An automatic medical coding normalization method and system for multi-modal medical data and a storage medium

Abstract

This invention discloses an automatic medical coding normalization method, system, and storage medium for multimodal medical data, belonging to the field of medical information technology. The method includes: collecting original textual terms and subject structured information of adverse events and / or concomitant medications, and performing text denoising, format standardization, and data deduplication; generating an initial candidate coding set based on a pre-trained medical language model; performing context-aware semantic disambiguation and rearranging of candidate codes in conjunction with subject structured information; retrieving historically confirmed coding cases from other research centers during the coding confirmation stage, outputting consistency prompts based on the differences between the current context and historical contexts, and receiving reconfirmation, coding modification, or review submission operations; and storing the final confirmation results and retrieving them to an electronic data acquisition system. This invention can improve the accuracy, processing efficiency, traceability, and multi-center coding consistency of medical coding.

Description

Technical Field

[0001] This invention belongs to the field of medical information technology, specifically relating to an automatic medical coding normalization method, system, and storage medium for multimodal medical data. Background Technology

[0002] In clinical research, pharmacovigilance, and medical data management, it is often necessary to medically code raw records such as adverse events and concomitant medications. This maps the free-text terms entered by researchers to standardized terms and corresponding codes in standard dictionaries, such as the standard codes in the Medical Dictionary of Regulatory Activities (MedDRA) and the World Health Organization Drug Dictionary (WHO Drug). These coding results typically serve as a crucial foundation for subsequent data cleaning, statistical analysis, and regulatory reporting, thus placing high demands on coding efficiency, accuracy, and consistency.

[0003] Current medical coding methods primarily rely on manual or semi-automatic retrieval, where coders match keywords from the original text terms in a dictionary and then manually select the appropriate standard terms. This approach is highly dependent on coders' experience and is inefficient when dealing with large volumes of clinical trial data. Furthermore, existing retrieval-based coding typically focuses on matching the original text terms themselves, failing to adequately consider contextual information such as the subject's pregnancy status, gender, medical history, concomitant medications, and examination indicators to differentiate candidate codes across different clinical scenarios. Consequently, it lacks sufficient support for differentiated coding of the same or similar descriptions in different clinical settings, potentially affecting the accuracy of recommendation results.

[0004] Furthermore, in multi-center clinical research scenarios, when different research centers and coders confirm the coding of the same or similar original text terms, existing technologies typically lack a processing mechanism that can combine previously confirmed coding cases and their corresponding contextual differences for pre-confirmation prompts and reviews. This makes it difficult to balance reasonable difference judgments in "same text, different contexts" situations with multi-center coding consistency control. Therefore, there is an urgent need for an automated medical coding normalization scheme that can, based on semantic matching of original text terms, incorporate subject contextual information for semantic disambiguation, and combine cross-center historical confirmed cases for difference prompts and consistency reviews during the coding confirmation stage. Summary of the Invention

[0005] The purpose of this invention is to provide an automatic medical coding normalization method, system, and storage medium for multimodal medical data, in order to solve the problems of low coding efficiency, insufficient utilization of subject context information, difficulty in accurately disambiguating the same original text terms in different clinical scenarios, and lack of consistency prompts and review mechanisms in the multi-center coding confirmation process by combining historical confirmed cases and their contextual differences, thereby improving the accuracy, processing efficiency, and multi-center coding consistency of medical coding.

[0006] To achieve the above-mentioned technical objectives, the present invention adopts the following technical solution: An automatic medical coding normalization method for multimodal medical data includes the following steps: Step 1: Collect the original text terms of adverse events and / or combined medications, as well as the subject's structured information. Perform text denoising, format standardization, and data deduplication on the data to obtain standardized data, and load it into a standard coding dictionary. Step 2: Based on the pre-trained medical language model, perform semantic representation on the normalized original text terms and standard terms, and generate an initial candidate encoding set based on semantic similarity; Step 3: Extract the subject structured information corresponding to the original text terms, jointly encode the original text terms and the subject structured information, perform context-aware disambiguation and rearrangement on the initial candidate encoding set, and output the candidate encoding results; Step 4: In the confirmation phase, retrieve historical confirmed coding cases from other research centers that are identical to or have been preprocessed with the original text terms from the coding log, extract the structured information of historical subjects, and perform a difference analysis with the structured information of current subjects. Step 5: When the difference analysis indicates that there is a contextual difference identified by the preset difference influence rules, and the current candidate coding result is at a different level or with a different code than the coding result in the historical confirmed coding case, output a consistency prompt message and receive reconfirmation, coding modification or manual review operations. Step 6: The final confirmed coding results are associated with the original text terms, subject structured information, research center information and confirmation operation information to form a coding log, and written into the electronic data acquisition system; Step 3 is used to rearrange candidate codes according to the subject's context, and steps 4 and 5 are used to provide pre-confirmation review prompts based on cross-center historical confirmed code cases.

[0007] Optionally, the structured information of the subject includes at least one or more of the following: age, gender, pregnancy status, medical history, concomitant medication information, and examination indicators; The standard coding dictionary library includes at least the MedDRA (Medical Dictionary of Regulatory Activities) and the WHO Drug Dictionary.

[0008] Optionally, the pre-trained medical language model is BioBERT; In step 3, the original text terms and the subject's structured information are jointly encoded. This includes concatenating the context features corresponding to the original text terms and the subject's structured information and inputting them into a context encoder to reorder the initial candidate encoding set and output the Top-N candidate encoding results with confidence scores.

[0009] Optionally, step 4, retrieving historically confirmed coded cases, includes: Using the normalized original text terms obtained after preprocessing the current original text terms as the search key, retrieve the same normalized original text terms from the encoding log; Retrieve historical coding results, historical subject structured information, historical research center information, and historical confirmation operation information corresponding to the standardized original text terms; A cross-center historical comparison case set was constructed based on the information from the historical research center.

[0010] Optionally, the difference analysis in step 5 shall at least compare the current subject's structured information with the historical subject's structured information in terms of pregnancy status, gender, medical history, concomitant medication information, and examination indicators. The consistency prompt information also includes a description of the differences between the current candidate coding result and the historical coding result; After the consistency prompt information is output, the coder is allowed to perform any of the following operations: keep the current code, change the code corresponding to the historical code result, or submit for manual review.

[0011] Optionally, when the original text terms are combined drug description information and lack the generic name of the drug, the initial candidate code set is generated based on the drug appearance description, usage description and / or administration-related description in the original text terms, and the candidate drug codes in WHO Drug are re-arranged with context-aware semantic disambiguation in combination with the subject's medical history information, so as to output the candidate drug code results for code confirmation.

[0012] To achieve the above-mentioned technical objectives, the present invention also adopts the following technical solution: An automated medical coding normalization system for multimodal medical data, comprising: The data acquisition and preprocessing module is used to perform the data acquisition, preprocessing, and standard encoding dictionary loading described in step 1 above. The semantic matching and candidate generation module is used to perform the semantic representation of the original text terms and standard terms and the generation of the initial candidate encoding set as described in step 2 above; The context disambiguation module is used to perform the subject structured information extraction, joint encoding, and candidate encoding rearrangement described in step 3 above; The consistency verification module is used to perform the cross-center historical confirmed coding case retrieval, context difference analysis, consistency prompt information output and confirmation interaction described in steps 4 and 5 above. The storage and retrieval module is used to perform the encoding log association storage and encoding result retrieval described in step 6 above; The model optimization module is used to update the pre-trained medical language model and / or the context disambiguation processing strategy based on the confirmation operation information in the encoded log; The context disambiguation module is used to sort the original text terms that are the same or have been preprocessed into different clinical scenarios of different subjects. The consistency verification module is used to output consistency prompts based on the cross-center historical confirmed coding cases and their contextual differences during the coding confirmation stage, so as to provide a pre-confirmation review prompt for the sorting of the differentiated candidate codes, thereby synergistically improving coding accuracy and multi-center coding consistency.

[0013] Optionally, the consistency verification module includes: The historical case retrieval unit is used to retrieve from the coding log historical confirmed coding cases that are the same as or the same as the normalized original text terms after preprocessing; The contextual difference analysis unit is used to perform difference analysis on the current subject's structured information and the historical subject's structured information according to the preset difference influence rule table, and to determine that there is a contextual difference affecting the coding selection when at least one of the pregnancy status, gender, medical history, accompanying medication information or examination indicators meets the corresponding rule. The prompting interaction unit is used to output the consistency prompt information when the triggering conditions are met, and to receive the coder's reconfirmation, code modification, or submission for manual review.

[0014] Optionally, the storage and retrieval module is used to associate and store the original text terms, the final confirmed coding results, the subject's structured information, the research center information, the coding time, the coder information, and the confirmation operation information to form a cross-center historical coding case library that can be called by the consistency verification module. The model optimization module is used to continuously optimize the processing results of the semantic matching and candidate generation module and / or the context disambiguation module based on the coder's coding modifications, submission for manual review, and final confirmation results.

[0015] To achieve the above-mentioned technical objectives, the present invention also adopts the following technical solution: A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed by the processor, it implements any of the above-described automatic medical code normalization methods. The computer program is used to enable the processor to perform context-aware semantic disambiguation processing based on subject structured information in step 3, and consistency prompting processing based on cross-center historical confirmed coded cases and their contextual differences in steps 4 and 5.

[0016] The main advantages of this invention compared to existing technologies are as follows: This invention first performs context-aware semantic disambiguation on candidate codes based on the original text terms and the subject's structured information. Then, during the coding confirmation stage, it combines cross-center historical confirmed coding cases and their contextual differences to provide consistency prompts and verification. This enables the same or similar original text terms to obtain coding results that are more consistent with the actual semantics in different clinical scenarios. At the same time, it avoids misjudgments caused by mechanically applying existing codes between multiple centers. This effectively solves the problems of insufficient utilization of subject context, difficulty in accurately distinguishing "same text, different context" situations, and insufficient control of multi-center coding consistency in existing technologies.

[0017] This invention utilizes a pre-trained medical language model to semantically represent terms in original texts and standard dictionary terms. It also rearranges candidate codes by combining contextual features such as the subject's pregnancy status, gender, medical history, accompanying medication information, and examination indicators. This approach can overcome the dependence of traditional keyword matching on surface words and improve the accuracy and intelligence of mapping original terms to standard codes.

[0018] This invention introduces a cross-center historical confirmed coding case comparison analysis mechanism in the coding confirmation stage. It can output difference prompts related to historical cases before the current coder confirms the coding, so that the coder can reconfirm, modify or submit the coding for review based on the contextual differences between the current case and the historical cases. This is beneficial to improving the consistency and controllability of coding processing in multi-center and multi-coder scenarios.

[0019] This invention associates and stores the final confirmed coding results with the original text terms, subject structured information, research center information, coding time, coder information, and confirmation operation information. This not only meets the traceability requirements of the medical coding process but also provides a data foundation for subsequent model optimization, rule updates, and unified management across centers.

[0020] This invention is applicable to various medical data coding scenarios, such as adverse events and combined medication, and can call results and interact with electronic data acquisition systems, demonstrating good engineering adaptability and application promotion value. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the automatic medical code normalization method described in this invention; Figure 2 This is a schematic diagram of the software interface of the context-aware semantic disambiguation module of the present invention. Figure 3 This is a schematic diagram of the software interface for the encoding consistency verification and alarm interface of the present invention. Figure 4 This is a schematic diagram of the module structure of the system of the present invention. Detailed Implementation

[0022] To make the technical solution, implementation logic, and beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that the following embodiments are only for illustrating the present invention and are not intended to limit the scope of protection of the present invention. Without departing from the concept of the present invention, those skilled in the art can make several equivalent modifications or substitutions, all of which should fall within the scope of protection of the present invention.

[0023] like Figure 4 As shown, this invention provides an automatic medical coding normalization system for multimodal medical data, comprising at least: a data acquisition and preprocessing module, a semantic matching and coding recommendation module, a context disambiguation module, a consistency verification module, a storage and retrieval module, and a model optimization module. These modules can be deployed on the same server or distributed functionally across an application server, a model server, and a database server. The application server hosts the interactive interface and business process control; the model server runs the pre-trained medical language model and context encoder; and the database server stores the standard coding dictionary, coding logs, historical cases, and model training data. The system can be integrated into an electronic data collection (EDC) system for automatic medical coding normalization of raw records such as adverse events and concomitant medications in clinical research. The data acquisition and preprocessing module is used to collect multimodal medical data during clinical research and to standardize the collected data, providing a unified input for subsequent semantic matching and context disambiguation. The input to this module includes at least: raw text terms of adverse events entered by the researcher, raw text terms of concomitant medications, basic information of the subjects, medical history information, information on accompanying medications, examination indicators, research center information, and timestamp information. The basic information of the subjects may include age, gender, pregnancy status, etc., and the examination indicators may include liver function indicators, kidney function indicators, or other structured fields related to clinical judgment. As specified in the briefing, multimodal medical data includes at least raw text terms and structured information of the subjects. Preprocessing includes text denoising, format standardization, and data deduplication, while simultaneously loading MedDRA and WHO Drug dictionaries to establish a standard coded dictionary library.

[0024] The text denoising is not simply deleting characters, but rather cleaning up information that affects the recognition of medical terms but does not carry effective medical semantics. In one embodiment, text denoising includes at least the following processing actions: 1. Standardize the use of full-width and half-width characters, and standardize the format of Chinese and English parentheses, commas, periods, forward slashes, hyphens, and other separators; 2. Remove duplicate spaces, consecutive punctuation marks, meaningless line breaks, repeated conjunctions, and template-like prompts that are clearly not involved in encoding decisions; 3. Correct common misspellings, such as replacing minor spelling errors caused by typing habits with standard writing forms; 4. Standardize and unify colloquial expressions, abbreviations, or synonyms and near-synonyms. For example, unify "want to vomit" and "feel a little nauseous" into standardized expressions that can be recognized by semantic models. 5. Noise words that do not affect medical semantic judgment are deleted, but core descriptions that may affect coding judgment are retained, such as descriptions of medication use, appearance, administration instructions, and disease course.

[0025] For example, the original input "The patient experienced stomach discomfort and nausea after taking the medication" can be transformed into the standardized text "Stomach discomfort and nausea" after text denoising. For "The patient reported taking small blue pills to lower blood pressure," descriptions such as "small blue pills" and "lower blood pressure" are retained because this information will affect drug coding judgment in subsequent candidate generation. Examples 1 and 2 both embody the above processing approach.

[0026] The format standardization is used to unify the expression of text fields and structured fields. In one embodiment, format standardization includes at least: 1. Standardize the original text terminology to a preset text encoding format; 2. Map structured fields to a unified field name and a unified set of values. For example, map "28 weeks pregnant", "in the womb", and "pregnant" to the pregnancy status field, and map "male" and "male" to the gender field. 3. Standardize the format of research center number, subject number, collection node number, and time field to form standardized fields that can be used for cross-center retrieval and correlation analysis; 4. Set preset tags for missing and abnormal fields to avoid ambiguity due to incomplete fields during subsequent joint encoding.

[0027] The data deduplication process is used to filter duplicate records and retain valid events. In one embodiment, the system can determine whether a record is duplicated based on a combination of subject identifier, data type, standardized original text terminology, collection time, and research center information. When the subject identifier, data type, standardized original text terminology are consistent, and the collection time falls within a preset approximation window, it can be determined as a duplicate record and merged or filtered out. Records with the same text but different occurrence times or research scenarios are retained to avoid the accidental deletion of genuinely duplicate events. Through the above processing, a standardized multimodal medical dataset can be obtained for subsequent steps.

[0028] In addition, the data acquisition and preprocessing module is also used to establish and update the standard coding dictionary library. Specifically, the MedDRA dictionary can be loaded to extract LLT layer terms and their associations with PT layer and higher levels; the WHO Drug dictionary can be loaded to extract drug names, active ingredients, ATC classifications, and corresponding codes. To ensure the timeliness of coding, the system can perform dictionary synchronization updates at preset intervals or when a new dictionary version is detected, and rebuild the dictionary retrieval index or dictionary vector index after the update.

[0029] The semantic matching and encoding recommendation module is used to semantically represent the terms in the normalized original text and the standard terms in the standard encoding dictionary, and generate an initial set of candidate codes. As specified in the briefing, this module can preferably use the BioBERT pre-trained medical language model to achieve semantic embedding of the original terms and standard terms, and obtain the top N candidate terms with the highest similarity and their corresponding codes by calculating the cosine similarity.

[0030] In one implementation, the processing logic of the semantic matching and encoding recommendation module is as follows: First, the preprocessed original text terms are input into the BioBERT model to obtain the original term semantic vectors. Second, the LLT layer terms from the MedDRA dictionary and the drug terms from the WHO Drug dictionary are input into the same model or a compatible model to obtain the standard term semantic vectors. Then, the similarity between the original term semantic vectors and the standard term semantic vectors is calculated to generate a candidate term list. Finally, the top N terms are sorted from high to low similarity and selected as the initial candidate encoding set for output to the subsequent context disambiguation module.

[0031] For adverse event terms, the system primarily generates an initial candidate set from the MedDRA dictionary; for combined medication terms, it primarily generates an initial candidate set from the WHO Drug dictionary. For mixed text containing both symptoms and medication instructions, the system can first identify the data type based on the field source or preset business rules, and then proceed to the corresponding candidate generation pathway.

[0032] In one implementation, to improve matching accuracy in clinical research scenarios, the BioBERT model can be fine-tuned before deployment based on historical coding data and a standard dictionary terminology corpus. The historical coding data includes matching pairs of original terms and their corresponding standard codes, while the standard dictionary terminology corpus includes the MedDRA terminology set and the WHO Drug terminology set. This fine-tuning allows the model to better adapt to common expression habits and coding contexts in clinical research.

[0033] like Figure 2 As shown, the context disambiguation module is used to incorporate subject-structured information to rearrange candidate results in a contextualized manner after the initial candidate encoding set is generated. This addresses the limitation of difficulty in distinguishing between "same text, different contexts" when relying solely on text matching. The briefing clarifies that this module can concatenate original terms with contextual features and then perform joint encoding through a context encoder. This prioritizes standard terms that align with the contextual logic and outputs Top-N recommendation results with confidence scores.

[0034] In one implementation, Figure 2 The software interface shown includes at least a raw text terminology display area, a subject key information display area, a Top-N candidate coding result display area, and a confidence score display area. The raw text terminology display area displays standardized raw text terms; the subject key information display area displays at least one of the following: pregnancy status, gender, medical history, concomitant medication information, and examination indicators; the Top-N candidate coding result display area displays the candidate coding results after context-aware semantic disambiguation and rearrangement; and the confidence score display area displays the confidence score corresponding to each candidate coding result.

[0035] In one embodiment, the context disambiguation module includes a context feature extraction unit, a joint encoding unit, a rearrangement scoring unit, and a confidence output unit.

[0036] The context feature extraction unit is used to extract context features related to the current coding task from the preprocessed structured information, such as pregnancy status, gender, medical history, accompanying medication information and examination indicators; the joint coding unit is used to concatenate the original text terms with the extracted context features and input them into the context encoder; the re-ranking and scoring unit is used to re-rank and score each candidate option in the initial candidate coding set according to the joint coding results; the confidence output unit is used to calculate the confidence of each candidate option according to the re-ranked scoring results and output the Top-N candidate coding results.

[0037] To make the above process clearer, in one embodiment, let the first... The initial semantic matching score of the candidate codes is , No. The context fit score between each candidate code and the current subject's structured information is: Then the first Rearrangement score of candidate codes It can be represented as:

[0038] in For the first The semantic matching weight coefficients corresponding to each candidate code. For the first The context adaptation weight coefficients corresponding to each candidate code.

[0039] Furthermore, the first Confidence scores of candidate codes It can be represented as:

[0040] in, This represents the number of candidate codes participating in the rearrangement. The expression is an exponential function. Through the above method, the system can fuse semantic matching results and context adaptation results to obtain Top-N candidate coding results that are more consistent with the current clinical scenario. The above formula is only one feasible implementation method, intended to illustrate the generation logic of confidence scores, and does not limit the invention to using this specific calculation form.

[0041] In practical implementation, context adaptation score The system can generate candidate codes based on the association rules between contextual features and candidate codes. For example, when a subject's structured information includes a pregnancy status of "yes" and the candidate standard term is related to pregnancy symptoms, the context fit score of the corresponding candidate code can be increased; when the subject is male or there is no pregnancy-related information in the context, the context fit score of pregnancy-related candidates can be decreased. Similarly, when a subject's medical history information includes information related to coronary artery disease or hypertension, higher context fit scores can be assigned to candidate codes related to antihypertensive drugs or cardiovascular treatment drugs. Through the above processing, the system can obtain differentiated candidate code rankings for the same or similar original text terms in different subjects' clinical scenarios.

[0042] The consistency verification module is used during the coding confirmation phase to review and prompt the current candidate coding results by combining historically confirmed coding cases across centers and their contextual differences, thereby achieving multi-center coding consistency control. As disclosed in the disclosure, this module detects different coding results for the same original term based on coding consistency rules within and across research centers, triggering alarms and supporting manual calibration. In line with the aforementioned claims, this embodiment further advances the consistency prompt to the coding confirmation phase, providing difference explanations and prompts before the coder formally submits the code.

[0043] In one embodiment, the consistency verification module includes a historical case retrieval unit, a context difference analysis unit, a trigger judgment unit, and a prompt interaction unit.

[0044] The historical case retrieval unit is used to retrieve historical confirmed coding cases outside the current research center from the coding log using the current standardized original text terms as the search key. The historical confirmed coding cases include at least historical original text terms, historical coding results, historical subject structured information, historical research center information, historical coding time, and historical coder confirmation operation information. The contextual difference analysis unit is used to compare the differences between the current subject's structured information and the historical subject's structured information in dimensions such as pregnancy status, gender, medical history, concomitant medication information, and examination indicators; The triggering judgment unit is used to determine whether the difference is sufficient to affect the candidate coding selection, and at the same time to determine whether the current candidate coding result and the historical coding result have different levels or different codes; The prompt interaction unit is used to output consistency prompt information when the trigger conditions are met, and to receive confirmation, code modification or submission for manual review by the coder.

[0045] In one embodiment, the triggering condition may include at least one of the following: 1. The standardized original text terminology of the current case is consistent with that of historical cases; 2. There are presupposed differences in the subject context between the current case and the historical cases, such as one case being pregnant while another is not, or one case having a specific medical history while another does not; 3. The current candidate coding result is at a different level or has a different code than the historically confirmed coding result; 4. The aforementioned contextual differences, after being judged by the rules, are sufficient to affect the choice of encoding path.

[0046] For example, if the standardized original text terms for both the current case and historical cases are "stomach discomfort and nausea," and if the historical case subject was pregnant and the historical coding result was a standard term consistent with pregnancy symptoms, while the current case is a non-pregnant male subject, and the current candidate coding result tends to indicate adverse gastrointestinal reactions or drug-induced gastrointestinal reactions, the system can determine that this difference belongs to a "same text, different context" scenario and output a prompt message to the current coder. The prompt message may include: the historical research center identifier, historical coding results, current candidate coding results, and a description of the contextual differences between the current and historical cases. The coder can choose to keep the current code, change to the code corresponding to the historical coding result, or submit it for manual review by the administrator based on the prompt message. In this way, the system does not unify mechanically consistent coding results, but rather adopts a consistent processing logic of "interpreting differences in context and conducting reviews."

[0047] The storage and retrieval module is used to associate and store the final confirmed coding results and related information, and to synchronize the coding results to the corresponding fields in the EDC system for subsequent data cleaning, statistical analysis, and regulatory reporting. As disclosed in the briefing, the coding log includes at least the original terminology, contextual auxiliary information, coding time, and coding personnel information, and can be used for traceability and model fine-tuning.

[0048] In one implementation, the storage and retrieval module is associated with storing at least the following data items: 1. Original text terminology and its normalization results; 2. Data type identification to distinguish between adverse events and combined medications; 3. Structured information of the subjects; 4. Research Center Information; 5. Candidate coding results and their confidence levels; 6. Finally confirm the encoding result; 7. Encoding time and coder information; 8. Consistency warning messages and their triggering reasons; 9. Confirmation information such as reconfirmation, code modification, or submission for manual review.

[0049] The aforementioned data, after being correlated, forms an encoding log. This log serves two purposes: it allows the consistency verification module to retrieve historical cases, and it can also be used by the model optimization module. After the log is stored, the system can synchronously write the final confirmed encoding result into the corresponding field of the EDC system, forming a standardized encoding record.

[0050] The model optimization module is used to incrementally optimize the BioBERT model and context disambiguation strategy based on the confirmation results and interaction behaviors in the encoding log. As clearly stated in the briefing, the coders' corrections or rejections can be fed back as positive or negative samples to the model training set for subsequent incremental learning and iterative model optimization.

[0051] In one implementation, the model optimization module can read the encoding log at a preset period and extract the following samples: 1. Highly reliable samples directly confirmed by the coder are used as positive samples; 2. Samples of candidate coding results modified by coders are used as samples of "initially inaccurate recommendations but corrected results"; 3. Samples that coders reject all recommended results and submit for manual review will be considered as samples awaiting manual judgment. 4. The unified result after manual review by the administrator is used as a cross-center consistency correction sample.

[0052] These samples allow for further optimization of the semantic matching and encoding recommendation module and the context disambiguation module, thereby continuously improving encoding accuracy and multi-center consistency.

[0053] like Figure 1 As shown, this invention provides an automatic medical coding normalization method for multimodal medical data, comprising steps 1 to 6. This method can be derived from... Figure 4 The system is executed as shown, with each step connected sequentially according to the data flow. Step 3 is used to perform context-aware semantic disambiguation on the original text terms in the current clinical scenario of the subject. Steps 4 and 5 are used to review and prompt the current candidate coding results in the coding confirmation stage by combining cross-center historical confirmed coding cases, thus forming a collaborative processing chain of "context disambiguation first, consistency prompts later".

[0054] Step 1: Multimodal medical data acquisition and preprocessing Step 1 involves collecting and standardizing multimodal medical data to create standardized input data that can be used for subsequent semantic matching, context disambiguation, and consistency checks. The multimodal medical data includes at least the original textual terms of adverse events and concomitant medications entered by the researcher, as well as structured information about the subjects. This structured information includes at least one or more of the following: age, gender, pregnancy status, medical history, concomitant medication information, and examination indicators.

[0055] In one implementation, data can be acquired from the adverse event table, combined medication table, subject basic information table, medical history table, concomitant medication table, and laboratory test table in the Electronic Data Acquisition (EDC) system. After reading the above data, the system first associates the raw text terms of the same subject under the same acquisition node with the corresponding structured information to form the raw sample record corresponding to the current term. The raw sample record includes at least: subject identifier, research center identifier, data type identifier, raw text term, acquisition time, and corresponding structured field content. By completing this association before preprocessing, it can be ensured that the context analysis in subsequent steps 3 and 5 is performed based on the same case data.

[0056] The preprocessing includes at least text denoising, format normalization, and data deduplication.

[0057] Text denoising is used to remove noise components that affect term recognition without compromising the original medical semantics. In one embodiment, text denoising is performed according to at least the following rules: 1. Convert full-width characters to half-width characters and standardize the format of Chinese and English punctuation, parentheses, spaces, forward slashes, and connectors; 2. Remove consecutive repeated spaces, repeated punctuation marks, meaningless line breaks, and purely template-based prompt text; 3. Based on a pre-set medical terminology error correction list, obvious typos are replaced; 4. Based on a pre-defined medical terminology alias mapping table, common abbreviations, colloquial expressions, and synonyms are standardized; 5. Delete redundant segments in the text, such as time words and degree adverbs, that do not affect the standard encoding judgment; 6. Keywords that may affect code recognition should be retained, including symptom keywords, drug appearance descriptions, usage descriptions, administration descriptions, and disease course prompts.

[0058] For example, for the statement "The patient experienced stomach discomfort and nausea after taking the medication", the system can retain "stomach discomfort and nausea" after removing template descriptions such as "patient" and "experienced after taking the medication"; for the statement "The patient reported taking blue pills to lower blood pressure", the system retains "blue pills" and "lower blood pressure" for subsequent drug candidate retrieval and sorting.

[0059] The format standardization is used to unify the representation of text fields and structured fields. In one embodiment, format standardization includes at least: 1. Standardize the original text terminology to a preset character encoding format; 2. Map structured fields to a unified field name and a unified set of values. For example, map "28 weeks pregnant", "in the womb", and "pregnant" to the "pregnant" field in the pregnancy status field. 3. Convert structured information such as gender, medical history, accompanying medications, and examination indicators into standard field formats; 4. Standardize the format for the fields of research center number, subject number, collection node number, and collection time; 5. Set default flags for missing and abnormal fields so that the context encoder can identify them later.

[0060] The data deduplication process is used to clean up duplicate entries without mistakenly deleting genuinely recurring events. In one implementation, the system uses "subject identifier + data type identifier + standardized original text terminology + collection time window + research center identifier" as the duplicate detection key. When the same subject is in the same research center, with the same data type, and the standardized original text terminology is the same, and the collection time falls within a preset time window, the latter record is marked as a duplicate record. Records with the same standardized original text terminology but different collection times or occurrence points are retained as independent events. This process avoids misjudging genuine recurrence events as duplicate data.

[0061] After completing the above preprocessing, the system loads a standard coding dictionary library. This library includes at least the MedDRA dictionary and the WHO Drug dictionary. The MedDRA dictionary is used to extract LLT layer terms and their higher-level hierarchical relationships, while the WHO Drug dictionary is used to extract drug names, active ingredients, ATC classifications, and corresponding codes. After loading the dictionaries, the system can create text indexes and vector indexes for the standard terms, which can then be used in step 2.

[0062] Step 2: Original term encoding recommendation based on semantic embedding Step 2 is used to generate an initial candidate code set based on the semantic association between the normalized original text terms and the standard terms in the standard coding dictionary. As specified in the briefing, the pre-trained medical language model preferably uses BioBERT, and the top N candidate terms and their corresponding codes are selected by calculating the cosine similarity between the semantic vectors of the original terms and the semantic vectors of the standard terms.

[0063] In one implementation, the system first inputs the normalized original text terms into the BioBERT model to obtain original term semantic vectors; then, it inputs MedDRA LLT layer terms and WHO Drug terms from the standard coding dictionary into the same model to obtain standard term semantic vectors; next, it calculates the cosine similarity between the original term semantic vectors and the semantic vectors of each standard term; finally, it sorts the similarity from high to low and selects the top N candidates to form an initial candidate coding set. For adverse event records, the system prioritizes generating a candidate set from the MedDRA dictionary; for combined medication records, the system prioritizes generating a candidate set from the WHO Drug dictionary.

[0064] In one implementation, the BioBERT model can be fine-tuned using a clinical research medical coding corpus before deployment. This medical coding corpus includes at least matching pairs of historical original terms and standard codes, as well as MedDRA / WHO Drug dictionary terminology, thereby enhancing the model's adaptability to medical expressions in clinical research scenarios. The initial candidate coding set output in step 2 includes at least candidate standard terms, candidate codes, and corresponding initial semantic matching scores.

[0065] In one implementation, when the original text terms contain combined drug description information but lack the generic name of the drug, the system first extracts at least one feature word from the original text terms, including drug appearance description, usage description, and administration-related description, to form a drug description feature set. Then, the drug description feature set is matched with drug names, active ingredients, ATC classifications in the WHO Drug dictionary, and a pre-established drug description mapping table to generate a candidate drug code set. The drug description mapping table records the correspondence between drug appearance, usage, or administration method and candidate drug codes. Subsequently, the system performs context-aware semantic disambiguation rearrangement on the candidate drug code set in conjunction with the subject's medical history information or current disease status to output candidate drug codes for coding confirmation.

[0066] Step 3: Context-aware semantic disambiguation like Figure 2 As shown, step 3 involves incorporating the structured information of the current subject into the initial candidate encoding set generated in step 2, and then rearranging the candidate results in a contextualized manner to address the limitation of distinguishing between "same text, different contexts" based solely on the original text terms. As explicitly stated in the briefing, the system can concatenate the original terms with contextual features, perform joint encoding through a context encoder, and output Top-N recommendation results with confidence scores.

[0067] To avoid insufficient disclosure, in this embodiment, the processing logic of step 3 is further refined as follows: 1. Contextual feature extraction; 2. Contextual feature vectorization; 3. Generation of candidate encoding context adaptation scores; 4. Candidate encoding rearrangement; 5. Confidence calculation and Top-N output.

[0068] Contextual feature extraction refers to extracting fields related to the current coding judgment from the structured information obtained in step 1. For example, for adverse event coding, pregnancy status, gender, medical history, concomitant medications, and examination indicators can be extracted; for combined medication coding, medical history, current disease status, concomitant medications, and administration information can be extracted.

[0069] In one implementation, context feature vectorization can be achieved as follows: Convert enumerated fields such as pregnancy status and gender into preset category vectors; Convert multi-value fields such as medical history and accompanying medications into multi-label feature vectors; Convert numerical fields such as inspection indicators into range labels or standardized numerical features; The above features are then concatenated in a preset order to form the contextual feature vector of the current case.

[0070] Subsequently, the normalized original text terms and the aforementioned context feature vectors are input into a context encoder. The context encoder can be an extended input form of the BioBERT model or a standalone attention network. The context encoder outputs a context representation corresponding to the current case and generates a context fit score for each candidate code based on this representation. To illustrate how this score is implemented, in one embodiment, the context fit score of the candidate code can be generated by combining rules and the model: First, a candidate code attribute label is pre-established for each candidate code. The candidate code attribute label is used to characterize the candidate code's attributes such as pregnancy-related, digestive system-related, drug response-related, antihypertensive treatment-related, or cardiovascular disease-related. In one embodiment, the candidate coding attribute labels can be generated based on the hierarchical relationships, term keywords, and preset scenario mapping rules in a standard coding dictionary. Specifically, a mapping relationship between candidate codes and preset scenario labels can first be established based on the system organ classification to which MedDRA terms belong, the PT / LLT hierarchical semantics, and the drug use, active ingredient, and ATC classification corresponding to WHO Drug terms. Then, combined with the keywords in the terms, attribute labels such as pregnancy-related, digestive system-related, drug response-related, antihypertensive treatment-related, or cardiovascular treatment-related are added to the candidate codes, and the generated attribute labels are stored in the candidate code attribute label table for use in step 3.

[0071] Second, match the contextual features of the current case with the candidate coded attribute labels; Third, for candidate codes whose matching results conform to the preset rules, their context adaptation scores are increased; for candidate codes that do not conform to the preset rules, their context adaptation scores are decreased.

[0072] For example, when the current case's pregnancy status is "pregnant" and a candidate code attribute label contains "pregnancy-related," the system increases the context adaptation score of that candidate code; when the current case's gender is "male" and a candidate code attribute label contains "pregnancy-related," the system decreases the context adaptation score of that candidate code; when the current case's medical history includes "coronary artery disease" and a candidate code attribute label contains "cardiovascular treatment-related," the system increases the context adaptation score of that candidate code. Through these rules, the system ensures that the context adaptation score has a clear business meaning, rather than being generated solely from an abstract model black box.

[0073] In one embodiment, let the first The initial semantic matching score of the candidate codes is The context adaptation score is Then the first Rearrangement score of candidate codes It can be represented as:

[0074] in For the first The semantic matching weight coefficients corresponding to each candidate code. For the first The context adaptation weight coefficients corresponding to each candidate code.

[0075] In one implementation, and It can be a preset constant, or it can be dynamically output by the fine-tuned context encoder based on the context features of the current case; when a preset constant is used, its value range can be determined by the verification samples in the development phase; when a dynamic output method is used, the weight coefficient is changed by the context encoder as the context of the current case changes.

[0076] Furthermore, to obtain a confidence score that can be presented to coders, the reordering scores of each candidate code can be used. Convert to normalized confidence Its expression is:

[0077] in, Indicates the candidate code number. For the summation variable, This indicates the number of candidate codes that participate in the rearrangement. Indicates the first The reordering score of each candidate code. In the process of summation, the first... The reordering score of each candidate code. Indicates the first Normalized confidence scores of candidate codes Represents an exponential function. Indicates the first Rearrangement score of candidate codes The exponential weight value obtained after exponential transformation is used to characterize the relative priority of the candidate code among all candidate codes. The Top-N candidate coding results are the top N candidate coding results selected after sorting them from high to low according to the normalized confidence scores. Through the above calculation, the system can output the Top-N candidate coding results with a total confidence score of 1, so that coders can intuitively see the relative priority of each candidate.

[0078] Finally, the system according to or The candidate codes are sorted from highest to lowest priority, and the Top-N candidate codes are output. For the original term "stomach upset and nausea," if the current subject is pregnant, candidate codes such as "nausea" and "vomiting," which are related to pregnancy reactions, are given higher priority; if the current subject is a non-pregnant male, candidate codes related to "indigestion" or drug-induced gastrointestinal reactions are given higher priority. Through the above processing, step 3 can provide differentiated candidate code rankings for the same original text terms in different subjects' clinical scenarios.

[0079] Step 4: Confirmation of Human-Machine Collaboration Coding Step 4 displays the Top-N candidate coding results output in Step 3 on the interactive interface and receives confirmation from the coder. The coder can perform operations such as keeping the current candidate, selecting other candidates, manually modifying, or submitting for manual review on the candidate coding results output by the system. This step not only completes the final coding confirmation but also provides an interactive entry point for the consistency prompt in Step 5.

[0080] In one implementation, the interactive interface displays at least the following: standardized original text terminology, a summary of key subject information, Top-N candidate coding results, corresponding confidence scores, and a prompt information area related to the current candidate result. When a coder selects a candidate coding result as the result to be confirmed, the system does not immediately write it to the final database, but instead proceeds to step 5 to perform a consistency check. If step 5 does not trigger a prompt, the system directly receives the confirmation operation and proceeds to step 6. If step 5 triggers a prompt, the system further receives the coder's reconfirmation, coding modification, or submission for manual review in the prompt interactive interface. Through this processing method, steps 4 and 5 form a pre-confirmation review link, rather than a post-audit link.

[0081] Step 5: Consistency check based on encoding path like Figure 3 As shown, step 5 is used to provide consistency prompts and explanations of differences for the current candidate coding results based on historically confirmed coding cases across centers before the coder's final confirmation. This aims to improve multi-center coding consistency and avoid mechanically applying historical coding results. The disclosure has already made public the consistency verification rules for the same center and cross-center research centers, coding divergence alarms, and the logic for manual unified calibration. Based on this, this implementation further refines the rules into four stages: "historical case retrieval—contextual difference analysis—trigger judgment—prompt interaction".

[0082] In one implementation, Figure 3 The software interface shown includes at least a historical research center information display area, a historical coding result display area, a current candidate coding result display area, a context difference explanation display area, and a confirmation operation area. The confirmation operation area includes controls for keeping the current code, selecting the code corresponding to the historical coding result, and submitting for manual review. This area is used to receive interactive operations from coders in response to consistency prompts.

[0083] First, the system uses the standardized original text terminology of the current case as the search key to retrieve historically confirmed coded cases from the coding log, excluding those from the current research center. Each historically confirmed coded case includes at least: standardized original text terminology, historical coding results, historical subject structured information, historical research center information, historical coding time, and historical confirmation operation information. If no corresponding historical case is found, the system does not trigger a consistency prompt and directly returns to step 4 to complete the current coding confirmation. If multiple historical cases are retrieved, the system can select one or more representative cases based on research project consistency, time priority, or historical confirmation validity to construct a historical control case set.

[0084] Secondly, the system performs a difference analysis on the structured information of the current case and historical cases. To make the determination of "sufficient to influence coding selection" executable, in one embodiment, the system presets a difference influence rule table, which includes at least: 1. Pregnancy status difference rule: When the pregnancy status of the current case is different from that of historical cases, and there are pregnancy-related codes in the candidate codes or historical codes, the difference is identified as a difference that affects the code selection; 2. Gender Difference Rule: When the current case is different from the historical cases in terms of gender, and the candidate code belongs to a category that is related to the gender-specific scenario, the difference is identified as a difference that affects the code selection; 3. Medical history difference rule: When there is a preset difference between the current case and historical cases in the medical history field, and the difference is associated with the category of the candidate code, the difference is identified as a difference that affects the code selection; 4. Rule on differences in concomitant medications: When there are differences in the concomitant medication information between the current case and historical cases, and these differences are sufficient to change the judgment on the cause of adverse reactions or the use of drugs, the differences are considered to affect the coding selection. 5. Examination indicator difference rule: When the examination indicators of the current case and historical cases differ from the preset threshold, and the difference is related to the category to which the candidate code belongs, the difference is identified as a difference that affects the code selection.

[0085] The system then uses the difference analysis results and the coding results together to trigger a judgment. In one implementation, the system triggers a consistency prompt only when the following conditions are met simultaneously: 1. The standardized original text terminology of the current case is consistent with that of historical cases; 2. At least one contextual difference between the current case and historical cases is identified as influencing coding selection according to the difference influence rule table; 3. The current candidate coding result to be confirmed is at a different level or has a different code than the coding result in the historical confirmed coding case.

[0086] If there are only minor differences that do not affect the encoding path, the system will not trigger a prompt to avoid excessive prompts interfering with the coder's operation. In this way, "sufficient to affect encoding selection" is no longer an abstract judgment, but is based on a preset rule table and triggering conditions.

[0087] When the triggering conditions are met, the system outputs a consistency prompt message. This message includes at least: a historical research center identifier, historical coding results, current candidate coding results awaiting confirmation, a description of the contextual differences between the current case and historical cases, and a suggestion for further verification by the coder. Based on this prompt message, the coder can perform any of the following actions: keep the current coding, select a code corresponding to a historical coding result, or submit for manual review. If the coder chooses to submit for manual review, the system sends the current candidate result, historical coding results, and a description of the differences to the coding administrator. After the coding administrator performs unified calibration based on clinical research rules, the system writes the administrator's result back to the current case and records the review result in the coding log.

[0088] For example, given the standardized original text term "stomach discomfort and nausea," the current case involves a non-pregnant male subject, and the current candidate code is "indigestion." A historical case involved a pregnant subject, and the historical code was "nausea." In this case, the system, based on the pregnancy status difference rules, identifies this difference as influencing code selection. Combined with the difference between the current candidate code and the historical code, a consistency prompt is triggered. The system alerts the coder that while the historical case uses the same original text terminology as the current case, the historical case involved a pregnant subject, while the current case does not. Therefore, the historical code cannot be mechanically applied to the current case. The coder can either maintain the current code or submit it for manual review. Through this processing logic, the system standardizes the rules for interpreting and reviewing differences, rather than forcibly standardizing the final coding results across all centers.

[0089] Step 6: Storage and Retrieval of Encoding Results Step 6 is used to associate and store the final confirmed coding results and related information, and to retrieve them from the corresponding fields in the EDC system to meet the needs of traceability management and subsequent data utilization. As clearly stated in the briefing, the coding log is used at least to trace the coding process and provide data support for subsequent model fine-tuning.

[0090] In one implementation, the system stores the following information as a complete log record: standardized original text terminology, current subject structured information, research center information, Top-N candidate coding results and confidence scores, final confirmed coding results, coding time, coder information, whether consistency prompts were triggered, prompt content, coder reconfirmation or coding modification results, and whether manual review was submitted. These log records are used for subsequent cross-center historical case retrieval and for the model optimization module to extract positive samples, modified samples, and reviewed samples.

[0091] After logging is completed, the system synchronizes the final confirmed encoding results to the corresponding data fields in the EDC system for subsequent data cleaning, statistical analysis, and regulatory reporting. Through the sequential execution of steps 1 to 6, this invention achieves a complete closed loop from raw data input, candidate generation, context disambiguation, pre-confirmation review prompts to result archiving and retrieval.

[0092] The following are examples related to several methods. Example 1: Automatic Coding of Adverse Events (AEs) This embodiment corresponds to the following: Figure 1 The method flow shown is as follows: Figure 2 , Figure 3 The core processing logic is shown below. First, an AE record is extracted from the EDC system, with the original term "patient experienced stomach discomfort and nausea after medication." Simultaneously, the subject's contextual data is obtained, including subject ID 001, being in the 28th week of pregnancy, and having no history of liver disease. Then, in step 1, the system performs text denoising and format standardization on the original term to obtain the standardized original text term "stomach discomfort and nausea," and forms structured contextual features, such as "yes" for pregnancy status and "no" for history of liver disease.

[0093] In step 2, the system inputs "stomach discomfort and nausea" into the BioBERT model to obtain the original term semantic vector. This vector is then retrieved from MedDRA's LLT layer vector library to obtain an initial candidate encoding set, including candidate standard terms such as "nausea," "vomiting," and "indigestion." Then, in step 3, "stomach discomfort and nausea" is concatenated with contextual features such as pregnancy status and input into the context encoder to rearrange the candidate options. Since the current subject is pregnant, the system increases the context fit score of candidate options that better align with pregnancy symptoms, resulting in "nausea" and "vomiting" receiving higher rearrangement scores than "indigestion." Finally, a Top-N recommendation result with confidence scores is output, with "nausea" having the highest confidence score.

[0094] In steps 4 and 5, when the coder is ready to confirm "nausea," the system retrieves previously confirmed coding cases from other research centers that match the current normalized original text term. If a search is conducted at another research center where the same normalized original text term corresponds to a subject who is not pregnant, but whose historical coding results are biased towards indigestion or drug-induced gastrointestinal reactions, the system will... Figure 3The alarm interface displayed indicates to the current coder: Although the original text terminology is consistent, the current case differs from historical cases in terms of pregnancy status, which may affect the coding path determination. The current coder, considering the subject's pregnancy, ultimately confirms the coding result as "nausea." This confirmation and its contextual information are stored in the coding log and synchronized to the EDC system. This embodiment demonstrates that the present invention can improve coding accuracy in "same text, different context" scenarios while maintaining cross-center consistency control.

[0095] Example 2: Automatic coding of combined medications This embodiment illustrates that the present invention is also applicable to scenarios involving combined medication. First, a combined medication record is extracted, with the original terminology being "patient reports taking small blue pills for lowering blood pressure." In step 1, the system retains descriptions meaningful for drug identification, such as "small blue pills" and "lowering blood pressure," and performs format standardization. Subsequently, in step 2, the system inputs the standardized original text terms into the BioBERT model and retrieves an initial set of candidate codes from the WHO Drug dictionary. Because the original terminology lacks a clear generic name, traditional keyword retrieval is difficult to directly match. However, through semantic vector matching, the system can retrieve relevant candidates for "blue tablets for treating hypertension" and prioritizes the code corresponding to amlodipine as a high-priority candidate result.

[0096] Further, in step 3, the system extracts the current subject's medical history information and other structured information for context disambiguation. If the subject's medical history clearly indicates coronary artery disease, the system can simultaneously retain other drug candidates related to cardiovascular disease treatment, such as atorvastatin, and calculate a context fit score based on the current drug use description "antihypertensive" and the medical history information, reordering the candidates. If the first candidate has the highest match with the drug use, it is output as the preferred recommendation; if there are multiple possibilities in the use description and medical history, the system outputs multiple Top-N candidate drug codes with confidence scores for coders to confirm. Finally, the coder confirms based on the primary use, and the confirmation result, along with the relevant context information, is written to the coding log. This embodiment shows that the present invention is not only applicable to AE coding but also to combined drug coding scenarios lacking generic names.

[0097] Example 3: System Deployment and Effect Verification In one implementation, the system can be deployed on a server and supports operating systems such as Windows or Linux. The BioBERT model can be deployed on a GPU server to improve the efficiency of semantic vector generation and similarity calculation; the EDC system interacts with the system through an interface to import raw data and write back the final encoded results. Figure 4The modules in the system shown can be deployed on the same server or on different servers and communicate with each other through an internal network.

[0098] To verify the effectiveness of this invention, 1000 original adverse event terms and 500 original terms related to concomitant medications from a multicenter clinical study were selected and compared using the method of this invention and a traditional semi-automatic coding method. The disclosure states that the average coding time of the method of this invention is approximately 0.5 minutes per term, while the traditional method takes approximately 1.2 minutes per term; the coding accuracy of the method of this invention is approximately 92.7%, while the traditional method is approximately 69.5%; the multicenter coding discrepancy rate of the method of this invention is approximately 1.8%, while the traditional method is approximately 12.3%. These results demonstrate that this invention can effectively improve the efficiency, accuracy, and multicenter consistency of medical coding.

[0099] Example 4: Example of a computer-readable storage medium This embodiment provides a computer-readable storage medium storing a computer program that can be loaded and executed by a processor to implement all the steps of the automated medical coding normalization method for multimodal medical data described in the foregoing embodiments. The computer-readable storage medium can be a server hard drive, solid-state drive, magnetic disk, optical disk, read-only memory (ROM), random access memory (RAM), flash memory, mobile storage card, or other tangible storage medium capable of storing a computer program and being read by a processor.

[0100] In one embodiment, the computer program includes data acquisition and preprocessing instructions, semantic matching and candidate generation instructions, context disambiguation instructions, consistency verification instructions, storage and retrieval instructions, and model optimization instructions. The data acquisition and preprocessing instructions control the processor to acquire multimodal medical data and perform text denoising, format standardization, data deduplication, and loading of a standard coding dictionary. The semantic matching and candidate generation instructions control the processor to perform semantic representation of the original text terms and standard terms based on a pre-trained medical language model and generate an initial candidate coding set. The context disambiguation instructions control the processor to extract the subject's structured information, perform context-aware semantic disambiguation rearrangement on the candidate coding set, and output Top-N candidate coding results with confidence scores. The consistency verification instructions control the processor to retrieve historical confirmed coding cases across centers during the coding confirmation stage, compare the differences between the current subject's structured information and the historical subject's structured information, output consistency prompts when trigger conditions are met, and receive reconfirmation, coding modification, or submission for manual review. The storage and retrieval instructions control the processor to associate and store the finally confirmed coding results and related information to form a coding log and retrieve it to the corresponding data fields of the electronic data acquisition system. The model optimization instructions control the processor to update the pre-trained medical language model and / or context disambiguation processing strategy based on the confirmation operation information in the coding log.

[0101] With the above settings, when the computer program is executed by the processor, it can realize the entire processing flow of the method described in the foregoing embodiments, enabling the same or similar original text terms to obtain differentiated candidate code ranking in different subjects' clinical scenarios, and outputting consistency prompt information based on cross-center historical confirmed coding cases and their contextual differences during the coding confirmation stage, thereby improving the accuracy, processing efficiency and multi-center coding consistency of medical coding.

[0102] In summary, such as Figure 1 As shown, this invention achieves a complete closed loop from multimodal medical data acquisition, terminological semantic matching, context disambiguation, pre-confirmation consistency prompts, result archiving to EDC invocation through steps 1 to 6; as Figure 2 As shown, the context disambiguation module undertakes the core task of differentiating and ranking candidate codes in the "same text, different context" scenario; for example... Figure 3 As shown, the consistency verification module provides prompts during the coding confirmation phase based on cross-center historical cases and their contextual differences; for example... Figure 4 As shown, the modules work together to complete the automatic medical code normalization process. Through the coordination of the above system structure and method, this invention can effectively solve the problems of low efficiency, insufficient context utilization, and poor multi-center consistency in the prior art.

[0103] This invention first performs context-aware semantic disambiguation on candidate codes based on the original text terms and the subject's structured information. Then, during the coding confirmation stage, it combines cross-center historical confirmed coding cases and their contextual differences to provide consistency prompts and verification. This enables the same or similar original text terms to obtain coding results that are more consistent with the actual semantics in different clinical scenarios. At the same time, it avoids misjudgments caused by mechanically applying existing codes between multiple centers. This effectively solves the problems of insufficient utilization of subject context, difficulty in accurately distinguishing "same text, different context" situations, and insufficient control of multi-center coding consistency in existing technologies.

[0104] This invention utilizes a pre-trained medical language model to semantically represent terms in original texts and standard dictionary terms. It also rearranges candidate codes by combining contextual features such as the subject's pregnancy status, gender, medical history, accompanying medication information, and examination indicators. This approach can overcome the dependence of traditional keyword matching on surface words and improve the accuracy and intelligence of mapping original terms to standard codes.

[0105] This invention introduces a cross-center historical confirmed coding case comparison analysis mechanism in the coding confirmation stage. It can output difference prompts related to historical cases before the current coder confirms the coding, so that the coder can reconfirm, modify or submit the coding for review based on the contextual differences between the current case and the historical cases. This is beneficial to improving the consistency and controllability of coding processing in multi-center and multi-coder scenarios.

[0106] This invention associates and stores the final confirmed coding results with the original text terms, subject structured information, research center information, coding time, coder information, and confirmation operation information. This not only meets the traceability requirements of the medical coding process but also provides a data foundation for subsequent model optimization, rule updates, and unified management across centers.

[0107] This invention is applicable to various medical data coding scenarios, such as adverse events and combined medication, and can call results and interact with electronic data acquisition systems, demonstrating good engineering adaptability and application promotion value.

[0108] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automatic medical coding normalization method for multimodal medical data, characterized in that, Includes the following steps: Step 1: Collect the original text terms of adverse events and / or combined medications, as well as the subject's structured information. Perform text denoising, format standardization, and data deduplication on the data to obtain standardized data, and load it into a standard coding dictionary. Step 2: Based on the pre-trained medical language model, perform semantic representation on the normalized original text terms and standard terms, and generate an initial candidate encoding set based on semantic similarity; Step 3: Extract the subject structured information corresponding to the original text terms, jointly encode the original text terms and the subject structured information, perform context-aware disambiguation and rearrangement on the initial candidate encoding set, and output the candidate encoding results; Step 4: In the confirmation phase, retrieve historical confirmed coding cases from other research centers that are identical to or have been preprocessed with the original text terms from the coding log, extract the structured information of historical subjects, and perform a difference analysis with the structured information of current subjects. Step 5: When the difference analysis indicates that there is a contextual difference identified by the preset difference influence rules, and the current candidate coding result is at a different level or with a different code than the coding result in the historical confirmed coding case, output a consistency prompt message and receive reconfirmation, coding modification or manual review operations. Step 6: The final confirmed coding results are associated with the original text terms, subject structured information, research center information and confirmation operation information to form a coding log, and written into the electronic data acquisition system; Step 3 is used to rearrange candidate codes according to the subject's context, and steps 4 and 5 are used to provide pre-confirmation review prompts based on cross-center historical confirmed code cases.

2. The automatic medical coding normalization method according to claim 1, characterized in that, The structured information of the subjects includes at least one or more of the following: age, gender, pregnancy status, medical history, concomitant medication information, and examination indicators. The standard coding dictionary library includes at least the MedDRA (Medical Dictionary of Regulatory Activities) and the WHO Drug Dictionary.

3. The automatic medical coding normalization method according to claim 1, characterized in that, The pre-trained medical language model is BioBERT; In step 3, the original text terms and the subject's structured information are jointly encoded. This includes concatenating the context features corresponding to the original text terms and the subject's structured information and inputting them into a context encoder to reorder the initial candidate encoding set and output the Top-N candidate encoding results with confidence scores.

4. The automatic medical coding normalization method according to claim 1, characterized in that, Step 4, which involves retrieving historically confirmed coded cases, includes: Using the normalized original text terms obtained after preprocessing the current original text terms as the search key, retrieve the same normalized original text terms from the encoding log; Retrieve historical coding results, historical subject structured information, historical research center information, and historical confirmation operation information corresponding to the standardized original text terms; A cross-center historical comparison case set was constructed based on the information from the historical research center.

5. The automatic medical coding normalization method according to claim 1, characterized in that, The difference analysis in step 5 compares at least the current subject's structured information with the historical subject's structured information in terms of pregnancy status, gender, medical history, concomitant medication information, and examination indicators. The consistency prompt information also includes a description of the differences between the current candidate coding result and the historical coding result; After the consistency prompt information is output, the coder is allowed to perform any of the following operations: keep the current code, change the code corresponding to the historical code result, or submit for manual review.

6. The automatic medical coding normalization method according to claim 1, characterized in that, When the original text terms contain combined drug description information and lack the generic name of the drug, the initial candidate code set is generated based on the drug appearance description, usage description and / or administration-related description in the original text terms. The candidate drug codes in WHO Drug are then re-arranged using context-aware semantic disambiguation in conjunction with the subject's medical history information to output the candidate drug codes for code confirmation.

7. An automatic medical coding normalization system for multimodal medical data, characterized in that, include: The data acquisition and preprocessing module is used to perform the data acquisition, preprocessing, and standard encoding dictionary loading described in step 1 of claim 1. The semantic matching and candidate generation module is used to perform the semantic representation of the original text terms and standard terms and the generation of the initial candidate encoding set as described in step 2 of claim 1; The context disambiguation module is used to perform the subject structured information extraction, joint encoding, and candidate encoding rearrangement described in step 3 of claim 1; The consistency verification module is used to perform the cross-center historical confirmed coding case retrieval, context difference analysis, consistency prompt information output and confirmation interaction as described in steps 4 and 5 of claim 1. The storage and retrieval module is used to perform the associated storage of the encoding log and the retrieval of the encoding result as described in step 6 of claim 1; The model optimization module is used to update the pre-trained medical language model and / or the context disambiguation processing strategy based on the confirmation operation information in the encoded log; The context disambiguation module is used to sort the original text terms that are the same or have been preprocessed into different clinical scenarios of different subjects. The consistency verification module is used to output consistency prompts based on the cross-center historical confirmed coding cases and their contextual differences during the coding confirmation stage, so as to provide a pre-confirmation review prompt for the sorting of the differentiated candidate codes, thereby synergistically improving coding accuracy and multi-center coding consistency.

8. The automatic medical coding normalization system according to claim 7, characterized in that, The consistency verification module includes: The historical case retrieval unit is used to retrieve from the coding log historical confirmed coding cases that are the same as or the same as the normalized original text terms after preprocessing; The contextual difference analysis unit is used to perform difference analysis on the current subject's structured information and the historical subject's structured information according to the preset difference influence rule table, and to determine that there is a contextual difference affecting the coding selection when at least one of the pregnancy status, gender, medical history, accompanying medication information or examination indicators meets the corresponding rule. The prompting interaction unit is used to output the consistency prompt information when the triggering conditions are met, and to receive the coder's reconfirmation, code modification, or submission for manual review.

9. The automatic medical coding normalization system according to claim 7, characterized in that, The storage and retrieval module is used to associate and store the original text terms, the final confirmed coding results, the subject's structured information, the research center information, the coding time, the coder information, and the confirmation operation information, so as to form a cross-center historical coding case library that can be called by the consistency verification module. The model optimization module is used to continuously optimize the processing results of the semantic matching and candidate generation module and / or the context disambiguation module based on the coder's coding modifications, submission for manual review, and final confirmation results.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the automatic medical code normalization method according to any one of claims 1 to 6; The computer program is used to enable the processor to perform context-aware semantic disambiguation processing based on subject structured information in step 3, and consistency prompting processing based on cross-center historical confirmed coded cases and their contextual differences in steps 4 and 5.

Images