Claim settlement auditing method and system based on knowledge base

Abstract

The invention relates to a knowledge base-based claim settlement auditing method and system, and the method comprises the steps: collecting and sorting structured data related to insurance claim settlement, carrying out the preprocessing, carrying out the matching mapping of the structured data and a preset standard field through a search engine, a deep learning algorithm or a large model analysis algorithm, and obtaining a plurality of matching results; performing fine adjustment and optimization on the plurality of matching results to construct and enrich a standard library; defining a corresponding label library and a relation library for the obtained standard library; obtaining to-be-audited data, performing structured processing, and performing matching mapping with the standard library to obtain standardized audit data; performing claim settlement auditing through a rule engine technology, a deep learning algorithm or a large model analysis algorithm in parallel in combination with a label library and a relation library to generate a plurality of reasoning conclusions; and giving a final auditing conclusion through an expert method, and completing claim settlement auditing. Compared with the prior art, the method has the advantages that the automation degree of claim settlement auditing is improved, and the result is accurate and reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of claims review systems, and in particular to a claims review method and system based on a knowledge base. Background Art

[0002] Claims review in the insurance field is an important part of insurance companies handling insurance compensation requests. After receiving a report and filing a case, an investigation will be conducted, followed by subsequent review work, including:

[0003] Review the authenticity and completeness of the insurance contract, identity certificate, relevant invoices and other supporting documents provided by the insured or beneficiary.

[0004] Check whether the insurance accident meets the claim conditions stipulated in the insurance contract, including the type of accident, time of occurrence, location, etc. At the same time, it will also review whether the insured or beneficiary has insurance rights and interests, and whether the policyholder or insured has violated the obligation to inform or notify.

[0005] Based on the contract provisions and loss circumstances, the insurance company will accurately calculate and review the claim to determine the amount of compensation.

[0006] Approval personnel conduct a comprehensive, objective and fair assessment of the claim application to ensure the accuracy and rationality of the claim case. The assessment results need to be clear so that the insured can understand the calculation method and basis of the compensation amount.

[0007] During the review process of the above insurance materials, an approval personnel is required to determine whether the claim conditions are met and whether the corresponding treatment or repair plan is reasonable. Manual processing brings a lot of work pressure and time cost. In addition, there is no standardized database and judgment rules in various insurance systems at present, and the overall degree of automation is not high. Therefore, there is an urgent need to provide an automatic claim review method to realize the rapid approval of insurance claims. Summary of the invention

[0008] The purpose of the present invention is to overcome the defects of the above-mentioned prior art and to provide a highly automated, accurate and reliable knowledge-based claims review method and system.

[0009] The purpose of the present invention can be achieved by the following technical solutions:

[0010] A claim review method based on a knowledge base includes the following steps:

[0011] Collect and organize structured data related to insurance claims and perform preprocessing, then match and map the structured data with preset standard fields through search engines, deep learning algorithms or large model analysis algorithms to obtain multiple matching results;

[0012] Calculate the similarity between the current structured data and each matching result for fine-tuning, thereby selecting the best matching result to achieve data mapping, so as to build and enrich the standard library;

[0013] Define the corresponding labels and relationships for assisting claims review for the obtained standard library to form a label library and a relationship library;

[0014] Obtain the data to be audited, perform structured processing, and then match and map it with the standard library to obtain standardized audit data;

[0015] The standardized audit data is processed in parallel using rule engine technology, deep learning algorithm or large model analysis algorithm, and combined with the label library and relationship library to conduct claims audit, generating multiple reasoning conclusions; each reasoning conclusion is evaluated based on experience through expert method, and the final audit conclusion is given to complete the claims audit.

[0016] Furthermore, the preprocessing rules include:

[0017] Case conversion processing, used to convert text to all lowercase or all uppercase;

[0018] Punctuation removal, used to remove punctuation marks from text;

[0019] Stop word removal, which is used to remove common words that are not helpful in understanding the meaning of pre-defined words;

[0020] Disease synonym processing, which is used to convert synonymous disease nouns into unified descriptions;

[0021] The preprocessing process also includes identifying pinyin and radical feature values, thereby improving the similarity weights of different terms for matching. The pinyin feature value identification process includes: converting both terms into pinyin, and if the pinyin is consistent, improving the corresponding similarity weight; the radical feature value identification process includes: extracting radicals from both terms, and if the radicals are consistent, improving the corresponding similarity weight.

[0022] Furthermore, the matching mapping process of the search engine includes:

[0023] Predefine multiple aliases corresponding to the standard fields, thereby constructing an alias library;

[0024] The structured data is searched and matched with various aliases in the alias library. If the match is successful, the standard field corresponding to the alias is mapped to obtain a matching result.

[0025] Furthermore, the matching and mapping process of the deep learning algorithm and the large model analysis algorithm both include:

[0026] Pre-acquire a plurality of corpora that are mapped with the standard fields to form a corpus;

[0027] Machine learning training is performed based on the corpus, and the structured data is matched and mapped using a trained deep learning algorithm or a large model analysis algorithm.

[0028] Furthermore, the fine-tuning is implemented by using an algorithm based on edit distance or an algorithm based on set theory;

[0029] The algorithm based on edit distance selects the best matching result by comparing the difference between the current structured data and each matching result.

[0030] The set theory-based algorithm selects the best matching result by calculating the similarity between the current structured data and each matching result;

[0031] The algorithm based on edit distance includes edit distance algorithm, Hamming distance algorithm or Damerau-Levenshtein distance algorithm;

[0032] The set theory-based algorithm calculates similarity through Jaccard similarity or Dice coefficient.

[0033] Furthermore, the standard library sets codes, standard names and classification fields for each relevant term for standardized storage;

[0034] The tag library is used to set tags for terms for specific classification;

[0035] The relationship library is used to set applicability restrictions, gender restrictions, amount restrictions and dosage restrictions.

[0036] Furthermore, the process of claim review using the rule engine technology is as follows:

[0037] Based on the standard library, multiple audit rules are constructed in advance in combination with the tag library and the relationship library;

[0038] Each of the audit rules is split into multiple conditional nodes. During the claim audit process of the standardized audit data, it is determined whether each conditional node is activated. The activated conditional nodes form a path. If all conditional nodes on the path are activated, the audit rule is deemed to meet the conditions and the corresponding reasoning conclusion is drawn.

[0039] Furthermore, the process of the deep learning algorithm or the large model analysis algorithm for claim review is specifically as follows:

[0040] Obtain standardized audit data and corresponding audit conclusions in advance to build training data;

[0041] Machine learning training is performed based on the training data, and the trained deep learning algorithm or large model analysis algorithm is used to perform claims audit on the standardized audit data to generate inference conclusions.

[0042] Furthermore, the search engine adopts TF-IDF algorithm, BM25 algorithm, BM25L algorithm, BM25+ algorithm, BM25T algorithm or BM25F algorithm;

[0043] The deep learning algorithm is a recurrent neural network, a BiLSTM-CRF model, or a pre-trained Transformer model;

[0044] The large model analysis algorithm is the GPT model.

[0045] The present invention also provides a claim review device based on a knowledge base, comprising a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of the method described above.

[0046] Compared with the prior art, the present invention has the following advantages:

[0047] (1) The present invention uses a search engine, a deep learning algorithm or a large model analysis algorithm to match and map the acquired structured data with standard fields; in the case where there are multiple standard field matching results, further fine-tuning is performed based on similarity evaluation to obtain a unique matching result, thereby realizing standardized database construction, and further setting up a label library and a relationship library to assist in claims review, thereby improving the standardization of terminology and providing data support for subsequent automatic claims review;

[0048] The structured data of claims are first matched and mapped with the standard library to obtain standardized audit data. Then, rule engine technology, deep learning algorithm and large model analysis algorithm are used in parallel, and the tag library and relationship library can be assisted to realize automatic audit of the structured data of claims and generate multiple reasoning conclusions. Finally, the final audit conclusion is obtained based on the expert's experience evaluation.

[0049] This solution has taken into account and optimized many aspects from the construction process of the standard library to the claims review process, making it more accurate and reliable, and overall improving the degree of automation of claims review.

[0050] (2) The present invention is based on a constructed database, and according to the various fields, tag library and relationship library therein, combined with rule engine technology, constructs and updates matching rules, thereby achieving a more complete and accurate audit rule expression based on a database with a high degree of standardization.

[0051] (3) The present invention pre-processes the collected data, taking into account multiple aspects such as case conversion, punctuation removal, stop word removal and disease synonym processing, and combines search engines, deep learning algorithms or large model analysis algorithms to achieve data matching mapping, has the ability to efficiently establish medical knowledge, and has generalization. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of a process of a claims review method based on a knowledge base is provided in an embodiment of the present invention;

[0053] Figure 2 A schematic diagram of a framework flow of a claim review method based on a knowledge base is provided in an embodiment of the present invention;

[0054] Figure 3 A schematic diagram of a standardized data review process provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings here can be arranged and designed in various different configurations.

[0056] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents selected embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0057] It should be noted that similar reference numerals and letters denote similar items in the following drawings, and therefore, once an item is defined in one drawing, further definition and explanation thereof is not required in subsequent drawings.

[0058] Example 1

[0059] like Figure 1 As shown, this embodiment provides a claim review method based on a knowledge base, comprising the following steps:

[0060] S1: Figure 2 As shown, the structured data related to insurance claims is collected and sorted and preprocessed, and then the structured data is matched and mapped with the preset standard fields through a search engine, a deep learning algorithm or a large model analysis algorithm to obtain multiple matching results;

[0061] The structured data therein is the data structure corresponding to the common rule forms in the field of insurance claims settlement. For example, the diagnosis list issued by a doctor directly records the disease name, treatment measures, etc. in the corresponding area, forming structured data;

[0062] S2: Calculate the similarity between the current structured data and each matching result for fine-tuning, so as to select the optimal matching result to achieve data mapping and construct and enrich the standard library;

[0063] S3: Define the corresponding tags and relationships for assisting claims settlement review for the obtained standard library to form a tag library and a relationship library;

[0064] S4: As Figure 3 shown, obtain the data to be reviewed, perform structured processing, and then match and map it with the standard library to obtain standardized review data;

[0065] S5: Pass the standardized review data through the rule engine technology, deep learning algorithm or large model analysis algorithm in parallel respectively, and combine the tag library and the relationship library to conduct claims settlement review to generate multiple inference conclusions; evaluate each inference conclusion according to experience through the expert method to give the final review conclusion and complete the claims settlement review.

[0066] The following is a specific description of the above steps:

[0067] In step S1, the preprocessing rules include:

[0068] Case conversion processing, which is used to convert the text to all lowercase or all uppercase for unified processing. For example, "N" and "n" can both be converted to lowercase "n";

[0069] Punctuation removal processing, which is used to remove the punctuation marks in the text because these symbols usually do not affect the meaning of the words;

[0070] Stop word removal processing, which is used to remove the common words that are not very helpful for understanding the meaning of the words, such as "de", "shi", etc.; however, note that for medical terms or professional vocabulary, common stop words may not be included;

[0071] Disease synonym processing, which is used to convert synonymous disease names into unified descriptions; the rule examples are as follows, and this rule needs to be updated manually regularly:

[0072] Table 2 Disease Synonym Processing

[0073] Malignant tumors cancer cancer cancer certificate disease Levy disease Psoriasis psoriasis Ⅻ XII ++ ++

[0074] Preferably, the algorithm mechanism of word frequency is commonly used to judge similar words. This method considers that in actual data, there are many misspelled words from stroke input methods and pinyin input methods, and for this feature, the pinyin and radical feature values are extracted.

[0075] Pinyin example:

[0076] Upper respiratory tract infection and upper respiratory tract infection----pinyinization----》After being converted to pinyin, both are shang hu xidao gan ran. If the pinyin is the same, the similarity weight will be increased.

[0077] Radical examples:

[0078] Escherichia coli enteritis and Bacillus enteritis ----radicalization----》Bacillus enteritis. If the radicals are the same after word segmentation, the similarity weight will be increased.

[0079] In step S1, the search engine's matching and mapping process includes:

[0080] Predefine multiple aliases corresponding to standard fields to build an alias library;

[0081] The structured data is searched and matched with various aliases in the alias library. If the match is successful, the standard field corresponding to the alias is mapped to obtain the matching result.

[0082] The matching and mapping process of deep learning algorithms and large model analysis algorithms includes:

[0083] Acquire multiple corpora that are mapped to standard fields in advance to form a corpus;

[0084] Perform machine learning training based on the corpus, and use the trained deep learning algorithm or large model analysis algorithm to match and map the structured data.

[0085] In step S2, the fine-tuning process of each matching result is implemented by using an algorithm based on edit distance or an algorithm based on set theory;

[0086] The algorithm based on edit distance selects the best inference conclusion by comparing the differences between the current structured data and each matching result.

[0087] The algorithm based on set theory selects the best inference conclusion by calculating the similarity between the current structured data and each matching result;

[0088] Algorithms based on edit distance include edit distance algorithm, Hamming distance algorithm or Damerau-Levenshtein distance algorithm;

[0089] Algorithms based on set theory calculate similarity through Jaccard similarity or Dice coefficient.

[0090] In step S2, the standard library established in this embodiment contains a library of medical terms commonly used in insurance claims, including diseases, surgeries, drugs, diagnosis and treatment, materials, and departments.

[0091] Taking the diagnosis and treatment database as an example, it contains fields such as code, standard name and classification. Among them, the most important fields are standard name and alias, which are used for standardization.

[0092] In step S3, based on the standard library, specific knowledge needs to be defined to assist in claim review. Knowledge includes two types: labels and relationships, forming a label library and a relationship library respectively.

[0093] Label library: For example, if a label is set to indicate that the disease "cholera" is a specific infectious disease, it will be excluded during the review process.

[0094] Relationship library: The following is an example of a relationship library, which shows that metformin is limited to diabetes. If only metformin but no diabetes is found during the audit, an audit prompt will be given.

[0095] Table 1 Relationship database instance

[0096] Drug Name Relationship Type Disease name Metformin Limited application diabetes

[0097] Relationship types also include gender restrictions, age restrictions, amount restrictions, dosage restrictions, etc.

[0098] In step S4, the data to be audited needs to be converted into standardized structured data and then matched and mapped with the established standard library. The matching and mapping process can adopt the same scheme as the standard library construction process to obtain standardized audit data for subsequent audit.

[0099] In step S5, the process of claim review using rule engine technology is as follows:

[0100] Based on the standard library, multiple audit rules are built in advance by combining the tag library and the relationship library;

[0101] Each audit rule is split into multiple conditional nodes. During the claim audit process of standardized audit data, it is determined whether each conditional node is activated. The activated conditional nodes form a path. If all conditional nodes on the path are activated, the audit rule is deemed to meet the conditions and the corresponding reasoning conclusion is drawn.

[0102] The specific process of claims review using deep learning algorithms or large model analysis algorithms is as follows:

[0103] Obtain standardized audit data and corresponding audit conclusions in advance to build training data;

[0104] Perform machine learning training based on training data, use the trained deep learning algorithm or large model analysis algorithm to conduct claims audit on standardized audit data and generate inference conclusions.

[0105] Specifically, the above-mentioned search engine technology, deep learning algorithm, large model analysis algorithm and fine-tuning algorithm are introduced as follows:

[0106] 1. Search engine technology (word meaning similarity matching engine)

[0107] TF-IDF

[0108] TF-IDF algorithm is the abbreviation of "Term Frequency-Inverse Document Frequency". This algorithm is mainly used to evaluate the importance of a word in a document set or a document in a corpus. It is a commonly used weighting technology in text mining and information retrieval. The following is a detailed description of the TF-IDF algorithm:

[0109] 1.1.1 Algorithm composition

[0110] The TF-IDF algorithm consists of two parts: TF (Term Frequency) and IDF (Inverse Document Frequency).

[0111] 1. TF (Term Frequency):

[0112] οDefinition: Indicates the frequency of a word appearing in a document.

[0113] Formula: TF(t,d) = the total number of terms in document d and the number of times term t appears in document d

[0114] ο Note: This number is usually normalized to prevent it from being biased towards longer documents. TF values ​​for different documents cannot be directly compared, but TF values ​​for different terms in the same document can be compared.

[0115] 2. IDF (Inverse Document Frequency):

[0116] οDefinition: It indicates the general importance of a word, that is, the degree of discrimination of the word in the document collection.

[0117] Formula: IDF(t,D) = log the number of documents containing term t + 1 the total number of documents

[0118] ο Note: The denominator is added with 1 to avoid the situation where the denominator is 0. Laplace smoothing is used to enhance the robustness of the algorithm. The higher the IDF value, the rarer the word is and the stronger the ability to distinguish documents is.

[0119] 1.1.2 Algorithm Principle

[0120] The basic idea of ​​the TF-IDF algorithm is to combine term frequency and inverse document frequency to evaluate the importance of a term in a document. Specifically, if a term appears frequently in a document and rarely in other documents, then the term has good discrimination and is suitable as a keyword for the document.

[0121] 1.1.3 Algorithm Application

[0122] 1. Feature extraction: Convert text data into feature vector representation to facilitate processing by machine learning algorithms.

[0123] 2. Feature screening: By calculating the TF-IDF value of a word, you can find important keywords in the text and help understand the text content.

[0124] 3. Text classification and clustering: Text classification and clustering tasks can be performed based on TF-IDF feature vectors.

[0125] 4. Similarity calculation: By calculating the TF-IDF similarity between documents, the similarity between documents can be measured.

[0126] 1.1.4 Algorithm Features

[0127] Simple and effective: The TF-IDF algorithm is simple to implement and has significant effects. It is one of the basic algorithms commonly used in text processing.

[0128] Strong interpretability: The TF-IDF value can intuitively reflect the importance of a word in a document and has strong interpretability.

[0129] High flexibility: The parameters in the TF-IDF algorithm (such as the logarithmic base in the IDF formula) can be adjusted according to actual conditions to adapt to different application scenarios.

[0130] BM25

[0131] The BM25 (Best Matching 25) algorithm is the mainstream text matching algorithm in the current information retrieval field. It is mainly used to calculate the relevance score between the query text (query) and each document in the document collection. BM25 can be regarded as an optimization of the TF-IDF algorithm. It adds several adjustable parameters on the basis of TF-IDF to make the algorithm more flexible and practical in application. The following is a detailed description of the BM25 algorithm:

[0132] 1.2.1 Algorithm Basics

[0133] TF-IDF (Term Frequency-Inverse Document Frequency): TF-IDF is a statistical method used to evaluate the importance of a word to a document set or a document in a corpus. The importance of a word increases in direct proportion to the number of times it appears in a document, but decreases in inverse proportion to the frequency of its appearance in the corpus.

[0134] Improvements of BM25 based on TF-IDF: Based on TF-IDF, the BM25 algorithm further refines and optimizes the calculation method of the relevance score by introducing additional parameters and calculation methods.

[0135] 1.2.2 Algorithm composition

[0136] The BM25 algorithm mainly consists of the following parts:

[0137] Importance of each word in the query: This is usually measured by IDF (Inverse Document Frequency), i.e. the less frequent a word appears in a document collection, the more important it is in a particular document.

[0138] The relevance between each word in the query and the document: BM25 optimizes the TF (term frequency) part and takes into account the length of the document. Specifically, it uses a nonlinear function to calculate the relationship between term frequency and relevance to avoid over-boosting the relevance score of long documents due to high term frequency.

[0139] The relevance of each word in the query to the query itself: This part is only used when the query is long and is used to characterize the relevance between words within the query.

[0140] 1.2.3 Algorithm formula

[0141] The general formula of the BM25 algorithm can be expressed as:

[0142] score(Q,d)=i=1∑nwi·R(qi,d)

[0143] in:

[0144] Q represents a query.

[0145] d represents the document to be matched.

[0146] qi represents the i-th word in the query.

[0147] wi represents the importance of qi, which is usually calculated using IDF or its variants.

[0148] R(qi,d) represents the correlation score between qi and document d, and its calculation formula takes into account factors such as word frequency and document length.

[0149] 1.2.4 Parameter Description

[0150] k1: used to control the saturation of word frequency, usually with a value between 1.2 and 2.

[0151] b: used to control the impact of document length on the score, with a value between 0 and 1. When b = 1, the document length has the greatest impact; when b = 0, the document length has no impact on the score.

[0152] The BM25 algorithm is widely used in information retrieval fields such as search engines and recommendation systems to estimate the relevance between documents and user queries and to sort documents accordingly.

[0153] 1.2.5 Advantages

[0154] Flexibility: By introducing adjustable parameters, the BM25 algorithm can be optimized according to different application scenarios and requirements.

[0155] Accuracy: Compared to the TF-IDF algorithm, BM25 shows higher accuracy when processing long documents and short queries.

[0156] In summary, the BM25 algorithm is an improved algorithm based on TF-IDF. By introducing additional parameters and calculation methods, it further improves the accuracy and flexibility of the relevance score calculation of text matching.

[0157] 1.3. BM25 Advanced

[0158] Based on BM25, the following algorithms are also used depending on the situation.

[0159] 1.BM25L (BM25 with Length Normalization):

[0160] ○This algorithm takes into account the impact of document length on the score based on BM25, and balances documents of different lengths by introducing a document length normalization term, with the aim of reducing the impact of document length on the relevance score.

[0161] 2.BM25+:

[0162] ○BM25+ is an improved BM25 algorithm that adds the calculation of query term weights to better handle important terms in the query. This penalty term is used to adjust the relevance score of longer documents to avoid longer documents occupying too large a proportion in the score.

[0163] 3.BM25T (BM25 with Term Weights):

[0164] ○BM25T is a method that introduces word weights into BM25, which determines the importance of each word in the text by considering features such as word frequency, inverse document frequency, and document length. It allows users to assign different weights to each word in the query to better reflect the importance of the query.

[0165] 4.BM25F:

[0166] ○BM25F is an improved algorithm that takes multiple fields into account. In information retrieval, there are usually multiple fields (such as title, body, tags, etc.) whose relevance needs to be scored. BM25F takes a weighted sum of the scores of multiple fields to better consider the impact of different parts of the document on the matching score, thereby deriving the final relevance score.

[0167] 2. Deep Learning Algorithms

[0168] In recent years, with the rapid development of deep learning technology, the performance of NER models based on deep learning has been significantly improved. The following are some commonly used deep learning models:

[0169] 1. RNN and its variants

[0170] Recurrent Neural Networks (RNNs) and their variants (such as LSTM and GRU) can process sequence data and capture long-distance dependencies in text. In NER tasks, RNNs are often used in combination with CRFs to form RNN-CRF models. This model first extracts feature representations of text sequences through RNNs, and then uses CRF layers for sequence annotation to take into account dependencies between tags.

[0171] 2. BiLSTM-CRF Model

[0172] The BiLSTM network combines LSTM in both the forward and backward directions, and can capture both the forward and backward information of the text sequence. The BiLSTM-CRF model performs well in the NER task and has become one of the current mainstream models.

[0173] 3. Pre-trained Model

[0174] Pre-trained Transformer models (such as BERT, GPT, etc.) have acquired rich language knowledge and context representation capabilities through self-supervised learning on large-scale corpora. These models can be fine-tuned to adapt to downstream tasks such as NER, significantly improving recognition performance.

[0175] Implementation steps of NER technology

[0176] The implementation of NER technology usually includes the following steps:

[0177] 1. Data preprocessing: Perform preprocessing operations such as text cleaning, word segmentation, and part-of-speech tagging so that subsequent algorithms can better process it.

[0178] 2. Feature extraction: Extract features useful for named entity recognition from the preprocessed text, such as vocabulary, part of speech, contextual information, etc.

[0179] 3. Model training: Use the labeled data set to train the NER model and learn the recognition rules and patterns of named entities.

[0180] 4. Entity recognition: Apply the trained model to new text to automatically identify and classify named entities.

[0181] 5. Post-processing: Post-process the recognition results, such as deduplication and sorting, for subsequent use.

[0182] 3. Large model analysis algorithm

[0183] 3.1 Model Architecture

[0184] Transformer model: GPT uses the Transformer model architecture, which is a neural network model based on the self-attention mechanism, which can maintain good performance when processing long texts. This model architecture enables GPT-4 to understand and process complex medical texts.

[0185] 3.2 Training Process

[0186] Unsupervised pre-training: GPT first conducts unsupervised pre-training on a large amount of general text corpus to learn the general features and rules of text. In this process, the model will learn basic information such as language grammar and semantics.

[0187] Supervised fine-tuning: Based on pre-training, GPT will perform supervised fine-tuning on specific datasets in the medical field. Through this process, the model can learn specific information such as expertise in the medical field and entity relationships, thereby improving its reasoning ability in the medical field.

[0188] 3.3 Reasoning Mechanism

[0189] Text generation: During the reasoning process, GPT can generate corresponding output text based on the input text. For example, when a description of a disease is input, the model can generate relevant information such as the definition, symptoms, and treatment methods of the disease.

[0190] Entity recognition and relationship reasoning: By utilizing data such as medical standard libraries and relationship libraries, GPT can identify medical entities in text (such as disease names, drugs, etc.) and the relationships between them (such as the association between diseases and symptoms, the relationship between drugs and treatment effects, etc.). Based on these entities and relationships, the model can perform complex reasoning and judgment.

[0191] Contextual understanding: GPT is able to consider contextual information when processing text, which helps it make more accurate inferences in the medical field. For example, when understanding the symptoms of a disease, the model will consider other relevant information about the disease (such as the patient's age, gender, medical history, etc.).

[0192] 3.4. Generation of Reasoning Conclusions

[0193] Comprehensive reasoning: Based on the above steps, GPT can integrate the input information, the knowledge learned by the model and the context information to generate reasoning conclusions. These conclusions may include diagnosis suggestions for diseases, recommendations for treatment plans, etc.

[0194] Credibility assessment: To improve the credibility of the inference conclusions, the conclusions generated by the model can be further evaluated and verified. For example, the conclusions can be compared and verified using an expert knowledge base to ensure their accuracy and reliability.

[0195] 4. Rule Engine Technology

[0196] The process of generating reasoning conclusions using rule engine technology specifically includes:

[0197] Construct matching rules, split the matching rules into multiple conditional nodes, and judge whether each conditional node is activated during the matching process of the input text to be reviewed. The activated conditional nodes form a path. If all conditional nodes on the path are activated, the matching rule is considered to meet the conditions.

[0198] Matching rules are built and updated based on the database.

[0199] In this embodiment, the Drools rule engine is used to dynamically implement the audit rules. The principles of the Drools rule engine can be summarized into the following core aspects:

[0200] 4.1 Basic Definition and Composition

[0201] Drools is an open source rule engine developed in Java language provided by JBoss. It is an implementation of Business Rule Management System (BRMS). Drools separates business decision logic from application code, allows business rules to be written using predefined semantic templates (such as rule scripts), and makes decisions based on these rules at runtime.

[0202] 4.2 Core Components and Concepts

[0203] 1. Rules:

[0204] ○ Rules are the key elements for expressing business logic in Drools, and consist of two parts: the condition (LHS) and the conclusion (RHS). The condition part defines the scenario that triggers the rule, while the conclusion part specifies the action to be performed when the condition is met.

[0205] 2. Fact:

[0206] ○Facts are data object instances passed to the rule engine, representing the current state or event. The rule engine puts the fact objects into the working memory and matches these facts against the condition part of the rule.

[0207] 3. Working Memory:

[0208] οWorking memory is where facts are temporarily stored in Drools and where the rule engine is executed. When facts are inserted into the working memory, the engine re-evaluates all rules to see if any of the rule conditions are met.

[0209] 4. Production Memory:

[0210] οIn Drools, production memory refers to where the rules themselves are stored, that is, the rule base. The rules here have been compiled and are ready to be executed.

[0211] 5. Agenda:

[0212] ο The agenda is responsible for tracking all the rules that meet the conditions and the order in which they are executed. When the condition part of the rule is met, the rule will be added to the agenda and wait for execution. Drools supports multiple execution strategies (such as forward links, reverse links, conflict resolution strategies, etc.) to determine the activation and execution order of rules.

[0213] 4.3 Rete Algorithm

[0214] Principle: Rete (Rapidly Exploring Random Trees) or its improved version ReteOO is an efficient pattern matching algorithm behind the Drools rule engine. The algorithm optimizes the rule matching process and reduces repeated calculations by building a data structure (Rete network). It maintains an internal network structure to track the relationship between conditions to speed up the response to facts changed in the working memory.

[0215] Process: First, split the rules into multiple conditional nodes and build a hierarchical structure in the network. When fact objects are inserted into the working memory, they propagate in the network and activate the nodes that match them. The activated nodes form a path, and if all nodes on this path are activated, the rule is considered to meet the conditions and is added to the agenda for execution.

[0216] 4.4 Workflow

[0217] 1. Rule file management: Define and manage rule files (usually in .drl format) through management tools (such as Drools Workbench).

[0218] 2. Rule compilation: Use the tools provided by Drools (such as KieServices) to compile the rule file into an executable form and store it in the production memory.

[0219] 3. Rule Enforcement:

[0220] ο At runtime, create a KieSession from production memory and insert real-time objects (i.e. facts) in the KieSession.

[0221] When a fact object is inserted or updated, the engine traverses the Rete network to find the rule conditions that match the fact.

[0222] ο If all conditions of a rule are met, the rule is activated and added to the agenda.

[0223] ο Trigger all matched rules to start executing by calling the kieSession.fireAllRules() method until the agenda is empty (that is, all eligible rules have been executed).

[0224] 5. Fine-tuning the algorithm

[0225] The fine-tuning process is implemented using an algorithm based on edit distance or an algorithm based on set theory;

[0226] The algorithm based on edit distance calculates the difference between the current structured data and each matching result, thereby selecting the best matching result;

[0227] The algorithm based on set theory selects the best matching result by calculating the similarity between the current structured data and each matching result.

[0228] 1) Algorithm based on edit distance

[0229] Edit distance (Levenshtein distance): Measures the difference between two strings by counting the minimum number of single-character edits (insertions, deletions, or substitutions) required to transform one string into the other.

[0230] Hamming distance: This is only applicable when the two strings are of the same length. It measures the difference by counting the number of different characters in corresponding positions of the two strings.

[0231] Damerau-Levenshtein distance: A variation of the edit distance that takes into account the swapping of adjacent characters in addition to insertions, deletions, and substitutions.

[0232] 2) Algorithms based on set theory

[0233] Jaccard similarity: It measures the similarity between two sets (or texts) by calculating the ratio of the size of the intersection of the two sets to the size of the union.

[0234] Dice coefficient: Similar to Jaccard similarity, but the denominator is 2 times the size of the set intersection plus the sum of the sizes of the individual sets, which is used to handle non-binary data.

[0235] This embodiment also provides a knowledge base-based claims review device, including a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of the knowledge base-based claims review method as described above.

[0236] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A claims review method based on a knowledge base, characterized in that: The following steps are involved: Collect and organize structured data related to insurance claims and perform preprocessing, then match and map the structured data with preset standard fields through search engines, deep learning algorithms or large model analysis algorithms to obtain multiple matching results; Calculate the similarity between the current structured data and each matching result for fine-tuning, thereby selecting the best matching result to achieve data mapping, so as to build and enrich the standard library; Define the corresponding labels and relationships for assisting claims review for the obtained standard library to form a label library and a relationship library; Obtain the data to be audited, perform structured processing, and then match and map it with the standard library to obtain standardized audit data; The standardized audit data is processed in parallel using rule engine technology, deep learning algorithm or large model analysis algorithm, and combined with the label library and relationship library to conduct claims audit, generating multiple reasoning conclusions; each reasoning conclusion is evaluated based on experience through expert method, and the final audit conclusion is given to complete the claims audit.

2. A method for claim review based on a knowledge base according to claim 1, characterized in that: The pre-processing rules include: Case conversion processing, used to convert text to all lowercase or all uppercase; Punctuation removal, used to remove punctuation marks from text; Stop word removal, which is used to remove common words that are not helpful in understanding the meaning of pre-defined words; Disease synonym processing, which is used to convert synonymous disease nouns into unified descriptions; The preprocessing process also includes identifying pinyin and radical feature values, thereby improving the similarity weights of different terms for matching. The pinyin feature value identification process includes: converting both terms into pinyin, and if the pinyin is consistent, improving the corresponding similarity weight; the radical feature value identification process includes: extracting radicals from both terms, and if the radicals are consistent, improving the corresponding similarity weight.

3. A method for claim review based on a knowledge base according to claim 1, characterized in that: The matching mapping process of the search engine includes: Predefine multiple aliases corresponding to the standard fields, thereby constructing an alias library; The structured data is searched and matched with various aliases in the alias library. If the match is successful, the standard field corresponding to the alias is mapped to obtain a matching result.

4. A method for claim review based on a knowledge base according to claim 1, characterized in that: The matching and mapping process of the deep learning algorithm and the large model analysis algorithm both include: Pre-acquire a plurality of corpora that are mapped with the standard fields to form a corpus; Machine learning training is performed based on the corpus, and the structured data is matched and mapped using a trained deep learning algorithm or a large model analysis algorithm.

5. A method for claim review based on a knowledge base according to claim 1, characterized in that: The fine tuning is implemented by using an algorithm based on edit distance or an algorithm based on set theory; The algorithm based on edit distance selects the best matching result by comparing the difference between the current structured data and each matching result. The set theory-based algorithm selects the best matching result by calculating the similarity between the current structured data and each matching result; The algorithm based on edit distance includes edit distance algorithm, Hamming distance algorithm or Damerau-Levenshtein distance algorithm; The set theory-based algorithm calculates similarity through Jaccard similarity or Dice coefficient.

6. A method for claim review based on a knowledge base according to claim 1, characterized in that: The standard library sets codes, standard names and classification fields for each relevant term for standardized storage; The tag library is used to set tags for terms for specific classification; The relationship library is used to set applicability restrictions, gender restrictions, amount restrictions and dosage restrictions.

7. A method for claim review based on a knowledge base according to claim 1, characterized in that: The specific process of claim review by the rule engine technology is as follows: Based on the standard library, multiple audit rules are constructed in advance in combination with the tag library and the relationship library; Each of the audit rules is split into multiple conditional nodes. During the claim audit process of the standardized audit data, it is determined whether each conditional node is activated. The activated conditional nodes form a path. If all conditional nodes on the path are activated, the audit rule is deemed to meet the conditions and the corresponding reasoning conclusion is drawn.

8. A method for claim review based on a knowledge base according to claim 1, characterized in that: The specific process of using the deep learning algorithm or large model analysis algorithm to review claims is as follows: Obtain standardized audit data and corresponding audit conclusions in advance to build training data; Machine learning training is performed based on the training data, and the trained deep learning algorithm or large model analysis algorithm is used to perform claims audit on the standardized audit data to generate inference conclusions.

9. A method for claim review based on a knowledge base according to claim 1, characterized in that: The search engine adopts TF-IDF algorithm, BM25 algorithm, BM25L algorithm, BM25+ algorithm, BM25T algorithm or BM25F algorithm; The deep learning algorithm is a recurrent neural network, a BiLSTM-CRF model, or a pre-trained Transformer model; The large model analysis algorithm is the GPT model.

10. A claim review device based on a knowledge base, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, and the processor calls the computer program to execute the steps of any one of the methods according to claims 1 to 9.

Images