Medical claim settlement material identification method and storage medium thereof

Abstract

The invention provides a medical claim settlement material identification method and a storage medium thereof, and belongs to the technical field of artificial intelligence, and the method comprises the steps: inputting a medical claim settlement image into a cross-modal network model, and obtaining a prediction feature vector of each candidate entity; screening out an effective standard entity from the claim element knowledge graph based on the context information to obtain a standard entity embedding vector; and determining a target standard entity of the target candidate entity according to the semantic distance between the two entities, and correcting the target candidate entity by using the information of the target standard entity to generate a structured claim element recognition result. According to the method, the search space of the knowledge graph is dynamically constrained through the spatio-temporal context, the business compliance of the to-be-calibrated range is ensured, the output of the model is forced to be close to the standard entity by using the vector semantic distance, the noise and illusion in cross-modal recognition are effectively corrected, and the accuracy and availability of medical claim element recognition are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for identifying medical claim materials and its storage medium. Background Technology

[0002] With the digital transformation of the insurance industry, the claims service involves a large number of heterogeneous and non-standardized medical material processing needs. There is an urgent need to use automation technology to accurately extract key claims elements such as drugs and treatment items from images such as outpatient medical records and invoices in order to improve claims efficiency and reduce operating costs.

[0003] Existing technologies primarily employ a pipelined approach combining Optical Character Recognition (OCR) and Named Entity Recognition (NER) for information extraction. Specifically, existing solutions typically first denoise and analyze the layout of medical document images to identify text regions; then, OCR technology is used to convert the image into a text sequence; finally, the text sequence is input into a sequence labeling model, such as BiLSTM-CRF or BERT-NER, to identify medical entities through contextual semantic analysis. Some improved solutions introduce simple layout analysis to assist in locating text blocks, or utilize a general knowledge base as a post-processing dictionary to perform simple string matching verification on the identified text.

[0004] The aforementioned existing technologies still have limitations in practical applications. On the one hand, due to limitations in image quality and font differences, the recognition results output by existing technologies often contain typos or non-standard expressions that deviate from standard medical terminology, making it difficult to directly use the extracted element data for standardized business processing. On the other hand, the final output claims element data often contains invalid or non-compliant information, reducing the accuracy and usability of the recognition results in the subsequent claims settlement process. Summary of the Invention

[0005] This invention provides a method for identifying medical claim materials and its storage medium, which solves the technical problems in the prior art where the identification results are prone to contain erroneous expressions that deviate from standard terminology, and the output claim elements are often mixed with invalid or non-compliant information in the current medical treatment environment, resulting in low identification accuracy and usability. It achieves high-precision and compliant identification and extraction of medical material elements in complex claim scenarios.

[0006] This invention provides a method for identifying medical claim materials, comprising: Determine the contextual information associated with the medical claim image to be identified; The medical claim image is input into a cross-modal network model to obtain at least one candidate entity output by the cross-modal network model and a predicted feature vector corresponding to each candidate entity; Based on the context information, all valid standard entities are selected from the pre-set claims element knowledge graph, and the standard entity embedding vector of each valid standard entity is obtained. Based on the semantic distance between the predicted feature vector and the standard entity embedding vector, a target standard entity matching each target candidate entity is determined from all the valid standard entities; the target candidate entity is any one of all the candidate entities. The target candidate entities that are matched with each target standard entity are modified using the information of each target standard entity to generate structured claims element identification results.

[0007] According to a method for identifying medical claim materials provided by the present invention, determining the context information associated with the medical claim image to be identified includes: The consultation time and consultation location information are extracted from the text content of the medical claim image; Alternatively, it can receive the patient's appointment time and location information input by the user in relation to the medical claim image.

[0008] According to a method for identifying medical claim materials provided by the present invention, the step of filtering all valid standard entities from a pre-set claim element knowledge graph based on the context information includes: Traverse the entity nodes in the knowledge graph of the claims elements; each entity node is configured with an effective time attribute and an applicable geographical attribute. Determine whether the time range defined by the effective time attribute includes the consultation time information in the context information to obtain a first determination result; Determine whether the geographical scope defined by the applicable geographical attribute includes the medical treatment geographical information in the context information to obtain a second determination result; Entity nodes whose first and second judgment results are both "yes" are determined as the valid standard entities.

[0009] According to a method for identifying medical claim materials provided by the present invention, the step of determining the target standard entity matching each target candidate entity from all the valid standard entities based on the semantic distance between the predicted feature vector and the standard entity embedding vector includes: For each of the target candidate entities, the cosine similarity value between its corresponding predicted feature vector and each of the standard entity embedding vectors is calculated. The valid standard entities are sorted in descending order of their cosine similarity scores. The valid standard entity that ranks first in the sorting is identified as the target standard entity that matches the target candidate entity.

[0010] According to a method for identifying medical claim materials provided by the present invention, before determining the first-ranked valid standard entity as the target standard entity matching the target candidate entity, the method further includes: Determine whether the cosine similarity value corresponding to the first-ranked valid standard entity is greater than a preset calibration threshold; If so, then the step of determining the first-ranked valid standard entity as the target standard entity that matches the target candidate entity is performed.

[0011] According to the medical claim material identification method provided by the present invention, the cross-modal network model includes a visual encoding branch and a text encoding branch set in parallel; The step of inputting the medical claim image into a cross-modal network model to obtain at least one candidate entity output by the cross-modal network model and a predicted feature vector corresponding to each candidate entity includes: The visual feature vector is obtained by extracting visual features from the medical claim image using the visual coding branch; the visual layout feature vector represents the table structure information and / or seal position information in the medical claim image. The text encoding branch is used to perform optical character recognition and semantic encoding on the medical claim image to obtain a text semantic feature vector. The visual layout feature vector and the text semantic feature vector are weighted and fused to generate the predicted feature vector.

[0012] According to a method for identifying medical claim materials provided by the present invention, the step of weighted fusion of the visual layout feature vector and the text semantic feature vector includes: Based on the clarity index of the medical claim image and the confidence index of the text semantic feature vector, calculate the first fusion weight of the visual layout feature vector and the second fusion weight of the text semantic feature vector. Based on the first fusion weight and the second fusion weight, the visual layout feature vector and the text semantic feature vector are weighted and concatenated to obtain the fused prediction feature vector.

[0013] According to the medical claim material identification method provided by the present invention, the cross-modal network model is obtained by iterative training through the following steps until a preset cutoff condition is reached: Obtain medical image samples containing labeled element tags, and obtain positive sample standard entity vectors corresponding to the element tags from the claims element knowledge graph; The medical image samples are input into the initial network model to be trained, and the initial network model is used to extract the predicted feature vector samples. A loss function is constructed based on the predicted feature vector samples and the positive sample standard entity vectors, and the parameters of the initial network model are updated using the loss function.

[0014] According to a method for identifying medical claim materials provided by the present invention, the step of constructing a loss function based on the predicted feature vector samples and the positive sample standard entity vectors includes: Based on the difference between the prediction result corresponding to the predicted feature vector sample and the element label, calculate the basic loss term for entity recognition; Calculate the cosine similarity between the predicted feature vector sample and the positive sample standard entity vector; A graph calibration loss term is constructed based on the cosine similarity, and the graph calibration loss term is configured to minimize the vector space distance between the predicted feature vector sample and the positive sample standard entity vector during training. The loss function is obtained by weighted summation of the map calibration loss term and the entity recognition basic loss term.

[0015] According to a method for identifying medical claim materials provided by the present invention, the step of extracting predicted feature vector samples using the initial network model includes: Visual layout feature vector samples of the medical image samples are extracted using the visual encoding branch of the initial network model; The text semantic feature vector samples of the medical image samples are extracted using the text encoding branch of the initial network model; Accordingly, the loss function also includes a cross-modal alignment loss term, which is configured to maximize the mutual information between the visual layout feature vector sample and the text semantic feature vector sample corresponding to the same medical image sample.

[0016] According to a method for identifying medical claim materials provided by the present invention, the pre-set claim element knowledge graph is dynamically updated through the following steps: Use targeted web crawlers to collect unstructured text data from external medical data sources; The knowledge quality score of the unstructured text data is calculated based on semantic similarity, source authority, and publication timeliness. If the knowledge quality score is greater than the preset entry threshold, then entity triples are extracted from the unstructured text data and updated in the claims element knowledge graph; The entity triple includes a head entity, a tail entity, and an edge defining the relationship between the two, and the head entity or the tail entity is configured with an effective time attribute and an applicable region attribute.

[0017] According to a method for identifying medical claim materials provided by the present invention, the dynamic updating step further includes: If the knowledge quality score remains below the entry threshold for a preset period, the webpage structure code of the external medical data source is obtained. Input the webpage structure code into a preset large language model and obtain the optimized crawling rules output by the large language model; Update the configuration of the targeted crawler using the optimized crawling rules.

[0018] According to a method for identifying medical claim materials provided by the present invention, the generation of structured claim element identification results includes: The corrected target standard entities are input into the relation classification layer to identify the logical relationships between the target standard entities. The target standard entity and its logical relationships are assembled into an entity triple containing a head entity, a relation edge, and a tail entity, which serves as the structured claims element identification result.

[0019] The present invention also provides a medical claims material identification device, comprising: An information collection unit is used to determine the contextual information associated with the medical claim image to be identified; An entity recognition unit is used to input the medical claim image into a cross-modal network model to obtain at least one candidate entity output by the cross-modal network model and a predicted feature vector corresponding to each candidate entity. The entity filtering unit filters out all valid standard entities from the preset claims element knowledge graph based on the context information, and obtains the standard entity embedding vector for each valid standard entity. An entity anchoring unit is configured to determine, based on the semantic distance between the predicted feature vector and the standard entity embedding vector, a target standard entity that matches each target candidate entity from all the valid standard entities; the target candidate entity is any one of all the candidate entities. The identification output unit is used to modify the target candidate entity that matches each target standard entity using the information of each target standard entity, and generate a structured claims element identification result.

[0020] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the medical claim material identification method as described above.

[0021] The medical claims material identification method and its storage medium provided by this invention ensure the business compliance of the scope to be calibrated by dynamically constraining the search space of the knowledge graph through spatiotemporal context, and use vector semantic distance to force the model output to move closer to the standard entity, effectively correcting noise and illusion in cross-modal recognition, and significantly improving the accuracy and usability of medical claims element identification. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a schematic diagram of the overall process of the medical claim material identification method provided by the present invention.

[0024] Figure 2 This is a flowchart illustrating the process of filtering valid standard entities from a knowledge graph of claims elements, as provided by this invention.

[0025] Figure 3 This is a schematic diagram of the process for determining target standard entities based on semantic distance, provided in an embodiment of the present invention.

[0026] Figure 4 This is a schematic diagram of the process for obtaining candidate entities and predicting feature vectors provided in an embodiment of the present invention.

[0027] Figure 5 This is a schematic diagram of the calculation process of the fused feature vector provided in the embodiment of the present invention.

[0028] Figure 6 This is a schematic diagram of the pre-training process of the cross-modal network model provided in an embodiment of the present invention.

[0029] Figure 7 This is a schematic diagram of the process of constructing the loss function in the pre-training of a cross-modal network model provided in an embodiment of the present invention.

[0030] Figure 8 This is a schematic diagram illustrating the process of dynamically updating the knowledge graph of claims elements provided in an embodiment of the present invention.

[0031] Figure 9 This is a flowchart illustrating the optimized targeted crawler configuration provided in an embodiment of the present invention.

[0032] Figure 10 This is a schematic diagram of the process for generating structured claims element identification results provided in an embodiment of the present invention.

[0033] Figure 11 This is a schematic diagram of the medical claim material identification device provided in an embodiment of the present invention.

[0034] Figure 12 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0036] It should be noted that, in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Those skilled in the art will understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0037] The terms "first," "second," etc., used in this invention are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0038] All actions involving the acquisition of medical claim images in this invention are carried out in compliance with the relevant data protection laws and regulations of the country where the claim is made, and with the authorization of the owner.

[0039] By analyzing existing technologies that combine OCR with NER based on a pipeline architecture for extracting medical document information, the following main shortcomings can be identified: (1) Poor accuracy of element extraction and lack of coordination between modalities.

[0040] Existing technologies have a gap between OCR and NER, relying solely on text information and failing to utilize contextual information such as visual layout, table boundaries, and seal positions in medical claim images for error correction and enhancement. They also exhibit extremely low robustness to blurry, handwritten, or formatted materials, resulting in low accuracy in feature extraction.

[0041] (2) Lack of knowledge calibration and difficulty in ensuring the timeliness of knowledge. Existing NER models lack domain knowledge constraints and cannot determine whether extracted entities (such as disease or drug names) are within a valid medical knowledge system. Furthermore, for time-sensitive elements such as the date of medical visit, it is impossible to verify whether their validity is consistent with the effective time of knowledge in the claims element knowledge graph. (3) Extracting isolated elements is not enough for direct claims settlement. Most existing technologies output an isolated list of claims elements, such as diagnosis name, drug list, total cost, etc., lacking logical connections between the various claims elements. This will cause the downstream claims settlement system to still need to consume a lot of computing power to match and judge the diagnosis and treatment behavior, and cannot directly enable in-depth claims settlement.

[0042] (4) Insufficient regional adaptability. Existing models are highly dependent on the document formats of specific regions or hospitals for training. Once the province or hospital is changed, the accuracy of the model will drop sharply due to changes in medical terminology, expense categories or table layout, requiring a large amount of re-labeling and retraining. Direct reuse is difficult due to many knowledge updates and conflicts, resulting in a 2-3 times increase in deployment costs.

[0043] In view of this, the present invention provides a method for identifying medical claim materials based on cross-modal fusion and knowledge graph calibration, and its storage medium. The following will describe the method in conjunction with... Figures 1-11 A detailed introduction will be provided.

[0044] The medical claim material identification method provided in this invention can be executed by a computing device equipped with computing resources such as a processor and memory, such as a server, personal computer, mobile terminal, or cloud computing platform. In all subsequent embodiments, the method will be described using an application to a processing system deployed on a server as an example.

[0045] Figure 1 This is a schematic diagram of the overall process of the medical claim material identification method provided by the present invention, as shown below. Figure 1 As shown, including but not limited to the following steps: Step 11: Determine the contextual information associated with the medical claim image to be identified.

[0046] In this embodiment, the medical claim image to be identified can be a digital image file of medical documents uploaded by the user during the insurance claim process through methods such as taking photos or scanning, such as photos or scans of outpatient medical records, invoices, discharge summaries, etc. issued by the hospital. The file format of the medical claim image can be various mainstream image or document formats such as JPEG, PNG, and PDF.

[0047] Contextual information refers to the associated information that can provide contextual constraints for subsequent recognition and calibration. In this embodiment, the contextual information can be determined by preprocessing and analyzing the medical claim image to be recognized itself. For example, optical character recognition technology can be used to identify fixed areas of the medical claim image, such as the header or signature area of ​​the document, and automatically parse out information such as the location of treatment and the date of issuance as the contextual information.

[0048] Step 12: Input the medical claim image to be identified into the cross-modal network model to obtain at least one candidate entity output by the cross-modal network model and the predicted feature vector corresponding to each candidate entity.

[0049] In this embodiment, the cross-modal network model can be a pre-trained deep learning model. Optionally, it is designed to simultaneously understand visual layout information in the image, such as the row and column structure of a table, the position of a seal, etc., as well as semantic content information.

[0050] The cross-modal network model receives the medical claim image to be identified as input and performs end-to-end feature extraction and analysis on it. The output candidate entity is a preliminary identification result of the key claim elements that may exist in the medical claim image, and its form can be a text string.

[0051] Considering the potential for poor image quality in medical claim images or the inherent limitations of cross-modal network models, the resulting candidate entities may contain errors. For example, for the clearly visible text "Amoxicillin" in a medical claim image, the cross-modal network model can generally accurately output the candidate entity "Amoxicillin." However, for the slightly blurred text "Cephalexin" in a medical claim image, the cross-modal network model may incorrectly output a similar-looking but non-existent word, such as "Cephalexin," which would be a candidate entity in this case.

[0052] The cross-modal network model identifies all possible claim elements in the medical claim image, and therefore typically outputs at least one candidate entity. The predicted feature vector output along with each candidate entity is a mathematical representation of the candidate entity in a specific high-dimensional semantic space.

[0053] The predicted feature vectors are generated by a cross-modal network model and contain rich contextual semantic information of the candidate entities and their visual layout information in the original medical claims image. Each candidate entity has a one-to-one corresponding predicted feature vector.

[0054] Step 13: Based on the context information, filter out all valid standard entities from the preset claims element knowledge graph, and obtain the standard entity embedding vector of each valid standard entity.

[0055] A pre-built claims knowledge graph can be a structured, authoritative knowledge base that stores a vast amount of standard claims elements and their interrelationships. These standard claims elements can include drug names, treatment items, disease diagnoses, medical materials, etc.

[0056] It should be noted that specific attributes can be configured for each entity in the claims element knowledge graph. These attributes can correspond to the context information obtained in step 11 above, such as the effective time attribute and the applicable geographical area attribute.

[0057] In this embodiment, the method for filtering out all valid standard entities can be as follows: the processing system uses the context information determined in step 11 as a query condition to perform a dynamic filter on the entire claims element knowledge graph. For example, if the context information indicates that the location of the visit is region XX and the visit time is December 2025, the processing system will filter out all claims elements that are valid in region XX at that time from the claims element knowledge graph.

[0058] The selected valid standard entities constitute the whitelist for the current scenario. Subsequently, the processing system retrieves the pre-stored, authoritative vector representations of these valid standard entities from the claims element knowledge graph, namely, the standard entity embedding vectors.

[0059] Step 14: Based on the semantic distance between the predicted feature vector and the standard entity embedding vector, determine the target standard entity that matches each target candidate entity from all the valid standard entities.

[0060] Semantic distance measures the similarity between a predicted feature vector and a standard entity embedding vector in the semantic space. The smaller the semantic distance, the closer the entities represented by the two vectors are in meaning. The semantic distance can be calculated in various ways, such as by calculating the cosine similarity or Euclidean distance between the two vectors.

[0061] In this embodiment, the process of determining the target standard entity that matches each target candidate entity can be as follows: The processing system traverses all candidate entities, and for each target candidate entity, it calculates the semantic distance between its corresponding predicted feature vector and the embedding vectors of all standard entities in the set of valid standard entities selected in step 13. Among all the calculated semantic distances, the valid standard entity with the smallest semantic distance can be selected and determined as the target standard entity that best matches the current target candidate entity.

[0062] For example, for the predicted feature vector of the candidate entity "cephaladine", the processing system calculates the semantic distance between it and the standard entity embedding vectors of all valid standard entities in the set of valid standard entities, such as "cephaladine" and "ibuprofen". It finds that the distance between it and the standard entity embedding vector of "cephaladine" is the smallest. Therefore, "cephaladine" will be identified as the target standard entity of "cephaladine".

[0063] Step 15: Use the information of each target standard entity to correct the target candidate entity that matches it, and generate a structured claims element identification result.

[0064] The correction method adopted for each target standard entity corresponding to the target candidate entity can be information replacement. Specifically, the processing system uses the standard name of the target standard entity determined in the previous step (e.g., "cefadroxil") to replace or correct the matching, possibly incorrect, candidate entity (e.g., "cefadroxil").

[0065] Ultimately, all corrected and confirmed claim elements will be integrated into a structured claim element identification result and output. The structured claim element identification result is a machine-readable, standardized data format, such as JSON or XML. It is not just a corrected list of text, but may also contain information such as the type and affiliation of the elements. It can be directly called by downstream claims processing rule engines or artificial intelligence agents to achieve full-process automation.

[0066] The medical claims material identification method provided by this invention ensures the business compliance of the scope to be calibrated by dynamically constraining the search space of the knowledge graph through spatiotemporal context, and uses vector semantic distance to force the model output to move closer to the standard entity, effectively correcting noise and illusions in cross-modal recognition, and significantly improving the accuracy and usability of medical claims element identification.

[0067] As an optional embodiment, the present invention further provides specific implementation methods for determining contextual information associated with the medical claim image to be identified, with two optional implementation paths, specifically including: One real-time approach is based on automatic parsing of image content, that is, parsing the consultation time and consultation location information from the text content of the medical claim image to be identified.

[0068] Specifically, the processing system first uses optical character recognition technology to scan the entire image or a specific area of ​​the medical claim image to be identified. By converting the pixel information in the image into computer-editable text data, the text content related to the medical claim image can be obtained.

[0069] Subsequently, the processing system can use preset semantic analysis rules or regular expressions to perform structured parsing of the text content.

[0070] For example, for information about the time of medical visits, look for fields in the search text that contain keywords such as "date", "time", or "day of medical visit", or match date format strings like "YYYY-MM-DD" to extract the specific date of medical visit, such as "March 15, 2025".

[0071] For example, regarding the location information for medical treatment, one can search for fields containing keywords such as "hospital," "clinic," and "address" in the text content and then match them using a pre-built administrative division dictionary or hospital name database. For instance, if the text content contains "XXX City YYY Hospital," entity linking technology can be used to identify the administrative region code or name of the hospital as "XXX City," thus confirming it as the location information for medical treatment.

[0072] The automatic parsing of context information provided in this embodiment can minimize manual data entry costs and automate the claims process.

[0073] Another real-time approach is based on proactive reception through user interaction, which involves proactively receiving information such as the time and location of medical treatment input by the user in relation to medical claim images.

[0074] In certain scenarios, such as when medical claim images are severely blurred, creased, or obscured, causing the automatic parsing in Path 1 above to fail to obtain results with sufficiently high confidence, or as a mandatory verification method, the processing system will initiate a user interaction process.

[0075] Specifically, the processing system will display a preview of the medical claim image on the front-end interactive interface, such as the insurance company's claims app or mini-program page, and simultaneously provide input controls related to the appointment time and city / region. Examples include a calendar selector and a province / city / district cascading selector. Users manually select or fill in the corresponding information based on their actual medical history. After receiving the data submitted by the user through the input controls, the system will directly identify it as the appointment time and location information.

[0076] The context information determination mechanism combining automatic image text parsing with manual user input provided in this invention balances the automation efficiency of the claims process with robustness in extreme scenarios. It ensures that regardless of the quality of the medical claims image to be identified, the processing system can obtain accurate information on the time and location of the visit. This precise spatiotemporal attribute data constitutes a key index for the subsequent dynamic slicing and filtering of the pre-set claims element knowledge graph, guaranteeing the effective execution of spatiotemporal constraint logic from the source and avoiding errors in the graph matching range due to missing spatiotemporal information.

[0077] After obtaining accurate contextual information such as the time and location of the visit, the processing system can use these spatiotemporal coordinates to perform precise segmentation filtering on the knowledge graph of claims elements, in order to eliminate interfering data that does not match the current medical scenario.

[0078] Figure 2 This is a flowchart illustrating the process of filtering valid standard entities from a knowledge graph of claims elements, as provided by this invention. Figure 2 As shown, the present invention further provides an implementation method for filtering all valid standard entities from a pre-set claims element knowledge graph based on the aforementioned context information, which mainly includes, but is not limited to, the following steps: Step 131: Traverse the entity nodes in the knowledge graph of the claims elements.

[0079] First, the underlying database of the claims element knowledge graph can be accessed, and the traversal process can be initiated. To ensure that the stored knowledge can accurately describe the differences in medical insurance policies under different time periods and administrative regions, the pre-built claims element knowledge graph incorporates spatiotemporal attribute fields during the schema design phase.

[0080] Specifically, for each entity node in the claims element knowledge graph, such as a certain drug, treatment item, or medical material, an effective time attribute and an applicable geographical attribute are pre-configured.

[0081] The effective time attribute typically includes key fields such as the effective date and the expiration date, which are used to define the legal existence window of the entity node in the official medical insurance catalog or hospital fee standards.

[0082] The applicable geographical attributes typically include administrative division codes or hospital identification fields, which define the geographical scope or specific medical institution scope to which the entity node applies.

[0083] Step 132: Determine whether the time range defined by the effective time attribute includes the consultation time information in the context information, and obtain the first determination result.

[0084] Step 133: Determine whether the geographical scope defined by the applicable geographical attribute includes the medical treatment geographical information in the context information, and obtain a second determination result.

[0085] For each traversed entity node, the processing system reads its effective time attribute and compares it with the consultation time information determined in step 11. Simultaneously or sequentially, it reads the applicable region attribute of the entity node and compares it with the consultation region information determined in step 11.

[0086] Specifically, the processing system checks whether the date indicated by the appointment time information is later than or equal to the effective date and earlier than or equal to the expiration date.

[0087] If the above time interval conditions are met, it means that the entity node is in a valid state at the time of the consultation, and the processing system marks the first judgment result as yes; otherwise, if the consultation time information exceeds the time range, for example, if the patient uses an old version of the medical service code that was abolished in 2023, it means that the entity has expired or has not yet become valid, and the processing system marks the first judgment result as no.

[0088] Furthermore, the process of comparing the location information for medical treatment supports the hierarchical inclusion logic of administrative divisions. For example, if the location information for medical treatment is XX City YY District, and the applicable location attribute of a certain entity node is configured as a code representing XX City (such as 110000) or a code representing the nationwide scope, then the geographical range includes the location information for medical treatment, and the processing system marks the second judgment result as yes.

[0089] Conversely, if a certain entity node is a supplementary drug that only exists in the KK City medical insurance catalog, then its geographical scope does not include the current medical treatment location information. The processing system will mark the second judgment result as no. This step effectively avoids misidentification caused by differences in medical insurance reimbursement in different places.

[0090] Step 134: Entity nodes whose first and second judgment results are both yes are determined as the valid standard entities.

[0091] The processing system performs a logical AND operation, meaning that an entity node is considered a valid standard entity only if both its first and second judgment results are true, indicating that the entity is valid in both the time and spatial dimensions. The set of all nodes identified as valid standard entities constitutes the specific valid standard entity candidate pool for the current claim case.

[0092] It should be noted that in the above embodiments, the spatio-temporal constraints of the entity nodes are specifically implemented as the effective time attribute and the applicable region attribute, which is only a common implementation method of the present invention. In actual applications, the storage and organization forms of the spatio-temporal information in the pre-set claim element knowledge graph are not limited to this. For example, the spatio-temporal constraint information can also be reflected by the relationship edges between entity nodes, such as attaching time attributes to the applicable edges, or implemented by other data structures such as version number mapping tables and multi-dimensional spatio-temporal index trees. In other words, as long as it is a configuration method that can perform validity screening on the entities in the claim element knowledge graph based on the visit time and visit region information, it should be included in the protection scope of the present invention.

[0093] After screening out the effective standard entity candidate pool in the current scenario, the processing system needs to solve the core disambiguation and error correction problems, that is, how to find the most correct target standard entity from the effective standard entity candidate pool.

[0094] Figure 3 It is a schematic flowchart of determining the target standard entity according to the semantic distance provided by an embodiment of the present invention. As Figure 3 shown, this embodiment uses the semantic distance in the high-dimensional vector space as a metric standard to implement the screening of the target standard entity, specifically including but not limited to the following steps: Step 141, for each of the target candidate entities, calculate the cosine similarity value between its corresponding predicted feature vector and each of the standard entity embedding vectors.

[0095] First, lock a target candidate entity that needs to be processed currently, such as the text "Cefradine" recognized by OCR, and read the predicted feature vector generated by the cross-modal network model and corresponding to this target candidate entity.

[0096] At the same time, traverse all the effective standard entities determined in step 13. For each effective standard entity, the processing system reads its pre-stored standard entity embedding vector.

[0097] Subsequently, perform vector operations, and calculate the cosine similarity value between the predicted feature vector and the standard entity embedding vector currently traversed. The obtained cosine similarity value can accurately reflect the consistency between the prediction result and the standard knowledge at the semantic level. The calculation formula is usually expressed as the dot product of two vectors divided by the product of their norms. The closer the value is to 1, the closer they are in the semantic space.

[0098] Step 142, sort the effective standard entities in descending order according to the cosine similarity value.

[0099] For the current target candidate entity, similarity scores between it and each effective standard entity will be obtained. Subsequently, all valid standard entities are arranged in descending order of cosine similarity scores. The valid standard entity at the top of the list represents the entity whose vector representation is closest to the feature vector predicted by the model, meaning that it is most likely the real object referred to by the candidate entity semantically.

[0100] Step 143: The valid standard entity that ranks first in the sorting is determined as the target standard entity that matches the target candidate entity.

[0101] The first element in the sorted list is selected directly, that is, the valid standard entity that is ranked first, and it is marked as the target standard entity that matches the current target candidate entity.

[0102] For example, regarding the erroneous candidate entity "cephalosporin," although the set of valid standard entities after spatiotemporal filtering may contain multiple drugs such as "cephalosporin," "amoxicillin," and "cephalosporin," the feature vector predicted by the cross-modal network model retains the semantic features of "cephalosporin antibiotics" and the glyphic features of the characters "cephalosporin" in the claims element knowledge graph (through cross-modal fusion). Therefore, its vector cosine similarity with the valid standard entity "cephalosporin" is usually the highest (e.g., 0.95), while its similarity with "amoxicillin" is lower (e.g., 0.2). Thus, the processing system accurately identifies "cephalosporin" as the target standard entity.

[0103] It should be noted that although this embodiment uses cosine similarity as an example, other metrics such as Euclidean distance and Manhattan distance can also be used to calculate semantic distance in other embodiments. This also falls within the scope of protection of this invention. When a distance metric is used, the sorting logic in step 142 will be adjusted accordingly.

[0104] This invention utilizes semantic distance in vector space as a metric to achieve a mapping from fuzzy prediction to precise knowledge. This matching mechanism based on embedding vectors is more robust than traditional matching based on text edit distance. It can not only correct simple typos, but also solve the problem of synonym or abbreviation matching by using semantic association, ensuring that the final target standard entity is the best result verified by deep semantic analysis, effectively eliminating model illusion.

[0105] Based on the above embodiments, as an optional embodiment, in scenarios where selecting the most similar valid standard entity as the final target standard entity solely based on the ranking result may be risky—for example, even the entity with the highest similarity may have a very low similarity value—this embodiment introduces a confidence judgment step. Specifically, before determining the valid standard entity ranked first as the target standard entity matching the target candidate entity, the following steps are also included: Determine whether the cosine similarity value corresponding to the first-ranked valid standard entity is greater than a preset calibration threshold; If so, then the step of determining the first-ranked valid standard entity as the target standard entity that matches the target candidate entity is performed.

[0106] The preset calibration threshold can be a pre-defined value used to measure the confidence of the match, such as 0.8 or 0.9. The processing system considers the match reliable only when the candidate entity predicted by the cross-modal network model has a sufficiently high semantic similarity to a valid standard entity in the medical claim image.

[0107] If the processing system determines that the calculated highest cosine similarity value is greater than the calibration threshold, it indicates that there is a high semantic correlation between the candidate entity and the first-ranked valid standard entity, and it can be confirmed that the valid standard entity is the correct matching object. Only then will the processing system continue with the subsequent determination steps.

[0108] Correspondingly, if the judgment result is negative, meaning the highest cosine similarity value has not reached the calibration threshold, the processing system will consider that the current candidate entity has not found a sufficiently credible match in the claims element knowledge graph. In this case, the processing system can adopt various processing strategies, such as marking the candidate entity as "unrecognizable" or "awaiting manual review," without performing erroneous forced correction. This effectively avoids the situation where a completely unrelated entity, but which happens to be the closest, is mistakenly designated as the target standard entity.

[0109] Continuing with the previous example, assuming the preset calibration threshold is set to 0.85, the following two different scenarios may exist: In scenario one, the cosine similarity score between the candidate entity "cephaladin" and the valid standard entity "cephaladin" is 0.95. Since 0.95 is greater than the calibration threshold of 0.85, the processing system confirms that this match is highly reliable and therefore identifies "cephaladin" as the target standard entity.

[0110] In scenario two, if a candidate entity in a medical claim image to be identified is identified as "Acyclovir" due to severe smudges by the cross-modal network model, the calculated cosine similarity between this candidate entity and "Acyclovir," which has the highest similarity in the set of valid standard entities, is only 0.7. However, since 0.7 is less than the calibration threshold of 0.85, the processing system will determine that the confidence level of this match is insufficient, thus suspending the correction operation for this candidate entity and avoiding possible erroneous associations.

[0111] This invention adds a calibration threshold judgment step, setting a confidence barrier for the reliability of semantic matching. This effectively filters out low-quality or low-confidence matching results, preventing the processing system from making forced error corrections without sufficient certainty. It ensures that entity calibration is only performed when semantics are highly consistent, thereby significantly improving the overall accuracy and reliability of the final generated structured claims element recognition results.

[0112] To effectively process the complex graphic and textual information in medical claim documents, the cross-modal network model provided in this invention innovatively adopts a dual-stream architecture with parallel visual encoding branches and text encoding branches.

[0113] Figure 4 This is a schematic diagram of the process for obtaining candidate entities and predicting feature vectors provided in an embodiment of the present invention, as shown below. Figure 4 As shown, the medical claim image is input into a cross-modal network model to obtain at least one candidate entity output by the cross-modal network model and a predicted feature vector corresponding to each candidate entity. This may include, but is not limited to: Step 121: Visual features are extracted from the medical claim image using the visual encoding branch to obtain a visual layout feature vector. The visual layout feature vector represents the table structure information and / or seal position information in the medical claim image.

[0114] First, the medical claim image is input into the visual encoding branch. This branch can be implemented using a deep learning model such as the Vision Transformer (ViT). Its core function is to extract non-textual, purely visual structured information from the medical claim image. By analyzing the image pixels, it is possible to capture the table structure information in the invoice or list, such as the alignment of rows and columns, cell boundaries, and the location information of the stamp on the document, such as the coordinate range and shape of the red stamp in the image. This structured information constitutes the visual layout feature vector.

[0115] Step 122: Use the text encoding branch to perform optical character recognition and semantic encoding on the medical claim image to obtain a text semantic feature vector.

[0116] Furthermore, the processing system will input the medical claim image into the text encoding branch, which will perform an optical character recognition step to convert the text pixels in the image into a preliminary text sequence.

[0117] The resulting text sequence is then input into a powerful semantic encoding model, such as BERT, a bidirectional encoder based on Transformer. The semantic encoding model extracts the deep semantic meaning of each word or character by deeply bidirectionally analyzing the context of the text, generating a text semantic feature vector to represent the medical or financial meaning of the text content itself.

[0118] Step 123: The visual layout feature vector and the text semantic feature vector are weighted and fused to generate the predicted feature vector.

[0119] After obtaining the visual layout feature vector and the text semantic feature vector respectively, the processing system performs a weighted fusion step to generate a highly robust prediction feature vector that simultaneously encodes both types of information.

[0120] Specifically, the processing system dynamically assigns different importance weights to the two feature vectors according to a preset strategy. For example, when the medical claim image is clear, the weight of the text semantic features may be higher, while when the medical claim image is blurry and the OCR result is unreliable, the weight of the visual layout features will be increased accordingly. In this way, even if the text information is incorrect, the processing system can still use visual layout information to assist in judgment and correction.

[0121] Since the predicted feature vector contains both visual and semantic information, the cross-modal network model can make more accurate recognitions. Therefore, by inputting the fused predicted feature vector into the downstream part of the cross-modal network model, such as the sequence labeling layer, the accurate candidate entities output by it can be obtained.

[0122] This invention, through the design of a dual-stream cross-modal architecture that combines vision and text, breaks through the information bottleneck caused by the traditional serial processing mode of OCR followed by NER, thereby greatly enhancing the robustness and accuracy of recognition of fuzzy text, handwritten text, complex tables, and non-standardized documents.

[0123] Figure 5 This is a schematic diagram of the calculation process of the fused feature vector provided in an embodiment of the present invention, as shown below. Figure 5 As shown, the present invention further provides an implementation method for weighted fusion of the visual layout feature vector and the text semantic feature vector, specifically through a cross-attention mechanism, including but not limited to the following steps: Step 1231: Based on the clarity index of the medical claim image and the confidence index of the text semantic feature vector, calculate the first fusion weight of the visual layout feature vector and the second fusion weight of the text semantic feature vector.

[0124] Step 1232: Based on the first fusion weight and the second fusion weight, the visual layout feature vector and the text semantic feature vector are weighted and concatenated to obtain the fused prediction feature vector.

[0125] In this embodiment, the processing system evaluates the quality of the input information before performing feature fusion to determine the relative reliability of the two modal features.

[0126] The clarity index is a quantitative assessment of the overall or local quality of a medical claim image. It can be obtained by calculating the image gradient, the variance of the Laplacian operator, or by using a specialized fuzziness assessment algorithm. The higher the clarity index, the better the image quality of the medical claim image and the more reliable the OCR recognition result.

[0127] The confidence index is typically output by the OCR engine in the text encoding branch when recognizing each character or word, representing its level of confidence in the current recognition result. A higher confidence index indicates that the recognized text content itself is more reliable.

[0128] The processing system can input these two metrics into a weight calculation module, such as a small neural network or a pre-defined mapping function, to calculate the first fusion weight assigned to the visual layout feature vector and the second fusion weight assigned to the text semantic feature vector, respectively. As can be seen from the above process, the first and second fusion weights are dynamically changing, and their sum can be 1.

[0129] The above operations are performed mainly because when the sharpness index or confidence index is low, it indicates that the image is blurry or the OCR result is uncertain. The processing system will automatically adjust the first fusion weight and the second fusion weight to more reasonably rely on visual typography information or trust text semantic information.

[0130] After determining the first and second fusion weights, the processing system performs a weighted concatenation operation, which can be achieved in various ways, such as weighted summation, gating mechanisms, or attention fusion, which will not be elaborated here.

[0131] Through weighted fusion, the final generated predicted feature vector can intelligently balance and integrate information from the two modalities, forming a highly robust and information-rich feature representation.

[0132] This invention, through the introduction of a mechanism that dynamically calculates fusion weights based on input quality, achieves adaptive fusion of cross-modal information, endowing the processing system with the ability to select the best input when faced with inputs of varying quality. When text information is damaged due to OCR recognition errors, it can automatically reduce its reliance on the text and instead utilize stable visual layout features as supplementary clues for inference, effectively resisting interference from OCR recognition errors and layout changes.

[0133] A high-performance cross-modal network model is the foundation for achieving the technical effects of this invention. This invention further provides an implementation method for pre-training a cross-modal network model.

[0134] Figure 6 This is a schematic diagram of the pre-training process of the cross-modal network model provided in an embodiment of the present invention, as shown below. Figure 6 As shown, the cross-modal network model is obtained through iterative training using the following steps until a preset cutoff condition is met: Step 21: Obtain medical image samples containing element labels, and obtain positive sample standard entity vectors corresponding to the element labels from the claims element knowledge graph.

[0135] Step 22: Input the medical image samples into the initial network model to be trained, and use the initial network model to extract the predicted feature vector samples.

[0136] Step 23: Construct a loss function based on the predicted feature vector samples and the positive sample standard entity vectors, and use the loss function to update the parameters of the initial network model.

[0137] Before pre-training begins, a large-scale training dataset is prepared, consisting of numerous real medical claim images with annotation information. The feature labels are standard answers that are manually or semi-automatically labeled, clearly indicating the key elements in each medical image sample, such as the text content of the drug "amoxicillin" and its location in the medical image sample.

[0138] The processing system queries the claims element knowledge graph to obtain the pre-stored embedding vector of the standard entity corresponding to each element label. This embedding vector is the positive sample standard entity vector. For example, for an element label labeled "Amoxicillin," the processing system will retrieve the standard entity embedding vector of "Amoxicillin" from the claims element knowledge graph. The significance of this step is to introduce external, authoritative knowledge graph information into the training data, providing knowledge anchors for subsequent model parameter optimization.

[0139] The initial network model to be trained refers to the state of the cross-modal network model before training begins. Its internal parameters can be randomly initialized or loaded with weights from a general pre-trained model (such as the general ViT and BERT models).

[0140] The processing system inputs the medical image samples prepared in step 21 into the initial network model to be trained in batches. The initial network model performs forward computation on the input medical image samples according to its internal two-stream parallel architecture, and outputs the corresponding predicted feature vector samples for the location of the corresponding feature labels.

[0141] In this embodiment, a loss function is constructed based on the predicted feature vector samples output by the initial network model and the obtained positive sample standard entity vectors. This loss function is designed to measure the difference or distance between the model's prediction and the standard answer.

[0142] Furthermore, the processing system can employ optimization algorithms such as gradient descent to backpropagate and update the parameters of the initial network model, such as the weights and biases of the neural network, using the calculated loss function. This update process will guide the model parameters to adjust in a direction that can reduce the loss value.

[0143] The entire training process is iterative, repeating steps 21 to 23 until the model performance on the validation set no longer improves or reaches a preset cutoff condition, such as completing a preset number of training rounds. After training, the resulting optimized initial network model is the cross-modal network model that can be used for actual recognition.

[0144] The cross-modal network model pre-training method provided in this invention not only uses the pre-set claims element knowledge graph as a query dictionary for post-processing, but also directly integrates it into the end-to-end training process as an embedding layer constraint of the deep learning model. By forcing the predicted feature vectors output by the model to approximate the standard entity vectors in the claims element knowledge graph in the semantic space, authoritative domain knowledge is solidified into the model parameters. This knowledge-injection training mechanism enables the final cross-modal network model to have the ability to perceive and align standard medical terms.

[0145] To achieve dual optimization of cross-modal network models in terms of traditional entity recognition capabilities and calibration capabilities for claims-related element knowledge graphs, this invention innovatively designs a hybrid loss function.

[0146] Figure 7 This is a flowchart illustrating the process of constructing the loss function during the pre-training of a cross-modal network model, as provided in an embodiment of the present invention. Figure 7 As shown, the step of constructing a loss function based on the predicted feature vector samples and the positive sample standard entity vectors includes: Step 231: Calculate the basic loss term for entity recognition based on the difference between the prediction result corresponding to the predicted feature vector sample and the element label.

[0147] Step 232: Calculate the cosine similarity between the predicted feature vector sample and the positive sample standard entity vector.

[0148] Step 233: Construct a graph calibration loss term based on the cosine similarity. The graph calibration loss term is configured to minimize the vector space distance between the predicted feature vector sample and the positive sample standard entity vector during training.

[0149] Step 234: The map calibration loss term and the entity recognition basic loss term are weighted and summed to obtain the loss function.

[0150] Step 231 is mainly used to calculate the traditional supervised learning loss, and the resulting entity recognition basic loss term is... Its role in training is to drive the model to learn basic text recognition and entity boundary segmentation capabilities, ensuring that it can extract text content from images as accurately as possible.

[0151] The processing system inputs the predicted feature vector samples into the classification layer or sequence labeling layer of the initial network model to obtain a specific prediction result (e.g., "cephalazoline"). Then, this prediction result is compared with the corresponding ground truth element label obtained in step 21 (e.g., "cephalazoline"). The difference between the two can be quantified using methods such as cross-entropy loss, and this quantified difference value constitutes the basic loss term for entity recognition. .

[0152] The purpose of step 232 is to construct the loss for the knowledge calibration component. This is done by calculating the cosine similarity between the predicted feature vector samples output from step 22 and the positive sample standard entity vectors obtained from the knowledge graph in step 21. The cosine similarity reflects the directional consistency of the two vectors in the semantic space; the closer the value is to 1, the closer the semantics. Therefore, the calculated cosine similarity can be used to construct the knowledge graph calibration loss term. L KG .

[0153] Furthermore, there are several ways to construct the map calibration loss term. One approach is to define the loss as 1 - cosine similarity. In this way, when the cosine similarity is close to 1, the map calibration loss term is close to 0; when the cosine similarity is low, the map calibration loss term... L KGThe distance is relatively large. By minimizing this graph calibration loss term during training, which is equivalent to maximizing the cosine similarity, the goal of minimizing the vector space distance between the predicted feature vector samples and the positive sample standard entity vectors is achieved.

[0154] Finally, the processing system performs a weighted sum of the loss terms from the two different objectives to form the final, complete hybrid loss function. Its formula can be expressed as: ; in, It is the final objective function value that the model needs to minimize in one training iteration; Identify the underlying loss term for the entity; For predicting feature vector samples; These are the standard entity vectors of positive samples; These are hyperparameters used for balancing. and exist Weights in Indicates calculation and The cosine similarity value between them; This represents the graph calibration loss term constructed based on the calculated cosine similarity values. L KG .

[0155] The embodiments of this invention employ a hybrid loss function that includes a basic loss term for entity recognition and a graph calibration loss term, thereby achieving dual guidance for the model training objective. This mechanism, which introduces embedding layer constraints during the training phase, enables the model to not only correct possible OCR errors and model illusions during the learning process, but also to actively learn the ability to align with the semantic space of authoritative knowledge.

[0156] Based on the above embodiments, as an optional embodiment, the step of extracting predicted feature vector samples using the initial network model mainly includes: Visual layout feature vector samples of the medical image samples are extracted using the visual encoding branch of the initial network model.

[0157] The text semantic feature vector samples of the medical image samples are extracted using the text encoding branch of the initial network model.

[0158] Accordingly, the loss function also includes a cross-modal alignment loss term, which is configured to maximize the mutual information between the visual layout feature vector sample and the text semantic feature vector sample corresponding to the same medical image sample.

[0159] During the forward propagation phase of pre-training the cross-modal network model, the processing system inputs a medical image sample in parallel into two main branches of the initial network model, including: extracting visual layout feature vector samples of the medical image sample using the visual encoding branch, and extracting textual semantic feature vector samples of the medical image sample using the text encoding branch.

[0160] Since vision and text are two different modalities, their encoded feature vectors are initially located in different, heterogeneous feature spaces, making direct and meaningful comparison impossible. To address this issue, this embodiment introduces a cross-modal alignment loss term in addition to the graph calibration loss term and entity recognition basic loss term defined in the aforementioned embodiments when constructing the loss function.

[0161] The determination of the cross-modal alignment loss term is typically based on a contrastive learning framework, such as the InfoNCE loss function. Specifically, for the same medical image sample, its extracted visual layout feature vector sample and textual semantic feature vector sample constitute a positive sample pair, meaning they should semantically point to the same concept. Meanwhile, the feature vector of the current medical image sample, whether it's a visual layout feature vector sample or a textual semantic feature vector sample, forms a negative sample pair with the feature vectors of other different medical image samples in the training batch.

[0162] The cross-modal alignment loss term, through its mathematical form, achieves the following effect during training: in a unified high-dimensional semantic space, it continuously narrows the distance between two vectors in a positive sample pair while simultaneously widening the distance between vectors in a negative sample pair. Through this optimization process of narrowing and widening, the initial network model is forced to learn a unified representation space in which visual layout features and textual semantic features from the same document can be aligned and mapped to each other.

[0163] Alternatively, the function expression for InfoNCE can be represented as: ; in, This is a temperature coefficient used to adjust the sharpness of the similarity score distribution; This is the text semantic feature vector corresponding to the medical image sample; Is with Visual layout feature vectors derived from the same medical image sample; This represents a set containing multiple vectors, including those from positive sample pairs. And multiple negative samples from other different medical image samples in the training batch; This represents the similarity calculation function; This represents the cross-modal alignment loss term.

[0164] Finally, the cross-modal alignment loss term obtained here will be weighted and summed together with the entity recognition basic loss term and the map calibration loss term defined in the previous embodiments to form a more complete multi-task final loss function, which will jointly guide the parameter update of the entire initial network model.

[0165] This invention addresses the fundamental problem of inconsistency between visual and textual heterogeneous feature spaces by introducing a cross-modal alignment loss term during model training. This significantly improves the robustness and accuracy of cross-modal network models when processing low-quality, blurry, or handwritten medical claim images.

[0166] The present invention further provides an implementation method for constructing a high-quality, dynamically updated knowledge graph of claims elements. Figure 8 This is a schematic diagram illustrating the process of dynamically updating the knowledge graph of claims elements provided in an embodiment of the present invention, such as... Figure 8 As shown, this embodiment achieves dynamic updating of the claims element knowledge graph through a closed-loop process that includes quality assessment and automatic extraction functions. The specific steps include: Step 31: Use targeted web crawlers to collect unstructured text data from external medical data sources.

[0167] Step 32: Calculate the knowledge quality score of the unstructured text data based on the semantic similarity dimension, the source authority dimension, and the publication timeliness dimension.

[0168] Step 33: If the knowledge quality score is greater than the preset entry threshold, then extract entity triples from the unstructured text data and update them in the claims element knowledge graph.

[0169] The processing system is equipped with a targeted web crawler module. This module can be a set of pre-defined software agents whose task is to periodically monitor and crawl information from multiple authoritative external medical data sources. These external medical data sources may include, but are not limited to: policy announcements published on national and local government websites, updated drug catalogs, fee adjustment notices published on the official websites of major hospitals, and publicly available information from other medical information service platforms. The targeted crawler collects unstructured text data, such as HTML webpage content, PDF announcements, or plain text documents.

[0170] To ensure the accuracy of the knowledge stored, the processing system conducts a rigorous quality assessment on each piece of unstructured text data before updating the claims element knowledge graph, such as calculating its Knowledge Quality Score (KQS).

[0171] Optionally, the calculated knowledge quality score mainly integrates dimensions such as semantic similarity, source authority, and publication timeliness.

[0172] The semantic similarity dimension refers to the semantic similarity calculated by the processing system using a deep language model (such as BERT) between the newly collected unstructured text data and the existing content in the claims element knowledge graph, in order to identify potential conflicting or redundant information.

[0173] The source authority dimension refers to the weight value assigned to different sources by the processing system based on a maintained list of data source authority weights. For example, the weight of government websites is higher than that of ordinary news websites.

[0174] The timeliness of publication refers to the publication date obtained by the processing system when parsing unstructured text data; the more recent the date, the higher the timeliness score.

[0175] Optionally, the formula for calculating the knowledge quality score (KQS) can be expressed as: ; in, This is the total knowledge quality score obtained after evaluating the newly crawled content; The semantic similarity score is calculated using models such as BERT to determine the degree of semantic similarity between new content and existing content in the claims element knowledge graph. The source authority score is a value assigned based on the authority level of the data source from which the information is published (such as government websites, industry portals, and general websites); the timeliness score is a score calculated based on the freshness of the information's publication date. , and These refer to the weights of semantic similarity score, source authority score, and publication timeliness score, respectively.

[0176] Furthermore, the processing system compares the calculated knowledge quality score with a preset entry threshold. Only when the knowledge quality score is greater than this threshold does the processing system consider the data to have sufficient value and credibility to be used to update the claims element knowledge graph.

[0177] For unstructured text data that passes quality verification, the processing system uses Large Language Model (LLM) technology, such as combining few-shot cues, to perform deep parsing and automatically extract structured entity triples from it.

[0178] It should be noted that the aforementioned entity triple includes a head entity, a tail entity, and an edge defining the relationship between the two. The head entity or the tail entity is configured with an effective time attribute and an applicable region attribute. During the extraction process, LLM not only identifies entities (such as drug A) and relationships (such as "for treatment"), but also parses the corresponding spatiotemporal metadata from the text context, such as the "Notice on Adjusting the 2025 Medical Insurance Drugs in XX City" in the announcement title, and directly appends it as the effective time attribute and applicable region attribute to newly generated or yet-to-be-updated entity nodes.

[0179] Finally, the processing system injects or updates these entity triples, which carry complete spatiotemporal attributes, into the claims element knowledge graph.

[0180] This invention provides an automated process for real-time updating of a claims element knowledge graph, addressing the technical pain points of traditional knowledge bases, such as high maintenance costs, untimely updates, and inconsistent information quality. Furthermore, the introduction of a knowledge quality score ensures that only high-quality, highly reliable information enters the claims element knowledge graph, guaranteeing the accuracy of the knowledge base from the source. The mechanism of automatically extracting and configuring spatiotemporal attributes using a large language model significantly reduces the cost of manually constructing knowledge and ensures that every entity in the graph possesses the basic data required for subsequent spatiotemporal filtering.

[0181] Based on the above embodiments, the present invention further provides an implementation method to address the problem of decreased crawling quality caused by redesign or structural changes of external medical data source websites.

[0182] Figure 9 This is a flowchart illustrating the optimized targeted crawler configuration provided in an embodiment of the present invention, such as... Figure 9 As shown, the dynamic update step further includes: Step 34: If the knowledge quality score remains below the entry threshold within a preset period, then obtain the webpage structure code of the external medical data source. Step 35: Input the webpage structure code into the preset large language model and obtain the optimized crawling rules output by the large language model; Step 36: Update the configuration of the targeted crawler using the optimized crawling rules.

[0183] In one optional embodiment, the processing system includes a self-monitoring mechanism that continuously tracks and records the knowledge quality score from a specific external medical data source. The processing system sets a preset tracking period, such as the last seven days or ten consecutive crawls.

[0184] If the processing system detects that the knowledge quality score from a specific data source is consistently lower than the inbound threshold within a tracking period, it will trigger a low-quality alert to inform that the website front-end structure of that specific data source may have changed, causing the original crawling rules (such as XPath paths) to become invalid and unable to accurately crawl valid information.

[0185] At this point, the processing system will automatically revisit the URL of the external medical data source and obtain its latest webpage structure code, which is usually the complete HTML document content.

[0186] Furthermore, the processing system takes the acquired webpage structure code as input and provides it to a pre-defined large language model. This large language model is specially tuned or configured with specific instruction prompts, enabling it to understand the webpage's DOM tree structure and generate crawler parsing rules. The processing system also provides the large language model with the type of information it expects to crawl as a target, such as finding announcement links whose titles contain drug catalogs. The large language model analyzes the input webpage structure code and automatically generates new, optimized crawling rules adapted to the current webpage structure based on the target. These crawling rules may include, but are not limited to, updated XPath paths, CSS selectors, or script logic for dynamic webpage crawling.

[0187] The system receives optimized crawling rules output by the large language model and uses them to automatically update the configuration file for that specific data source in the targeted crawler module. After the update is complete, the targeted crawler will use the new rules to collect data the next time it executes a task.

[0188] This invention provides a self-repairing method for crawler rules driven by a large language model, which realizes intelligent maintenance of the knowledge graph data source collection process and solves the pain point of traditional crawlers frequently failing due to website redesign and requiring manual intervention to modify the code.

[0189] Figure 10 This is a schematic diagram of the process for generating structured claims element recognition results provided in an embodiment of the present invention, such as... Figure 10 As shown, the generated structured claims element identification results include, but are not limited to: Step 151: Input the corrected target standard entities into the relation classification layer to identify the logical relationships between the target standard entities.

[0190] Step 152: Assemble the target standard entity and its logical relationships into the entity triplet containing the head entity, relation edge and tail entity, as the structured claim element identification result.

[0191] The processing system has used the knowledge graph to correct all candidate entities, resulting in an accurate and compliant list of target standard entities, such as "amoxicillin", "upper respiratory tract infection", and "20 yuan".

[0192] This set of target standard entities and their contextual information in the medical claim image can be input into a pre-trained relation classification layer. The relation classification layer can be an independent neural network model or a downstream module of a cross-modal network model, and its task is to determine the possible logical relationships between entity pairs.

[0193] For example, this relation classification layer can identify: There is a relationship between "amoxicillin" and "upper respiratory tract infection" as "used for treatment".

[0194] There is a relationship of "incurring a fee" between "amoxicillin" and "20 yuan".

[0195] There is a "belonging" relationship between "amoxicillin" and "Class A medical insurance".

[0196] After identifying the logical relationships, the processing system performs an assembly operation, turning the discrete elements and the identified logical relationships into structured entity triples.

[0197] The entity triple is a standard data representation consisting of <head entity, relation edge, tail entity>. Based on the example above, the processing system will generate the following triples: Amoxicillin is used to treat upper respiratory tract infections. <Amoxicillin, cost 20 yuan>; Amoxicillin is covered by Category A medical insurance. This set of entity triples constitutes the final, information-rich structured claim element identification result, which can be output in machine-readable formats such as JSON, XML, or RDF.

[0198] The method for generating structured claims element recognition results provided in this embodiment achieves a leap from isolated element recognition to structured knowledge construction. The output result is a structured triple containing complete logical relationships, which can be directly read and calculated by downstream claims rule engines or AI inference agents. For example, the claims engine can directly verify whether the relationship <drug A, used for treatment, diagnosis B> conforms to medical logic and medical insurance regulations.

[0199] Figure 11 This is a schematic diagram of the medical claim material identification device provided in an embodiment of the present invention, as shown below. Figure 11 As shown, it mainly includes, but is not limited to: Information collection unit 1 is used to determine the contextual information associated with the medical claim image to be identified; Entity recognition unit 2 is used to input the medical claim image into a cross-modal network model to obtain at least one candidate entity output by the cross-modal network model and a predicted feature vector corresponding to each candidate entity.

[0200] Entity filtering unit 3 filters out all valid standard entities from the preset claims element knowledge graph based on the context information, and obtains the standard entity embedding vector of each valid standard entity.

[0201] The entity anchoring unit 4 is used to determine the target standard entity that matches each target candidate entity from all the valid standard entities based on the semantic distance between the predicted feature vector and the standard entity embedding vector; the target candidate entity is any one of the candidate entities.

[0202] The identification output unit 5 is used to modify the target candidate entity that matches it by using the information of each target standard entity, and generate a structured claim element identification result.

[0203] It should be noted that the medical claim material identification device provided by the present invention can execute the medical claim material identification method described in any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0204] The medical claims material recognition device provided by this invention ensures the business compliance of the scope to be calibrated by dynamically constraining the search space of the knowledge graph through spatiotemporal context, and uses vector semantic distance to force the model output to move closer to the standard entity, effectively correcting noise and illusions in cross-modal recognition, and significantly improving the accuracy and usability of medical claims element recognition.

[0205] Figure 12 This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 12As shown, the electronic device may include: a processor 1210, a communications interface 1220, a memory 1230, and a communication bus 1240, wherein the processor 1210, the communications interface 1220, and the memory 1230 communicate with each other through the communication bus 1240. The processor 1210 can call logical instructions in the memory 1230 to execute a medical claim material recognition method. The method includes: determining contextual information associated with a medical claim image to be recognized; inputting the medical claim image into a cross-modal network model to obtain at least one candidate entity output by the cross-modal network model and a predicted feature vector corresponding to each candidate entity; filtering all valid standard entities from a preset claim element knowledge graph based on the contextual information and obtaining a standard entity embedding vector for each valid standard entity; determining a target standard entity matching each target candidate entity from all valid standard entities based on the semantic distance between the predicted feature vector and the standard entity embedding vector; the target candidate entity being any one of all the candidate entities; and modifying the matching target candidate entity using information from each target standard entity to generate a structured claim element recognition result.

[0206] Furthermore, the logical instructions in the aforementioned memory 1230 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0207] On the other hand, the present invention also provides a computer program product, the computer program product including a computer program stored on a non-transitory computer-readable storage medium, the computer program including program instructions, when the program instructions are executed by a computer, the computer is able to execute the medical claim material recognition method provided in the above embodiments, the method including: determining context information associated with a medical claim image to be identified; inputting the medical claim image into a cross-modal network model, obtaining at least one candidate entity output by the cross-modal network model and a predicted feature vector corresponding to each candidate entity; filtering all valid standard entities from a preset claim element knowledge graph based on the context information, and obtaining a standard entity embedding vector for each valid standard entity; determining a target standard entity matching each target candidate entity from all the valid standard entities according to the semantic distance between the predicted feature vector and the standard entity embedding vector; the target candidate entity is any one of all the candidate entities; and correcting the target candidate entity matching it using the information of each target standard entity respectively, generating a structured claim element recognition result.

[0208] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the medical claim material recognition method provided in the above embodiments. The method includes: determining contextual information associated with a medical claim image to be identified; inputting the medical claim image into a cross-modal network model to obtain at least one candidate entity output by the cross-modal network model and a predicted feature vector corresponding to each candidate entity; filtering all valid standard entities from a preset claim element knowledge graph based on the contextual information, and obtaining a standard entity embedding vector for each valid standard entity; determining a target standard entity matching each target candidate entity from all the valid standard entities based on the semantic distance between the predicted feature vector and the standard entity embedding vector; the target candidate entity being any one of all the candidate entities; and modifying the matching target candidate entity using information from each target standard entity to generate a structured claim element recognition result.

[0209] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0210] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0211] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for identifying medical claim materials, characterized in that, include: Determine the contextual information associated with the medical claim image to be identified; The medical claim image is input into a cross-modal network model to obtain at least one candidate entity output by the cross-modal network model and a predicted feature vector corresponding to each candidate entity; Based on the context information, all valid standard entities are selected from the pre-set claims element knowledge graph, and the standard entity embedding vector of each valid standard entity is obtained. Based on the semantic distance between the predicted feature vector and the standard entity embedding vector, a target standard entity matching each target candidate entity is determined from all the valid standard entities; the target candidate entity is any one of all the candidate entities. The target candidate entities that are matched with each target standard entity are modified using the information of each target standard entity to generate structured claims element identification results.

2. The method for identifying medical claim materials according to claim 1, characterized in that, The contextual information associated with the medical claim image to be identified includes: The consultation time and consultation location information are extracted from the text content of the medical claim image; Alternatively, it can receive the patient's appointment time and location information input by the user in relation to the medical claim image.

3. The method for identifying medical claim materials according to claim 2, characterized in that, The process of filtering all valid standard entities from a pre-set claims element knowledge graph based on the context information includes: Traverse the entity nodes in the knowledge graph of the claims elements; each entity node is configured with an effective time attribute and an applicable geographical attribute. Determine whether the time range defined by the effective time attribute includes the consultation time information in the context information to obtain a first determination result; Determine whether the geographical scope defined by the applicable geographical attribute includes the medical treatment geographical information in the context information to obtain a second determination result; Entity nodes whose first and second judgment results are both "yes" are determined as the valid standard entities.

4. The method for identifying medical claim materials according to claim 1, characterized in that, The step of determining the target standard entity that matches each target candidate entity from all the valid standard entities based on the semantic distance between the predicted feature vector and the standard entity embedding vector includes: For each of the target candidate entities, the cosine similarity value between its corresponding predicted feature vector and each of the standard entity embedding vectors is calculated. The valid standard entities are sorted in descending order of their cosine similarity scores. The valid standard entity that ranks first in the sorting is identified as the target standard entity that matches the target candidate entity.

5. The method for identifying medical claim materials according to claim 4, characterized in that, Before determining the first-ranked valid standard entity as the target standard entity that matches the target candidate entity, the method further includes: Determine whether the cosine similarity value corresponding to the first-ranked valid standard entity is greater than a preset calibration threshold; If so, then the step of determining the first-ranked valid standard entity as the target standard entity that matches the target candidate entity is performed.

6. The method for identifying medical claim materials according to claim 1, characterized in that, The cross-modal network model includes visual encoding branches and text encoding branches set in parallel; The step of inputting the medical claim image into a cross-modal network model to obtain at least one candidate entity output by the cross-modal network model and a predicted feature vector corresponding to each candidate entity includes: The visual feature vector is obtained by extracting visual features from the medical claim image using the visual coding branch; the visual layout feature vector represents the table structure information and / or seal position information in the medical claim image. The text encoding branch is used to perform optical character recognition and semantic encoding on the medical claim image to obtain a text semantic feature vector. The visual layout feature vector and the text semantic feature vector are weighted and fused to generate the predicted feature vector.

7. The method for identifying medical claim materials according to claim 6, characterized in that, The weighted fusion of the visual layout feature vector and the text semantic feature vector includes: Based on the clarity index of the medical claim image and the confidence index of the text semantic feature vector, calculate the first fusion weight of the visual layout feature vector and the second fusion weight of the text semantic feature vector. Based on the first fusion weight and the second fusion weight, the visual layout feature vector and the text semantic feature vector are weighted and concatenated to obtain the fused prediction feature vector.

8. The method for identifying medical claim materials according to claim 1, characterized in that, The cross-modal network model is obtained through iterative training using the following steps until a preset cutoff condition is met: Obtain medical image samples containing labeled element tags, and obtain positive sample standard entity vectors corresponding to the element tags from the claims element knowledge graph; The medical image samples are input into the initial network model to be trained, and the initial network model is used to extract the predicted feature vector samples. A loss function is constructed based on the predicted feature vector samples and the positive sample standard entity vectors, and the parameters of the initial network model are updated using the loss function.

9. The method for identifying medical claim materials according to claim 8, characterized in that, The step of constructing a loss function based on the predicted feature vector samples and the positive sample standard entity vectors includes: Based on the difference between the prediction result corresponding to the predicted feature vector sample and the element label, calculate the basic loss term for entity recognition; Calculate the cosine similarity between the predicted feature vector sample and the positive sample standard entity vector; A graph calibration loss term is constructed based on the cosine similarity, and the graph calibration loss term is configured to minimize the vector space distance between the predicted feature vector sample and the positive sample standard entity vector during training. The loss function is obtained by weighted summation of the map calibration loss term and the entity recognition basic loss term.

10. The method for identifying medical claim materials according to claim 8, characterized in that, The step of extracting predicted feature vector samples using the initial network model includes: Visual layout feature vector samples of the medical image samples are extracted using the visual encoding branch of the initial network model; The text semantic feature vector samples of the medical image samples are extracted using the text encoding branch of the initial network model; Accordingly, the loss function also includes a cross-modal alignment loss term, which is configured to maximize the mutual information between the visual layout feature vector sample and the text semantic feature vector sample corresponding to the same medical image sample.

11. The method for identifying medical claim materials according to claim 1, characterized in that, The pre-set knowledge graph of claims elements is dynamically updated through the following steps: Use targeted web crawlers to collect unstructured text data from external medical data sources; The knowledge quality score of the unstructured text data is calculated based on semantic similarity, source authority, and publication timeliness. If the knowledge quality score is greater than the preset entry threshold, then entity triples are extracted from the unstructured text data and updated in the claims element knowledge graph; The entity triple includes a head entity, a tail entity, and an edge defining the relationship between the two, and the head entity or the tail entity is configured with an effective time attribute and an applicable region attribute.

12. The method for identifying medical claim materials according to claim 11, characterized in that, The dynamic update steps also include: If the knowledge quality score remains below the entry threshold for a preset period, the webpage structure code of the external medical data source is obtained. Input the webpage structure code into a preset large language model and obtain the optimized crawling rules output by the large language model; Update the configuration of the targeted crawler using the optimized crawling rules.

13. The method for identifying medical claim materials according to claim 12, characterized in that, The generation of structured claims element identification results includes: The corrected target standard entities are input into the relation classification layer to identify the logical relationships between the target standard entities. The target standard entity and its logical relationships are assembled into an entity triple containing a head entity, a relation edge, and a tail entity, which serves as the structured claims element identification result.

14. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the medical claim material identification method as described in any one of claims 1 to 13.

Images