An AI large model-based disease association named entity recognition method and system

Abstract

The application provides a disease association named entity recognition method based on an AI large model, comprising the following steps: constructing a high-quality medical corpus, wherein the high-quality medical corpus comprises electronic medical record text data and standard medical language data corresponding to the electronic medical record text data; extracting and coding features of the electronic medical record text data and the standard medical language data corresponding thereto through a large model respectively; fine-tuning the large model through a minimization multi-modal alignment loss function according to the electronic medical record coding features and the standard medical language coding features corresponding thereto; performing unsupervised entity extraction on authoritative medical literature through the fine-tuned large model, and screening a dynamic updated entity library through rule filtering and confidence sorting threshold screening; performing initial named entity recognition on a target electronic medical record, performing multi-dimensional semantic matching on the initial entity recognition result and the entity library, and taking the matching result as a final named entity recognition result. The application has the advantages of improving data efficiency, enhancing medical field knowledge adaptability and the like.

Description

Technical Field

[0001] This invention belongs to the field of biomedical information processing, and in particular relates to a method and system for disease-related named entity recognition based on a large AI model. Background Technology

[0002] Currently, the mainstream methods for named entity recognition (NER) in electronic medical records (EMR) mainly consist of traditional sequence labeling models (such as BiLSTM-CRF) and pre-trained language models (such as BERT). Traditional sequence labeling models rely on local context modeling, making it difficult to effectively distinguish the semantic relationship between entities and disease diagnoses. Furthermore, they lack the ability to handle the large amount of unstructured descriptions, variations of technical terms, and cross-paragraph semantic relationships in EERs, easily leading to the mis-extraction of entities unrelated to disease diagnoses. They also heavily rely on large-scale labeled data. While pre-trained language model-based methods can capture global semantics, their self-attention mechanism lacks adaptability to medical domain knowledge, also resulting in the mis-extraction of irrelevant entities when faced with complex medical texts. They also still require a large amount of labeled data for training. Both types of methods have significant limitations, including limited generalization ability, a lack of targeted recognition capabilities for disease-related entities, and poor data efficiency, making it difficult to meet the practical needs of medical scenarios where labeled resources are scarce and complex semantic relationships require processing. Summary of the Invention

[0003] To address the problems existing in the background art, one aspect of the present invention provides a disease-related named entity recognition method based on a large AI model, comprising:

[0004] S1: Construct a high-quality medical corpus, which includes: electronic medical record text data, and standard medical terminology data corresponding to the electronic medical record text data;

[0005] S2: Extract coding features from electronic medical record text data and its corresponding standard medical terminology data using a large model;

[0006] S3: Fine-tune the large model by minimizing the multimodal alignment loss function based on the electronic medical record coding features and their corresponding standard medical terminology coding features;

[0007] S4: Unsupervised entity extraction is performed on authoritative medical literature using a fine-tuned large model. A dynamically updated entity thesaurus is constructed by filtering based on rules and confidence ranking thresholds.

[0008] S5: Perform initial named entity recognition on the target electronic medical record, perform multi-dimensional semantic matching between the initial entity recognition results and the entity thesaurus, and use the matching results as the final named entity recognition results.

[0009] Another aspect of the present invention provides a disease-related named entity recognition system based on an AI large model, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the aforementioned disease-related named entity recognition method based on an AI large model.

[0010] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for disease-related named entity recognition based on an AI large model.

[0011] The present invention has at least the following beneficial effects

[0012] This invention constructs a high-quality medical corpus, utilizes a large model to extract and fine-tune the coding features of electronic medical records and standard medical terms, then performs unsupervised entity extraction on authoritative medical literature to construct a dynamic lexicon, and finally matches the initial recognition results of the target electronic medical record with the multi-dimensional semantics of the lexicon to obtain the final result. This invention can solve the problems of traditional methods in electronic medical record named entity recognition, such as dependence on large-scale labeled data, insufficient ability to handle term variations and cross-paragraph semantic associations, and has advantages such as improving data efficiency and enhancing the adaptability of medical knowledge. Attached Figure Description

[0013] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0014] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0015] Please see Figure 1 One aspect of the present invention provides a disease-related named entity recognition method based on a large AI model, comprising:

[0016] S1: Construct a high-quality medical corpus, which includes: electronic medical record text data, and standard medical terminology data corresponding to the electronic medical record text data;

[0017] In this embodiment, the electronic medical record text data includes outpatient medical records, inpatient medical records (such as chief complaint, present illness, past medical history, physical examination, auxiliary examinations, diagnosis, and treatment plan), examination reports (such as imaging reports and the text descriptions of laboratory reports), and surgical records, all of which have undergone rigorous de-identification processing. These electronic medical record texts naturally contain numerous variations of professional terminology (e.g., "hypertension" can be expressed as "high-pressure disease," "elevated blood pressure," or "primary hypertension," or "heart failure" can be abbreviated as "heart failure," "CHF," or "heart failure") and cross-paragraph semantic relationships (e.g., mentioning "sudden chest tightness" in the "chief complaint" while describing "chest pain radiating to the left shoulder, accompanied by profuse sweating" in the "present illness"—both in different paragraphs pointing to a potential association of "myocardial ischemia"). Standard medical terminology data can adopt standardized conceptual identifiers for diseases, symptoms, drugs, etc., derived from medical ontology or terminology sets such as SNOMED CT and ICD-10, or their preferred names.

[0018] S2: Extract coding features from electronic medical record text data and its corresponding standard medical terminology data using a large model;

[0019] Preferably, step S2 includes:

[0020] S21: Constructing a text feature encoder based on a generative large model f text and standard medical terminology feature encoder f term , where the text feature encoder f text and standard medical terminology feature encoder f term Generative large models are constructed using two identical network architectures;

[0021] S22: Input the electronic medical record text data into the text feature encoder f text Extract the text encoding features of electronic medical records; input the standard medical terminology data into the standard medical terminology feature encoder f term Extract standard medical terminology coding features, wherein the electronic medical record text coding features and standard medical terminology coding features are the output embeddings of specific layers of the corresponding generative large model.

[0022] In this embodiment, a text feature encoder f is constructed based on a generative large model. text and standard medical terminology feature encoder f term , where f text It is based on the output of a specific layer of a large model (the embedding of the [CLS] token) used to encode text fragments x and f in electronic medical records. term It is specifically designed for encoding standard medical terminology t, which also utilizes a large-model encoder architecture but is independent of f. text The aim is to provide refined, context-free semantic representations for standardized medical concepts.

[0023] Step S2 extracts encoded features from electronic medical record text data and standard medical terminology data by constructing a text feature encoder and a standard medical terminology feature encoder. This provides a crucial semantic representation foundation for the medical domain-oriented fine-tuning of large models (such as DeepSeek). Specifically, this step utilizes a generative large model architecture to encode features of unstructured text in electronic medical records (such as variants of the term "high blood pressure") and standard medical terms (such as "hypertension"), enabling the semantic information of both to be measured in a unified feature space. Through subsequent optimization of the multimodal alignment loss function, this process achieves semantic alignment between variant terms in electronic medical records and standard terms (e.g., bringing the feature vector of "hypertension" closer to the standard feature vector of "hypertension"). This enhances the large model's ability to identify variant medical terms and strengthens its ability to capture cross-paragraph semantic associations (such as the implicit connection between symptoms and disease diagnosis). This lays the semantic understanding foundation for subsequent dynamic lexicon construction and high-precision entity screening, effectively solving the problems of insufficient handling of terminology variants and weak cross-paragraph semantic association recognition capabilities in traditional methods.

[0024] S3: Fine-tune the large model by minimizing the multimodal alignment loss function based on the electronic medical record coding features and their corresponding standard medical terminology coding features;

[0025] Preferably, the multimodal alignment loss function includes:

[0026] L align =E x,t [[1-cos(f text ((x),f term ((t))]

[0027] Among them, L align E represents the multimodal alignment loss function. x,t f represents expectation text ((x) represents the electronic medical record coding characteristics; f term (t) represents the coding features of standard medical terminology; x represents the electronic medical record text data; t represents the standard medical terminology data; cos() represents the cosine similarity function.

[0028] In this embodiment, step S3 fine-tunes the large model by minimizing the multimodal alignment loss function, achieving semantic alignment between the electronic medical record coding features and the standard medical terminology coding features. By optimizing the loss function, the large model can establish strong associations between unstructured text in the electronic medical record (such as variants of terms like "hypertension" and "heart failure") and standard medical terms (such as "hypertension" and "heart failure") in the feature space, significantly improving the model's ability to identify variants of medical terms. Simultaneously, this fine-tuning process strengthens the large model's ability to capture cross-paragraph semantic associations. For example, it aligns scattered information such as "fever" in the chief complaint and "elevated white blood cell count" in the test report with the standard terminology feature of "infection." This enables the model to accurately understand the implicit connections between symptoms, test results, and disease diagnoses scattered across different paragraphs in the electronic medical record, effectively solving the problems of misidentification of terminology variants and insufficient cross-paragraph semantic association processing capabilities caused by a lack of domain knowledge adaptation in traditional methods.

[0029] S4: Unsupervised entity extraction is performed on authoritative medical literature using a fine-tuned large model. A dynamically updated entity thesaurus is constructed by filtering based on rules and confidence ranking thresholds.

[0030] Preferably, step S4 includes:

[0031] S41: Design a Prompt based on authoritative medical literature. Input the Prompt into the fine-tuned large model to extract entities and obtain an entity list. The entity list includes: entity, disease type to which the entity belongs, entity type and confidence level.

[0032] In this embodiment, the fine-tuned large model adopts the large model fine-tuned in step S3. The disease type to which the entity belongs is obtained based on the classification of authoritative medical literature, which belongs to prior knowledge. The Prompt designed for authoritative medical literature can be designed as follows:

[0033] Please extract all mentioned disease names from the following medical text and assign a confidence score (0-1, accurate to two decimal places) to each disease name. Do not extract if you are unsure. Text: {Literature Excerpt}

[0034] Model output example:

[0035]

[0036] S42: Filter the entity list by length, keeping entities with a length of 2-15 characters;

[0037] S43: Filter the entity list after step S42 using regular expressions to remove strings that do not meet the requirements;

[0038] In this embodiment, regular expression filtering is performed by defining some regular expression patterns to filter out strings that contain too many numbers or special symbols or that obviously do not conform to medical nomenclature (e.g., filtering out the part in parentheses in "C-reactive protein elevated (>10mg / L)" and keeping only "C-reactive protein elevated").

[0039] S44: Sort the entities in the entity list filtered in step S43 in descending order of confidence, set a threshold θ, and add entities with a confidence level greater than the threshold to the dynamic lexicon.

[0040] In this embodiment, step S4 uses a fine-tuned large model to perform unsupervised entity extraction from authoritative medical literature and construct a dynamic lexicon, achieving automated accumulation and updating of medical knowledge. Specifically, this step uses Prompt to guide the large model to extract entities from authoritative literature. After length filtering, regular expression filtering, and confidence threshold screening, a dynamically updated entity lexicon is formed. This process can continuously supplement long-tail medical terms (such as names of emerging diseases) without manual annotation, improving the model's coverage of rare disease terms. At the same time, confidence screening ensures the accuracy of entities in the lexicon. The dynamic lexicon can be updated in real time with new literature (such as adding terms like "viral pneumonia"), effectively solving the problem of domain knowledge lag in traditional methods. It provides a comprehensive and accurate terminology benchmark for subsequent multi-dimensional semantic matching, ultimately improving the response speed of entity recognition to emerging medical concepts.

[0041] S5: Perform initial named entity recognition on the target electronic medical record, perform multi-dimensional semantic matching between the initial entity recognition results and the entity thesaurus, and use the matching results as the final named entity recognition results.

[0042] Preferably, the step of performing multi-dimensional semantic matching between the initial entity recognition result and the entity lexicon includes:

[0043] S51: Represent the entity in the initial entity recognition result as e j The entity thesaurus contains disease types that are the same as the target electronic medical record, and are also related to entity e. j Entities of the same type are treated as entity e j The entity to be matched;

[0044] In this embodiment, the initial named entity recognition of the target electronic medical record includes: using a real-time named entity recognition model to perform initial named entity recognition on the target electronic medical record. The real-time named entity recognition model includes: a sequence labeling model (such as BiLSTM-CRF) or a pre-trained language model (such as BERT). In this embodiment, the disease type of the target electronic medical record is also obtained based on the classification of the target electronic medical record, which belongs to the prior knowledge of the target electronic medical record.

[0045] S52: Calculate entity ej Its word similarity score to each entity to be matched:

[0046]

[0047] Among them, sim_form(e j ,e ji ) represents the j-th entity e in the initial entity recognition result. j With its i-th entity e ji Word form similarity score; |e j | represents entity e j The character length; |e ji | represents entity e ji The character length; dlev(e j ,e ji ) represents entity e j and entity e ji Lewinstein distance;

[0048] S53: Obtain entity e via BioBERT j Given the semantic vector of the entity to be matched, and calculate entity e j Semantic similarity score with the entity to be matched:

[0049]

[0050] Among them, sim_context(e j ,e ji ) represents the j-th entity e in the initial entity recognition result. j With its i-th entity e ji semantic similarity score; v j Represents entity e j semantic vector; v ji Represents entity e ji The semantic similarity score; |||| represents the Euclidean norm;

[0051] S54: Design a Prompt that incorporates contextual information from the target electronic medical record to predict entity e using a fine-tuned large model. j The relevance score between it and the entity to be matched;

[0052] In this embodiment, the fine-tuned large model adopts the large model fine-tuned in step S3. The example of designing the Prompt by combining the contextual information of the target electronic medical record is as follows:

[0053] The original text of the electronic medical record mentions Where '{e j}'Whether it conforms to standard medical concepts'{e jiDoes 'refer to the same thing?' Please give a relevance score (0.0-1.0). Entity e extracted from the electronic medical record j The context content.

[0054] S55: Transfer entity e j The multi-dimensional semantic matching scores between the entity and the entity to be matched are weighted and fused to obtain a comprehensive score. If entity e exists... j If the overall score between entity e and the entity it is to be matched is higher than the set filtering threshold τ, then entity e will be matched. j This serves as the final named entity recognition result.

[0055] Preferably, the comprehensive score includes:

[0056] S(e j ,e ji )=α·sim_form(e j ,e ji )+β·sim_context((e j ,e ji )+γ·score_model((e j ,e ji )

[0057] Among them, S(e j ,e ji ) represents entity e j With its i-th entity e ji The overall score; α, β, and γ represent weight parameters, and α + β + γ = 1; score_model((e j ,e ji ) represents entity e j With its i-th entity e ji The correlation score.

[0058] In this embodiment, initial named entity recognition of the target electronic medical record is performed and multi-dimensional semantic matching with a dynamic lexicon is performed. A comprehensive score is calculated by weighted fusion of word form similarity, semantic vector, and large model correlation score. Combined with disease type screening and confidence threshold filtering, the accurate mapping from variant terms to standard terms is achieved. This effectively solves the problems of cross-paragraph semantic association recognition and non-standard term processing, improves entity recognition accuracy, and reduces the false extraction rate of irrelevant entities. At the same time, incremental learning of domain knowledge is achieved by relying on the dynamic lexicon, which enhances the model's ability to recognize emerging diseases and long-tail terms, providing reliable technical support for the accurate identification of medical entities.

[0059] Another aspect of the present invention provides a disease-related named entity recognition system based on an AI large model, the system comprising a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the aforementioned disease-related named entity recognition method based on an AI large model.

[0060] Another aspect of the present invention provides a computer storage medium storing a remote monitoring program, which, when executed by a processor, implements the aforementioned method for disease-related named entity recognition based on an AI large model.

[0061] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0062] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A disease-related named entity recognition method based on a large AI model, characterized in that, include: S1: Construct a high-quality medical corpus, which includes: electronic medical record text data, and standard medical terminology data corresponding to the electronic medical record text data; S2: Extract coding features from electronic medical record text data and its corresponding standard medical terminology data using a large model; S3: Fine-tune the large model by minimizing the multimodal alignment loss function based on the electronic medical record coding features and their corresponding standard medical terminology coding features; The multimodal alignment loss function includes: in, This represents the multimodal alignment loss function. Expressing expectations, Indicates the coding characteristics of electronic medical records; This represents the coding characteristics of standard medical terminology; Represents electronic medical record text data; Data representing standard medical terminology; Represents the cosine similarity function; S4: Unsupervised entity extraction is performed on authoritative medical literature using a fine-tuned large model. A dynamically updated entity thesaurus is constructed by filtering based on rules and confidence ranking thresholds. S5: Perform initial named entity recognition on the target electronic medical record, perform multi-dimensional semantic matching between the initial entity recognition results and the entity thesaurus, and use the matching results as the final named entity recognition results; The step of performing multi-dimensional semantic matching between the initial entity recognition result and the entity lexicon includes: S51: Represent the entities in the initial entity recognition results as... The entity thesaurus contains disease types that are the same as the target electronic medical record, and are also related to the entity... Entities of the same entity type are treated as entities The entity to be matched; S52: Computational Entity Its word similarity score to each entity to be matched: in, This indicates the first entity in the initial entity recognition result. individual entities Its first One entity to be matched Word form similarity score; Representing entities The length of the characters; Representing entities The length of the characters; Representing entities and entity Lewinstein distance; S53: Obtaining Entities via BioBERT Given the semantic vector of the entity to be matched, and calculate the entity... Semantic similarity score with the entity to be matched: in, This indicates the first entity in the initial entity recognition result. individual entities Its first One entity to be matched Semantic similarity score; Representing entities semantic vector; Representing entities Semantic similarity score; Denotes the Euclidean norm; S54: Design a prompt that incorporates contextual information from the target electronic medical record to predict entities using a fine-tuned large model. The relevance score between it and the entity to be matched; S55: Transfer entity The multi-dimensional semantic matching scores between the entity to be matched and the entity to be matched are weighted and fused to obtain a comprehensive score. If an entity exists... If the overall score between the entity and the entity to be matched is higher than the set filtering threshold τ, then the entity will be... This serves as the final named entity recognition result.

2. The method for disease-related named entity recognition based on a large AI model according to claim 1, characterized in that, Step S2 includes: S21: Constructing a text feature encoder based on a generative large model and standard medical terminology feature encoder Among them, the text feature encoder and standard medical terminology feature encoder Generative large models are constructed using two identical network architectures; S22: Input the electronic medical record text data into the text feature encoder Extract the text encoding features of electronic medical records; input standard medical terminology data into the standard medical terminology feature encoder. Extract standard medical terminology coding features, wherein the electronic medical record text coding features and standard medical terminology coding features are the output embeddings of specific layers of the corresponding generative large model.

3. The method for disease-related named entity recognition based on a large AI model according to claim 1, characterized in that, Step S4 includes: S41: Design a Prompt based on authoritative medical literature. Input the Prompt into the fine-tuned large model to extract entities and obtain an entity list. The entity list includes: entity, disease type to which the entity belongs, entity type and confidence level. S42: Filter the entity list by length, keeping entities with a length of 2-15 characters; S43: Filter the entity list after step S42 using regular expressions to remove strings that do not meet the requirements; S44: Sort the entities in the entity list filtered in step S43 in descending order of confidence, set a threshold θ, and add entities with a confidence level greater than the threshold to the dynamic lexicon.

4. The method for disease-related named entity recognition based on a large AI model according to claim 1, characterized in that, The initial named entity recognition of the target electronic medical record includes: using a real-time named recognition model to perform initial named entity recognition on the target electronic medical record, wherein the real-time named recognition model includes: a sequence labeling model or a pre-trained language model.

5. The method for disease-related named entity recognition based on a large AI model according to claim 1, characterized in that, The overall score includes: in, Representing entities Its first One entity to be matched Overall score; , and Represents the weight parameters, and + + =1; Representing entities Its first One entity to be matched The correlation score.

6. A disease-related named entity recognition system based on a large AI model, characterized in that, The system includes a memory and a processor; the memory is used to store an application program; the processor is used to run the application program and execute the disease association named entity recognition method based on an AI large model as described in any one of claims 1 to 5.

7. A computer storage medium, characterized in that, The computer storage medium stores a remote monitoring program, which, when executed by a processor, implements a disease-related named entity recognition method based on an AI large model as described in any one of claims 1 to 5.

Images