A medical insurance three catalog intelligent matching method based on a Med-BERT model

Abstract

The application relates to the field of data processing, in particular to a medical insurance three-catalogue intelligent matching method based on a Med-BERT model, which comprises the following steps: S1, providing a medical insurance three-catalogue intelligent matching system based on the Med-BERT model; S2, establishing a high-dimensional semantic vector database of a medical insurance three-catalogue library; S3, matching medical data; S4, manually inputting; S5, perfecting the model; through a multi-region catalog library integration vectorization+Med-BERT small model semantic matching+automatic training learning closed-loop whole-process architecture, the problems that the existing scheme is rigid, depends on external interfaces and has high deployment cost are solved, and accurate, efficient and safe matching between a medical list and the medical insurance three-catalogue is realized.

Description

Technical Field

[0001] This invention relates to the field of data processing, specifically to an intelligent matching method for the three medical insurance catalogs based on the Med-BERT model. Background Technology

[0002] With the rapid development of the health and accident insurance industry, the efficiency and accuracy of claims services have become one of the core competitive advantages. In the medical expense claims process, matching the medical bill details with the three medical insurance catalogs (drug catalog, treatment item catalog, and material catalog) is a crucial process that directly determines the compliance and reasonableness of the claims amount calculation.

[0003] The three medical insurance catalogs exhibit significant regional differences. Each province and city develops its own catalog based on local medical resources and economic levels, and the catalog content (such as name descriptions, medical insurance levels, and reimbursement ratios) is updated regularly. However, the medical lists provided by insured individuals often suffer from inconsistent terminology and non-standard expressions. For example, drug names may contain redundant information such as brand names, abbreviations, specification suffixes, and punctuation marks; treatment items and material names may contain regional colloquialisms or abbreviations, making accurate matching between the list details and the catalog names quite difficult. Existing mainstream matching solutions rely on OCR recognition + multi-round regularization processing + exact match as their core, but they suffer from the following problems: rigid matching logic and low accuracy: existing solutions depend on the exact match core logic, which can only recognize names with completely identical characters and cannot handle semantically related scenarios of medical terms. For example, amoxicillin capsules in the list may fail to match the generic name of amoxicillin in the catalog, and intravenous infusion (common) in the list may fail to match the standard name of intravenous infusion in the catalog due to incomplete character consistency, even if semantically equivalent. This results in a large amount of valid information being misjudged as not being covered by medical insurance. Regular expression processing has strong limitations and poor adaptability: it can only mechanically remove redundant information in fixed formats and lacks semantic understanding capabilities in the medical field. For non-standardized expressions (such as the brand name in cefixime dispersible tablets (Sefos), or the abbreviation for acupuncture in traditional Chinese medicine), it cannot achieve effective simplification and standardization, and it cannot handle newly added variant terms (such as common names for new materials), leading to poor preprocessing results. High labor costs and low efficiency: due to low matching accuracy, a large amount of details that could be matched through semantic association need to be manually entered. This not only increases the workload of claims personnel but also relies on their medical expertise, making human error prone to occur. Furthermore, the long manual processing cycle severely impacts claims efficiency. No iterative optimization mechanism: The regular expression rules and matching logic of the existing solution are fixed settings, which cannot be optimized by actual claims data (such as matching results manually added). After long-term use, the matching effect is difficult to improve and cannot adapt to the needs of medical insurance catalog updates and medical terminology evolution.

[0004] Therefore, it is essential to propose an intelligent matching method for the three medical insurance catalogs based on the Med-BERT model. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent matching method for the three medical insurance catalogs based on the Med-BERT model. This method uses vector matching to achieve efficient and accurate matching between medical data and medical insurance catalog data, thereby addressing the shortcomings and unmet technical requirements of existing technologies.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for intelligent matching of the three medical insurance catalogs based on the Med-BERT model, comprising the following steps: I. A medical insurance three-category intelligent matching system based on Med-BERT model is provided, including a data receiving module, a preprocessing module, a first storage module (MySQL), a data processing module, a second storage module (Milvus), a verification and output module, and an automatic training module; In this application, the data receiving module is the core front-end processing module of the medical insurance data intelligent processing system. It is used to realize cross-source integration of medical insurance data, end-to-end extraction of medical list information and standardized preprocessing. It mainly obtains medical insurance-related data information from the medical security bureau or relevant official websites.

[0007] Module Input: This module receives two types of input data: ① Input from the Medical Security Bureau: Batch-imported medical insurance data and real-time medical insurance data obtained from system integration (data formats include structured tables and semi-structured text); ② Input from the Service Caller: Medical records (carrier types include scanned copies of paper medical lists, electronic medical list PDF files, and photos of medical lists taken with a mobile phone), in JPG, PNG, and PDF formats.

[0008] Module Output: This module outputs two types of processed data: ① Integrated medical insurance data: a structured medical insurance data set that is categorized and integrated by region and year; ② Enhanced list text: medical list text with extracted details and amount information and standardized preprocessing (including semantic enhancement features).

[0009] The preprocessing module is the core collaborative unit connecting the data receiving module with the storage and training modules. It includes standardized processing and cleaning, manual supplementation of catalog data and manual entry and association matching of list data and catalog data nodes. It realizes the standardized organization of medical insurance data, the supplementation of missing information and the supply of training data, and provides data support for subsequent vectorization and matching processes. The first storage module serves as a structured storage medium for the three medical insurance catalogs, used for persistent storage of standardized drug, treatment, and material catalog data, supporting subsequent data processing modules. This module enables structured storage, data updates, and query support for the three medical insurance catalogs (drugs, treatments, and materials), ensuring data integrity and accessibility. The data processing module is the core algorithm execution module of the medical insurance data intelligent processing system. It is responsible for high-dimensional semantic vectorization, vector storage, and similarity matching of medical insurance three-category database data and list data, providing accurate candidate matching results for the verification module. It realizes high-dimensional semantic vectorization processing of medical insurance three-category database data (drugs, treatments, materials) and structured list data, stores the generated catalog semantic vectors in the Milvus vector database, and completes matching calculations using the vector similarity algorithm of the Med-BERT model. It then returns the top K records sorted by relevance to support accurate verification by the subsequent verification module. The second storage module is a dedicated vector data storage unit for the medical insurance data intelligent processing system. It is built based on the Milvus vector database and is used to efficiently store high-dimensional semantic vectors from the three medical insurance catalogs, providing low-latency and high-throughput vector retrieval support for similarity matching in the data processing module. The main functions are as follows: Persistent storage of vector data: providing structured storage for high-dimensional semantic vectors of the catalog and list text, ensuring data integrity and persistence; Efficient vector index construction: building optimized indexes based on vector data features, supporting millisecond-level near nearest neighbor (ANN) search; Multi-condition mixed query support: supporting compound queries combining vector similarity and scalar metadata (such as region and type) to meet complex business retrieval needs; Dynamic data update and management: providing interfaces for inserting, updating, and deleting vector data, supporting full and incremental updates of medical insurance catalog and list data; The verification and output module is the decision-making center and quality control core of the medical insurance data intelligent processing system. By constructing a multi-level verification system, including basic information verification, three-level progressive actuarial verification of amounts, and business logic and policy compliance verification, combined with the decision-making logic of step-by-step verification, step-by-step standard improvement, and violation tracing and downgrading, it achieves accurate verification of highly similar candidate records and high-credibility comprehensive decision-making, ensuring the accuracy, compliance and traceability of medical insurance settlement data.

[0010] Compliance verification is conducted at each level: detailed list level (strictest), project summary level (medium), and invoice summary level (lenient). The stringency of verification decreases with each level, with zero tolerance for errors at the detailed list level and reasonable error accumulation allowed at the summary level. The system integrates three categories of rules: basic information matching, actuarial calculation of amounts, and policy compliance, covering all dimensions of medical insurance settlement verification needs. Detailed list violations are directly classified as high-risk, while project / invoice level violations are penalized with lenient standards to avoid invalidating the validity of individual details due to summary errors.

[0011] The automatic training module is the central hub for model iteration in the medical insurance data intelligent processing system. It is responsible for receiving medium / low confidence data (valid samples after manual review, supplementation, and correlation) from the preprocessing module. Through standardized processing, incremental training, and effect evaluation, it continuously optimizes the Med-BERT model while ensuring that the original accuracy of the model does not decrease.

[0012] II. Establishment of a High-Dimensional Semantic Vector Database for the Three Medical Insurance Catalogs 2.1) The data receiving module receives medical insurance data from multiple regions and processes the data to generate a structured integrated medical insurance dataset. In this application, the above-mentioned content requires the collection of complete data from the three medical insurance catalogs of 34 provincial-level administrative regions (including provinces, autonomous regions, municipalities, and special administrative regions) nationwide, covering the three major categories of catalogs: Drug catalog: Includes chemical drugs, traditional Chinese medicine preparations, and Chinese herbal medicine slices, etc. The core fields are generic name, brand name, dosage form, specifications, medical insurance level, reimbursement ratio, reimbursement restrictions (applicable diseases, frequency of use), unit price limit, regional identifier, catalog version number, and update timestamp; Medical Service Item Catalog: Includes examination items, treatment items, surgical items, etc. The core fields are item name, medical department classification, clinical use, medical insurance level, reimbursement ratio, reimbursement restrictions, region identifier, catalog version number, and update timestamp; Medical Materials Catalog: Includes disposable medical materials, implantable medical materials, etc. The core fields are material name, material type, specifications, medical insurance level, reimbursement ratio, reimbursement restrictions, unit price limit, region identifier, catalog version number, and update timestamp.

[0013] 2.2) The structured medical insurance integrated dataset is cleaned and standardized through the preprocessing module to generate a standardized three-category library; 2.3) The first storage module uses a MySQL database to store a standardized three-category library (such as reimbursement level and payment level are unified as cost level). The standardized three-category library distinguishes the directory type, fills in area_code to specify the medical insurance region, and enters it into the original database to form a structured database of medical insurance three-category library. The standardized three-category library is categorized by medicine_type, with the following table fields: medicine_id (unique identifier for catalog data), import_type (data import type), medicine_type (catalog type: 1-medicine, 2-treatment, 3-materials), area_code (medical insurance area code), item_name (item name), item_code (item code), fee_type (fee level), medicine_classify (item category), code_no (code), medicine_name (medicine / treatment / material name), name_mnemonic (trade name mnemonic code), common_name (generic name), common_mnemonic (generic name mnemonic code), english_name (English name), dosage_form (dosage form, only valid for medicines), unit_num (quantity per package, only valid for medicines / materials), unit (package unit), content (content), content_unit (Content Unit), ethical_flag (Prescription Drug Flag), manufacturer_id (Pharmaceutical ID), manufacturer_name (Manufacturer Name), include (Item Included), except (Excluded Items), selfcontrol_agenti_flag (Self-made Reagent Flag), selfcontrol_agenti_org (Self-made Reagent Application Agency), self_pay_ratio (Self-payment Ratio), standard_unit_price (Standard Unit Price, Valid Only for Diagnostic / Materials), price_limit (Price Limit), reimbursement_limit (Reimbursement Limit, Valid Only for Pharmaceuticals), remark (Remark), state (Data Status), created_at (Creation Date), updated_at (Last Update Date), operator_id (Operator ID), operator_name (Operator Name).

[0014] Data writing: Receives standardized three-category data from the preprocessing module, distinguishes the catalog type by medicine_type, fills in area_code to specify the medical insurance region, and completes batch data writing; Data updating: Synchronizes manually entered drug, treatment, and material data, matches existing records based on medicine_id, and updates corresponding field information; Query support: Provides query interfaces by area_code (medical insurance region), medicine_type (catalog type), medicine_name (project name), etc., providing data retrieval support for the manual data entry module below.

[0015] 2.4) The processing module identifies newly added or modified records in the structured data of the three medical insurance catalogs (including drug, treatment, and material catalogs, with fields including generic name, dosage form, specification, cost level, co-payment ratio, price limit, reimbursement restriction, remarks, region, etc.) based on the updated_at field of the first storage module, and obtains incremental data. After classifying the incremental data by item type (drug / treatment / material), it is then processed in a standardized manner (text concatenation, standardization, and removal of invalid information) to generate a standardized and enhanced medical insurance list text. 2.5) Input the enhanced list text of medical insurance regulations into the Med-BERT model in the processing module for high-dimensional semantic vectorization transformation to generate high-dimensional semantic vectors of medical insurance text; 2.6) The high-dimensional semantic vector of medical insurance text is stored in the catalog vector set through the second storage module, and used as the catalog entry vector in the high-dimensional semantic vector database of the three catalogs of medical insurance. A special index is built on the catalog vector set. In the high-dimensional semantic vector database of medical insurance text, each record in the medical insurance three-category database has a corresponding 768-dimensional semantic vector, along with metadata such as medicine_id (unique identifier of catalog data), medicine_type (catalog type: 1-drug, 2-treatment, 3-material), and area_code (medical insurance area code).

[0016] Medical directory vectors

[0017] Index type: IVF_FLAT index is used. This index divides the vector space into multiple clusters using the K-means algorithm. During a query, it performs an exact search only in the clusters closest to the target vector, significantly improving query speed while ensuring high recall, making it very suitable for medical insurance data retrieval scenarios.

[0018] Core parameter configuration: metric_type: IP. Considering that the Med-BERT output vectors have already undergone L2 normalization, using IP is equivalent to using cosine similarity, but more computationally efficient. nlist: Number of cluster centers, set to 1024. This parameter strikes a balance between retrieval speed and accuracy; this value is suitable for datasets with millions of vectors.

[0019] Index building timing: Full build: Manually trigger index building after the first batch insertion of a large amount of data into the collection (such as initializing the medical insurance catalog). Incremental build: For continuously inserted incremental data, Milvus supports automatically merging it into the existing index in the background, or periodically rebuilding the index according to a data volume threshold (every 10,000 new vectors) to maintain retrieval performance.

[0020] Data operation interface The module provides standard CRUD (Create, Read, Update, Delete) interfaces for data processing modules to call.

[0021] Vector insertion: Supports batch insertion of vectors and their metadata.

[0022] Vector Query: Input one or more query vectors; specify the search scope; set search parameters (e.g., metric_type, params={"nprobe":16}, where nprobe determines the number of clusters to access during the search; a larger value results in higher accuracy but slower speed). Set output parameters (e.g., limit=50, returning the Top-50 results).

[0023] Vector update: Updates the value of a vector or scalar field based on the primary key ID.

[0024] Vector deletion: Delete vectors based on the primary key ID or an expression in the scalar field (e.g., area_code in ["110000", "310000"]).

[0025] III. Medical Data Matching 3.1) The system receives medical records containing image information through the data receiving module, extracts and standardizes the image information from the medical records, and generates a medical-enhanced list text. 3.2) Input the medical enhanced list text into the Med-BERT model in the processing module for high-dimensional semantic vectorization transformation to generate a high-dimensional semantic vector of the list text (deleted after matching is completed). High-dimensional semantic vector of the list text: A 768-dimensional semantic vector corresponding to each list data item, with accompanying metadata such as list_id (unique identifier of list data), list_type (list type: medicine / treatment / materials), area_code (area code of the treatment area), etc. 3.3) Then, the Med-BERT model is used to calculate the similarity between the high-dimensional semantic vector of the list text and the high-dimensional semantic vector of the three medical insurance catalogs in step 2.6), and the matched catalog entry vectors are sorted from high to low semantic similarity. The semantic association between vectors is calculated using the cosine similarity algorithm, as shown in the following formula:

[0026]

[0027] 3.4) Take the first K matching records from 3.3) (containing information such as list_id, medicine_id, similarity_score, medicine_type, and area_code, i.e., including: the unique identifier of the target vector (e.g., list_id); the unique identifier of the matched directory vector (medicine_id); the similarity score (e.g., cosine similarity); and the metadata of the matching vector (e.g., medicine_type, area_code), as candidate matching results, and perform consistency verification on the candidate matching results; 3.5) The verification and output module obtains the medical insurance catalog details corresponding to the catalog ID and candidate matching result from the first storage module, performs three-layer verification on the candidate matching result, outputs the verification result, and integrates the semantic similarity with the three-layer verification result; 3.6) Calculate a comprehensive confidence score for each record based on semantic similarity and the three-level verification results. The comprehensive confidence score is divided into three types: high, medium, and low. Based on the comprehensive confidence score, the unique optimal matching result selected from the candidate matching results is taken as the matching success result. If the comprehensive confidence score is high, the matching is determined to be successful. If the comprehensive confidence score is medium or low, the matching is determined to be unsuccessful. 3.7) After a match fails, the verification and output module sends the medical list data that fails to match as a vector to the preprocessing module for matching with the lower confidence list text data. The reason for the match failure is manually determined, and the manual operation in step four is performed according to the reason. The content after confirmation, modification or supplementation is used as the matching result. 3.8) Output the successfully matched results and the associated training dataset as the results, and output the confidence score and verification report of the matched records; IV. Manual Operation 4.1) When a match fails as determined in 3.6), the matching records with medium confidence are manually reviewed based on the low confidence list text data of the vector matching. The review operation includes confirmation, modification, or supplementation. When supplementation is required, manual supplementation is performed in the preprocessing module. After manual operation, the final matching result is changed. 4.2) When a match fails as determined in 3.6), the low-confidence matching records are manually corrected based on the low-confidence list text data of the vector matching. The correction operation includes modification and supplementation. When supplementation is required, manual supplementation is performed in the preprocessing module. After manual correction, the final matching result is changed. 4.3) Based on the vector matching of the low-confidence list text data, the manual search is performed on the medical insurance catalog data in the first storage module using (ES) word segmentation search based on the content of the failed reconfirmation and the low-confidence list text data to identify whether the medical insurance catalog data failed to match or was missing during the claim process due to the lack of coverage of drugs, treatment items and medical materials. If the missing data is identified, the operation of step 4.4) is continued. If the corresponding matching information is found, the operation of step 4.5) is performed directly. 4.4) Based on vector matching of low-confidence list text data, manually entered data information is added to the medical insurance catalog data. (Preprocessing module) New medical insurance catalog data is generated and stored in the medical insurance three-catalog structured database in the first storage module. The specific content generated from the new medical insurance catalog data can be understood as follows: Missing item location: Identify drugs, treatment items, and medical materials not covered in the three-directory database of the storage module (MySQL); Accurate field completion: Refer to the field format of the three catalogs and manually enter the complete information of missing items (such as the co-payment ratio and price limit of drugs). Catalog update: Synchronize the supplementary data to the medical insurance catalog data in the storage module (MySQL) to generate new medical insurance catalog data; The first storage module uses a MySQL database to uniformly store the new three-category database and the standardized three-category database (such as reimbursement level and payment level being unified as cost level). The standardized three-category database distinguishes the catalog type by medicine_type, fills in area_code to specify the medical insurance region, completes batch data writing, synchronizes the manually supplemented drug, diagnosis and treatment and material data in the new medical insurance catalog data, matches the original records according to medicine_id, updates the corresponding field information, and provides query support, that is, provides query interfaces by area_code (medical insurance region), medicine_type (catalog type), medicine_name (project name) and other dimensions, and provides data retrieval support for the manual data entry module.

[0028] In this application, the fields in the drug database are: generic name, drug name, dosage form, cost level, co-payment ratio, price limit, reimbursement restriction, and remarks; the fields in the treatment database are: treatment name, cost level, standard unit price, price limit, co-payment ratio, and remarks; and the fields in the material database are: material name, cost level, standard unit price, price limit, co-payment ratio, and remarks.

[0029] 4.5) Manually retrieve the medical insurance catalog data of the current region stored in the first storage module, and associate the list text data with the low confidence of vector matching with the new medical insurance catalog data of the current region (medical insurance catalog data after manual supplementation) or directly associate it with the medical insurance catalog data with uncovered data that has not been identified, to generate list text-medical insurance catalog data association pairs; This can be understood as manually referencing the medical insurance catalog data for the current region in the first storage module, and then matching the vector-matched unrecognized list text data with the medical insurance catalog data to generate a list text-medical insurance catalog data association pair; 4.6) Then (medium confidence processed samples and low confidence processed samples) perform two-level verification on the list text-medical insurance catalog data association pair and the data that has been successfully reconfirmed. Mark the content that has been successfully verified at both levels as training samples to generate an association training dataset. In simple terms, after a failure is detected, during the manual cause identification process: the matching records with medium confidence are reconfirmed. If the match is correct, but the model only determines its confidence level to be medium, then it is considered a successful match without modification or supplementation. After passing two sets of verification, it is used as a sample for model training and output as the result. If the matching record with medium confidence is indeed a mismatch, but a corresponding matching result exists in the vector database, and the mismatch is the cause of the match failure, then no supplementation is needed. The matching result is directly modified, and after passing two sets of verification, it is used as a sample for model training and the final result is output. For low-confidence matching results, if there is a corresponding matching result in the vector database, the matching result is manually modified, and after passing two levels of verification, it is output as the result and then used as a sample for model training. Manual supplementation is performed for medium-confidence matching failures caused by a lack of matching objects in the vector database, as well as for low-confidence outputs due to a lack of matching objects in the vector database or no match found (which is considered a low-confidence match). Essentially, if a match with medium confidence fails, there is a reconfirmation process before determining whether to directly modify or supplement the data. However, for low confidence matches, no reconfirmation is required; the data can be directly modified or supplemented manually.

[0030] Additionally, it should be noted that the preprocessing module includes a supplementary entry and matching process. The logic of this process is as follows: if the matching fails, manual verification is performed. If it is not confirmed, that is, if modification or supplementary entry is required, modification means that there is corresponding data in the original database, so matching is performed directly to obtain the associated matching pair. If no corresponding matching content is found after the search, that is, the uncovered data is identified, then manual supplementary entry is performed, and then manual association matching is performed.

[0031] The first storage module actually needs to receive two responses from the preprocessing module. The first is to manually identify whether there is any uncovered data in the database, and the second is to provide data information when matching is performed after the supplementary recording is completed.

[0032] Then, the verified list text—medical insurance catalog data association pair and the confirmed medium-confidence matching content are labeled as training samples and matching results to generate an association training dataset; The generation of the associated training dataset can be understood as follows: Associative data entry: Based on vector matching of unrecognized list data, manually refer to the medical insurance catalog data of the current region in the first storage module, and associate and match the list text with the medical insurance catalog data to generate a list text-medical insurance catalog data association pair; Specifically, the case claims system provides an input interface with an embedded regional medical insurance catalog drop-down function. When inputting data, the system automatically retrieves catalog data from the first storage module (MySQL) of the current region, assisting operators in associating and matching the list text with the medical insurance catalog data.

[0033] The list text and medical insurance catalog data associated with the manually supplemented data are labeled as training samples to generate an associated training dataset.

[0034] V. Model Improvement 5.1) Preprocessing and sample augmentation of the associated training dataset using the automatic training module; The automatic training module receives manually processed data from the preprocessing module, containing two types of samples (both with manually labeled matching tags): Medium confidence level processing samples: Approved samples: list data + original matching catalog data + marked as valid match; Adjusted samples: list data + manually replaced catalog data + marked as valid match.

[0035] The meaning of "review adjustment" in the medium confidence level processing sample is equivalent to including the two processing methods in the low confidence level processing sample below, namely existing content and new content.

[0036] Low-confidence samples are handled as follows: Existing samples in the catalog: list data + manually selected catalog data + valid label matching; New samples in the catalog: list data + manually maintained (new) catalog data + valid label matching.

[0037] All samples must include the following core fields: list_text (list-enhanced text), directory_text (directory-enhanced text), data_type (data type: medicine / treatment / materials), area_code (region code), etc.

[0038] 5.2) Divide the processed data into training and validation sets, and input the training set into the Med-BERT model for incremental update training; In this application, the partitioning of the training set and the sample set is specifically handled as follows (using a stratified random partitioning method to ensure that the distribution of various types of samples in the training / validation set is consistent with the real business scenario): The training and validation sets are divided according to a set ratio, while satisfying the following constraints: Stratification dimensions: Two-tier stratification based on data type (drugs / treatments / materials) and sample source (medium confidence / low confidence) to ensure that the proportion of each type of sample in the training and validation sets deviates from the original sample distribution by ≤5%; New catalog sample bias: Low confidence - the proportion of new catalog samples in the validation set is increased to 35% (higher than the basic proportion), which strengthens the assessment of the matching ability of the new catalog; Evenly distributed across regions: The proportion of samples coded for each region in the training and validation sets varies by ≤3%, avoiding excessive concentration of samples in a particular region; Incremental update training includes: Training task configuration: Main task: Input list text + directory text concatenation sequence, predict matching label (1 = valid match), directly align with business goal, i.e. learn the matching relationship between list and directory, and ensure that training goal and online inference goal are consistent; Auxiliary task: Masked Language Modeling (MLM) task (reuse the medical terminology masking strategy in the pre-training stage, accounting for 10%), to enhance the model's understanding of medical terms. Hyperparameters and training strategies: Parameter freezing: The first 6 Transformer layers are frozen, and only the last 6 layers, pooling layer, and classification layer are fine-tuned. The first 6 layers learn the semantics of general medical terminology, which is the basic knowledge for the medical insurance scenario. Freezing these layers can prevent the original general knowledge from being overwritten. The last 6 layers learn scenario-specific features, and fine-tuning can accurately adapt to the scenario-specific needs of newly added samples. At the same time, only some parameters are fine-tuned, improving training efficiency.

[0039] Learning rate: Given the small number of new medical insurance samples, a large learning rate can easily lead to overfitting the model to these new samples. Cosine annealing decay allows the learning rate to decrease smoothly, ensuring both rapid learning of new samples in the early stages and stable convergence in the later stages, avoiding oscillations. The initial learning rate is 1e-6, gradually decaying to 5e-7 during training.

[0040] Batch weight skew: The weight of newly added directory samples is multiplied by 1.2. Low confidence - newly added directory samples are the core objective of this training. Increasing their weight within the batch allows the model to learn the features of these samples first, ensuring the matching ability of new directories.

[0041] Knowledge distillation: The original model is a stable model validated online. Distillation loss is used to make the output of the new model as close as possible to the original model. The original model is used as the teacher model, and distillation loss (weight 0.3) is added to the loss function to avoid catastrophic forgetting.

[0042] Early stopping mechanism: This prevents the model from overfitting to new samples and ensures generalization ability. If training continues, the model will remember the details of the new samples. Training automatically stops when the semantic matching accuracy on the validation set fails to improve for two consecutive rounds.

[0043] 5.3) The trained model is evaluated using a three-tiered approach: basic metrics, specific metrics, and comparative testing. (This ensures that the core metrics of the new model (i.e., semantic matching accuracy) are not lower than 98% of the baseline, allowing for minor fluctuations to balance iterative effectiveness and business stability.) If the evaluation fails, training is automatically terminated, and the original model version is retained. If the evaluation passes, the trained model version is retained.

[0044] Preferably, in step 2.1), the medical insurance data covers three major categories of catalogs: drug catalog, treatment item catalog, and medical material catalog. By constructing a region-year dual-dimensional classification label system and using dictionary mapping rules, medical insurance data of different formats are classified and integrated according to their respective regions to form a structured medical insurance integrated dataset containing region-year dual-dimensional classification labels. This application can be understood as follows: Based on the dimensional differences in medical insurance policies across regions, a two-dimensional classification label system of region and year is constructed. Through dictionary mapping rules, medical insurance data in different formats are classified and integrated according to their region (e.g., Beijing, Shanghai) + the year of data generation (e.g., 2024, 2025) to form a structured intermediate set of medical insurance data. The fields obtained are as follows: medical catalog code, generic name of drug, registered name, specifications, packaging quantity, packaging unit, cost level, cost ratio, reimbursement restrictions, etc.

[0045] In step 2.2), the cleaning and standardization process includes: 2.2.1) Unified field format: Date fields are converted to a unified format; numeric fields (reimbursement ratio, unit price limit, etc.) retain two decimal places; and text fields remove full-width / half-width spaces and special characters (such as ×, #, @). 2.2.2) Unified Terminology Mapping: Based on the "National Medical Service Price Item Specification" and "Naming Principles for Generic Drug Names", a terminology mapping dictionary is constructed to unify regional characteristic terms, clinical colloquialisms, and abbreviations into nationally common standard terms. For example, intramuscular injection is mapped to intramuscular injection, intravenous drip is mapped to intravenous infusion, and the separation of cefixime dispersible tablets (Cefixime) retains the association between cefixime dispersible tablets (generic name) and Cefixime (trade name). 2.2.3) Classification and splitting: The structured medical insurance integrated dataset is split into three subsets according to type: drugs, treatment items, and medical materials; 2.2.4) Deduplication and Merging: Merge duplicate data with the same name and specifications within the same subset (such as drugs with the same generic name and specifications); 2.2.5) Field unification: Standardize heterogeneous field names from different regions; 2.2.6) Catalog construction: Organize subsets according to the fields required for claims to form a standardized catalog (such as unifying reimbursement level and payment level as expense level).

[0046] In addition, a version management design for the catalog library is added to record the update operations of the catalog in each region (adding entries, deleting entries, modifying field content), and to add version tags and timestamps to each catalog data; it supports incremental updates and historical version backtracking of the database (catalog data at any point in time can be queried for review of historical claims cases), forming a version management mechanism.

[0047] Preferably, the standardization process in step 2.4) includes text splicing, standardization processing, and invalid information removal, specifically: 2.4.1) Text splicing and standardization processing: The structured data and core fields of the three medical insurance catalogs are concatenated into enhanced text. Among them, the concatenation of drug categories is as follows: generic name, dosage form, specifications, cost level, co-payment ratio, price limit, reimbursement restrictions, and region (example: Amoxicillin capsules 0.5g / tablet, Category B, 0.250 yuan, limited to inpatient care in Beijing). The concatenation of diagnosis and treatment / material categories is as follows: item name, cost level, standard unit price, price limit, co-payment ratio, and region (example: blood routine examination, Category B, 20 yuan, 20 yuan, 0.1 yuan in Shanghai). The medical bill data is concatenated with the core fields to form enhanced text. The core fields are: item name, dosage form, specifications, usage scenario, treatment area, single item amount, and total cost (example: Amoxicillin capsules 0.5g / tablet, hospitalization in Beijing, 10 yuan, 50 yuan). 2.4.2) Invalid Information Removal: Delete invalid content such as special symbols, redundant spaces, advertising logos, and hospital internal codes from the text; The specific content of step 2.5) is as follows: 2.5.1) Load the Med-BERT model that has been pre-trained in the medical field, with architectural modifications and embedding layer enhancements. 2.5.2) The preprocessed enhanced text is segmented into words. The segmented sequence is input into the Med-BERT model encoder (INT8 quantization optimization is used to reduce memory usage) for deep semantic encoding and feature fusion to capture the core semantic information in the text. 2.5.3) The Med-BERT model adopts a weighted pooling strategy optimized for the medical field to generate 768-dimensional semantic vectors; the generated vectors are subjected to L2 normalization to ensure uniform vector magnitude.

[0048] In this application, a Med-BERT custom word segmenter is used to segment the preprocessed enhanced text, accurately identifying medical terms, drug names, specifications, etc.

[0049] Preferably, in step 3.1), the standardization process specifically includes: 3.1.3) Invalid information removal: Use regular expressions to match and delete irrelevant content such as advertising logos, hospital internal codes, doctor signatures, and remarks from the text; 3.1.4) Terminology Standardization: Call the pre-set medical terminology standardization dictionary to map non-standard expressions to standard terms (e.g., complete cephalosporin as cephalosporin antibiotics, and standardize surgical sutures as medical sutures). 3.1.5) Semantic enhancement processing: Concatenate auxiliary information such as dosage form, specifications, and usage scenarios with the name text to generate enhanced list text (Example: Amoxicillin Capsules 0.5g / tablet, Hospitalization in Beijing). The content of step 3.3) is as follows: 3.3.1) Similarity Algorithm Selection: The cosine similarity algorithm is used to calculate the semantic association between vectors; 3.3.2) Matching process: Read the list text vector to be matched. In the medical_directory_vectors set, filter according to area_code (prioritize matching the same region) and medicine_type (ensure the type is consistent) to narrow down the matching range. For the filtered directory vectors, use the cosine similarity algorithm to calculate the similarity score with the list text vector. 3.3.3) Sorting: Sort the matching results from high to low according to the similarity score.

[0050] Preferably, the specific content of step 3.4) is as follows: 3.4.1) Take the first K records as candidate matching results, including information such as list_id, medicine_id, similarity_score, medicine_type, and area_code; 3.4.2) Vector Dimension Validation: Check if the generated vector has 768 dimensions. If there is a dimension anomaly, re-vectorize the vector. 3.4.3) Meta-information matching and verification: Verify whether the meta-information such as medicine_id, list_id, and area_code corresponding to the vector are consistent with the original data. If they are inconsistent, re-vectorize the data. 3.4.4) Validate storage results: Query the Milvus vector database to confirm that the vector data has been successfully written and the index is built normally. If it is not normal, re-vectorize the data. 3.4.5) Matching result verification: Randomly select some matching results and manually verify the reasonableness of the similarity scores to ensure the accuracy of the matching algorithm.

[0051] Preferably, step 3.5) includes the following: 3.5.1) The basic information verification includes: strong type matching, project type matching, region priority matching, and data integrity. As a pre-filtering layer, it quickly eliminates fundamentally erroneous records, laying the foundation for subsequent amount verification.

[0052]

[0053] 3.5.2) Three-level progressive amount actuarial verification: The calculation is carried out at three levels: detailed list level, project summary level and invoice summary level, with the strictness decreasing. Violations at the detailed level have the greatest impact, while reasonable error accumulation is allowed at the summary level.

[0054] Level 1: List Detail Level Validation Accurate verification of individual details is the core guarantee for the accuracy of medical insurance settlement; violations exceeding the tolerance limit will result in significant point deductions.

[0055]

[0056] Level 2: Project Summary Level Validation The summary verification of all details under the same project means that violations will affect the confidence level of all details under the project, but the penalty is lighter than that at the detail level.

[0057]

[0058] Level 3: Invoice Summary Level Verification The core requirements for closed-loop verification of the entire invoice amount are consistency of the total amount and basic accuracy of the classification and summary, with the lightest penalties for violations.

[0059]

[0060] 3.5.3) Business logic and policy compliance verification includes the following: Reimbursement restrictions, correlation between diagnosis and treatment and drugs, verification of the validity period of the medical insurance catalog, compliance of medication / treatment frequency and verification of special drugs; A supplementary verification layer is added to verify whether the matching results conform to the logic of medical insurance policies, thereby further improving the credibility of decision-making.

[0061]

[0062] 3.5.4) Assign weights to the verification results: The specific operation is as follows: the weight ratio of the three-level amount actuarial verification is 0.7, the level score = (1 - penalty coefficient) × level weight; the total amount score = the sum of the scores of each level, the weight ratio of the basic information verification is 0.15, the basic score = the sum of the scores of each rule × 0.15, the weight ratio of the business logic compliance verification is 0.15, the compliance score = the sum of the scores of each rule (single rule score = (1 - deduction ratio) × rule weight ratio).

[0063] Preferably, the specific content of step 3.6) is as follows: 3.6.1) Calculation of overall confidence level The total rule matching score = Level 3 actuarial score + basic information score + business logic compliance score, ranging from 0 to 1; Weighted semantic similarity = semantic similarity score × total rule matching score; Correction item for violation transmission = 1 - (List-level violation deduction ratio × 0.6 + Project-level violation deduction ratio × 0.3 + Invoice-level violation deduction ratio × 0.1); Final overall confidence score = semantic similarity weighted × violation propagation correction term; 3.6.2) Setting dynamic thresholds and decision logic The confidence levels at each level are configured based on the complexity of the policy claims, the strictness of the region, and the precision of the insurance company's verification, to arrive at the final decision state. The default system confidence threshold is set as follows: high confidence threshold ≥ 0.90, medium confidence threshold 0.80-0.90, and low confidence threshold < 0.80. Among them, high confidence: score ≥ high threshold, decision status = successful match, can be directly used for case claim entry and exit without manual review; Medium confidence level: Medium threshold ≤ score < high threshold, decision status = matching failure, manual review required, push verification report, mark the project / invoice level violation or compliance verification failure points; Low confidence: Score < medium threshold, decision status = match failed, highlight the list of violations for manual correction; 3.6.3) Tie handling: When the score difference is ≤0.01, a tie-breaking mechanism is adopted. Priority 1: No violations at the list level > Within the list level tolerance > Exceeding the list level tolerance; Priority 2: No violations during compliance verification > Minor violations only > Major violations; Priority 3: Regional exact match > Nationwide general match; Priority 4: The latest updated catalog record.

[0064] Preferably, the specific content of step 4.5) is as follows: The specific content of the two-level verification in step 4.6) is as follows: 4.6.1) Field integrity verification: Force verification of core fields; if any field is missing, it will be manually entered automatically, with specific prompts. 4.6.2) Format conformity verification: In terms of text format, the text must be fully aligned with the vectorized input format of the data processing module; otherwise, a pre-check of text standardization will be automatically triggered. In terms of label format, match_label only allows valid matches, data_type only allows drugs / treatments / materials, and area_code must be a standard 6-digit administrative division code; otherwise, manual correction will be required.

[0065] The specific content of step 5.1) is as follows: 5.1.1) Standardized alignment, the specific content of which is as follows: Text processing: The same text processing logic is reused with the data processing module for cleaning and standardization (removing redundant spaces, unifying punctuation formats, and standardizing unit expressions, such as unifying g as the abbreviation of gram). Label mapping: The sample type labels such as medium confidence - after review and adjustment and low confidence - newly added directory are mapped to specific labels used for model training, which facilitates subsequent classification evaluation; 5.1.2) Targeted sample augmentation, the specific content of which is as follows: For low-confidence - newly added catalog samples: perform medical term replacement enhancement on the list text (e.g., replace "admission" with "hospitalization", "0.5g / tablet" with "0.5g per tablet"), generate 2 enhanced samples, and improve the model's generalization ability to the new catalog; For samples with medium confidence level after review and adjustment: retain the original samples without enhancement (to avoid interfering with the accurate matching relationship after manual correction); 5.1.3) Deduplication filtering: Compare with the historical training set (the effective sample library of the last 3 training sessions), and determine duplicate samples by text similarity (cosine similarity ≥ 0.95). Duplicate samples are directly deleted and do not enter the training process. Preferably, the specific content of step 5.3) is as follows: 5.3.1) Basic evaluation metric: Semantic matching accuracy on the validation set ≥ 98% of the accuracy of the original model; 5.3.2) Specific evaluation indicators: Regarding the accuracy of newly added catalog samples: the matching accuracy of low-confidence newly added catalog samples is ≥92%; Regarding the accuracy of the adjusted sample: the matching accuracy of the adjusted sample at medium confidence level is ≥95%; 5.3.3) Comparison test with the original model: Using the model's validation dataset, randomly select several high-confidence samples and run the new model and the original model simultaneously. Compare the vector similarity score distribution and the consistency of the Top5 matching results between the two models. Among them, the Top5 matching results of samples with a mean deviation of scores between the new model and the original model ≤ a preset threshold (0.01-0.05) and a Top5 consistency ≥ a preset threshold (95%-98%) are consistent with the original model.

[0066] Additionally, it should be noted that all preset thresholds in this application exist independently, rather than taking the same value; they are simply a collective term.

[0067] 5.3.4) Secondary verification of sample labeling: Before the model is automatically evaluated, 10% (10%-20%) of the training samples are selected, prioritizing newly added catalogs and data that have been reviewed and adjusted, and pushed to senior medical insurance auditors for secondary manual annotation and review. Review content: Confirm whether the matching relationship between list_text and directory_text is correct, and whether the sample type labels are accurate; Review result processing: If the review finds that the annotation error rate is ≥ the preset threshold (4%-8%), the samples will be re-screened and the manual annotation review process will be repeated; if the error rate is < the preset threshold (4%-8%), the erroneous samples will be corrected and the automatic evaluation will continue.

[0068] 5.3.5) Implement standardized version management and canary deployment for the new model that has passed the evaluation to ensure system stability: Version number format: V + date + sample size + accuracy (e.g., V20260115_2000_95.2 represents training on January 15, 2026, 2000 samples, and an accuracy of 95.2%). Archived content includes: model files, training data (training set + validation set), evaluation report, and hyperparameter configuration, all stored in a Git repository, supporting retrieval and backtracking by version number.

[0069] Among them, the optimized Med-BERT model is updated to the data processing module for subsequent vectorization processing of medical insurance data; Model training report: includes training data size, training / validation set split results, model evaluation metrics (accuracy, F1 score, etc.), and version identifier; Training data archiving: Store the divided training / validation sets in a specified database for model backtracking and retraining.

[0070] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention uses a self-developed Med-BERT model (Medical Insurance Data Semantic Vectorization Processing Model) as its core algorithm module. Independent of the data flow process, it specifically implements semantic vectorization processing of medical insurance three-category database data and medical list identification details, and provides standardized vector similarity scoring capabilities for the verification module. Through customized modifications for the medical field, the module addresses the shortcomings of the general BERT model in understanding the semantics of medical terminology and its weak ability to extract features from short texts, achieving high-precision semantic representation and correlation calculation of medical insurance data.

[0071] 2. This application solves the problems of rigid matching, reliance on external interfaces, and high deployment costs of existing solutions by integrating multi-regional catalog databases into vectorization, using Med-BERT small model semantic matching, and automatic training learning closed-loop architecture. This achieves accurate, efficient, and secure matching of medical list details with the three medical insurance catalogs. Attached Figure Description

[0072] Figure 1 This is a logical diagram showing the information interaction between the modules in this invention; Figure 2 This is an overall flowchart of the method in this invention; Detailed Implementation The following will refer to the appendices in the embodiments of the present invention. Figure 1-2 The technical solutions in the embodiments of the present invention are clearly and completely described herein. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0073] Please see Figure 1-2 Embodiments of the present invention: Example: In this embodiment, specifically: (a) Domain Adaptive Pre-training 1.1) Basic model selection and lightweight adaptation BERT-Base-Chinese was chosen as the base model, with the following core parameters: 110M parameters, 768 hidden layer dimensions, 12 Transformer encoder layers, 12 attention heads, and a maximum input sequence length of 512. To meet local deployment requirements, INT8 quantization was used to optimize the base model for lightweight processing. After quantization, the model's memory usage is ≤16GB, and inference latency is reduced by 40%, satisfying the performance requirements for batch data processing.

[0074] 1.2) Corpus Construction and Preprocessing in the Medical Field A 320GB pre-trained corpus specifically for the medical field was constructed, covering all dimensions of text in medical insurance scenarios, and standardized preprocessing was completed, including: ① Core fields; ② Core medical journal research literature: Based on the PubMedCentral Chinese dataset, CNKI dataset, etc., non-clinically related content is filtered out, and paragraphs related to diagnosis and treatment and pharmacology are retained; ③ National / Provincial / Municipal Medical Insurance Catalog Data: Standardize field naming and expand regional coding information; ④ Clinical Practice Guidelines / Claims Cases: Structured analysis to extract text related to treatment items, indications, and medical insurance policies, as well as anonymized insurance claims data containing personal privacy information.

[0075] After preprocessing, the corpus was deduplicated using TF-IDF (similarity threshold ≥ 0.95), and finally 220GB of effective corpus was retained to ensure the diversity and effectiveness of the pre-training data.

[0076] 1.3) Customized Masked Language Modeling (MLM) Training The core strategies for optimizing MLM tasks based on medical text features are as follows: ① Entity-priority masking rules: The masking probability of core entities such as drug name and treatment item name is increased to 15% (10% for general text); the masking probability of key attributes such as dosage form, specification, and region is set to 10%; whole word masking accounts for 70% and single word masking accounts for 30% to avoid invalid splitting of amoxicillin capsules into amoxicillin, xicillin, etc.

[0077] ② Training hyperparameter configuration: learning rate 2e-5, batch size=32, number of training epochs=10, optimizer selected as AdamW, weight decay coefficient 0.01, and the accuracy was verified by understanding medical terminology after training (accuracy ≥95%).

[0078] 1.4) Incremental pre-training mechanism To address the annual updates to the medical insurance catalog, an incremental pre-training strategy was designed: incremental MLM training was performed annually based on newly added catalog data, with the learning rate reduced to 1e-6 and the number of training rounds reduced to 2. This approach aims to prevent the model from forgetting historical knowledge and maintain its adaptability to the latest medical insurance policies.

[0079] (II) Professionalization of Architecture In response to the characteristic that medical insurance texts are mainly short texts (average length ≤ 20 characters), the original BERT architecture was deeply customized: 2.1 Removal of redundant modules By removing the original BERT's sentence-level pooling layer and next sentence prediction (NSP) task module, the interference of long text dependency mechanism on short text feature extraction is reduced, and the model's inference speed is improved by 15%.

[0080] 2.2 Design of Multi-Granularity Phrase Pooling Mechanism A new multi-granularity phrase pooling module for entity perception has been added, and its technical implementation is as follows: ① Medical entity recognizer training: Based on the labeled medical entity dataset (labeled with 6 types of entities: drug ingredients, dosage form, specifications, scene, and region), the entity recognizer is trained using BiLSTM-CRF. The entity recognition F1 score is ≥94%, accurately locating the core entity positions in short texts.

[0081] ② Multi-granularity pooling strategy This strategy is a multi-granularity phrase pooling process for short medical insurance texts in the Med-BERT model. Specifically, it takes the token-level semantic features of the medical insurance text output by the Med-BERT encoder as input and feeds them into the entity recognizer to identify and locate the core medical entities in the text (including drug ingredients, dosage forms / specifications, scenarios / regions, etc.). Then, a differentiated pooling strategy is adopted for different types of core entities—character-level pooling is performed on drug ingredient entities to retain their semantic details, word-level pooling is performed on dosage form / specification entities to fuse related semantic information, and phrase-level pooling is performed on scenario / region entities to retain contextual features. Subsequently, the features after these three types of pooling are input into the feature weighting and fusion module. Through learnable weight parameters, the integration of multi-granular features is completed, and finally, short text semantic features that fuse the information of each core entity are generated and output, thereby achieving accurate semantic representation of short medical insurance texts.

[0082] Drug ingredients (e.g., amoxicillin): character-level pooling preserves semantic details of the ingredients; Dosage form / specification (e.g., 0.5g / tablet): word-level pooling, fusing dose-unit semantics; Scenario / Region (e.g., hospitalization in Beijing): Phrase-level pooling to preserve contextual relevance; Weighted fusion: Initial weights are assigned to features of different granularities (0.5 for ingredients, 0.3 for dosage form / specification, and 0.2 for scene / region), and the weights are automatically optimized during training.

[0083] ③ Attention mechanism optimization: Adjust the Transformer attention mask for short texts: limit the attention window size to 16 (covering the maximum length of short medical insurance texts), reduce invalid attention calculations, and increase the attention weight of core entities (e.g., increase the attention weight of drug names by 20%).

[0084] 2.3 Embedding Layer Enhancement Based on the standard BERT embedding layer, a new entity type embedding layer is added to construct a 3D embedding and fusion system: ① Embedded layer architecture design

[0085] ② Embedding, fusion, and mapping The three types of embeddings are concatenated according to their dimensions to obtain concatenated features with dimensions of 768+768+128=1664. The concatenated features are mapped back to 768 dimensions through a linear transformation layer. An independent learning rate (5e-5) is set for the entity type embedding layer to ensure that it effectively learns medical entity type features during training.

[0086] ③ Embedding layer effect verification Comparative experiments show that after adding entity type embedding, the model's semantic differentiation accuracy between amoxicillin capsules (dosage form) and amoxicillin granules (dosage form) is improved, effectively solving the problem of confusion between medical entity types.

[0087] 2.4 Feature Fusion Optimization Instead of simple feature concatenation, a hierarchical intelligent weighted fusion strategy is designed to achieve the optimal combination of global semantics and domain features: ① Feature hierarchy division Extract features from different levels of the Transformer encoder and divide them into three core feature categories:

[0088] ② Hierarchical weighted fusion model Learnable weight parameters α (global semantics), β (entity perception), and γ (hierarchical aggregation) are assigned to the three types of features, satisfying α+β+γ=1. The weights are automatically optimized using the loss function of a medical text semantic matching task. After fusion, the feature representation capability is enhanced by ReLU activation function and Dropout (probability 0.1).

[0089] (III) Medical Data Vector Scoring System (The initial scoring system was manually formulated and then entered into the system) Manual scoring is not directly used as a vector parameter, but is instead used for: ① labeling high-quality training data to optimize the Med-BERT model; ② calibrating the similarity threshold to improve the accuracy of the verification module.

[0090] 3.1 Professional scoring team A professional scoring team of 7 members was formed, with the following composition and responsibilities: ① Three clinical physicians with the title of associate chief physician or above (one each from internal medicine, surgery, and traditional Chinese medicine): responsible for clinical functional equivalence scoring; ② Two specialists with over 5 years of experience in medical insurance claims: responsible for scoring semantic similarity of names and regional suitability; ③ Two medical data processing engineers: responsible for scoring the relevance of terminology standardization, and also organizing the scoring data as labels for model training.

[0091] 3.2 Multi-dimensional scoring criteria Multi-dimensional scoring standard development: The "Multi-dimensional Scoring Standard for Medical Data Correlation" was developed, clarifying four core scoring dimensions, their corresponding weights, scoring ranges, and rules, as detailed below: Name semantic similarity (weight 0.4, score range 0-10): 10 points for completely identical core terms (e.g., amoxicillin capsules and amoxicillin capsules); 8 points for identical core terms but different modifiers (dosage form, specifications) (e.g., amoxicillin capsules and amoxicillin dispersible tablets); 4 points for partially identical core terms (e.g., cefixime and cefdinir); 0 points for no overlapping core terms (e.g., amoxicillin and ibuprofen). Clinical functional equivalence (weight 0.3, scoring range 0-10 points): 10 points for completely identical clinical use and therapeutic effect (e.g., intravenous infusion and intravenous drip); 6 points for partial overlap of clinical functions (e.g., ordinary acupuncture and electroacupuncture); 0 points for no correlation in clinical functions (e.g., blood routine examination and B-ultrasound examination). Terminology standardization relevance (weight 0.2, scoring range 0-10 points): 10 points for fully conforming to the national unified medical terminology standard; 7 points for conforming to the regional medical insurance catalog standard terminology; 4 points for non-standard terminology but widely used in clinical practice; 0 points for obscure expressions without any standard basis. Regional adaptability (weight 0.1, score range 0-10): 10 points for entries that are completely consistent with the region of the list; 7 points for general entries that are applicable across regions; 3 points for region-specific entries (included only in one province).

[0092] Medical data vector quantization calculation: A professional team scores the pairwise correlation of all entries in the catalog, and the final correlation score for each piece of related data is calculated using a weighted average formula. S = W1×S1 + W2×S2 + W3×S3 + W4×S4 Where S is the final association score (ranging from 0 to 10), W1-W4 are the weights of the four scoring dimensions (0.4, 0.3, 0.2, 0.1), and S1-S4 are the scores for the corresponding dimensions.

[0093] Example: If the scores for catalog item A (Amoxicillin Capsules, Beijing area, Class A) and item B (Amoxicillin Dispersible Tablets, nationwide use, Class A) are S1=8, S2=10, S3=10, and S4=7, then the final association score is S=0.4×8+0.3×10+0.2×10+0.1×7=8.9, which is the medical data vector parameter of the two items.

[0094] The text-to-human score was used as supervised data to fine-tune the Med-BERT model and improve the accuracy of vector representation. A similarity threshold was set based on the human score (e.g., a score ≥ 8 points corresponds to a cosine similarity ≥ 0.9) for the result filtering of the verification module.

[0095] Additionally, it should be noted that, Figure 1 Service caller and Figure 2 The insurance claims system can be understood as a client application.

[0096] Simulation verification results To verify the effectiveness of vector matching and candidate set selection, a comparative simulation experiment was designed. The experimental conditions and results are as follows: Experimental data: 1,000 real medical bill details were selected (350 drugs, 330 treatment items, and 320 medical supplies), covering various scenarios including standardized terminology, clinical colloquialisms, and regionally distinctive expressions; The experimental results of this application are as follows: Candidate set screening accuracy (the proportion of candidates containing correctly matched entries): This scheme achieves 92.3%; Average length of candidate set (after threshold filtering): 12.6 in this scheme; False negative rate (the proportion of correctly matched entries that do not enter the candidate set): This scheme has a false negative rate of only 1.2%.

[0097] Experimental results demonstrate that the candidate set screening effect of this scheme is significantly better than that of existing technologies, ensuring both high coverage and effectively reducing the size of the candidate set.

[0098] Comparative Example 1: An experiment was conducted using the following existing techniques: Medical bill information collection: Key information on the medical bill is identified through a multimodal big data model, and the names of drugs, treatment items or medical materials are extracted. At the same time, the regional information of the invoices corresponding to the bill is collected to determine the target region's medical insurance three-category database to be accessed.

[0099] Name text regular expression preprocessing: Multiple rounds of regular expression processing with fixed rules are performed on the name text identified by the large model. Each round focuses on removing specific redundant information. First round of processing: Remove punctuation marks (such as commas, periods, and pauses) from the name; Second round of processing: Remove parentheses and the content within them (such as (imported) (domestic) (10 tablets / plate), etc.). Third round of processing: Delete pure numbers and combinations of numbers and units (such as 0.5g, 10ml, 200mg, etc.). The fourth round of processing: unify text case, remove spaces and special characters (such as ×, #, etc.).

[0100] Multi-round exact match: After each round of regular expression processing, the processed name text is matched against the generic name, drug name, treatment name or material name in the target region's medical insurance catalog (i.e., the characters must be completely identical for a match to be considered successful).

[0101] Result processing: If a match is successful after a round of regular expression processing, the corresponding medical insurance level (Class A, B, C), medical insurance ratio, reimbursement restrictions, unit price limit, and other information are retrieved from the directory database; if no matching directory name is found after all rounds of regular expression processing, the match is deemed to have failed and is handed over to manual input for supplementary matching.

[0102] Candidate set screening accuracy (the proportion of candidates containing correctly matched entries): This pair's accuracy was 68.7%; Average length of candidate set (after threshold filtering): 28.3 items on average in this comparison. False negative rate (the proportion of correctly matched entries that did not enter the candidate set): 18.5% for this pair.

[0103] Experimental results demonstrate that the candidate set screening effect of Example 1 is significantly better than that of Comparative Example 1, which can ensure high coverage and effectively reduce the size of the candidate set.

[0104] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0105] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for intelligent matching of the three medical insurance catalogs based on the Med-BERT model, characterized in that, Includes the following steps: I. A medical insurance three-category intelligent matching system based on the Med-BERT model is provided, including a data receiving module, a preprocessing module, a first storage module, a data processing module, a second storage module, a verification and output module, and an automatic training module; II. Establishment of a High-Dimensional Semantic Vector Database for the Three Medical Insurance Catalogs 2.1) The data receiving module receives medical insurance data from multiple regions and processes the data to generate a structured integrated medical insurance dataset. 2.2) The structured medical insurance integrated dataset is cleaned and standardized through the preprocessing module to generate a standardized three-category library; 2.3) The standardized three-directory library is stored in the first storage module using a MySQL database. The standardized three-directory library is distinguished by directory type, the medical insurance region is identified, and it is entered into the original database to form a structured database of the medical insurance three-directory library. 2.4) The processing module identifies newly added or modified records in the structured data of the three medical insurance catalogs based on the updated_at field of the first storage module, obtains incremental data, classifies the incremental data according to the project type, and then performs standardization processing to generate a medical insurance standardized enhanced list text. 2.5) Input the enhanced list text of medical insurance regulations into the Med-BERT model in the processing module for high-dimensional semantic vectorization transformation to generate high-dimensional semantic vectors of medical insurance text; 2.6) The high-dimensional semantic vector of the medical insurance text is stored in the catalog vector set through the second storage module, and used as the catalog entry vector. A special index is built on the catalog vector set. III. Medical Data Matching 3.1) The system receives medical records containing image information through the data receiving module, extracts and standardizes the image information from the medical records, and generates a medical-enhanced list text. 3.2) Input the medical enhanced checklist text into the Med-BERT model in the processing module for high-dimensional semantic vectorization transformation to generate high-dimensional semantic vectors of the checklist text; 3.3) Then, the Med-BERT model is used to calculate the similarity between the high-dimensional semantic vector of the list text and the high-dimensional semantic vector of the three medical insurance catalogs in step 2.6), and the matched catalog entry vectors are sorted from high to low semantic similarity. 3.4) Take the first K matching records from 3.3) as candidate matching results, and perform consistency verification on the candidate matching results; 3.5) The verification and output module obtains the medical insurance catalog details corresponding to the catalog ID and candidate matching result from the first storage module, performs three-layer verification on the candidate matching result, outputs the verification result, and integrates the semantic similarity with the three-layer verification result; 3.6) Calculate a comprehensive confidence score for each record based on semantic similarity and the three-level verification results. The comprehensive confidence score is divided into three types: high, medium, and low. Based on the comprehensive confidence score, the unique optimal matching result selected from the candidate matching results is taken as the matching success result. If the comprehensive confidence score is high, the matching is determined to be successful. If the comprehensive confidence score is medium or low, the matching is determined to be unsuccessful. 3.7) After a match fails, the verification and output module sends the medical list data that fails to match as a vector to the preprocessing module for matching with the lower confidence list text data. The reason for the match failure is manually determined, and the manual operation in step four is performed according to the reason. The content after confirmation, modification or supplementation is used as the matching result. 3.8) Output the successfully matched results and the associated training dataset as the matching results, and output the confidence score and verification report of the matched records; IV. Manual Operation 4.1) When a match fails as determined in 3.6), the matching records with medium confidence are manually reviewed based on the low confidence list text data of the vector matching. The review operation includes confirmation, modification, or supplementation. When supplementation is required, manual supplementation is performed in the preprocessing module. After manual operation, the final matching result is changed. 4.2) When a match fails as determined in 3.6), the low-confidence matching records are manually corrected based on the low-confidence list text data of the vector matching. The correction operation includes modification and supplementation. When supplementation is required, manual supplementation is performed in the preprocessing module. After manual correction, the final matching result is changed. 4.3) Based on the vector matching of the low-confidence list text data, the manual search is performed on the medical insurance catalog data in the first storage module through word segmentation search based on the content of the failed reconfirmation and the low-confidence list text data to identify whether the medical insurance catalog data failed to match or was missing during the claim process due to the lack of coverage of drugs, treatment items and medical materials. If the missing data is identified, the operation of step 4.4) is continued. If the corresponding matching information is found, the operation of step 4.5) is performed directly. 4.4) Based on vector matching of low-confidence list text data, manually entered data information is added to the medical insurance catalog data to generate new medical insurance catalog data, which is then stored in the structured database of the three medical insurance catalogs in the first storage module. 4.5) Manually retrieve the medical insurance catalog data of the current region stored in the first storage module, and associate the list text data with the new medical insurance catalog data of the current region with the vector matching with the list text data with low confidence, or directly associate it with the medical insurance catalog data with uncovered data that has not been identified, to generate list text-medical insurance catalog data association pairs; 4.6) Then, perform two-level verification on the list text-medical insurance catalog data association pair and the data that has been successfully reconfirmed. Mark the content that has been successfully verified at both levels as training samples to generate an association training dataset. V. Model Improvement 5.1) Preprocessing and sample augmentation of the associated training dataset using the automatic training module; 5.2) Divide the processed data into training and validation sets, and input the training set into the Med-BERT model for incremental update training; 5.3) The trained model is evaluated in three layers: basic indicators, specific indicators, and comparative tests. If the evaluation fails, training is automatically terminated and the original model version is retained. If the evaluation passes, the trained model version is retained.

2. The intelligent matching method for the three medical insurance catalogs based on the Med-BERT model according to claim 1, characterized in that, In step 2.1), the medical insurance data covers three major categories of catalogs: drug catalog, treatment item catalog, and medical material catalog. By constructing a region-year dual-dimensional classification label system and using dictionary mapping rules, medical insurance data of different formats are classified and integrated according to their respective regions to form a structured medical insurance integrated dataset containing region-year dual-dimensional classification labels. In step 2.2), the cleaning and standardization process includes: 2.2.1) Standardize field formats: Date fields are converted to a standardized format, numeric fields are retained to two decimal places, and redundant content is removed from text fields; 2.2.2) Unified Terminology Mapping: Construct a terminology mapping dictionary to unify regionally distinctive terms, clinical colloquialisms, and abbreviations into nationally accepted standard terms; 2.2.3) Classification and splitting: The structured medical insurance integrated dataset is split into three subsets according to type: drugs, treatment items, and medical materials; 2.2.4) Deduplication and Merging: Merge duplicate data with the same name and specifications within the same subset; 2.2.5) Field unification: Standardize heterogeneous field names from different regions; 2.2.6) Directory construction: Organize subsets according to the fields required for claims to form a standardized directory.

3. The intelligent matching method for the three medical insurance catalogs based on the Med-BERT model according to claim 2, characterized in that, The standardization process in step 2.4) includes text splicing, standardization processing, and invalid information removal, specifically: 2.4.1) Text splicing and standardization processing: The structured data and core fields of the three medical insurance catalogs are concatenated into enhanced text. Specifically, for drugs, the concatenation includes generic name, dosage form, specifications, cost level, co-payment ratio, price limit, reimbursement restrictions, and region. For medical treatment / materials, the concatenation includes item name, cost level, standard unit price, price limit, co-payment ratio, and region. The medical bill data is concatenated with the core fields to form enhanced text. The core fields are: item name, dosage form, specifications, usage scenario, treatment location, single item amount, and total cost. 2.4.2) Invalid Information Removal: Delete invalid content such as special symbols, redundant spaces, advertising logos, and hospital internal codes from the text; The specific content of step 2.5) is as follows: 2.5.1) Load the Med-BERT model that has been pre-trained in the medical field, with architectural modifications and embedding layer enhancements. 2.5.2) The preprocessed enhanced text is segmented into words, and the segmented sequence is input into the Med-BERT model encoder for deep semantic encoding and feature fusion to capture the core semantic information in the text; 2.5.3) The Med-BERT model adopts a weighted pooling strategy optimized for the medical field to generate 768-dimensional semantic vectors; the generated vectors are subjected to L2 normalization to ensure uniform vector magnitude.

4. The intelligent matching method for the three medical insurance catalogs based on the Med-BERT model according to claim 3, characterized in that, In step 3.1), the specific content of the standardization process is as follows: 3.1.3) Invalid Information Removal: Use regular expressions to match and delete irrelevant content from the text; 3.1.4) Terminology Standardization: Call the pre-built medical terminology standardization dictionary to map non-standard expressions to standard terms; 3.1.5) Semantic enhancement processing: Concatenate auxiliary information such as dosage form, specifications, and usage scenarios with the name text to generate enhanced list text; The content of step 3.3) is as follows: 3.3.1) Similarity Algorithm Selection: The cosine similarity algorithm is used to calculate the semantic association between vectors; 3.3.2) Matching process: Read the list text vectors to be matched from the medical_list_vectors collection. In the medical_directory_vectors collection, filter according to area_code and medicine_type to narrow down the matching range. For the filtered directory vectors, use the cosine similarity algorithm to calculate the similarity score with the list text vectors. 3.3.3) Sorting: Sort the matching results from high to low according to the similarity score.

5. A method for intelligent matching of the three medical insurance catalogs based on the Med-BERT model according to claim 1, 2, 3 or 4, characterized in that the specific content of step 3.4) is as follows: 3.4.1) Take the first K records as candidate matching results, including information such as list_id, medicine_id, similarity_score, medicine_type, and area_code; 3.4.2) Vector Dimension Validation: Check if the generated vector has 768 dimensions. If there is a dimension anomaly, re-vectorize the vector. 3.4.3) Meta-information matching and verification: Verify whether the meta-information such as medicine_id, list_id, and area_code corresponding to the vector are consistent with the original data. If they are inconsistent, re-vectorize the data. 3.4.4) Storage result verification: Query the Milvus vector database to confirm that the vector data has been successfully written and the index is built normally. If not, re-vectorize the data. 3.4.5) Matching result verification: Randomly select some matching results and manually verify the reasonableness of the similarity score.

6. The intelligent matching method for the three medical insurance catalogs based on the Med-BERT model according to claim 5, characterized in that the specific content of step 3.5) includes: 3.5.1) The basic information verification includes: Strong type matching: The cost type of the list details must be exactly the same as the cost level in the catalog. If they are not the same, the candidate record will be excluded. Project type matching: The project to which the list details belong must match the project category marked in the catalog. If they do not match, the project type will be marked as mismatched, and points will be deducted subsequently. Region-priority matching: Prioritize matching records with identical region codes. If none are found, match the national general directory. If a cross-regional match is found, mark it and reduce the score for this dimension. Data integrity: Core monetary fields such as cost level / self-payment ratio / region code / limit have no empty values, are non-negative, and are in compliant format. Missing core fields will be directly classified as low confidence. 3.5.2) Three-level progressive amount actuarial verification: Calculated level by level, according to the three levels of list details, project summary and invoice summary; 3.5.3) Business logic and policy compliance verification includes the following: Reimbursement restriction matching: The reimbursement restrictions marked in the catalog are matched with the type of visit / diagnosis information on the list. If they do not match, the auxiliary rule score is deducted. Relevance between diagnosis and treatment and medication: The usage scenarios of the medication details are related to the main diagnosis and treatment items of this visit in the clinical knowledge graph. If there is no correlation, the score of the auxiliary rule will be deducted. Medical insurance catalog validity period verification: The date of visit on the list must be within the validity period of the catalog. If the catalog has no expiration date, it is valid by default. If it is expired, the auxiliary rule score will be deducted and the catalog will be marked as expired. Medication / treatment frequency compliance: The frequency of use of the details corresponding to the same catalog ID within one calendar month shall not exceed the upper limit stipulated by the medical insurance policy. If the frequency is exceeded, the auxiliary rule score will be deducted and the rationality needs to be manually verified. Special drug verification: Drugs used for special diseases marked in the catalog must have a corresponding disease diagnosis in the invoice medical record disease diagnosis information. If there is no corresponding disease, the auxiliary rule score will be deducted, and the drug will be directly marked and requires manual review. 3.5.4) Assign weights to the verification results: The specific operation is as follows: the weight ratio of the three-level amount actuarial verification is 0.7, the weight ratio of the basic information verification is 0.15, and the weight ratio of the business logic compliance verification is 0.

15.

7. The intelligent matching method for the three medical insurance catalogs based on the Med-BERT model according to claim 6, characterized in that the specific content of step 3.6) is as follows: 3.6.1) Calculation of overall confidence level The total rule matching score = Level 3 actuarial score + basic information score + business logic compliance score, ranging from 0 to 1; Weighted semantic similarity = semantic similarity score × total rule matching score; Correction item for violation transmission = 1 - (List-level violation deduction ratio × 0.6 + Project-level violation deduction ratio × 0.3 + Invoice-level violation deduction ratio × 0.1); Final overall confidence score = semantic similarity weighted × violation propagation correction term; 3.6.2) Setting dynamic thresholds and decision logic The confidence levels at each level are configured based on the complexity of the policy claims, the strictness of the region, and the precision of the insurance company's verification, to arrive at the final decision state. The default system confidence threshold is set as follows: high confidence threshold ≥ 0.90, medium confidence threshold 0.80-0.90, and low confidence threshold < 0.

80. Among them, high confidence: score ≥ high threshold, decision status = successful match, can be directly used for case claim entry and exit without manual review; Medium confidence level: Medium threshold ≤ score < high threshold, decision status = matching failure, manual review required, push verification report, mark the project / invoice level violation or compliance verification failure points; Low confidence: Score < medium threshold, decision status = match failed, highlight the list of violations for manual correction; 3.6.3) Tie handling: When the score difference is ≤0.01, a tie-breaking mechanism is adopted. Priority 1: No violations at the list level > Within the list level tolerance > Exceeding the list level tolerance; Priority 2: No violations during compliance verification > Minor violations only > Major violations; Priority 3: Regional exact match > Nationwide general match; Priority 4: The latest updated catalog record.

8. A method for intelligent matching of the three medical insurance catalogs based on the Med-BERT model, as described in claims 1, 2, 3, 4, 6, or 7, characterized in that, The specific content of the two-level verification in step 4.6) is as follows: 4.6.1) Field integrity verification: Force verification of core fields; if any field is missing, it will automatically return for manual entry, with specific prompts. 4.6.2) Format conformity verification: In terms of text format, the text must be fully aligned with the vectorized input format of the data processing module; otherwise, a pre-check of text standardization will be automatically triggered. In terms of label format, match_label only allows valid matches, data_type only allows drugs / treatments / materials, and area_code must be a standard 6-digit administrative division code; otherwise, manual correction will be required.

9. The intelligent matching method for the three medical insurance catalogs based on the Med-BERT model according to claim 8, characterized in that, The specific content of step 5.1) is as follows: 5.1.1) Standardized alignment, the specific content of which is as follows: Text processing: The same set of text processing logic is reused with the data processing module for cleaning and standardization. Label mapping: Map sample type labels such as medium confidence - after review and adjustment and low confidence - newly added directory to specific labels used for model training; 5.1.2) Targeted sample augmentation, the specific content of which is as follows: For low-confidence - newly added catalog samples: perform medical terminology synonym replacement enhancement on the list text, generate 2 enhanced samples, and improve the model's generalization ability to the new catalog; For the medium confidence level - adjusted sample: retain the original sample, do not enhance it; 5.1.3) Deduplication filtering: By comparing with the historical training set, duplicate samples are determined by text similarity. Duplicate samples are directly deleted and do not enter the training process.

10. The intelligent matching method for the three medical insurance catalogs based on the Med-BERT model according to claim 9, characterized in that, The specific content of step 5.3) is as follows: 5.3.1) Basic evaluation metric: Semantic matching accuracy on the validation set ≥ 98% of the accuracy of the original model; 5.3.2) Specific evaluation indicators: Regarding the accuracy of newly added catalog samples: the matching accuracy of low-confidence newly added catalog samples is ≥92%; Regarding the accuracy of the adjusted sample: the matching accuracy of the adjusted sample at medium confidence level is ≥95%; 5.3.3) Comparison test with the original model: Using the model's validation dataset, run the new model and the original model simultaneously, and compare the vector similarity score distribution and the consistency of the Top5 matching results between the two. Among them, the mean deviation of the scores between the new model and the original model is ≤0.02, and the Top5 matching results of samples with a Top5 consistency of ≥95% are consistent with the original model.

Images