An intelligent agent for medical insurance audit and a medical insurance audit management system

Abstract

The application discloses a medical insurance audit intelligent agent and a medical insurance audit management system, and belongs to the technical field of medical insurance audit, and specifically can realize the following steps: obtaining multi-source heterogeneous data and performing data cleaning to obtain document data; based on a medical insurance audit knowledge base, meta data related to rules is screened and summarized from the text data; a large language model performs reasoning based on audit prompt words integrated with the medical insurance audit knowledge base and the meta data and a CoT reasoning mechanism, and outputs an audit result. The medical insurance audit intelligent agent is used for implementing document quality control on medical records in the diagnosis and treatment process, and is also used for the audit and complaint process after medical insurance settlement, realizing real-time reminding intervention and post-complaint feedback on medical insurance settlement, and improving the accuracy and timeliness of medical insurance audit.

Description

Technical Field

[0001] This invention belongs to the field of medical insurance review technology, specifically relating to a medical insurance review intelligent agent and a medical insurance review management system. Background Technology

[0002] Under stringent regulatory policies, hospital medical insurance management faces significant challenges. Firstly, the manual review model is unsustainable, specifically manifested in the following ways: (1) Low review efficiency: There are limited medical insurance review personnel. Faced with a massive number of medical records, the coverage of manual spot checks is extremely low, making it impossible to achieve full review. A large number of violations are only exposed after settlement.

[0003] (2) High professional threshold: Medical insurance reviewers need to be proficient in medical insurance policies, clinical medicine and medical records at the same time. The training cycle is long and the talent shortage problem is prominent.

[0004] (3) The cost of correcting errors after the fact is high: Once the expenses are settled and deducted by the medical insurance bureau, the recovery process is complicated and the financial loss is difficult to recover.

[0005] Secondly, the hospital urgently needs internal quality control, specifically manifested in the following ways: (1) Reduce illegal deductions: In order to avoid medical insurance deductions, hospitals have a strong incentive to conduct internal quality control before reporting expenses, so as to discover and intervene in illegal issues in the hospital as early as possible. (2) Improving medical quality: Compliance review is inherently unified with the standardization of clinical pathways and the improvement of medical record quality. Using AI quality control to drive the standardization of medical record writing and diagnosis and treatment behavior is an inevitable requirement for achieving refined hospital management.

[0006] Furthermore, existing traditional auditing methods are inadequate to meet the demands of modern regulation. This mainly includes: (1) Blind spots in structured data review: Traditional rule engines can only process structured data such as ICD diagnostic codes and billing item codes. However, most clinical decision-making is based on unstructured text (such as medical records, surgical records, and examination reports). This leads to the inability to identify deeper violations, such as correct diagnostic codes but lack of support in medical records or a disconnect between treatment and diagnostic logic.

[0007] (2) Lack of semantic understanding ability, unable to achieve connotation quality control: The system cannot understand the meaning of medical texts. For example, the rule engine can only check whether there is a surgical code for the drug "limited to surgical patients". However, AI quality control can determine whether the surgical process and the necessity of postoperative medication are recorded in detail in the medical record. The rule engine cannot identify that laparotomy and open laparotomy are the same operation, which leads to misjudgment that the surgical record and the billing items are inconsistent.

[0008] (3) The problem of data silos is serious and cross-validation is impossible: Patients' medical information is scattered across dozens of systems such as HIS, EMR, LIS, and PACS, and traditional technologies are unable to achieve cross-modal consistency verification. For example, it is impossible to automatically detect the typical problem of "the surgical record does not mention the use of a certain high-value consumable, but the billing list has been charged". Summary of the Invention

[0009] In view of the above, the purpose of this invention is to provide a medical insurance review intelligent agent and a medical insurance review management system. This medical insurance review intelligent agent can process multi-source data of various types, and, by combining a clinical medical knowledge base and a medical insurance policy rule base, performs medical insurance review through large language model reasoning. Furthermore, this medical insurance review intelligent agent can be used during the diagnosis and treatment process to implement document quality control of medical records, and also in the review and appeal process after medical insurance settlement, enabling real-time reminders and interventions during the medical insurance settlement process and providing feedback on post-appeals, thereby improving the accuracy and timeliness of medical insurance review.

[0010] To achieve the above-mentioned objectives, an embodiment provides a medical insurance review intelligent agent, comprising the following steps: Obtain multi-source heterogeneous data and perform data cleaning to obtain document data; Based on the medical insurance review knowledge base, metadata related to the rules is aggregated and filtered from text data. The large language model performs reasoning based on the review prompt words that are integrated into the medical insurance review knowledge base and metadata, and incorporates the CoT (CoT) reasoning mechanism to output the review results.

[0011] Preferably, the multi-source heterogeneous data comes from the Hospital Information System (HIS), Electronic Medical Record System (EMR), Laboratory Information System (LIS), Image Archiving System (PACS), Communication System (PACS), and Medical Record Homepage System, and is obtained from the above systems using the distributed scheduling platform DolphinScheduler and the heterogeneous data source synchronization tool DataX.

[0012] Preferably, the step of summarizing and filtering rule-related metadata from text data based on the medical insurance review knowledge base includes: For patient test and medical order data, the Reranker model is used to match the rules in the medical insurance review knowledge base with the patient test and medical order data to perform relevance retrieval scoring, and only data with high scores are adopted as metadata related to the rules.

[0013] Preferably, the metadata related to the rules is aggregated and filtered from the text data based on the medical insurance review knowledge base, including: For medical records and examination data, the non-thinking mode of LLM is adopted to match the rules in the medical insurance audit knowledge base with the medical records and examination data to perform relevance scoring, and only the data with high scores is adopted as metadata related to the rules.

[0014] Preferably, the large language model performs reasoning based on audit prompt words integrated into the medical insurance audit knowledge base and metadata, and incorporates a CoT (CoT) reasoning mechanism, including: The medical insurance review knowledge base and the selected metadata are used to construct dynamic review prompts, which are then input into a large language model. The large language model performs CoT reasoning based on the review prompts and outputs the review results, which include compliance and violation evidence chains.

[0015] Preferably, when the large language model performs reasoning based on the audit prompt words and incorporates the CoT (CoT) reasoning mechanism, it adopts batch processing and dynamically adjusts the batch size.

[0016] To achieve the above-mentioned objectives, this invention also provides a medical insurance review and management system, characterized in that the medical insurance review intelligent agent is used for clinical medical insurance decision support, that is, to perform real-time document quality control on medical records acquired in real time, and generate review results to remind medical workers in real time to provide decision support.

[0017] Preferably, the aforementioned medical insurance auditing intelligent agent is used for intelligent auditing after medical insurance settlement. During the auditing process, it performs quality control on the medical insurance settlement statement and generates audit results to provide a basis for appeals.

[0018] Preferably, the review results are visualized as document quality control prompts, including the text source, review basis and policy requirements, and providing options for whether to review an appeal and whether to pay for it out of pocket.

[0019] Preferably, the text reviewed by the medical review intelligent agent is automatically distinguished and marked, and supports copying.

[0020] Compared with the prior art, the beneficial effects of the present invention include at least the following: To address the unique characteristics of medical documents, a parsing framework combining named entity recognition, semantic vectorization, and hybrid sorting was adopted. This framework effectively solves the problems of complex semantics and diverse terminology in medical texts, enabling accurate parsing of document data and providing a reliable data foundation for subsequent review and reasoning using large language models.

[0021] We designed a dedicated prompt word template for the medical insurance supervision field. This template incorporates the requirements of medical insurance supervision policies, making the audit results output by the large language model more in line with business needs.

[0022] The constructed intelligent medical insurance audit system can be applied to clinical medical insurance decision support and intelligent audit after medical insurance settlement, forming a three-in-one regulatory system of intelligent understanding, knowledge collaboration and process intervention. This promotes the transformation of the regulatory model from passive error correction to proactive early warning and source blocking, and improves the efficiency and accuracy of medical insurance audit. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart of the method steps for implementing the medical insurance review intelligent agent provided in the embodiment; Figure 2 This is a visual example of in-process AI document quality control in clinical settings provided in the implementation example; Figure 3 This is a visual example of AI-powered document quality control in the management terminal provided in the implementation example; Figure 4 This is a visual example of AI-powered intelligent appeals provided in the implementation example. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and do not limit the scope of protection of this invention.

[0026] The embodiment provides a medical insurance review intelligent agent, which can understand various complex semantics in the medical field and combine them with a clinical medical knowledge base and a medical insurance policy rule base, and use the powerful reasoning ability of a large language model to achieve accurate review of medical insurance data.

[0027] like Figure 1 As shown, the intelligent medical insurance review agent can perform the following steps: S1: Obtain heterogeneous data from multiple sources and perform data cleaning to obtain document data.

[0028] The heterogeneous data used for medical insurance review originates from multiple medical systems, including but not limited to Hospital Information System (HIS), Electronic Medical Record (EMR), Laboratory Information System (LIS), image archiving, Picture Archiving System (PACS), and medical record front page system. Specifically, the data acquisition process utilizes a distributed scheduling platform, DolphinScheduler, and a heterogeneous data source synchronization tool, DataX, to integrate multimodal data such as patient basic information, structured cost data, and unstructured medical record texts (admission records, progress notes, and surgical records). After data acquisition, automatic data cleaning is performed to remove sensitive data and adjust formatting, resulting in document data.

[0029] S2, based on the medical insurance review knowledge base, summarizes and filters metadata related to the rules from text data. The large language model performs reasoning based on the review prompt words integrated with the medical insurance review knowledge base and metadata and incorporates the CoT reasoning mechanism to output the review results.

[0030] In this example, to verify the medical insurance validity of text data, the first step is to filter metadata related to the current rules from medical records, examinations, tests, and medical orders, i.e., to remove redundant data. Specifically, this includes: First, a medical insurance review knowledge base is constructed. This knowledge base includes rules that abstract information such as diagnoses, surgeries, and drugs that the rules depend on. Specifically, this information can be presented as a list, with each entry containing the rule name, rule content, and rule requirements. This knowledge base serves as the basis for medical insurance review, used to filter metadata related to medical insurance rules to eliminate redundant data.

[0031] Secondly, for factual data such as patient test results and medical orders, the Reranker model is used to match the rules in the medical insurance audit knowledge base with the patient test results and medical orders to perform relevance retrieval scoring, and only data with high scores are adopted as metadata related to the rules.

[0032] For medical records, examination data, etc., the non-thinking mode of LLM is adopted to match the rule content in the medical insurance review knowledge base with the medical records and examination data to perform relevance scoring, and only the data with high scores is adopted as metadata related to the rules.

[0033] Finally, the filtered metadata is integrated, and the CoT (Cooperation of Knowledge) of medical insurance review is incorporated into the prompts of the large language model. This makes the inference output interpretable review results, which include compliance and violation evidence chains. This approach breaks through the limitations of traditional rule engines and achieves intrinsic quality control over the rationality of medical behavior, that is, medical insurance review is carried out starting from the connotation of medical records.

[0034] The review prompts incorporate metadata and the medical insurance review knowledge base to construct dynamic prompts. Specifically, this can be achieved by: You are an omniscient and all-knowing clinical medical expert, tasked with helping the hospital recover losses from medical insurance violations while adhering to legal and compliant regulations. I need you to perform a professional analysis based on the following structured input. Here is a brief description of the task: 1. Task Content (1) The hospital received a violation notice regarding the charge item: [{{feeItem}}] (hereinafter referred to as "the item under review") + [{{feetime}}] (hereinafter referred to as "the time of occurrence of the item under review"). I will give you the medical records of the patient who violated the regulations later. Please analyze the medical records of this patient and analyze whether the violation content is reasonable. (2) The following is a breakdown of the medical insurance department's restrictions on payment for the reviewed item (hereinafter referred to as "medical insurance restrictions"). Please strictly follow each sub-rule in your analysis and answer in the prescribed format: Medical records (including at least one of the following): admission record, progress notes, doctor's orders, examination / lab reports, discharge summary, electronic medical record, medical record cover sheet, settlement statement, outpatient medical record, and detailed expense list. Medical record structure: Medical records will be provided in standard JSON format. The following is an example of the input parameters: [ 'Document Category 1': [{'Source': 'Document Title 1', 'Content': 'Document Content 1', 'Document ID': '1'},{'Source': 'Document Title 2', 'Content': 'Document Content 2', 'Document ID': '2'},...], 'Document Category 2': [{'Source': 'Document Title 3', 'Content': 'Document Content 3', 'Document ID': '3'},{'Source': 'Document Title 4', 'Content': 'Document Content 4', 'Document ID': '4'},...],... 'Inspection Results': [{'Inspection Findings': 'Inspection Findings 1','Inspection Impressions': 'Inspection Impressions 1','Inspection Items': 'Item Name 1','Document ID': '5',''},...], 'Test Results': [...], 'Hospitalization orders': [...], 'Outpatient Prescription': [...], 'Diagnosis':{'Current Visit':[...],'Patient History':[...]}, 'Basic Information': [...], ] 2. Core Principles of Review (1) Reversal of the burden of proof: The hospital bears the burden of proof for the compliance of the audited item; if the key conditions are not recorded, it is considered that the conditions are not met (not "cannot be determined"); permissibility rule: not recorded = not met = non-compliant. (2) In case of doubt, the hospital shall be deemed to have no evidence: Only when there is **directly conflicting objective evidence** in the medical records (there is both compliant and non-compliant evidence) can a compliance judgment be made in favor of the hospital. (3) Principle of reasonable default diagnosis: all diagnoses in the medical record are assumed to be valid and established, and the correctness of the diagnosis itself is prohibited from being questioned or reviewed; (4) Project use default principle: The project being reviewed is already in use by default, and the review of "whether it is in use" is prohibited.

[0035] 3. Analysis Process (1) Solve the rules: Clarify the logical relationship of all conditions in the [Medical Insurance Limitation Rules]: If the [Medical Insurance Limitation Rules] contain the statement "XX should meet any of the following conditions", it is an "OR" relationship; if the [Medical Insurance Limitation Rules] contain the statement "XX should meet the following conditions at the same time", it is an "AND" relationship; for conditions such as time and frequency, the calculation basis and nodes need to be clearly defined; it is forbidden to extend the limitation conditions on your own.

[0036] (2) Finding evidence: All judgments must be based on objective records in medical records; priority should be given to relevant records of the current visit as evidence. When there is no evidence for the current visit record, historical visit records can be used as evidence (limited to chronic disease-related diagnosis and treatment information); the examination / test report used to support compliance judgment must be earlier than the occurrence time of the audited item, otherwise it cannot be used as valid evidence. For "frequency" violation items, the actual execution records (such as medical records and operation records) need to be checked; systematically search medical records: the complexity of the medical record structure cannot be used to determine that it is "not recorded".

[0037] 4. Conclusion: (1) Rule of law: All judgments must be strictly based on the [Medical Insurance Restriction Rules] as the sole criterion, and personal medical knowledge or speculation outside the rules is prohibited; (2) Prohibition of indirect speculation: It is prohibited to use indirect evidence as the main basis for judgment, unless the [Medical Insurance Restriction Rules] explicitly allow inference based on such evidence; (3) Priority of direct evidence: When both direct and indirect evidence exist, the judgment shall be based on the direct evidence in the medical records. (4) Priority of confirmed diagnosis: When both confirmed and suspected diagnoses exist in the medical record, the confirmed diagnosis should be used as the main basis for judging whether it meets the rules; when there is no clear confirmed diagnosis in the medical record, the suspected diagnosis (such as "XX to be investigated", "Suspected XX", "Possible XX", "XX?") can be used as valid evidence. (5) Key points of review for examination and testing items: For the [reviewed items] of examination and testing, the core of the review is whether there are objective indicators in the medical record that conform to the application scenario of the [medical insurance restriction rules] (e.g., specific symptoms, signs, diagnoses, preoperative routines, other auxiliary examination results, etc.). It is prohibited to review whether the purpose or reason for issuing the item has been fully recorded. (5) Prohibition of circular reasoning: It is prohibited to use abnormal results of the audited item (examination and testing items) as the basis for the compliance of the audited item's charges, and it is prohibited to use the use of the audited item (treatment items) as the basis for inferring that the patient has the indication for the audited item. Compliance must be proved by other evidence in the medical record that is independent of the audited item. (6) Principle of Judgment on Doubt: When the evidence is contradictory, leading to doubt about whether the rules are met, the principle of "presumption of innocence" must be followed, and an interpretation favorable to the hospital should be made, that is, a judgment of "compliance" should be made. (7) Consistency of total score: If it is an "or" condition, the whole is compliant if any sub-condition is met, and the final conclusion and judgment basis should be interpreted in a way that is favorable to the hospital; (8) Reverse restrictions are prohibited: The [item under review] itself cannot be used as a condition for determining whether a patient meets the restrictions of other treatment options.

[0038] 5. Source indicated: Each conditional statement must be clearly cited (e.g., "chief complaint section of admission record"); - the cited content must directly support compliance checks; the specific output specification is: please return a standard JSON object, in the following format: { "Conditional Analysis": [ { "Condition Number": "Condition 1", Original Policy Text: "Decomposition of Medical Insurance Restrictions" / / The analysis of the same condition can output a maximum of four most relevant conditions. "Clinical Evidence": { "Source": "The specific location of the clinical evidence (document name and title in the original `Source` section) and the original record date (omitting hours, minutes, and seconds). If there are multiple sources of clinical evidence, connect them with `+`, such as Admission Record (2000-01-01) + Examination Report (2000-01-01)". "Content": "Direct citations of medical records from the source must meet the principle of necessity and sufficiency." "Compliance": "Compliant / Uncompliant" "Numerical Basis":[{"Name":"Item 1", "Value":"Value 1"},{"Name":"Item 2", "Value":"Value 2"},...] , / / This should only be filled in if the rule is determined to be compliant by quantitative values ​​in an independent test report document (the clinical evidence's source contains the **test report**) and the conclusion is 'compliant'. Units are not allowed in the values; laboratory indicators and values ​​from outpatient and emergency medical records, admission records, discharge records, progress notes, etc., which are not from test reports, are prohibited. When the values ​​come from non-test reports or the conclusion is 'non-compliant', an empty array [] will be output.

[0039] "Key Evidence in Documents":[{"Document Evidence ID":"Document Evidence ID1", "Key Evidence":"Key Evidence 1"},{"Document Evidence ID":"Document Evidence ID2", "Key Evidence":"Key Evidence 2"},...] / / This is filled in when the rule determines compliance based on non-inspection report content (such as outpatient / emergency medical records, admission records, discharge records, progress notes, etc.) and the conclusion is "compliant"; Key evidence must be 100% derived from continuous string truncation in the original text (capable of precise keyword matching in the original text), and a single key evidence should be limited to 20 characters. Output an empty array [] when not applicable.

[0040] } } ], "Violation Judgment": { Final Conclusion: Compliant / Non-compliant "Judgment Basis": "The reasoning for the judgment derived after considering all conditions must cite specific evidence, and the chain of evidence must meet the principle of necessity and sufficiency." "Reasons for Appeal": "When the [Final Conclusion] is 'Compliant', list only the specific clinical evidence that meets the criteria, separated by commas (directly use key evidence from the conditional analysis, and refrain from including condition numbers, conclusive descriptions, sources of evidence, or process analysis); when the [Final Conclusion] is 'Uncompliant', briefly explain which conditions were not met and the reasons." Confidence level: High / Medium / Low } } Special attention should also be paid to: (1) Output consistency requirements: The comprehensive judgment conclusion must be logically consistent with the analysis results of each condition; avoid contradictions or logical jumps; (2) Standardization: The judgment of each medical insurance limitation condition must be based on the full medical record; it is strictly forbidden to fabricate clinical evidence and judgment basis, otherwise you will be laid off.

[0041] Please analyze the patient's medical records according to the requirements above and in the required format.

[0042] The large language model uses prompts and incorporates a CoT (Coding inference) mechanism to reason about the input content. During logical reasoning, the large language model also employs batch processing and dynamically adjusts the batch size. Specifically, each task in the batch processing task is pre-computed using tokens, and then sorted in ascending order based on the number of tokens to ensure that the requested token lengths are similar during batch processing. Access to the large model is scheduled according to concurrent threads. A vLLM inference engine is used, which implements continuous batch processing and pagination attention mechanisms to improve batch processing request speed.

[0043] By combining the aforementioned large language model with dual knowledge base collaborative decision-making, and further employing multimodal data fusion and consistency verification methods, this invention establishes a cross-modal logical association chain of "diagnosis-operation-fee-document record," and performs consistency verification and auditing on it. For example, it can identify hidden violations such as "the surgical record does not mention the use of a certain high-value consumable, but it has been billed in the expense details" or "the diagnosis name does not match the indications for the surgery / drug."

[0044] The embodiment also provides a medical insurance review and management system that applies the aforementioned medical insurance review intelligent agent. Specifically, it includes: The aforementioned intelligent medical insurance review agent enables in-process intervention within the medical insurance review management system, pioneering a real-time quality control application scenario. This involves directly embedding the intelligent medical insurance review agent as a review node into doctors' clinical workflows for clinical medical insurance decision support, allowing for real-time intervention before patient discharge. Specifically, it performs real-time document quality control on acquired medical records, generating review results to provide real-time reminders to medical professionals for decision support. These real-time prompts act like a readily available online medical insurance policy expert, assisting doctors in developing compliant treatment habits and improving the quality of medical care itself.

[0045] The aforementioned intelligent medical insurance review agent also implements post-event error correction in the medical insurance review management system, which is used in the application scenario of intelligent review after medical insurance settlement. During the review process, it performs quality control on medical insurance settlement statements and generates review results to provide a basis for appeals.

[0046] In both of these application scenarios, the focus of the review has shifted from the cost item itself to the underlying clinical practices and documentary evidence supporting that cost, thus improving the depth and quality of the review. It can also handle the review of violations in the following situations: (1) Detection of hidden violations: It can identify deep-seated problems that traditional rule engines cannot detect, such as ordering expensive examinations despite insufficient diagnostic evidence, and discrepancies between treatment records and surgical procedures. For example, even if the diagnosis code is correct, this invention can determine whether the description in the medical record is sufficient to support the diagnosis; (2) Semantic consistency verification: In the scenario of consistency between surgical charges and records, it can not only match keywords, but also understand medical synonyms (such as "exploratory laparotomy" and "exploratory laparotomy"), and determine whether the document records reflect the core medical behavior represented by the charge item; (3) Logical chain judgment: It can analyze the logical relationship between diagnosis-treatment-drug-consumables and judge its clinical rationality.

[0047] In this embodiment, the review results output by the large language model are visualized in real time as document quality control prompts, including the text source, review basis, and policy requirements, and provide options for whether to review appeals and whether to pay for them out of pocket. Figure 2 As shown, Figure 3 as well as Figure 4 As shown, auditors or medical professionals can clearly identify the problems, and medical professionals, in particular, can make corresponding adjustments and modifications in real time based on the audit results.

[0048] In this embodiment, the text reviewed by the medical review intelligent agent is automatically distinguished and marked, and supports copying, making it easy to view and edit.

[0049] The following example illustrates the review process of the medical review intelligent agent of this invention: Example 1: In-process review of the rationality of the use of "Nintedanib Ethylenediate Soft Capsules" Data Input: The doctor prescribes "nintedanib ethoxylate soft capsules" to the patient. The system captures this instruction in real time and automatically retrieves multimodal data such as the patient's current diagnosis, medical history, and imaging examination reports (e.g., HRCT) from the EMR system.

[0050] Semantic parsing and reasoning: Analyzing medical records. For example, a diagnosis record might be found to be "pulmonary fibrosis," but the imaging report describes it as "ground-glass opacities," without typical idiopathic pulmonary fibrosis (IPF) features such as "reticular" or "honeycomb" patterns. Knowledge base collaborative verification: The parsing results (diagnosis: pulmonary fibrosis; imaging: atypical IPF features) are compared with the medical insurance rules knowledge base (this drug is limited to IPF, SSc-ILD, etc.) and the clinical knowledge base (IPF diagnostic criteria).

[0051] Decision-making and intervention: The large language model, based on reasoning, determined that the medication use might not meet the medical insurance reimbursement criteria (lack of direct evidence supporting the IPF diagnosis). The system immediately popped up a real-time alert on the doctor's workstation, prompting "The current diagnostic basis may not meet the drug reimbursement restrictions. Please supplement the relevant diagnostic basis for IPF or select an alternative," thereby preventing potential illegal billing.

[0052] (If deduction has already occurred) Smart Appeal: If the expense has already been deducted by the medical insurance bureau, the system will initiate a smart appeal process. It will automatically read through the entire medical record, accurately locate all paragraphs that mention "pulmonary fibrosis" but lack specific descriptions, and generate an appeal report that clearly states: "The medical record lacks direct imaging or pathological evidence to support the diagnosis of IPF, and does not meet the drug's limited payment conditions," and attach written records as evidence.

[0053] The medical review intelligent agent and medical insurance review management system provided in the above embodiments can realize a knowledge closed loop: the cases intercepted in the process and the results of the appeals afterward can be used as training data to continuously optimize all models, making the in-process warnings more and more accurate, forming a positive cycle that gets smarter the more it is used.

[0054] It can also achieve synergistic effects: in-process quality control reduces general violations, allowing medical insurance experts to focus on handling complex and difficult cases in the intelligent appeal system, maximizing the efficiency of human-machine collaboration; It also enables a unified platform: avoiding the problems of data incompatibility and inconsistent rules among multiple systems, and providing hospitals with a unified and efficient medical insurance compliance management platform.

[0055] The specific embodiments described above illustrate the technical solution and beneficial effects of the present invention in detail. It should be understood that the above description is only the most preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart agent for medical insurance review, characterized in that, It can achieve the following steps: Obtain multi-source heterogeneous data and perform data cleaning to obtain document data; Based on the medical insurance review knowledge base, metadata related to the rules is aggregated and filtered from text data. The large language model performs reasoning based on the review prompt words that are integrated into the medical insurance review knowledge base and metadata, and incorporates the CoT (CoT) reasoning mechanism to output the review results.

2. The intelligent agent for medical insurance review according to claim 1, characterized in that, The multi-source heterogeneous data comes from the Hospital Information System (HIS), Electronic Medical Record System (EMR), Laboratory Information System (LIS), Image Archive, Communication System (PACS), and Medical Record Homepage System, and is obtained from the above systems using the distributed scheduling platform DolphinScheduler and the heterogeneous data source synchronization tool DataX.

3. The intelligent agent for medical insurance review according to claim 2, characterized in that, The metadata related to the rules, which is aggregated and filtered from text data based on the medical insurance review knowledge base, includes: For patient test and medical order data, the Reranker model is used to match the rules in the medical insurance review knowledge base with the patient test and medical order data to perform relevance retrieval scoring, and only data with high scores are adopted as metadata related to the rules.

4. The intelligent agent for medical insurance review according to claim 1, characterized in that, The metadata related to the rules, which is aggregated and filtered from text data based on the medical insurance review knowledge base, includes: For medical records and examination data, the non-thinking mode of LLM is adopted to match the rules in the medical insurance audit knowledge base with the medical records and examination data to perform relevance scoring, and only the data with high scores is adopted as metadata related to the rules.

5. The intelligent agent for medical insurance review according to claim 1, characterized in that, The large language model is based on review prompts integrated with the medical insurance review knowledge base and metadata, and incorporates a CoT (CoT) reasoning mechanism for reasoning, including: The medical insurance review knowledge base and the selected metadata are used to construct dynamic review prompts, which are then input into a large language model. The large language model performs CoT reasoning based on the review prompts and outputs the review results, which include compliance and violation evidence chains.

6. The intelligent agent for medical insurance review according to claim 1, characterized in that, When the large language model performs inference based on the review prompt words and incorporates the CoT (CoT) reasoning mechanism, it adopts batch processing and dynamically adjusts the batch size.

7. A medical insurance review and management system, characterized in that, The medical insurance review intelligent agent according to any one of claims 1-6 is used for clinical medical insurance decision support, that is, to perform real-time document quality control on medical records obtained in real time and generate review results to remind medical workers in real time to provide decision support.

8. The medical insurance review and management system according to claim 7, characterized in that, The medical insurance auditing intelligent agent according to any one of claims 1-6 is used for intelligent auditing after medical insurance settlement, and performs quality control on the medical insurance settlement statement during the auditing process, generating audit results to provide a basis for appeals.

9. The medical insurance review and management system according to claim 7 or 8, characterized in that, The review results are visualized as document quality control prompts, including the text source, review basis and policy requirements, and provide options for whether to review an appeal and whether to pay for it yourself.

10. The medical insurance review and management system according to claim 7 or 8, characterized in that, Text reviewed by the medical review agent is automatically distinguished and marked, and supports copying.

Images