Operation term multi-level normalization method, system, equipment and medium

Abstract

The invention provides a multi-level normalization method, system and device for operation terms and a medium, and belongs to the technical field of medical information processing. The method comprises the following steps: firstly, carrying out attribute entity disassembly through a large language model, normalizing each attribute expression based on a term knowledge graph of each attribute, and then carrying out knowledge reasoning by using a plurality of normalized attributes based on a standard operation knowledge graph to screen a normalized operation entity, thereby effectively avoiding a calculation error caused by text similarity. And under the condition that the normalization result cannot be obtained by the standard operation knowledge graph, screening the optimally matched operation normalization entity through the similarity. Finally, under the condition that knowledge reasoning of the standard operation knowledge map and similarity screening operation normalization entities are invalid, possible candidate operation entities are summarized, and the most suitable operation entity normalization result is selected from the candidate operation entities through the reasoning ability of a large language model. And the uncertainty caused by model prediction or large model reasoning is reduced.

Description

Technical Field

[0001] This invention belongs to the field of medical information processing technology, and more specifically relates to a method, system, device and medium for multi-level normalization of surgical terminology. Background Technology

[0002] With the deepening development of medical informatization, medical institutions have accumulated massive amounts of surgical-related data. However, due to multiple factors such as regional differences, inconsistent standards among medical institutions, diverse writing habits of doctors, and heterogeneous information systems, the same standard surgical procedure can be described in dozens or even hundreds of different textual ways in clinical practice. This diversity in terminology severely restricts the accuracy and efficiency of key applications such as intelligent medical insurance review, clinical research analysis, and medical quality assessment based on surgical data. Therefore, the precise normalization of surgical terminology has become an urgent need in the field of medical information standardization.

[0003] Current mainstream normalization methods primarily rely on calculating text similarity between large language models and candidate standard entities. These methods use text matching algorithms such as cosine similarity and edit distance to filter entities from the candidate standard surgical procedure list that are most similar to the input description. However, relying solely on text similarity has significant limitations: on the one hand, it is prone to false positives due to differences in expression length and modifiers; on the other hand, traditional text similarity algorithms struggle to accurately capture the semantic equivalence relationships of the numerous synonyms and near-synonyms present in the medical field. Furthermore, the effectiveness of various custom similarity calculation methods is highly dependent on the completeness of the algorithm design, and they often perform inconsistently in real-world, complex medical scenarios.

[0004] Another common approach is to build specialized models based on manually labeled training data. This method requires a significant investment of professional manpower for data labeling, which is not only costly but also directly impacts model performance due to the quality of the labeling. More challenging is the uncertainty in the model's generalization ability caused by biases in the distribution of training data, potentially leading to systematic biases in the output results. From a long-term maintenance perspective, surgical terminology standards are updated periodically. Faced with new standard terminology, methods based on training data require re-labeling, retraining, and re-testing and optimizing the data. This results in a long update cycle, poor timeliness, and difficulty in adapting to the rapid iteration of medical standards.

[0005] In summary, existing technologies have significant shortcomings in terms of accuracy, interpretability, and maintenance costs. In particular, they lack an effective solution to combine the semantic understanding capabilities of large language models with the structured reasoning capabilities of medical knowledge graphs, failing to achieve multi-level, interpretable normalization processing from surgical attribute decomposition to knowledge reasoning. Therefore, there is an urgent need for a novel surgical terminology normalization method that can overcome these deficiencies. Summary of the Invention

[0006] To address the above problems, the present invention aims to provide a method, system, device, and medium for multi-level normalization of surgical terminology. By leveraging the semantic understanding capabilities of a large language model, the structured reasoning capabilities of a knowledge graph, vector similarity calculation, and a multi-level decision-making mechanism, it achieves high-precision, interpretable, and automated normalization of surgical terminology from free text to standard terminology.

[0007] To achieve the above objectives, the present invention employs the following technical solution: In a first aspect, embodiments of this application provide a method for multi-level normalization of surgical terminology, including: Raw surgical description data is collected and processed into structured attribute entity data through a large language model. The structured attribute entity data includes location entity, surgical procedure entity, approach entity, and disease nature entity. The structured attribute entity data is obtained and processed based on the standard terminology knowledge graph corresponding to each attribute to generate standardized attribute data. The standardized attribute data includes standard site terms, standard surgical procedure terms, standard approach terms, and standard disease nature terms. The standardized attribute data is obtained, and a search is performed in the standard surgical knowledge graph to generate search result data. The search result data includes a matching status identifier, and when the matching status identifier is multiple candidate matches or no match, it also includes a list of candidate surgical standard terms. Based on the matching status identifier in the search results data, when the matching status identifier is multiple candidate matches or no match, the candidate surgical standard terminology list and standardized attribute data are processed by vector similarity calculation to generate a candidate surgical standard terminology list with similarity scores. Based on the candidate surgical standard terminology list with similarity scores, structured attribute entity data, and original surgical description data, normalized surgical standard terminology data is generated by constructing medical knowledge reasoning prompts and processing them using a large language model.

[0008] In an optional implementation, the acquisition of raw surgical description data, and its processing into structured attribute entity data using a large language model, includes: Construct an attribute decomposition prompt template based on the CO-STAR specification, where the role is defined as a medical information standardization expert, the target attribute list includes location, surgical procedure, approach and disease nature, and the output format is specified as JSON structure; The raw surgical description data and the prompt word template are combined and then input into the large language model to obtain the initial response data; Perform JSON parsing on the initial response data to extract the values ​​of the site entity field, surgical procedure entity field, approach entity field, and disease nature entity field; The parsed entity data is validated for integrity. The integrity validation is considered to be successful when there are at least two valid entities in the site entity field, surgical procedure entity field, approach entity field, and disease nature entity field. When validation passes, the system outputs structured attribute entity data containing valid entity data and marks missing fields as null values; when validation fails, the system outputs a disassembly exception flag and records the reason for the failure.

[0009] In one optional implementation, the step of acquiring the structured attribute entity data and processing it based on the standard terminology knowledge graph corresponding to each attribute to generate standardized attribute data includes: Four independent attribute terminology knowledge graphs were established, which were used to store standard terminology sets for site of origin, surgical procedure, approach, and disease nature, respectively. For each attribute entity in the structured attribute entity data, a three-level mapping process is performed in the corresponding attribute term knowledge graph. First, a string exact match is performed in the corresponding standard term set. If the exact match is successful, the mapping result is recorded. If the exact match fails, the word vectors of the attribute entity and each candidate term in the standard term set are calculated, and the cosine similarity is calculated. Standard terms with a cosine similarity exceeding the first threshold are selected through similarity matching. When similarity matching fails, text edit distance similarity is calculated, and standard terms with a text edit distance similarity exceeding the second threshold are selected. Record the mapping results for each attribute entity, including the selected standard terms, mapping type, and confidence score; Generate standardized attribute data containing all attribute standard terminology, and retain the original text for attributes that fail to be mapped and mark them as unstandardized.

[0010] In one optional implementation, the step of acquiring the standardized attribute data, performing retrieval processing in a standard surgical knowledge graph, and generating retrieval result data includes: A standard surgical knowledge graph is constructed, wherein the standard surgical knowledge graph uses standard surgical terminology nodes as nodes, each node is associated with a standard surgical code and a standard surgical name, and is connected to site nodes, surgical procedure nodes, approach nodes and disease nature nodes through edges; Convert standardized attribute data into a graph query statement and set four attribute standard terms as query conditions; Perform a graph query and analyze the query results; when the query results of the four attribute standard terms uniquely identify a standard surgical term node, output the standard surgical code and standard surgical name of that node as the final normalized result, and set the matching status flag to exact match; When multiple standard surgical terminology nodes are found, the standard surgical code, standard surgical name, and attribute information of each standard surgical terminology node are output as a candidate surgical standard terminology list, and the matching status is set to multiple candidate matches; when the query result is empty, an empty candidate list is output, and the matching status is set to no match.

[0011] In an optional implementation, the step of processing the candidate surgical standard terminology list and standardized attribute data through vector similarity calculation to generate a candidate surgical standard terminology list with similarity scores includes: A pre-built standard surgical terminology vector library is generated as follows: for each standard surgical terminology node, its associated standard site terms, standard procedure terms, standard approach terms, and standard disease nature terms are concatenated into descriptive text in a fixed order, and then converted into a descriptive vector using a text embedding model. The query text is constructed by concatenating the standard terms of each attribute in the standardized attribute data in a fixed order, and then converted into a query vector using a text embedding model. Calculate the cosine similarity between the query vector and the description vector of the candidate surgical standard term node as a similarity score, and generate a similarity pairing list of candidate terms; The similarity pairing list of candidate terms is sorted in descending order based on similarity score, and a similarity threshold of similarity_threshold is applied for filtering, retaining candidate terms with similarity scores exceeding similarity_threshold. Output a list of candidate surgical standard terms with similarity scores. Each list item includes the standard surgical code, standard surgical name, and similarity score.

[0012] In an optional implementation, the step of generating normalized surgical standard terminology data by constructing medical knowledge reasoning prompts and processing them using a large language model, based on a candidate list of surgical standard terms with similarity scores, structured attribute entity data, and original surgical description data, includes: A medical knowledge reasoning prompt template is constructed, which includes a system role definition section, a task description section, a context information section, and an output format section; the context information section is used to integrate a list of candidate surgical standard terms with similarity scores, structured attribute entity data, and raw surgical description data; Based on the candidate surgical standard terminology list, sorted in descending order of similarity score, select the top N candidate terms and their scores, fill them into the corresponding positions in the prompt word template, and generate prompt word data; Call the large language model API interface, pass in the prompt word data, and obtain the response output data of the large language model; The response output data is structured and parsed to verify whether the output format conforms to the agreed JSON specification, and the standard surgical code field is extracted. When parsing is successful and the standard surgical code exists in the list of candidate surgical standard terms, the standard surgical name corresponding to the code is output as normalized surgical standard terminology data; when parsing fails or the standard surgical code does not exist, the standard surgical name of the candidate term with the highest similarity score is selected as a temporary result, and a manual review process is triggered.

[0013] In one alternative implementation, the first threshold is 0.9 and the second threshold is 0.7.

[0014] Secondly, embodiments of this application also provide a multi-level normalization system for surgical terminology, including: The attribute decomposition module is used to collect raw surgical description data and process it into structured attribute entity data through a large language model. The structured attribute entity data includes location entity, surgical procedure entity, approach entity, and disease nature entity. The terminology standardization module is used to acquire the structured attribute entity data, process it based on the standard terminology knowledge graph corresponding to each attribute, and generate standardized attribute data. The standardized attribute data includes standard site terms, standard surgical procedure terms, standard approach terms, and standard disease nature terms. The graph retrieval module is used to obtain the standardized attribute data, perform retrieval processing in the standard surgical knowledge graph, and generate retrieval result data. The retrieval result data includes a matching status identifier, and when the matching status identifier is multiple candidate matches or no match, it also includes a list of candidate surgical standard terms. The similarity filtering module is used to generate a candidate surgical standard terminology list with similarity scores by processing the candidate surgical standard terminology list and standardized attribute data according to the matching status identifier in the search result data. When the matching status identifier is multiple candidate matches or no match, the module is used to process the data through vector similarity calculation based on the candidate surgical standard terminology list and standardized attribute data. The intelligent reasoning module is used to generate normalized surgical standard terminology data by constructing medical knowledge reasoning prompts and calling a large language model to process the candidate surgical standard terminology list with similarity scores, structured attribute entity data, and original surgical description data.

[0015] Thirdly, embodiments of this application also provide 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 program to implement the steps of the surgical terminology multi-level normalization method as described in any of the above.

[0016] Fourthly, embodiments of this application also provide a storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the surgical terminology multi-level normalization method as described in any of the above claims.

[0017] As can be seen from the above technical solutions, the present invention has the following advantages: The multi-level normalization method for surgical terminology provided in this application firstly decomposes attribute entities using a large language model, then normalizes the descriptions of each attribute based on the terminology knowledge graph of each attribute. Next, it uses multiple normalized attributes to perform knowledge reasoning and filter normalized surgical entities based on a standard surgical knowledge graph, minimizing computational errors caused by text similarity. When the standard surgical knowledge graph cannot provide normalization results, the best-matching normalized surgical entity is selected using similarity (e.g., text cosine similarity). Finally, when both knowledge reasoning from the standard surgical knowledge graph and similarity-based selection of normalized surgical entities fail, possible candidate surgical entities are aggregated, and the reasoning capability of the large language model is used to select the most suitable normalized surgical entity from the candidate entities, reducing the uncertainty caused by model prediction or large model reasoning.

[0018] This application establishes a dual guarantee mechanism by combining the deep semantic understanding capabilities of a large language model with the precise structured reasoning of a knowledge graph. It integrates five progressive decision-making levels—attribute decomposition, terminology standardization, graph retrieval, similarity calculation, and intelligent reasoning—to effectively overcome the limitations of traditional single-technology solutions. It demonstrates higher accuracy and reliability when handling complex surgical descriptions, synonym expressions, and variations in professional terminology, and significantly reduces the probability of mismatches and missed matches.

[0019] This application constructs a complete transparent processing pipeline, preserving clear processing traces and reasoning basis for each decision-making step. From the attribute decomposition of the original description to the determination of the final standard terminology, the attribute mapping relationships, graph matching paths, similarity calculation results, and reasoning logic throughout the process are traceable and verifiable, fully meeting the stringent requirements of interpretability in medical scenarios and providing reliable support for medical quality control and auditing.

[0020] This application employs a modular architecture design, making knowledge graph maintenance and algorithm model updates relatively independent. When surgical terminology standards change, only the terminology nodes and relationships in the knowledge graph need to be updated, without retraining the large language model or adjusting the overall architecture. This greatly reduces system maintenance costs, ensures rapid response to dynamic updates of medical standard terminology, and effectively solves the problem of lagging updates in traditional training model methods.

[0021] This application achieves fully automated processing from raw descriptions to standard terminology by constructing an intelligent, multi-level decision-making pipeline. A manual review mechanism is only triggered when all automated steps fail to achieve a high-confidence result. This design significantly reduces the workload of manual intervention, improves the efficiency of surgical terminology normalization, and is particularly suitable for the batch processing needs of large-scale medical data.

[0022] This application fully leverages the synergistic advantages of various technologies, including large language models, knowledge graphs, and vector computation. Large language models provide deep semantic understanding capabilities, knowledge graphs ensure the accuracy and structure of professional knowledge, and vector computation enables fine-grained similarity matching. These technologies complement and verify each other, forming a complete technological closed loop. Attached Figure Description

[0023] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the description will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 A flowchart illustrating the multi-level normalization method for surgical terminology provided in this application.

[0025] Figure 2 A schematic diagram of the structure of the multi-level normalization system for surgical terminology provided in this application.

[0026] Figure 3 A schematic diagram of the structure of the electronic device provided in this application. Detailed Implementation

[0027] Various embodiments of this disclosure will be described more fully in the detailed steps of the multi-level normalization method for surgical terminology described below. This disclosure may have various embodiments, and adjustments and changes may be made therein. However, it should be understood that there is no intention to limit the various embodiments of this disclosure to the specific embodiments disclosed herein, but rather this disclosure should be understood to cover all adjustments, equivalents, and / or alternatives falling within the spirit and scope of the various embodiments of this disclosure.

[0028] In the following, the terms “comprising” or “may include”, which may be used in various embodiments of this disclosure, indicate the presence of the disclosed functions, operations, or elements, and do not limit the addition of one or more functions, operations, or elements. Furthermore, as used in various embodiments of this disclosure, the terms “comprising,” “having,” and their cognates are intended only to indicate a particular feature, number, step, operation, element, component, or combination of the foregoing, and should not be construed as primarily excluding the presence of one or more other features, numbers, steps, operations, elements, components, or combinations of the foregoing, or the possibility of adding one or more combinations of the foregoing.

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Please see Figure 1 The diagram shows a flowchart of a multi-level normalization method for surgical terminology in a specific embodiment. The method includes: S1: Collect raw surgical description data and process it into structured attribute entity data through a large language model.

[0031] In specific implementations, structured attribute entity data includes location entities, surgical procedure entities, approach entities, and disease-related entities.

[0032] First, a prompt word template is constructed. Following the CO-STAR specification, the large language model is designated as a medical information standardization expert. The target attribute list is defined to cover location, surgical procedure, approach, and disease nature, and the output format is specified as JSON. For example, the template provides a sample input: laparoscopic cholecystectomy. The output would be: body_part = gallbladder, method = resection, approach = laparoscopy, and pathology = null.

[0033] Then, the original surgical description data, such as laparoscopic cholecystectomy, is combined with prompt word templates to form the input data for the large language model. Initial response data is obtained by calling the API of a large language model such as GPT-4 or a medical-specific model. Next, the initial response data is parsed as JSON, extracting the values ​​of the entity fields for location, procedure, approach, and disease nature, and performing integrity verification to check that at least two fields contain valid entities (i.e., non-null values). If the verification passes, structured attribute entity data is output, for example, body_part for gallbladder, method for resection, approach for laparoscopy, and pathology for null values; missing fields are marked as null. If the verification fails, a decomposition anomaly flag, such as DECOMPOSE_ERROR, is output, and the reason for the failure, such as too many missing fields, is recorded for subsequent anomaly handling. This step utilizes the generalization ability of the large language model to accurately decompose attribute entities from free text, improving the accuracy and standardization of the decomposition and reducing manual intervention.

[0034] In this step, the large language model is a large language model that has been fully pre-trained on both general and medical corpora. Preferred model types include, but are not limited to: models specifically optimized for Chinese medical scenarios (such as Hua Tuo GPT and Bian Que large models), open-source models supporting efficient fine-tuning (such as LLaMA-2 / 3 and the ChatGLM series), or commercial APIs with powerful medical inference capabilities (such as GPT-4 and corresponding versions of Claude). The core criteria for model selection include the depth of understanding of medical terminology, the length of the context window, and the accuracy of instruction compliance.

[0035] To adapt to the application scenarios of this step, domain-specific enhancement processing needs to be performed on the selected large language model, specific parameters and design of the project need to be provided, and output normalization and post-processing are required. The specific processing methods are as follows: 1. Domain-specific enhancement processing: Using large-scale, high-quality medical text corpora (such as medical textbooks, clinical guidelines, and academic papers) and structured surgical record data, the selected large language model was further pre-trained in a domain-adaptive manner. Through this continued pre-training, the model was able to deeply integrate standard terms and their relationships from the medical knowledge graph, thereby improving its ability to represent surgical-related entities, relationships, and context.

[0036] A fine-tuning dataset for instructions is constructed, consisting of paired data of "(surgical description, structured attributes)". Data sources include: standard examples generated from official names and standard attribute combinations based on standard surgical libraries (such as the ICD-9-CM surgical coding library); variant examples generated by synonym substitution, colloquial rewriting, adding modifiers, or omitting some attributes of standard surgical descriptions; and examples sampled and annotated from real, desensitized surgical records on the front page of medical records. Supervised fine-tuning on this dataset allows the model to accurately grasp the task requirements of extracting target attributes from free text.

[0037] 2. Provide specific parameters and design details for the project: The CO-STAR framework is used to construct prompt word templates. Specific parameter examples are shown below: C (Context): "You are an expert system proficient in medical terminology standardization and surgical coding. Your task is to accurately extract structured attribute information from a non-standard surgical description." O (Objective): "Please identify and output the following four attributes: body_part (surgical site), method (surgical procedure), approach (surgical approach), and pathology (pathology of the disease). If any information cannot be inferred from the description, output an empty string for the corresponding field." S (Style, Constraint): "The output must be a strictly valid JSON object containing only the four specified keys. Key names must be in English. Do not add any explanatory text." T (Tone): "Ensure that the terminology used in the output is as accurate and standardized as possible." A (Audience): "The output will be directly parsed and used by downstream automation systems." R (Response): Provides 3-5 high-quality few-shot examples, covering scenarios with varying levels of complexity and completeness. For example: Input: "Laparoscopic cholecystectomy" Output: {"body_part":"Cholecyst","method":"Cholecystectomy","approach":"Laparoscopy","pathology":""} Input: "Perform a subtotal gastrectomy on the patient" Output: {"body_part":"stomach","method":"subtotal gastrectomy","approach":"open abdomen","pathology":""} Input: "Right lung nodule wedge resection" Output: {"body_part":"Right lung","method":"wedge resection","approach":"","pathology":"nodule"} Additionally, when generating the final answer, a lower temperature parameter (e.g., temperature=0.1) is set to reduce the randomness of the output and ensure the stability and consistency of attribute extraction. The three most semantically similar examples to the current input surgical description are dynamically inserted into the prompt words as samples for in-context learning, improving the model's analogical reasoning ability for the current task. Similarity is calculated by encoding the description and examples into the input sentence model.

[0038] 3. Output normalization and post-processing: When calling the model API, enable JSON output mode (such as OpenAI's response_format={"type":"json_object"}), or add an output parsing layer afterward to ensure that the returned content conforms to the predefined JSON structure. If parsing fails, trigger a retry mechanism or fallback processing.

[0039] In addition to extracting attribute values, the scalable prompts require the model to simultaneously output a confidence score for each attribute extraction. If the confidence score of any key attribute (such as location or procedure) is lower than a set threshold, or if the extraction result fails consecutively in the atlas mapping of subsequent steps, the decomposition is marked as a low-confidence result, and a specific anomaly handling process is activated.

[0040] S2: Obtain the structured attribute entity data, process it based on the standard terminology knowledge graph corresponding to each attribute, and generate standardized attribute data.

[0041] In specific implementations, standardized attribute data includes standard site terminology, standard surgical procedure terminology, standard approach terminology, and standard disease nature terminology.

[0042] First, four independent attribute terminology knowledge graphs are constructed, which respectively store the standard terminology set for the location, such as gallbladder and liver; the standard terminology set for the surgical procedure, such as resection and anastomosis; the standard terminology set for the approach, such as laparoscopic approach and open approach; and the standard terminology set for the nature of the disease, such as benign and malignant. Each graph contains terminology nodes and semantic relationships, such as synonyms and hierarchical relationships.

[0043] Then, for each attribute entity, a three-level mapping is performed in the corresponding knowledge graph: First, exact matching is performed, searching for completely identical strings in the standard terminology set; for example, inputting "gallbladder" matches "gallbladder". If exact matching fails, cosine similarity is calculated using the word vectors of the attribute entity and each candidate term in the standard terminology set. Terms with a cosine similarity exceeding the first threshold of 0.9 are selected; for example, inputting "cholecystectomy" might match "cholecystectomy surgery". If similarity matching fails, text edit distance similarity is calculated, and terms with a similarity exceeding the second threshold of 0.7 are selected; for example, inputting "laparoscopy" might match "laparoscopic approach".

[0044] Finally, the mapping results for each attribute are recorded, including standard terms, mapping type, and confidence score, generating standardized attribute data. For attributes that fail to map, the original text is retained and marked as unstandardized; for example, outputting "body_part" as gallbladder, "method" as resection, "approach" as laparoscopic approach, and "pathology" as null. This step, through a multi-level mapping strategy, accurately normalizes attribute entities to standard terms, improving the accuracy and recall of subsequent knowledge graph retrieval and reducing bias in text similarity calculation.

[0045] It should be noted that in this step, the first threshold of 0.9 and the second threshold of 0.7 are determined based on performance evaluation on the validation set and a balance with business requirements. The specific determination method is as follows: First, use a labeled validation set (containing matching pairs of "original description - standard procedure"). Run the system at different thresholds and calculate precision, recall, and F1 score. By plotting PR curves or F1-threshold curves, select the threshold point that maximizes the F1 score or meets specific business needs (such as prioritizing high precision).

[0046] The similarity score distributions for "correct matches" and "incorrect matches" in the validation set are calculated using distribution analysis. The ideal threshold should be located outside or at the point of least overlap between the two distributions to best distinguish between positive and negative examples. For example, if 95% of correct matches have a similarity score higher than 0.75, while 95% of incorrect matches have a similarity score lower than 0.65, then Th_sim can be initially set at around 0.7, and then fine-tuned according to business risk preferences.

[0047] Regarding the selection of the first threshold of 0.9 and the second threshold of 0.7, the optimal operating point is selected on the curve based on the specific task objective.

[0048] For the first threshold of 0.9: This is applied to edit distance similarity matching, aiming for extremely high precision, ensuring that the matched terms are almost certainly correct synonyms or abbreviations, even at the cost of some recall (missing some synonyms with significant spelling differences). This threshold is typically chosen on the PR curve near a position with high precision.

[0049] The second threshold of 0.7 is applied to word vector similarity matching or as a broad recall in the final candidate set. The goal is to ensure a high recall rate, to avoid missing potential relevant candidates as much as possible, and to tolerate a certain number of false positives, because there are more rigorous screening steps (such as spectral retrieval and large model inference) to correct for these errors.

[0050] Furthermore, all thresholds in this method can be designed as configurable hyperparameters. In actual deployment, system administrators can calibrate and adjust these thresholds based on the data characteristics of different medical institutions (such as differences in writing standards) or different application scenarios (such as medical insurance review scenarios requiring higher precision, and scientific research retrieval scenarios requiring higher recall) to maintain optimal system performance.

[0051] S3: Obtain the standardized attribute data, perform retrieval processing in the standard surgical knowledge graph, and generate retrieval result data.

[0052] In a specific implementation, the search result data includes a matching status identifier, and when the matching status identifier is multiple candidate matches or no match, it also includes a list of candidate surgical standard terms.

[0053] First, a standard surgical knowledge graph is constructed using a graph database such as Neo4j. Nodes include standard surgical terminology nodes, each node is associated with a standard surgical code and name, such as the code ICD-9-CM-51.23 and the name laparoscopic cholecystectomy, location nodes, surgical procedure nodes, approach nodes, and disease nature nodes. Edges represent the relationships, such as having location and surgical procedure.

[0054] Then, the standardized attribute data is converted into a graph query statement, such as a Cypher query. The query conditions require that at least one other attribute standard term is matched. For example, the query statement is: MATCHs is Surgery, s uses the procedure Method and the Method name is Resection, s has a body part and the body part name is Gallbladder, and returns the id and name of s.

[0055] Next, the query results are analyzed: If the query result for the attribute standard term uniquely identifies a standard surgical term node, the standard surgical code and name of that node are output as the final normalized result, and the matching status is set to exact match; if multiple nodes are found, the codes, names, and attribute information of all nodes are output as a candidate surgical standard term list, and the matching status is set to multiple candidate match; if the query result is empty, an empty list is output, and the matching status is set to no match. In this case, the query conditions can be relaxed, such as matching only surgical procedure terms, and a new query can be performed. Finally, the retrieval result data is output, including the matching status and candidate list, if applicable. For example, the status is output as multiple_candidates, and the candidates contain the code ICD-9-CM-51.23 and the name "laparoscopic cholecystectomy". This step utilizes the structured reasoning capabilities of the knowledge graph to deeply understand the connotation features of surgery, accurately retrieve candidate surgical entities, and reduce the bias caused by simple text similarity.

[0056] S4: Based on the matching status identifier in the search results data, when the matching status identifier is multiple candidate matches or no match, the candidate surgical standard terminology list and standardized attribute data are processed by vector similarity calculation to generate a candidate surgical standard terminology list with similarity scores.

[0057] In a specific implementation, firstly, for each surgical terminology node in the standard surgical knowledge graph, the associated standard location, surgical procedure, approach, and disease nature terms are concatenated into descriptive text in a fixed order, such as cholecystectomy, laparoscopic approach, and null value. A text embedding model such as Sentence-BERT is used to generate a 1024-dimensional descriptive vector, which is then stored in the vector database milvus. Then, the standard terms for each attribute in the standardized attribute data are concatenated into query text in the same order, and the same embedding model is used to generate query vectors.

[0058] Next, the cosine similarity between the query vector and each candidate surgical description vector is calculated, using the formula: A candidate term-similarity pairing list is generated. The pairing list is then sorted in descending order based on the similarity score, and a similarity threshold of 0.8 is applied for filtering, retaining candidate terms whose similarity exceeds the threshold.

[0059] Finally, the output includes a list of candidate surgical standard terms with similarity scores. Each list item includes the standard surgical code, name, and similarity score. For example, the output contains the code ICD-9-CM-51.23, the name "laparoscopic cholecystectomy," and a score of 0.85. This step uses vector similarity calculation to finely filter candidate surgical entities, compensating for the shortcomings of knowledge graph retrieval and improving the accuracy of normalization.

[0060] In this step, a Transformer-based sentence encoding model is used. Its core function is to map variable-length text sentences into a fixed-dimensional dense vector (e.g., 768-dimensional or 384-dimensional), ensuring that semantically similar sentences are close in distance within the vector space. Typical implementations include Siamese / triple network models based on BERT or RoBERTa architectures, such as Sentence-BERT or its variant all-MiniLM-L6-v2. These models are chosen because they are trained on large-scale corpora, possess excellent semantic representation capabilities, and offer efficient vector computation performance, making them particularly suitable for large-scale semantic similarity matching.

[0061] To ensure that the embedding vectors accurately capture subtle semantic differences in medical terminology (especially surgical-related terminology), this invention requires that the selected text embedding model be specifically trained in the medical field, taking into account the application scenario. Specifically, the following methods can be used: This approach directly utilizes sentence-encoded versions of embedding models pre-trained on biomedical literature (such as PubMed) and clinical texts, such as SapBERT, BioBERT, or ClinicalBERT. These models have an inherent advantage in representing medical entities and concepts.

[0062] If a general-domain sentence encoding model (such as all-MiniLM-L6-v2) is used, it needs to be fine-tuned through contrastive learning using medical text pairs. The training data consists of positive sample pairs (sentences with different expressions but referring to the same standard surgical procedure) and negative sample pairs (sentences referring to different standard surgical procedures). For example, ("laparoscopic cholecystectomy", "laparoscopic cholecystectomy") constitutes a positive sample pair, while ("laparoscopic cholecystectomy", "open subtotal gastrectomy") constitutes a negative sample pair. Through this training, the model can learn that "laparoscopic" and "open" are distinguishing features, while "gallbladder" and "stomach" are core entity features.

[0063] S5: Based on the candidate surgical standard terminology list with similarity scores, structured attribute entity data, and original surgical description data, normalized surgical standard terminology data is generated by constructing medical knowledge reasoning prompts and calling a large language model for processing.

[0064] In a specific implementation, a medical knowledge reasoning prompt template is first constructed, including a system role defined as "clinical surgical expert", a task description of selecting the most appropriate surgical standard term from a candidate list based on medical knowledge, context information integrating the original surgical description, structured attribute entity data, candidate surgical standard term list (with similarity scores) and definition description of each candidate entity, and an output format requirement of JSON format, including "selected_code" and "reasoning" fields.

[0065] Then, the candidate list is sorted in descending order of similarity score. The top N (e.g., top 5) candidate terms and their scores are selected and filled into the prompt word template. For example, the input data includes the original description "laparoscopic cholecystectomy", attribute entity data, and the candidate list [{code:"ICD-9-CM-51.23",name:"laparoscopic cholecystectomy",score:0.85}]. Next, the prompt word data is passed in by calling the API of the large language model to obtain the response output data. Finally, the response output data is parsed in a structured manner, the JSON format is verified, and the standard surgical code field is extracted. If the parsing is successful and the code exists in the candidate list, the standard surgical name corresponding to the code is output as the normalized surgical standard term data (e.g., "laparoscopic cholecystectomy"). If the parsing fails or the code does not exist, the candidate term with the highest similarity score is selected as the temporary result, and the manual review process is triggered (the reason is recorded as "LLM inference failed"). When large language models, knowledge graphs, and similarity calculations all fail, this step achieves the final decision through medical knowledge reasoning, ensuring the accuracy and interpretability of the normalized results and improving recall.

[0066] In this embodiment, a multi-layered surgical terminology normalization process is constructed by organically integrating the semantic understanding capabilities of a large language model, the structured reasoning capabilities of a knowledge graph, and the similarity matching technology of vector computation. This achieves accurate conversion from free text descriptions to standard terms. The method employs a progressive processing strategy, significantly improving the accuracy and recall of normalization processing through five closely interconnected stages: attribute decomposition, terminology mapping, knowledge graph retrieval, vector filtering, and intelligent reasoning. Its modular design ensures the system's maintainability and scalability, greatly reducing the need for manual intervention. The fully automated processing and traceability mechanism not only guarantee processing efficiency but also meet the stringent requirements of medical scenarios for process interpretability, providing a reliable technical foundation for medical information applications such as medical insurance review and clinical research.

[0067] like Figure 2As shown, the following are embodiments of the surgical terminology multi-level normalization system provided in this disclosure. This system and the surgical terminology multi-level normalization methods in the above embodiments belong to the same inventive concept. For details not described in detail in the embodiments of the surgical terminology multi-level normalization system, please refer to the embodiments of the surgical terminology multi-level normalization method described above.

[0068] A multi-level normalization system for surgical terminology includes: The attribute decomposition module is used to collect raw surgical description data and process it into structured attribute entity data through a large language model. The structured attribute entity data includes location entity, surgical procedure entity, approach entity, and disease nature entity.

[0069] The terminology standardization module is used to acquire the structured attribute entity data, process it based on the standard terminology knowledge graph corresponding to each attribute, and generate standardized attribute data. The standardized attribute data includes standard site terms, standard surgical procedure terms, standard approach terms, and standard disease nature terms.

[0070] The graph retrieval module is used to acquire the standardized attribute data, perform retrieval processing in the standard surgical knowledge graph, and generate retrieval result data. The retrieval result data includes a matching status identifier, and when the matching status identifier is multiple candidate matches or no match, it also includes a list of candidate surgical standard terms.

[0071] The similarity filtering module is used to generate a candidate surgical standard terminology list with similarity scores by processing the candidate surgical standard terminology list and standardized attribute data according to the matching status identifier in the search result data. When the matching status identifier is multiple candidate matches or no match, the module is used to process the data through vector similarity calculation.

[0072] The intelligent reasoning module is used to generate normalized surgical standard terminology data by constructing medical knowledge reasoning prompts and calling a large language model to process the candidate surgical standard terminology list with similarity scores, structured attribute entity data, and original surgical description data.

[0073] The surgical terminology multi-level normalization system provided in this embodiment firstly utilizes the deep semantic understanding capabilities of a large model to effectively decompose and standardize unstructured and complex surgical descriptions, ensuring the accuracy of the parsing from the source. Secondly, by combining precise retrieval from a knowledge graph with vectorized similarity calculation, it balances the rigor of structured rules with the flexibility of semantic matching, significantly improving the recall and precision of the normalization results. Finally, when traditional methods fail, a large model is introduced as an intelligent decision-maker, ensuring the method's final processing capability when dealing with marginal cases and ambiguous descriptions, thus forming a comprehensive, progressive, and self-verifiable efficient normalization scheme.

[0074] Figure 3 A schematic diagram of the hardware structure of an electronic device for implementing various embodiments of the present invention.

[0075] The surgical terminology multi-level normalization method provided in this application can be applied to electronic devices. Those skilled in the art will understand that the electronic device structure involved in the embodiments of this invention does not constitute a limitation on the electronic device. An electronic device may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. In the embodiments of this invention, electronic devices include, but are not limited to, laptop computers, desktop computers, workbenches, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. Electronic devices may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the embodiments of this application described and / or claimed herein.

[0076] Electronic devices may include processors, external memory interfaces, internal memory, universal serial bus (USB) interfaces, charging management modules, power management modules, batteries, wireless communication modules, audio modules, speakers, microphones, sensor modules, buttons, cameras, displays, and SIM card interfaces, etc.

[0077] A processor may include one or more processing units, such as: a central processing unit (CPU), an application processor (AP), a modem processor, a graphics processing unit (GPU), an image signal processor (ISP), a controller, memory, a video codec, a digital signal processor (DSP), a baseband processor, and / or a neural network processing unit (NPU). Different processing units may be independent devices or integrated into one or more processors.

[0078] The processor can serve as the nerve center and command center of an electronic device. The controller can generate operation control signals based on the instruction opcode and timing signals to control the fetching and execution of instructions.

[0079] The processor may also include memory for storing instructions and data. In some embodiments, the memory in the processor is a cache memory. This memory can store instructions or data that the processor has just used or that are used repeatedly. If the processor needs to use the instruction or data again, it can retrieve it directly from this memory. This avoids repeated accesses, reduces processor latency, and thus improves system efficiency.

[0080] An external storage interface (ESI) can be used to connect external memory cards, such as microSD cards, to expand the storage capacity of electronic devices. The external memory card communicates with the processor through the ESI to perform data storage functions, such as saving music and video files on the external memory card.

[0081] Internal memory can be used to store computer executable program code, which includes instructions. The processor executes various functional applications and data processing of electronic devices by running the instructions stored in internal memory. Internal memory can include a program storage area and a data storage area. Internal memory can include high-speed random access memory, and can also include non-volatile memory, such as at least one disk storage device, flash memory device, universal flash storage (UFS), etc.

[0082] Wireless communication functionality in electronic devices can be achieved through antennas, wireless communication modules, modem processors, and baseband processors.

[0083] Wireless communication modules can provide solutions for wireless communication applications in electronic devices, including wireless local area networks (WLANs) (such as wireless fidelity (Wi-Fi) networks), Bluetooth (BT), global navigation satellite system (GNSS), frequency modulation (FM), near field communication (NFC), and infrared (IR) technologies.

[0084] Electronic devices can implement audio functions through audio modules, speakers, receivers, microphones, headphone jacks, and application processors.

[0085] Electronic devices can achieve shooting functions through ISPs, cameras, video codecs, GPUs, displays, and application processors.

[0086] Electronic devices can achieve display functions through GPUs, displays, and application processors.

[0087] A GPU is a microprocessor for image processing, connected to the display screen and application processor. GPUs are used to perform mathematical and geometric calculations for graphics rendering. A processor may include one or more GPUs, which execute program instructions to generate or modify display information.

[0088] A display screen is used to display images, videos, etc. A display screen includes a display panel.

[0089] The aforementioned electronic device realizes the multi-level normalization method for surgical terminology of this application by organically combining the semantic understanding ability of large language models, the structured reasoning ability of knowledge graphs, vector similarity calculation and multi-level decision-making mechanisms. It achieves a comprehensive and beneficial effect of ensuring high-precision normalization while making the processing process interpretable, the system easy to maintain, highly automated, and fully leveraging the synergistic advantages of technology.

[0090] The storage medium provided in this application stores a program product capable of implementing a multi-level normalization method for surgical terminology.

[0091] Multi-level normalization methods for surgical terminology include: Raw surgical description data is collected and processed into structured attribute entity data through a large language model. The structured attribute entity data includes location entity, surgical procedure entity, approach entity, and disease nature entity. The structured attribute entity data is obtained and processed based on the standard terminology knowledge graph corresponding to each attribute to generate standardized attribute data. The standardized attribute data includes standard site terms, standard surgical procedure terms, standard approach terms, and standard disease nature terms. The standardized attribute data is obtained, and a search is performed in the standard surgical knowledge graph to generate search result data. The search result data includes a matching status identifier, and when the matching status identifier is multiple candidate matches or no match, it also includes a list of candidate surgical standard terms. Based on the matching status identifier in the search results data, when the matching status identifier is multiple candidate matches or no match, the candidate surgical standard terminology list and standardized attribute data are processed by vector similarity calculation to generate a candidate surgical standard terminology list with similarity scores. Based on the candidate surgical standard terminology list with similarity scores, structured attribute entity data, and original surgical description data, normalized surgical standard terminology data is generated by constructing medical knowledge reasoning prompts and processing them using a large language model.

[0092] In some possible implementations, the surgical terminology multi-level normalization method of this disclosure can be implemented as a program product comprising program code that, when run on a terminal device, causes the terminal device to perform the steps described in the "Exemplary Methods" section of this specification according to various exemplary embodiments of this disclosure.

[0093] The storage medium disclosed herein may be any combination of one or more readable media. A readable medium may be a readable signal medium or a readable storage medium. A readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples (a non-exhaustive list) of readable storage media include: an electrical connection having one or more wires, a portable disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof.

[0094] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for multi-level normalization of surgical terminology, characterized in that, include: Raw surgical description data is collected and processed into structured attribute entity data through a large language model. The structured attribute entity data includes location entity, surgical procedure entity, approach entity, and disease nature entity. The structured attribute entity data is obtained and processed based on the standard terminology knowledge graph corresponding to each attribute to generate standardized attribute data. The standardized attribute data includes standard site terms, standard surgical procedure terms, standard approach terms, and standard disease nature terms. The standardized attribute data is obtained, and a search is performed in the standard surgical knowledge graph to generate search result data. The search result data includes a matching status identifier, and when the matching status identifier is multiple candidate matches or no match, it also includes a list of candidate surgical standard terms. Based on the matching status identifier in the search results data, when the matching status identifier is multiple candidate matches or no match, the candidate surgical standard terminology list and standardized attribute data are processed by vector similarity calculation to generate a candidate surgical standard terminology list with similarity scores. Based on the candidate surgical standard terminology list with similarity scores, structured attribute entity data, and original surgical description data, normalized surgical standard terminology data is generated by constructing medical knowledge reasoning prompts and processing them using a large language model.

2. The method for multi-level normalization of surgical terminology according to claim 1, characterized in that, The collected raw surgical description data is processed into structured attribute entity data using a large language model, including: Construct an attribute decomposition prompt template based on the CO-STAR specification, where the role is defined as a medical information standardization expert, the target attribute list includes location, surgical procedure, approach and disease nature, and the output format is specified as JSON structure; The raw surgical description data and the prompt word template are combined and then input into the large language model to obtain the initial response data; Perform JSON parsing on the initial response data to extract the values ​​of the site entity field, surgical procedure entity field, approach entity field, and disease nature entity field; The parsed entity data is validated for integrity. The integrity validation is considered to be successful when there are at least two valid entities in the site entity field, surgical procedure entity field, approach entity field, and disease nature entity field. When validation passes, the system outputs structured attribute entity data containing valid entity data and marks missing fields as null values; when validation fails, the system outputs a disassembly exception flag and records the reason for the failure.

3. The method for multi-level normalization of surgical terminology according to claim 2, characterized in that, The process of acquiring the structured attribute entity data and processing it based on the standard terminology knowledge graph corresponding to each attribute to generate standardized attribute data includes: Four independent attribute terminology knowledge graphs were established, which were used to store standard terminology sets for site of origin, surgical procedure, approach, and disease nature, respectively. For each attribute entity in the structured attribute entity data, a three-level mapping process is performed in the corresponding attribute term knowledge graph. First, a string exact match is performed in the corresponding standard term set. If the exact match is successful, the mapping result is recorded. If the exact match fails, the word vectors of the attribute entity and each candidate term in the standard term set are calculated, and the cosine similarity is calculated. Standard terms with a cosine similarity exceeding the first threshold are selected through similarity matching. When similarity matching fails, text edit distance similarity is calculated, and standard terms with a text edit distance similarity exceeding the second threshold are selected. Record the mapping results for each attribute entity, including the selected standard terms, mapping type, and confidence score; Generate standardized attribute data containing all attribute standard terminology, and retain the original text for attributes that fail to be mapped and mark them as unstandardized.

4. The method for multi-level normalization of surgical terminology according to claim 3, characterized in that, The process of acquiring the standardized attribute data, performing retrieval processing within the standard surgical knowledge graph, and generating retrieval result data includes: A standard surgical knowledge graph is constructed, wherein the standard surgical knowledge graph uses standard surgical terminology nodes as nodes, each node is associated with a standard surgical code and a standard surgical name, and is connected to site nodes, surgical procedure nodes, approach nodes and disease nature nodes through edges; Convert standardized attribute data into a graph query statement and set four attribute standard terms as query conditions; Perform a graph query and analyze the query results; when the query results of the four attribute standard terms uniquely identify a standard surgical term node, output the standard surgical code and standard surgical name of that node as the final normalized result, and set the matching status flag to exact match; When multiple standard surgical terminology nodes are found, the standard surgical code, standard surgical name, and attribute information of each standard surgical terminology node are output as a candidate surgical standard terminology list, and the matching status is set to multiple candidate matches; when the query result is empty, an empty candidate list is output, and the matching status is set to no match.

5. The method for multi-level normalization of surgical terminology according to claim 4, characterized in that, The process involves calculating vector similarity based on the candidate surgical standard terminology list and standardized attribute data to generate a candidate surgical standard terminology list with similarity scores, including: A pre-built standard surgical terminology vector library is generated as follows: for each standard surgical terminology node, its associated standard site terms, standard procedure terms, standard approach terms, and standard disease nature terms are concatenated into descriptive text in a fixed order, and then converted into a descriptive vector using a text embedding model. The query text is constructed by concatenating the standard terms of each attribute in the standardized attribute data in a fixed order, and then converted into a query vector using a text embedding model. Calculate the cosine similarity between the query vector and the description vector of the candidate surgical standard term node as a similarity score, and generate a similarity pairing list of candidate terms; The similarity pairing list of candidate terms is sorted in descending order based on similarity score, and a similarity threshold of similarity_threshold is applied for filtering, retaining candidate terms with similarity scores exceeding similarity_threshold. Output a list of candidate surgical standard terms with similarity scores. Each list item includes the standard surgical code, standard surgical name, and similarity score.

6. The method for multi-level normalization of surgical terminology according to claim 5, characterized in that, The process involves generating normalized surgical standard terminology data based on a list of candidate surgical standard terms with similarity scores, structured attribute entity data, and original surgical description data. This is achieved by constructing medical knowledge reasoning prompts and processing them using a large language model. The data includes: A medical knowledge reasoning prompt template is constructed, which includes a system role definition section, a task description section, a context information section, and an output format section; the context information section is used to integrate a list of candidate surgical standard terms with similarity scores, structured attribute entity data, and raw surgical description data; Based on the candidate surgical standard terminology list, sorted in descending order of similarity score, select the top N candidate terms and their scores, fill them into the corresponding positions in the prompt word template, and generate prompt word data; Call the large language model API interface, pass in the prompt word data, and obtain the response output data of the large language model; The response output data is structured and parsed to verify whether the output format conforms to the agreed JSON specification, and the standard surgical code field is extracted. When parsing is successful and the standard surgical code exists in the list of candidate surgical standard terms, the standard surgical name corresponding to the code is output as normalized surgical standard terminology data; when parsing fails or the standard surgical code does not exist, the standard surgical name of the candidate term with the highest similarity score is selected as a temporary result, and a manual review process is triggered.

7. The method for multi-level normalization of surgical terminology according to claim 3, characterized in that, The first threshold is 0.9, and the second threshold is 0.

7.

8. A multi-level normalization system for surgical terminology, characterized in that, The system employs the surgical terminology multi-level normalization method as described in any one of claims 1 to 7; The system includes: The attribute decomposition module is used to collect raw surgical description data and process it into structured attribute entity data through a large language model. The structured attribute entity data includes location entity, surgical procedure entity, approach entity, and disease nature entity. The terminology standardization module is used to acquire the structured attribute entity data, process it based on the standard terminology knowledge graph corresponding to each attribute, and generate standardized attribute data. The standardized attribute data includes standard site terms, standard surgical procedure terms, standard approach terms, and standard disease nature terms. The graph retrieval module is used to obtain the standardized attribute data, perform retrieval processing in the standard surgical knowledge graph, and generate retrieval result data. The retrieval result data includes a matching status identifier, and when the matching status identifier is multiple candidate matches or no match, it also includes a list of candidate surgical standard terms. The similarity filtering module is used to generate a candidate surgical standard terminology list with similarity scores by processing the candidate surgical standard terminology list and standardized attribute data according to the matching status identifier in the search result data. When the matching status identifier is multiple candidate matches or no match, the module is used to process the data through vector similarity calculation based on the candidate surgical standard terminology list and standardized attribute data. The intelligent reasoning module is used to generate normalized surgical standard terminology data by constructing medical knowledge reasoning prompts and calling a large language model to process the candidate surgical standard terminology list with similarity scores, structured attribute entity data, and original surgical description data.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the surgical terminology multi-level normalization method as described in any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the surgical terminology multi-level normalization method as described in any one of claims 1 to 7.

Images