A complaint text generation method and related apparatus

Abstract

This application discloses a method and related apparatus for generating appeal texts, relating to the field of artificial intelligence technology. The method includes: acquiring medical insurance violation materials from a target medical record; retrieving an appeal evidence chain related to the medical insurance violation materials from a pre-constructed evidence graph; decomposing the appeal evidence chain according to review dimensions to obtain appeal evidence for at least one review dimension; and generating a medical insurance violation appeal text based on the medical insurance violation materials and the appeal evidence for the at least one review dimension using a medical big data model fine-tuned from medical data. This application improves the comprehensiveness and reliability of the appeal evidence chain by constructing an evidence graph adapted to multiple review dimensions used in the medical insurance violation review process; and by automatically generating medical insurance violation appeal texts through a medical big model, it reduces the professional competence requirements for the appellant and improves appeal efficiency.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method and apparatus for generating appeal text. Background Technology

[0002] Appeals against medical insurance audit results are a crucial way to protect the legitimate rights and interests of insured individuals and healthcare providers. Currently, the appeal process for medical insurance audit results is mostly conducted manually, where multiple sources of evidence are manually integrated into a single appeal document. However, this manual appeal model has several problems: Firstly, obtaining and integrating evidence from multiple sources is difficult and requires a high level of expertise. Insufficient evidence may result in fewer appeals collected due to the user's limited expertise, or difficulty in integrating collected evidence from multiple sources, such as non-medical users struggling to integrate data from medical records with medical insurance policies or medical knowledge. Secondly, manual operation is time-consuming, labor-intensive, and inefficient. Summary of the Invention

[0003] In view of the above problems, this application provides a method and related apparatus for generating appeal texts, so as to achieve the purpose of automatically generating more comprehensive and reliable appeal texts for medical insurance violations. The specific solution is as follows:

[0004] The first aspect of this application provides a method for generating appeal text, including:

[0005] Obtain medical insurance violation materials for the target medical record, wherein the medical insurance violation materials are materials describing the medical insurance violations that exist in the target medical record;

[0006] Retrieve the appeal evidence chain related to the medical insurance violation materials from the pre-constructed evidence graph, which is a knowledge graph constructed based on evidence data from multiple review dimensions on which the medical insurance violation review process is based;

[0007] The chain of evidence for the appeal is broken down according to the review dimensions to obtain appeal evidence for at least one review dimension;

[0008] A medical big data model fine-tuned using medical data is used to generate a medical insurance violation appeal text based on the medical insurance violation materials and the appeal evidence from at least one review dimension.

[0009] In one possible implementation, the medical big data model fine-tuned using medical field data generates a medical insurance violation appeal text based on the medical insurance violation materials and the appeal evidence from at least one review dimension, including:

[0010] Obtain the review priority weight for each of the review dimensions, and using the medical big data model, generate the medical insurance violation appeal text based on the medical insurance violation materials, the appeal evidence for at least one review dimension, and the review priority weight. The review priority weight is used to determine the proportion of the corresponding appeal evidence in the medical insurance violation appeal text.

[0011] In one possible implementation, the medical insurance violation materials include a medical insurance violation notice and the original medical record corresponding to the medical insurance violation notice;

[0012] Before retrieving the chain of evidence for the appeal related to the medical insurance violation materials from the pre-constructed evidence map, the method further includes:

[0013] The aforementioned medical big data model is used to extract violation elements related to medical insurance violations from the medical insurance violation notices.

[0014] The local audit engine is used in conjunction with the original medical record to conduct a local audit of the violation elements, resulting in local audit results and audit process data.

[0015] The aforementioned violation elements, the local violation review results, and the review process data are used as the processed medical insurance violation materials.

[0016] The step of retrieving the chain of evidence for appeals related to the medical insurance violation materials from the pre-constructed evidence map includes:

[0017] Retrieve the chain of evidence for appeals related to the processed medical insurance violation materials from the evidence map.

[0018] In one possible implementation, the step of using a local auditing engine in conjunction with the original medical record to perform local violation auditing on the violation elements, obtaining local violation audit results and audit process data, includes:

[0019] The local review engine is used to obtain the local rules pointed to by the violation elements. Based on the local rules and / or preset regular expressions, medical record fragments related to the violation elements are matched from the original medical record text. The local violation review result is determined based on the local rules and the medical record fragments. The local rules are logical judgment rules derived from existing medical policy norms.

[0020] The local rules and the medical record fragments are used as the data for the review process.

[0021] In one possible implementation, the process of breaking down the chain of evidence for the appeal according to review dimensions to obtain appeal evidence for at least one review dimension includes:

[0022] Data related to medical insurance policies are extracted from the aforementioned chain of evidence for the appeal, serving as evidence for the appeal based on policy principles.

[0023] And / or, extract the correspondence between medical record entities and diagnosis and treatment relationships from the chain of evidence for the appeal, sort and map the correspondence according to the timeline of diagnosis first, treatment then test, and perform policy numerical matching analysis on the diagnosis and treatment facts obtained from the sorting and mapping to obtain fact matching data, which serves as the appeal evidence for the fact matching dimension.

[0024] And / or, extract appeal cases from the appeal evidence chain as appeal evidence for the case corroboration dimension;

[0025] And / or, extract medical knowledge from the aforementioned chain of evidence for the appeal as evidence for the appeal in the dimension of clinical consensus.

[0026] In one possible implementation, the medical big data model fine-tuned using medical field data generates a medical insurance violation appeal text based on the medical insurance violation materials and the appeal evidence from at least one review dimension, including:

[0027] Obtain a pre-configured appeal template, which includes at least the error in the violation determination, the basis for the appeal, and the final request;

[0028] The medical big data model is used to generate specific content of the violation judgment error points based on the medical insurance violation materials, to generate specific content of the appeal basis based on the appeal evidence of at least one review dimension, and to generate specific content of the final appeal based on the specific content of the violation judgment error points and the appeal basis, so as to obtain the medical insurance violation appeal text corresponding to the appeal template.

[0029] In one possible implementation, the violation determination error point includes the violation determination source, the error determination type, and the initial rebuttal direction. The violation determination source is used to describe the determination subject and number, the error determination type is used to describe the specific violation item determined, and the initial rebuttal direction is used to describe the determination error point.

[0030] And / or, the basis for the appeal includes a description of the appeal evidence for each review dimension, and the description consists of three parts: the source of the evidence, the content of the evidence, and the relationship between the content of the evidence and the error.

[0031] And / or, the appeal template also includes a reasoning chain visualization diagram, which is a flowchart composed of the correlation logic between the specific content of the violation judgment error point, the appeal basis, and the final request.

[0032] One possible implementation also includes:

[0033] The medical insurance violation appeal text is subjected to at least one of the following checks: evidence source integrity check, entity association consistency check, rule weight reasonableness check, and time logic self-consistency check, and the check result is obtained.

[0034] Specifically, the evidence source integrity check is used to verify whether the medical insurance violation appeal text covers all the appeal evidence of all review dimensions; the entity association consistency check is used to verify whether the entities in the medical insurance violation appeal text can be found in the evidence graph; the rule weight reasonableness check is used to verify whether the proportion of appeal evidence of each review dimension in the medical insurance violation appeal text matches the review priority weight; and the time logic self-consistency check is used to verify whether the time logic of time-related entities in the medical insurance violation appeal text is reasonable.

[0035] One possible implementation also includes:

[0036] Obtain error correction suggestions corresponding to the error information in the evidence graph;

[0037] The evidence map is updated based on the error correction suggestion information to obtain the updated evidence map;

[0038] Obtain a test set, and conduct an appeal text generation test based on the medical insurance violation materials in the test set and the updated evidence map. Determine the evidence retrieval accuracy of the medical insurance violation appeal text obtained in the test compared with the appeal text tags in the test set.

[0039] If the accuracy rate of the evidence retrieval is greater than or equal to the preset accuracy threshold, the updated evidence map will be used to generate medical insurance violation appeal text in the future.

[0040] If the evidence retrieval accuracy is less than the accuracy threshold, the rationality of the error correction suggestion information is analyzed backtracking and unreasonable parts are corrected to obtain updated error correction suggestion information. Then, the evidence map is updated according to the error correction suggestion information to obtain the updated evidence map.

[0041] One possible implementation also includes:

[0042] Obtain a training problem set, wherein each training problem in the training problem set includes: a training medical insurance violation appeal text, a problem description of the problems existing in the training medical insurance violation appeal text, and a problem type label to which the problem description belongs;

[0043] The problem type labels are selected from the training problem set to retrieve the first training problem subset that has been missed in association. Based on the problem descriptions in the first training problem subset, the new associations that have been missed in the evidence graph are determined, and the new associations are updated to the evidence graph.

[0044] And / or, when the evidence retrieval model is used in the retrieval process of the appeal evidence chain, the training medical insurance violation appeal texts in the first training question subset are updated according to the new association, and the updated texts are used as training labels. The first training medical insurance violation materials corresponding to the training labels are obtained, and the evidence retrieval model is trained according to the first training medical insurance violation materials and the training labels.

[0045] And / or, filter the problem type labels from the training problem set to form a second training problem subset that generates text with missing fields, determine the target field missing in the appeal template based on the problem description in the second training problem subset, and add the target field to the appeal template;

[0046] And / or, select a third subset of training problems labeled with logical errors from the training problem set, take the training medical insurance violation appeal texts in the third training problem subset as error texts, obtain the correct texts corresponding to the error texts and the second training medical insurance violation materials corresponding to the error texts, and fine-tune the medical big model based on the second training medical insurance violation materials, the correct texts and the error texts.

[0047] A second aspect of this application provides a computer program product including computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the appeal text generation method of the first aspect or any implementation thereof.

[0048] A third aspect of this application provides an electronic device, comprising at least one processor and a memory connected to the processor, wherein:

[0049] The memory is used to store computer programs;

[0050] The processor is used to execute the computer program so that the electronic device can implement the appeal text generation method of the first aspect or any implementation thereof.

[0051] A fourth aspect of this application provides a computer storage medium carrying one or more computer programs, which, when executed by an electronic device, enable the electronic device to implement the appeal text generation method of the first aspect or any implementation thereof.

[0052] Using the above technical solution, the appeal text generation method provided in this application considers that the core of generating an effective medical insurance violation appeal text lies in providing evidence to prove that the target medical record is not in violation. To obtain appeal evidence proving the target medical record's non-violation, this application can first obtain the medical insurance violation materials for the target medical record, and then retrieve the appeal evidence chain related to the medical insurance violation materials from a pre-constructed evidence graph. Since the evidence graph is a knowledge graph constructed based on evidence data from multiple review dimensions used in the medical insurance violation review process, and the appeal evidence used in the medical insurance violation appeal process is precisely the review evidence used in the medical insurance violation review process, the evidence chain retrieved from the evidence graph can serve as the appeal evidence chain proving the target medical record is not in violation. Because the evidence graph can establish connections between related evidence data, compared to manually collecting fragmented appeal evidence, the appeal evidence chain retrieved based on the evidence graph is more comprehensive and reliable, resulting in a more accurate and effective medical insurance violation appeal text subsequently generated.

[0053] Furthermore, since the medical insurance violation review process involves reviewing evidence from multiple review dimensions in descending order of importance, this application can decompose the appeal evidence chain according to review dimensions to obtain appeal evidence for at least one review dimension. Then, using a medical big data model fine-tuned from medical data, a medical insurance violation appeal text can be generated based on the medical insurance violation materials and the appeal evidence for at least one review dimension. Because the medical big data model has grasped basic medical knowledge after fine-tuning from medical data, it can organize the appeal evidence from each review dimension in a more accurate and logical manner, ensuring the validity of the medical insurance violation appeal text. The entire process requires no manual intervention, reducing the professional competence requirements for the appellant and improving efficiency. Attached Figure Description

[0054] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent when taken in conjunction with the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and the originals and elements are not necessarily drawn to scale.

[0055] Figure 1 A schematic diagram of a system architecture provided for this application;

[0056] Figure 2 A flowchart illustrating a method for generating appeal text provided in this application;

[0057] Figure 3 A schematic diagram of a hardware framework for an appeal text generation process provided in this application;

[0058] Figure 4 A schematic diagram of a hardware framework for post-human-computer interaction verification provided in this application;

[0059] Figure 5 A schematic diagram of an appeal text generation device provided in this application;

[0060] Figure 6 This is a schematic diagram of the structure of an electronic device provided in this application. Detailed Implementation

[0061] The embodiments of this application are described below with reference to the accompanying drawings. The terminology used in the implementation section of this application is for explaining specific embodiments only and is not intended to limit the scope of this application.

[0062] The embodiments of this application will now be described with reference to the accompanying drawings. Those skilled in the art will recognize that, with technological advancements and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are equally applicable to similar technical problems.

[0063] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms are interchangeable where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or apparatus that comprises a series of elements is not necessarily limited to those elements but may include other elements not explicitly listed or inherent to those processes, methods, products, or apparatuses.

[0064] This application provides a method and related apparatus for generating appeal texts, which can be applied to scenarios where appeals are required regarding the results of medical insurance violation reviews.

[0065] For example, if an insured person or medical service provider (such as a hospital) believes that one or more violations pointed out in the medical insurance violation notice are incorrectly judged, they can generate a medical insurance violation appeal text according to this application plan to initiate an appeal.

[0066] Optionally, the appeal text generation method provided in this application can be applied to, for example... Figure 1 The system architecture shown includes a terminal 100 and a server 200. The server 200 may include one or more servers (…). Figure 1 (This example uses a server as an illustration).

[0067] Either terminal 100 or server 200 can be used independently to execute the appeal text generation method provided in the embodiments of this application. Alternatively, terminal 100 and server 200 can also be used collaboratively to execute the appeal text generation method provided in the embodiments of this application.

[0068] For example, an insured person or a medical service provider inputs medical insurance violation materials on the user interface provided by terminal 100. Terminal 100 transmits the medical insurance violation materials to server 200, which automatically generates a medical insurance violation appeal text. Finally, server 200 feeds back the medical insurance violation appeal text to terminal 100 for output.

[0069] The following description Figure 1 The product form of the mid-terminal 100;

[0070] The terminal 100 in this application embodiment can be a mobile phone, tablet computer, wearable device, vehicle device, augmented reality (AR) / virtual reality (VR) device, laptop computer, ultra-mobile personal computer (UMPC), netbook, personal digital assistant (PDA), etc., and this application embodiment does not impose any restrictions on it.

[0071] To enable those skilled in the art to better understand this application, the method for generating appeal text according to embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0072] Reference Figure 2 , Figure 2 This application provides a flowchart illustrating a method for generating appeal text, as shown in the embodiments below. Figure 2 As shown, the appeal text generation method may include:

[0073] Step S201: Obtain medical insurance violation materials for the target medical record.

[0074] Here, medical insurance violation materials refer to materials describing the medical insurance violations found in the target medical record.

[0075] Optionally, the materials related to medical insurance violations should include at least the medical insurance violation notice and the original medical record corresponding to the medical insurance violation notice.

[0076] Among them, the medical insurance violation notice refers to any form of violation information issued by the two designated medical insurance platforms, such as a written notice or verbal notice, that carries the medical insurance violation content of the target medical record. For example, "Patient **, Department of admission **, Settlement time **, According to the medical insurance charge catalog, recombinant human thrombopoietin injection is limited to thrombocytopenia caused by chemotherapy for solid tumors or primary immune thrombocytopenia. The prescribed drug exceeds the medical insurance limit payment scope, involving a violation amount of 2,367 yuan."

[0077] The original medical record includes, but is not limited to, the following information: basic patient information (such as identity information, social information, cost information, etc.), core records of diagnosis and treatment (such as chief complaint, present illness history, past medical history, physical examination data, auxiliary examination data, etc.), diagnosis and treatment process and decision data (admission diagnosis data, differential diagnosis data, medical progress records, etc.), treatment results and follow-up guidance data.

[0078] It should be noted that, in addition to medical insurance violation notices and original medical records, other materials related to medical insurance violations may also be included, such as transfer records, etc. This application does not specify any particular restrictions.

[0079] Step S202: Retrieve the appeal evidence chain related to medical insurance violation materials from the pre-constructed evidence graph. The evidence graph is a knowledge graph constructed based on evidence data under multiple review dimensions of the medical insurance violation review process.

[0080] To facilitate the collection of comprehensive evidence materials in the future, this embodiment can pre-construct an evidence map. Since the appeal evidence used in the medical insurance violation appeal process is also the audit evidence used in the medical insurance violation review process, in order to generate a medical insurance violation appeal text that is compatible with the medical insurance violation review process in the future, this embodiment can construct an evidence map based on the evidence data under multiple audit dimensions on which the medical insurance violation review process is based.

[0081] Optionally, the process of constructing the evidence graph may include: extracting entities from evidence data under multiple review dimensions based on the medical insurance violation review process, determining the relationships between the extracted entities based on preset entity relationship extraction rules and / or semantic relationships between entities, determining edge weights based on information such as the frequency of occurrence and semantic strength of entity relationships, and finally constructing the evidence graph based on entities, entity relationships and edge weights.

[0082] Of course, there are other ways to construct the evidence graph, and this application does not impose specific limitations. For example, in order to obtain a more accurate chain of evidence for the appeal through evidence retrieval, and to improve retrieval efficiency, the entities in the evidence graph can be divided into multiple communities based on the closeness between the entities. This allows for the retrieval of relevant communities first, and then the search of the chain of evidence for the appeal within each community.

[0083] In this embodiment, the evidence graph can serve as an external knowledge base, providing more comprehensive and reliable evidence for generating medical insurance violation appeal texts. Therefore, when it is necessary to collect appeal evidence based on medical insurance violation materials, the evidence graph can be used to retrieve a chain of appeal evidence proving that the medical insurance violation content pointed out in the medical insurance violation notice was not actually a violation.

[0084] Optionally, the aforementioned review dimensions may include: policy basis, factual matching, case evidence, and clinical consensus. Therefore, the evidence graph could be a knowledge graph integrating medical insurance policies, complete medical records, appeal cases, and medical knowledge.

[0085] Of course, different audit dimensions may be used in medical insurance violation audits in different regions or at different levels. Therefore, the evidence map may contain evidence data of fewer or more than the above four audit dimensions. This application does not make specific limitations.

[0086] Additionally, it should be noted that evidence data for a particular review dimension may be updated. In order for this application to retrieve a valid chain of evidence for the appeal, a specific update strategy can be adopted to update the evidence map in relatively real-time. For example, a specific update strategy could be weekly updates, monthly updates, or updating the policy map along with the evidence data, etc. This application does not specify any particular strategy.

[0087] Step S203: Decompose the appeal evidence chain according to the review dimensions to obtain appeal evidence for at least one review dimension.

[0088] Considering that the medical insurance violation review process usually follows the order of importance of the review dimensions from high to low, the appeal evidence for each dimension is reviewed one by one. For example, first, it is reviewed whether there are sufficient national and local medical insurance policies to support whether the declared content is in violation. Then, the medical behavior is objectively compared with the policy provisions to verify the logical consistency. After that, high-risk patterns are identified by referring to historical violation cases, and medical knowledge based on clinical consensus is used to determine the violation.

[0089] In order to present evidence materials that are compatible with the review process in the medical insurance violation appeal text, this embodiment can first decompose the appeal evidence chain in a structured way. More specifically, it can be decomposed step by step according to the review dimensions to obtain the appeal evidence for each review dimension.

[0090] It should be understood that when the evidence map constructed above contains all or nearly all evidence data, for common medical records, the appeal evidence chain can generally include appeal evidence from all review dimensions. However, there may be exceptions, resulting in the appeal evidence chain containing only appeal evidence from some review dimensions. For example, there may be medical insurance violations that have never occurred before, which may lack relevant appeal cases or medical insurance policies, or the appeal evidence chain may lack appeal evidence from certain review dimensions due to the slow update speed of the evidence map. Therefore, when the appeal evidence chain is broken down, appeal evidence from at least one of multiple review dimensions may be obtained.

[0091] Step S204: Using a medical big data model fine-tuned from medical data, generate a medical insurance violation appeal text based on medical insurance violation materials and appeal evidence from at least one audit dimension.

[0092] In this embodiment, considering that when manually drafting medical insurance violation appeal texts, the writer needs to have extensive knowledge of medical-related knowledge, policies, and historical appeal cases in order to find the evidence materials required for the target medical record from a large amount of evidence data, and at the same time, the found multi-source evidence materials need to be organized together in an accurate and reasonable logic, the whole process requires a lot of manpower and time costs, and the efficiency is very low.

[0093] To leverage the powerful semantic understanding and deep reasoning capabilities of large language models, this embodiment can pre-fine-tune a general-purpose large language model (such as the iFlytek Spark Large Model or the Deep Search Large Model) using medical domain data (such as medical-related knowledge, policies, historical appeal cases, medical records, etc.) to obtain a medical large model. Therefore, a medical insurance violation appeal text can be generated based on medical insurance violation materials and appeal evidence from at least one review dimension.

[0094] Optionally, in order to distinguish evidence of different importance in medical insurance violation appeal texts, this embodiment can pre-assign review priority weights to each review dimension based on the actual judgment priority of medical insurance review (such as policy basis > fact matching > case evidence > clinical consensus), to ensure that the appeal content is focused and conforms to the review logic.

[0095] Based on this, we can first obtain the review priority weight of each review dimension, and then use the medical big model to generate a medical insurance violation appeal text based on the medical insurance violation materials, the appeal evidence of at least one review dimension, and the review priority weight. The review priority weight is used to determine the proportion of the corresponding appeal evidence in the medical insurance violation appeal text.

[0096] It should be noted that the length ratio here refers to the length ratio of the corresponding appeal evidence relative to the appeal evidence of all review dimensions. Preferably, the higher the review priority weight of a review dimension, the greater the length ratio of the appeal evidence of that review dimension compared to the appeal evidence of other dimensions.

[0097] This embodiment utilizes a large medical model to organize the appeal evidence from various review dimensions in a more accurate and logical manner. During the organization process, more detailed descriptions can be generated for appeal evidence with higher review priority weights. Thus, when reviewers review medical insurance violation appeal texts, their limited attention can be precisely guided to the core evidence that determines the success or failure of the appeal, thereby improving the appeal success rate.

[0098] It should also be noted that, optionally, the process of dismantling the chain of evidence for the appeal in step S103 can also be implemented using a large medical model to improve efficiency.

[0099] The appeal text generation method provided in this application considers that the core of generating a valid medical insurance violation appeal text lies in providing evidence to prove that the target medical record is not in violation. To obtain appeal evidence proving the target medical record's non-violation, this application first obtains the medical insurance violation materials for the target medical record, and then retrieves the appeal evidence chain related to the medical insurance violation materials from a pre-constructed evidence graph. Since the evidence graph is a knowledge graph constructed based on evidence data from multiple review dimensions used in the medical insurance violation review process, and the appeal evidence used in the medical insurance violation appeal process is precisely the review evidence used in the medical insurance violation review process, the evidence chain retrieved from the evidence graph can serve as the appeal evidence chain proving the target medical record's non-violation. Because the evidence graph can establish connections between related evidence data, compared to manually collecting fragmented appeal evidence, the appeal evidence chain retrieved based on the evidence graph is more comprehensive and reliable, resulting in a more accurate and effective medical insurance violation appeal text.

[0100] Furthermore, since the medical insurance violation review process is based on evidence from multiple review dimensions, reviewed in descending order of importance, this application can decompose the appeal evidence chain according to review dimensions to obtain appeal evidence for at least one review dimension. Then, using a medical big data model fine-tuned from medical data, a medical insurance violation appeal text can be generated based on the medical insurance violation materials and the appeal evidence for at least one review dimension. Because the medical big data model has grasped basic medical knowledge after fine-tuning from medical data, it can organize the appeal evidence from each review dimension in a more accurate and logical manner, ensuring the validity of the medical insurance violation appeal text. The entire process requires no manual intervention, reducing the professional competence requirements for the appellant (such as the writer mentioned above) and improving efficiency.

[0101] In some embodiments of this application, considering that most medical institutions nowadays have local review engines to improve the success rate of medical insurance reviews, they can perform local reviews of input materials, such as reviewing whether a certain drug used in a medical record is covered by medical insurance. Since local review is essentially a review of medical insurance violations, the relevant data from local review can be used as reference information when retrieving evidence to find more accurate evidence materials.

[0102] See Figure 3 This is a schematic diagram of the hardware framework for an appeal text generation process provided in this application. Figure 3 In this embodiment, before retrieving the appeal evidence chain related to medical insurance violation materials from the pre-constructed evidence map in step S202, the medical big model is first used to extract violation elements related to medical insurance violations from the medical insurance violation notice. For example, given the following prompt words, structured violation elements including violation type, violation item, and violation amount are extracted.

[0103] "As an expert in analyzing medical insurance violation information, please extract structured data from the given violation information text based on the following rule definitions."

[0104] ### Partial Analysis of Project Definitions:

[0105] 1. Violation Rules: These refer to restrictive provisions in medical insurance policies regarding the use of medical items. Output types may include "Excessive drug use, duplicate charges, excessive dosage, incompatible indications, etc."

[0106] - Excessive drug use: Specifically refers to the use of drugs that exceed the indications, populations, disease stages, or treatment lines specified in the medical insurance catalog (a first-level rule type). ......

[0108] Violation items: refers to the specific names of drugs or treatment items involved in the violation. The full name must be extracted accurately, excluding descriptions of the violation type.

[0109] ### Text of violation information to be parsed:

[0110] "Patient **, Department Admitted **, Settlement Time **. According to the medical insurance reimbursement catalog, recombinant human thrombopoietin injection is limited to thrombocytopenia caused by chemotherapy for solid tumors or primary immune thrombocytopenia. The prescribed medication exceeds the medical insurance reimbursement limit, involving a violation amount of 2367 yuan."

[0111] ### Output Requirements:

[0112] Please return the following structured information in JSON format, including the following fields:

[0113] - violationRule (Violation rule, must be aligned with the local engine rule format)

[0114] - violationItem (The type and name of the drug must be specified in the violation item)

[0115] - basicInfo (basic information, including the admitting department, settlement time, desensitized sensitive information, and amount of violation)).

[0116] Taking the medical insurance violation notice from the previous example as an example, the extracted violation elements can be: {“violationRule”:“Excessive drug use”:“violationItem”:“Recombinant human thrombopoietin injection”:“basicInfo”:“......”}.

[0117] Next, the local audit engine can be used in conjunction with the original medical records to conduct local audits of the violation elements, and the local audit results and audit process data can be obtained.

[0118] Optionally, the local violation review process can be as follows: using the local review engine to obtain the local rules pointed to by the violation elements, matching medical record fragments related to the violation elements from the original medical record text according to the local rules and / or preset regular expressions, and determining the local violation review result based on the local rules and medical record fragments.

[0119] The aforementioned local rules refer to computer-executable logical judgment rules derived from existing medical policies and regulations. For example, national standards regarding medication can be transformed into local rules that can be used to extract medical record fragments in this application.

[0120] In practical applications, different local engine rule bases can be configured for different violation items. Based on the violation items and rules extracted above, the corresponding local engine rule base can be located to obtain the local rules within the base. Alternatively, all local rules can be stored in a single rule base and retrieved when needed. This application does not impose specific limitations on this approach.

[0121] In this embodiment, the local review engine can match medical record segments related to the violation elements from the original medical record text based on local rules, or it can locate medical record segments related to the violation elements from the original medical record text based on preset regular expressions. By filtering medical records, interference from segments unrelated to the violation elements can be reduced in subsequent local violation review.

[0122] Finally, the system can combine the matched medical record fragments with local rules to provide a local violation review result, indicating whether the violation is in violation or not.

[0123] For example, regarding the aforementioned violation elements {"violationRule": "excessive drug use": "violationItem": "recombinant human thrombopoietin injection": "basicInfo": "......"}, after processing by the local audit engine, the output content can be as follows:

[0124] "Local Rules: Rule Combination Logic: ****; Rule 1: The diagnosis list includes diagnoses related to solid tumors; Rule 2: The medical record includes chemotherapy-related records; Rule 3: The examination and testing section includes platelet count indicators, and the value is less than 100 × 10⁻⁶." 9 / L; Rule 4: The diagnostic list includes diagnoses related to essential thrombocytopenia...

[0125] Medical record excerpt: "2025-06-05 Laboratory report: Platelet count 65×10⁻⁵" 9 / L; 2025-06-06 Medical Record: The patient has primary myelofibrosis, diagnosis list: ......

[0126] Local violation review result: Violation.

[0127] The aforementioned local rules and medical record fragments constitute the data for the review process. Figure 3 Local violation review results and review process data are used together as local review clues.

[0128] In one possible implementation, if the local violation review result is "no violation", the violation elements and local review clues can be used as processed medical insurance violation materials. Then, in step S202, the appeal evidence chain related to the processed medical insurance violation materials can be retrieved from the evidence map.

[0129] Given the limited accuracy of local review engines, which may produce incorrect results that could impact subsequent appeals, and considering that appeals for medical insurance violations are typically initiated only after expert analysis confirms that the alleged violations are true, the local review engine's results should only be used as a reference. That is, regardless of whether the local review engine determines a violation or not, the violation elements and local review clues are treated as processed medical insurance violation materials. In step S202, the evidence chain related to the processed medical insurance violation materials is then retrieved from the evidence graph.

[0130] Compared to the approach of large medical models that directly generate medical insurance violation appeal texts from original medical records, this application incorporates pre-generated medical record fragments from a local review engine. This allows the model to focus more on key content of the medical records and reduces model illusions.

[0131] Given the numerous highly complex and specialized procedures involved in the medical insurance violation appeal process, while common retrieval paradigms (such as RAG (Retrieval-Augmented Generation)) are suitable, GraphRAG offers advantages such as deep association and multi-hop reasoning, globally consistent understanding, overcoming semantic similarity limitations, and high interpretability and credibility. GraphRAG, in particular, is well-suited for medical insurance violation appeal scenarios and can retrieve more accurate search results. Therefore, please refer to [further details omitted]. Figure 3 Optionally, the evidence retrieval process in step S202 may employ the GraphRAG retrieval paradigm.

[0132] Taking the evidence map integrating the aforementioned medical insurance policies, the full original medical records, appeal cases, and medical knowledge as an example, based on the unique identifier of the target medical record, the aforementioned violation elements such as "excessive drug use, recombinant human thrombopoietin injection, etc.", and the output of the local review engine "patient platelet count 65×10 9 "Information such as / L, diagnosis of primary myelofibrosis, etc." The chain of evidence for the appeal obtained through GraphRAG retrieval can be:

[0133] "Medical insurance policy: The reimbursement restrictions for recombinant human thrombopoietin injection in the National Basic Medical Insurance, Work Injury Insurance and Maternity Insurance Drug Catalog (2024) clearly stipulate that the medication must meet the condition of 'thrombocytopenia after chemotherapy for solid tumors (platelet count < 50 × 10⁻⁶)'." 9 / L) or primary immune thrombocytopenic purpura;

[0134] Appeal Cases: Case No. ******, a patient appealed regarding "the use of recombinant human thrombopoietin injection for primary myelofibrosis," with the appeal opinion being ******. The Medical Insurance Bureau's review opinion was "the diagnosis does not meet the requirements, and the appeal failed"; Case No. ******, a patient appealed regarding "the use of recombinant human thrombopoietin injection for primary myelofibrosis," with the appeal opinion being ******. The Medical Insurance Bureau's review opinion was "the appeal was successful."

[0135] Patient's medical record: 2025-07-15 10:00, Recombinant human thrombopoietin injection 15000 IU, subcutaneous injection; Primary myelofibrosis (ICD-10: D75.802), etc.

[0136] Medical knowledge: Primary myelofibrosis is a myeloproliferative neoplasm, and its pathogenesis and treatment requirements differ from those of thrombocytopenia following chemotherapy in solid tumors. Recombinant human thrombopoietin injection is not a routine treatment for primary myelofibrosis. Patients with primary myelofibrosis have a platelet count <30 × 10⁻⁶. 9 When the thrombopoietin level is 90 / L, thrombopoietin receptor agonists may be considered, but recombinant human thrombopoietin is not the first choice.

[0137] The evidence chain of the appeal retrieved via GraphRAG is then fed into the medical big data model. Within the medical big data model, the evidence chain is decomposed and the appeal text is generated, thus obtaining the medical insurance violation appeal text.

[0138] Compared to methods that do not incorporate local review clues and GraphRAG retrieval, this embodiment introduces local review clues, allowing GraphRAG retrieval to refer to the review mode of the local review engine, thus improving the accuracy of the retrieval.

[0139] In some embodiments of this application, the process of step S203, "disassembling the appeal evidence chain according to the review dimensions to obtain appeal evidence of at least one review dimension", is described.

[0140] As mentioned above, multiple review dimensions may include policy basis dimension, fact matching dimension, case evidence dimension and clinical consensus dimension. Based on this, the evidence chain decomposition process in this embodiment may include the decomposition process of at least one of the following review dimensions.

[0141] First, extract relevant data on medical insurance policies from the chain of evidence in the appeal, and use it as evidence in the appeal based on policy principles.

[0142] Optionally, the extracted medical insurance policy-related data can be sorted by "policy level (national > local), effective time (current > expired), and clause type (drug restriction / treatment guidelines)". This allows the current national policy clauses to be prioritized as the core basis when generating medical insurance violation appeal texts in the future.

[0143] Second, extract the correspondence between medical record entities and diagnosis and treatment relationships from the chain of evidence for the appeal. Organize and map the correspondence according to the timeline of diagnosis first, then treatment, and finally testing. Then, conduct policy numerical matching analysis on the diagnosis and treatment facts obtained from the sorting and mapping to obtain fact matching data, which serves as the appeal evidence for the fact matching dimension.

[0144] Specifically, to obtain evidence for the appeal based on factual matching, the following two steps are required: First, link the "medical record entity - treatment relationship" in the appeal evidence chain, and sort out the facts according to the timeline of "diagnosis -> treatment -> testing". For example, the ICD code of medical record MED_20250715 is C34.9 (lung cancer), the chemotherapy regimen is (paclitaxel + carboplatin), and the platelet count is 42×10. 9 " / L" maps to the clinical facts: "The patient was admitted to the hospital for 'right lung adenocarcinoma (ICD-10: C34.901)', and underwent paclitaxel + carboplatin chemotherapy on July 10, 2025. A follow-up platelet count on July 12 was 42 × 10⁻⁶." 9 / L, which meets the diagnosis of thrombocytopenia after chemotherapy; the second step is to match the diagnosis and treatment facts with the policy value thresholds, and obtain fact matching data based on the matching results.

[0145] For example, in the second step of processing, judgment terms such as "clearly meets," "exceeds," and "does not meet" can be used to quantify the matching relationship between the numerical entities in the medical records and the policy numerical thresholds in the diagnosis and treatment facts. For example, "platelet count 42×10" can be used as a quantification. 9 / L" and the policy "<50×10" 9 The relation mapping for " / L" is the matching relation "platelet count 42×10" 9 / L, lower than the policy limit of 50×10 9 The / L threshold fully meets the medication requirements.

[0146] Therefore, the above-mentioned diagnostic and treatment facts and matching relationships constitute the fact-matching data.

[0147] Third, extract appeal cases from the appeal evidence chain to serve as supporting evidence for the appeal.

[0148] Optionally, the extracted appeal cases can be sorted by "regional level (provincial level > local level), case closing time, appeal success rate (success > failure)", giving priority to successful cases at the provincial level and above, and giving priority to cases within the past year.

[0149] Fourth, extract medical knowledge from the chain of evidence in the appeal to serve as evidence in the dimension of clinical consensus.

[0150] Optionally, the extracted medical knowledge can be sorted by the degree of consensus (i.e., weight), and some medical knowledge with low degree of consensus can be filtered out.

[0151] In this embodiment, by structurally decomposing and arranging the appeal evidence chain according to the core judgment dimensions of medical insurance review, it is ensured that the appeal evidence for each review dimension corresponds to the key concerns of medical insurance review. This makes the organization logic of evidence in subsequent medical insurance violation appeal texts more consistent with the logic of medical insurance violation review, thereby improving the readability of medical insurance violation appeal texts and the success rate of appeals.

[0152] In some other embodiments of this application, the process of step S204 above, "using medical data to fine-tune a large medical model, generating a medical insurance violation appeal text based on medical insurance violation materials and appeal evidence from at least one audit dimension," is described.

[0153] To further improve the readability of medical insurance violation appeal texts and to adapt to the review requirements of medical insurance regulators, this embodiment can pre-configure appeal templates, enabling the medical big data model to generate structured medical insurance violation appeal texts according to the appeal templates.

[0154] In this embodiment, the appeal template includes at least three parts: the error point in the violation determination, the basis for the appeal, and the final demand. The error point in the violation determination describes the erroneous content pointed out in the medical insurance violation notice. The basis for the appeal clarifies the internal logic of the aforementioned erroneous content based on the appeal evidence from each review dimension. The final demand describes the user's ultimate goal or requirement.

[0155] For example, the final appeal could be: "In summary, the patient's use of recombinant human thrombopoietin injection complies with the core policy restrictions of the 2024 National Reimbursement Drug List, is supported by medical records, and is supported by similar cases and clinical guidelines. The 'excessive use' judgment by the designated medical insurance platform is erroneous. We hereby formally apply for: 1. Revocation of the violation judgment (No.: ****); 2. Restoration of the drug's eligibility for reimbursement by the medical insurance fund. Our unit can provide complete original medical records, copies of test reports, and other materials to assist in the review at any time."

[0156] Optionally, the error in the violation determination can include three parts: the source of the violation determination, the type of the error determination, and the direction of the preliminary rebuttal. The source of the violation determination describes the determining entity and the number (e.g., a violation determination issued by the two-tier platform on August 10, 2025, with the number Y20250810xxx). The type of the error determination describes the specific violation item (e.g., determining that the patient's use of recombinant human thrombopoietin injection is an over-limit drug use). The direction of the preliminary rebuttal describes the error in the determination (i.e., the erroneous content) (e.g., the actual patient scenario fully meets the conditions limited by the medical insurance policy, and the determination has a factual deviation).

[0157] Optionally, the basis for the appeal includes a description of the appeal evidence for each review dimension. This description consists of three parts: the source of the evidence, the content of the evidence, and the relationship between the content of the evidence and the error.

[0158] Optionally, in scenarios where the review priority weights of each review dimension are referenced, the proportion of the description of each review dimension in the appeal basis is determined by the review priority weight. For example, if the review priority weights are 40% for policy basis, 30% for fact matching, 20% for case evidence, and 10% for clinical consensus, then the length of the policy basis dimension in the appeal basis will account for 40% of the total length of the appeal basis section, the length of the fact matching dimension will account for 30% of the total length of the appeal basis section, the length of the case evidence dimension will account for 20% of the total length of the appeal basis section, and the length of the clinical consensus dimension will account for 10% of the total length of the appeal basis section.

[0159] Optionally, the appeal template may also include a reasoning chain visualization diagram, which is a flowchart composed of the logical connections between the specific content of the violation judgment error, the basis for the appeal, and the final demand. The weight values ​​of the evidence at each stage may also be marked in the reasoning chain visualization diagram.

[0160] After obtaining the above-mentioned appeal template, this embodiment can use a medical big data model to generate specific content of the error points in the medical insurance violation judgment based on the medical insurance violation materials, generate specific content of the appeal basis based on the appeal evidence of at least one review dimension, and generate specific content of the final appeal based on the specific content of the error points in the violation judgment and the appeal basis, so as to obtain the medical insurance violation appeal text corresponding to the appeal template.

[0161] In this embodiment, the judgment logic of medical insurance review is replicated as the core. The appeal content that meets the normative requirements of medical insurance supervision is generated in three steps: "evidence disassembly and matching to align with the core dimensions of review, weight adjustment of the review priority, and document organization that adapts to the format requirements of medical insurance supervision". This ensures that the appeal reasons are highly aligned with the focus of medical insurance review and improves the effectiveness of medical insurance violation appeal text.

[0162] In some further embodiments of this application, in order to improve the reliability of the medical insurance violation appeal text generated above, this embodiment provides a post-human-computer interaction verification mechanism, such as... Figure 4 This is a schematic diagram of a hardware framework for post-verification of human-computer interaction provided in this application. In this mechanism, medical large models can be trained and evidence maps can be updated by combining expert feedback.

[0163] Considering the fully automated generation of medical insurance violation appeal texts described above, due to the high level of specialization and complexity of medical knowledge, there may be instances where the medical insurance violation appeal texts are unusable. To avoid appeal failures caused by unusable medical insurance violation appeal texts, this embodiment can verify the medical insurance violation appeal texts after they are generated, and obtain the verification results.

[0164] Optionally, this embodiment provides the following four-dimensional verification rules. In practical applications, one or more of these verification rules can be used to verify medical insurance violation appeal texts.

[0165] Dimension 1: Completeness of Evidence Sources. That is, this embodiment can verify the completeness of evidence sources for medical insurance violation appeal texts to check whether the appeal texts cover all the appeal evidence across all review dimensions.

[0166] It should be understood that in order to improve the success rate of appeals, the evidence in the medical insurance violation appeal text should be as sufficient as possible. Therefore, this embodiment can perform evidence source integrity verification on the medical insurance violation appeal text to verify whether the medical insurance violation appeal text covers the appeal evidence of all review dimensions.

[0167] For example, the system verifies whether the evidence materials in the appeal basis section of the medical insurance violation appeal text cover the four core types of evidence: "policy, medical records, cases, and knowledge". The first verification result is obtained under the evidence source completeness dimension. If the first verification result indicates that a certain type of evidence is missing (such as no case is cited), the system will automatically generate a prompt message: "Evidence in the xx dimension is missing and needs to be supplemented by retrieval".

[0168] Dimension Two: Entity Association Consistency. That is, this embodiment can perform entity association consistency verification on medical insurance violation appeal texts to verify whether the entities in the medical insurance violation appeal texts can be found in the evidence graph.

[0169] For example, the consistency between the entity descriptions in the medical insurance violation appeal text and the entity descriptions in the evidence map is verified to obtain the second verification result under the entity association consistency dimension. If the second verification result indicates that a certain entity description in the medical insurance violation appeal text is inconsistent with the evidence map, such as "type 2 diabetes" not being mapped to the ICD-10 code "E11.9", then the prompt message "Please supplement the ICD code annotation to ensure consistency with the description of the entity ENT_00456 (ENT_00456 is the entity identifier) ​​in the map" is automatically generated.

[0170] Dimension 3: Reasonableness of Rule Weights. That is, this embodiment can verify the reasonableness of rule weights in medical insurance violation appeal texts to check whether the proportion of appeal evidence for each review dimension in the medical insurance violation appeal text matches the review priority weight.

[0171] For example, if the description of the appeal evidence in the case evidence dimension exceeds the proportion of the policy basis dimension in terms of length, an automatic warning message will be generated: "The weight of case evidence accounts for 35%, which exceeds the standard threshold of 20%, and the expression of policy basis needs to be optimized and strengthened."

[0172] Dimension Four: Temporal Logical Consistency. That is, this embodiment can perform temporal logical consistency verification on medical insurance violation appeal texts to verify whether the temporal logic of time-related entities in the medical insurance violation appeal texts is reasonable.

[0173] Considering that medical insurance policies may be updated, the medical insurance application and review process needs to be based on the policy in effect at the time of the illness and treatment of the target medical record. This means the policy's effective date must precede the treatment date. Additionally, other temporal logic needs to be ensured to be reasonable, such as the treatment date preceding the testing date. Therefore, a time logic self-consistency rule can be set to verify the reasonableness of the time logic of time-related entities in the medical insurance violation appeal text.

[0174] For example, if the policy takes effect on January 1, 2024, and the treatment date is December 10, 2023, a prompt will be automatically generated stating "The policy does not cover the treatment date, and the 2023 version of the medical insurance catalog terms need to be replaced."

[0175] Optionally, the medical insurance violation appeal text can be corrected based on the above verification results to obtain the corrected medical insurance violation appeal text.

[0176] For example, if the evidence source integrity verification process finds that the medical insurance violation appeal text lacks case-based supporting evidence, it can automatically trigger a GraphRAG supplementary search, update the appeal basis based on the supplementary searched evidence, and output the corresponding update message, such as "No similar case-based supporting evidence was detected. We have re-searched for cases related to 'thrombocytopenia after lung cancer chemotherapy' and added a new case CASE_20250112 (appeal successful)".

[0177] Considering that some verification results can be automatically corrected, while others require manual intervention, this embodiment can optionally generate a link integrity verification report based on the verification results, marking passed items, warning items, and error items. Warning items refer to items that require manual confirmation, and error items refer to items that require mandatory correction. For example, appeal texts can only be submitted after all error items have been corrected.

[0178] In this embodiment, the four-dimensional verification rules can be used to automatically verify whether the evidence chain of the appeal content is complete and whether the logic is self-consistent. This allows for timely correction when errors are found, avoiding the submission of incorrect medical insurance violation appeal texts and preventing the appeal from failing.

[0179] It is understandable that the imperfections of evidence graphs, evidence retrieval models (used to implement the appeal evidence chain retrieval process in step S102 above, such as GraphRAG retrieval models, or BioELECTRA (Pretrained Biomedical Text Encoder using Discriminators) models), appeal templates, and large medical models may prevent the generation of reliable appeal texts through automated verification and correction.

[0180] For example, if the evidence map contains entity errors, relationship errors, or unreasonable weighting, it may result in the inability to obtain a complete and accurate chain of evidence for the appeal, thus rendering the appeal text unusable.

[0181] Therefore, this embodiment provides the following update process.

[0182] First, expert-driven updates to the evidence map (passive updates).

[0183] In this embodiment, the evidence graph can be manually or automatically analyzed based on the verification results of the medical insurance violation appeal text to identify errors such as entity errors, relationship errors (e.g., "mistakenly associating 'hypertension' with 'chemotherapy complications'"), and unreasonable weights (e.g., "a certain appeal case highly related to medical insurance violation materials was not prioritized for retrieval due to the low weight of the associated edge").

[0184] Meanwhile, dynamic changes in evidence data such as medical insurance policies may lead to errors in certain entities or relationships in the evidence graph, or unreasonable edge weights. These errors can be identified through feedback from medical insurance experts and clinicians.

[0185] To correct the aforementioned errors, experts can provide correction suggestions. Therefore, this embodiment can obtain error correction suggestions corresponding to the errors in the evidence graph, and then update the evidence graph based on these suggestions to obtain an updated evidence graph. Optionally, the error correction suggestions may include: entity identifiers (such as entity IDs), relationship identifiers (such as relationship IDs), and correction suggestions.

[0186] For example, if there is an error in an entity or relationship, a partial update can be triggered, updating only the erroneous node and its associated edges; if the weights are unreasonable, the edge weight calculation strategy can be adjusted to recalculate the edge weights.

[0187] To verify the effectiveness of the updated evidence map, this embodiment can obtain a test set, which includes several (e.g., 100) medical insurance violation materials and corresponding appeal text tags. Then, based on the medical insurance violation materials in the test set and the updated evidence map, an appeal text generation test can be performed (the specific process is described in steps S202-S204 above). The medical insurance violation appeal texts obtained in the test are then compared with the appeal text tags to obtain the evidence retrieval accuracy rate of the medical insurance violation appeal texts obtained in the test compared to the appeal text tags in the test set.

[0188] If the evidence retrieval accuracy is greater than or equal to the preset accuracy threshold (e.g., 5%), the updated evidence map will be used to generate subsequent medical insurance violation appeal texts.

[0189] Conversely, if the evidence retrieval accuracy is less than the accuracy threshold, the rationality of the error correction suggestions is analyzed retrospectively and unreasonable parts are corrected to obtain updated error correction suggestions. The evidence map is then updated based on the error correction suggestions to obtain the updated evidence map.

[0190] That is, if the evidence retrieval accuracy is low, it may be that the error correction suggestion information provided above is incorrect. In this case, the error correction suggestion information can be corrected to obtain the updated error correction suggestion information. Then, the evidence map in step S202 is updated again, and so on, forming a closed loop iteration.

[0191] Optionally, to prevent the iteration from getting stuck in an infinite loop, a maximum iteration count threshold can be set so that the iteration loop will exit in time when the maximum iteration count threshold is reached.

[0192] Optionally, after exiting the loop, a prompt message can be generated, such as "The evidence map may contain errors and requires further review," so that human intervention can be carried out in a timely manner to avoid the use of erroneous evidence maps in the future, which could lead to inaccurate evidence chains in the appeal.

[0193] Second, the adaptive iterative evidence graph, evidence retrieval model, appeal template, and medical big data model are automatically updated.

[0194] To enable the self-learning and updating of the evidence graph, evidence retrieval model, appeal template, and medical big data model, this embodiment can also form a dynamic optimization link based on the accumulation of historical appeal results and manual feedback, thereby achieving automatic optimization of the evidence graph, medical big data model, appeal template, and evidence retrieval model.

[0195] Specifically, this embodiment can obtain a training problem set, and each training problem in the training problem set includes: training medical insurance violation appeal text, a problem description of the problems existing in the training medical insurance violation appeal text, and a problem type label to which the problem description belongs.

[0196] Here, the problem description of the training medical insurance violation appeal text refers to the review result provided by the appeal reviewer after the appeal is submitted based on the training medical insurance violation appeal text, which describes the problems existing in the training medical insurance violation appeal text (such as the policy basis provided during diagnosis and treatment not being in effect).

[0197] In this embodiment, the above-mentioned problem descriptions can be automatically tagged, and the tag types include but are not limited to: missing search associations, missing fields in generated text, and logical errors.

[0198] If the problem description is "The use of the drug 'cefopiperidine-sulbactam' should not be based solely on the diagnosis of 'lung infection', but should also be based on clear evidence of etiology or indications of severe illness", this problem description suggests that the atlas may have only established a simple link "lung infection -> can be used -> cefoperazone-sulbactam", but lacks deeper restrictive relationships. Therefore, this problem description can be labeled as "missing association in search".

[0199] Of course, there may be new problem descriptions without preset rules, such as the first occurrence of "mismatch between rare disease medical insurance policy". In this case, it can be pushed to the expert terminal for manual annotation (according to statistics, manual annotation takes ≤5 minutes / item).

[0200] After obtaining the training question set, the following automatic iteration can be performed based on the question type labels in the training question set (automatic iteration can be triggered periodically or when there is frequent human feedback):

[0201] First, select the first training problem subset from the training problem set, using problem type labels as the first training problem subset to retrieve missing associations. Based on the problem descriptions in the first training problem subset, determine the new associations that were missed in the evidence graph and update the evidence graph with the new associations. This update includes adding the new associations to the graph and adjusting the corresponding edge weights.

[0202] Second, when using an evidence retrieval model in the process of retrieving the evidence chain of the appeal, the training medical insurance violation appeal texts in the first training question subset are updated according to the new association, and the updated texts are used as training labels. The first training medical insurance violation materials corresponding to the training labels are obtained, and the evidence retrieval model is trained based on the first training medical insurance violation materials and the training labels.

[0203] In other words, this embodiment can use the first training medical insurance violation materials and corresponding training labels of the missing association type to fine-tune the evidence retrieval model in order to improve the semantic matching accuracy of similar entities (entities corresponding to new associations).

[0204] Third, select a second training problem subset from the training problem set whose problem type labels are missing fields in the generated text. Based on the problem descriptions in the second training problem subset, determine the target fields missing in the appeal template and add the target fields to the appeal template.

[0205] Specifically, in this embodiment, the missing target field in the appeal template can be determined based on the problem description in the second training problem subset, and then placeholders for the target field can be added to the appeal template to obtain the updated appeal template.

[0206] Fourth, a third subset of training problems labeled with logical errors is selected from the training problem set. The training medical insurance violation appeal texts in the third training problem subset are used as error texts. The correct texts corresponding to the error texts and the second training medical insurance violation materials corresponding to the error texts are obtained. Based on the second training medical insurance violation materials, the correct texts, and the error texts, the large medical model is fine-tuned.

[0207] Here, the correct text corresponding to the erroneous text can be automatically generated by a more accurate large language model based on the erroneous text, or it can be provided manually; this application does not impose any specific limitations.

[0208] Optionally, after iteration, data can be automatically and randomly selected as a test set to verify the effectiveness of the updated evidence graph, evidence retrieval model, medical big data model, and appeal template. Update events can also be recorded in the log or a new version can be generated (forming a version convention function).

[0209] It should also be noted that in practical applications, any one or more of the above-mentioned automatic iteration methods can be implemented, and this application does not impose any specific limitations.

[0210] Optionally, iterative updates can be performed according to the entire process of "feedback collection -> data annotation -> model iteration -> effect verification -> map update", without the need for repeated manual intervention, thus achieving continuous optimization of the system's retrieval accuracy and generation quality.

[0211] The above describes a method for generating appeal text according to embodiments of this application. The following describes the apparatus for performing the above-described method for generating appeal text.

[0212] Please see Figure 5 , Figure 5 This is a schematic diagram of an appeal text generation device provided in an embodiment of this application. Figure 5 As shown, the appeal text generation device may include:

[0213] The data acquisition unit 501 is used to acquire medical insurance violation materials of the target medical record. The medical insurance violation materials are materials describing the medical insurance violations that exist in the target medical record.

[0214] The data retrieval unit 502 is used to retrieve the appeal evidence chain related to medical insurance violation materials from the pre-constructed evidence graph. The evidence graph is a knowledge graph constructed based on evidence data under multiple review dimensions based on the medical insurance violation review process.

[0215] The data decomposition unit 503 is used to decompose the appeal evidence chain according to the review dimensions to obtain appeal evidence of at least one review dimension.

[0216] Data generation unit 504 is used to generate medical insurance violation appeal text based on medical insurance violation materials and appeal evidence from at least one audit dimension, using a medical big model fine-tuned with medical data.

[0217] Each module in the aforementioned appeal text generation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0218] This application also provides an electronic device, which may include at least one processor and a memory connected to the processor, wherein:

[0219] Memory is used to store computer programs;

[0220] The processor is used to execute computer programs to enable the electronic device to implement any of the appeal text generation methods provided in the embodiments of this application.

[0221] refer to Figure 6 The diagram illustrates a structural schematic suitable for implementing the electronic device in the embodiments of this application. The electronic device in the embodiments of this application may include, but is not limited to, fixed terminals such as mobile phones, laptops, PDAs (personal digital assistants), PADs (tablet computers), desktop computers, etc. Figure 6 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0222] like Figure 6 As shown, the electronic device may include a processing unit (e.g., a central processing unit, a graphics processing unit, etc.) 601, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 602 or a program loaded from a storage device 608 into a random access memory (RAM) 603. When the electronic device is powered on, the RAM 603 also stores various programs and data required for the operation of the electronic device. The processing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0223] Typically, the following devices can be connected to I / O interface 605: input devices 606 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 607 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 608 including, for example, memory cards, hard drives, etc.; and communication devices 609. Communication device 609 allows electronic devices to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 6 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown. More or fewer devices may be implemented or have alternatively.

[0224] This application also provides a computer program product including computer-readable instructions, which, when executed on an electronic device, cause the electronic device to implement any of the appeal text generation methods provided in this application.

[0225] This application also provides a computer-readable storage medium that carries one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the appeal text generation methods provided in this application.

[0226] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and 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. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0227] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, 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 is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0228] In the above embodiments, the implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, in the form of a computer program product.

[0229] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

Claims

1. A method for generating appeal text, characterized in that, include: Obtain medical insurance violation materials for the target medical record, wherein the medical insurance violation materials are materials describing the medical insurance violations that exist in the target medical record; Retrieve the appeal evidence chain related to the medical insurance violation materials from the pre-constructed evidence graph, which is a knowledge graph constructed based on evidence data from multiple review dimensions on which the medical insurance violation review process is based; The chain of evidence for the appeal is broken down according to the review dimensions to obtain appeal evidence for at least one review dimension; A medical big data model fine-tuned using medical data is used to generate a medical insurance violation appeal text based on the medical insurance violation materials and the appeal evidence from at least one review dimension.

2. The appeal text generation method according to claim 1, characterized in that, The medical big data model, fine-tuned using medical field data, generates a medical insurance violation appeal text based on the medical insurance violation materials and the appeal evidence from at least one review dimension, including: Obtain the review priority weight for each of the review dimensions, and using the medical big data model, generate the medical insurance violation appeal text based on the medical insurance violation materials, the appeal evidence for at least one review dimension, and the review priority weight. The review priority weight is used to determine the proportion of the corresponding appeal evidence in the medical insurance violation appeal text.

3. The method for generating appeal text according to claim 1 or 2, characterized in that, The medical insurance violation materials include the medical insurance violation notice and the original medical record corresponding to the medical insurance violation notice; Before retrieving the chain of evidence for the appeal related to the medical insurance violation materials from the pre-constructed evidence map, the method further includes: The aforementioned medical big data model is used to extract violation elements related to medical insurance violations from the medical insurance violation notices. The local audit engine is used in conjunction with the original medical record to conduct a local audit of the violation elements, resulting in local audit results and audit process data. The aforementioned violation elements, the local violation review results, and the review process data are used as the processed medical insurance violation materials. The step of retrieving the chain of evidence for appeals related to the medical insurance violation materials from the pre-constructed evidence map includes: Retrieve the chain of evidence for appeals related to the processed medical insurance violation materials from the evidence map.

4. The appeal text generation method according to claim 3, characterized in that, The process of using a local auditing engine in conjunction with the original medical record to perform local violation audits on the violation elements, obtaining local violation audit results and audit process data, includes: The local review engine is used to obtain the local rules pointed to by the violation elements. Based on the local rules and / or preset regular expressions, medical record fragments related to the violation elements are matched from the original medical record text. The local violation review result is determined based on the local rules and the medical record fragments. The local rules are logical judgment rules derived from existing medical policy norms. The local rules and the medical record fragments are used as the data for the review process.

5. The method for generating appeal text according to claim 1 or 2, characterized in that, The process of breaking down the chain of evidence for the appeal according to review dimensions yields appeal evidence for at least one review dimension, including: Data related to medical insurance policies are extracted from the aforementioned chain of evidence for the appeal, serving as evidence for the appeal based on policy principles. And / or, extract the correspondence between medical record entities and diagnosis and treatment relationships from the chain of evidence for the appeal, sort and map the correspondence according to the timeline of diagnosis first, treatment then test, and perform policy numerical matching analysis on the diagnosis and treatment facts obtained from the sorting and mapping to obtain fact matching data, which serves as the appeal evidence for the fact matching dimension. And / or, extract appeal cases from the appeal evidence chain as appeal evidence for the case corroboration dimension; And / or, extract medical knowledge from the aforementioned chain of evidence for the appeal as evidence for the appeal in the dimension of clinical consensus.

6. The appeal text generation method according to claim 1, characterized in that, The medical big data model, fine-tuned using medical field data, generates a medical insurance violation appeal text based on the medical insurance violation materials and the appeal evidence from at least one review dimension, including: Obtain a pre-configured appeal template, which includes at least the error in the violation determination, the basis for the appeal, and the final request; The medical big data model is used to generate specific content of the violation judgment error points based on the medical insurance violation materials, to generate specific content of the appeal basis based on the appeal evidence of at least one review dimension, and to generate specific content of the final appeal based on the specific content of the violation judgment error points and the appeal basis, so as to obtain the medical insurance violation appeal text corresponding to the appeal template.

7. The method for generating appeal text according to claim 6, characterized in that, The violation determination error points include the source of the violation determination, the type of error determination, and the initial rebuttal direction. The source of the violation determination is used to describe the determining subject and number, the type of error determination is used to describe the specific violation item determined, and the initial rebuttal direction is used to describe the determination error point. And / or, the basis for the appeal includes a description of the appeal evidence for each review dimension, and the description consists of three parts: the source of the evidence, the content of the evidence, and the relationship between the content of the evidence and the error. And / or, the appeal template also includes a reasoning chain visualization diagram, which is a flowchart composed of the correlation logic between the specific content of the violation judgment error point, the appeal basis, and the final request.

8. The method for generating appeal text according to claim 1, characterized in that, Also includes: The medical insurance violation appeal text is subjected to at least one of the following checks: evidence source integrity check, entity association consistency check, rule weight reasonableness check, and time logic self-consistency check, and the check result is obtained. Specifically, the evidence source integrity check is used to verify whether the medical insurance violation appeal text covers all the appeal evidence of all review dimensions; the entity association consistency check is used to verify whether the entities in the medical insurance violation appeal text can be found in the evidence graph; the rule weight reasonableness check is used to verify whether the proportion of appeal evidence of each review dimension in the medical insurance violation appeal text matches the review priority weight; and the time logic self-consistency check is used to verify whether the time logic of time-related entities in the medical insurance violation appeal text is reasonable.

9. The method for generating appeal text according to claim 1, characterized in that, Also includes: Obtain error correction suggestions corresponding to the error information in the evidence graph; The evidence map is updated based on the error correction suggestion information to obtain the updated evidence map; Obtain a test set, and conduct an appeal text generation test based on the medical insurance violation materials in the test set and the updated evidence map. Determine the evidence retrieval accuracy of the medical insurance violation appeal text obtained in the test compared with the appeal text tags in the test set. If the accuracy rate of the evidence retrieval is greater than or equal to the preset accuracy threshold, the updated evidence map will be used to generate medical insurance violation appeal text in the future. If the evidence retrieval accuracy is less than the accuracy threshold, the rationality of the error correction suggestion information is analyzed backtracking and unreasonable parts are corrected to obtain updated error correction suggestion information. Then, the evidence map is updated according to the error correction suggestion information to obtain the updated evidence map.

10. The method for generating appeal text according to claim 1, characterized in that, Also includes: Obtain a training problem set, wherein each training problem in the training problem set includes: a training medical insurance violation appeal text, a problem description of the problems existing in the training medical insurance violation appeal text, and a problem type label to which the problem description belongs; The problem type labels are selected from the training problem set to retrieve the first training problem subset that has been missed in association. Based on the problem descriptions in the first training problem subset, the new associations that have been missed in the evidence graph are determined, and the new associations are updated to the evidence graph. And / or, when the evidence retrieval model is used in the retrieval process of the appeal evidence chain, the training medical insurance violation appeal texts in the first training question subset are updated according to the new association, and the updated texts are used as training labels. The first training medical insurance violation materials corresponding to the training labels are obtained, and the evidence retrieval model is trained according to the first training medical insurance violation materials and the training labels. And / or, filter the problem type labels from the training problem set to form a second training problem subset that generates text with missing fields, determine the target field missing in the appeal template based on the problem description in the second training problem subset, and add the target field to the appeal template; And / or, select a third subset of training problems labeled with logical errors from the training problem set, take the training medical insurance violation appeal texts in the third training problem subset as error texts, obtain the correct texts corresponding to the error texts and the second training medical insurance violation materials corresponding to the error texts, and fine-tune the medical big model based on the second training medical insurance violation materials, the correct texts and the error texts.

11. A computer program product, characterized in that, It includes computer-readable instructions that, when executed on an electronic device, cause the electronic device to implement the appeal text generation method as described in any one of claims 1 to 10.

12. An electronic device, characterized in that, It includes at least one processor and a memory connected to the processor, wherein: The memory is used to store computer programs; The processor is used to execute the computer program to enable the electronic device to implement the appeal text generation method as described in any one of claims 1 to 10.

13. A computer storage medium, characterized in that, The storage medium carries one or more computer programs that, when executed by an electronic device, enable the electronic device to implement the appeal text generation method as described in any one of claims 1 to 10.

Images