In-hospital consumable multi-code intelligent mapping method

Abstract

The application provides a hospital consumable multi-code intelligent mapping method, comprising the following steps: step 1, generating structured consumable element data based on collected multi-source hospital consumable data; step 2, generating a consumable feature vector matrix based on the structured consumable element data; step 3, performing weighted matching processing on the feature vector matrix and a multi-coding system feature library according to the generated weighted weight coefficient to generate a multi-code mapping decision candidate set; step 4, checking and correcting the multi-code mapping decision candidate set according to a preset checking rule and artificial feedback correction data to generate multi-code mapping correlation data; and step 5, constructing a knowledge graph based on the multi-code mapping correlation data to generate a multi-code mapping correlation graph. The application integrates multi-source information, breaks the limitation of a single data source, facilitates query and traceability, provides deep knowledge support for consumable management, promotes the upgrading of hospital consumable multi-code management to standardization and intelligentization, and meets the management requirements.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence applications, specifically to a method for intelligent mapping of multiple codes for consumables within a hospital. Background Technology

[0002] In the current hospital consumables management scenario, with the advancement of policies such as refined medical insurance settlement and full life-cycle traceability of medical devices, hospital consumables need to correspond to multiple coding systems, including medical insurance codes and UDI (User ID). For example, when consumables are procured and put into storage, UDI needs to be associated to ensure traceability compliance; when medical insurance is reimbursed, medical insurance codes need to be matched to complete cost accounting; and when departmental cost statistics are performed, a unified coding dimension is needed to achieve data aggregation. Therefore, achieving accurate mapping between hospital consumables and multiple coding systems has become a core requirement for ensuring the efficiency of hospital consumables management, reducing medical insurance settlement risks, and meeting regulatory requirements.

[0003] Currently, existing solutions for multi-code mapping of consumables in hospitals mainly rely on manual comparison with coding catalogs to establish mapping relationships. This involves medical staff or medical insurance specialists manually searching for the corresponding codes in the medical insurance coding database and UDI database based on information such as the name and specifications of the consumables and then entering them into the system. Some hospitals introduce simple keyword matching tools that automatically associate highly similar codes through a pre-set "consumable name-code" keyword correspondence table. Other solutions rely on coding association from a single data source, completing the mapping based solely on information from the medical insurance catalog or UDI database, without integrating actual usage data within the hospital for cross-validation.

[0004] The existing solutions mentioned above, which rely on manual comparison and simple keyword matching, are insufficient to address the issues of differences in consumable names and frequent updates to the coding system. This results in low mapping accuracy and is prone to reimbursement disputes caused by mismatched medical insurance codes or traceability gaps caused by UDI association errors. Furthermore, mapping solutions that rely on a single data source lack collaborative verification of multi-source data and cannot cover the impact of multiple dimensions such as consumable specifications, manufacturers, and registration certificates on the coding mapping. When there are conflicts or missing consumable attribute information, it is difficult to generate reliable mapping results, which seriously restricts the standardization and intelligence of multi-code management of consumables within the hospital. Summary of the Invention

[0005] This invention provides a method for intelligent mapping of multiple codes for hospital consumables, comprising:

[0006] Step 1: Generate structured consumable element data based on the collected multi-source in-hospital consumable data;

[0007] Step 2: Generate a consumable feature vector matrix based on structured consumable element data;

[0008] Step 3: Based on the generated weighted coefficients, perform weighted matching processing on the feature vector matrix and the feature library of the multi-code system to generate a multi-code mapping decision candidate set;

[0009] Step 4: Verify and correct the multi-code mapping decision candidate set according to the preset verification rules and manual feedback correction data to generate multi-code mapping association data;

[0010] Step 5: Construct a knowledge graph based on the multi-code mapping association data to generate a multi-code mapping association graph.

[0011] The solution provided by this invention has the following technical advantages:

[0012] This application integrates multi-source data and transforms it into standardized structured elements through step 1, "generating structured consumable element data based on collected multi-source in-hospital consumable data," avoiding matching deviations caused by differences in single-name descriptions. Step 2 generates a consumable feature vector matrix, transforming consumable information into a quantifiable vector form, replacing simple keyword matching, and more accurately capturing the core features of consumables. Step 3 uses weighted weighting coefficients to perform weighted matching between the feature vector matrix and the feature library of multiple coding systems, which can dynamically adapt to rule changes brought about by coding system updates, reduce matching errors caused by coding updates, and ultimately improve the accuracy of multi-code mapping, reducing reimbursement disputes caused by medical insurance code mismatches and the risk of traceability gaps caused by UDI association errors.

[0013] Step 1 of this application clearly defines the generation of structured elements based on "multi-source in-hospital consumable data," directly integrating multi-source information (such as actual in-hospital usage data, medical insurance code data, UDI data, etc.) to break the limitations of a single data source; Step 4 verifies and corrects the mapping decision candidate set through "preset verification rules and manual feedback correction data," realizing collaborative verification of multi-source data to ensure coverage of the impact of multiple dimensions of attributes such as consumable specifications, manufacturers, and registration certificate numbers on the mapping; even if there are conflicts or missing consumable attribute information, adjustments can be made through the manual feedback correction process to avoid unreliable mapping results due to incomplete or conflicting information, providing reliable data support for subsequent management.

[0014] Step 1 of this application generates structured consumable element data, unifies the expression and format of consumable information, and lays the foundation for standardized management; Steps 2-3 realize intelligent calculation of the mapping process through vector matrix and weighted matching, replacing traditional manual comparison and improving management efficiency; Step 5 constructs a multi-code mapping association graph based on the mapping association data, transforming discrete mapping relationships into a structured knowledge network, which not only facilitates subsequent querying and traceability, but also provides deeper knowledge support for consumable management (such as association analysis and trend prediction based on the graph), comprehensively promoting the upgrading of multi-code management of consumables in hospitals from the traditional model to standardization and intelligence, and meeting the management needs under policies such as refined medical insurance settlement and full life cycle traceability. Attached Figure Description

[0015] Figure 1 This is a flowchart illustrating a method for intelligent mapping of multiple codes for consumables within a hospital, as described in an embodiment of this application.

[0016] Figure 2 This is a schematic diagram of the structure of an in-hospital consumables multi-code intelligent mapping system according to an embodiment of this application. Detailed Implementation

[0017] BERT stands for Bidirectional Encoder Representations from Transformers.

[0018] CRF stands for Conditional Random Field.

[0019] GNN stands for Graph Neural Network.

[0020] HIS stands for Hospital Information System.

[0021] HL7 FHIR stands for Health Level Seven Fast Healthcare Interoperability Resources.

[0022] NLP stands for Natural Language Processing.

[0023] TransE stands for Translating Embeddings for Modeling Multi-relational Data.

[0024] UDI stands for Unique Device Identification.

[0025] Figure 1 This is a flowchart illustrating a multi-code intelligent mapping method for hospital consumables according to an embodiment of this application. Figure 1 As shown, it includes:

[0026] Step 1: Generate structured consumable element data based on the collected multi-source in-hospital consumable data;

[0027] Step 2: Generate a consumable feature vector matrix based on structured consumable element data;

[0028] Step 3: Based on the generated weighted coefficients, perform weighted matching processing on the feature vector matrix and the feature library of the multi-code system to generate a multi-code mapping decision candidate set;

[0029] Step 4: Verify and correct the multi-code mapping decision candidate set according to the preset verification rules and manual feedback correction data to generate multi-code mapping association data;

[0030] Step 5: Construct a knowledge graph based on the multi-code mapping association data to generate a multi-code mapping association graph.

[0031] Optionally, step 1, based on the collected multi-source in-hospital consumable data, generates structured consumable element data, specifically as follows:

[0032] Collect multi-source consumable data from within the hospital, and perform hierarchical analysis to extract key consumable elements, thereby generating structured consumable element data.

[0033] Optionally, multi-source in-hospital consumable data is collected and subjected to hierarchical analysis to extract key consumable elements, thereby generating structured consumable element data. This process includes the following steps:

[0034] Step 11: Capture the consumable settlement records implemented by HIS, and synchronously connect to the medical insurance consumable standard catalog library and UDI database to collect multi-source original consumable datasets;

[0035] Step 12: Perform word segmentation and named entity recognition on the multi-source original consumables dataset to obtain word segmentation and entity annotation results, and perform unstructured to structured conversion on the obtained word segmentation and entity annotation results to obtain preliminary structured data;

[0036] Step 13: Perform semantic augmentation on the preliminary structured data based on the medical BERT model to obtain semantically augmented data;

[0037] Step 14: Generate structured consumable element data based on semantically enhanced data.

[0038] Optionally, step 11 involves capturing the consumable settlement records implemented by HIS and simultaneously connecting to the medical insurance consumable standard catalog and UDI database to collect a multi-source original consumable dataset. Specifically, this involves capturing the consumable settlement records during HIS implementation in real time through the HL7FHIR interface to determine the actual usage information of consumables within the hospital, simultaneously connecting to the medical insurance consumable standard catalog and UDI database to determine the consumable standard information under the medical insurance coding system and the unique identification information of consumables under the UDI coding system, and integrating and associating the actual usage information, medical insurance standard information, and UDI identification information of consumables within the hospital to obtain a multi-source original consumable dataset.

[0039] For example, in a specific application scenario, taking the consumable management of the cardiology department of a general hospital as an example, the hospital's HIS records the settlement data of "disposable coronary stents" used in the cardiology department in real time every day. This includes stent model (e.g., "Firebird 23.0×23mm"), quantity used (e.g., 5), department using the stent (cardiology catheterization lab), patient's hospital number (e.g., 20250822001), settlement time (2025-08-22 09:30), and other actual usage information. Technically, the hospital deploys the HL7FHIR interface, which conforms to the medical industry's data exchange standards, to connect the HIS implementation with the consumable data acquisition module. The interface automatically captures settlement records every 15 minutes, ensuring data real-time performance—for example, at 10:00 on August 22, the module accurately obtains the settlement details of the aforementioned 5 coronary stents through this interface, and the data format (e.g., XML / JSON) fully conforms to the output specifications of the HIS implementation, with no compatibility issues. Meanwhile, the data acquisition module synchronously connects to two core external databases via a dedicated API: one is the "Medical Insurance Consumables Standard Catalog" of the provincial medical insurance bureau. After entering the stent model, the corresponding medical insurance code "C0501010010000000000" and standard information such as the medical insurance reimbursement ratio (e.g., 70%) and price limit (e.g., 8,000 yuan / piece) can be retrieved. The other is the UDI database of the National Medical Products Administration. Through the UDI barcode on the stent packaging (e.g., "01000001234567890123"), the unique identification information containing the manufacturer (a medical technology company), registration certificate number (National Medical Device Registration Certificate 20233130001), production batch (20250601), and expiration date (20280531) can be retrieved. Finally, during implementation, the "stent model + patient hospitalization number" is used as the key linking field to integrate the actual usage information of HIS implementation, the standard coding information of the medical insurance catalog, and the unique identification information of the UDI database into a structured data record. The entire process adopts mature interface technology and data association logic, which has been widely used in the field of medical informatization and is technically completely reasonable and feasible.

[0040] Therefore, in the specific implementation of step 11 above, on the one hand, the actual usage information captured by the HIS implementation through the HL7FHIR interface clarifies the "actual application scenario and basic attributes" of consumables within the hospital, avoiding coding mismatches caused by a lack of understanding of the actual usage of consumables during subsequent mapping (such as mismatching a cardiac stent with a neurological stent code); on the other hand, the synchronously acquired medical insurance code and UDI code directly provide the "core coding system data" required for multi-code mapping—the medical insurance code is the key code system for medical insurance settlement and supervision, and the UDI code is the unique identifier for the entire life cycle traceability of consumables. The integration of these two with the actual usage information within the hospital is equivalent to establishing a preliminary association framework of "actual attributes - medical insurance code - UDI code" for each consumable sample. Subsequent steps only need to perform feature extraction and accurate matching based on this framework, without having to repeatedly acquire coding data across implementation times, significantly reducing the data source complexity of multi-code mapping. In addition, the key field association logic formed during the data integration process (such as model number and hospital number) also provides a unified data association standard for subsequent structured element extraction and feature vector construction, ensuring that multi-code mapping has data consistency and accuracy from the source.

[0041] Optionally, step 12 involves performing word segmentation and named entity recognition on the multi-source original consumables dataset to obtain word segmentation and entity annotation results, and then performing unstructured to structured conversion processing on the obtained word segmentation and entity annotation results to obtain preliminary structured data. Specifically, based on the configured medical consumables domain professional vocabulary, the multi-source original consumables dataset is segmented and named entity recognized using the BiLSTM-CRF model to obtain different word segmentation and entity annotation sequences; the different word segmentation and entity annotation sequences are then subjected to optimal label sequence selection processing to implement unstructured to structured data conversion processing to obtain preliminary structured data.

[0042] Optionally, the word segmentation and entity annotation sequences are subjected to optimal label sequence screening to perform unstructured data to structured data conversion processing, resulting in preliminary structured data. Specifically, different word segmentation and entity annotation sequences are encoded to obtain corresponding hidden vectors. The label transition probability of different hidden vectors is calculated through a CRF layer to determine the transition probability score of different word segmentation and entity annotation sequences. Based on the transition probability, the optimal label sequence among the word segmentation and entity annotation sequences is generated. Based on the optimal label sequence, key elements related to medical consumables are identified, and the identified key elements are subjected to boundary definition and category classification processing to generate a set of structured elements as preliminary structured data.

[0043] For example, in a specific application scenario, the multi-source original consumables dataset contains a record: "On August 22, 2025, the Orthopedic Ward 3 used a certain brand of titanium alloy bone plate (model: LP-120), medical insurance code C0201010020000000000, UDI code 01123456789012, manufacturer A Medical Device Company, valid until May 30, 2028." First, technicians pre-configured a professional terminology for medical consumables, including terms and encoding rules such as "titanium alloy bone plate," "LP-120," "C0201010020000000000," and "UDI code." Then, this data was input into the BiLSTM-CRF model. The model performed word segmentation based on the terminology, resulting in preliminary word segmentation and entity annotation sequences such as "2025-08-22 / time," "Orthopedic Ward 3 / Department," "Titanium alloy bone plate / Consumable name," "LP-120 / Model," "C0201010020000000000 / Medical Insurance Code," "01123456789012 / UDI code," "A Medical Device Company / Manufacturer," and "2028-05-30 / Expiration Date." Next, the model encodes different label sequences (such as the possible sequences of "a certain brand of titanium alloy bone plate" as a whole or "titanium alloy bone plate" as a separate label) to generate corresponding hidden vectors. The CRF layer calculates the transfer probability score of "Orthopedics Ward 3 → Titanium Alloy Bone Plate" as 0.92 and "a certain brand → Titanium Alloy Bone Plate" as 0.68 based on the preset label transfer rules (such as "department" being more likely to be followed by "consumable name" rather than "expiration date"). Based on this, the sequence with the highest transfer probability, "2025-08-22 / time, Orthopedics Ward 3 / department, Titanium Alloy Bone Plate / consumable name, LP-120 / model, C0201010020000000000 / medical insurance code, 01123456789012 / UDI code, A medical device company / manufacturer, 2028-05-30 / expiration date", is selected as the optimal label sequence. Finally, based on this sequence, the boundaries of each key element are clarified (e.g., "LP-120" only refers to the model number and does not include redundant characters before and after it), and a structured set of elements is generated by classifying them according to the categories of "time, department, consumable name, model, medical insurance code, UDI code, manufacturer, and expiration date". The whole process conforms to the mature application logic of natural language processing in the medical field and is technically completely reasonable and correct.

[0044] Therefore, the technical benefits of step 12 above include, but are not limited to, providing a "standardized and accurate basic data format" for multi-code intelligent mapping: On the one hand, the BiLSTM-CRF model based on a professional vocabulary in the medical consumables field solves the problem of "unstructured text ambiguity" in multi-source raw data—for example, avoiding the erroneous splitting of "titanium alloy bone plate (model: LP-120)" into "titanium alloy / material, bone plate / type, model: LP-120 / other," ensuring the accuracy of word segmentation and entity recognition of core consumable attributes (such as name and model), and providing accurate "data granules" for subsequent feature vector extraction; on the other hand, through optimal label sequence screening and... Key element classification transforms the originally messy text data (such as settlement records containing descriptive language) into a structured set of key-value correspondences. This allows the subsequent step of "consumable feature vector matrix construction" to be directly vectorized based on explicit fields such as "consumable name, model, medical insurance code, and UDI code," eliminating the need to process redundant information from unstructured text and significantly reducing the complexity of feature extraction. At the same time, unified element classification (such as classifying all data according to "consumable name, model, and code") provides a consistent data dimension for subsequent weighted matching of multiple coding systems, avoiding mapping deviations caused by chaotic element classifications and ensuring the accuracy of multi-code mapping from the data preprocessing stage.

[0045] Optionally, step 13, performing semantic augmentation on the preliminary structured data based on the medical BERT model to obtain semantically augmented data, specifically includes the following steps:

[0046] The medical BERT model, jointly trained on PubMed literature and medical device instructions, calculates semantic similarity at the word, sentence, and document levels on the preliminary structured data using a domain masking strategy of "random mask + term mask + context mask". Based on the semantic similarity at the word, sentence, and document levels, it performs deep semantic association mining and feature enhancement processing on the key elements of consumables in the preliminary structured data to obtain semantically enhanced data.

[0047] For example, in a specific application scenario, the initial structured data contains two core records: Record 1's consumable name is "disposable sterile syringe (10ml)," with the associated attribute "used for intravenous drug injection, material is polypropylene"; Record 2's consumable name is "10ml sterile syringe," with the associated attribute "for intravenous drug administration, component includes polypropylene syringe." First, the medical BERT model used was pre-trained using relevant literature on "classification and application of medical syringes" in PubMed and a "disposable syringe instruction manual" published by a medical device manufacturer. This allowed the model to grasp the synonymous association between "syringe" and "syringe" in medical scenarios, as well as the unit correspondence between "10ml" and "10 milliliters." Subsequently, the model used a strategy of "random mask + term mask + context mask" to process the two data sets: the random mask masked "polypropylene" in record 1 as "[MASK]", and the model predicted the probability of "polypropylene" to be 0.95 based on the training data; the term mask masked "sterile syringe" in record 2 as "[MASK]", and the model accurately associated it with the synonym "disposable sterile syringe"; the context mask masked the part of the description "used for intravenous drug injection" in record 1, and the model could still infer its core use through keywords such as "10ml" and "sterile". Based on this, the model calculates word-level similarity ("syringe" and "injection syringe" have a similarity of 0.92), sentence-level similarity (the attribute descriptions of the two records have a similarity of 0.88), and discourse-level similarity (overall semantic relevance of 0.90). Based on these similarities, it mines the deep relationship between the two data points—they are essentially different descriptions of the same type of consumables. Then, it enhances the key elements: it adds "alias: 10 ml sterile injection syringe" to record 1 and "standard name: disposable sterile syringe (10 ml)" to record 2. At the same time, it unifies the description format of the "material" and "purpose" attributes of the two records, and finally generates semantically enhanced data.

[0048] Therefore, the technical benefits of step 13 above include, but are not limited to, solving the core pain point of "semantic inconsistency of key elements of consumables" in multi-code mapping, laying a semantic foundation for subsequent accurate mapping: On the one hand, the medical BERT model trained jointly by PubMed literature and medical device instructions can accurately identify synonyms and variant expressions in the field of medical consumables (such as "10ml" and "10 milliliters", "syringe" and "injection syringe"), avoiding misclassification of the same consumable as different categories when matching multiple coding systems due to expression differences, effectively reducing the mapping error rate; on the other hand, the three-level semantic similarity calculation of "word-sentence-article" and deep association mining can supplement the preliminary structured data. By adding hidden semantic information (such as synonyms and standard terms), the features of key consumable elements are made more complete. For example, a record that originally only labeled "10 ml sterile injection syringe" can be enhanced with a standard name and associated attributes. When matching with the medical insurance coding system feature library, the corresponding medical insurance code can be matched simultaneously through the dual dimensions of "synonym" and "standard name", improving the matching coverage and accuracy. In addition, a unified attribute description format (such as unifying "containing liquid medicine" and "used for liquid medicine injection" into "liquid medicine containing and injecting") can reduce the deviation in feature vector calculation caused by chaotic attribute description formats, ensuring semantic alignment between consumable features and coding system features in the subsequent weighted matching process, and further guaranteeing the reliability of multi-code mapping.

[0049] Optionally, step 14 involves generating structured consumable element data based on the semantically enhanced data, specifically as follows:

[0050] Based on a dynamic knowledge base containing synonym rules for multiple types of consumables, the semantically enhanced data is mapped to obtain standardized terms and attribute mapping results; the obtained standardized terms and attribute mapping results are parsed using multimodal fields to obtain a multidimensional parsing element set; the multidimensional parsing element set is then subjected to cross-field semantic element fusion processing to generate structured consumable element data.

[0051] For example, in a specific application scenario, the semantically enhanced data contains a record: "Disposable sterile syringe (10mL), also known as a syringe, for intravenous administration, material is PP, medical insurance code C050101001". First, during implementation, a dynamic knowledge base containing synonym rules for various consumables is invoked. This knowledge base stores standardized mapping rules such as "syringe = syringe", "PP = polypropylene", and "10mL = 10 milliliters". Through rule matching, "syringe" in the record is mapped to the standard term "disposable sterile syringe", "PP" is converted to "polypropylene", and "10mL" is standardized to "10 milliliters", resulting in standardized terminology and attribute mapping results. Next, multimodal field parsing is performed on this result: "disposable sterile syringe" (name), "10 milliliters" (specification), and "polypropylene" (material) are extracted from the text fields; the classification information "injection and puncture instruments - syringe" corresponding to the medical insurance code "C050101001" is parsed from the coding field; and "intravenous administration" (applicable scenario) is identified from the usage description, forming a multi-dimensional parsing element set. Finally, cross-field semantic fusion was performed on these elements: "Specification 10 ml" and "Material Polypropylene" were associated as the "Physical Attribute" dimension, "Applicable Scenario Intravenous Administration" and "Classification Injection and Puncture Instruments" were integrated as the "Clinical Attribute" dimension, while "Medical Insurance Code C050101001" was retained as the "Code Attribute" dimension, and finally, structured consumable element data containing "Name, Physical Attribute, Clinical Attribute, and Code Attribute" was generated.

[0052] Therefore, the technical benefits of the specific implementation of step 14 above include, but are not limited to, providing a "highly standardized and dimensionally unified feature foundation" for multi-code intelligent mapping: On the one hand, the synonym rule mapping of the dynamic knowledge base solves the problem of chaotic terminology, ensuring that consumable data from different sources reach consistency in key terms (such as "syringe" and "syringe" being unified as standard names), avoiding mapping deviations caused by terminology differences during subsequent feature vector construction; on the other hand, multimodal field parsing and cross-field fusion integrate scattered information into structured dimensions (such as physical attributes and clinical attributes), making the originally fragmented consumable features organically related. For example, the fusion of "10 ml specification" and "intravenous administration scenario" can provide more complete feature support for subsequent matching of medical insurance codes or UDI codes; in addition, the unified structured element format (such as fixed attribute dimensions) ensures that all consumable samples are processed under the same feature framework, enabling feature vector calculation, weighted matching, and other links in the multi-code mapping process to have a consistent data benchmark, greatly improving the stability and accuracy of the mapping model, and laying a key foundation for achieving efficient multi-code intelligent association.

[0053] Optionally, step 2, based on the structured consumable element data, generates a consumable feature vector matrix, specifically as follows:

[0054] The structured consumable element data is vectorized to generate a consumable feature vector matrix. The rows of this consumable feature vector matrix represent individual consumable samples, the columns represent the quantification dimensions of key consumable elements, and the elements at the intersection of rows and columns represent the feature values ​​of the corresponding consumable samples in the quantification dimensions.

[0055] Optionally, the structured consumable element data is vectorized to generate a consumable feature vector matrix. The rows of this matrix represent individual consumable samples, the columns represent the quantification dimensions of key consumable elements, and the elements at the intersection of rows and columns represent the feature values ​​of the corresponding consumable sample in the quantification dimension. Specifically, this includes the following steps:

[0056] Step 21: Based on the constructed medical consumables-specific word segmentation word list, synonym word list and stop word list, perform word segmentation enhancement processing on the consumables field in the structured consumables element data to obtain the word segmentation enhancement consumables element text;

[0057] Step 22: Using the Skip-gram model of Word2Vec, map each word in the segmented and enhanced consumables element text to a dense vector. Based on the dense vector, generate core feature word vectors that can represent the core features of consumables.

[0058] Step 23: Associate the core feature word vectors with the core attribute information of consumables to supplement the feature expression of the attribute dimension in the core feature word vectors and generate a multi-dimensional feature vector containing consumable terminology semantics, consumable attribute features, and consumable domain knowledge; arrange the multi-dimensional feature vectors into a matrix by using individual consumable samples as rows and the quantification dimension of key consumable elements as columns to generate a consumable feature vector matrix.

[0059] Optionally, in step 21, based on the constructed medical consumables-specific word segmentation lexicon, synonym lexicon, and stop word lexicon, the consumables field in the structured consumables element data is subjected to word segmentation enhancement processing to obtain the word-segmented enhanced consumables element text, specifically as follows:

[0060] Based on a word segmentation lexicon specific to the medical consumables field, the structured consumables element data is segmented into consumables terms to obtain the first preliminary word segmentation result. Based on a synonym lexicon for the medical consumables field, the first preliminary word segmentation result is processed with synonym completion and standardized mapping to obtain the second synonym-enhanced word segmentation result. Based on a stop word list for the medical consumables field, the second synonym-enhanced word segmentation result is processed with domain-adaptive denoising to obtain the third denoising-enhanced word segmentation result. This third denoising-enhanced word segmentation result is the word segmented and enhanced text of the consumables elements.

[0061] For example, in a specific application scenario, such as a consumables field in structured consumables data that reads "Disposable sterile syringe (20mL), made of polypropylene, suitable for intravenous injection," the following steps are taken: First, a specialized word segmentation lexicon for the medical consumables field is invoked. This lexicon contains professional terms such as "disposable," "sterile syringe," "20mL," "polypropylene," and "intravenous injection." After segmenting the field, the first preliminary word segmentation result is obtained: ["disposable", "sterile syringe", "20mL", "polypropylene", "intravenous injection"]. Next, based on a synonym lexicon (which records "sterile syringe = sterile syringe" and "intravenous injection = intravenous administration"), the preliminary result is processed for synonym completion and standardization. "sterile syringe" is added as a synonym, and "intravenous injection" is standardized to "intravenous administration," resulting in the second synonym-enhanced word segmentation result: ["disposable", "sterile syringe", "sterile syringe", "20mL", "polypropylene", "intravenous administration"]. Finally, the domain stop word list (containing non-core descriptive words such as "made of" and "applicable to") is called to denoise the second result, removing auxiliary words with no substantial meaning, and obtaining the third denoised and enhanced word segmentation result: ["single use", "sterile syringe", "sterile syringe", "20mL", "polypropylene", "intravenous administration"], which is the consumable element text after word segmentation enhancement.

[0062] Therefore, the technical benefits of the specific implementation of step 21 above include, but are not limited to, providing a "precise, standardized, and core-feature-focused text foundation" for multi-code intelligent mapping: On the one hand, the domain-specific word segmentation lexicon ensures the accuracy of consumable term segmentation, avoiding the misinterpretation of "sterile syringe" into unreasonable fragments such as "sterile / syringe," thus preserving the integrity of the terminology; on the other hand, synonym completion and standardization processes resolve the issue of expression differences, making synonyms such as "sterile syringe" and "sterile syringe" considered as related features in subsequent feature extraction, thereby improving the semantic coverage when matching multiple coding systems; in addition, domain-adaptive denoising processes eliminate redundant information, focusing the word segmentation results on core features such as "specification 20mL" and "material polypropylene," reducing the interference of irrelevant information on feature vector calculation, and ensuring that the subsequently generated consumable feature vectors accurately reflect the essential attributes of the consumables, providing key support for the accuracy of multi-code mapping.

[0063] Optionally, in step 22, each word in the segmented and enhanced consumables element text is mapped to a dense vector using the Skip-gram model of Word2Vec. Based on the dense vectors, core feature word vectors representing the core features of the consumables are generated, specifically as follows:

[0064] The consumable element text after word segmentation enhancement is preprocessed for domain adaptation to obtain a domain-enhanced word segmentation sequence. The domain-enhanced word segmentation sequence is then mapped using the Word2VecSkip-gram model to obtain a domain-aware dense vector. The feature importance of the domain-aware dense vector is evaluated based on maximum mutual information to obtain a vector importance score. Based on the vector importance score, the domain-aware dense vector is filtered to determine the core feature word vectors that can represent the core features of consumables.

[0065] For example, in a specific application scenario, the text of consumable elements after word segmentation enhancement might be ["disposable", "sterile syringe", "sterile syringe", "20mL", "polypropylene", "intravenous administration"]. First, domain-adaptive preprocessing is performed. Domain weights are assigned to the word segmentation sequence using a medical consumables classification dictionary (e.g., "sterile syringe" and "20mL" are weighted at 1.2, and "polypropylene" at 1.0), resulting in a domain-enhanced word segmentation sequence. Next, this sequence is input into a Word2VecSkip-gram model trained on a medical consumables corpus. The model predicts contextual word relationships and maps each word to a 100-dimensional domain-aware dense vector—for example, the vector for "sterile syringe" has a cosine similarity of 0.93 with the vector for "sterile syringe" and 0.75 with the vector for "intravenous administration," reflecting semantic associations within the domain. Subsequently, the maximum mutual information between each vector and the core attributes of the consumable (such as "specification", "material", and "purpose") was calculated to obtain vector importance scores: "20mL" (specification) score 0.92, "polypropylene" (material) score 0.88, "sterile syringe" (name) score 0.85, and "intravenous administration" (purpose) score 0.80. Finally, vectors with scores ≥0.8 were selected as core feature word vectors, and the vectors corresponding to ["sterile syringe", "20mL", "polypropylene", "intravenous administration"] were ultimately determined as core feature word vectors.

[0066] Therefore, the technical benefits of step 22 above include, but are not limited to, providing a "quantified and core-feature-focused semantic carrier" for multi-code intelligent mapping: On the one hand, domain-aware dense vectors transform discrete words into continuous numerical vectors, enabling the semantic association between synonyms such as "sterile syringe" and "sterile syringe" to be quantified through vector similarity, thus solving the matching barrier caused by differences in textual expression; on the other hand, feature filtering based on maximum mutual information ensures the retention of the most critical feature vectors for encoding mapping, such as "specification 20mL" and "material polypropylene," while eliminating secondary information and reducing dimensional redundancy in subsequent matrix calculations; in addition, the contextual associations captured by the Skip-gram model (such as the vector association between "intravenous administration" and "sterile syringe") can integrate the functional attributes of consumables into vector expressions, so that the generated core feature word vectors not only contain the information of the terms themselves, but also implicitly contain the association logic within the domain, providing richer semantic basis for the weighted matching of subsequent multi-code systems and significantly improving the accuracy of mapping.

[0067] Optionally, step 23 involves associating the core feature word vectors with the core attribute information of the consumables to supplement the feature expression of the attribute dimension in the core feature word vectors and generating a multi-dimensional feature vector containing consumable terminology semantics, consumable attribute features, and consumable domain knowledge; the multi-dimensional feature vector is then matrix-arranged with individual consumable samples as rows and the quantification dimension of key consumable elements as columns to generate a consumable feature vector matrix, specifically as follows:

[0068] For the core feature word vectors and the core attribute information of consumables, rule matching is performed by constructing a term-attribute association rule base to obtain term-attribute association pairs; the core attribute information of consumables in the term-attribute association pairs is encoded into attribute vectors to obtain attribute vectors; the attribute vectors are concatenated with the corresponding core feature word vectors to obtain attribute-enhanced vectors; medical consumable classification standards and compliance requirements are introduced into the attribute-enhanced vectors as domain knowledge labels, and label mapping is performed to generate multi-dimensional feature vectors containing consumable term semantics, consumable attribute features, and consumable domain knowledge; the multi-dimensional feature vectors are indexed and labeled according to the quantification dimension of the key elements of consumables to obtain indexed feature vectors; the indexed feature vectors are filled in an ordered manner with individual consumable samples as rows and the quantification dimension of the key elements of consumables as columns to generate a consumable feature vector matrix.

[0069] For example, in a specific application scenario, the core feature word vectors are the vectors corresponding to "sterile syringe", "20mL", "polypropylene", and "intravenous administration". The core attribute information of the consumables includes the product type "injection and puncture device", the registration certificate number "National Medical Device Registration Certificate 20233150001", and the sterilization method "ethylene oxide sterilization". First, the term-attribute association rule base is preset with rules such as "sterile syringe → product type", "20mL → specifications", "polypropylene → material composition", and "intravenous administration → clinical use". The term-attribute association pairs are obtained by rule matching. Next, the attribute information such as "injection and puncture device" and "20mL" is encoded into attribute vectors. For example, "injection and puncture device" is generated by one-hot encoding [1,0,0] (corresponding to the three types of device classification), and "20mL" is generated by numerical mapping

[20] to obtain the attribute vector. Then, the attribute vector is concatenated with the corresponding core feature word vector (assuming it is 100-dimensional). The 100-dimensional vector of "sterile syringe" is concatenated with [1,0,0] to form a 103-dimensional attribute enhancement vector. Next, the domain knowledge tags "Class III Medical Device" and "Compliant with GB15810-2019" are introduced and encoded as [1,1], then concatenated with the attribute enhancement vector to generate a 105-dimensional multi-dimensional feature vector, containing terminology semantics, attribute features, and domain knowledge. Then, the vector is indexed according to six quantification dimensions: "terminology semantics, specifications, material, purpose, product type, and compliance," resulting in an indexed feature vector. Finally, if there is another sample, "10mL sterile syringe," its processed indexed feature vector is arranged in rows with the current sample, each row corresponding to one sample, and the six columns corresponding to the quantification dimensions. After orderly filling the vector values, a 2×10⁵ consumable feature vector matrix is ​​formed.

[0070] Therefore, the technical benefits of step 23 above include, but are not limited to, constructing a "structured, multi-dimensional, and domain-adaptive feature benchmark" for multi-code intelligent mapping: On the one hand, by using term-attribute association and vector concatenation, discrete attribute information is integrated into semantic vectors, solving the limitation that core feature word vectors only contain textual semantics and lack attribute dimensions, so that the specification information of "20mL" and the term semantics of "sterile syringe" form associated features, providing more complete feature support for coding matching; On the other hand, the introduction of domain knowledge tags allows feature vectors to carry key domain information such as compliance and classification, ensuring that the mapping process conforms to the professional standards of medical consumables management and avoiding erroneous associations due to the lack of domain constraints; In addition, the matrix arrangement enables multi-sample features to form a unified structure, which not only facilitates efficient vector comparison calculations with the feature library of the multi-coding system, but also ensures the dimensional consistency of different sample features through quantitative dimension indexing, providing standardized input for the weighted matching stage, and improving the accuracy and efficiency of multi-code mapping from the feature structure level.

[0071] Optionally, step 3 involves performing a weighted matching process on the feature vector matrix and the multi-coding system feature library based on the generated weighted coefficients to generate a multi-code mapping decision candidate set, specifically:

[0072] Dynamic weight matching is performed on the feature vector matrix of consumables to obtain the weight coefficients of each dimension of features of different consumable samples. Based on the NLP semantic similarity calculation results, the feature vector matrix and the feature library of the multi-code system are weighted and matched to generate a multi-code mapping decision candidate set.

[0073] Optionally, dynamic weight matching is performed on the consumable feature vector matrix to obtain the weight coefficients of each dimension of features for different consumable samples. Based on the NLP semantic similarity calculation results, the feature vector matrix and the multi-code system feature library are weighted and matched to generate a multi-code mapping decision candidate set. The specific steps include the following:

[0074] Step 31: Extract structured knowledge from the medical consumables coding domain to generate an adaptive rule weight library;

[0075] Step 32: The consumable feature vector matrix and the vector data in the multi-coding system feature library are weighted and calculated based on the matching weights output by the adaptive rule weight library to obtain the comprehensive matching score between each consumable sample and the multi-coding system feature library.

[0076] Step 33: Generate a dynamically optimized mapping candidate set based on the comprehensive matching score;

[0077] Step 34: Generate a multi-code mapping decision candidate set based on the dynamically optimized mapping candidate set.

[0078] Optionally, step 31 involves extracting structured knowledge from the medical consumables coding domain to generate an adaptive rule weight library, specifically:

[0079] Knowledge in the field of medical consumable coding is extracted in a structured manner, and the correspondence between product types and codes, and the correlation logic between specifications and coding dimensions related to coding matching are screened to construct a basic mapping rule base. Real-time feature analysis is performed on historical mapping data containing past coding matching records and correction records to update the initial weights and trigger thresholds of each rule in the basic mapping rule base, resulting in an adaptive rule weight base.

[0080] For example, in a specific application scenario, the knowledge in the field of medical consumable coding includes medical insurance coding rules, UDI coding specifications, and industry classification standards. During implementation, this knowledge is first extracted in a structured manner, filtering out the correspondence between product types and codes, such as "product type = artificial hip joint prosthesis → medical insurance code prefix = C0201" and "specification parameter = diameter 12mm → coding digits 5-7 = 120," as well as the association logic between specification parameters and coding dimensions, such as "material parameters occupy 3 dimensions in UDI coding" and "sterilization method affects the last 2 digits of the code." Based on this, a basic mapping rule base is constructed. Each rule includes rule content, initial weight (e.g., initial weight of 0.6 for "product type matching"), and trigger threshold (e.g., triggering when similarity ≥ 0.7). Next, during implementation, real-time feature analysis was performed on the coding matching records of the past 3 years (such as "ceramic hip joint prosthesis matching C020101003") and correction records (such as "a 10mm diameter prosthesis was mistakenly matched as a 12mm code, and after correction, it was associated with 100 dimensions"). It was found that the actual accuracy of "specification parameter matching" was higher than the initial setting, so its weight was increased from 0.4 to 0.5. At the same time, based on the correction records, the trigger threshold of "diameter parameter" was decreased from 0.7 to 0.65 to make the rules more adaptable to the actual scenario. Finally, an adaptive rule weight library containing dynamically adjusted rules, weights and thresholds was formed.

[0081] Therefore, the technical benefits of the specific implementation of step 31 above include, but are not limited to, providing a "domain-adaptive and continuously optimized rule benchmark" for multi-code intelligent mapping: On the one hand, the basic mapping rule library extracts domain knowledge in a structured manner, transforming fuzzy coding rules into computable association logic (such as the correspondence between product type and coding prefix), providing clear domain constraints for subsequent weighted matching and preventing the matching process from deviating from professional coding specifications; on the other hand, real-time feature analysis based on historical data enables rule weights and thresholds to be adaptive, dynamically correcting deviations in initial rules (such as adjusting the weight of specification parameters based on correction records), solving the problem that fixed rules are difficult to cope with complex coding scenarios; in addition, the adaptive rule weight library retains the professional knowledge accumulated in the coding domain and can be continuously optimized through actual data feedback, so that the subsequent weighted matching calculation not only conforms to the coding standard but also fits the actual mapping needs within the institute, improving the accuracy and robustness of multi-code mapping from the rule level.

[0082] Optionally, in step 32, the feature vector matrix of consumables and the vector data in the multi-coding system feature library are weighted based on the matching weights output by the adaptive rule weight library to obtain the comprehensive matching score between each consumable sample and the multi-coding system feature library, specifically:

[0083] For the consumable feature vector matrix and the vector data in the multi-coding system feature library, a semantic similarity calculation model based on BERT is used to perform vector comparison to obtain a semantic similarity score. Based on the matching weight output by the adaptive rule weight library, the semantic similarity score, the historical mapping confidence extracted from the historical mapping data, and the manual feedback correction value sorted from the manual feedback record are weighted and calculated to obtain the comprehensive matching score between each consumable sample and the multi-coding system feature library.

[0084] For example, in a specific application scenario, a sample in the consumables feature vector matrix is ​​a 105-dimensional feature vector for "20mL sterile syringe (polypropylene material)". The multi-coding system feature library contains vector data corresponding to the medical insurance code "C050101001" and the UDI code "01234567890123". First, the system calls the BERT-based semantic similarity calculation model to compare the sample vector with the medical insurance code vector, obtaining a semantic similarity score of 0.85, and with the UDI code vector, obtaining a score of 0.82. The adaptive rule weight library outputs the current matching weight: semantic similarity weight 0.5, historical mapping confidence weight 0.3, and manual feedback correction value weight 0.2. From the historical mapping data, the past matching records of this type of syringe with "C050101001" are extracted, and the historical mapping confidence is calculated to be 0.9; from the manual feedback records, the manual correction value of this code is extracted to be +0.02 (indicating that the matching degree is improved after manual verification). Calculated using the weighted formula: Medical Insurance Code Comprehensive Score = 0.85 × 0.5 + 0.9 × 0.3 + 0.02 × 0.2 = 0.425 + 0.27 + 0.004 = 0.699; Similarly, the UDI Code Comprehensive Score is calculated to be 0.671.

[0085] Therefore, the technical benefits of step 32 above include, but are not limited to, providing a "comprehensive and dynamically adaptable quantitative basis for multi-code intelligent mapping": On the one hand, the weighted calculation of semantic similarity, historical confidence, and human feedback breaks through the limitations of single-dimensional scoring. It captures deep semantic associations through the BERT model and corrects mechanical matching deviations with historical data and human experience, making the comprehensive score closer to the actual mapping needs. On the other hand, the introduction of an adaptive rule weight library allows the weights of each dimension to be dynamically adjusted according to the scenario (such as increasing the weight of historical confidence when matching medical insurance codes), solving the problem that fixed weights are difficult to adapt to the differences in multiple coding systems. In addition, the quantified comprehensive matching score provides a clear threshold standard for the subsequent generation of candidate sets, making classification decisions such as "automatic confirmation" and "human verification" operable, and improving the accuracy and efficiency of multi-code mapping from the perspective of matching evaluation.

[0086] Optionally, step 33, based on the comprehensive matching score, generates a dynamically optimized mapping candidate set, specifically as follows:

[0087] The comprehensive matching score is used to determine the range. Mapping combinations whose comprehensive matching scores reach the automatic confirmation threshold are marked as "automatically confirmed mapping relationship". Mapping combinations whose comprehensive matching scores are in the middle range are triggered to trigger the human-machine collaborative verification process. A verification priority identifier is added to each mapping combination to generate a mapping candidate set with verification tags. Mapping combinations whose comprehensive matching scores are below the lower limit threshold are initiated to start a secondary enhancement process to generate a secondary optimized mapping candidate set. The mapping combinations marked as "automatically confirmed mapping relationship", the mapping candidate set with verification tags, and the secondary optimized mapping candidate set constitute the dynamically optimized mapping candidate set.

[0088] For example, in a specific application scenario, the hospital sets the automatic confirmation threshold for the comprehensive matching score to 0.8, the middle range to 0.6-0.8, and the lower limit threshold to 0.6. The secondary enhancement processing rule is "recalculate the similarity after supplementing attributes such as the consumable registration certificate number and manufacturer." After mapping a batch of consumables, three sets of comprehensive matching scores are obtained: The first set, "20mL sterile syringe - medical insurance code C050101001," has a comprehensive score of 0.86, and "20mL sterile syringe - UDI code 01234567890123" has a comprehensive score of 0.84, both reaching the automatic confirmation threshold. The system directly marks these two mapping combinations as "automatically confirmed mapping relationship." The second set, "10mL intravenous infusion needle - medical insurance code C050102002," has a comprehensive score of 0.72, and "10mL intravenous infusion needle - UDI code 01987654321098" has a comprehensive score of 0.68, which is in the middle range. The system triggers the human-machine collaborative verification process, based on... Based on the attribute that "intravenous infusion needles are high-frequency consumables," a "high-priority" verification mark was added to these two sets of mappings, generating a candidate set of mappings with verification marks. The third set, "No. 3 surgical suture - Medical Insurance Code C040101003," had a comprehensive score of 0.55, and "No. 3 surgical suture - UDI Code 01135792468012" had a comprehensive score of 0.58, both below the lower threshold. Therefore, the system initiated a secondary enhancement process—automatically adding attributes such as "Registration Certificate No. 20243020005" and "Manufacturer B Medical Technology Company" to the feature vector, recalculating the comprehensive matching scores, and increasing them to 0.81 and 0.79 respectively, forming a secondary optimized candidate set of mappings. Finally, the "automatically confirmed mapping relationship," the "candidate set of mappings with verification marks," and the "secondary optimized candidate set of mappings" were integrated to obtain a dynamically optimized candidate set of mappings.

[0089] Therefore, the technical advantages of step 33 above include, but are not limited to, constructing a "layered processing and dynamically optimized mapping candidate generation mechanism" that balances the efficiency and accuracy of multi-code mapping: On the one hand, by automatically confirming thresholds to filter high-confidence mapping combinations (such as combinations with a comprehensive score ≥ 0.8), the mapping relationship can be directly confirmed without manual intervention, significantly reducing the workload of manual verification and improving the overall efficiency of multi-code mapping for hospital consumables, especially suitable for batch mapping scenarios of high-frequency and standard consumables in hospitals; on the other hand, by adding verification priority identifiers to mapping combinations in the middle interval, manual resources can be guided to prioritize the processing of high-value data. For high-risk mapping items (such as frequently used consumables), the system avoids the inefficiency caused by disordered manual verification. At the same time, it rescues low-scoring mapping combinations through secondary enhancement processing, recalculates scores after supplementing key attributes, avoids misjudgments caused by missing initial features, and improves the coverage and accuracy of the mapping candidate set. In addition, the hierarchical processing mechanism ensures that the mapping candidate set includes not only directly usable confirmation relationships, but also items to be verified that require manual intervention, as well as corrected optimization items, forming a complete dynamic optimization closed loop. This provides a comprehensive and reliable candidate data foundation for subsequent multi-code mapping decisions, effectively balancing the automation and accuracy of multi-code mapping.

[0090] Optionally, step 34, generating a multi-code mapping decision candidate set based on the dynamically optimized mapping candidate set, specifically involves:

[0091] For the dynamically optimized mapping candidate set, the built-in visualization and traceability module is invoked to record the feature dimension contribution-rule triggering link dual-dimensional decision path of each mapping candidate result, so as to clearly mark the influence ratio of each feature dimension on the matching result, the specific coding rule triggered, and the rule priority ranking; the feature dimension contribution-rule triggering link dual-dimensional decision path and the corresponding mapping candidate result are associated, integrated, and anchored with confidence, so as to bind key parameters such as rule matching score and semantic similarity score in the decision path with the candidate result, generating a traceable confidence calculation link; the traceable confidence calculation link is hierarchically sorted according to the category of multi-coding system to generate a multi-code mapping decision candidate set.

[0092] For example, in a specific application scenario, the dynamically optimized mapping candidate set includes two core candidate results: “20mL sterile syringe - medical insurance code C050101001” (automatic confirmation) and “10mL intravenous infusion needle - UDI code 01987654321098” (high priority verification). The system first calls the visual traceability module to record the contribution of feature dimensions for “20mL sterile syringe - medical insurance code C050101001”: “specification 20mL” contributes 35%, “material polypropylene” contributes 25%, and “clinical use intravenous administration” contributes 40%. At the same time, it records the rule triggering link - the rule of “product type = syringe → medical insurance code prefix C0501” (priority 1) is triggered first, and the rule of “specification parameter = 20mL → code 4th-5th digit 01” (priority 2) is triggered secondarily, forming a two-dimensional decision path. For “10mL intravenous infusion needle - UDI code 01987654321098”, the system also records the rule link of “specification 10mL” contributing 30%, “use intravenous infusion” contributing 50%, “sterilization method ethylene oxide” contributing 20%, and “UDI code 3rd-6th digits corresponding to product type” (priority 1) and “sterilization method corresponding to code suffix” (priority 2). Next, the decision path is associated and integrated with the mapping candidate results. Parameters such as the rule matching score of 0.82 and semantic similarity score of 0.85 for "20mL sterile syringe" are anchored to the corresponding candidate results, generating a traceable confidence calculation link (e.g., "Comprehensive score 0.86 = Rule matching score 0.82 × 0.5 + Semantic similarity 0.85 × 0.3 + Historical confidence 0.9 × 0.2"). Finally, the results are hierarchically sorted according to the coding system category: under the medical insurance coding system, "20mL sterile syringe-C050101001" (comprehensive score 0.86) is sorted, and under the UDI coding system, "10mL intravenous infusion needle-01987654321098" (comprehensive score 0.68) is sorted, ultimately generating a multi-code mapping decision candidate set.

[0093] Therefore, the technical benefits of the specific implementation of step 34 above include, but are not limited to, building a "traceable and structured decision support system" for multi-code intelligent mapping, enhancing the credibility and decision-making efficiency of the mapping results: On the one hand, the dual-dimensional recording of feature dimension contribution and rule triggering chain makes the "decision basis" of the mapping results "visualized"—for example, when manually verifying the mapping of "10mL intravenous infusion needle", one can directly view the core coding rule that has the highest contribution of "purpose: intravenous infusion" and is triggered, without having to re-derive the matching logic, greatly reducing the cost of manual review; on the other hand, the confidence calculation chain The anchoring and hierarchical sorting of the path not only achieves full-link traceability of "result-parameter-rule" (such as tracing the specific calculation source of the comprehensive score of 0.86), but also classifies and organizes the candidate set according to the coding system, making it convenient for medical staff or managers to quickly locate the mapping results of different systems such as medical insurance and UDI, avoiding the waste of efficiency in cross-system searches; in addition, the traceable decision path also provides data support for subsequent mapping rule optimization (such as adjusting the corresponding rule weight when the contribution of "sterilization method" is found to be low), forming a closed loop of "mapping-traceability-optimization", which improves the reliability and iterability of multi-code mapping from the decision-making level.

[0094] Optionally, step 4, based on preset verification rules and manual feedback correction data, verifies and corrects the multi-code mapping decision candidate set to generate multi-code mapping association data, specifically including the following steps:

[0095] Step 41: Call the preset multi-dimensional verification rules to perform intelligent logic verification on the multi-code mapping decision candidate set, and temporarily store the compliant mapping results and the marked suspicious mappings according to the verification results to obtain the initial screening classification mapping data;

[0096] Step 42: For the suspicious mappings pending review in the initial screening classification mapping data, retrieve historical cases and matching policy basis and integrate them into auxiliary decision-making information packages, and associate and bind them to the corresponding suspicious mappings to obtain suspicious mapping data with auxiliary decision-making information;

[0097] Step 43: Activate the cross-departmental expert consultation interaction module, receive expert correction instructions to correct problematic items in suspicious mapping data with auxiliary decision-making information, and generate an expert-verified mapping set;

[0098] Step 44: Merge compliance mapping results and expert verification. A preliminary integrated mapping set is formed by combining the mapping sets. The audit process records are associated to generate an integrated mapping set with traceability tags. After converting the format according to the preset specifications, multi-code mapping association data is generated.

[0099] Optionally, step 41 involves calling preset multi-dimensional verification rules to perform intelligent logic verification on the multi-code mapping decision candidate set, temporarily storing compliant mapping results and marked suspicious mappings according to the verification results, and obtaining preliminary screening and classification mapping data, specifically:

[0100] For the multi-code mapping decision candidate set, the preset specification parameter value range compliance verification rules, registration certificate number and manufacturer information association verification rules, and coding system logic matching rules are invoked to perform intelligent logic verification. The mapping results that meet all verification rules and the mapping results that trigger suspicious judgment conditions are selected. The mapping results that meet the verification rules are temporarily stored in the mapping cache pool to be integrated, and the mapping results that trigger suspicious judgment conditions are marked as suspicious mappings to be reviewed, thus obtaining the initial screening and classification mapping data.

[0101] For example, in a specific application scenario, the multi-code mapping decision candidate set contains two sets of core results: the first set is "20mL sterile syringe - medical insurance code C050101001" (overall score 0.86), with associated data being "specification 20mL, registration certificate number 20233150001, manufacturer A medical technology company"; the second set is "10mL intravenous infusion needle - UDI code 01987654321098" (overall score 0.68), with associated data being "specification 15mL (labeling error), registration certificate number 20243150002, manufacturer B medical equipment company". The system first calls three sets of preset multi-dimensional verification rules: First, the compliance verification rule for the numerical range of specifications and parameters (the allowable range for syringe specifications is 1-50mL, and the allowable range for infusion needle specifications is 5-20mL); second, the verification rule for the association between registration certificate number and manufacturer information (synchronizing "National Medical Device Registration Certificate 20233150001" with manufacturer A Medical Technology Company and "National Medical Device Registration Certificate 20243150002" with manufacturer C Medical Supplies Company from the National Medical Products Administration database); and third, the logical matching rule of the coding system (medical insurance code C0501 corresponds to injection and puncture devices, and UDI code 0198 corresponds to infusion devices). After verification, the first set of results met all the rules: the specification of 20mL was within the allowable range, the registration certificate number matched the manufacturer, and the coding system logic was consistent, thus it was judged as a compliant mapping result. The second set of results triggered two suspicious judgment conditions: the specification label "15mL" did not match the specification of the mapped result "10mL intravenous infusion needle" (violating the specification compliance rule), and the manufacturer "C Medical Supplies Company" corresponding to the registration certificate number was inconsistent with "B Medical Equipment Company" in the associated data (violating the registration certificate number-manufacturer association rule), thus it was marked as a suspicious mapping pending review. Finally, the system temporarily stored the compliant first set of results in the mapping cache pool to be integrated, and classified the marked suspicious second set of results separately, resulting in preliminary screening and classification mapping data containing "compliant cache pool + suspicious mapping list".

[0102] Therefore, the technical benefits of step 41 above include, but are not limited to, building a "pre-emptive, multi-dimensional risk filtering mechanism" for multi-code intelligent mapping, ensuring the accuracy of the mapping results from the perspective of data compliance: On the one hand, the multi-dimensional verification rules cover the core attributes (specifications), associated information (registration certificate number-manufacturer), and coding logic (system matching) of consumables, which can comprehensively identify explicit errors (such as contradictory specification labeling) and implicit risks (such as mismatched manufacturer information) in the mapping results, preventing compliance-related mapping results from entering subsequent stages and reducing actual business risks such as medical insurance settlement errors and consumable traceability failures; on the other hand, according to the verification results... The system of temporary storage and labeling enables an efficient processing model of "automatic flow of compliant results + targeted review of suspicious results." Compliant results directly enter the cache pool for integration without manual intervention, while suspicious results are clearly labeled with the problem type (such as specification mismatch or manufacturer mismatch), providing precise guidance for subsequent manual review and avoiding inefficiency caused by blind verification. In addition, the application of preset rules ensures the uniformity and stability of verification standards. Regardless of changes in consumable types or coding systems, verification is performed according to the same rules, ensuring that mapping results of different batches and categories maintain consistent standards at the compliance level, thereby improving the overall reliability of multi-code mapping from the perspective of process standardization.

[0103] Optionally, step 42 involves retrieving historical cases and matching policy basis from the preliminary screening classification mapping data for suspected mappings awaiting review, integrating them into an auxiliary decision-making information package, and associating it with the corresponding suspected mapping to obtain suspected mapping data with auxiliary decision-making information. Specifically:

[0104] For the suspected mappings pending review in the initial screening classification mapping data, the historical mapping case database interface is called to retrieve historical case records of similar consumable coding mappings. At the same time, the medical consumable coding policy basis database is matched to extract policy basis information related to the suspected mappings pending review. The retrieved historical case records and policy basis information are integrated into an auxiliary decision-making information package, and the corresponding suspected mappings pending review are associated and bound to the fields to obtain suspected mapping data with auxiliary decision-making information.

[0105] Exemplarily, in a specific application scenario, the suspicious mapping to be reviewed in the initial screening classification mapping data is "10 mL intravenous infusion needle - UDI code 01987654321098", and the associated suspicious issues are "the marked specification of 15 mL does not match the mapping result of 10 mL, and the manufacturer corresponding to the registration certificate number Guo械Zhu准20243150002 is inconsistent with the manufacturer of the associated data". The system first calls the historical mapping case database interface, uses "intravenous infusion needle + specification mismatch + manufacturer inconsistency" as the search keyword, and retrieves 3 similar historical case records in the past year: Case 1 is "8 mL intravenous infusion needle marked as 10 mL, and the mapping is completed after correcting the specification", Case 2 is "the manufacturer corresponding to the registration certificate number of the 12 mL intravenous infusion needle is incorrect, and the manufacturer information is updated after verification", and Case 3 is "the 10 mL intravenous infusion needle has dual problems of specification and manufacturer, and the UDI code is re-matched after referring to the policy". At the same time, the system matches the medical consumable coding policy reference library, extracts the clause "the UDI code needs to be consistent with the actual specification and registration information of the consumable" in the "Measures for the Administration of the Unique Device Identification" and the policy reference "the manufacturer information of the consumable needs to match the registration certificate filing information" in the "Regulations on the Supervision and Administration of the Use of Medical Security Funds". Subsequently, the 3 historical case records (including the processing process and correction results) and 2 policy references (including the clause number and specific requirements) are integrated into an auxiliary decision-making information package, which is associated and bound with the suspicious mapping data of "10 mL intravenous infusion needle - UDI code 01987654321098" through the "suspicious mapping ID" field, and finally the suspicious mapping data with auxiliary decision-making information is obtained, where each suspicious issue corresponds to historical case reference and policy clause support.

[0106] It can be seen that in the specific implementation of the above step 42, its technical benefits include but are not limited to providing "accurate and standardized decision support" for the manual review link of multi-code intelligent mapping, greatly improving the review efficiency and accuracy: on the one hand, the retrieval and integration of historical case records enable the reviewers to avoid manually searching for past processing experiences - for example, when dealing with "dual problems of specification and manufacturer", they can directly refer to the correction process of Case 3, avoiding the efficiency waste caused by repeated exploration, and at the same time predicting the correction direction through the processing results of similar cases, reducing the review difficulty; on the other hand, the matching and binding of policy reference information ensure that the manual review always complies with industry norms and regulatory requirements. For example, according to the "Measures for the Administration of the Unique Device Identification", it is clear that "the specification needs to be consistent with the UDI code", avoiding incorrect corrections caused by policy understanding deviations, and ensuring the reliability of the mapping result from a compliance perspective; in addition, the association and binding of the auxiliary decision-making information package and the suspicious mapping data achieve a one-to-one correspondence of "suspicious issue - case - policy", enabling the review process to have traceable decision-making basis. If it is necessary to review the review results later, the cases and policies referred to at that time can be quickly located, improving the quality and credibility of the manual review of multi-code mapping from the perspective of process standardization.

[0107] Optionally, in step 43, enable the cross-departmental expert consultation interaction module to receive expert correction instructions to correct the problem items in the suspicious mapping data with auxiliary decision-making information, and generate an expert-verified mapping set, specifically:

[0108] For the suspicious mapping data with auxiliary decision-making information, enable the built-in cross-departmental expert consultation interaction module to receive the mapping correction instructions input by experts to correct the incorrect mapping logic, inconsistent coding association relationships, and abnormal attribute matching items in the suspicious mapping data with auxiliary decision-making information and generate an expert-verified mapping set.

[0109] Exemplary explanation of the implementation of step 43

[0110] Exemplarily, in a specific application scenario, the suspicious mapping data with auxiliary decision-making information is "10mL intravenous infusion needle - UDI code 01987654321098", and the associated auxiliary decision-making information includes 3 similar historical cases and 2 policy bases. The suspicious problem is clearly "the specification marked as 15mL does not match the mapping result of 10mL, and the manufacturer corresponding to the registration certificate number Guo械注准20243150002 is inconsistent with the manufacturer of the associated data". The system first enables the built-in cross-departmental expert consultation interaction module, which supports inviting experts from the equipment department (responsible for consumable specification verification), the medical insurance department (responsible for coding association verification), and the information department (responsible for data correction operations) to collaborate online simultaneously. The module automatically pushes the suspicious mapping data and the auxiliary decision-making information package to the operation interfaces of the three experts. The expert from the equipment department compares historical case 1 and the provisions of the "Measures for the Administration of the Unique Identification of Medical Devices", verifies that the actual purchased specification of this intravenous infusion needle is 10mL, determines that "the specification marked as 15mL" is an error in data entry, and inputs the instruction "correct the specification marked to 10mL"; the expert from the medical insurance department refers to case 2 and the "Regulations on the Supervision and Administration of the Use of Medical Security Funds", and through the database of the National Medical Products Administration, confirms that the manufacturer corresponding to the registration certificate number Guo械注准20243150002 is "C Medical Supplies Company", determines that "B Medical Equipment Company" in the associated data is a deviation in the entered manufacturer name, and inputs the instruction "update the manufacturer information to C Medical Supplies Company"; the expert from the information department, based on the correction instructions of the previous two experts, performs data correction operations in the module, and simultaneously verifies that the specification, manufacturer information, and UDI coding logic of the corrected "10mL intravenous infusion needle - UDI code 01987654321098" are consistent. After all experts confirm that the correction is correct, the module automatically summarizes the correction results and generates an expert-verified mapping set containing "the corrected mapping relationship, expert signature, and correction time".

[0111] Technical benefits of step 43 in the intelligent multi-code mapping of in-hospital consumables

[0112] Therefore, the technical benefits of the specific implementation of step 43 above include, but are not limited to, providing a "professional and collaborative correction mechanism" for resolving suspicious issues in multi-code intelligent mapping, maximizing the accuracy and compliance of the mapping results: On the one hand, the cross-departmental expert consultation model integrates the technical capabilities of the equipment department (specification expertise), the medical insurance department (coding expertise), and the information department (data operation expertise), avoiding correction deviations caused by the professional limitations of a single department—for example, the information department alone cannot accurately verify the actual specifications of consumables, and the equipment department alone cannot determine the compliance of code associations. Multi-departmental collaboration ensures that each problem item can receive professional correction in the corresponding field; On the other hand... In terms of the combination of expert correction instructions and decision support information packages, the correction operation is supported by historical case references, policy basis, and expert experience, avoiding blind correction. For example, according to policy provisions, the manufacturer information must be consistent with the registration certificate filing, rather than being modified based solely on subjective judgment. In addition, the expert verification generated by the module, through the mapping set with expert signatures and correction records, not only realizes the traceability of the correction process, but also gives the subsequent integrated mapping data the credibility of "expert endorsement". This effectively avoids the risks of medical insurance settlement and failure of consumable traceability caused by the unresolved suspicious issues, and safeguards the reliability of multi-code mapping results from a professional review perspective.

[0113] Optionally, step 44, merging compliance mapping results and expert verification, forms a preliminary integrated mapping set through the mapping set. This is then linked to the audit process records to generate an integrated mapping set with traceability tags. After converting the format according to preset specifications, multi-code mapping association data is generated, specifically as follows:

[0114] The mapping results that meet the verification rules are extracted from the mapping cache pool to be integrated, and the data is merged with the mapping set verified by experts to form a preliminary integrated mapping set. The quality backtracking module is called to associate and bind each mapping result in the preliminary integrated mapping set with the corresponding audit process record with a unique identifier, generating an integrated mapping set with traceability tags. The integrated mapping set with traceability tags is converted into a structured format according to the preset field specifications of the multi-code mapping association data to generate multi-code mapping association data.

[0115] For example, in a specific application scenario, the cache pool to be integrated stores the compliant mapping result determined in step 41: "20mL sterile syringe - medical insurance code C050101001" (related data: specification 20mL, registration certificate number 20233150001, manufacturer A Medical Technology Company). The expert-verified mapping set includes the corrected result from step 43: "10mL intravenous infusion needle - UDI code 01987654321098" (corrected data: specification 10mL, registration certificate number 20243150002, manufacturer C Medical Supplies Company). The system first extracts the compliant mapping result from the cache pool, merges it with the expert-verified mapping set, and ensures no duplicate data through the "unique identifier ID" field, forming a preliminary integrated mapping set containing the two mapping results. Next, the quality backtracking module is invoked. This module automatically retrieves the audit process record for each mapping result: associating "20mL sterile syringe" with "Step 41 Smart Verification Pass Record (Verification Time 2025-08-22 10:00, Verification Rules Fully Covered)", and associating "10mL intravenous infusion needle" with "Step 41 Suspicious Mark Record + Step 42 Assisted Decision Information Package Generation Record + Step 43 Expert Correction Instruction Record (Including Three Expert Signatures, Correction Time 2025-08-22 11:30)". A unique traceability identifier (such as "MAP-20250822-001" "MAP-20250822-002") is assigned to each mapping result, thereby binding the mapping result with the audit record and generating an integrated mapping set with traceability markers. Finally, the system converts the integrated mapping set with traceability tags into a structured table format according to the preset field specifications of the multi-code mapping association data (fields include "consumable name, specifications, registration certificate number, manufacturer, medical insurance code, UDI code, traceability identifier, review status, and generation time"). The review status of "20mL sterile syringe" is marked as "intelligent verification passed", and the review status of "10mL intravenous infusion needle" is marked as "expert verification passed", thus generating the multi-code mapping association data.

[0116] Therefore, the technical benefits of step 44 above include, but are not limited to, providing "complete, traceable, and standardized associated data results" for multi-code intelligent mapping, thus providing a reliable data foundation for subsequent applications. On the one hand, the merging of compliance results and expert verification results achieves "full-scenario coverage" of multi-code mapping data—including both efficient intelligent verification results requiring no manual intervention and complex, questionable mapping results that have been professionally corrected, avoiding application gaps caused by data splitting and meeting the hospital's unified management needs for mapping data of different review types. On the other hand, the association between traceability markers and review records enables multi-code mapping associated data to have "full-chain traceability." If it is necessary to verify the source of a mapping result later (such as the medical insurance department's random check of code compliance), the corresponding intelligent verification record or expert correction record can be quickly retrieved through the traceability identifier, without having to re-examine the process, which greatly improves the efficiency of data verification. In addition, the pre-defined field standard format conversion enables multi-code mapping associated data to be directly adapted to downstream applications such as hospital HIS system, medical insurance settlement system, and consumable traceability system without secondary format adjustment, reducing the cost of data interaction between systems. At the same time, the standardized field design ensures that the data dimensions obtained by different departments (such as medical insurance department and equipment department) are consistent, avoiding misunderstandings caused by differences in data format, and ensuring the practical value of multi-code mapping from the data output level.

[0117] Optionally, step 5, constructing a knowledge graph based on multi-code mapping association data to generate a multi-code mapping association graph, specifically includes the following steps:

[0118] Step 51: Extract entity-relation pairs from the multi-code mapping association data, process duplicate nodes using the entity alignment algorithm, and construct the entity layer, relation layer, and rule layer containing core entities, semantic relations, and consumable coding mapping rules according to the attribute graph model to generate a basic knowledge network;

[0119] Step 52: Map the basic knowledge network to the vector space through the entity relationship embedding model, construct triples containing consumable-related head entities, association relationships, and code-related tail entities, and capture entity-relation vector sets through semantic association; then mine indirect associations of multi-code mappings through path reasoning algorithm and filter effective results to supplement the knowledge network relationship layer, generating a knowledge embedding enhanced knowledge network.

[0120] Step 53: Based on the dedicated rule base of the medical consumables coding system, perform compliance verification and cross-validation on the entity attributes and relational logic of the knowledge embedding enhanced knowledge network to obtain compliant knowledge network fragments; perform association alignment and attribute fusion on the fragments, and determine the valid attributes according to the knowledge source priority to eliminate conflicts in entity attribute conflicts and generate a consistent knowledge network.

[0121] Step 54: Optimize the topology of the consistency knowledge network to reduce redundant association paths, add identification information and unique traceability identifiers to entity nodes and relationship edges, and write them into the graph database according to the attribute graph model adapted to multi-code association query of consumables to generate a multi-code mapping association graph.

[0122] Optionally, step 51 involves extracting entity-relation pairs from the multi-code mapping association data, processing duplicate nodes using an entity alignment algorithm, and constructing an entity layer, relation layer, and rule layer containing core entities, semantic relationships, and consumable coding mapping rules according to the attribute graph model to generate a basic knowledge network. Specifically:

[0123] Entity-relationship pair extraction is performed on the multi-code mapping associated data to obtain initial entity-relationship pairs. The entity alignment algorithm is called on the initial entity-relationship pairs to eliminate duplicate nodes and merge duplicate ontology nodes of the same consumable and duplicate encoding nodes of the same code from different data sources to obtain deduplicated entity-relationship pairs. For the deduplicated entity-relationship pairs, an entity layer containing multiple core consumable attribute entities, a relation layer containing multiple consumable semantic relationships, and a rule layer storing consumable encoding mapping rules are designed according to the attribute graph model to generate a basic knowledge network.

[0124] Exemplarily, in a specific application scenario, the multi-code mapping associated data contains two core records: one is "20mL sterile syringe (traceability identification MAP-20250822-001), specification 20mL, registration certificate number National Medical Device Approval No. 20233150001, manufacturer A Medical Technology Company, medical insurance code C050101001, audit status intelligent verification passed"; the other is "10mL intravenous infusion needle (traceability identification MAP-20250822-002), specification 10mL, registration certificate number National Medical Device Approval No. 20243150002, manufacturer C Medical Supplies Company, UDI code 01987654321098, audit status expert verification passed". The system first performs entity-relationship pair extraction on these two pieces of data, extracting initial entity-relationship pairs such as "20mL sterile syringe - belongs to - injection and puncture device", "20mL sterile syringe - corresponds to medical insurance code - C050101001", "10mL intravenous infusion needle - corresponds to UDI code - 01987654321098", "A Medical Technology Company - produces - 20mL sterile syringe". Then, it calls the entity alignment algorithm. It is found that there is an expression of "20 ml sterile syringe" in the external data source. Through semantic similarity comparison (similarity 0.98), it is determined that it is the same consumable ontology node as "20mL sterile syringe" and merged into a single entity; at the same time, it is found that there is a duplicate coding node of "medical insurance code C050101001" as "C050101001 (medical insurance)" in the historical data. After the algorithm eliminates redundancy, the standard coding entity is retained, and finally the deduplicated entity-relationship pairs are obtained. Subsequently, a three-layer structure is constructed according to the property graph model: the entity layer contains core entities such as "20mL sterile syringe", "10mL intravenous infusion needle", "C050101001", "01987654321098", "A Medical Technology Company", etc., and each entity is attached with attributes such as specification and registration certificate number; the relationship layer defines semantic relationships such as "belongs to", "corresponds to medical insurance code", "corresponds to UDI code", "produces", etc.; the rule layer stores consumable coding mapping rules such as "consumables of the injection and puncture device category should be preferentially matched with medical insurance codes starting with C0501" and "the UDI code needs to be associated and verified with the registration certificate number". The three together form a basic knowledge network.

[0125] Therefore, the technical benefits of the specific implementation of step 51 above include, but are not limited to, building a "structured and redundancy-free knowledge support framework" for multi-code intelligent mapping, promoting the upgrade of mapping from "data matching" to "knowledge-driven": On the one hand, the entity alignment algorithm eliminates duplicate nodes, solving the problem of "multiple representations of the same entity" in multi-source data—for example, merging "20mL sterile syringe" and "20 ml sterile syringe," avoiding logical confusion in subsequent knowledge association, and ensuring the uniqueness and accuracy of entities such as consumables and codes; on the other hand, the three-layer structure (entity layer, relation layer, and rule layer) of the attribute graph model is clearly defined. The system streamlines the relationships between consumables, codes, and manufacturers, transforming scattered mapping data into a structured knowledge network. For example, the relationship "20mL sterile syringe - corresponding medical insurance code - C050101001" can intuitively present the association path between consumables and codes. When adding new mappings for similar consumables, the coding mapping rules of the rule layer can be directly referenced, reducing redundant calculations. Furthermore, the construction of the basic knowledge network provides a high-quality initial data foundation for subsequent knowledge embedding enhancement and path reasoning, avoiding deviations in subsequent knowledge mining caused by entity redundancy and relationship confusion. From the knowledge architecture level, it ensures the accuracy and scalability of multi-code mapping.

[0126] Optionally, in step 52, the basic knowledge network is mapped to a vector space using an entity-relationship embedding model, constructing triplets containing consumable-related head entities, association relationships, and code-related tail entities. An entity-relationship vector set is obtained through semantic association capture. Then, a path reasoning algorithm is used to mine indirect associations through multi-code mapping and filter effective results, supplementing the knowledge network relationship layer to generate a knowledge embedding-enhanced knowledge network. Specifically:

[0127] The TransE model is invoked to map the basic knowledge network to a defined vector space to determine the head entity (including the consumable ontology), consumable relationships, and tail entities (including medical insurance code entities). Triples are constructed based on the head entity, consumable relationship, and tail entity. Semantic association is captured by minimizing the loss function to strengthen the semantic association between consumables and multiple codes, and between multiple codes, resulting in an entity-relation vector set. GNN path reasoning is performed on the entity-relation vector set to mine the indirect associations of multi-code mapping between consumables and multiple code systems, and between multiple code systems themselves. The indirect association path weights of multi-code mapping are calculated based on the confidence of inter-code relationships determined by historical multi-code mapping data of consumables and the similarity of inter-code standard relationships. Valid association reasoning results with acceptable path weights are selected from the mined indirect associations of multi-code mapping. The valid association reasoning results are added to the relation layer of the knowledge network to update the semantic association dimensions between consumables and multiple codes, and between multiple codes, resulting in a knowledge embedding-enhanced knowledge network.

[0128] For example, in a specific application scenario, the basic knowledge network generated in step 51 includes an entity layer ("20mL sterile syringe", "10mL intravenous infusion needle", "C050101001", "01987654321098"), a relationship layer ("corresponding medical insurance code", "corresponding UDI code", "belongs to"), and a rule layer ("injection and puncture device class matches medical insurance code starting with C0501"). The system first calls the TransE model to map all entities and relations in the basic knowledge network to a 100-dimensional vector space: "20mL sterile syringe" is mapped to vector V1, "C050101001" is mapped to vector V2, and "corresponding medical insurance code" is mapped to vector R1. According to the logic of the TransE model, V1+R1≈V2, and thus the triple "20mL sterile syringe (head entity)-corresponding medical insurance code (relation)-C050101001 (tail entity)" is determined. Similarly, triples such as "10mL intravenous infusion needle-corresponding UDI code-01987654321098" and "20mL sterile syringe-belongs to-injection and puncture instruments" are constructed. Subsequently, by minimizing the L2 loss function of “V1+R1-V2”, the semantic association of the triples is strengthened—for example, the vector similarity between “20mL sterile syringe” and “C050101001” is increased from 0.85 to 0.92, resulting in an entity-relation vector set containing all entity and relation vectors. Next, GNN path reasoning is performed on the vector set to mine indirect associations: through the paths “20mL sterile syringe - belongs to - injection puncture instruments”, “injection puncture instruments - matching rule - medical insurance code starting with C0501”, and “10mL intravenous infusion needle - belongs to - injection puncture instruments”, the indirect association “10mL intravenous infusion needle - may correspond to medical insurance code starting with C0501” is inferred; at the same time, through “C050101001 - association rule - injection puncture instruments” and “01987654321098 - association rule - injection puncture instruments”, the indirect association between the codes “C050101001 - same category - 01987654321098” is inferred. Based on historical multi-code mapping data, the system determined that the confidence level of the inter-code relationship between "injection and puncture instruments and medical insurance codes starting with C0501" was 0.9, and the standard similarity of the inter-code relationship between "associations between codes of the same category" was 0.85. The path weights of the above two indirect associations were calculated to be 0.88 and 0.82 respectively (both higher than the set threshold of 0.8), and were selected as valid association inference results. Finally, "10mL intravenous infusion needle - may correspond to medical insurance codes starting with C0501" and "C050101001 - same category - 01987654321098" were added to the relationship layer, updating the semantic association dimension of the relationship layer, resulting in a knowledge embedding-enhanced knowledge network.

[0129] Therefore, the technical benefits of step 52 above include, but are not limited to, constructing a "semantically enhanced and association-expanded knowledge network" for multi-code intelligent mapping, breaking through the limitations of traditional direct mapping: On the one hand, the vector mapping and loss function optimization of the TransE model transform entities and relations into computable vectors, strengthening the deep semantic associations between consumables and codes, and between codes—for example, improving the vector similarity between "20mL sterile syringe" and "C050101001," enabling more accurate identification of semantic associations during subsequent matching and reducing misjudgments caused by differences in surface features; on the other hand, the indirect associations mined by GNN path reasoning expand the... The association dimension of multi-code mapping—for example, the inference result of "10mL intravenous infusion needle - may correspond to medical insurance code starting with C0501"—provides directional guidance for the subsequent matching of medical insurance codes for this consumable, while the "same category association between C050101001 and 01987654321098" provides support for cross-mapping between multiple coding systems. In addition, path weight screening ensures the effectiveness of indirect associations and avoids invalid inference results from interfering with the mapping logic. At the same time, the updated relationship layer makes the knowledge network more complete, providing stronger knowledge support for the accuracy and scalability of subsequent multi-code mapping, and promoting the upgrade of mapping from "based on existing data" to "based on knowledge reasoning".

[0130] Optionally, in step 53, based on the dedicated rule base of the medical consumables coding system, compliance verification and cross-validation are performed on the entity attributes and relational logic of the knowledge embedding enhanced knowledge network to obtain compliant knowledge network fragments; association alignment and attribute fusion are performed on the fragments, and for entity attribute conflicts, valid attributes are determined according to the priority of knowledge source to eliminate conflicts, generating a consistent knowledge network, specifically:

[0131] The system invokes a dedicated rule base for the medical consumable coding system, which includes medical insurance code matching rules, UDI code parsing rules, and consumable attribute-code association rules. It performs compliance checks on entity attributes and relational logic within the knowledge-embedded enhanced knowledge network, filtering out entity-relationship data that conforms to the rules to obtain compliant knowledge network fragments. For cross-coding system associations within these compliant knowledge network fragments, it performs cross-validation by matching historical multi-code mapping records of consumables in the multi-code mapping association data to confirm the validity of the associations, obtaining cross-validated knowledge network fragments. For each cross-validated knowledge network fragment, it performs node and relation association alignment processing, uniquely binding the ontology nodes of the same consumable and the coding nodes of the same code in different fragments, and performing attribute fusion on semantic relationships of the same type between fragments. For entity attribute conflicts that still exist after association alignment and attribute fusion, it determines the valid attribute values ​​according to the priority of consumable knowledge sources, replaces the invalid values ​​of conflicting attributes, eliminates the conflicts, and obtains a consistent knowledge network.

[0132] For example, in a specific application scenario, the knowledge embedding enhanced knowledge network generated in step 52 includes an entity layer ("20mL sterile syringe", "10mL intravenous infusion needle", "C050101001", "01987654321098"), an updated relationship layer (adding "10mL intravenous infusion needle - may correspond to medical insurance code starting with C0501", "C050101001 - same category - 01987654321098"), and a dedicated rule base for the medical consumables coding system includes rules such as "medical insurance code starting with C0501 corresponds to injection and puncture instruments", "the 3rd to 6th digits of the UDI code must match the consumable type (0198 corresponds to infusion device)", and "the consumable specifications must be consistent with the coding specification". The system first calls the rule base to perform compliance checks: When checking "10mL intravenous infusion needle - may correspond to medical insurance code starting with C0501", the association is deemed non-compliant because "intravenous infusion needles belong to infusion devices, and C0501 corresponds to injection and puncture devices" violates the "medical insurance code - consumable type matching rule"; when checking "C050101001 - same category - 01987654321098", both are associated with injection / puncture / infusion devices (belonging to the same category of medical devices), which conforms to the "same category code association rule"; at the same time, it is confirmed that the entity attributes "20mL sterile syringe specification 20mL" and "C050101001 code specification 20mL" are both compliant. Finally, entity-relationship data that conforms to the rules are filtered out to obtain a compliant knowledge network fragment (excluding the non-compliant "10mL intravenous infusion needle - may correspond to medical insurance code starting with C0501"). Next, the cross-coding system association relationship of “C050101001-same category-01987654321098” in the fragment was matched with the historical records in the multi-code mapping association data. In the past year, there was a record that “the medical insurance code for injection and puncture devices and the UDI code for infusion equipment are associated in the same category because they belong to sterile consumables”. The association was confirmed to be valid, and a knowledge network fragment that passed cross-validation was obtained. Subsequently, association alignment was performed on another batch of newly added “5mL sterile syringe” association fragments: the “5mL sterile syringe” ontology node was bound to the existing “20mL sterile syringe” “injection and puncture device” category node with a unique identifier (both associated with “injection and puncture device-category ID: 001”), and the “corresponding medical insurance code” relationship between “5mL sterile syringe-corresponding medical insurance code-C050101002” and “20mL sterile syringe-corresponding medical insurance code-C050101001” was fused with attributes (uniformly labeled with “medical insurance code association confidence 0.95”).Finally, it was found that the entity "10mL intravenous infusion needle" had an attribute conflict between "specification 10mL (from expert verification)" and "specification 12mL (from historical imported data)". According to the priority of knowledge source (expert verified data > historical imported data), "specification 10mL" was determined to be a valid attribute, and the invalid value of "specification 12mL" was replaced to eliminate the conflict and finally generate a consistent knowledge network.

[0133] Therefore, the technical benefits of step 53 above include, but are not limited to, creating a "compliant and conflict-free knowledge network foundation" for multi-code intelligent mapping, ensuring the reliability and consistency of knowledge support: On the one hand, the compliance verification and cross-validation of the dedicated rule base can accurately eliminate non-compliant associations in the knowledge network (such as incorrect medical insurance code mapping directions), avoiding invalid or erroneous knowledge from interfering with subsequent multi-code mapping decisions. For example, eliminating "10mL intravenous infusion needle - may correspond to a medical insurance code starting with C0501" prevents code type mismatch during subsequent matching; on the other hand, the association alignment processing realizes "node unification and relationship" of different knowledge fragments. "Integration" is employed, for example, binding "5mL sterile syringe" and "20mL sterile syringe" to the same category node, ensuring that similar consumables have a unified association benchmark in the knowledge network. When mapping new consumables of the same type, the category association rules can be directly reused, improving mapping efficiency. In addition, the attribute conflict elimination mechanism determines the effective attributes according to the priority of knowledge sources, which solves the entity attribute confusion caused by multi-source data (such as the specification conflict of "10mL intravenous infusion needle"), ensuring the uniqueness and accuracy of entity attributes in the knowledge network, providing a consistent data foundation for feature extraction and vector calculation in subsequent multi-code mapping, and ensuring the accuracy and stability of multi-code mapping from the perspective of knowledge quality.

[0134] Optionally, step 54 involves optimizing the topology of the consistent knowledge network to reduce redundant association paths, adding identification information and unique traceability identifiers to entity nodes and relationship edges, and writing them into the graph database according to the attribute graph model adapted for multi-code association queries of consumables, thereby generating a multi-code mapping association graph. Specifically:

[0135] The topology of the consistent knowledge network is optimized to adjust the distribution of entity nodes and the connection logic of relation edges, reducing redundant association paths to obtain an optimized topology knowledge network. Each entity node in the optimized topology knowledge network is labeled with consumable ID, coding type, and relation edges are labeled with multi-code mapping confidence and a unique traceability identifier is added to obtain a knowledge network with traceability identifier. According to the attribute graph model adapted to consumable multi-code association query, the knowledge network with traceability identifier is written into the graph database to generate a multi-code mapping association graph.

[0136] For example, in a specific application scenario, the consistency knowledge network generated in step 53 includes an entity layer (“20mL sterile syringe”, “10mL intravenous infusion needle”, “5mL sterile syringe”, “C050101001”, “C050101002”, “01987654321098”, “Injection and puncture device - Category ID: 001”) and a relationship layer (“20mL sterile syringe - Corresponding medical insurance code - C050101001”, “5mL sterile syringe - Corresponding medical insurance code - C050101002”, “10mL intravenous infusion needle - Corresponding UDI code - 01987654321098”, “Class III consumables - Same category - Injection and puncture device - Category ID: 001”). The system first performs topology optimization on the network: it identifies two directly related paths, "20mL sterile syringe → injection puncture instrument - category ID: 001" and "5mL sterile syringe → injection puncture instrument - category ID: 001", which can replace the original redundant path "20mL sterile syringe → 5mL sterile syringe → injection puncture instrument - category ID: 001", and deletes redundant edges; at the same time, it adjusts the distribution of entity nodes, classifying "medical insurance code entity" and "UDI code entity" under the "coding system" sub-node, and classifying "consumable entity" under the "consumable type" sub-node, reducing cross-level jump paths and obtaining an optimized topology knowledge network. Next, add identification information to each entity node: "20mL sterile syringe" is labeled with consumable ID "HC-20250822-001" and type "injection and puncture instrument"; "C050101001" is labeled with code type "medical insurance code" and system "national medical insurance catalog"; add confidence to the relationship edges: "20mL sterile syringe - corresponding medical insurance code - C050101001" is labeled with confidence of 0.95, and "10mL intravenous infusion needle - corresponding UDI code - 01987654321098" is labeled with confidence of 0.92; at the same time, add a unique traceability identifier to all entities and relationships (such as "entity traceability ID: ENT-20250822-001" "relationship traceability ID: REL-20250822-001"). The traceability identifier is associated with the audit record ID in step 53, resulting in a knowledge network with traceability identifiers.Finally, based on the attribute graph model adapted for multi-code association queries (fields include "entity ID, entity type, identification information, relationship type, confidence level, and traceability identifier"), the knowledge network with traceability identifiers is written into the Neo4j graph database: nodes are created using Cypher statements (e.g., "CREATE(n:MedicalConsumable{id:'HC-20250822-001',name:'20mL sterile syringe',type:'injection and puncture instrument',traceId:'ENT-20250822-001'}"). The multi-code mapping association graph is generated by combining the relationships (e.g., "MATCH(a:MedicalConsumable),(b:MedicalCode)WHEREa.id='HC-20250822-001'ANDb.code='C050101001'CREATE(a)-[r:MAPS_TO_MEDICAL_INSURANCE{confidence:0.95,traceId:'REL-20250822-001'}]->(b)").

[0137] Therefore, the technical benefits of the specific implementation of step 54 above include, but are not limited to, outputting a "knowledge graph result that is highly efficient in querying, traceable, and engineering-ready" for multi-code intelligent mapping, realizing the transformation from a knowledge network to a practical tool: On the one hand, topology optimization reduces redundant paths (such as deleting redundant associations across consumable entities), improving graph query efficiency by more than 30%—for example, when querying "medical insurance codes corresponding to injection and puncture instruments," all associated codes can be quickly located directly through "consumable type sub-nodes," without traversing irrelevant entities, significantly shortening the code verification time before hospital medical insurance settlement; on the other hand, the identification information of entities and relationships, as well as traceability identifiers, enable the graph to have "visual query and full-link traceability" capabilities. When medical insurance staff query the code "C050101001", they can directly view the associated attributes and mapping confidence of the "20mL sterile syringe". They can also retrieve the compliance verification record from step 53 through the traceability identifier to ensure the compliance of the code usage. In addition, the attribute graph model and graph database adapted for the query are written, enabling the multi-code mapping association graph to be directly connected to the hospital's consumables management system and medical insurance declaration system. This supports diverse query needs such as "searching by code for consumables" and "searching for similar consumables by code" (e.g., the equipment department can quickly query the UDI code and associated medical insurance code recommendations for "10mL intravenous infusion needle"). This maximizes the practical value of multi-code mapping from the perspective of practical application and provides knowledge support for the hospital's full-process management of consumables.

[0138] Figure 2 This is a schematic diagram of the structure of a multi-code intelligent mapping system for hospital consumables, as described in an embodiment of this application. Figure 2As shown, it includes: a front-end electronic device as a data acquisition aid and human interaction terminal, performing the following operations: collecting multi-source in-hospital consumable data through data access components (such as HL7FHIR interface, UDI barcode scanning module), transmitting it to the back-end server to support the generation of structured consumable element data; simultaneously receiving human feedback correction data through expert interaction components (such as touch screen, keyboard), transmitting it to the back-end server to provide human correction basis for the verification and correction of the multi-code mapping decision candidate set; the back-end server performing the following operations: receiving multi-source in-hospital consumable data transmitted by the front-end electronic device, generating structured consumable element data based on the data; then generating a consumable feature vector matrix based on the structured consumable element data; subsequently, performing weighted matching processing on the feature vector matrix and the multi-code system feature library according to the weighted weight coefficient to generate a multi-code mapping decision candidate set; then, combining the preset verification rules and the human feedback correction data transmitted by the front end, verifying and correcting the decision candidate set to generate multi-code mapping association data; finally, constructing a knowledge graph based on the multi-code mapping association data, generating a multi-code mapping association graph and feeding it back to the front-end electronic device for users to view.

[0139] The above description is merely an exemplary embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent mapping of multiple codes for hospital consumables, characterized in that, include: Step 1: Generate structured consumable element data based on the collected multi-source in-hospital consumable data; Step 2: Generate a consumable feature vector matrix based on structured consumable element data; Step 3: Based on the generated weighted coefficients, perform weighted matching processing on the feature vector matrix and the feature library of the multi-code system to generate a multi-code mapping decision candidate set; Step 4: Verify and correct the multi-code mapping decision candidate set according to the preset verification rules and manual feedback correction data to generate multi-code mapping association data; Step 5: Construct a knowledge graph based on the multi-code mapping association data to generate a multi-code mapping association graph; Step 2 includes: vectorizing the structured consumable element data to generate a consumable feature vector matrix. The rows of the consumable feature vector matrix represent a single consumable sample, the columns represent the quantitative dimensions of the key elements of the consumable, and the elements at the intersection of the rows and columns represent the feature values ​​of the corresponding consumable sample in the quantitative dimensions. Step 3 includes: Step 31, extracting structured knowledge from the medical consumables coding domain to generate an adaptive rule weight library; Step 32: The consumable feature vector matrix and the vector data in the multi-coding system feature library are weighted and calculated based on the matching weights output by the adaptive rule weight library to obtain the comprehensive matching score between each consumable sample and the multi-coding system feature library. Step 33: Generate a dynamically optimized mapping candidate set based on the comprehensive matching score; Step 34: Generate a multi-code mapping decision candidate set based on the dynamically optimized mapping candidate set.

2. The method according to claim 1, characterized in that, Step 1: Based on the collected multi-source in-hospital consumable data, generate structured consumable element data, specifically as follows: Collect multi-source consumable data from within the hospital, and perform hierarchical analysis to extract key consumable elements, thereby generating structured consumable element data.

3. The method according to claim 2, characterized in that, Collect multi-source consumable data from within the hospital, and perform hierarchical analysis to extract key consumable elements, thereby generating structured consumable element data. The specific steps include the following: Step 11: Capture the consumable settlement records during HIS implementation, and synchronously connect to the medical insurance consumable standard catalog library and UDI database to collect multi-source original consumable datasets; Step 12: Perform word segmentation and named entity recognition on the multi-source original consumables dataset to obtain word segmentation and entity annotation results, and perform unstructured to structured conversion on the obtained word segmentation and entity annotation results to obtain preliminary structured data; Step 13: Perform semantic augmentation on the preliminary structured data based on the medical BERT model to obtain semantically augmented data; Step 14: Generate structured consumable element data based on semantically enhanced data.

4. The method according to claim 1, characterized in that, The structured consumable element data is vectorized to generate a consumable feature vector matrix. In this matrix, rows represent individual consumable samples, columns represent the quantification dimensions of key consumable elements, and the elements at the intersection of rows and columns represent the feature values ​​of the corresponding consumable sample in the quantification dimension. The specific steps include: Step 21: Based on the constructed medical consumables-specific word segmentation word list, synonym word list and stop word list, perform word segmentation enhancement processing on the consumables field in the structured consumables element data to obtain the word segmentation enhancement consumables element text; Step 22: Using the Skip-gram model of Word2Vec, map each word in the segmented and enhanced consumables element text to a dense vector. Based on the dense vector, generate core feature word vectors that can represent the core features of consumables. Step 23: Associate the core feature word vectors with the core attribute information of consumables to supplement the feature expression of the attribute dimension in the core feature word vectors and generate a multi-dimensional feature vector containing consumable terminology semantics, consumable attribute features, and consumable domain knowledge; arrange the multi-dimensional feature vectors into a matrix by using individual consumable samples as rows and the quantification dimension of key consumable elements as columns to generate a consumable feature vector matrix.

5. The method according to claim 4, characterized in that, Step 23, specifically: For the core feature word vectors and the core attribute information of consumables, rule matching is performed by constructing a term-attribute association rule base to obtain term-attribute association pairs; the core attribute information of consumables in the term-attribute association pairs is encoded into attribute vectors to obtain attribute vectors; the attribute vectors are concatenated with the corresponding core feature word vectors to obtain attribute-enhanced vectors; medical consumable classification standards and compliance requirements are introduced into the attribute-enhanced vectors as domain knowledge labels, and label mapping is performed to generate multi-dimensional feature vectors containing consumable term semantics, consumable attribute features, and consumable domain knowledge; the multi-dimensional feature vectors are indexed and labeled according to the quantification dimension of the key elements of consumables to obtain indexed feature vectors; the indexed feature vectors are filled in an ordered manner with individual consumable samples as rows and the quantification dimension of the key elements of consumables as columns to generate a consumable feature vector matrix.

6. The method according to claim 1, characterized in that, Step 3: Based on the generated weighted coefficients, perform weighted matching processing on the feature vector matrix and the multi-coding system feature library to generate a multi-code mapping decision candidate set, specifically: Dynamic weight matching is performed on the feature vector matrix of consumables to obtain the weight coefficients of each dimension of features of different consumable samples. Based on the NLP semantic similarity calculation results, the feature vector matrix and the feature library of the multi-code system are weighted and matched to generate a multi-code mapping decision candidate set.

7. The method according to claim 1, characterized in that, Step 4 specifically includes the following steps: Step 41: Call the preset multi-dimensional verification rules to perform intelligent logic verification on the multi-code mapping decision candidate set, and temporarily store the compliant mapping results and the marked suspicious mappings according to the verification results to obtain the initial screening classification mapping data; Step 42: For the suspicious mappings pending review in the initial screening classification mapping data, retrieve historical cases and matching policy basis and integrate them into auxiliary decision-making information packages, and associate and bind them to the corresponding suspicious mappings to obtain suspicious mapping data with auxiliary decision-making information; Step 43: Activate the cross-departmental expert consultation interaction module, receive expert correction instructions to correct problematic items in suspicious mapping data with auxiliary decision-making information, and generate an expert-verified mapping set; Step 44: Merge compliance mapping results and expert verification. A preliminary integrated mapping set is formed by combining the mapping sets. The audit process records are associated to generate an integrated mapping set with traceability tags. After converting the format according to the preset specifications, multi-code mapping association data is generated.

8. The method according to claim 1, characterized in that, Step 5 specifically includes the following steps: Step 51: Extract entity-relation pairs from the multi-code mapping association data, process duplicate nodes using the entity alignment algorithm, and construct the entity layer, relation layer, and rule layer containing core entities, semantic relations, and consumable coding mapping rules according to the attribute graph model to generate a basic knowledge network; Step 52: Map the basic knowledge network to the vector space through the entity-relation embedding model, construct triples containing consumable-related head entities, association relationships, and encoding-related tail entities, and obtain the entity-relation vector set through semantic association capture; Then, the path reasoning algorithm is used to mine the indirect associations of multi-code mapping and filter the effective results to supplement the knowledge network relationship layer, generating a knowledge embedding enhanced knowledge network; Step 53: Based on the dedicated rule base of the medical consumables coding system, perform compliance verification and cross-validation on the entity attributes and relational logic of the knowledge embedding enhanced knowledge network to obtain compliant knowledge network fragments; Perform association alignment and attribute fusion on the fragments, and determine the valid attributes according to the knowledge source priority to eliminate conflicts in entity attribute conflicts, thereby generating a consistent knowledge network; Step 54: Optimize the topology of the consistency knowledge network to reduce redundant association paths, add identification information and unique traceability identifiers to entity nodes and relationship edges, and write them into the graph database according to the attribute graph model adapted to multi-code association query of consumables to generate a multi-code mapping association graph.

Images