Intelligent claim settlement processing method and system based on large model-decision tree coevolution

Abstract

The invention relates to the technical field of claim settlement processing, and discloses an intelligent claim settlement processing method and system based on large model-decision tree coevolution, and the method comprises the steps: analyzing a historical claim settlement case based on a large language model, extracting claim settlement information, generating a decision rule set, and constructing a decision tree; obtaining a new claim settlement case, and extracting a case feature vector; traversing the case feature vector from a root node of the decision tree to a termination node, and outputting a claim decision result and evaluation confidence; comparing the evaluation confidence coefficient with a preset confidence coefficient threshold value, and determining whether the result is manually audited or not; comparing a claim decision result with a manual auditing result; and when a comparison result has a difference, determining a new decision rule, and incrementally integrating the new decision rule into the decision tree. The claim settlement processing period is shortened, the manual participation proportion is reduced, and the overall processing efficiency of the claim settlement service is improved.

Description

Technical Field

[0001] This invention relates to the field of claims processing technology, and more specifically, to an intelligent claims processing method and system based on large model-decision tree co-evolution. Background Technology

[0002] With the continuous improvement of informatization and intelligentization in the insurance industry, the volume of claims for medical and health insurance, accident insurance, and other types of insurance is showing a rapid growth trend. Claims processing, as a crucial link in the insurance business chain, directly affects the operational efficiency, risk control capabilities, and user service experience of insurance companies. However, existing claims processing methods still have many shortcomings in practical applications, urgently requiring more intelligent, efficient, and sustainably evolving technological solutions for improvement.

[0003] Currently, claims processing primarily employs a combination of manual review and automated review using rule engines. On one hand, manual review relies on claims personnel's experience and judgment regarding medical policies, insurance terms, and actual cases. While offering high flexibility, it suffers from low processing efficiency, high labor costs, significant subjective variability, and difficulty in scaling. On the other hand, automated claims systems based on rule engines typically rely on pre-configured, fixed rules. While effective for highly structured cases with relatively stable scenarios, they are prone to insufficient rule coverage and increased misjudgment rates when facing complex and diverse medical practices and varying material quality in actual claims scenarios. In existing technologies, machine learning and deep learning are increasingly being applied to intelligent claims processing, for example, using classification models or neural networks to predict claims outcomes. While such methods improve automation to some extent, they generally suffer from the following shortcomings: the models are mostly "black box" structures, the decision-making process lacks interpretability, and it is difficult to meet the explanation needs of the insurance industry in compliance audits, liability determination, and dispute resolution; the models usually rely on large-scale labeled data for offline training, and when medical policies, insurance terms, or data distribution change, the model update cost is high and the adaptability is insufficient; the models are difficult to effectively integrate human review experience and cannot conduct structured summaries and continuous optimization of erroneous decisions.

[0004] Therefore, it is necessary to design an intelligent claims processing method and system based on large model-decision tree co-evolution to solve the problems existing in the current technology. Summary of the Invention

[0005] In view of this, the present invention proposes an intelligent claims processing method and system based on large model-decision tree co-evolution, which aims to solve the problems of difficulty in integrating human review experience and inability to structurally summarize and continuously optimize erroneous decisions.

[0006] In one aspect, this invention proposes an intelligent claims processing method based on large model-decision tree co-evolution, comprising: Historical claims cases are analyzed using a large language model to extract claims information and generate a set of decision rules; a decision tree is then constructed based on the set of decision rules. New claims are acquired and case feature vectors are extracted. Starting from the root node of the decision tree, the case feature vectors are compared with the condition thresholds of each decision node in the decision tree. The matching path is traversed to the termination node, and the claims decision result is output. The assessment confidence level is determined based on the decision path length, node matching accuracy, and historical verification accuracy. The assessed confidence level is compared with a preset confidence level threshold; if the assessed confidence level is greater than or equal to the confidence level threshold, the new claim case is approved; if the assessed confidence level is less than the confidence level threshold, the new claim case is manually reviewed, and the manual review result is determined. The claims decision results are compared with the manual review results. When there are differences in the comparison results, multi-dimensional clustering analysis and semantic induction are performed on the new claims based on the case feature vector and decision path log to determine new decision rules. The new decision rules are then incrementally integrated into the decision tree.

[0007] Furthermore, when analyzing historical claims cases based on large language models, extracting claims information, and generating decision rule sets, the following are included: Based on the aforementioned large language model, semantic understanding is performed on historical claims cases to extract claims information fields, including diagnosis codes, treatment methods, cost details, and liability attribution. Based on these claims information fields, a structured prompt template is constructed, which includes historical claims case samples and decision rule format requirements. The claims information fields and structured prompt templates are input into the large language model. The correlation pattern between the input features and the decision results is analyzed to generate decision condition statements and corresponding result judgment parameters. The decision condition statements include feature condition combinations and logical operators. The result judgment parameters include claims decision type and compensation calculation parameters. The decision condition statements are iteratively optimized in multiple rounds. The feature conditions, logical relationships and threshold parameters in the decision condition statements are parsed and converted into condition expressions. Based on the logical dependencies between the decision condition statements, a priority system for decision rules is constructed. The condition expressions, result judgment parameters and priority system are organized to form the decision rule set.

[0008] Furthermore, when constructing a decision tree based on the set of decision rules, the following steps are included: The conditional expressions are sorted according to their complexity and feature relevance to determine the node construction order of the decision tree. Rule conflicts are resolved according to the priority system, and node weights are assigned to each conditional expression. The sorted conditional expressions are converted into decision tree nodes, and the feature conditions, logical relationships, and threshold parameters in the conditional expressions are mapped to condition judgment codes. The branch structure of the decision tree is optimized according to the information gain or Gini coefficient of the feature conditions. The decision tree nodes are organized hierarchically to form a tree structure with a root node, internal decision nodes, and a terminal node. Result judgment parameters are added to the terminal node. The decision tree is then constructed.

[0009] Furthermore, when acquiring new claims cases and extracting case feature vectors, the following are included: Obtain the claim application materials for new claims, including medical bills, diagnostic certificates, and expense lists; perform optical character recognition (OCR) processing on the claim application materials to generate structured text data; perform data cleaning on the structured text data to remove noise and fill in missing values; based on preset feature extraction rules, extract diagnostic codes, treatment method categories, expense detail types, medical institution levels, material completeness scores, and OCR recognition confidence from the structured text data, and encode them according to predefined vector dimensions to generate a case feature vector.

[0010] Furthermore, when determining the assessment confidence level based on decision path length, node matching accuracy, and historical validation accuracy, the following factors are considered: The decision path length is determined based on the number of nodes in the decision path; the node matching accuracy is determined by weighted averaging of the matching degree of all nodes in the decision path; the historical verification accuracy is determined by statistically analyzing the proportion of correct decisions made by the decision path in historical claims cases; and the evaluation confidence level is output by comprehensively calculating the decision path length, node matching accuracy, and historical verification accuracy.

[0011] Furthermore, when comparing the claims decision result with the manual review result, the following steps are included: The claims decision results and manual review results are analyzed in a structured manner to extract comparison elements. The comparison elements include decision type, compensation amount, liability ratio, and reason for rejection. Based on the comparison elements, the claims decision results and manual review results are compared for consistency, and a comparison result is generated.

[0012] Furthermore, based on the case feature vector and decision path log, multi-dimensional clustering analysis and semantic induction are performed on the new claims cases to determine new decision rules. When these new decision rules are incrementally integrated into the decision tree, the process includes: The feature vectors of the new claims are integrated with the feature vectors of the erroneous cases in the historical error case database to generate a feature vector set. The historical error case database stores the feature vectors and decision path logs of claims that were previously determined to have substantial differences. Based on the feature vector set, dimensional feature data is extracted, including diagnostic code distribution, cost type features, material integrity index, OCR confidence, and decision path features. The feature vector set is grouped based on the density clustering algorithm, and cases with similar feature patterns are grouped into the same cluster, forming three types of error pattern clusters: bill recognition error cluster, disease liability determination fuzzy cluster, and cost calculation deviation cluster.

[0013] Furthermore, based on the case feature vector and decision path log, when performing multi-dimensional clustering analysis and semantic induction on the new claims case to determine new decision rules, and incrementally integrating the new decision rules into the decision tree, the process also includes: For each error pattern cluster, common and differential features of cases within the cluster are extracted. These common and differential features are then input into the large language model. Based on the large language model, the root causes of errors are analyzed, and an error type description is generated. The error type description includes error triggering conditions, typical features, scope of impact, and severity. The error type description is then associated and mapped with uncovered nodes or error branches of the decision tree to construct an error pattern knowledge base.

[0014] Furthermore, based on the case feature vector and decision path log, when performing multi-dimensional clustering analysis and semantic induction on the new claims case to determine new decision rules, and incrementally integrating the new decision rules into the decision tree, the process also includes: Based on the error pattern knowledge base and the large language model, new decision condition statements and new result determination parameters are generated; the new decision condition statements are converted into new condition expressions; and the new condition expressions and new result determination parameters are integrated into the decision tree.

[0015] Compared with existing technologies, the advantages of this invention are as follows: By introducing a large language model for deep semantic understanding and rule extraction of historical claims cases, and combining it with decision trees for structured reasoning, automatic judgment of new claims cases is achieved. Compared with methods that rely entirely on manual review, this invention shortens the claims processing cycle, reduces the proportion of manual intervention, and improves the overall processing efficiency of claims business while ensuring accuracy. The technical approach of using a large language model to generate rules and decision trees to execute decisions combines the knowledge induction capability of the large model with the path traceability of the decision tree. The claims decision-making process can be clearly explained through decision path logs, avoiding the problem of insufficient interpretability of pure black-box models in financial insurance scenarios, and meeting the needs of regulatory compliance and business auditing. By introducing an assessment confidence mechanism, the decision path length, node matching accuracy, and historical verification accuracy are comprehensively evaluated. Low-confidence cases are automatically triggered for manual review, thereby reducing the risk of erroneous claims and denials. At the same time, the historical verification accuracy is used to constrain the decision results, enhancing stability and reliability in complex and edge scenarios. When there is a discrepancy between the automated claims decision and the manual review result, it does not simply revert to manual processing. Instead, it performs multi-dimensional clustering analysis based on case feature vectors and decision path logs, and combines this with semantic induction using a large language model to automatically discover error patterns and generate new decision rules, thus achieving incremental updates and co-evolution of the decision tree.

[0016] On the other hand, this application also provides an intelligent claims processing system based on large model-decision tree co-evolution, used to apply the above-mentioned intelligent claims processing method based on large model-decision tree co-evolution, including: The construction unit is configured to analyze historical claims cases based on a large language model, extract claims information and generate a set of decision rules; and construct a decision tree based on the set of decision rules. The output unit is configured to acquire new claims cases and extract case feature vectors; starting from the root node of the decision tree, compare the case feature vectors with the condition thresholds of each decision node in the decision tree, traverse along the matching path to the termination node, and output the claims decision result; and determine the assessment confidence level based on the decision path length, node matching accuracy, and historical verification accuracy. The judgment unit is configured to compare the assessed confidence level with a preset confidence level threshold; if the assessed confidence level is greater than or equal to the confidence level threshold, the new claim case is approved; if the assessed confidence level is less than the confidence level threshold, the new claim case is manually reviewed, and the manual review result is determined. The optimization unit is configured to compare the claims decision results with the manual review results; when there are differences in the comparison results, based on the case feature vector and decision path log, perform multi-dimensional clustering analysis and semantic induction on the new claims case, determine new decision rules, and incrementally integrate the new decision rules into the decision tree.

[0017] It is understandable that the above-mentioned intelligent claims processing method and system based on large model-decision tree co-evolution have the same beneficial effects, and will not be elaborated further here. Attached Figure Description

[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart illustrating the intelligent claims processing method based on large model-decision tree co-evolution provided in an embodiment of the present invention; Figure 2 This is a functional block diagram of an intelligent claims processing system based on large model-decision tree co-evolution, provided in an embodiment of the present invention. Detailed Implementation

[0019] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, embodiments and features in the embodiments of the present invention can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] In some embodiments of this application, see Figure 1 As shown, this application proposes an intelligent claims processing method based on large model-decision tree co-evolution, including: S100: Analyze historical claims cases based on a large language model, extract claims information and generate a set of decision rules; construct a decision tree based on the set of decision rules.

[0021] S200: Obtain new claims cases and extract case feature vectors; starting from the root node of the decision tree, compare the case feature vectors with the condition thresholds of each decision node in the decision tree, traverse along the matching path to the termination node, and output the claims decision result; and determine the assessment confidence level based on the decision path length, node matching accuracy, and historical verification accuracy.

[0022] S300: Compare the assessed confidence level with the preset confidence level threshold; if the assessed confidence level is greater than or equal to the confidence level threshold, the new claim case is approved; if the assessed confidence level is less than the confidence level threshold, the new claim case is manually reviewed, and the result of the manual review is determined.

[0023] S400: Compare the claims decision results with the manual review results; when there are differences in the comparison results, perform multi-dimensional clustering analysis and semantic induction on new claims cases based on case feature vectors and decision path logs, determine new decision rules, and integrate the new decision rules incrementally into the decision tree.

[0024] Specifically, in step S100, historical claims cases stored in the insurance business are first centrally acquired. These historical claims cases include claims application materials, claims conclusions, and corresponding manual review records. A large language model is introduced to perform semantic understanding and knowledge extraction on the unstructured text content of these historical claims cases, automatically identifying and extracting key information fields related to claims decisions, such as diagnosis codes, treatment methods, cost composition, liability attribution, and material completeness. Based on this, a structured prompt template containing historical case samples and rule output format constraints is constructed. Claim information fields are input into the large language model to analyze the correlation between case characteristics and the final claims result, generating decision condition statements with clear condition combinations and logical relationships, along with corresponding result judgment parameters. Further, the generated decision condition statements undergo multiple rounds of parsing and optimization, transforming them into standardized condition expressions. Based on the logical dependencies between rules and business priorities, a structured set of decision rules is formed. Then, based on the decision rule set, a decision tree structure is constructed according to indicators such as condition complexity, feature relevance, and information gain or Gini coefficient, ensuring the decision tree has clear hierarchical relationships and executable judgment logic.

[0025] In step S200, when a new claim is received, the claim application materials for that new claim are processed, including optical character recognition (OCR) of medical invoices, diagnostic certificates, and expense lists to generate structured text data. This structured text data is then cleaned, denoised, and imputed for missing values. Subsequently, based on preset feature extraction rules, features such as diagnostic codes, treatment method categories, expense detail types, hospital level, material completeness scores, and OCR recognition confidence are extracted from the structured text data and encoded according to predefined vector dimensions to form a case feature vector. Starting from the root node of the decision tree, the case feature vector is matched against conditional thresholds at each level of the decision tree, traversing downwards along branches that meet the conditions until the termination node is reached, at which point the corresponding claim decision result is output. Simultaneously, the decision process is quantitatively evaluated by comprehensively calculating the decision path length, the matching accuracy of each node, and the verification accuracy of the path in historical cases, thereby obtaining an evaluation confidence level used to measure the reliability of the decision.

[0026] In step S300, the assessed confidence level is compared with a pre-set confidence level threshold. When the assessed confidence level is greater than or equal to the confidence level threshold, it indicates that the current claims decision path is stable and reliable, and the new claims case is automatically approved and enters the subsequent payment process. When the assessed confidence level is less than the confidence level threshold, it is determined that the case has uncertainty or potential risk, and the new claims case is automatically transferred to the manual review process. The claims personnel review the case based on their business experience and the insurance terms, and generate the corresponding manual review result.

[0027] In step S400, the automatically generated claims decision results and the manually reviewed results are structurally analyzed to extract comparison elements including decision type, compensation amount, liability ratio, and reasons for rejection. A consistency comparison analysis is then performed between the two. When the comparison results show substantial differences, the new claims case is marked as an abnormal learning sample. Combining its case feature vector and complete decision path log with historical error case data, multi-dimensional clustering analysis and semantic induction are performed to identify common features and error patterns leading to misjudgments. Based on this, a large language model is used to generate new decision condition statements and result judgment parameters, which are then converted into executable conditional expressions and incrementally integrated into the existing decision tree. This achieves continuous supplementation and optimization of decision rules, promoting the collaborative evolution and performance improvement of the claims decision model in actual business operations.

[0028] Specifically, the confidence threshold is determined based on the consistency between automated decisions and manual review results in historical claims cases. By statistically analyzing the accuracy of automated claims decision paths in historical cases, the false positive rate and human intervention rate are assessed across different confidence intervals. The confidence interval that maximizes the proportion of automated processing while ensuring a low risk of erroneous claims is selected. This threshold should ensure strong reliability of claims decisions at high confidence levels and avoid hasty automated decisions at low confidence levels, thus reducing the risk of errors. The decision path length is determined based on the depth distribution of historical claims decision paths. By analyzing the decision path depth of different cases in historical data, the most representative and comprehensive path length interval is selected, avoiding paths that are too short, resulting in a lack of detailed analysis, or paths that are too long, leading to computational complexity and runtime delays. The path length should be moderate to balance decision accuracy and processing efficiency. Node matching accuracy is set by assessing the impact of each node in the historical case data on the claims outcome. The contribution of each feature condition to decision accuracy is weighted, and the matching accuracy of each node is quantitatively evaluated using standards such as information gain or the Gini coefficient. Nodes with higher predictive power and discriminative power are assigned higher accuracy weights to ensure that key nodes in the decision-making process provide more information and reduce the probability of misjudgment. Historical verification accuracy is determined by reviewing recent claims cases and calculating the consistency ratio between automated decisions and manual reviews. The statistical period for accuracy is typically selected from claims data within the past quarter or six months to reflect the effectiveness and accuracy of current decision-making rules. If a decision-making rule shows high accuracy in historical cases, it will be prioritized to ensure the accuracy of new claims processing. In cluster analysis, clustering is performed using the feature vectors of historical error cases. Parameters such as the clustering distance threshold and minimum sample size are dynamically set based on the distribution density of the feature data and the number of error cases. The distance threshold ensures that cases with similar characteristics are correctly classified into the same category, while the minimum sample size prevents unstable clustering results due to insufficient samples. By adjusting these parameters, different error patterns are accurately distinguished while avoiding over-clustering or under-clustering. When discrepancies are found between the results of manual review and automated decision-making, new decision condition statements are generated based on the common and dissimilar characteristics of historical error cases. The generation of these new decision condition statements follows rules based on semantic analysis of case features using a large language model. By analyzing the root cause of each error pattern, a clear error type description is generated, and the decision logic is adjusted accordingly. This approach not only resolves misjudgments in current cases but also continuously accumulates new decision rules to enhance future decision-making capabilities.

[0029] Understandably, combining the semantic understanding and knowledge induction capabilities of large language models with the structured reasoning and interpretability advantages of decision trees improves the automation level and decision accuracy of claims processing. By automatically generating decision rules and constructing decision trees through in-depth analysis of historical claims cases, rapid and stable judgment of new claims is achieved, shortening the claims processing cycle and reducing manual review costs. The introduction of a confidence assessment mechanism based on decision path length, node matching accuracy, and historical verification accuracy automatically triggers manual review for low-confidence cases, controlling the risk of erroneous claims and denials, and enhancing the reliability and security of claims decisions. When discrepancies exist between automated claims results and manual review results, cluster analysis and semantic induction based on case feature vectors and decision path logs can be performed to automatically generate and incrementally integrate new decision rules, enabling continuous learning and co-evolution of the decision model. This allows claims to dynamically adapt to changes in business rules, medical policies, and data distribution.

[0030] In some embodiments of this application, when analyzing historical claims cases based on a large language model, extracting claims information, and generating a set of decision rules, the following steps are included: Based on a large language model, semantic understanding of historical claims cases is performed to extract claims information fields, including diagnosis codes, treatment methods, cost details, and liability attribution. Based on the claims information fields, a structured prompt template is constructed, which includes historical claims case samples and decision rule format requirements. Input the claims information fields and structured prompt templates into the large language model, analyze the correlation pattern between input features and decision results, and generate decision condition statements and corresponding result judgment parameters. The decision condition statements include feature condition combinations and logical operators, and the result judgment parameters include claims decision type and compensation calculation parameters. The decision condition statements are iteratively optimized in multiple rounds. The characteristic conditions, logical relationships and threshold parameters in the decision condition statements are parsed and converted into condition expressions. Based on the logical dependencies between decision condition statements, a priority system for decision rules is constructed. The condition expressions, result judgment parameters and priority system are organized into a set of decision rules.

[0031] Specifically, semantic understanding of historical claims cases is performed based on a large language model. Historical claims cases typically contain a large amount of unstructured or semi-structured text information, such as claim application instructions, medical record summaries, diagnostic certificates, expense lists, and manual review comments. By utilizing the natural language understanding capabilities of the large language model, contextual semantic parsing and entity recognition are performed on the above text content to automatically identify key claim information fields closely related to the claim decision. Claim information fields include at least the diagnostic code (such as the standard code corresponding to the disease or injury), treatment method (such as outpatient treatment, inpatient treatment, type of surgery, etc.), expense details (such as drug costs, examination fees, surgical fees and their amount distribution), liability attribution (such as whether it falls within the scope of insurance liability, and the application of exclusion clauses), and the completeness of medical materials (such as whether invoices, diagnostic certificates, and medical records are complete). Through the above semantic understanding process, the original historical claims cases are transformed into a structured or semi-structured data foundation that can be used for rule analysis. Structured prompt templates are then constructed based on the claim information fields. Structured prompt templates are used to standardize the input and output formats of the large language model. Their content includes not only extracted claims information fields but also typical historical claims case samples and explicit decision rule format requirements. For example, the prompt template may specify that decision rules must be output in the form of "If condition ×× is met, then execute the ×× claims result," and requires explicit provision of condition combinations, logical operators, and result parameters. By introducing structured prompt templates, the generation scope of the large language model can be constrained, reducing the uncertainty of free text output and improving the stability, consistency, and parsability of generated rules. Claims information fields and structured prompt templates are input into the large language model to perform rule induction analysis on historical claims cases. Based on its pre-trained knowledge and contextual reasoning capabilities, the large language model comprehensively analyzes the correlation between different case characteristics and the final claims result, automatically summarizing decision condition statements with discriminative power and generating corresponding result judgment parameters for each decision condition statement. The decision condition statement describes the combination relationship between multiple feature conditions, including feature condition items and their logical operators, such as "AND", "OR", and "NOT". The result judgment parameter describes the type of claim decision to be output when the condition is met (such as normal payment, partial payment, and claim rejection) and the corresponding payment calculation parameters or formulas (the payment calculation parameters or formulas are determined through the analysis of several historical claim cases), so that the generated rules can directly guide the execution of claim decisions. The decision condition statement undergoes multiple rounds of iterative optimization and standardization. By parsing the initially generated decision condition statement, the feature conditions, logical relationships, and threshold parameters it contains are decomposed, and ambiguous, redundant, or conflicting conditions are corrected and simplified, converting the natural language form of the condition description into a standardized condition expression that can be directly executed.Simultaneously, based on the logical dependencies, applicable scope, and business importance among different decision condition statements, a priority system for decision rules is constructed to ensure that when multiple rules are simultaneously satisfied, they can be matched and executed in a predetermined order, avoiding rule conflicts or overriding errors. Finally, the condition expressions, result judgment parameters, and priority system are uniformly organized and encapsulated to form a structured and manageable set of decision rules. This set of decision rules serves as the foundational input for subsequently building the decision tree. It not only retains the business knowledge summarized from historical cases by the large language model but also ensures the executability and stability of the rules through standardization and priority constraints.

[0032] Understandably, by accurately extracting and correlating key information fields such as diagnosis codes, treatment methods, cost details, liability attribution, and the completeness of medical materials, the coverage and matching accuracy of claims rules for actual business scenarios have been improved, reducing omissions and deviations in manual rule design. With the help of structured prompt templates and multi-round iterative optimization mechanisms, the generated decision condition statements are more standardized, stable, and easy to parse in terms of logical expression, threshold setting, and result judgment parameters. Furthermore, by constructing a rule priority system, rule conflicts have been avoided, thus forming a set of decision rules that is clear in hierarchy and highly interpretable.

[0033] In some embodiments of this application, constructing a decision tree based on a set of decision rules includes: The conditional expressions are sorted according to their complexity and feature relevance to determine the node construction order of the decision tree. Rule conflicts are resolved based on the priority system, and node weights are assigned to each conditional expression. The sorted conditional expressions are converted into decision tree nodes, and the feature conditions, logical relationships, and threshold parameters in the conditional expressions are mapped to condition judgment codes. The branch structure of the decision tree is optimized according to the information gain or Gini coefficient of the feature conditions. The decision tree nodes are organized hierarchically to form a tree structure with a root node, internal decision nodes, and a terminal node. Result judgment parameters are added to the terminal node. The decision tree is then constructed.

[0034] Specifically, the conditional expressions are first sorted according to their complexity and feature relevance to determine the node construction order of the decision tree. Conditional complexity measures the number of features involved, the depth of logical operators, and the fineness of threshold judgments. Conditions with lower complexity and stronger discriminative power are prioritized for placement closer to upper-level nodes in the decision tree. Feature relevance is assessed by evaluating the correlation or information contribution between each feature and the claim decision result in historical claims data; highly relevant features are placed closer to the root node. This sorting method allows the decision tree to effectively distinguish most cases at a shallower level, improving overall reasoning efficiency. Based on the priority system of decision rules, rules with overlapping applicability or potential conflicts are uniformly coordinated, and each conditional expression is assigned a corresponding node weight. Node weights reflect the importance of the condition in the decision-making process, and their values ​​can comprehensively consider factors such as the reliability of the rule source, historical verification accuracy, and business risk level. When multiple conditions are simultaneously met or conflicting judgments occur, the condition with the higher node weight is executed first, ensuring the consistency and stability of the decision logic and avoiding overlapping or contradictory conclusions between different rules. The sorted conditional expressions are converted into decision tree nodes one by one, and the feature conditions, logical relationships, and threshold parameters contained in the conditional expressions are mapped into directly executable condition judgment code. Specifically, each conditional expression corresponds to an internal decision node, which encapsulates the corresponding feature judgment logic. For example, it performs Boolean judgments on whether the diagnostic code belongs to a specified set, whether the cost amount exceeds a preset threshold, or whether the material integrity score meets the requirements. This enables the decision tree to quickly match and select branches from the input case feature vector at runtime. The branch structure of the decision tree is optimized according to the information gain or Gini coefficient of each feature condition in historical claims samples. When constructing branches, features that can minimize sample uncertainty or classification confounding are prioritized as classification criteria, thereby making the case distribution on different branches more concentrated and reducing invalid branches and redundant judgments. This process helps improve the discriminative ability and generalization performance of the decision tree and reduces the risk of misjudgment caused by noisy data or abnormal samples. The decision tree nodes are organized according to hierarchical relationships, and a tree structure containing a root node, internal decision nodes, and a terminal node is constructed layer by layer. The system is structured with a root node for making the most basic and comprehensive conditional judgments, internal decision nodes for progressively refining and differentiating case characteristics, and a termination node for outputting the final claims settlement conclusion. This clear hierarchical structure ensures the decision-making path is traceable and interpretable, facilitating subsequent auditing and manual review. The termination node incorporates corresponding result determination parameters, including the specific claims settlement decision type and its corresponding compensation calculation formula or parameter configuration.When the case feature vector reaches the terminal node along the decision path, the claim result can be directly output based on the result judgment parameters configured in that node, and the compensation amount can be calculated. This completes the mapping from rules to decisions, and ultimately builds a decision tree model that is clear in structure, efficient in execution, and easy to maintain.

[0035] Understandably, by sorting conditional expressions according to their complexity and feature relevance, and assigning node weights based on a rule priority system, rule conflicts and redundant judgments are reduced. This allows key features to be prioritized at higher levels of the decision tree, thereby reducing the decision path length and improving inference efficiency. Optimizing the branch structure based on information gain or the Gini coefficient gives the decision tree stronger discriminative power and generalization performance, reducing misjudgments and invalid branches. By directly configuring the claims decision type and compensation calculation formula in the termination node, the decision result output is highly consistent with the compensation calculation process, ensuring clear logic.

[0036] In some embodiments of this application, when acquiring a new claim case and extracting the case feature vector, the process includes: The system obtains claim application materials for new claims, including medical bills, diagnostic certificates, and expense lists. It then performs optical character recognition (OCR) processing on these materials to generate structured text data. The structured text data is cleaned to remove noise and fill in missing values. Based on predefined feature extraction rules, the system extracts diagnostic codes, treatment method categories, expense detail types, hospital level, material completeness scores, and OCR recognition confidence from the structured text data, and encodes these features according to predefined vector dimensions to generate a case feature vector.

[0037] Specifically, this involves acquiring claim application materials for new claims. Claim application materials are submitted by the insured or policyholder through online or offline channels, typically including original materials in various forms such as medical bills, diagnostic certificates, and expense lists. These materials may include scanned copies of paper documents, photographs, and possibly electronic documents or invoices. These materials come from diverse sources and have inconsistent formats, serving as the primary basis for subsequent claim assessments. Therefore, centralized collection and unified management are necessary to provide complete data input for subsequent processing. Optical Character Recognition (OCR) processing is then performed on the claim application materials to generate structured text data. For medical bills, diagnostic certificates, and expense lists in image or scanned form, the OCR module is invoked to recognize and parse the text, numbers, and tabular information. The recognition results are then converted into structured text data according to predefined data templates, such as mapping item names, amounts, dates, and hospital names from the bills to corresponding fields. This step converts image data that was originally uncalculated and unanalyzable into text and field data that can be processed by algorithms. Finally, the structured text data undergoes data cleaning to improve data quality and usability. The system identifies and removes obviously erroneous or invalid noise data, such as duplicate records, abnormal characters, and incorrectly formatted fields, through rule-based validation and statistical analysis. Simultaneously, it fills in missing key information fields. Filling methods include supplementing information based on other materials within the same case, making reasonable estimates based on historical statistical distributions, or marking fields as missing, to avoid biases in subsequent feature extraction or decision-making due to incomplete data. After data cleaning, core features for claims decision-making are extracted from structured text data based on pre-defined feature extraction rules. These rules are formulated based on insurance terms, medical management rules, and historical claims experience, and are used to abstract complex text information into discriminative feature variables. Specifically, this includes: extracting diagnostic codes from diagnostic certificates to characterize disease or injury types; identifying treatment methods, such as outpatient, inpatient, or surgical types, from medical records; summarizing expense details and their composition ratios from expense lists; extracting the level of medical institutions from medical information to reflect differences in medical institution qualifications; and calculating a material completeness score based on the completeness of submitted materials, combined with the recognition confidence score output by the OCR module as an auxiliary feature to measure the reliability of text recognition. Finally, the extracted features are uniformly encoded according to predefined vector dimensions and encoding methods to generate case feature vectors. Encoding methods can include numerical encoding, interval mapping, or one-hot encoding, enabling different types of features to be represented in the same vector space. By generating case feature vectors in a unified format, new claims cases can be used as standardized input.

[0038] Understandably, by performing optical character recognition and data cleaning on materials such as medical bills, diagnostic certificates, and expense lists, the cost of manual data entry is reduced, the interference caused by noisy data and missing information on claims judgment is reduced, and the accuracy and consistency of the original data are improved. Based on preset feature extraction rules, key factors such as diagnostic codes, treatment methods, expense types, level of medical institutions, material completeness scores, and OCR recognition confidence are uniformly encoded to form case feature vectors. This allows complex claims information to participate in model calculations in a unified dimension, thereby improving the stability, repeatability, and automation level of the claims decision-making process.

[0039] In some embodiments of this application, when determining the assessment confidence level based on decision path length, node matching accuracy, and historical verification accuracy, the following steps are included: The decision path length is determined by the number of nodes in the decision path; the node matching accuracy is determined by weighted averaging of the matching degree of all nodes in the decision path; the historical verification accuracy is determined by statistically analyzing the proportion of correct decisions made by the decision path in historical claims cases; and the decision path length, node matching accuracy, and historical verification accuracy are comprehensively calculated to output the assessment confidence level.

[0040] Specifically, the length of the decision path is determined based on the number of nodes in the decision path. A decision path refers to the complete judgment process where the case feature vector starts from the root node of the decision tree, passes through multiple internal decision nodes, and finally reaches the terminal node. The length of the decision path is obtained by counting the number of decision nodes included in the path. Generally, the length of the decision path reflects the level of detail in the decision-making process. A path that is too short may indicate insufficient basis for judgment and a coarse decision, while a path that is too long may introduce too many marginal conditions or noisy judgments. Therefore, path length, as one of the important indicators for measuring decision stability, is included in the calculation of confidence level. The matching degree of all nodes in the decision path is weighted and averaged to determine the node matching accuracy. The node matching degree is used to measure the degree of matching between the case feature vector and the condition judgments of each decision node, such as the closeness of feature values ​​to node thresholds and the consistency of category judgments. For each node in the decision path, the corresponding matching degree is calculated, and different weight coefficients are assigned according to the node's hierarchical position in the decision tree, node weight, or feature importance. The matching degrees of each node are then weighted and averaged to obtain the node matching accuracy, which reflects the consistency of judgments along the entire decision path. This indicator effectively reflects the logical consistency and credibility of the current case's decision-making process. The proportion of correct decisions made by the decision-making path in historical claims cases is statistically analyzed to determine the historical verification accuracy rate. A sample of cases that have followed the same or highly similar decision-making paths in the historical claims case database is selected, and the proportion of cases where the automated claims decision result matches the final manual review or confirmed result is calculated. This is used as the historical verification accuracy rate of the decision-making path. This indicator reflects the performance level of the path in past actual business operations and can assess the reliability of the current decision-making path from an experience perspective. The decision-making path length, node matching accuracy, and historical verification accuracy rate are comprehensively calculated to output the assessment confidence level. Following a preset weighting rule or function model, the above three indicators are normalized and then weighted and fused to form an assessment confidence level value that characterizes the credibility of the current claims decision.

[0041] Understandably, by comprehensively considering the decision path length, node matching accuracy, and historical verification accuracy, the system reflects the level of detail in the decision-making process, the degree of compliance between the current case and the rule, and the successful experience of the path in historical business, thus identifying low-confidence or potentially abnormal cases. By using quantitative assessment confidence as a criterion, it can automatically distinguish between high-confidence cases that can be directly approved and low-confidence cases that require manual review, thereby improving the efficiency and accuracy of claims processing and reducing the risk of erroneous claims and denials.

[0042] In some embodiments of this application, comparing the claims decision result with the manual review result includes: The claims decision results and manual review results are analyzed in a structured manner to extract comparison elements. The comparison elements include decision type, compensation amount, liability ratio and reason for rejection. Based on the comparison elements, the claims decision results and manual review results are compared for consistency and a comparison result is generated.

[0043] Specifically, the system first performs unified structured parsing on both automatically generated claims decision results and manually reviewed results. This maps the results, which might otherwise exist in text, table, or field formats, into a standardized data structure, eliminating format differences caused by different sources and expression methods. During the structured parsing process, key comparison elements are extracted for comparative analysis. These elements include at least the claims decision type (e.g., approved, partially approved, or rejected), the final confirmed compensation amount, the liability ratio of each responsible party, and the description of the rejection reason in the rejection scenario. This ensures that the automated decision results and manually reviewed results are comparable across the same semantic and field dimensions. Based on this, a consistency comparison analysis is conducted item by item based on the comparison elements. For numerical elements (e.g., compensation amount and liability ratio), consistency is determined through numerical difference calculations and threshold judgments. For categorical or textual elements (e.g., decision type and rejection reason), differences are determined through rule matching or semantic consistency judgments. Finally, the comparison conclusions of all comparison elements are synthesized to generate a structured comparison result, which characterizes the degree of consistency and type of difference between the automated claims decision and the manually reviewed conclusion.

[0044] Understandably, by comparing core elements such as decision type, compensation amount, liability ratio, and reasons for claim rejection item by item, the deviation location and type of difference between automatic claims decision and manual review conclusion can be accurately identified, avoiding the neglect of key details by relying solely on the overall result. The generated comparison results can serve as an important basis for anomaly sample identification and rule learning, providing high-quality feedback data for subsequent optimization of decision rules, adjustment of decision tree structure, and co-evolution of large models.

[0045] In some embodiments of this application, when performing multi-dimensional clustering analysis and semantic induction on new claims cases based on case feature vectors and decision path logs to determine new decision rules, and incrementally integrating the new decision rules into the decision tree, the process includes: The feature vectors of new claims are integrated with the feature vectors of erroneous cases in the historical error case database to generate a feature vector set. The historical error case database stores the feature vectors and decision path logs of claims that were previously determined to have substantial differences. Based on the feature vector set, dimensional feature data is extracted, including diagnostic code distribution, cost type features, material integrity index, OCR confidence, and decision path features. Based on the density clustering algorithm, the feature vector set is grouped, and cases with similar feature patterns are clustered into the same cluster, forming three types of error pattern clusters: bill recognition error cluster, disease liability determination fuzzy cluster, and cost calculation deviation cluster.

[0046] In some embodiments of this application, when performing multi-dimensional clustering analysis and semantic induction on new claims cases based on case feature vectors and decision path logs to determine new decision rules, and incrementally integrating the new decision rules into the decision tree, the method further includes: For each error pattern cluster, common and dissimilar features of cases within the cluster are extracted. These features are then input into a large language model. Based on the large language model, the root causes of errors are analyzed, and error type descriptions are generated. The error type descriptions include error triggering conditions, typical features, scope of impact, and severity. The error type descriptions are then mapped to uncovered nodes or error branches of the decision tree to construct an error pattern knowledge base.

[0047] In some embodiments of this application, when performing multi-dimensional clustering analysis and semantic induction on new claims cases based on case feature vectors and decision path logs to determine new decision rules, and incrementally integrating the new decision rules into the decision tree, the method further includes: Based on the error pattern knowledge base and using a large language model, new decision condition statements and new result determination parameters are generated; the new decision condition statements are converted into new condition expressions; and the new condition expressions and new result determination parameters are integrated into the decision tree.

[0048] Specifically, when performing multi-dimensional clustering analysis and semantic induction on new claims cases based on case feature vectors and decision path logs to determine new decision rules and achieve incremental integration, the first step is to integrate the case feature vectors corresponding to the new claims cases with the feature vectors of claims cases in the historical error case database that have been confirmed to have substantial differences, forming a feature vector set for analysis. The historical error case database not only stores the feature vectors of each error case but also synchronously saves the corresponding decision path logs to fully reconstruct the judgment process when the error occurred. Based on this, multi-dimensional feature data is extracted from the feature vector set. These multi-dimensional feature data include diagnostic coding distribution features to characterize differences in disease types and applicable terms, expense type features reflecting the reasonableness of expense composition, material integrity indicators to measure the compliance of materials, OCR recognition confidence features reflecting the reliability of text recognition, and decision path features reflecting the rule matching process, thus providing comprehensive data support for subsequent error pattern analysis. Subsequently, the feature vector set is automatically grouped based on a density clustering algorithm, aggregating claims cases with similar distribution characteristics and similar misjudgment behaviors in the multi-dimensional feature space into the same cluster, forming a cluster of invoice recognition errors, a cluster of ambiguous disease liability determination, and a cluster of expense errors. Error pattern clusters of different types, such as bias clusters, are calculated to distinguish decision biases from different sources and mechanisms. After classifying the error pattern clusters, common and differential features of cases within each cluster are extracted. These features are then input into a large language model for semantic analysis and inductive reasoning to identify the root causes of recurring errors. This process generates a structured error type description, which details the error triggering conditions, typical characteristics, scope of impact, and severity. This error type description is then mapped to uncovered decision nodes or identified error branches in the decision tree. The system gradually builds a sustainable and expandable error pattern knowledge base. Finally, based on the error pattern knowledge base and with the help of a large language model, it automatically generates targeted new decision condition statements and corresponding new result judgment parameters. The new decision condition statements are then parsed and converted into new condition expressions that can be directly executed. After completing conflict detection, priority verification, and historical data verification with existing rules, the new condition expressions and new result judgment parameters are incrementally integrated into the existing decision tree. This allows for the continuous supplementation and optimization of decision rules without disrupting the original stable structure, and promotes the self-learning, self-correction, and collaborative evolution of the intelligent claims decision system in actual business operations.

[0049] Understandably, by integrating the feature vectors of new claims cases with historical erroneous cases, and using density clustering algorithms to identify different types of error pattern clusters such as invoice recognition errors, ambiguous disease liability determinations, and cost calculation deviations, the common patterns and underlying causes of claims misjudgments can be systematically revealed, avoiding the instability caused by rule adjustments driven by single cases. Combining large language models to perform semantic analysis and summarization of error patterns, an error pattern knowledge base is constructed, and targeted decision condition statements and result judgment parameters are automatically generated. This allows new rules to accurately fill in the original uncovered or error-prone judgment nodes in the decision tree, and smoothly integrate them into the existing decision tree structure in an incremental manner. Thus, without affecting the existing stable decision-making capabilities, the adaptability and decision accuracy of intelligent claims to complex, ambiguous, and abnormal scenarios are improved, promoting the development of the claims rule system towards continuous evolution and self-learning.

[0050] Based on another preferred embodiment described above, see [link to preferred embodiment]. Figure 2 As shown, this embodiment provides an intelligent claims processing system based on large model-decision tree co-evolution, used to apply the above-mentioned intelligent claims processing method based on large model-decision tree co-evolution, including: The building unit is configured to analyze historical claims cases based on a large language model, extract claims information and generate a set of decision rules; and build a decision tree based on the set of decision rules. The output unit is configured to acquire new claims cases and extract case feature vectors; starting from the root node of the decision tree, the case feature vectors are compared with the condition thresholds of each decision node in the decision tree, and the matching path is traversed to the termination node to output the claims decision result; and the assessment confidence level is determined based on the decision path length, node matching accuracy, and historical verification accuracy. The judgment unit is configured to compare the assessed confidence level with a preset confidence level threshold; if the assessed confidence level is greater than or equal to the confidence level threshold, the new claim case is approved; if the assessed confidence level is less than the confidence level threshold, the new claim case is subject to manual review, and the manual review result is determined. The optimization unit is configured to compare the claims decision results with the manual review results; when there are differences in the comparison results, it performs multi-dimensional clustering analysis and semantic induction on new claims cases based on case feature vectors and decision path logs, determines new decision rules, and integrates the new decision rules incrementally into the decision tree.

[0051] In summary, by introducing a large language model for deep semantic understanding and rule extraction of historical claims cases, and combining it with decision trees for structured reasoning, automatic judgment of new claims cases is achieved. Compared to methods relying entirely on manual review, this invention shortens the claims processing cycle, reduces the proportion of manual intervention, and improves the overall processing efficiency of claims business while ensuring accuracy. The technical approach of using a large language model to generate rules and decision trees to execute decisions combines the knowledge induction capability of the large model with the path traceability of the decision tree. The claims decision-making process can be clearly explained through decision path logs, avoiding the problem of insufficient interpretability of pure black-box models in financial insurance scenarios, and meeting the requirements of regulatory compliance and business auditing. By introducing an assessment confidence mechanism, the decision path length, node matching accuracy, and historical verification accuracy are comprehensively evaluated. Cases with low confidence are automatically triggered for manual review, thereby reducing the risk of erroneous claims and denials. At the same time, the historical verification accuracy is used to constrain the decision results, enhancing stability and reliability in complex and edge scenarios. When there is a discrepancy between the automated claims decision and the manual review result, it does not simply revert to manual processing. Instead, it performs multi-dimensional clustering analysis based on case feature vectors and decision path logs, and combines this with semantic induction using a large language model to automatically discover error patterns and generate new decision rules, thus achieving incremental updates and co-evolution of the decision tree.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. An intelligent claims processing method based on large model-decision tree co-evolution, characterized in that, include: Historical claims cases are analyzed using a large language model to extract claims information and generate a set of decision rules; a decision tree is then constructed based on the set of decision rules. New claims are acquired and case feature vectors are extracted. Starting from the root node of the decision tree, the case feature vectors are compared with the condition thresholds of each decision node in the decision tree. The matching path is traversed to the termination node, and the claims decision result is output. The assessment confidence level is determined based on the decision path length, node matching accuracy, and historical verification accuracy. The assessed confidence level is compared with a preset confidence level threshold; if the assessed confidence level is greater than or equal to the confidence level threshold, the new claim case is approved; if the assessed confidence level is less than the confidence level threshold, the new claim case is manually reviewed, and the manual review result is determined. The claims decision results are compared with the manual review results. When there are differences in the comparison results, multi-dimensional clustering analysis and semantic induction are performed on the new claims based on the case feature vector and decision path log to determine new decision rules. The new decision rules are then incrementally integrated into the decision tree.

2. The intelligent claims processing method based on large model-decision tree co-evolution as described in claim 1, characterized in that, When analyzing historical claims cases based on large language models, extracting claims information, and generating a set of decision rules, the following are included: Based on the aforementioned large language model, semantic understanding is performed on historical claims cases to extract claims information fields, including diagnosis codes, treatment methods, cost details, and liability attribution. Based on these claims information fields, a structured prompt template is constructed, which includes historical claims case samples and decision rule format requirements. The claims information fields and structured prompt templates are input into the large language model. The correlation pattern between the input features and the decision results is analyzed to generate decision condition statements and corresponding result judgment parameters. The decision condition statements include feature condition combinations and logical operators. The result judgment parameters include claims decision type and compensation calculation parameters. The decision condition statements are iteratively optimized in multiple rounds. The feature conditions, logical relationships and threshold parameters in the decision condition statements are parsed and converted into condition expressions. Based on the logical dependencies between the decision condition statements, a priority system for decision rules is constructed. The condition expressions, result judgment parameters and priority system are organized to form the decision rule set.

3. The intelligent claims processing method based on large model-decision tree co-evolution as described in claim 2, characterized in that, When constructing a decision tree based on the set of decision rules, the following is included: The conditional expressions are sorted according to their complexity and feature relevance to determine the node construction order of the decision tree. Rule conflicts are resolved according to the priority system, and node weights are assigned to each conditional expression. The sorted conditional expressions are converted into decision tree nodes, and the feature conditions, logical relationships, and threshold parameters in the conditional expressions are mapped to condition judgment codes. The branch structure of the decision tree is optimized according to the information gain or Gini coefficient of the feature conditions. The decision tree nodes are organized hierarchically to form a tree structure with a root node, internal decision nodes, and a terminal node. Result judgment parameters are added to the terminal node. The decision tree is then constructed.

4. The intelligent claims processing method based on large model-decision tree co-evolution as described in claim 3, characterized in that, When acquiring new claims cases and extracting case feature vectors, the following are included: Obtain the claim application materials for new claims, including medical bills, diagnostic certificates, and expense lists; perform optical character recognition (OCR) processing on the claim application materials to generate structured text data; perform data cleaning on the structured text data to remove noise and fill in missing values; based on preset feature extraction rules, extract diagnostic codes, treatment method categories, expense detail types, medical institution levels, material completeness scores, and OCR recognition confidence from the structured text data, and encode them according to predefined vector dimensions to generate a case feature vector.

5. The intelligent claims processing method based on large model-decision tree co-evolution as described in claim 4, characterized in that, When determining the assessment confidence level based on decision path length, node matching accuracy, and historical validation accuracy, the following are included: The decision path length is determined based on the number of nodes in the decision path; the node matching accuracy is determined by weighted averaging of the matching degree of all nodes in the decision path; the proportion of correct decisions made by the decision path in historical claims cases is statistically analyzed, and the historical verification accuracy is determined; the decision path length, node matching accuracy, and historical verification accuracy are comprehensively calculated to output the evaluation confidence level.

6. The intelligent claims processing method based on large model-decision tree co-evolution as described in claim 5, characterized in that, When comparing the claims decision results with the manual review results, the following is included: The claims decision results and manual review results are analyzed in a structured manner to extract comparison elements. The comparison elements include decision type, compensation amount, liability ratio, and reason for rejection. Based on the comparison elements, the claims decision results and manual review results are compared for consistency, and a comparison result is generated.

7. The intelligent claims processing method based on large model-decision tree co-evolution as described in claim 6, characterized in that, Based on the case feature vector and decision path log, multi-dimensional clustering analysis and semantic induction are performed on the new claims cases to determine new decision rules. When the new decision rules are incrementally integrated into the decision tree, the process includes: The feature vectors of the new claims are integrated with the feature vectors of the erroneous cases in the historical error case database to generate a feature vector set. The historical error case database stores the feature vectors and decision path logs of claims that were previously determined to have substantial differences. Based on the feature vector set, dimensional feature data is extracted, including diagnostic code distribution, cost type features, material integrity index, OCR confidence, and decision path features. The feature vector set is grouped based on the density clustering algorithm, and cases with similar feature patterns are grouped into the same cluster, forming three types of error pattern clusters: bill recognition error cluster, disease liability determination fuzzy cluster, and cost calculation deviation cluster.

8. The intelligent claims processing method based on large model-decision tree co-evolution as described in claim 7, characterized in that, Based on the case feature vector and decision path log, when performing multi-dimensional clustering analysis and semantic induction on the new claims case to determine new decision rules, and incrementally integrating the new decision rules into the decision tree, the process further includes: For each error pattern cluster, common and differential features of cases within the cluster are extracted. These common and differential features are then input into the large language model. Based on the large language model, the root causes of errors are analyzed, and an error type description is generated. The error type description includes error triggering conditions, typical features, scope of impact, and severity. The error type description is then associated and mapped with uncovered nodes or error branches of the decision tree to construct an error pattern knowledge base.

9. The intelligent claims processing method based on large model-decision tree co-evolution as described in claim 8, characterized in that, Based on the case feature vector and decision path log, when performing multi-dimensional clustering analysis and semantic induction on the new claims case to determine new decision rules, and incrementally integrating the new decision rules into the decision tree, the process further includes: Based on the error pattern knowledge base and the large language model, new decision condition statements and new result determination parameters are generated; the new decision condition statements are converted into new condition expressions; and the new condition expressions and new result determination parameters are integrated into the decision tree.

10. An intelligent claims processing system based on large model-decision tree co-evolution, used to apply the intelligent claims processing method based on large model-decision tree co-evolution as described in any one of claims 1-9, characterized in that, include: The construction unit is configured to analyze historical claims cases based on a large language model, extract claims information and generate a set of decision rules; and construct a decision tree based on the set of decision rules. The output unit is configured to acquire new claims cases and extract case feature vectors; starting from the root node of the decision tree, the case feature vectors are compared with the condition thresholds of each decision node in the decision tree, and the matching path is traversed to the termination node to output the claims decision result; The assessment confidence level is determined based on the decision path length, node matching accuracy, and historical verification accuracy. The judgment unit is configured to compare the assessed confidence level with a preset confidence level threshold; if the assessed confidence level is greater than or equal to the confidence level threshold, the new claim case is approved; if the assessed confidence level is less than the confidence level threshold, the new claim case is manually reviewed, and the manual review result is determined. The optimization unit is configured to compare the claims decision results with the manual review results; when there are differences in the comparison results, based on the case feature vector and decision path log, perform multi-dimensional clustering analysis and semantic induction on the new claims case, determine new decision rules, and incrementally integrate the new decision rules into the decision tree.

Images