System and method for extracting key information from medical documents using deep NLP models

Abstract

The present invention discloses a system and method for extracting key information from medical documents using a deep NLP model, relating to the field of natural language processing technology. The method comprises the following steps: collecting multi-source data, cleaning noise, and aligning it with a standard terminology library to construct an original corpus, a standardized entity dictionary, and a key information ontology library; adapting and adjusting the annotated dataset using the pre-trained model BioBERT, training a multi-task NLP model based on the task, and optimizing the model weights based on high-frequency terms; processing new documents with the optimized model, updating the ontology library after threshold filtering, cluster analysis, and expert verification, and supplementing new terms and annotations; mixing new and old annotated data, and incrementally updating the model using the EWC algorithm; parsing new documents to construct a dynamic medical knowledge graph. This invention can effectively improve the existing situation in which new domain knowledge and new terminology cannot be quickly incorporated into the model, resulting in difficulty in timely model updates.

Description

Technical Field

[0001] The present invention relates to the field of natural language processing technology, and specifically to a system and method for using a deep NLP model to extract key information from medical documents. Background Art

[0002] In the pharmaceutical field, information is experiencing explosive growth. Various medical documents, as crucial carriers of knowledge, cover a wide range of topics, from basic research to clinical applications. Documents such as drug inserts, medical literature, and clinical guidelines contain vast amounts of valuable information, but most exist in unstructured form, making it difficult for computers to efficiently utilize them. Natural language processing (NLP) technology, through text cleaning and standardization, and the application of appropriate models for training and analysis, can transform unstructured medical text into structured knowledge, thereby constructing a medical knowledge graph, providing strong support for research, decision-making, and application in the pharmaceutical field.

[0003] In current medical knowledge management systems, ontology libraries are a core component. They define concepts, relationships, and attributes in the medical field, providing a framework for organizing and understanding knowledge. However, existing technologies have significant shortcomings in updating ontology libraries. With the rapid development of medical science, new knowledge, such as new drug molecules, disease diagnostic methods, and treatments, continues to emerge, and new medical terminology has also emerged. However, existing technologies are often unable to capture these changes in a timely manner, resulting in a lag in ontology updates. Summary of the Invention

[0004] The purpose of the present invention is to provide a system and method for extracting key information from medical documents using a deep NLP model to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for extracting key information from medical documents using a deep NLP model, the method comprising the following steps:

[0006] Step 1: Collect multi-source data, clean the noise, and align it with the standard terminology library to build the original corpus, standardized entity dictionary, and key information ontology library;

[0007] Step 2: Use the pre-trained model BioBERT to adapt the labeled dataset, train a multi-task NLP model based on the task, and optimize the model weights based on high-frequency terms;

[0008] Step 3: Process new documents with the optimized model, perform threshold filtering, cluster analysis, and expert verification, then update the ontology library and add new terms and annotations;

[0009] Step 4: Mix the new and old annotated data and use the EWC algorithm to incrementally update the model; parse the new documents and build a dynamic medical knowledge graph.

[0010] In step 1, various unstructured medical documents are collected, including not only common drug instructions and medical literature, but also clinical guidelines, case reports, pharmaceutical company R&D reports and other data sources to ensure comprehensive and rich knowledge.

[0011] Identify and clean various types of noise in original documents; noise types include but are not limited to formatting errors, irrelevant metadata, and duplicate content. Use regular expression matching, text format verification tools, and customized cleaning rules to comprehensively clean the original documents and obtain the original corpus.

[0012] For non-standard terms in the original corpus, a method based on rule matching and semantic similarity calculation is used to align them with the standard terminology database. The similarity between non-standard terms and terms in the standard terminology database is calculated using the edit distance algorithm. At the same time, the consistency of terms is determined by combining the lexical and syntactic rules specific to the medical field, thus obtaining a standardized medical entity dictionary.

[0013] Organize a team of medical experts to define entities, relationships, and attributes in detail based on standardized dictionaries and original corpora using specially developed annotation tools; and obtain a structured key information ontology library;

[0014] Collect public medical data sets, screen and evaluate them to ensure data quality and compatibility with this solution; integrate the screened public data sets with expert-annotated data, and form an annotated data set through operations such as data deduplication and format unification.

[0015] In step 2, BioBERT was selected as the base model for domain adaptation on the annotated dataset. BioBERT has a certain pre-training foundation in the biomedical field, but further optimization is needed to meet the specific medical knowledge graph construction requirements of this project. Fine-tuning the model parameters and feature representation on the annotated dataset was performed to better align it with the language characteristics and knowledge structure of the medical field, resulting in an optimized text encoder.

[0016] Use the optimized text encoder to encode the labeled dataset and convert the text data into a vector representation suitable for model processing;

[0017] A multi-task NLP model is constructed by jointly training a CRF (Conditional Random Field), a multi-head attention mechanism, and a question-answer decoder. During training, the roles and collaboration of each component are clearly defined. The CRF is used to handle label dependencies in sequence labeling tasks, the multi-head attention mechanism enhances the model's ability to focus on different parts of the text, and the question-answer decoder is used to extract specific knowledge from the text.

[0018] The attention distribution of the multi-task NLP model is analyzed. Based on the high-frequency terms in the standardized dictionary, the model's attention to key terms is enhanced by adjusting the attention weights. The frequency of occurrence and importance scores of high-frequency terms in the text are calculated, and the weights in the model's attention mechanism are adjusted according to the scores, so that the model focuses more on key terms when processing text, resulting in an optimized NLP model.

[0019] In step 3, the optimized NLP model is used to process the newly uploaded document. The model analyzes and predicts the document content. In the prediction result processing phase, the model prediction results are filtered by a preset threshold, and the parts with prediction confidence higher than the threshold are screened out to obtain a set of potential new information candidates.

[0020] For the context of the new information candidate set, the Sentence-BERT model is used to convert the text into semantic vectors, and then the K-means clustering algorithm is used to cluster these vectors. After the clustering operation, new term candidates with semantic grouping are obtained;

[0021] The clustering results are submitted to a specially established team of experts for verification. The experts clearly display relevant information about the clustering results, including the text content, contextual information, and cluster category of the new term candidate. The experts review and judge the new term candidates based on their professional knowledge and domain experience. Detailed verification standards and processes are developed. New terms and related annotation data that pass expert verification are updated to the existing ontology library, enabling dynamic expansion and improvement of the ontology library.

[0022] In step 4, the labeled data set and the newly labeled data are mixed, and the data playback method is used to ensure that the proportion of each type of data in the mixed data set is reasonable; by analyzing the distribution of different categories of data, the data is balanced using techniques such as oversampling or undersampling to obtain a balanced data set.

[0023] For the optimized NLP model parameters, the elastic weight consolidation (EWC) algorithm is used to protect important knowledge such as entities, relationships, and attributes that the model has learned in the annotated dataset; thus, an incrementally updated NLP model is obtained.

[0024] For newly uploaded documents, the LayoutLM model is used for parsing. The LayoutLM model parses the document into structured text or tabular data to obtain accurate structured output.

[0025] The parsed text is input into the incrementally updated model, which extracts entities, relationships, and attributes from the text. A medical knowledge graph is constructed using the Neo4j graph database, with the extracted entities as nodes, relationships as edges, and attributes as node or edge features. Through graph structure design and data storage methods, a dynamically updated medical knowledge graph is achieved.

[0026] A system for extracting key information from medical documents using a deep NLP model. The system includes a data preprocessing and annotation module, a multi-task model training and optimization module, a new term discovery and verification module, and an incremental learning and knowledge graph construction module.

[0027] The data preprocessing and annotation module is used to collect multi-source data, clean noise and align it according to the standard terminology library, and construct the original corpus, standardized entity dictionary and key information ontology library; the multi-task model training and optimization module is used to adapt and adjust the annotated dataset using the pre-trained model BioBERT, train the multi-task NLP model according to the task, and optimize the model weights based on high-frequency terms; the new term discovery and verification module is used to process new documents with the optimized model, update the ontology library after threshold filtering, cluster analysis, and expert verification, and supplement new terms and annotations; the incremental learning and knowledge graph construction module is used to mix new and old annotated data and incrementally update the model using the EWC algorithm; parse new documents and construct a dynamic medical knowledge graph;

[0028] The output end of the data preprocessing and labeling module is connected to the input end of the multi-task model training and optimization module; the output end of the multi-task model training and optimization module is connected to the input end of the new terminology discovery and verification module; the output end of the new terminology discovery and verification module is connected to the input end of the incremental learning and knowledge graph construction module and the data preprocessing and labeling module; the output end of the incremental learning and knowledge graph construction module is connected to the input end of the new terminology discovery and verification module.

[0029] The data preprocessing and annotation module includes a multi-source data acquisition unit, a data cleaning and standardization unit, and an ontology library construction and annotation unit;

[0030] The multi-source data acquisition unit is used to collect medical documents and public data sets; the data cleaning and standardization unit is used to clean noisy data, align the standard terminology library, and generate the original corpus and standardized entity dictionary; the ontology library construction and annotation unit is used by experts to annotate entities, relationships and attributes to form a structured key information ontology library and annotated data set;

[0031] The output end of the multi-source data acquisition unit is connected to the input end of the data cleaning and standardization unit; the output end of the data cleaning and standardization unit is connected to the input end of the ontology library construction and annotation unit; the output end of the ontology library construction and annotation unit is connected to the input end of the multi-task model training and optimization module.

[0032] The multi-task model training and optimization module includes a domain adaptation pre-training unit, a multi-task model joint training unit and a high-frequency term weight optimization unit;

[0033] The domain adaptation pre-training unit is used to use BioBERT to perform domain adaptation fine-tuning on medical field texts and optimize the text encoder; the multi-task model joint training unit is used to jointly train CRF entity recognition, multi-head attention relationship extraction, and question-answer decoder attribute extraction; the high-frequency term weight optimization unit is used to enhance the model attention weight based on high-frequency terms in the standardized dictionary and improve the accuracy of key information extraction;

[0034] The output end of the domain adaptation pre-training unit is connected to the input end of the multi-task model joint training unit; the output end of the multi-task model joint training unit is connected to the input end of the high-frequency term weight optimization unit; the output end of the high-frequency term weight optimization unit is connected to the input end of the new term discovery and verification module.

[0035] The new term discovery and verification module includes a new document information extraction unit, a semantic clustering analysis unit, and an expert verification and update unit;

[0036] The new document information extraction unit is used to process new documents using the optimized model and generate a set of potential new term candidates through filtering by a preset threshold; the semantic clustering analysis unit is used to cluster the context of candidate terms using Sentence-BERT and K-means to generate semantic groupings; the expert verification and update unit is used for expert review of the clustering results and to supplement the new terms and annotation data into the key information ontology library;

[0037] The output end of the new document information extraction unit is connected to the input end of the semantic clustering analysis unit; the output end of the semantic clustering analysis unit is connected to the input end of the expert verification and update unit; the output end of the expert verification and update unit is connected to the input end of the incremental learning and knowledge graph construction module and the data preprocessing and annotation module.

[0038] The incremental learning and knowledge graph construction module includes a hybrid data playback unit, an EWC incremental model update unit and a dynamic knowledge graph construction unit;

[0039] The hybrid data playback unit is used to mix new and old annotated data and balance the data set distribution to support incremental training. The EWC incremental model update unit is used to update the model using the elastic weight solidification EWC algorithm to protect historical knowledge from being forgotten. The dynamic knowledge graph construction unit is used to parse the new document structure, extract entity relationships, and update the knowledge graph using Neo4j.

[0040] The output end of the hybrid data playback unit is connected to the input end of the EWC incremental model update unit; the output end of the EWC incremental model update unit is connected to the input end of the dynamic knowledge graph construction unit; the output end of the dynamic knowledge graph construction unit is connected to the input end of the new term discovery and verification module.

[0041] Compared with the existing technology, the beneficial effects of the present invention are: the present invention collects common drug instructions and medical literature, and also incorporates multiple data sources such as clinical guidelines, case reports, and pharmaceutical company R&D reports to ensure the comprehensiveness and richness of knowledge. Compared with the existing technology that may only rely on a single or a few data sources, it can cover a wider range of medical field knowledge and provide a more solid data foundation for subsequent analysis and application; the present invention processes new documents, and after threshold filtering, cluster analysis and expert verification, updates new terms and related annotation data to the ontology library, thereby realizing the dynamic expansion and improvement of the ontology library, which can timely reflect new knowledge and new changes in the medical field and maintain the timeliness and integrity of the knowledge system. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 Schematic diagram of the steps of the method for extracting key information from medical documents using the deep NLP model of the present invention;

[0043] Figure 2 Schematic diagram of the process of using the deep NLP model of the present invention to extract key information from medical documents. DETAILED DESCRIPTION

[0044] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0045] Example: Figure 1-Figure 2 As shown, the present invention provides a technical solution, a method for extracting key information from medical documents using a deep NLP model, characterized in that the method comprises the following steps:

[0046] Step 1: Collect multi-source data, clean the noise, and align it with the standard terminology library to build the original corpus, standardized entity dictionary, and key information ontology library;

[0047] Step 2: Use the pre-trained model BioBERT to adapt the labeled dataset, train a multi-task NLP model based on the task, and optimize the model weights based on high-frequency terms;

[0048] Step 3: Process new documents with the optimized model, perform threshold filtering, cluster analysis, and expert verification, then update the ontology library and add new terms and annotations;

[0049] Step 4: Mix the new and old annotated data and use the EWC algorithm to incrementally update the model; parse the new documents and build a dynamic medical knowledge graph.

[0050] Step 1: Multi-source data collection and cleaning;

[0051] Data sources: drug instructions (1,000 copies, including PDF and text formats); medical journal articles (5,000 articles, from public medical databases); hospital desensitization clinical reports (200 copies, including records of efficacy and side effects);

[0052] Cleaning operations: Remove non-text content (such as tables and images in PDFs) and extract plain text; use regular expressions to filter special symbols (such as "®" and "©") and garbled characters; split text into paragraphs and remove duplicate paragraphs (such as the "General Warnings" section of drug instructions);

[0053] Example data snippet: A Phase II clinical trial of Drug A (main ingredient: chemical X) showed a 75% remission rate for Disease B, with common adverse reactions including rash (incidence 15%) and dizziness (incidence 8%).

[0054] Terminology standardization and alignment;

[0055] Standard library matching: Use the "National Drug Standard Terminology Library" to align non-standard names; for example: "Drug A" → standard name "Chemical A_National Drug Standard H2022001", "Disease B" → standard code "ICD-11:8A45.0";

[0056] Unmatched term processing: Mark as pending expert review item and temporarily store in a temporary library;

[0057] Expert annotation and ontology library construction:

[0058] Labeling rules: Entity type: drug, disease, ingredient, adverse reaction, test indicator;

[0059] Relationship definitions: drug-treatment-disease, drug-induces-adverse-reaction, ingredient-belongs-to-drug;

[0060] Annotation result example: {"text":"Drug A has a 75% remission rate for disease B, and adverse reactions include rash.","entity":[{"name":"Drug A","type":"Drug","standard code":"H2022001"},{"name":"Disease B","type":"Disease","standard code":"8A45.0"},{"name":"Rash","type":"Adverse reaction","standard code":"ADR_0032"}],"relationship":[{"subject":"H2022001","object":"8A45.0","type":"Treatment"},{"subject":"H2022001","object":"ADR_0032","type":"Cause"}]};

[0061] Step 2: Model training and optimization;

[0062] Model selection: BioBERT, a deep learning model pre-trained on medical text;

[0063] Input format: Divide the text into 512-character blocks, retaining key position information such as drug names and disease names; Parameter adjustment: Initial learning rate: 0.00002; Training rounds: 3 rounds;

[0064] Multi-task joint training:

[0065] Task design: Entity recognition task: labeling the start and end locations of drugs, diseases, and adverse reactions; relation extraction task: determining whether there are relationships such as "treatment" and "initiation" in a sentence; attribute extraction task: extracting numerical attributes from text (such as "remission rate 75%");

[0066] Model structure optimization: In the model attention layer, increase the weight of key entities such as drug names and disease names (for example, entity position attention weight × 1.5);

[0067] Using a dynamic loss balancing strategy, the entity recognition loss weight accounts for 60%, the relationship extraction loss weight accounts for 30%, and the attribute extraction loss weight accounts for 10%;

[0068] High-frequency term enhancement: Count the top 10% of terms in the standard term library (such as "chemotherapy" and "tumor"). During model training, if the text contains high-frequency terms, the loss function of the sample is weighted (for example, weight coefficient × 2).

[0069] Step 3: New term discovery and dynamic update:

[0070] Input example: A fragment from a new drug package insert: "Long-term use of drug C may cause elevated blood potassium (incidence 5%). Regular monitoring of electrolytes is recommended."

[0071] Model prediction results: Entities: "Drug C" (drug, confidence 0.94), "Increased blood potassium" (adverse reaction, confidence 0.68); Relationship: "Drug C-induced-Increased blood potassium" (confidence 0.72); Threshold filtering: retain results with confidence ≥ 0.65 ("Increased blood potassium" is retained);

[0072] Semantic Clustering and Expert Review:

[0073] Clustering process: Extract the context of "elevated blood potassium" (e.g., "long-term use may lead to elevated blood potassium"); use a semantic encoding model to convert the context into a vector; group by vector similarity and classify it into the "abnormal biochemical indicators" category along with "hyponatremia" and "elevated creatinine";

[0074] Expert decision: Confirm "increased blood potassium" as a new adverse reaction, code it as "ADR_0157"; add the relationship: "Drug C-induced-ADR_0157";

[0075] Ontology library and model iteration:

[0076] Data replay: Mix the newly added 200 annotated data items (including "elevated blood potassium") with the original 6,000 data items; randomly sample at a ratio of 5:1 to avoid imbalance between new and old data;

[0077] Incremental model training: Use elastic weight solidification technology to lock the model's recognition ability for old terms (such as "rash" and "dizziness"); only allow small adjustments to model parameters in the direction of newly added terms (the learning rate is set to 0.00001).

[0078] Step 4: Dynamic construction of knowledge graph;

[0079] Structured parsing and storage:

[0080] Table parsing results: {"Drug":"Drug C","Attribute":[{"Name":"Recommended Dose","Value":"50mg / d"},{"Name":"Half-Life","Value":"12 hours"}]};

[0081] Map update: New nodes added: Drug C (ID: H2023001), adverse reaction "increased blood potassium" (ID: ADR_0157); new relationship added: H2023001-trigger-ADR_0157; new attributes added: recommended dose and half-life of Drug C.

[0082] A system for extracting key information from medical documents using a deep NLP model. The system includes a data preprocessing and annotation module, a multi-task model training and optimization module, a new term discovery and verification module, and an incremental learning and knowledge graph construction module.

[0083] The data preprocessing and annotation module is used to collect multi-source data, clean noise and align it according to the standard terminology library, and construct the original corpus, standardized entity dictionary and key information ontology library; the multi-task model training and optimization module is used to adapt and adjust the annotated dataset using the pre-trained model BioBERT, train the multi-task NLP model according to the task, and optimize the model weights based on high-frequency terms; the new term discovery and verification module is used to process new documents with the optimized model, update the ontology library after threshold filtering, cluster analysis, and expert verification, and supplement new terms and annotations; the incremental learning and knowledge graph construction module is used to mix new and old annotated data and incrementally update the model using the EWC algorithm; parse new documents and construct a dynamic medical knowledge graph;

[0084] The output end of the data preprocessing and labeling module is connected to the input end of the multi-task model training and optimization module; the output end of the multi-task model training and optimization module is connected to the input end of the new terminology discovery and verification module; the output end of the new terminology discovery and verification module is connected to the input end of the incremental learning and knowledge graph construction module and the data preprocessing and labeling module; the output end of the incremental learning and knowledge graph construction module is connected to the input end of the new terminology discovery and verification module.

[0085] The data preprocessing and annotation module includes a multi-source data acquisition unit, a data cleaning and standardization unit, and an ontology library construction and annotation unit;

[0086] The multi-source data acquisition unit is used to collect medical documents and public data sets; the data cleaning and standardization unit is used to clean noisy data, align the standard terminology library, and generate the original corpus and standardized entity dictionary; the ontology library construction and annotation unit is used by experts to annotate entities, relationships and attributes to form a structured key information ontology library and annotated data set;

[0087] The output end of the multi-source data acquisition unit is connected to the input end of the data cleaning and standardization unit; the output end of the data cleaning and standardization unit is connected to the input end of the ontology library construction and annotation unit; the output end of the ontology library construction and annotation unit is connected to the input end of the multi-task model training and optimization module.

[0088] The multi-task model training and optimization module includes a domain adaptation pre-training unit, a multi-task model joint training unit and a high-frequency term weight optimization unit;

[0089] The domain adaptation pre-training unit is used to use BioBERT to perform domain adaptation fine-tuning on medical field texts and optimize the text encoder; the multi-task model joint training unit is used to jointly train CRF entity recognition, multi-head attention relationship extraction, and question-answer decoder attribute extraction; the high-frequency term weight optimization unit is used to enhance the model attention weight based on high-frequency terms in the standardized dictionary and improve the accuracy of key information extraction;

[0090] The output end of the domain adaptation pre-training unit is connected to the input end of the multi-task model joint training unit; the output end of the multi-task model joint training unit is connected to the input end of the high-frequency term weight optimization unit; the output end of the high-frequency term weight optimization unit is connected to the input end of the new term discovery and verification module.

[0091] The new term discovery and verification module includes a new document information extraction unit, a semantic clustering analysis unit, and an expert verification and update unit;

[0092] The new document information extraction unit is used to process new documents using the optimized model and generate a set of potential new term candidates through filtering by a preset threshold; the semantic clustering analysis unit is used to cluster the context of candidate terms using Sentence-BERT and K-means to generate semantic groupings; the expert verification and update unit is used for expert review of the clustering results and to supplement the new terms and annotation data into the key information ontology library;

[0093] The output end of the new document information extraction unit is connected to the input end of the semantic clustering analysis unit; the output end of the semantic clustering analysis unit is connected to the input end of the expert verification and update unit; the output end of the expert verification and update unit is connected to the input end of the incremental learning and knowledge graph construction module and the data preprocessing and annotation module.

[0094] The incremental learning and knowledge graph construction module includes a hybrid data playback unit, an EWC incremental model update unit and a dynamic knowledge graph construction unit;

[0095] The hybrid data playback unit is used to mix new and old annotated data and balance the data set distribution to support incremental training. The EWC incremental model update unit is used to update the model using the elastic weight solidification EWC algorithm to protect historical knowledge from being forgotten. The dynamic knowledge graph construction unit is used to parse the new document structure, extract entity relationships, and update the knowledge graph using Neo4j.

[0096] The output end of the hybrid data playback unit is connected to the input end of the EWC incremental model update unit; the output end of the EWC incremental model update unit is connected to the input end of the dynamic knowledge graph construction unit; the output end of the dynamic knowledge graph construction unit is connected to the input end of the new term discovery and verification module.

[0097] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

Claims

1. A method for extracting key information from medical documents using a deep NLP model, characterized by: The method comprises the following steps: Step 1: Collect multi-source data, clean the noise, and align it with the standard terminology library to build the original corpus, standardized entity dictionary, and key information ontology library; Step 2: Use the pre-trained model BioBERT to adapt the labeled dataset, train a multi-task NLP model based on the task, and optimize the model weights based on high-frequency terms; Step 3: Process new documents with the optimized model, perform threshold filtering, cluster analysis, and expert verification, then update the ontology library and add new terms and annotations; Step 4: Mix the new and old annotated data and incrementally update the model using the EWC algorithm; parse the new documents and build a dynamic medical knowledge graph; In step 1, unstructured medical documents, including instructions, reports, and literature, are collected; For the original document, clean the noise to get the original corpus; For non-standard terms in the original corpus, the standard term base is used for alignment to obtain a standardized entity dictionary; For the standardized dictionary and original corpus, define entities, relationships and attributes through expert annotation to obtain a structured key information ontology library; Integrate public datasets and expert-annotated data to form an annotated dataset; In step 2, for the labeled dataset, the pre-trained model BioBERT is used for domain adaptation to obtain an optimized text encoder; Encode the annotated dataset using an optimized text encoder; For the encoded text, a multi-task NLP model is obtained by jointly training CRF, multi-head attention, and question-answer decoder; For the attention distribution of the multi-task NLP model, the weights are enhanced based on the high-frequency terms in the standardized dictionary to obtain an optimized NLP model; In step 3, the optimized NLP model is used to process new documents. The model prediction results are filtered based on a preset threshold to obtain a set of potential new information candidates. For the context of the new information candidate set, the Sentence-BERT model is used to convert the text into semantic vectors, and then the K-means clustering algorithm is used to cluster the semantic vectors to obtain semantic groups. The semantic groups are clustered to obtain new term candidates. The clustering results are verified through the expert platform to obtain an updated ontology library, including newly added terms and annotation data; In step 4, the labeled data set and the newly labeled data are mixed using data replay to obtain a balanced data set; For the optimized NLP model parameters, an elastic weight solidification algorithm is used to protect the entities, relationships, and attributes learned by the optimized NLP model in the annotated dataset, thereby obtaining an incrementally updated NLP model. For newly uploaded documents, use LayoutLM to parse them and obtain structured text or table data; For the parsed text, use the incrementally updated model to obtain entity, relationship or attribute results; For the extracted results, Neo4j is used to build a graph to obtain a dynamically updated medical knowledge graph.

2. A system for extracting key information from medical documents using a deep NLP model, applied to the method for extracting key information from medical documents using a deep NLP model as described in claim 1, characterized in that: The system includes a data preprocessing and annotation module, a multi-task model training and optimization module, a new term discovery and verification module, and an incremental learning and knowledge graph construction module; The data preprocessing and annotation module is used to collect multi-source data, clean noise and align it according to the standard terminology library, and construct the original corpus, standardized entity dictionary and key information ontology library; the multi-task model training and optimization module is used to adapt and adjust the annotated dataset using the pre-trained model BioBERT, train the multi-task NLP model according to the task, and optimize the model weights based on high-frequency terms; the new term discovery and verification module is used to process new documents with the optimized model, update the ontology library after threshold filtering, cluster analysis, and expert verification, and supplement new terms and annotations; the incremental learning and knowledge graph construction module is used to mix new and old annotated data and incrementally update the model using the EWC algorithm; parse new documents and construct a dynamic medical knowledge graph; The output end of the data preprocessing and annotation module is connected to the input end of the multi-task model training and optimization module; the output end of the multi-task model training and optimization module is connected to the input end of the new terminology discovery and verification module; the output end of the new terminology discovery and verification module is connected to the input end of the incremental learning and knowledge graph construction module and the data preprocessing and annotation module; the output end of the incremental learning and knowledge graph construction module is connected to the input end of the new terminology discovery and verification module; The data preprocessing and annotation module includes a multi-source data acquisition unit, a data cleaning and standardization unit, and an ontology library construction and annotation unit; The multi-source data acquisition unit is used to collect medical documents and public data sets; the data cleaning and standardization unit is used to clean noisy data, align the standard terminology library, and generate the original corpus and standardized entity dictionary; the ontology library construction and annotation unit is used by experts to annotate entities, relationships and attributes to form a structured key information ontology library and annotated data set; The output end of the multi-source data acquisition unit is connected to the input end of the data cleaning and standardization unit; the output end of the data cleaning and standardization unit is connected to the input end of the ontology library construction and annotation unit; the output end of the ontology library construction and annotation unit is connected to the input end of the multi-task model training and optimization module; The multi-task model training and optimization module includes a domain adaptation pre-training unit, a multi-task model joint training unit and a high-frequency term weight optimization unit; The domain adaptation pre-training unit is used to use BioBERT to perform domain adaptation fine-tuning on medical field texts and optimize the text encoder; the multi-task model joint training unit is used to jointly train CRF entity recognition, multi-head attention relationship extraction, and question-answer decoder attribute extraction; the high-frequency term weight optimization unit is used to enhance the model attention weight based on high-frequency terms in the standardized dictionary and improve the accuracy of key information extraction; The output end of the domain adaptation pre-training unit is connected to the input end of the multi-task model joint training unit; the output end of the multi-task model joint training unit is connected to the input end of the high-frequency term weight optimization unit; the output end of the high-frequency term weight optimization unit is connected to the input end of the new term discovery and verification module; The new term discovery and verification module includes a new document information extraction unit, a semantic clustering analysis unit, and an expert verification and update unit; The new document information extraction unit is used to process new documents using the optimized model and generate a set of potential new term candidates through filtering by a preset threshold; the semantic clustering analysis unit is used to cluster the context of candidate terms using Sentence-BERT and K-means to generate semantic groupings; the expert verification and update unit is used for expert review of the clustering results and to supplement the new terms and annotation data into the key information ontology library; The output end of the new document information extraction unit is connected to the input end of the semantic clustering analysis unit; the output end of the semantic clustering analysis unit is connected to the input end of the expert verification and update unit; the output end of the expert verification and update unit is connected to the input end of the incremental learning and knowledge graph construction module and the data preprocessing and annotation module; The incremental learning and knowledge graph construction module includes a hybrid data playback unit, an EWC incremental model update unit and a dynamic knowledge graph construction unit; The mixed data playback unit is used to mix new and old annotated data to balance the data set distribution to support incremental training; the EWC incremental model update unit is used to use elastic weight solidification EWC algorithm to update the model to protect historical knowledge from being forgotten; The dynamic knowledge graph construction unit is used to parse the new document structure, extract entity relationships, and update the knowledge graph using Neo4j; The output end of the hybrid data playback unit is connected to the input end of the EWC incremental model update unit; the output end of the EWC incremental model update unit is connected to the input end of the dynamic knowledge graph construction unit; the output end of the dynamic knowledge graph construction unit is connected to the input end of the new term discovery and verification module.

Images