept, and all attribute groups of the first concept have a parent-child relationship with a certain attribute group of the second concept, then the first concept is the parent concept of the second concept.

[0103] According to the present invention, a medical terminology graph reasoning and construction apparatus constructs or updates a medical terminology graph based on a verified and corrected set of attribute groups, comprising: When the medical entity type of the original medical term is a simple entity or an indivisible and combinable entity, the original medical term is normalized to the concept nodes in the medical terminology graph by means of string similarity or retrieval and ranking methods of large language models. If the medical entity type of the original medical term is a separable and combinable entity, and the original medical term does not have a precise normalization result found in the medical terminology graph, a new concept node is created based on the nearest parent concept of the original medical term and the verified and corrected attribute group.

[0104] Figure 6 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 6 As shown, the electronic device may include: a processor 610, a communications interface 620, a memory 630, and a communication bus 640, wherein the processor 610, the communications interface 620, and the memory 630 communicate with each other via the communication bus 640. The processor 610 can call logical instructions in the memory 630 to execute a medical terminology graph reasoning construction method, which includes: Based on the combination rule system, a large-scale language model is used to pre-extract attributes from the input original medical terms and output the attribute group of the original medical terms. The attribute group includes each medical entity type of the original medical terms, as well as the attribute category and attribute value corresponding to each medical entity type. Based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the attribute group is validated and corrected; Based on the verified and corrected attribute groups, perform reasoning retrieval in the medical terminology graph; and / or, based on the verified and corrected attribute groups, construct or update the medical terminology graph; Among them, the combination rule system classifies various medical entity types into simple entities, separable combination entities, and indivisible combination entities; separable combination entities are defined with attribute categories and value ranges, and the value range of each attribute category is anchored to the defined medical concept and the sub-concepts of the defined medical concept; In a medical terminology graph, nodes include concept nodes and terminology nodes. Each concept node represents a medical concept, and each terminology node represents a string with medical meaning. Each concept node is connected to at least one terminology node, and concept nodes are connected to each other through parent-child relationships and attribute relationships.

[0105] Furthermore, the logical instructions in the aforementioned memory 630 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0106] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the medical terminology atlas reasoning construction method provided by the above methods, the method comprising: Based on the combination rule system, a large-scale language model is used to pre-extract attributes from the input original medical terms and output the attribute group of the original medical terms. The attribute group includes each medical entity type of the original medical terms, as well as the attribute category and attribute value corresponding to each medical entity type. Based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the attribute group is validated and corrected; Based on the verified and corrected attribute groups, perform reasoning retrieval in the medical terminology graph; and / or, based on the verified and corrected attribute groups, construct or update the medical terminology graph; Among them, the combination rule system classifies various medical entity types into simple entities, separable combination entities, and indivisible combination entities; separable combination entities are defined with attribute categories and value ranges, and the value range of each attribute category is anchored to the defined medical concept and the sub-concepts of the defined medical concept; In a medical terminology graph, nodes include concept nodes and terminology nodes. Each concept node represents a medical concept, and each terminology node represents a string with medical meaning. Each concept node is connected to at least one terminology node, and concept nodes are connected to each other through parent-child relationships and attribute relationships.

[0107] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the medical terminology atlas reasoning construction method provided by the above methods, the method comprising: Based on the combination rule system, a large-scale language model is used to pre-extract attributes from the input original medical terms and output the attribute group of the original medical terms. The attribute group includes each medical entity type of the original medical terms, as well as the attribute category and attribute value corresponding to each medical entity type. Based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the attribute group is validated and corrected; Based on the verified and corrected attribute groups, perform reasoning retrieval in the medical terminology graph; and / or, based on the verified and corrected attribute groups, construct or update the medical terminology graph; Among them, the combination rule system classifies various medical entity types into simple entities, separable combination entities, and indivisible combination entities; separable combination entities are defined with attribute categories and value ranges, and the value range of each attribute category is anchored to the defined medical concept and the sub-concepts of the defined medical concept; In a medical terminology graph, nodes include concept nodes and terminology nodes. Each concept node represents a medical concept, and each terminology node represents a string with medical meaning. Each concept node is connected to at least one terminology node, and concept nodes are connected to each other through parent-child relationships and attribute relationships.

[0108] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0109] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0110] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a medical terminology atlas, characterized in that, include: Based on the combination rule system, a large-scale language model is used to pre-extract attributes from the input original medical terms and output the attribute groups of the original medical terms. The attribute group includes each medical entity type of the original medical term, as well as the attribute category and attribute value corresponding to each medical entity type; Based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the attribute group is verified and corrected; Based on the verified and corrected attribute groups, reasoning retrieval is performed in the medical terminology atlas; And / or, based on the validated and corrected attribute groups, construct or update the medical terminology map; The combination rule system classifies various medical entity types into simple entities, separable combination entities, and indivisible combination entities. The separable combination entities are defined with attribute categories and value ranges, and the value range of each attribute category is anchored to the defined medical concept and the sub-concepts of the defined medical concept. In the medical terminology graph, the nodes include concept nodes and terminology nodes; each concept node represents a medical concept, each terminology node represents a string with medical meaning, each concept node is connected to at least one terminology node, and the concept nodes are connected to each other through parent-child relationships and attribute relationships.

2. The medical terminology atlas reasoning construction method according to claim 1, characterized in that, Based on each medical entity type of the original medical terminology, and the attribute category and attribute value corresponding to each medical entity type, the attribute group is validated, including: Based on each medical entity type of the original medical terminology, the attribute category and attribute value corresponding to each medical entity type, and the verification rules, the attribute group is verified. The verification rules include at least one of the following: parent concept type verification rules, entity-attribute matching rules, attribute-value range matching rules, and logical consistency rules; The parent concept type validation rules include: If the medical entity type is the parent concept and the medical entity type is a detachable and combinable entity, the medical entity type validation fails. The entity-attribute matching rules include: If the medical entity type and the attribute category do not belong to a pre-defined association relationship, the verification of the medical entity type and the attribute category fails. The attribute-range matching rules include: If the attribute value does not belong to the value range of the attribute category, the attribute category and the attribute value verification fails. The logical consistency rules include: If the relationship between any two of the medical entity type, the attribute category, and the attribute value does not satisfy medical logic, the verification of any two items will fail.

3. The method for constructing a medical terminology atlas according to claim 1 or 2, characterized in that, Based on each medical entity type of the original medical terminology, and the corresponding attribute category and attribute value for each medical entity type, the attribute group is modified, including: The attribute groups that failed the validation will be pushed to the medical terminology expert platform for manual correction. The manually modified attribute groups are added to the fine-tuning dataset, and the large language model is incrementally trained based on the fine-tuning dataset.

4. The medical terminology atlas reasoning construction method according to claim 1, characterized in that, Based on the validated and corrected attribute groups, inference retrieval is performed in the medical terminology atlas, including: Based on the method of comparing the attribute groups after verification and correction and the parent-child relationship, the parent concept of the original medical term is searched in a breadth-first manner, starting from the root node of the medical terminology graph. If there exists any concept that is the parent concept of the original medical term, and none of the child concepts of the original medical term are the parent concepts of the original medical term, then the arbitrary concept shall be regarded as the nearest parent concept of the original medical term. When the nearest parent concept of the original medical term is a child concept of the original medical term, the nearest parent concept is used as the accurate normalized result obtained by retrieving the original medical term.

5. The medical terminology atlas reasoning construction method according to claim 4, characterized in that, The methods for comparing father-son relationships include: If all parent concepts of the first concept have a parent-child relationship or are the same as a parent concept of the second concept, and all attribute groups of the first concept have a parent-child relationship with a certain attribute group of the second concept, then the first concept is the parent concept of the second concept.

6. The method for constructing a medical terminology atlas according to claim 4 or 5, characterized in that, Based on the validated and corrected attribute groups, construct or update the medical terminology atlas, including: When the medical entity type of the original medical term is a simple entity or an indivisible and combinable entity, the original medical term is normalized to the concept nodes in the medical terminology graph by means of string similarity or retrieval and rearrangement methods of large language models. If the medical entity type of the original medical term is a separable and combinable entity, and the original medical term does not have a precise normalization result found in the medical terminology graph, a new concept node is created based on the nearest parent concept of the original medical term and the verified and corrected attribute group.

7. A medical terminology atlas reasoning and construction device, characterized in that, include: The attribute extraction module is used to pre-extract attributes from the input raw medical terms based on the combination rule system and a large language model, and output the attribute group of the raw medical terms; the attribute group includes each medical entity type of the raw medical terms, as well as the attribute category and attribute value corresponding to each medical entity type; The verification and correction module is used to verify and correct the attribute group based on each medical entity type of the original medical term and the attribute category and attribute value corresponding to each medical entity type. The reasoning construction module is used to perform reasoning retrieval in the medical terminology graph based on the verified and corrected attribute groups. And / or, based on the validated and corrected attribute groups, construct or update the medical terminology map; The combination rule system classifies various medical entity types into simple entities, separable combination entities, and indivisible combination entities. The separable combination entities are defined with attribute categories and value ranges, and the value range of each attribute category is anchored to the defined medical concept and the sub-concepts of the defined medical concept. In the medical terminology graph, the nodes include concept nodes and terminology nodes; each concept node represents a medical concept, each terminology node represents a string with medical meaning, each concept node is connected to at least one terminology node, and the concept nodes are connected to each other through parent-child relationships and attribute relationships.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the medical terminology atlas reasoning construction method as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the medical terminology graph reasoning construction method as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the medical terminology graph reasoning construction method as described in any one of claims 1 to 6.

Images