Medical clinical data coding method based on large language model and artificial code multiplexing

Abstract

The invention discloses a medical clinical data coding method based on a large language model and artificial code multiplexing. The medical clinical data coding method comprises the following steps: acquiring medical clinical original data to be coded and identifying a coding type; automatically generating a data cleaning dynamic template, and inputting a data cleaning instruction into the large language model to obtain cleaned data; calling a preset automatic coding rule base to execute automatic coding, wherein the automatic coding rule base comprises matching rules sorted according to priorities; if the automatic coding is not hit, entering a manual coding process and entering a synonym library for semantic association matching of subsequent automatic coding; after the medical coding dictionary is updated, a preset upgrading rule base is called to automatically process coded data, data to be upgraded are recognized, a coding upgrading suggestion is generated in combination with a retrieval enhancement generation technology, and data upgrading is achieved; and exporting a coding result. According to the invention, the efficiency and accuracy of medical clinical data coding can be improved, and manpower and time costs are reduced.

Description

Technical Field

[0001] This invention relates to the field of medical data coding technology, and in particular to a medical clinical data coding method based on the reuse of large language models and manual coding. Background Technology

[0002] Medical coding is a core component of the standardized processing of clinical research data. Its core function is to transform the diverse medical information generated in clinical trials, including adverse event names, investigational drug names, patient symptoms and signs, clear disease diagnoses, past medical history, and contraindications and indications of investigational drugs, into standardized coding terms. This eliminates data confusion caused by differences in the descriptions of different recorders and lays the foundation for subsequent data statistical analysis, cross-study data comparison, and international multi-center research collaboration.

[0003] Throughout the entire clinical research process, medical coding is characterized by both complexity and precision, and is crucial for ensuring research quality. On the one hand, by standardizing terminology, it ensures that the recording, transmission, and aggregation of medical information follow fixed rules, promoting the standardization of clinical work. On the other hand, it requires staff to be proficient in coding classification systems such as MedDRA (International Dictionary of Medical Terms) and WhoDrug (World Health Organization Drug Dictionary), verifying original medical records for complex medical information to ensure accurate correspondence between codes and actual information, thus guaranteeing data quality and reliability. At the same time, standardized coding enables my country's clinical research data formats to align with internationally accepted standards such as ICH-GCP, facilitating the internationalization of medical management.

[0004] Current mainstream medical coding technology solutions are based on a core logic of automation combined with manual completion and cyclical review. Technicians use professional tools such as SAS to extract raw data from the electronic data acquisition system, call MedDRA or WhoDrug standard dictionaries to perform automatic term matching, and output two types of results: successful match and unsuccessful match. For unsuccessful matches, professional medical coders consult clinical trial protocols, medical record reports, and other materials to manually screen and determine standard codes. The two types of code results are then merged and submitted for review. If they fail, they are returned for correction. When the dictionary version is updated or new data is added, the entire process must be repeated.

[0005] However, existing technologies have significant drawbacks, making it difficult to meet the needs of efficient clinical research: First, the automatic coding hit rate is low. Raw data often contains blank characters, redundant number annotations, or date information due to input habits. Furthermore, medical terminology has many synonyms, and there are differences between colloquial and professional expressions. Existing technologies can only achieve literal matching and cannot deeply understand the semantics, requiring a large amount of manual preprocessing. Second, coding results cannot be reused. There is a lack of a dedicated knowledge accumulation mechanism. When the same unmatched term reappears in different clinical trials or different stages of the same trial, the coding process must be repeated, resulting in redundant work. Third, dictionary version updates are difficult to adapt to. MedDRA and WhoDrug dictionaries are updated around March and September each year. Existing technologies cannot automatically associate new and old versions of terms and codes, requiring manual review of all historical data, which incurs huge learning costs and repetitive workloads. Fourth, multilingual support is lacking. International multicenter projects require multilingual coding results, but existing technologies can only output the dictionary's default language, requiring manual conversion, which is prone to errors and delays.

[0006] Therefore, developing a medical clinical data encoding method that takes into account intelligent data cleaning, reuse of encoding results, and efficient dictionary updates has become an urgent technical problem to be solved in this field. Summary of the Invention

[0007] The problem to be solved by this invention is to provide a medical clinical data encoding method based on large language models and manual coding reuse, so as to overcome the shortcomings of existing medical clinical data encoding, which suffer from low hit rate, inability to reuse encoding results, and difficulty in adapting to dictionary version updates, resulting in a large amount of manual workload.

[0008] The technical solution adopted by this invention to solve its technical problem is: a medical clinical data encoding method based on the reuse of large language models and manual coding, comprising the following steps: S1, Obtain the raw medical clinical data to be encoded, and identify the encoding type of the raw data; S2, automatically generate a dynamic template for data cleaning based on the identified encoding type, generate data cleaning instructions by combining predefined prompt words, and input the data cleaning instructions into a large language model to perform normalization processing on the original data to obtain cleaned data; S3, invoke a preset automatic encoding rule library, which contains matching rules sorted by priority, and perform automatic encoding based on the encoding type and the cleaned data; S4. If the automatic coding fails, the manual coding process will begin. The standard coding and coding basis determined by the manual coding process will be obtained. The unmatched terms corresponding to the automatic coding failure will be entered into the thesaurus and the coding type and applicable dictionary version will be marked for semantic association matching in subsequent automatic coding. S5: When the medical coding dictionary is updated, the pre-set upgrade rule base is called to automatically process the coded data, identify the data that needs to be upgraded, and generate coding upgrade suggestions for the data to be upgraded by combining "Retrieval-Augmented Generation" (RAG) technology, thereby realizing data upgrade; S6. Set filtering conditions according to user needs, filter target encoded data, generate and export the encoding result corresponding to the target encoded data.

[0009] As a further improvement of the present invention, the encoding type includes MedDRA encoding type and WhoDrug encoding type; in step S1, when importing the original data to be encoded, the data type is automatically determined to be MedDRA encoding type or WhoDrug encoding type by reading the core field information of the original data; wherein, the core field information includes at least one of the following: generic name of drug, drug dosage form, drug manufacturer, adverse event name and disease diagnosis.

[0010] As a further improvement of the present invention, in step S2, the data cleaning dynamic template is used to clarify the standard format of different types of data. The large language model can automatically remove redundant information in the original data based on the data cleaning instructions, and at the same time correct irregular medical expressions in the original data into expressions that conform to the standard medical dictionary specifications, so as to obtain cleaned data with uniform format and clear semantics.

[0011] As a further improvement of the present invention, in step S3, the matching rules of the automatic encoding rule base are arranged in descending order of priority as precise matching rules, semantic association matching rules, and fuzzy association matching rules. When performing automatic encoding, the highest priority matching rule is used first to match the cleaned data: if the match is successful, the automatic encoding success result is output; if the match fails, the system automatically switches to the next lower priority matching rule and re-executes the matching until all priority matching rules have been executed.

[0012] As a further improvement of the present invention, the automatic encoding rule base is also configured with a matching detail library: the matching detail library corresponding to the precise matching rule includes a terminology literal consistency judgment standard, the matching detail library corresponding to the semantic association matching rule includes a common medical synonym comparison table, and the matching detail library corresponding to the fuzzy association matching rule includes terminology core word matching technology.

[0013] As a further improvement of the present invention, in step S4, the manual coding process includes coding lifecycle management. The coding lifecycle management is triggered when entering the manual coding process. By performing full lifecycle management on the coding data, the creation time, update records, and applicable scenarios of each piece of coding data are recorded, making the coding data traceable.

[0014] As a further improvement of the present invention, the coding lifecycle management includes a pending coding state, a coding in progress state, a pending review state, and a reviewed state. When data initially enters the manual coding process, it is in the pending coding state. After the coder claims the data, it switches to the coding in progress state. After the coder completes the manual coding and submits it, the data status switches to the pending review state. After the reviewer confirms that the coding is correct, the data status switches to the reviewed state, and the data is simultaneously entered into the thesaurus, completing the manual coding. If the review fails, the data status reverts to the pending coding state.

[0015] As a further improvement of the present invention, in step S5, the upgrade rule base includes the following predefined rules: if the level, content or code of the term coding data information in the old version of the medical coding dictionary changes, the term is automatically identified as data to be upgraded; a dictionary change knowledge base is constructed and RAG is used in conjunction with a large language model to efficiently generate coding upgrade suggestions.

[0016] As a further improvement of the present invention, in step S6, when generating the encoding result, the pre-established multilingual medical coding dictionary association data can be called according to the user's language requirements to generate the encoding result in the corresponding language.

[0017] As a further improvement of the present invention, in step S6, the filtering conditions include one or more of data type, data encoding status, and time range; the exported content of the encoding result includes one or more of the original terminology, standard encoding, encoding type, lifecycle status, and export time; the exported format of the encoding result includes Excel format or CSV format.

[0018] The beneficial effects of this invention are as follows: This invention provides a medical clinical data coding method based on the reuse of large language models and manual coding. Through the design of a complete process including coding type identification, intelligent data cleaning using large language models, priority automatic coding, reuse of manual coding, rule-based upgrade strategies, and large language model-assisted dictionary version updates and on-demand filtering and export, it effectively solves the core pain points of existing technologies such as low automatic coding hit rate, inability to reuse coding results, and reliance on a large amount of manual work for dictionary updates. It improves the standardization of data by leveraging large language models, realizes the accumulation and reuse of manual coding knowledge through a thesaurus, and reduces version update and maintenance costs by combining large language models to generate upgrade coding suggestions. Ultimately, it significantly improves the efficiency and accuracy of medical clinical data coding, reduces manpower and time costs, and provides an innovative solution for medical clinical data coding that combines high efficiency, accuracy, and universality. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly described below. Obviously, the drawings described below are merely some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without any creative effort.

[0020] Figure 1 This is a flowchart illustrating the steps of the medical clinical data encoding method based on the reuse of large language models and manual coding in this invention. Figure 2 This is a schematic diagram illustrating the process of processing raw data based on a large language model according to the present invention; Figure 3 This is a schematic diagram of the synonym encoding process in this invention; Figure 4 This is a schematic diagram of the coding data upgrade suggestion process in this invention; Figure 5 This is a schematic diagram of the multilingual data result export process in this invention. Detailed Implementation

[0021] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. This application can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] Additionally, specific details are provided in the following description to facilitate a thorough understanding of the examples. Those skilled in the art will understand that the practice can be carried out without these specific details.

[0023] The technical solutions provided by the various embodiments of this application are described below with reference to the accompanying drawings.

[0024] See Figures 1 to 5 This invention provides a medical clinical data encoding method based on large language models and manual coding reuse, comprising the following steps: S1: Obtain the raw medical clinical data to be encoded and identify the encoding type of the raw data.

[0025] S2 automatically generates a dynamic template for data cleaning based on the identified encoding type, generates data cleaning instructions by combining predefined prompt words, and inputs the data cleaning instructions into a large language model to standardize the original data and obtain cleaned data.

[0026] S3 calls the preset automatic encoding rule library, which contains matching rules sorted by priority, and performs automatic encoding based on the encoding type and the cleaned data.

[0027] S4. If the automatic coding fails, the manual coding process begins. The standard code and coding basis determined by the manual coding are obtained. The unmatched terms corresponding to the automatic coding failure are entered into the thesaurus and marked with their coding type and applicable dictionary version for semantic association matching in subsequent automatic coding. When a medical term to be coded that matches the unmatched term is subsequently obtained, the corresponding standard code is directly retrieved from the thesaurus to complete the automatic coding of the medical term to be coded.

[0028] S5, when the medical coding dictionary is updated, calls the pre-set upgrade rule base to automatically process the coded data, identify the data that needs to be upgraded, and generate coding upgrade suggestions for the data to be upgraded in combination with RAG, thereby realizing the data upgrade.

[0029] S6 allows users to set filtering conditions according to their needs, filter target encoded data, and generate and export the encoding results corresponding to the target encoded data.

[0030] This invention effectively addresses the core pain points of existing technologies, such as low automatic coding hit rate, unusable coding results, and reliance on extensive manual intervention for dictionary updates, by integrating coding type identification, intelligent data cleaning using a large language model, priority-based automatic coding, manual coding reuse, rule-based upgrade strategies, and a large language model-assisted dictionary version update and on-demand filtering and export process. It leverages a large language model to improve data standardization, utilizes a thesaurus to achieve knowledge accumulation and reuse of manually coded data, and combines the large language model to generate upgrade coding suggestions to reduce version update and maintenance costs. Ultimately, it significantly improves the efficiency and accuracy of medical clinical data coding, reduces manpower and time costs, and provides an innovative solution for medical clinical data coding that combines high efficiency, accuracy, and universality.

[0031] Combining theoretical analysis and practical verification, this method not only significantly improves the automatic coding hit rate by more than 10% and the reuse rate of encoded data by more than 70%, but also reduces the manual workload during dictionary updates by 50%, greatly reducing the repetitive work of technical personnel and coders, and lowering labor and time costs.

[0032] Optionally, the large language model described in this invention is a pre-trained language model with the ability to understand the semantics of medical terms, such as DeepSeek, ChatGPT, and Tongyi Qianwen.

[0033] The encoding types include MedDRA encoding type and WhoDrug encoding type.

[0034] In step S1, when importing the raw data to be encoded, the core field information of the raw data is read to automatically determine whether the data type is MedDRA or WhoDrug. The core field information includes at least one of the following: drug generic name, drug dosage form, drug manufacturer, adverse event name, and disease diagnosis. WhoDrug data includes drug generic name, drug dosage form, and drug manufacturer, while MedDRA data includes adverse event name and disease diagnosis. By reading the core field information, the data type can be automatically determined to be WhoDrug encoded data or MedDRA encoded data. This provides accurate type matching for subsequent data cleaning and automatic encoding processes, ensuring that subsequent processes are tailored to the characteristics of different encoding types, thus laying the foundation for the efficient advancement of the entire encoding method.

[0035] like Figure 2As shown, in step S2, the dynamic data cleaning template is used to clarify the standardized format of different types of data. For example, redundant information such as numerical values, units, and dates in adverse event terms need to be removed, while retaining the core symptom descriptions. By combining predefined prompts (such as identifying and deleting blank characters, date, and number annotations in terms, and uniformly correcting "tooth" to "teeth"), the large language model is guided to deeply understand the semantics of medical terms, rather than just processing them literally. Based on the data cleaning instructions, the large language model can automatically remove redundant information from the original data (such as processing "2024-06-01 fever 39℃" to "fever"), while correcting irregular medical expressions in the original data to expressions that conform to the standard medical dictionary specifications (such as correcting "diarrhea" to "diarrhea"), resulting in cleaned data with a unified format and clear semantics.

[0036] This invention clarifies the standardized format of different types of data through dynamic data cleaning templates and executes data cleaning instructions using a large language model. This achieves automatic removal of redundant information in the original data and standardized correction of irregular medical expressions. It effectively solves the problem of low automatic coding hit rate caused by chaotic original data format and inconsistent expression in the prior art. The cleaned data has a unified format and clear semantics, which improves the matching success rate of subsequent automatic coding from the source, reduces the workload of manual preprocessing, and reduces the risk of deviation caused by human operation.

[0037] In step S3, the matching rules in the automatic coding rule base are ordered from highest to lowest priority as follows: exact matching rules, semantic association matching rules, and fuzzy association matching rules. During automatic coding, rules with higher priority are executed first, and only one matching rule applies to the same data. Once a rule is successfully executed, subsequent rule matching stops. Specifically, the highest priority matching rule, i.e., the exact matching rule, is used first to match the cleaned data: if the match is successful, an automatic coding success result is output; if the match fails, the system automatically switches to the next lower priority matching rule and re-executes the matching until all priority matching rules have been executed; if all rules fail to match, the manual coding process begins.

[0038] This step prioritizes the matching rules in the auto-coding rule base according to the logic of "precise matching rules take precedence over semantic association matching rules, and semantic association matching rules take precedence over fuzzy association matching rules." During encoding, the rules are matched sequentially according to priority, ensuring that the accuracy of the encoding matching is prioritized and avoiding encoding errors caused by prioritizing low-precision rules. At the same time, the hierarchical matching logic maximizes the coverage of scenarios that can be automatically encoded, reduces the probability of automatic encoding misses, and further reduces the workload of manual encoding. This improves the efficiency of automatic encoding while ensuring encoding quality.

[0039] It is worth mentioning that the automatic coding rule base is also configured with a matching detail library. Specifically, the matching detail library for exact matching rules includes a literal consistency judgment standard for terms, which includes judgment logic that ignores differences in capitalization and full-width / half-width character differences. The matching detail library for semantic association matching rules includes a table of common medical synonyms, which is dynamically updated based on historical data stored in the thesaurus. The matching detail library for fuzzy association matching rules includes a matching technology based on core terms. This invention, by configuring corresponding matching detail libraries for the automatic coding rule base, clarifies the literal consistency judgment standard for exact matching, the table of common medical synonyms for semantic association matching, and the core term matching technology for fuzzy association matching. This makes automatic coding an ordered decision-making process, rather than a simple fuzzy matching. This design significantly improves the accuracy and coverage of automatic coding, reduces the amount of data flowing into the manual coding stage, and reduces workload.

[0040] In step S4, the manual coding process also includes coding lifecycle management. When all matching rules in the automatic coding rule base fail to match in step S3, the data enters the manual coding process, which triggers coding lifecycle management. By managing the entire lifecycle of the coded data, the creation time, update records, and applicable scenarios of each piece of coded data are recorded, making the coded data traceable. This effectively solves the problems of chaotic and untraceable coded data management in existing technologies, ensuring that the source and change process of each piece of coded data are clearly traceable, facilitating subsequent data verification and maintenance. At the same time, it provides accurate scenario references for the reuse of coded data, ensuring the accuracy of reused codes.

[0041] Specifically, the coding lifecycle management includes four states: pending coding, coding in progress, pending review, and reviewed. Data initially enters the manual coding process in the pending coding state; after a coder claims the data, it switches to the coding in progress state. After the coder manually matches the data to the standard dictionary (WhoDrug / MedDRA) in the system, the data state switches to the pending review state. Once the reviewer confirms the coding is correct, the data state switches to the reviewed state, and the data is entered into the thesaurus, completing the manual coding. If the review fails, the data state reverts to the pending coding state.

[0042] This step clarifies the specific status transition logic of the coding lifecycle management, including the pending coding status, coding in progress status, pending review status, and reviewed status. It clearly defines the responsibilities of coders and reviewers and the data flow path, making the manual coding process standardized and transparent. This avoids coding errors caused by data omissions, duplicate processing, or missing review processes, and ensures that each step of manual coding has a clear status label and responsibility attribution. While improving the accuracy of manual coding, it also reduces process communication costs and improves the efficiency of manual coding.

[0043] like Figure 3 As shown, based on the manually encoded "word-by-word terms"—that is, the original unmatched terms that the automatic encoding failed to find—coders identify the standard synonyms corresponding to these terms from candidate words recommended by the system or manually entered. After the coder confirms the synonym, it is marked as "synonym pending review" and transferred to the reviewer role. The reviewer conducts a compliance review of the associated data, focusing on verifying the "semantic consistency between the synonym and the word-by-word term" and the "accuracy of matching the synonym with the standard encoding," to determine whether the review passes. If the review fails, it returns to the coder to reconfirm the synonym and initiates the review process again; if the review passes, it is added to the current dictionary version's synonym list.

[0044] like Figure 4 As shown, during dictionary version iteration, to alleviate the problems of low efficiency, high manual verification costs, and delayed version change adaptation in the traditional encoding upgrade mode, RAG is introduced to build an intelligent assisted encoding upgrade system. Through the entire encoding-assisted upgrade process of "rule filtering - suggestion construction - vector retrieval - intelligent generation - manual verification," the system achieves accurate and efficient encoding upgrades. The specific implementation process in step S5 is as follows: First, the built-in rule engine is activated, and the full set of coded data is automatically scanned based on the preset verification rule base to identify the target coded data that needs to be upgraded. The rule base verification includes, but is not limited to: checking whether there are any changes to the text description, code identifier and hierarchical structure of the low-level term (LLT) and its associated parent terms at all levels corresponding to the old version code; and whether the drug name, drug code and selected ATC of the old term code have been updated in the new version.

[0045] Subsequently, based on the located code data to be upgraded, it is dynamically integrated with the preset standardized prompt word template to generate structured prompt words. The template must clearly include the code identifier to be upgraded, the source dictionary version, the target dictionary version, and the core requirements (such as "to upgrade code 'xxx' from dictionary version A to dictionary version B, accurate code upgrade suggestions and change basis must be provided"), to ensure the clarity of the prompt words and the completeness of the information.

[0046] Following this, the structured prompts are converted into high-dimensional vector data using a pre-trained text embedding model (such as Sentence-BERT). The vector data is then input into a vector retrieval engine, where cosine similarity is calculated with the vector indices of historical change records, version difference descriptions, and encoding mapping relationships pre-stored in the dictionary change database. The top-K reference data with the highest semantic relevance (such as historical upgrade cases of similar codes and encoding rule descriptions of the target version) are retrieved, and a structured context is constructed based on the relevance weights.

[0047] Finally, the structured prompts and weighted contextual information are input into a large language model (e.g., Tongyi Qianwen). The model combines the version change logic and encoding mapping rules in the contextual information to generate accurate encoding upgrade suggestions. The suggestions must include key information such as the target encoding, the basis for the change, and the hierarchical attribute adaptation status (e.g., "Encoding 'xxx' has been updated to 'XXX' in dictionary version B, the basis for the change is 'XXX' specific upgrade information: XXX"). After generation, the coder verifies the compliance of the upgrade suggestions based on their professional knowledge and business scenarios. If the verification passes, the "use this encoding" operation can be triggered directly to complete the upgrade; if the verification fails, the prompts can be adjusted based on the verification opinions (e.g., supplementing business constraints) and the retrieval and generation process can be re-executed, or manual secondary encoding can be performed directly, forming a closed-loop control of "machine assistance - manual verification".

[0048] Furthermore, existing medical clinical data coding suffers from a lack of multilingual support. International multicenter projects require multilingual coding results, but current technology can only output the default dictionary language, necessitating manual secondary conversion, which is error-prone and delays progress.

[0049] like Figure 5 As shown, to address the problem that existing technologies cannot output multilingual results, this invention embeds a multilingual dictionary association mechanism into the encoding process. Specifically, the system pre-establishes term-code correspondences for different language versions of the dictionary (e.g., Chinese version of MedDRA, English version of MedDRA). When generating encoding results, the system can call the pre-established multilingual medical coding dictionary association data according to the user's language requirements to generate the corresponding language encoding results. For example, it can simultaneously output "headache" and its corresponding codes. This design can meet the multilingual data requirements of international multicenter clinical trials, eliminating the need for manual secondary conversion of the encoding results, avoiding errors and time losses caused by manual conversion, ensuring the accuracy and consistency of multilingual encoding results, and enhancing the applicability and competitiveness of this invention in international scenarios.

[0050] In step S6 of this invention, the filtering conditions include one or more of data type, data encoding status, and time range; the exported content of the encoding result includes one or more of the original terminology, standard code, encoding type, lifecycle status, and export time; the exported format of the encoding result includes Excel or CSV format. This step, by clearly defining the specific types of filtering conditions, the exported content of the encoding result, and the export format, allows users to filter target data according to their actual needs, improving the practicality and convenience of the encoding result. Users can directly obtain accurate data after filtering as needed without additional manual processing. The universal export format facilitates subsequent data statistical analysis and sharing, further reducing the subsequent cost of data use and improving the user-friendliness of the technology.

[0051] The same or similar parts between the various embodiments in this specification can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments.

[0052] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A medical clinical data coding method based on large language models and manual coding reuse, characterized in that, Includes the following steps: S1, Obtain the raw medical clinical data to be encoded, and identify the encoding type of the raw data; S2, automatically generate a dynamic template for data cleaning based on the identified encoding type, generate data cleaning instructions by combining predefined prompt words, and input the data cleaning instructions into a large language model to perform normalization processing on the original data to obtain cleaned data; S3, invoke a preset automatic encoding rule library, which contains matching rules sorted by priority, and perform automatic encoding based on the encoding type and the cleaned data; S4. If the automatic coding fails, the manual coding process will begin. The standard coding and coding basis determined by the manual coding process will be obtained. The unmatched terms corresponding to the automatic coding failure will be entered into the thesaurus and the coding type and applicable dictionary version will be marked for semantic association matching in subsequent automatic coding. S5: When the medical coding dictionary is updated, the pre-set upgrade rule base is called to automatically process the coded data, identify the data that needs to be upgraded, and generate coding upgrade suggestions for the data to be upgraded by combining the "retrieval enhancement generation technology" to achieve data upgrade. S6. Set filtering conditions according to user needs, filter target encoded data, generate and export the encoding result corresponding to the target encoded data.

2. The medical clinical data encoding method based on large language model and manual coding reuse according to claim 1, characterized in that, The encoding types include MedDRA encoding type and WhoDrug encoding type; in step S1, when importing the original data to be encoded, the data type is automatically determined to be MedDRA encoding type or WhoDrug encoding type by reading the core field information of the original data; wherein, the core field information includes at least one of the following: generic name of drug, drug dosage form, drug manufacturer, adverse event name and disease diagnosis.

3. The medical clinical data encoding method based on large language model and manual coding reuse according to claim 1, characterized in that, In step S2, the data cleaning dynamic template is used to clarify the standard format of different types of data. The large language model can automatically remove redundant information in the original data based on the data cleaning instructions, and at the same time correct irregular medical expressions in the original data into expressions that conform to the standard medical dictionary specifications, so as to obtain cleaned data with uniform format and clear semantics.

4. The medical clinical data coding method based on large language model and manual coding reuse according to claim 1, characterized in that, In step S3, the matching rules in the automatic encoding rule base are ranked from high to low as follows: exact matching rules, semantic association matching rules, and fuzzy association matching rules. When performing automatic encoding, the highest priority matching rule is used first to match the cleaned data. If the match is successful, the automatic encoding success result is output. If the match fails, the system automatically switches to the next lower priority matching rule and re-executes the matching until all priority matching rules have been executed.

5. The medical clinical data coding method based on large language model and manual coding reuse according to claim 4, characterized in that, The automatic encoding rule base is also configured with a matching detail base; the matching detail base corresponding to the precise matching rule includes a terminology literal consistency judgment standard, the matching detail base corresponding to the semantic association matching rule includes a common medical synonym comparison table, and the matching detail base corresponding to the fuzzy association matching rule includes terminology core word matching technology.

6. The medical clinical data encoding method based on large language model and manual coding reuse according to claim 1, characterized in that, In step S4, the manual coding process includes coding lifecycle management. The coding lifecycle management is triggered when entering the manual coding process. By managing the entire lifecycle of the coding data, the creation time, update records, and applicable scenarios of each piece of coding data are recorded, making the coding data traceable.

7. The medical clinical data coding method based on large language model and manual coding reuse according to claim 6, characterized in that, The coding lifecycle management includes four states: pending coding, coding in progress, pending review, and reviewed. When data first enters the manual coding process, it is in the pending coding state. After the coder claims the data, it switches to the coding in progress state. After the coder completes the manual coding and submits it, the data status switches to the pending review state. After the reviewer confirms that the coding is correct, the data status switches to the reviewed state, and the data is entered into the thesaurus to complete the manual coding. If the review fails, the data status returns to the pending coding state.

8. The medical clinical data coding method based on large language model and manual coding reuse according to claim 1, characterized in that, In step S5, the upgrade rule base includes the following predefined rules: if the level, content, or code of the term coding data information in the old version of the medical coding dictionary changes, the term is automatically identified as data to be upgraded; while constructing a dictionary change knowledge base, the "retrieval enhancement generation technology" is used in conjunction with a large language model to efficiently generate coding upgrade suggestions.

9. The medical clinical data encoding method based on large language model and manual coding reuse according to claim 1, characterized in that, In step S6, when generating the encoding result, the pre-established multilingual medical coding dictionary association data can be called according to the user's language requirements to generate the encoding result in the corresponding language.

10. The medical clinical data encoding method based on large language model and manual coding reuse according to claim 1, characterized in that, In step S6, the filtering conditions include one or more of data type, data encoding status, and time range; the exported content of the encoding result includes one or more of the original terminology, standard encoding, encoding type, lifecycle status, and export time; the export format of the encoding result includes Excel format or CSV format.

Images