A method and system for generating a medical insurance audit database based on deep learning

Abstract

The present application provides a method and system for generating a medical insurance audit database based on deep learning, which relates to the technical field of databases. The method includes steps such as data collection, obtaining audit requirements, obtaining rules, building a database, rule matching, result judgment, output and optimization; the system includes a data collection module, an audit requirement acquisition module, a rule acquisition module, a database building module, a rule matching module, a result judgment module, an output module and an optimization module. The present application can adapt to changes in the audit policy of medical insurance, adjust the audit rules in a timely manner, and associate the audit rules with historical audit data.

Description

Technical Field

[0001] The present application relates to the technical field of databases, and in particular to a method and system for generating a medical insurance review database based on deep learning. Background Art

[0002] During the medical insurance application process, a large amount of data to be processed and reviewed will be generated. These data include medical insurance payment information, identity certificates, relationship certificates and other types of documents and materials submitted by enterprises or individuals. These data need to be strictly reviewed to ensure their authenticity, accuracy and compliance.

[0003] The business rules of medical insurance audit refer to a series of standards and regulations that medical insurance institutions follow when auditing data. These rules include but are not limited to the review of insurance eligibility, the verification of payment base, the calculation of payment ratio, and the processing of supplementary and recovery payments. The formulation and implementation of medical insurance audit rules are directly related to the balance of income and expenditure of medical insurance funds and the protection of the rights and interests of insured persons. However, due to the complexity and diversity of medical insurance policies, medical insurance audit rules are also constantly being updated and improved.

[0004] A Chinese invention patent with publication date of September 7, 2021 and publication number CN113360671A provides a medical insurance document review method and system based on a knowledge graph. It discloses the use of relationship extraction algorithms to connect entities and set the spatiotemporal attributes of relationships, construct quadruple groups, and use new medical insurance policy documents to dynamically update the medical insurance review knowledge graph. However, the intelligence level of its review decision-making is limited by the completeness and accuracy of the knowledge graph. Summary of the invention

[0005] In order to adapt to changes in medical insurance review policies and improve the intelligence and accuracy of review decisions, the present application provides a method and system for generating a medical insurance review database based on deep learning.

[0006] In the first aspect, the present application provides a method for generating a medical insurance review database based on deep learning, which adopts the following technical solutions:

[0007] A method for generating a medical insurance review database based on deep learning, comprising:

[0008] Data collection: collecting historical audit data and pending audit data, and recording the pending audit data as first data;

[0009] Obtain audit requirements: Obtain audit requirements for medical insurance;

[0010] Obtain rules: formulate business rules according to audit requirements;

[0011] Database construction: Build a knowledge base based on historical audit data and business rules;

[0012] Rule matching: matching the first data with the business rules to obtain a matching result;

[0013] Result judgment: judge whether the matching result meets expectations. If yes, execute the output step; if not, execute the optimization step;

[0014] Output: Output matching results;

[0015] Optimization: including correction, first modeling, data generation, calculation of relevance and storage;

[0016] Correction: Feedback the first data and the matching results to the operation and maintenance personnel, who formulate a business rule based on the first data, which is recorded as the first rule;

[0017] First modeling: Building a data generation model based on deep learning algorithms;

[0018] Data generation: using a data generation model to generate second data that conforms to the first rule;

[0019] Calculate the correlation: calculate the correlation between the second data and the historical audit data, obtain the historical audit data whose correlation is greater than the preset correlation threshold, and record it as the third data;

[0020] Storage: Create the same index for the second data, the third data and the first rule, store them in the knowledge base, and execute the output step.

[0021] By adopting the above technical solution, this application formulates business rules according to the audit requirements of medical insurance, improves flexibility and intelligence, and reduces the burden of manual review by matching the audit data to be processed (first data) with the formulated business rules to obtain matching results. This application also determines whether the matching results meet expectations. If not, it means that the business rules in the knowledge base are not applicable to the audit data to be processed, and the operation and maintenance personnel are required to formulate a corresponding business rule to process the audit data to be processed. In this way, direct human-computer interaction can quickly respond to problems and adjust business rules to adapt to new audit requirements. When the matching result does not meet expectations, the present application further generates second data that meets the business rules (first rules) formulated by the operation and maintenance personnel based on the first data using a deep learning algorithm, obtains historical audit data (i.e., third data) with a high degree of correlation by calculating the correlation between the second data and the historical audit data, and establishes a correlation between the second data, the third data, and the business rules (first rules) formulated by the operation and maintenance personnel, that is, establishes a correlation between the audit data to be processed, the historical audit data, and the first rule, and stores the three and the established correlation in the knowledge base. This data-driven method helps to discover potential laws and trends, provides strong support for subsequent audit work, helps to build more complete and accurate audit rules, and enriches the content of the knowledge base, and improves the rationality and accuracy of audit decisions. The newly generated data, the screened historical audit data, and the new business rules are stored in the knowledge base in the same indexed manner, which realizes centralized management and rapid retrieval of data. At the same time, by establishing the correlation between these data, it provides rich reference information for subsequent audit work, improves the reusability and sharing of data, and reduces data redundancy.

[0022] Optionally, after executing the step of building a database and before executing the step of matching rules, the method further includes:

[0023] Integration: Select a rule engine based on business rules and integrate the rule engine into the medical insurance business system;

[0024] Data processing: processing the first data to obtain processed first data, and recording the processed first data as new first data;

[0025] Triggering rules: The first data is passed to the rule engine, which traverses the business rules in the knowledge base and triggers the business rules according to the content of the first data.

[0026] By adopting the above technical solution, by integrating the selected rule engine into the existing medical insurance business system, the separation of business logic and audit rules is achieved, making the medical insurance business system more modular and scalable, reducing the coupling between the medical insurance business system and the rule engine, so that changes in business rules will not directly affect other parts of the business system, reducing maintenance costs. Through the data processing steps, the first data is converted into a format suitable for processing by the rule engine to ensure that the rule engine can correctly understand and apply the business rules. After the first data is passed to the rule engine, the rule engine automatically traverses the business rules stored in the knowledge base and triggers the corresponding business rules according to the content of the first data. This process does not require human intervention, realizes the automation of the audit process, and improves the timeliness and accuracy of the audit results. Due to the flexibility of the rule engine, business rules can be easily added, modified or deleted to adapt to different audit needs and scenarios.

[0027] Optionally, after the step of performing result judgment and before the step of performing optimization, the method further includes:

[0028] Monitoring: Monitor the matching results in real time and obtain the business rules when the matching results do not meet expectations, which are recorded as the second rules;

[0029] Integration: Integrate all second rules into the first rule set;

[0030] Identification: by means of statistical analysis, identify the second rule with the highest frequency in the first rule set, which is recorded as the third rule;

[0031] Delete: Delete the third rule in the knowledge base and re-execute the monitoring steps.

[0032] By adopting the above technical solution, the present application obtains the business rules (second rules) when the matching results do not meet expectations by setting monitoring steps, integrates all the second rules into the first rule set, and identifies the second rule (i.e., the third rule) with the highest frequency of occurrence in the first rule set through statistical analysis methods, which helps to find the second rule (i.e., the third rule) with the lowest matching success rate, and further deletes the third rule, which helps to clean up invalid or erroneous business rules and improve the accuracy and reliability of the knowledge base. Deleting the third rule can also reduce interference and misjudgment caused by these business rules during the review process.

[0033] Optionally, after executing the identifying step and before executing the deleting step, the method further includes:

[0034] Rule judgment: judge whether there is an error in the third rule. If so, execute the modification step; if not, execute the deletion step;

[0035] Modification: Modify the incorrect third rule and add the modified third rule to the knowledge base.

[0036] By adopting the above technical solution, through the rule judgment step, it is possible to more accurately identify whether the third rule (i.e., the second rule with the highest frequency that does not meet expectations) is indeed wrong. If so, the third rule with errors is corrected through the modification step so that it can correctly reflect business needs and audit standards, thereby improving the accuracy and reliability of the audit process. The modified third rule is added to the knowledge base, realizing the dynamic update and optimization of the knowledge base, which helps to keep the knowledge base synchronized with business needs.

[0037] Optionally, after executing the modifying step and before executing the deleting step, the method further includes:

[0038] Extract the third rule: integrate the wrong third rule and the modified third rule into the second rule set;

[0039] Second modeling: Building LSTM model based on deep learning algorithm;

[0040] Model training: Use the second rule set as the first training set to train the LSTM model to obtain the first LSTM model.

[0041] By adopting the above technical solution, this application integrates the erroneous third rule and the modified third rule into the second rule set, which helps to unify the erroneous third rule and its revised version, and provides a complete data foundation for subsequent analysis and modeling. The LSTM model performs well in processing sequence data and problems with long-term dependencies. In the application scenarios of business rules, there are complex dependencies and associations between rules. The LSTM model can better capture these relationships. The second rule set is used as the first training set, and the first training set is used to train the LSTM model, so that the LSTM model can learn the laws and characteristics in the second rule set and has the ability to correct the erroneous third rule.

[0042] Optionally, after executing the model training step and before executing the deletion step, the method further includes:

[0043] Constructing a training set: Integrate the correct third rule in the rule judgment step into the second training set;

[0044] Retraining: The first LSTM model is retrained using the second training set to obtain a second LSTM model.

[0045] By adopting the above technical solution, the third rule confirmed to be correct in the rule judgment step is integrated into the second training set, and the second training set is used to retrain the first LSTM model, so that the first LSTM model can learn more features and patterns about correct rules. Combined with the model training steps, the LSTM model has the ability to modify incorrect business rules and learn the features and patterns of correct business rules, which indirectly enables the LSTM model to have the ability to judge right and wrong.

[0046] Optionally, after the monitoring step and before the integration step, the method further includes:

[0047] Input: The second rule obtained in the monitoring step is used as the input of the second LSTM model, and the output of the second LSTM model is obtained, which is recorded as the fourth rule;

[0048] First judgment: judging whether the fourth rule is the same as the second rule, if so, executing the optimization step; if not, executing the replacement step;

[0049] Replacement: Use the fourth rule to replace the corresponding second rule.

[0050] By adopting the above technical solution, the present application uses the second rule obtained in the monitoring step as the input of the second LSTM model to obtain the fourth rule. By judging whether the fourth rule is the same as the second rule, judging whether the second rule is wrong, if there is an error, modifying it using the second LSTM model, outputting the fourth rule, and replacing the second rule with the fourth rule, the automatic update and optimization of business rules is realized, which helps to maintain the accuracy and timeliness of the knowledge base.

[0051] Optionally, after executing the identifying step and before executing the deleting step, the method further includes:

[0052] Word segmentation: perform word segmentation on the third rule to obtain keywords;

[0053] Split: Set a split point at the keyword, split the third rule at the split point, and obtain m sub-rules;

[0054] Fill: Add the sub-rule to the knowledge base.

[0055] By adopting the above technical solution, the third rule is decomposed into a series of keywords through word segmentation processing, and split points are set at the keywords to split the third rule into multiple sub-rules. This split helps to decompose the complex third rule into sub-rules that are easier to manage and understand, and improve the flexibility and scalability of the system. The split sub-rules can be applied independently to different scenarios or conditions, which enhances the applicability and coverage of the rule base. The split sub-rules are added to the knowledge base to realize the dynamic update and expansion of the knowledge base.

[0056] Optionally, after executing the splitting step and before executing the filling step, the method further includes:

[0057] Extract sub-rules: extract the i-th sub-rule;

[0058] Calculate similarity: Calculate the similarity between the ith sub-rule and all business rules in the knowledge base in turn, and record it as the fourth data;

[0059] Similarity judgment: judge whether the fourth data is greater than a preset similarity threshold. If so, take the i-th sub-rule as the new third rule and execute the deletion step; if not, execute the filling step.

[0060] By adopting the above technical solution, the present application calculates the similarity between the i-th sub-rule and all business rules in the knowledge base, and can identify repeated business rules or business rules that are highly similar to the sub-rules. If there are repeated or highly similar business rules, the sub-rule is used as a new third rule and deleted from the knowledge base, thereby keeping the knowledge base clean and accurate.

[0061] In the second aspect, the present application provides a system for generating a medical insurance review database based on deep learning, which adopts the following technical solutions:

[0062] A medical insurance review database generation system based on deep learning, comprising:

[0063] A data collection module, used for collecting historical audit data and audit data to be processed, and recording the audit data to be processed as first data;

[0064] Obtain audit requirements module, used to obtain audit requirements for medical insurance;

[0065] The rule acquisition module is connected to the audit requirement acquisition module and is used to formulate business rules according to the audit requirements;

[0066] The database building module is connected to the data collection module and the rule acquisition module to build a knowledge base based on historical audit data and business rules;

[0067] A rule matching module, which is in communication with the data acquisition module and the rule acquisition module, and is used to match the first data with the business rules to obtain a matching result;

[0068] A result judgment module, which is in communication with the rule matching module and is used to judge whether the matching result meets expectations;

[0069] An output module, which is in communication with the result judgment module and is used to output the matching result;

[0070] The optimization module is connected to the data acquisition module, the output module and the database building module in communication, and includes a correction unit, a first modeling unit, a data generation unit, a correlation calculation unit and a storage unit;

[0071] The correction unit is used to feed back the first data and the matching result to the operation and maintenance personnel, and the operation and maintenance personnel formulate a business rule based on the first data, which is recorded as the first rule;

[0072] A first modeling unit, used for establishing a data generation model based on a deep learning algorithm;

[0073] A data generating unit, configured to generate second data conforming to the first rule by using a data generating model;

[0074] A correlation calculation unit, used to calculate the correlation between the second data and the historical audit data, and obtain the historical audit data whose correlation is greater than a preset correlation threshold value, which is recorded as the third data;

[0075] The storage unit is used to establish the same index for the second data, the third data and the first rule, and store them in the knowledge base.

[0076] By adopting the above technical solution, the present application formulates business rules based on the audit requirements obtained by the audit requirements acquisition module, builds a knowledge base based on historical audit data and business rules, and provides rich data support for subsequent rule matching and result judgment. The present application matches the audit data to be processed (first data) with the business rules to obtain matching results, judges the matching results, and confirms whether they meet expectations. When the matching results do not meet expectations, it proves that the business rules in the current knowledge base are not applicable to the audit data to be processed. The audit data to be processed is a new data type caused by policy changes and needs to be fed back to the operation and maintenance personnel to give a business rule. There is a certain correlation between the audit data to be processed and other historical audit data. Therefore, a data generation model is established through a deep learning algorithm to generate second data that meets the first rule. By calculating the correlation between the second data and the historical audit data, the correlation between the two is established. Finally, the same index is established for the second data, the third data and the first rule, and they are stored in the knowledge base, realizing unified management and efficient use of data.

[0077] In summary, the present application includes at least one of the following beneficial technical effects:

[0078] 1. This application formulates business rules according to the audit requirements of medical insurance, which improves flexibility and adaptability. This application obtains matching results by matching the audit data to be processed (first data) with the formulated business rules, thereby reducing the burden of manual review. This application also determines whether the matching results meet expectations. If not, it means that the business rules in the knowledge base are not applicable to the audit data to be processed, and the operation and maintenance personnel are required to formulate a corresponding business rule to process the audit data to be processed. In this way, direct human-computer interaction can quickly respond to problems and adjust business rules to adapt to new audit requirements. When the matching result does not meet expectations, the present application further generates second data that meets the business rules (first rules) formulated by the operation and maintenance personnel based on the first data using a deep learning algorithm, obtains historical audit data (i.e., third data) with a high degree of correlation by calculating the correlation between the second data and the historical audit data, and establishes a correlation between the second data, the third data, and the business rules (first rules) formulated by the operation and maintenance personnel, that is, establishes a correlation between the audit data to be processed, the historical audit data, and the first rule, and stores the three and the established correlation in the knowledge base. This data-driven method helps to discover potential laws and trends, provides strong support for subsequent audit work, helps to build more complete and accurate audit rules, and enriches the content of the knowledge base, and improves the rationality and accuracy of audit decisions. The newly generated data, the screened historical audit data, and the new business rules are stored in the knowledge base in the same indexed manner, which realizes centralized management and rapid retrieval of data. At the same time, by establishing the correlation between these data, it provides rich reference information for subsequent audit work, improves the reusability and sharing of data, and reduces data redundancy.

[0079] 2. This application integrates the erroneous third rule and the modified third rule into the second rule set, which helps to unify the erroneous third rule and its revised version, and provides a complete data foundation for subsequent analysis and modeling. The LSTM model performs well in processing sequence data and problems with long-term dependencies. In the application scenarios of business rules, there are complex dependencies and associations between rules. The LSTM model can better capture these relationships. The second rule set is used as the first training set to train the LSTM model, so that the LSTM model can learn the laws and characteristics in the second rule set and has the ability to correct the erroneous third rule.

[0080] 3. The third rule confirmed to be correct in the rule judgment step is integrated into the second training set, and the second training set is used to retrain the first LSTM model, so that the first LSTM model can learn more features and patterns about correct rules. Combined with the model training steps, the LSTM model has the ability to modify incorrect business rules and learn the features and patterns of correct business rules, which indirectly enables the LSTM model to have the ability to judge right and wrong. BRIEF DESCRIPTION OF THE DRAWINGS

[0081] Figure 1 is a flow chart of the method of Example 1 of the present application;

[0082] Figure 2 is a flow chart of the method of Example 2 of the present application;

[0083] Figure 3 This is a flow chart of the method from S910 monitoring to S915 deletion in Example 3 of the present application;

[0084] Figure 4 This is a flowchart of the method in which S914 is modified to be replaced by S923 in Example 3 of the present application;

[0085] Figure 5 This is a flow chart of the method from S9121 word segmentation to S9126 filling in Example 3 of the present application. DETAILED DESCRIPTION

[0086] The following combination Figures 1 to 5 This application is described in further detail.

[0087] Embodiment 1: This embodiment discloses a method for generating a medical insurance audit database based on deep learning, such as Figure 1 As shown, the method includes the following steps: S1 data collection, S2 obtaining audit requirements, S3 obtaining rules, S4 database construction, S5 rule matching, S6 result judgment, S7 output and S8 optimization. First, historical audit data and audit data to be processed are collected, and then the audit requirements of medical insurance are obtained. Then, business rules are formulated based on the audit requirements, and then matching is performed based on historical audit data and business rules. Then, it is determined whether the matching result meets expectations. If so, S7 output is executed, and if not, S8 optimization is executed. The specific process of this application is as follows:

[0088] S1 data collection, collect historical audit data through internal database, external API interface, file upload and other methods. These historical audit data cover all successful and failed audit cases in the past, including the specific content, results and feedback of the audit. This step also requires real-time collection of pending audit data, which will be specially marked as "first data". In order to improve the efficiency and accuracy of subsequent processing, the first data needs to be initially cleaned and formatted to ensure the consistency and integrity of the first data.

[0089] S2 obtains audit requirements and obtains relevant policy information from local medical insurance official websites through web crawler technology. It also deeply analyzes the specific requirements of medical insurance audits, including but not limited to policies and regulations, audit standards, special terms, etc. It can also communicate with relevant departments or experts to achieve an accurate understanding of the audit requirements.

[0090] Before performing a web crawler, you need to first determine whether the target website complies with the robots.txt protocol and relevant laws and regulations, then determine the list of medical insurance official websites that need to be crawled, then design the crawler, then crawl the web pages, then store the crawled data, then clean the data, remove HTML tags and redundant information in the crawled data, then use Python's Pandas library to organize and analyze the data, extract the specific requirements of medical insurance review, such as reimbursement ratio, reimbursement scope, reimbursement process, etc., and then store the processed data.

[0091] Designing a crawler involves using your browser’s developer tools (such as Chrome DevTools) to view the HTML structure of the target web page, determine the location of content required by policy or audit, and then write the program code.

[0092] S3 obtains rules. Based on the audit requirements obtained in S2, combined with business logic and actual conditions, a set of detailed business rules are formulated. These business rules will cover all aspects of the audit.

[0093] For example, the compliance of the reimbursement form is reviewed and business rules are set. The set business rules include: business rule 1, business rule 2, business rule 3, business rule 4 and business rule 5.

[0094] The compliance of the reimbursement form means that the reimbursement items must comply with the reimbursement scope stipulated by the medical insurance policy, and the information on the reimbursement document (such as date, amount, project name, etc.) must be complete and accurate.

[0095] Business rule 1: Establish a document receiving registration system to record information such as the submission time, submitter, document type and quantity of each reimbursement document.

[0096] Business Rule 2: Conduct a preliminary review of each document to check whether it is complete and whether it is an original or a certified copy. For any documents that are missing or do not meet the requirements, promptly notify the submitter to supplement or correct them.

[0097] Business Rule 3: Calculate the reimbursement amount for each reimbursement document based on the reimbursement standards and maximum limits stipulated in the medical insurance policy.

[0098] Business rule 4: Compare the actual consumption amount on the reimbursement document with the amount to be reimbursed to ensure that the two are consistent. For the amount exceeding the amount to be reimbursed, no reimbursement will be made and the reason will be stated.

[0099] S4 builds a knowledge base based on historical audit data and business rules. The knowledge base not only stores audit cases and business rules, but also includes data analysis and mining results. As the system runs deeper, the knowledge base is constantly updated and optimized to always reflect the latest audit requirements and business rules.

[0100] S5 rule matching, matching the first data with the business rules to obtain matching results. Checking whether the first data meets each business rule one by one in a certain order (such as logical order, order of importance, etc.), and checking whether the first data meets multiple business rules, so as to improve matching efficiency.

[0101] During the matching process, the first data is parsed to extract key information related to the business rules, and the extracted key information is compared with the conditions in the business rules to determine whether the first data meets these conditions. For each business rule, the matching result (such as passed, failed, not applicable, etc.) is recorded.

[0102] For example, the first data is as follows:

[0103] Document number: 20230401-001;

[0104] Reimbursement items: hospitalization expenses;

[0105] Reimbursement amount: 5,000 yuan;

[0106] Document date: April 1, 2023;

[0107] Document type: Original.

[0108] The business rule matching process is as follows:

[0109] A. Document Receipt and Preliminary Review

[0110] Business rule A1: Establish a document receipt registration system (this rule is more about process management and does not directly match the first data, but confirms that the first data has been received and registered).

[0111] Business Rule A2: Conduct a preliminary review of each document.

[0112] Matching process: Check whether the document type in the first data is the original and confirm that the document type meets the requirements.

[0113] Matching result: Matching is successful, the document type is original, and the output matching result is: passed.

[0114] B. Compliance Review

[0115] Business rule B1: Check reimbursement items one by one.

[0116] Matching process: Check whether the reimbursement items in the first data are within the reimbursement range specified by the medical insurance policy. The reimbursement range specified by the local medical insurance policy is 2,000-6,000 yuan.

[0117] Matching result: Matching is successful, the reimbursement item is compliant, and the output matching result is: passed.

[0118] Business rule B2: Check the information on the reimbursement document.

[0119] Matching process: Check whether the document date, reimbursement amount and other information in the first data are complete and accurate.

[0120] Matching result: Assuming that all this information is complete and accurate, the match is successful and the matching result is output: passed.

[0121] C. Calculation of reimbursement amount

[0122] Business Rule C1: Based on the reimbursement standards and maximum limits stipulated in the medical insurance policy.

[0123] Matching process: Compare the reimbursement amount in the first data (5,000 yuan) with the standards and maximum limits stipulated in the medical insurance policy.

[0124] Matching result: 5,000 yuan does not exceed the specified standard and maximum limit, the match is successful, and the matching result is output: passed.

[0125] S6 result judgment, judge whether the matching result meets the expectation, if so, execute S7 output; if not, execute S8 optimization.

[0126] S7 output, output matching results.

[0127] S8 optimization, including S81 correction, S82 first modeling, S83 data generation, S84 calculation of correlation and S85 storage.

[0128] S81 is revised to feed back the first data and matching results to the operation and maintenance personnel (when the operation and maintenance personnel need to refer to the business rules, the business rules that match the first data will also be fed back). The operation and maintenance personnel will analyze the problems and deficiencies in the audit process based on the first data, combined with the actual situation and business needs. Subsequently, the operation and maintenance personnel will formulate a new business rule (recorded as the first rule).

[0129] S82 first modeling, based on deep learning algorithm, establish a data generation model.

[0130] S83 data generation uses a data generation model to generate second data that conforms to the first rule. These second data will be used as supplementary or test data to further verify and optimize the audit capabilities of the system.

[0131] S84 calculates the correlation, calculates the correlation between the second data and the historical audit data, obtains the historical audit data whose correlation is greater than a preset correlation threshold, and records it as the third data.

[0132] By calculating the correlation, the historical audit data (recorded as the third data) that is highly relevant to the second data can be screened out, providing a data basis for S85 storage.

[0133] For example, the correlation is set to be calculated based on whether the two fields "Reimbursement Item" and "Reimbursement Amount" are the same. If they are the same, the correlation value of the field is 1, otherwise, the correlation value of the field is 0.

[0134] Historical audit data includes:

[0135] Reimbursement item: eye examination, reimbursement amount: 2,000 yuan, submitted by: Wang Wu, submission date: April 15, 2023;

[0136] Reimbursement item: dental treatment, reimbursement amount: 2,500 yuan, submitted by: Zhao Liu, submission date: May 5, 2023;

[0137] Reimbursement item: dental treatment, reimbursement amount: 3,000 yuan, submitter: Li Si, submission date: June 10, 2022.

[0138] The correlation is calculated as follows:

[0139] ;

[0140] Among them, D is the total correlation; is the degree of association with the "Reimbursement Item" field; The weight of the association with the "Reimbursement Item" field; is the degree of association with the "Reimbursement Amount" field; is the weight of the association with the "reimbursement amount" field. .

[0141] For the first piece of historical review data, the reimbursement items and reimbursement amounts do not match, and the corresponding correlation is 0.

[0142] For the second piece of historical audit data, the reimbursement items match but the reimbursement amounts do not match, and the corresponding correlation degree is 0.5.

[0143] For the third historical audit data, both the reimbursement items and the reimbursement amounts match, and the corresponding correlation degree is 1.

[0144] S85 stores, establishes the same index for the second data, the third data and the first rule, stores the second data, the third data, the first rule and the correlation between them in the knowledge base, and executes S7 output to output the matching result.

[0145] Take the example of "A company needs to review the medical insurance payment status of its employees to ensure that all employees have paid the corresponding medical insurance fees in accordance with national regulations." for further explanation.

[0146] a. Data collection: collect the audit data to be processed (first data), including: a series of employees' personal information and medical insurance payment records, such as name, age, date of employment, salary income, medical insurance payment base and amount, etc.

[0147] b. Obtain the audit requirements by using web crawler technology to obtain the medical insurance audit requirements on the official website of medical insurance in the prefecture-level city where the company is located.

[0148] c. Obtaining rules. Based on the medical insurance review requirements obtained in b. Obtaining review requirements, some of the business rules formulated are as follows:

[0149] Rule 1: Employees must be within the legal working age.

[0150] Rule 2: The medical insurance contribution base of employees should be between the minimum and maximum standards prescribed by the state.

[0151] Rule 3: The amount of an employee’s monthly medical insurance contribution should be proportional to his or her salary income.

[0152] d. Build a database, build a knowledge base based on historical audit data and business rules, and store the historical audit data and business rules in the knowledge base in the form of a hash table.

[0153] e rule matching, including the following:

[0154] e1 data analysis, analyzes the first data and extracts relevant information such as each employee’s age, date of employment, salary income, and medical insurance payment status.

[0155] e2 condition comparison:

[0156] For rule one, check whether each employee's age is within the legal working age range.

[0157] For Rule 2, compare each employee's medical insurance contribution base to see if it is between the minimum and maximum standards prescribed by the state.

[0158] For Rule 3, the amount of medical insurance that employees should pay is calculated based on their salary income and compared with the actual amount paid.

[0159] e3 result record, for each rule, record the matching result. For example, if an employee's age does not meet the requirements, the result of rule 1 is "failed".

[0160] f Result judgment: obtain the matching result through e3 result record. If the matching result is a negative word such as "failed" or "unsuitable", it is judged as not meeting expectations. Otherwise, it is in line with expectations.

[0161] g output, output matching results.

[0162] h optimization, including h1 correction, h2 first modeling, h3 data generation, h4 calculation of correlation and h5 storage.

[0163] h1 is revised, and the first data (i.e. the employee's medical insurance payment record) and the matching result (i.e. the result of the system judgment) are fed back to the operation and maintenance personnel. After analysis, the operation and maintenance personnel found that the current business rules could not handle certain special situations, so they formulated a new business rule (the first rule), which defines the various situations and requirements for medical insurance payment in more detail. The newly formulated first rule may include special provisions for the medical insurance payment amount for new employees during the probation period, and adjustments to the upper limit of the medical insurance payment base for high-income employees.

[0164] h2 modeling, based on deep learning algorithms, establishes data generation models.

[0165] The data generation model uses Generative Adversarial Networks (GANs), which consists of two parts: a generator and a discriminator. The generator is responsible for generating new data that is as close to the real data as possible, while the discriminator is responsible for distinguishing the generated data from the real data.

[0166] The structure of the data generation model is as follows:

[0167] Generator: Design a neural network whose input is random noise or latent variables and whose output is the generated medical insurance payment record. The goal of the generator is to learn the distribution of real data and generate new data similar to it.

[0168] Discriminator: This involves another neural network whose input is either real data or data generated by the generator and outputs the probability that the data is real. The goal of the discriminator is to distinguish between real data and generated data.

[0169] The training process of the data generation model is as follows:

[0170] During the training process, the generator and the discriminator compete with each other. The generator tries to trick the discriminator into mistaking the generated data for real data, while the discriminator strives to improve its ability to distinguish between real data and generated data.

[0171] By continuously adjusting the parameters of the generator and the discriminator, the two can make progress together in the confrontation until the generator can generate new data that is close enough to the real data.

[0172] h3 data generation, the second data generated using the trained GANs model includes: medical insurance payment records of multiple simulated employees, which meet the requirements of the first rule in terms of age, salary, medical insurance base, etc.

[0173] h4 calculates the correlation between the second data and the medical insurance payment status of similar employees in the historical audit data, and finds a batch of historical records that are highly similar to the second data in multiple key indicators as the third data.

[0174] h5 storage, in the knowledge base, establishes the same index for the second data, the third data and the first rule, and marks the relationship between them. In this way, when similar audits are required in the future, relevant data and rules can be quickly retrieved as a reference.

[0175] This embodiment obtains a matching result by matching the audit data to be processed with pre-established business rules, and further determines whether the matching result meets expectations to determine whether the audit data to be processed is new data affected by the new policy. If it is new data, the operation and maintenance personnel need to provide a new business rule, namely the first rule, based on the audit data to be processed and the matching result, and generate second data that meets the first rule through a data generation model, and establish the same index for the second data, the third data and the first rule by calculating the correlation between the second data and the historical audit data, and store them in the knowledge base, thereby achieving rapid response to problems, adjusting business rules to adapt to new audit requirements, and quickly establishing their correlation with historical audit data.

[0176] Example 2: Figure 2 As shown, the difference between this embodiment and embodiment 1 is that after executing S4 database building and before executing S5 rule matching, it also includes:

[0177] S41 Integration: Select a rule engine based on business rules and integrate the rule engine into the medical insurance business system, for example, select a mature rule engine such as Drools, JBoss Rules, Jess, etc.

[0178] The medical insurance business system is an existing system for handling medical insurance.

[0179] Integrating the rule engine into the medical insurance business system includes:

[0180] API connection: According to the API provided by the rule engine, write the corresponding code to call the functions of the rule engine.

[0181] Data interface: defines the data exchange format and interface between the rule engine and the medical insurance business system to ensure smooth data transmission.

[0182] Rule deployment: Deploy business rules to the rule engine so that they can be executed when needed.

[0183] Test and Verification: Test and verify the integrated rule engine to ensure that it can correctly execute business rules and work seamlessly with the medical insurance business system.

[0184] S42 data processing: processing the first data to obtain processed first data, and recording the processed first data as new first data.

[0185] The processing of the first data includes: data cleaning and data conversion, wherein the data is converted into data results required by the rule engine and the data structure of the medical insurance business system, and necessary conversion is performed on the first data.

[0186] S43 triggers the rule, calls the API of the rule engine to pass the first data to the rule engine, and the rule engine traverses the business rules in the knowledge base. The rule engine triggers the corresponding business rules according to the content of the first data and generates corresponding results or actions.

[0187] For example, if the rule engine indicates that an employee's medical insurance payment base does not comply with regulations, the rule engine will generate a warning or error message and trigger the corresponding processing flow.

[0188] In this embodiment, through the collaborative work of S41, S42 and S43, the medical insurance business system can efficiently process complex business rules and data, thereby improving the accuracy and efficiency of business processing.

[0189] Embodiment 3, as Figure 3 As shown, the difference between this embodiment and embodiment 1 is that after executing S6 result judgment and before executing S8 optimization, it also includes:

[0190] S910 monitoring, using log recording, event monitoring and other technical means, monitors the matching results in real time, and obtains the business rules when the matching results do not meet expectations, which are recorded as the second rules.

[0191] S911 integration integrates all second rules into the first rule set.

[0192] S912 identification, through statistical analysis methods, counts the number of times each second rule is triggered during the monitoring process, identifies the second rule with the highest frequency in the first rule set, and records it as the third rule. The purpose of this step is to identify business rules that may cause frequent matching errors or do not meet expectations.

[0193] S913 rule determination, determining whether the third rule has an error, if so, executing S914 modification; if not, indicating that the third rule is correct but may be outdated, executing S915 deletion.

[0194] S914: modifying the erroneous third rule and adding the modified third rule to the knowledge base.

[0195] For example, the third rule is: If the reimbursement amount exceeds 6,000 yuan, additional medical certification documents must be provided, and the reimbursement item must be hospitalization expenses or major surgical expenses.

[0196] After review by the operation and maintenance personnel, it was found that the rule too strictly restricted the combination of reimbursement amounts and reimbursement items, and excluded some reasonable but high-reimbursement non-hospitalization or non-major surgical expense items.

[0197] The operation and maintenance staff modified the incorrect third rule to: If the reimbursement amount exceeds 5,000 yuan, additional medical certification documents must be provided according to the nature of the reimbursement item. For hospitalization expenses or major surgery expenses, a detailed expense list and diagnosis certificate issued by the hospital must be provided; for other high-cost reimbursement items (such as expensive drugs, special treatments, etc.), a doctor's prescription, treatment plan or certification documents from relevant medical institutions must be provided.

[0198] S915 deletes, deletes the third rule in the knowledge base, and re-executes S910 monitoring.

[0199] In this embodiment, by monitoring the matching results, the second rule with the highest frequency of occurrence when the matching results do not meet expectations is obtained, and it is determined whether there is an error. If the matching failure is caused by an error in the second rule itself, the error is modified and the modified business rule is added to the knowledge base. On the contrary, if it is outdated due to policy changes, which leads to a high probability of matching failure, it is directly deleted from the knowledge base to reduce its resource usage and shorten the matching time.

[0200] like Figure 4 As shown, in other embodiments, after executing S914 modification and before executing S915 deletion, the following is further included:

[0201] S916 extracts the third rule, and integrates the erroneous third rule and the modified third rule into a second rule set.

[0202] The second modeling of S917 is to establish an LSTM model based on the deep learning algorithm and build the basic framework of the LSTM model according to business needs and data characteristics.

[0203] S918 model training uses the second rule set as the first training set to train the LSTM model. Through continuous iteration and optimization, the LSTM model can learn how to modify erroneous business rules to obtain the first LSTM model.

[0204] S919 constructs a training set, and integrates the correct third rule in the rule judgment step into a second training set.

[0205] S920 retrains, using the second training set to retrain the first LSTM model to obtain a second LSTM model, so that the second LSTM model learns a correct business rule pattern.

[0206] Through steps S916 to S920, this embodiment establishes an LSTM model. By training the LSTM model with the erroneous third rule and the modified third rule, the LSTM model learns how to modify the erroneous business rules, and then retrains the trained LSTM model with the correct business rules to learn the correct business rule pattern, indirectly enabling it to have the ability to judge whether the business rules are incorrect. If there are erroneous business rules, the erroneous business rules can be modified and the correct business rules can be output, thereby reducing the review time and improving the review efficiency.

[0207] After executing the S910 monitoring step again, the following solution is implemented:

[0208] S921 inputs the second rule obtained in the monitoring of S910 as the input of the second LSTM model, and outputs the fourth rule based on the pattern or relationship learned internally.

[0209] S922 is the first judgment, judging whether the fourth rule is the same as the second rule. If so, execute S915 to delete, take the fourth rule as the new third rule, and delete the new third rule in the knowledge base (when the fourth rule and the second rule are the same, the second rule, the third rule, and the fourth rule actually refer to one business rule); if not, it indicates that the second rule is wrong, and execute S923 to replace.

[0210] S923 Replacement: Use the fourth rule to replace the corresponding second rule.

[0211] After the above steps, after obtaining the second LSTM model and executing S910 monitoring again, there is no need to execute the steps of integrating S911 into S914 modification. The second rule obtained in S910 monitoring can be directly used as the input of the second LSTM model to obtain the correct business rule, that is, the fourth rule, and by judging whether the second rule and the fourth rule are the same, decide whether to replace the second rule.

[0212] like Figure 5 As shown, in other embodiments, after executing S912 identification and before executing S915 deletion, the following is further included:

[0213] S9121: word segmentation, performing word segmentation processing on the third rule to obtain keywords. The keywords include "and", "or", "comma", etc., which are intended to split the multiple sub-rules included in the third rule.

[0214] S9122 Split, set a split point at the keyword, and split the third rule at the split point to obtain m sub-rules.

[0215] S9123 extracts sub-rules, and extracts the i-th sub-rule from the m sub-rules obtained by splitting as the object currently being processed.

[0216] S9124 calculates the similarity between the ith sub-rule and all business rules in the knowledge base, and records the value of the similarity as fourth data.

[0217] S9125 similarity judgment, judging whether the fourth data is greater than a preset similarity threshold, if so, taking the i-th sub-rule as a new third rule and executing S915 deletion; if not, executing S9126 filling.

[0218] S9126 is filled in, adding the sub-rule corresponding to the fourth data not greater than the preset similarity threshold in the similarity judgment in S9125 to the knowledge base.

[0219] Splitting helps to analyze the third rule in more detail and find parts that are more similar to existing business rules in the knowledge base. In order to compare the similarity of each sub-rule with the business rules in the knowledge base one by one, the sub-rules need to be extracted one by one for processing. Through similarity calculation, the similarity between the sub-rule and the existing business rules in the knowledge base can be evaluated to determine whether there are duplicate or highly similar business rules. Through the filling step, new and unique sub-rules can be incorporated into the knowledge base, thereby enriching the content of the knowledge base and improving its coverage and accuracy.

[0220] Embodiment 4: This embodiment discloses a system for generating a medical insurance review database based on deep learning, the system comprising:

[0221] The data collection module is used to collect historical audit data and audit data to be processed, and record the audit data to be processed as first data.

[0222] The module for obtaining audit requirements is used to obtain the audit requirements for medical insurance.

[0223] The rule acquisition module is connected to the audit requirement acquisition module and is used to formulate a series of business rules based on the audit requirements.

[0224] The database building module communicates with the data acquisition module and the rule acquisition module to build a knowledge base based on historical audit data and business rules. The knowledge base is a database that stores a large amount of historical data and rule information, which can provide strong support for subsequent rule matching and result judgment.

[0225] The rule matching module is connected to the data acquisition module and the rule acquisition module for matching the first data with the business rules and obtaining a matching result.

[0226] The result judgment module is connected to the rule matching module for judging whether the matching result meets expectations.

[0227] The output module is connected to the result judgment module for outputting the matching result.

[0228] The optimization module is connected in communication with the data acquisition module, the output module and the database building module, and includes a correction unit, a first modeling unit, a data generation unit, a correlation calculation unit and a storage unit.

[0229] The correction unit is used to feed back the first data and the matching result to the operation and maintenance personnel. The operation and maintenance personnel formulate a business rule based on the first data, which is recorded as the first rule.

[0230] The first modeling unit is used to establish a data generation model based on a deep learning algorithm.

[0231] The data generating unit is used to generate second data conforming to the first rule by adopting the data generating model.

[0232] The correlation calculation unit is used to calculate the correlation between the second data and the historical audit data. By comparing the similarities and differences between the two, the historical audit data (third data) that is correlated with the second data can be found.

[0233] The storage unit is used to establish the same index for the second data, the third data and the first rule, and store them in the knowledge base.

[0234] This embodiment provides a medical insurance audit database generation system based on deep learning. The data acquisition module collects historical and pending audit data, the audit requirement acquisition module clarifies the medical insurance audit standards, and the rule acquisition module formulates business rules based on this. The database construction module uses historical data and rules to build a knowledge base, and the rule matching module matches the pending data with the rules to obtain the results. The result judgment module evaluates the matching results, and the output module displays the results. The optimization module modifies the erroneous business rules through the correction unit, the first modeling unit generates new data using deep learning, the data generation unit generates data that conforms to the new business rules, the correlation calculation unit finds relevant data, and finally the storage unit indexes the new data, related data and new rules and stores them in the knowledge base to achieve continuous optimization and expansion of the system.

[0235] The above are all preferred embodiments of the present application, and the protection scope of the present application is not limited thereto. Therefore, any equivalent changes made according to the structure, shape, and principle of the present application should be included in the protection scope of the present application.

Claims

1. A method for generating a medical insurance review database based on deep learning, comprising: Data collection: collecting historical audit data and pending audit data, and recording the pending audit data as first data; Obtaining audit requirements: Obtaining audit requirements for medical insurance; characterized in that the method further comprises: Obtain rules: formulate business rules according to audit requirements; Database construction: Build a knowledge base based on historical audit data and business rules; Rule matching: matching the first data with the business rules to obtain a matching result; Result judgment: judge whether the matching result meets expectations. If yes, execute the output step; if not, execute the optimization step; Output: Output matching results; Optimization: including correction, first modeling, data generation, calculation of relevance and storage; Correction: Feedback the first data and the matching results to the operation and maintenance personnel, who formulate a business rule based on the first data, which is recorded as the first rule; First modeling: Based on the deep learning algorithm, a data generation model is established. The data generation model adopts a generative adversarial network. The generative adversarial network includes: a generator and a discriminator. The generator includes a neural network for learning the distribution of real data and generating new data. The discriminator includes another neural network for judging whether the input data is real data or new data output by the generator. Data generation: using a data generation model to generate second data that conforms to the first rule; Calculate the correlation: calculate the correlation between the second data and the historical audit data, obtain the historical audit data whose correlation is greater than the preset correlation threshold, and record it as the third data; Storage: Create the same index for the second data, the third data and the first rule, store them in the knowledge base, and execute the output step.

2. The method for generating a medical insurance audit database based on deep learning according to claim 1 is characterized in that: After executing the database building step and before executing the rule matching step, it also includes: Integration: Select a rule engine based on business rules and integrate the rule engine into the medical insurance business system; Data processing: processing the first data to obtain processed first data, and recording the processed first data as new first data; Triggering rules: The first data is passed to the rule engine, which traverses the business rules in the knowledge base and triggers the business rules according to the content of the first data.

3. The method for generating a medical insurance audit database based on deep learning according to claim 1, characterized in that: After the step of executing result judgment and before the step of executing optimization, the method further includes: Monitoring: Monitor the matching results in real time and obtain the business rules when the matching results do not meet expectations, which are recorded as the second rules; Integration: Integrate all second rules into the first rule set; Identification: by means of statistical analysis, identify the second rule with the highest frequency in the first rule set, which is recorded as the third rule; Delete: Delete the third rule in the knowledge base and re-execute the monitoring steps.

4. The method for generating a medical insurance audit database based on deep learning according to claim 3 is characterized in that: After the step of identifying and before the step of deleting, the following steps are also included: Rule judgment: judge whether there is an error in the third rule. If so, execute the modification step; if not, execute the deletion step; Modification: Modify the incorrect third rule and add the modified third rule to the knowledge base.

5. The method for generating a medical insurance audit database based on deep learning according to claim 4 is characterized in that: After executing the modification step and before executing the deletion step, it also includes: Extract the third rule: integrate the wrong third rule and the modified third rule into the second rule set; Second modeling: Building LSTM model based on deep learning algorithm; Model training: Use the second rule set as the first training set to train the LSTM model to obtain the first LSTM model.

6. The method for generating a medical insurance audit database based on deep learning according to claim 5 is characterized in that: After the model training step and before the deletion step, the following steps are also included: Constructing a training set: Integrate the correct third rule in the rule judgment step into the second training set; Retraining: The first LSTM model is retrained using the second training set to obtain a second LSTM model.

7. The method for generating a medical insurance audit database based on deep learning according to claim 6, characterized in that: After the monitoring step and before the integration step, the following steps are also included: Input: The second rule obtained in the monitoring step is used as the input of the second LSTM model, and the output of the second LSTM model is obtained, which is recorded as the fourth rule; First judgment: judging whether the fourth rule is the same as the second rule, if so, executing the optimization step; if not, executing the replacement step; Replacement: Use the fourth rule to replace the corresponding second rule.

8. The method for generating a medical insurance audit database based on deep learning according to any one of claims 3 to 7, characterized in that: After the step of identifying and before the step of deleting, the following steps are also included: Word segmentation: perform word segmentation on the third rule to obtain keywords; Split: Set a split point at the keyword, split the third rule at the split point, and obtain m sub-rules; Fill: Add the sub-rule to the knowledge base.

9. The method for generating a medical insurance audit database based on deep learning according to claim 8, characterized in that: After the splitting step and before the filling step, the following steps are also included: Extract sub-rules: extract the i-th sub-rule; Calculate similarity: Calculate the similarity between the ith sub-rule and all business rules in the knowledge base in turn, and record it as the fourth data; Similarity judgment: judge whether the fourth data is greater than a preset similarity threshold. If so, take the i-th sub-rule as the new third rule and execute the deletion step; if not, execute the filling step.

10. A system for generating a medical insurance audit database based on deep learning, comprising: A data collection module, used for collecting historical audit data and audit data to be processed, and recording the audit data to be processed as first data; The module for obtaining audit requirements is used to obtain the audit requirements of medical insurance; characterized in that the system is used to implement the method according to any one of claims 1 to 9, and the system further includes: The rule acquisition module is connected to the audit requirement acquisition module and is used to formulate business rules according to the audit requirements; The database building module is connected to the data collection module and the rule acquisition module to build a knowledge base based on historical audit data and business rules; A rule matching module, which is in communication with the data acquisition module and the rule acquisition module, and is used to match the first data with the business rules to obtain a matching result; A result judgment module, which is in communication with the rule matching module and is used to judge whether the matching result meets expectations; An output module, which is in communication with the result judgment module and is used to output the matching result; An optimization module is connected to the data acquisition module, the output module and the database building module in communication, and the optimization module includes a correction unit, a first modeling unit, a data generation unit, a correlation calculation unit and a storage unit; The correction unit is used to feed back the first data and the matching result to the operation and maintenance personnel, and the operation and maintenance personnel formulate a business rule based on the first data, which is recorded as the first rule; A first modeling unit is used to establish a data generation model based on a deep learning algorithm, wherein the data generation model adopts a generative adversarial network, and the generative adversarial network includes: a generator and a discriminator, wherein the generator includes a neural network for learning the distribution of real data and generating new data, and the discriminator includes another neural network for judging whether the input data is real data or new data output by the generator; A data generating unit, configured to generate second data conforming to the first rule by using a data generating model; A correlation calculation unit, used to calculate the correlation between the second data and the historical audit data, and obtain the historical audit data whose correlation is greater than a preset correlation threshold value, which is recorded as the third data; The storage unit is used to establish the same index for the second data, the third data and the first rule, and store them in a knowledge base.

Images