Claim case data structured information collection method, system, device and medium

By performing text recognition and layout analysis on image materials of claims cases, establishing associated indexes, and performing cross-image field inheritance and conversion, the problem of the inability to retrieve associated fields across images in existing technologies has been solved, achieving automated and accurate field collection.

CN123093333APending Publication Date: 2026-11-13NUANWA (SHANGHAI) INFORMATION TECHNOLOGY CO LTD
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202611282515.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-21
Publication Date
2026-11-13

AI Technical Summary

Technical Problem

In the current claims process, the image materials of claims cases cannot retrieve related fields across images. When fields are missing, they need to be manually entered, which is prone to introducing errors in the assignment of values ​​from irrelevant images. Existing technology cannot achieve code-free expansion and accurate field mapping.

Method used

By performing text recognition and layout analysis on multiple image materials of the same claim case, an image data database is established and an associated index is built. The target field information is extracted first and its completeness and consistency are checked. If it is not qualified, cross-image candidate fields are selected from the associated index for field inheritance and transformation or association reasoning to generate the final text extraction result.

Benefits of technology

It enables automated cross-image field supplementation, reduces manual intervention, avoids errors caused by irrelevant images, and improves the accuracy and efficiency of field acquisition.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123093333A_ABST
    Figure CN123093333A_ABST
Patent Text Reader

Abstract

This application discloses a method, system, device, and medium for collecting structured information from claims case data, relating to the field of computer technology. This application sequentially generates single-image layout data, structured fields, and field data for all images in the same claims case. Image metadata, various original images, and related features are stored in an image data database, and a related index is established. All image recognition data is integrated based on the case, enabling cross-image information retrieval. For target field data, text extraction is performed first to obtain the first field information, and verification is conducted. If the verification is successful, the current image field is directly used, reducing cross-image calculations. If the verification fails, candidate field data across images is filtered based on the related index, and the second field information is obtained through field inheritance transformation or related reasoning. After re-verification, the target field extraction result is determined. No manual image retrieval and supplementation are required; only related images are retrieved for calculation, avoiding field assignment errors caused by irrelevant images.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method, system, device and medium for collecting structured information on claims data. Background Technology

[0002] In health insurance and medical insurance claims, a single claim typically involves multiple types of image materials, including medical invoices, medical insurance settlement statements, expense lists, outpatient and emergency room medical records, inpatient medical records, discharge summaries, diagnostic certificates, examination and test reports, prescriptions, bank statements, and identity and account verification documents. Downstream review, calculation, and risk control systems not only need to obtain the structured fields of each image itself, but also need to generate case-level information that reflects the same medical visit, the same expenses, and the same insured person.

[0003] Existing solutions typically treat each image as an independent recognition object. The system can configure the fields to be output for a specific image type through fixed templates, dedicated OCR (Optical Character Recognition) models, or a combination of general OCR and a large model. This type of solution can solve the problem of "which fields need to be collected for a certain type of image", but the range of values ​​for its fields is usually limited to the current image.

[0004] In real claims materials, a specific target field may not be reliably obtained from the current image due to image blurriness, obstruction, missing pages, OCR recognition errors, or the fact that it is not recorded on the invoice itself. Meanwhile, the same field may clearly exist in other images of the same case, or it may be inferred from fields in other images and invoice data. For example, the patient's name on the invoice may be blurry, but the name may be clearly visible on the medical insurance settlement statement or medical record; the invoice may not clearly indicate outpatient or inpatient status, but medical records, discharge summaries, or expense lists from the same hospital and on the same date can indicate the type of treatment.

[0005] If the above cross-image supplementation logic is implemented item by item through business code, then each time a new target field, candidate image type, association condition or inference caliber is added, the code needs to be modified and republished. If all images of the same case are directly provided to the large model, a large amount of irrelevant material will be introduced, increasing the context length, model cost and the risk of incorrect association. Especially when the same case contains materials from multiple visits or multiple hospitals, it is easy to assign incorrect values ​​to fields that do not belong to the same medical event.

[0006] Therefore, a technical solution is needed that maintains the general OCR and structured extraction capabilities of large models, while elevating the processing unit from a single image to a claims case, and defining the preferred extraction, failure rollback, candidate filtering, field inheritance, correlation reasoning, conflict handling, and source tracing processes of target fields within the scope of multiple images in a case through declarative configuration rather than material-specific code. Summary of the Invention

[0007] This application provides a method, system, device, and medium for collecting structured information on claims data, which at least solves the problems of existing claims data identification which only processes a single image independently, cannot retrieve related fields across images, requires manual supplementation of missing or conflicting fields, and is prone to introducing irrelevant images that cause assignment errors.

[0008] This application provides a method for collecting structured information from claims case data, including: Receive multiple data images of the same claim case, and perform text recognition and layout parsing on each data image to obtain single image layout data; Based on the data type, perform single-image structured extraction processing on the data image to obtain the structured fields and field data of the data image; The metadata, single image layout data, structured fields, field data, and related features of each image are merged and stored into an image data database, and a related index is established. The target field data is extracted by first-choice text extraction to obtain the first field information, and the first field information is then checked for completeness, consistency, conflict and confidence. If the verification is successful, the information in the first field will be used as the text extraction result of the target field data. If the verification fails, a rollback is triggered. Based on the association index, cross-image candidate field data related to the target field data is filtered from the image data database, and field inheritance and transformation or association reasoning is performed on the cross-image candidate field data to obtain the second field information. The second field information is checked for consistency, conflict, and confidence. If the checks are successful, the second field information is used as the text extraction result of the target field data.

[0009] Furthermore, the step of performing text recognition and layout parsing processing on each of the aforementioned data images to obtain single-image layout data includes: Acquire a single data image, perform text recognition on the single data image, and obtain the full text of the image, text block coordinates, and recognition confidence score; Based on the full text of the image and the coordinates of the text blocks, table parsing, page number recognition, seal, signature, and checkbox recognition are performed to integrate and output single image layout data.

[0010] Furthermore, the step of performing single-image structured extraction processing on the data image according to the data type to obtain the structured fields and field data data of the data image includes: Obtain the data type configuration corresponding to the current data image. The data type configuration includes one or more of the following: fields to be collected, field descriptions, output modes, field aliases, and validation rules. Obtain the single image layout data, and combine it with the document type configuration to call a language model, table model, rule model, multimodal large language model or a combination thereof to perform semantic extraction; Integrate the extracted field values, field confidence scores, original field images, source page numbers, and coordinates to output structured field data and field information.

[0011] Furthermore, the establishment of the associated index includes: Obtain the associated features of each of the aforementioned data images, wherein the associated features include one or more of the following: person entity, medical institution, image date range, outpatient number, inpatient number, invoice number, and settlement slip number; Based on association features, the images are divided into multiple association groups, and only images within the same association group are traversed during cross-image retrieval. The index dimensions are: data type, personnel entity, medical institution, image date range, visit identification, invoice number, and field code. The index dimension is bound to its corresponding image metadata, single image layout data, structured fields, and field data to generate a searchable cross-image data association index.

[0012] Furthermore, the step of filtering cross-image candidate field data related to the target field data data from the image data database based on the association index includes: Obtain a preset whitelist of candidate image types and associated constraints; the associated constraints include one or more of the following: same case, same person, same hospital, date distance not exceeding a preset number of days, same outpatient number, same inpatient number, and same expense event; Read all associated indexes in the image data database and filter cross-image candidate field data that belong to the same association group as the target field data and meet the association constraints. For each cross-image candidate field data, a correlation score is calculated by weighted summation based on case consistency, personnel consistency, institution consistency, date proximity, medical visit identification relationship, and semantic fit between candidate images and target fields; Retain cross-image candidate field data with correlation scores higher than the score threshold.

[0013] Furthermore, the step of obtaining the second field information by performing field inheritance and transformation or association reasoning on the cross-image candidate field data includes: Extract source fields that are synonymous with or convertible to the target field from the cross-image candidate field data; If a source field that is synonymous with or convertible to the target field can be extracted, the inheritance confidence of the target field is calculated together based on the source field confidence, candidate image association score, and conversion rule reliability. When the inheritance confidence of the target field is greater than a preset inheritance confidence threshold, field inheritance and conversion are performed to obtain the second field information. The process of field inheritance and conversion includes source field reading, field type checking, standardization conversion, association condition verification, and target field assignment. If a source field that is synonymous with or convertible to the target field cannot be extracted, then perform association reasoning, output the target field value, and combine the target field values ​​generated from multiple cross-image candidate field data to form the second field information.

[0014] Furthermore, the method also includes: Obtain the final text extraction result of the target field. The final text extraction result includes one or more of the following: target field code, field value, confidence level, target image identifier, evaluation type, source image identifier, source field, single image layout data, source field image page number and coordinates, candidate association score, configuration version used, model or operator version, verification status, and conflict handling result. The evaluation type is divided into current image extraction, cross-image inheritance, format conversion, and association inference. Integrate the final text extraction results to generate a cross-image field source chain and output it synchronously; Based on the aforementioned cross-image field source chain, the same field can be re-evaluated for historical cases after configuration or model upgrades to pinpoint the source of the result change.

[0015] This application also provides a structured information collection system for claims case data, including: The case material input module is used to receive multiple data images of the same claim case; The general text recognition and layout parsing module performs text recognition and layout parsing on each of the aforementioned data images to obtain single image layout data; The single-image structured extraction module is used to perform single-image structured extraction processing on the data image according to the data type to obtain the structured fields and field data data of the data image; The case-level image evidence pool is used to merge and store the metadata, single image layout data, structured fields, field data and related features of each image into an image data database, and to establish a related index. The case-level field evaluation engine is used to perform preferred text extraction on the target field data to obtain the first field information, and to perform integrity, consistency, conflict and confidence checks on the first field information; if the checks are successful, the first field information is used as the text extraction result of the target field data. The cross-image candidate selection module is used to trigger a rollback when the verification fails, and to filter cross-image candidate field data related to the target field data data from the image data database according to the association index. The field inheritance and conversion module is used to inherit and convert the cross-image candidate field data to obtain the second field information; The association reasoning module is used to perform association reasoning on the cross-image candidate field data to obtain the second field information; The consistency verification and conflict handling module is used to perform consistency, conflict, and confidence verification on the second field information. The case-level structured output module is used to extract the text of the target field data by using the information of the second field as the text extraction result when the verification is successful.

[0016] This application also provides an electronic device, including: a memory for storing a computer program; and a processor for executing the computer program to implement the steps of any of the above-described methods for collecting structured information on claims data.

[0017] This application also provides a computer-readable storage medium storing a computer program, wherein when the computer program is executed by a processor, it implements the steps of any of the above-described methods for collecting structured information on claims data.

[0018] This application achieves this by performing text recognition and layout parsing on all images of documents under the same claim case to obtain single-image layout data. Then, based on the document type, structured fields and field data are extracted from each single image. Metadata, single-image layout data, structured fields, field data, and related features corresponding to all images are uniformly stored in an image data database, and a related index is built. This allows for the integration of all original recognition data and structured data at the image level within a complete claim case as a unified processing unit. It overcomes the limitations of existing technologies that only store single images independently and cannot retrieve related information across images, fully preserving the original recognition basis of each image and providing a unified and searchable data foundation for subsequent cross-image field supplementation. Based on this, the target field data is first extracted using preferred text extraction to obtain the first field information. Simultaneously, the first field information undergoes multi-dimensional verification of completeness, consistency, conflict, and confidence, prioritizing the use of the current document image. The extracted field results ensure that field collection prioritizes the original information on the ticket, reducing the additional computational overhead caused by cross-image reasoning. When the first field information extracted from the current image fails the verification, the pre-built association index accurately filters matching cross-image candidate field data from the image data database. Then, field inheritance and transformation or association reasoning processing is performed on the filtered cross-image candidate field data to generate second field information. Subsequently, consistency, conflict, and confidence checks are performed on the second field information again, and the best result is selected as the final text extraction result of the target field. The entire process can automatically solve the business pain point of not being able to complete reliable field collection when a single document image has blurry images, missing information, low recognition confidence, or conflicting fields. There is no need for manual intervention to retrieve other images of the same case to manually supplement fields. At the same time, relying on the association index constraint, only related images within the same claim case are retrieved for evaluation, which can avoid the problem of field assignment errors caused by irrelevant images participating in field derivation. Attached Figure Description

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

[0020] Figure 1 This is a flowchart illustrating a method for collecting structured information on claims data in one embodiment of this application; Figure 2 This is a structural block diagram of a structured information collection system for claims case data in one embodiment of this application; Figure 3 This is a schematic diagram illustrating the upgrade path from single-image structured acquisition to case-level structured acquisition in one embodiment of this application; Figure 4 This is a schematic diagram illustrating the process of converting field configuration from a business statement to an executable plan in one embodiment of this application; Figure 5 This is a schematic diagram of a general rollback process after a preferred extraction failure in one embodiment of this application; Figure 6 This is a schematic diagram of the case-level reasoning process in one embodiment of this application when the target field does not directly exist in the invoice; Figure 7 This is an internal structural diagram of a computer device in one embodiment of this application. Detailed Implementation

[0021] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of this application.

[0022] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover 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. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0023] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0024] To address the issues raised in the background technology, CN111444792A presents a technical approach that separates the document template from the text recognition process. This involves locating the area to be recognized using document features or a recognition area template, and then performing text recognition. This solution reduces the coupling between OCR (Optical Character Recognition) and specific text recognition capabilities. CN121582955A discloses a technical solution for recognizing and parsing claims documents based on a multimodal large-scale model, an optical character recognition model, and an inference large-scale model. This solution can utilize multiple models to obtain document information. Furthermore, existing no-code or low-code extraction systems typically allow configuration of material types, field names, field descriptions, output schemas, prompts, and validation rules, thereby expanding the structured fields of a single document without modifying some business code.

[0025] The shortcomings of existing technologies are as follows: the processing object is still centered on a single image or a single document. When the current image field is missing or unreliable, it is impossible to automatically reference other images in the same case according to the field-level configuration; cross-image supplementation logic is usually fixed in business code, process orchestration or special prompt words. Adding new field associations still requires development and release, and true no-code expansion cannot be achieved; simply copying the same field from other images cannot determine whether the candidate images belong to the same hospital, the same visit or the same expense event, which is prone to incorrect associations; inputting all images of a case into a large model at once will introduce irrelevant context and excessively long input, increase the model call cost and reduce the stability of target field inference; some target fields do not exist directly in any single image and need to be combined with fields or ticket data of multiple images for inference. Existing field mapping configurations are difficult to express this evaluation process; if the source image, source image location, association conditions and evaluation actions of cross-image generated fields are not saved at the same time, it is difficult to review, replay and locate errors.

[0026] like Figure 1 , Figure 3 As shown, an embodiment of this application provides a method for collecting structured information from claims case data, including the following steps: Step S1: Receive multiple data images of the same claim case, and perform text recognition and layout parsing on each data image to obtain single image layout data; Step S2: Perform single-image structured extraction processing on the data image according to the data type to obtain the structured fields and field data data of the data image; Step S3: Combine and store the metadata, single image layout data, structured fields, field data and related features of each image into an image data database, and establish a related index. Step S4: Extract the first field information from the target field data using the preferred text, and then verify the completeness, consistency, conflict and confidence of the first field information. Step S5: If the verification is successful, the first field information is used as the text extraction result of the target field data. Step S6: If the verification fails, a rollback is triggered. Cross-image candidate field data related to the target field data is filtered from the image data database according to the association index. Field inheritance and transformation or association reasoning is performed on the cross-image candidate field data to obtain the second field information. Step S7: Perform consistency, conflict, and confidence checks on the second field information. If the checks are successful, the second field information is used as the text extraction result of the target field data.

[0027] The rollback trigger conditions include any one or more combinations of the following: field is empty, confidence level is below the threshold, format validation fails, inconsistency with the basic insured information of the case, and conflict with other fields of the current image.

[0028] Understandably, if the consistency, conflict, and confidence checks on the second field fail, it will be subject to manual review.

[0029] The process involves sequentially generating single-image layout data, structured fields, and field data for all images in the same claim case. Image metadata, various original images, and related features are stored in the image data database, and a related index is established. All image recognition data is integrated based on the case, enabling cross-image information retrieval and retaining complete original recognition data as supplementary support for fields. For target field data, text extraction is performed first to obtain the first field information, and completeness, consistency, conflict, and confidence checks are conducted. If the checks are successful, the current image field is directly used, reducing cross-image calculations. If the checks fail, candidate field data across images is filtered based on the related index, and the second field information is obtained through field inheritance transformation or related reasoning. After a second check, the target field extraction result is determined. This process can automatically solve the collection problems of image blurring, missing fields, low confidence, and data conflicts, eliminating the need for manual image retrieval and supplementation. Only related images are retrieved for calculation, avoiding field assignment errors caused by irrelevant images.

[0030] Furthermore, the step of performing text recognition and layout parsing processing on each of the aforementioned data images to obtain single-image layout data includes: Acquire a single data image, perform text recognition on the single data image, and obtain the full text of the image, text block coordinates, and recognition confidence score; Based on the full text of the image and the coordinates of the text blocks, table parsing, page number recognition, seal, signature, and checkbox recognition are performed to integrate and output single image layout data.

[0031] Furthermore, the step of performing single-image structured extraction processing on the data image according to the data type to obtain the structured fields and field data data of the data image includes: Obtain the data type configuration corresponding to the current data image. The data type configuration includes one or more of the following: fields to be collected, field descriptions, output modes (Schema), field aliases, and validation rules. Obtain the single image layout data, and combine it with the document type configuration to call a language model, table model, rule model, multimodal large language model or a combination thereof to perform semantic extraction; Integrate the extracted field values, field confidence scores, original field images, source page numbers, and coordinates to output structured field data and field information.

[0032] Furthermore, the establishment of the associated index includes: Obtain the associated features of each of the aforementioned data images, wherein the associated features include one or more of the following: person entity, medical institution, image date range, outpatient number, inpatient number, invoice number, and settlement slip number; Based on association features, the images are divided into multiple association groups, and only images within the same association group are traversed during cross-image retrieval. The index dimensions are: data type, personnel entity, medical institution, image date range, visit identification, invoice number, and field code. The index dimension is bound to its corresponding image metadata, single image layout data, structured fields, and field data to generate a searchable cross-image data association index.

[0033] The associated index does not pre-merge fields with the same name from different images; it is only dynamically read when the field is evaluated.

[0034] Furthermore, the step of filtering cross-image candidate field data related to the target field data data from the image data database based on the association index includes: Obtain a preset whitelist of candidate image types and associated constraints; the associated constraints include one or more of the following: same case, same person, same hospital, date distance not exceeding a preset number of days, same outpatient number, same inpatient number, and same expense event; Read all associated indexes in the image data database and filter cross-image candidate field data that belong to the same association group as the target field data and meet the association constraints. For each cross-image candidate field data, a correlation score is calculated by weighted summation based on case consistency, personnel consistency, institution consistency, date proximity, medical visit identification relationship, and semantic fit between candidate images and target fields; Retain cross-image candidate field data with correlation scores higher than the score threshold.

[0035] Specifically, for each cross-image candidate field data set, a correlation score R_j is calculated: R_j = W1 × C_case + W2 × C_person + W3 × C_org + W4 × C_time + W5 × C_visit + W6 × C_semantic; where C_case represents case consistency, C_person represents personnel consistency, C_org represents institutional consistency, C_time represents date proximity, C_visit represents visitation identification relationship, and C_semantic represents the semantic fit between the candidate image and the target field; W1, W2, W3, W4, W5, and W6 are weights; cross-image candidate field data sets with a correlation score R_j higher than the score threshold are retained.

[0036] Furthermore, the step of obtaining the second field information by performing field inheritance and transformation or association reasoning on the cross-image candidate field data includes: Extract source fields that are synonymous with or convertible to the target field from the cross-image candidate field data; If a source field that is synonymous with or convertible to the target field can be extracted, the inherited confidence of the target field is calculated based on the source field confidence, candidate image association score, and conversion rule reliability: C_target = C_source × R_relation × C_transform; where C_source is the source confidence, R_relation is the image association score, and C_transform is the conversion reliability coefficient; when the inherited confidence of the target field is greater than a preset inherited confidence threshold, field inheritance and conversion are performed to obtain the second field information. The process of field inheritance and conversion includes reading the source field, checking the field type, standardizing the conversion, verifying the association conditions, and assigning a value to the target field. If a source field that is synonymous with or convertible to the target field cannot be extracted, then perform association reasoning, output the target field value, and combine the target field values ​​generated from multiple cross-image candidate field data to form the second field information.

[0037] Furthermore, the method also includes: Obtain the final text extraction result of the target field. The final text extraction result includes one or more of the following: target field code, field value, confidence level, target image identifier, evaluation type, source image identifier, source field, single image layout data, source field image page number and coordinates, candidate association score, configuration version used, model or operator version, verification status, and conflict handling result. The evaluation type is divided into current image extraction, cross-image inheritance, format conversion, and association inference. Integrate the final text extraction results to generate a cross-image field source chain and output it synchronously; Based on the aforementioned cross-image field source chain, the same field can be re-evaluated for historical cases after configuration or model upgrades to pinpoint the source of the result change.

[0038] In the aforementioned method for collecting structured information from claims case data, text recognition and layout parsing are performed on all data images under the same claims case to obtain single-image layout data. Then, based on the data type, structured fields and field data are extracted from the single images. Metadata, single-image layout data, structured fields, field data, and related features corresponding to all images are uniformly stored in an image data database, and a related index is built. This method integrates all original recognition data and structured data from all image levels within a complete claims case as a unified processing unit, breaking the limitations of existing technologies that only store single data images independently and cannot retrieve related information across images. It completely preserves the original recognition basis of each image, providing a unified and searchable data foundation for subsequent cross-image field supplementation. Based on this, the target field data data is first subjected to preferred text extraction to obtain the first field information. Simultaneously, the first field information undergoes multi-dimensional verification of completeness, consistency, conflict, and confidence, prioritizing the use of... The field results extracted from the current image data ensure that field collection prioritizes the original information on the invoice, reducing the additional computational overhead caused by cross-image reasoning. When the first field information extracted from the current image fails the verification, the system uses a pre-built association index to accurately filter matching cross-image candidate field data from the image data database. Then, it performs field inheritance and transformation or association reasoning on the filtered cross-image candidate field data to generate the second field information. Subsequently, it performs consistency, conflict, and confidence checks on the second field information again and selects the best one as the final text extraction result of the target field. The entire process can automatically solve the business pain point of not being able to complete reliable field collection when a single image data has blurry images, missing information, low recognition confidence, or conflicting fields. There is no need for manual intervention to retrieve other images of the same case to manually supplement fields. At the same time, relying on the association index constraint, it only retrieves related images within the same claim case for evaluation, which can avoid the problem of field assignment errors caused by irrelevant images participating in field derivation.

[0039] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0040] In one embodiment, such as Figure 2 As shown, a structured information collection system for claims cases is provided, including: The case material input module is used to receive multiple data images of the same claim case; The general text recognition and layout parsing module performs text recognition and layout parsing on each of the aforementioned data images to obtain single image layout data; The single-image structured extraction module is used to perform single-image structured extraction processing on the data image according to the data type to obtain the structured fields and field data data of the data image; The case-level image evidence pool is used to merge and store the metadata, single image layout data, structured fields, field data and related features of each image into an image data database, and to establish a related index. The case-level field evaluation engine is used to perform preferred text extraction on the target field data to obtain the first field information, and to perform integrity, consistency, conflict and confidence checks on the first field information; if the checks are successful, the first field information is used as the text extraction result of the target field data. The cross-image candidate selection module is used to trigger a rollback when the verification fails, and to filter cross-image candidate field data related to the target field data data from the image data database according to the association index. The field inheritance and conversion module is used to inherit and convert the cross-image candidate field data to obtain the second field information; The association reasoning module is used to perform association reasoning on the cross-image candidate field data to obtain the second field information; The consistency verification and conflict handling module is used to perform consistency, conflict, and confidence verification on the second field information. The case-level structured output module is used to extract the text of the target field data by using the information of the second field as the text extraction result when the verification is successful.

[0041] For a description of the features in the corresponding embodiment of the structured information collection system for claims case data, please refer to the relevant description of the corresponding embodiment of the structured information collection method for claims case data, which will not be repeated here.

[0042] In the aforementioned structured information collection system for claims cases, text recognition and layout parsing are performed on all images of the same claims case to obtain single-image layout data. Then, based on the data type, structured fields and field data are extracted from each single image. Metadata, single-image layout data, structured fields, field data, and related features corresponding to all images are uniformly stored in the image data database, and a related index is built. This system integrates all original recognition data and structured data from all images within a complete claims case as a unified processing unit, breaking the limitations of existing technologies that only store single images independently and cannot retrieve related information across images. It completely preserves the original recognition basis of each image, providing a unified and searchable data foundation for subsequent cross-image field supplementation. Based on this, the system first performs preferred text extraction on the target field data to obtain the first field information, and simultaneously performs multi-dimensional verification of the first field information, including completeness, consistency, conflict, and confidence, prioritizing the use of... The field results extracted from the current image data ensure that field collection prioritizes the original information on the invoice, reducing the additional computational overhead caused by cross-image reasoning. When the first field information extracted from the current image fails the verification, the system uses a pre-built association index to accurately filter matching cross-image candidate field data from the image data database. Then, it performs field inheritance and transformation or association reasoning on the filtered cross-image candidate field data to generate the second field information. Subsequently, it performs consistency, conflict, and confidence checks on the second field information again and selects the best one as the final text extraction result of the target field. The entire process can automatically solve the business pain point of not being able to complete reliable field collection when a single image data has blurry images, missing information, low recognition confidence, or conflicting fields. There is no need for manual intervention to retrieve other images of the same case to manually supplement fields. At the same time, relying on the association index constraint, it only retrieves related images within the same claim case for evaluation, which can avoid the problem of field assignment errors caused by irrelevant images participating in field derivation.

[0043] To facilitate understanding of the solution, the following explanation uses a field evaluation DSL (engine) approach. This DSL can be implemented using JSON, YAML, XML, database configuration, rule tables, visualization workflows, or other declarative methods. Furthermore, general OCR and layout parsing can utilize cloud-based OCR, local OCR, multimodal OCR, layout analysis models, tabular models, or multi-model fusion. Single-image structured extraction and cross-image association reasoning can employ large text models, multimodal models, domain models, rule engines, traditional machine learning, or combinations thereof. Case-level image evidence pools can be implemented using relational databases, document databases, search indexes, graph databases, vector databases, or in-memory data structures. Candidate associations can employ hard rules, weighted scoring, learned ranking, vector recall, or a combination of rules and models, as long as they can filter original image data related to the target field based on image relationships within the case. In addition to inheritance and reasoning, field evaluation actions can also include concatenation, aggregation, summation, date range calculation, encoding mapping, entity normalization, and multi-value merging.

[0044] like Figure 3 As shown, the structured information collection system for claims case data includes at least a case material input module, a general OCR (Optical Character Recognition) and layout parsing module, a single image structured extraction module, a case-level image evidence pool, a no-code field evaluation configuration center, a case-level field evaluation engine (DSL), a cross-image candidate selection module, a field inheritance and conversion module, an association reasoning module, a consistency verification and conflict handling module, and a case-level structured output module.

[0045] The case material input module receives one or more images of the same claim case and their optional material types; the general OCR and layout parsing module outputs text, coordinates, tables, page numbers, and recognition confidence; the single image structure extraction module generates the current image fields based on the material type configuration, using a large model or other models; the case-level image evidence pool stores the OCR (text recognition) source image, structured fields, and associated features of each image; the case-level field evaluation engine (DSL) defines the cross-image evaluation strategy for each target field; at runtime, when the current image extraction fails, other images are filtered, field inheritance, transformation, or association reasoning is performed, and case-level structured results that can be directly used by downstream reviewers are output.

[0046] Figure 3 This illustrates the upgrade path of this application from single-image structured acquisition to case-level structured acquisition. OCR and the large model still handle basic literal recognition and semantic extraction, but whether a field needs to be supplemented across images, which images to reference, and how to evaluate it are controlled by the field-level DSL configuration and the general evaluation engine.

[0047] This application does not merely decouple OCR from the large model, nor does it simply add field configurations or multi-model scheduling. Instead, it treats claims cases as unified information collection units and configures a complete "field evaluation strategy" for each target field. This strategy includes at least the following: preferred extraction of the current image, failure trigger conditions, other image types that can be referenced, candidate image association conditions, direct field inheritance or association inference actions, candidate conflict handling, confidence propagation, and source chain output. The configuration is parsed by a general runtime to form an executable flow, therefore, adding cross-image field relationships does not require adding material-specific code.

[0048] This application uses claims cases as the boundary for a single field evaluation. Each case has a case identifier and contains one or more image objects. Each image object includes at least an image identifier, material type, page number information, original image or PDF reference, OCR result, layout structure, single-image structured fields, and field data.

[0049] A single case may involve multiple visits, multiple hospitals, or multiple expense events. To avoid erroneous associations based solely on case identifiers, the system can also establish association features for images, such as personnel, medical institutions, date ranges, outpatient numbers, inpatient numbers, invoice numbers, and settlement slip numbers, and form one or more visit association groups or expense association groups accordingly. When evaluating fields, priority is given to searching among candidate images belonging to the same association group as the current image.

[0050] The general OCR and layout parsing module does not rely on fixed field coordinates of specific materials. It performs text recognition, text block localization, table parsing, page number recognition, and optional seal, signature, and checkbox recognition on the input image, and outputs the full OCR text, text block coordinates, table row and column structure, and recognition confidence.

[0051] The single-image structured extraction module retains the material type-level configuration capability. Business personnel can configure the fields to be collected, field descriptions, output schemas, field aliases, and validation rules for materials such as invoices, medical records, and medical insurance settlement statements. The module generates single-image structured fields from the current image using a large model, tabular model, rule model, or a combination thereof. Each field stores at least the field value, field confidence level, original image of the field, source page number, and coordinates.

[0052] The case-level image evidence pool is used to uniformly store the original OCR (Optical Character Recognition) source images and single-image structured results of each image within a case, and to establish an index that can be retrieved by the field evaluation engine. The index dimensions can include at least material type, personnel entity, medical institution, image date or date range, visit identification, invoice number, field code, and source image location.

[0053] The database retains the image to which the fields belong, and does not pre-merge fields with the same name from different materials. Only when executing the evaluation strategy for a specific target field does the system determine whether to reference other images and how to map source fields to target fields based on the DSL, thereby avoiding the premature mixing of different image information without business semantics.

[0054] A Field Evaluation DSL is a declarative configuration used to describe the complete evaluation process of a target field across multiple images in a case. Business or operations personnel express field retrieval strategies through configuration items, without needing to implement conditional branches, database queries, model calls, or field assignment code. DSLs can be saved in JSON, YAML, forms, regular expressions, or visual workflow formats.

[0055] Each field configuration can include at least the following: Target fields: target image type, field code, field type, output path, and whether it is required; Preferred source: Extract from the current image OCR, the fields already extracted from the current image, or a specified single image model; Rollback trigger conditions: field is empty, confidence level is below the threshold, format validation fails, inconsistent with case basic information or conflict with other fields; Candidate image range: Allowed material types, field paths, and original OCR (Optical Character Recognition) source image range; Related constraints: same case, same person, same hospital, date distance, same outpatient or inpatient number, same expense event, etc. Evaluation actions: field inheritance, field mapping, normalization transformation, set aggregation, or association reasoning based on images from multiple sources; Conflict strategies: Prioritize source, correlation strength, time proximity, confidence level, and constraint fulfillment, or transfer to manual intervention; Output requirements: Field confidence level, source image, source field or original image of the field, page coordinates, configuration version, and evaluation trajectory.

[0056] The DSL parsing module converts field configurations into field evaluation execution plans composed of common operation nodes. These operation nodes include at least the current image extraction node, trigger condition judgment node, original case image query node, candidate image filtering node, field inheritance node, format conversion node, correlation reasoning node, consistency check node, conflict handling node, and result output node.

[0057] Before configuration release, the system performs checks on field path validity, field type, candidate material type, dependency fields, circular dependencies, no-termination rollback, and output path conflicts. Configurations that pass these checks are assigned a version number and loaded by the general runtime. When adding fields or adjusting cross-image relationships, only a new configuration needs to be released; no additional material-specific code is required.

[0058] Figure 4 This illustrates the process of converting field configurations from business statements into executable plans. The configuration describes "under what conditions, from which images, which data, and what evaluation actions to perform," while the general runtime is responsible for performing database retrieval, field mapping, and model invocation.

[0059] For each target field, the system first performs a preferred extraction from the current image according to the DSL configuration. The preferred extraction can directly use the single-image structured fields of the current image, or it can perform rule, table, or large model extraction again from the current image OCR results according to the field configuration.

[0060] When the preferred result meets the requirements of non-empty fields, confidence level, format, and consistency, the system directly outputs the current image result. When there are situations such as missing fields, low recognition confidence, format verification failure, inconsistency with the insured person's information, or conflict with other fields of the current image, cross-image rollback is triggered. Different fields can be configured with different trigger conditions and thresholds.

[0061] After cross-image rollback is triggered, the system does not directly provide all images of the case to the model. Instead, it filters candidate images from the case-level database based on the DSL field. Candidate selection includes at least the following: material type whitelist, personnel consistency, medical institution consistency, date or date range relationship, outpatient number or inpatient number relationship, invoice number relationship, and expense event relationship.

[0062] In one implementation, the system calculates an association score R_j for candidate images D_j: R_j = W1 × C_case + W2 × C_person + W3 × C_org + W4 × C_time + W5 × C_visit + W6 × C_semantic. Where C_case represents case consistency, C_person represents personnel consistency, C_org represents institutional consistency, C_time represents date proximity, C_visit represents visitor identification relationship, and C_semantic represents the semantic fit between the candidate image and the target field; W1, W2, W3, W4, W5, and W6 are determined by field configuration or system configuration. The system retains only the top K candidate images that meet the hard association conditions and have a score higher than a threshold.

[0063] Through dynamic candidate selection, the system can prioritize images belonging to the same medical event as the current image when the same case contains materials from multiple visits or multiple hospitals, thereby reducing cross-event erroneous assignment.

[0064] When other images have reliably extracted fields that are synonymous with or convertible to the target field, the system can perform direct inheritance according to the DSL. The inheritance process includes reading the source field, checking the field type, standardizing the transformation, validating the association conditions, and assigning values ​​to the target field.

[0065] For example, when a patient's name is missing from a medical invoice, the configuration can specify that the patient's name field be selected sequentially from the medical insurance settlement statement, outpatient medical record, or discharge summary. The candidate image must match the hospital on the current invoice, the date must be within a preset window, and the source name must match the insured person's name. After verification, the system standardizes the source field and assigns it to the patient's name on the invoice.

[0066] The inherited confidence score of the target field can be calculated jointly based on the source field confidence score, candidate image association score, and transformation rule reliability. For example, C_target = C_source × R_relation × C_transform; where C_source is the source confidence score, R_relation is the image association score, and C_transform is the transformation reliability coefficient. The inherited result will not override the preferred result for the current image that meets the high confidence condition, unless explicitly specified in the field configuration.

[0067] When a target field cannot be directly obtained from a field with the same name in other images, the system can perform association inference. The association inference node reads the specified field, local OCR text, table cells, or original visual images from the filtered candidate images based on the DSL, and inputs the original visual images within the specified range into the rule engine, text big model, multimodal big model, or domain model to generate the target field.

[0068] For example, the target field is "outpatient or inpatient attribute of the current invoice". The system first infers from the invoice's own OCR and field results; when the invoice does not explicitly state this, the DSL can be configured to select medical records, discharge summaries, expense lists, and medical insurance settlement statements from the same hospital, on the same date, or with dates falling within the expense range, and reference original image data such as the "Outpatient Medical Record" title, outpatient number, inpatient number, admission date, discharge date, and inpatient department. The model generates outpatient or inpatient conclusions based on these cross-image original image data.

[0069] The output of association inference includes at least the target field value, the image identifier involved in the inference, the reference field or the OCR (Optical Character Recognition) source image, the inference rule or model identifier, the association score, and the result confidence. If the output cannot be bound to the candidate images and original image data allowed by the configuration, it will not be considered a valid result.

[0070] Figure 5 The general fallback process after the initial extraction fails is shown. The same runtime can perform both simple field inheritance and correlation reasoning based on other image fields and ticket data.

[0071] When multiple candidate images provide different field values, or when the directly inherited result is inconsistent with the associated inference result, the system performs conflict resolution based on the DSL. Conflict resolution can comprehensively consider material type priority, association score, source field confidence, original image completeness, temporal proximity, and consistency with constraints on case basic information and other fields to select the final result.

[0072] For conflicts that cannot be automatically resolved, the system does not force the generation of a single field value. Instead, it outputs candidate values, their respective source images, correlation conditions, and conflict reasons, and then sends them for manual review. The manual results can be rewritten as new field values, but the original candidates and evaluation trajectory are still retained.

[0073] The final field results should at least save: target field code, field value, confidence level, target image identifier, evaluation type, source image identifier, source field or OCR (Optical Character Recognition) source image, source field image page number and coordinates, candidate association score, configuration version used, model or operator version, verification status and conflict handling results.

[0074] The evaluation type can at least distinguish between direct extraction from the current image, cross-image field inheritance, cross-image field transformation, and cross-image association reasoning. By preserving the complete source chain, the system can explain why the target field does not originate from the current image, and can re-execute the same field evaluation plan for historical cases after configuration or model upgrades to pinpoint the source of the result change.

[0075] This application elevates the information collection unit from a single image to a claims case. Even when current image fields are ambiguous, missing, or not directly recorded, other images from the same case can still be used to generate fields. Cross-image fallback conditions, candidate sources, and evaluation methods for fields are all configured via DSL. Adding new field relationships requires no development of document-specific code or republishing of business programs. Candidate images are dynamically filtered through association constraints such as hospital, date, personnel, visit number, and expense events, reducing erroneous field associations between multiple visits for the same case. Only candidate images and partial original images related to the target field are provided to the model, avoiding inputting all images of the case into the model and reducing irrelevant context, token consumption, and inference interference. It also supports direct field inheritance and cross-image association inference, handling scenarios where other images have fields with the same name, as well as scenarios where the target field requires comprehensive judgment from multiple documents. Field results retain the source image, source field or ticket coordinates, association conditions, and evaluation trajectory, allowing reviewers to verify them and supporting historical playback after configuration and model upgrades. It retains the original ability to configure structured fields by image type, so it can be used for single-image no-code acquisition of any newly added material, or further configured with case-level cross-image acquisition strategies.

[0076] Example 1: Cross-image field completion for patient name on invoice.

[0077] The upstream system receives a claim case, which includes medical invoices, medical insurance settlement statements, and outpatient medical records. The system performs general OCR and single-image structured extraction on each image. The "Patient Name" field of the medical invoice did not yield results due to image blurring or a recognition confidence level below 0.80.

[0078] For the field invoice.patient_name, the DSL configuration for evaluating the field sets the current invoice as the preferred source; a rollback is triggered when the field is empty, has low confidence, or does not match the name of the insured in the case; the candidate material types are, in order, medical insurance settlement statement, medical record, and discharge summary; the candidate images must meet the following requirements: same case, same hospital, image date and invoice date not exceeding three days apart, and the source name must match the name of the insured.

[0079] The system retrieves candidate images from the case database, obtaining the patient's name field and its page coordinates from the medical insurance settlement statement. The confidence score for this source field is 0.98, and the candidate image matches both the hospital and date on the invoice, with a correlation score exceeding the configured threshold. The system performs space cleanup and character standardization on the name before assigning it to the patient's name on the invoice, and outputs the evaluation type "Cross-Image Field Inheritance," the source image identifier, the source field, the source field image page number and coordinates, and the configured version. The entire process requires no additional cross-image processing code.

[0080] Example 2: Cross-image association reasoning of outpatient or inpatient attributes of invoices.

[0081] The upstream system receives medical invoices, medical records, discharge summaries, expense lists, and medical insurance settlement statements. The target field is `invoice.visit_type`, which identifies whether the current invoice is for outpatient or inpatient expenses.

[0082] The system first retrieves original images such as "outpatient," "inpatient," "inpatient number," and "outpatient number" from the invoice's own OCR. When the invoice does not clearly state this or the model's output confidence level is lower than the configured threshold, the system selects other materials from the same hospital, from the same person, and whose dates match the invoice dates, according to the DSL.

[0083] Among the candidate materials, the discharge summary includes the admission and discharge dates, the inpatient medical record includes the hospital number, and the expense list covers the invoice date within the expense range. The system provides these fields and the corresponding OCR (Optical Character Recognition) source images to the associated inference node. The inference node outputs "Hospitalized" and returns the material identifier and original image used in the inference. The system further verifies whether the invoice date falls within the hospitalization range and whether the hospital name matches. After successful verification, it outputs "Hospitalization Invoice".

[0084] Figure 6 This demonstrates the case-level reasoning process when the target field does not directly exist on the invoice. Instead of indiscriminately inputting all case materials into the model, the system only uses candidate images associated with the current invoice.

[0085] Example 3: Direct inheritance and conflict handling of hospital admission numbers across images.

[0086] In one case, the hospital number area on the medical invoice was obscured. Both the inpatient medical record and discharge summary extracted the hospital number, but the two results differed by one character. The DSL field specifies the priority of candidate materials as discharge summary, inpatient medical record, and expense list, while also requiring consistency in the hospital and hospitalization date range.

[0087] The system calculates the correlation scores between the two source images and the current invoice, and compares the field OCR confidence, the integrity of the source image page number and coordinates, and the consistency of the hospitalization date range. If the discharge summary source meets the higher priority and has a higher field confidence, then the hospitalization number is adopted; if it is still uncertain, two candidates and their original images are output for manual review, instead of generating a result through simple majority voting.

[0088] Example 4: Candidate image isolation under the condition of multiple visits for the same case.

[0089] This case involves medical records from two separate visits by the same insured person at two different hospitals on different dates. The patient's name is missing from the current invoice. Searching solely based on the case identifier may incorrectly reference a different medical record.

[0090] This application uses the DSL field as a hard condition, first requiring hospital consistency, and then calculates the correlation score based on date distance, outpatient or inpatient number, cost range, etc., only considering images belonging to the current visit correlation group as candidates. This avoids assigning fields from other medical events in the same case to the current invoice.

[0091] Example 5: No-code release of new materials and cross-image strategies.

[0092] The system needs to add "Pharmacy Purchase Invoice" data and output the purchaser's name, drug name, and prescription / in-purchase attributes. Business personnel first configure the single-image field and extraction instructions for this data; then, they configure a DSL for evaluating the "Prescription / In-Purchase Attributes": if the invoice itself can be clearly identified, it is output directly; otherwise, prescriptions or outpatient medical records from the same case with similar dates are selected, and correlation reasoning is performed based on the drug name and prescription information.

[0093] The configuration is published after passing field path, candidate material and termination condition checks. The general runtime automatically executes the corresponding OCR, single image extraction, cross-image candidate selection and inference process, without the need to add cross-image business code specifically for pharmacy invoices.

[0094] This application retains the basic capabilities of "general OCR to extract literal and layout information and large models to achieve structured understanding" and expands the single image field configuration to a case-level field evaluation configuration. This enables the system to dynamically select the original image from other images in the same case based on the configuration when the target field cannot be reliably obtained from the current image, and generate the target field through field inheritance, format conversion or correlation reasoning.

[0095] First, taking claims cases as the processing unit, the OCR results, single image structured fields, image metadata, and source fields of image page numbers and coordinates of each image within the case are uniformly indexed to form a case-level image evidence pool.

[0096] Second, provide a field evaluation DSL, which declares the preferred source, candidate image type, evaluation action, and conflict strategy for the target field.

[0097] Third, the DSL is compiled into a field evaluation execution plan by a general parser, and the execution process is completed by a general operator, so there is no need to write special code for new materials or new cross-image relationships.

[0098] Fourth, it supports two types of cross-image evaluation methods: direct inheritance and associative reasoning. It can map identified fields in other images to the target field, and it can also combine fields from other images and local ticket data to infer fields that are not directly recorded in the current image.

[0099] Fifth, when outputting field values, the source image, source field or original image of the field, page coordinates, association conditions, evaluation actions, confidence level and configuration version are output simultaneously to form a verifiable cross-image field source chain.

[0100] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the steps in any of the above embodiments of the structured information collection method for claims case data.

[0101] In one embodiment, the electronic device may be a server, and its internal structure diagram may be as follows: Figure 7 As shown, the electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores structured information collected from claims cases. The network interface communicates with external terminals via a network. When the computer program is executed by the processor, it implements a method for collecting structured information from claims cases.

[0102] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above embodiments of the structured information collection method for claims case data.

[0103] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0104] The embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the structured information collection method for claims case data.

[0105] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps in any of the above embodiments of the structured information collection method for claims case data.

[0106] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0107] The foregoing has provided a detailed description of the structured information collection method, system, equipment, and medium for claims case data provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only intended to aid in understanding the method and core ideas of this application. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A method for collecting structured information from claims case data, characterized in that, include: Receive multiple data images of the same claim case, and perform text recognition and layout parsing on each data image to obtain single image layout data; Based on the data type, perform single-image structured extraction processing on the data image to obtain the structured fields and field data of the data image; The metadata, single image layout data, structured fields, field data, and related features of each image are merged and stored into an image data database, and a related index is established. The target field data is extracted by first-choice text extraction to obtain the first field information, and the first field information is then checked for completeness, consistency, conflict and confidence. If the verification is successful, the information in the first field will be used as the text extraction result of the target field data. If the verification fails, a rollback is triggered. Based on the association index, cross-image candidate field data related to the target field data is filtered from the image data database, and field inheritance and transformation or association reasoning is performed on the cross-image candidate field data to obtain the second field information. The second field information is checked for consistency, conflict, and confidence. If the checks are successful, the second field information is used as the text extraction result of the target field data.

2. The method for collecting structured information from claims case data according to claim 1, characterized in that, The step of performing text recognition and layout parsing processing on each of the aforementioned data images to obtain single-image layout data includes: Acquire a single data image, perform text recognition on the single data image, and obtain the full text of the image, text block coordinates, and recognition confidence score; Based on the full text of the image and the coordinates of the text blocks, table parsing, page number recognition, seal, signature, and checkbox recognition are performed to integrate and output single image layout data.

3. The method for collecting structured information on claims cases according to claim 1, characterized in that, The step of performing single-image structured extraction processing on the data image according to the data type to obtain the structured fields and field data data of the data image includes: Obtain the data type configuration corresponding to the current data image. The data type configuration includes one or more of the following: fields to be collected, field descriptions, output modes, field aliases, and validation rules. Obtain the single image layout data, and combine it with the document type configuration to call a language model, table model, rule model, multimodal large language model or a combination thereof to perform semantic extraction; Integrate the extracted field values, field confidence scores, original field images, source page numbers, and coordinates to output structured field data and field information.

4. The method for collecting structured information on claims case data according to claim 1, characterized in that, The establishment of the associated index includes: Obtain the associated features of each of the aforementioned data images, wherein the associated features include one or more of the following: person entity, medical institution, image date range, outpatient number, inpatient number, invoice number, and settlement slip number; Based on association features, the images are divided into multiple association groups, and only images within the same association group are traversed during cross-image retrieval. The index dimensions are: data type, personnel entity, medical institution, image date range, visit identification, invoice number, and field code. The index dimension is bound to its corresponding image metadata, single image layout data, structured fields, and field data to generate a searchable cross-image data association index.

5. The method for collecting structured information on claims case data according to claim 1, characterized in that, The step of filtering cross-image candidate field data related to the target field data data from the image data database according to the association index includes: Obtain a preset whitelist of candidate image types and associated constraints; the associated constraints include one or more of the following: same case, same person, same hospital, date distance not exceeding a preset number of days, same outpatient number, same inpatient number, and same expense event; Read all associated indexes in the image data database and filter cross-image candidate field data that belong to the same association group as the target field data and meet the association constraints. For each cross-image candidate field data, a correlation score is calculated by weighted summation based on case consistency, personnel consistency, institution consistency, date proximity, medical visit identification relationship, and semantic fit between candidate images and target fields; Retain cross-image candidate field data with correlation scores higher than the score threshold.

6. The method for collecting structured information on claims case data according to claim 1, characterized in that, The step of obtaining the second field information by performing field inheritance and transformation or association reasoning on the cross-image candidate field data includes: Extract source fields that are synonymous with or convertible to the target field from the cross-image candidate field data; If a source field that is synonymous with or convertible to the target field can be extracted, the inheritance confidence of the target field is calculated together based on the source field confidence, candidate image association score, and conversion rule reliability. When the inheritance confidence of the target field is greater than a preset inheritance confidence threshold, field inheritance and conversion are performed to obtain the second field information. The process of field inheritance and conversion includes source field reading, field type checking, standardization conversion, association condition verification, and target field assignment. If a source field that is synonymous with or convertible to the target field cannot be extracted, then perform association reasoning, output the target field value, and combine the target field values ​​generated from multiple cross-image candidate field data to form the second field information.

7. The method for collecting structured information on claims case data according to claim 1, characterized in that, The method further includes: Obtain the final text extraction result of the target field. The final text extraction result includes one or more of the following: target field code, field value, confidence level, target image identifier, evaluation type, source image identifier, source field, single image layout data, source field image page number and coordinates, candidate association score, configuration version used, model or operator version, verification status, and conflict handling result. The evaluation type is divided into current image extraction, cross-image inheritance, format conversion, and association inference. Integrate the final text extraction results to generate a cross-image field source chain and output it synchronously; Based on the aforementioned cross-image field source chain, the same field can be re-evaluated for historical cases after configuration or model upgrades to pinpoint the source of the result change.

8. A structured information collection system for claims case data, characterized in that, include: The case material input module is used to receive multiple data images for the same claim case; The general text recognition and layout parsing module performs text recognition and layout parsing on each of the aforementioned data images to obtain single-image layout data. The single-image structured extraction module is used to perform single-image structured extraction processing on the data image according to the data type to obtain the structured fields and field data data of the data image; The case-level image evidence pool is used to merge and store the metadata, single image layout data, structured fields, field data and related features of each image into an image data database, and to establish a related index. The case-level field evaluation engine is used to extract the first field information from the target field data and perform integrity, consistency, conflict and confidence checks on the first field information. If the verification is successful, the information in the first field will be used as the text extraction result of the target field data. The cross-image candidate selection module is used to trigger a rollback when the verification fails, and to filter cross-image candidate field data related to the target field data data from the image data database according to the association index. The field inheritance and conversion module is used to inherit and convert the cross-image candidate field data to obtain the second field information; The association reasoning module is used to perform association reasoning on the cross-image candidate field data to obtain the second field information; The consistency verification and conflict handling module is used to perform consistency, conflict, and confidence verification on the second field information. The case-level structured output module is used to extract the text of the target field data by using the information of the second field as the text extraction result when the verification is successful.

9. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of the method for collecting structured information on claims data as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program is executed by a processor, it implements the steps of the structured information collection method for claims case data as described in any one of claims 1 to 7.

Citation Information

Patent Citations

  • Claim settlement case report processing method and device, electronic equipment and storage medium

    CN121582955A