Causal federation element reinforcement learning intelligent decision-making method and system for people-benefiting insurance claim settlement

Abstract

The invention discloses a causal federation element reinforcement learning intelligent decision-making method and system for people-benefiting insurance claim settlement, and relates to the field of insurance intelligent claim settlement, and the method comprises the steps: building and adaptively generating a regional causal knowledge graph based on a multi-dimensional policy text; a causal federation element reinforcement learning framework is adopted, through element strategy network initialization, and federation learning is utilized to carry out fine adjustment on small samples in an encryption mode on the premise that data does not go out of a domain, so that local intention recognition model adaptation is realized; according to the claim consultation text of the user, performing intention recognition in combination with neural symbol double-track reasoning and knowledge graph causal chain matching; and training the decision tree, and generating a visual report through anti-factual reasoning to assist in decision making. According to the method, the user intention recognition accuracy and the rule interpretability are enhanced, the response adaptation efficiency of policy dynamic updating is improved, privacy is guaranteed on the premise that data are not out of a domain, small sample learning is achieved, then the people-benefiting insurance claim settlement processing efficiency and the customer satisfaction degree are optimized, and the operation cost is reduced.

Description

Technical Field

[0001] This application relates to the field of intelligent insurance claims, and in particular to a causal federated meta-reinforcement learning intelligent decision-making method and system for claims processing of inclusive insurance. Background Technology

[0002] As a universal commercial health insurance program, Huiminbao covers a wide population, has significant policy differences, and involves complex claims scenarios (including over 50 sub-categories of intent such as special drug reimbursement, pre-existing condition recognition, and cross-provincial medical treatment registration). Its claims processing efficiency and customer satisfaction directly impact the promotion and operation of the insurance product. Existing Huiminbao claims intent identification and triage technologies have the following core shortcomings: 1. Broken causal logic in rules, resulting in poor policy adaptability: Existing technologies mostly employ relational knowledge graphs (such as GraphRAG dynamic graphs) or fixed rule engines, which can only encode "entity-relationships" and cannot capture the "cause-effect" logical constraints in policy provisions (such as the necessary condition relationship between "special drug access" and "material requirements"). When policies iterate (such as adding special drug catalogs or adjusting rules for cross-regional medical treatment registration), the original relational rules are prone to becoming invalid, resulting in an intent recognition error rate as high as 18%, and the policy update response cycle requires more than 48 hours, which cannot meet the needs of dynamic adaptation.

[0003] 2. The contradiction between data privacy and sample sparsity is prominent in many regions: Huiminbao has significant regional policy differences. Due to privacy protection requirements, claims data in various regions cannot be centrally shared, resulting in insufficient new cases in some regions (especially newly added pilot areas). The traditional incremental learning model that relies on centralized big data is completely ineffective, and the intent recognition model has poor generalization ability and is difficult to adapt to the differentiated needs of regions.

[0004] 3. "Black box reasoning" in intent recognition, low decision transparency: Existing intent recognition technologies based on graph neural networks (GNN) and deep learning can only output intent labels and routing results, but cannot explain "why the intent was determined" or "why the routing path was chosen". Manual review lacks visual logic support, which can easily lead to customer disputes. The dispute rate of existing technologies is 15%-20%, and manual decision-making accounts for 40% of the total claims processing time.

[0005] 4. Homogeneous technical solutions and low barriers to innovation: Existing patents mostly apply technologies such as GraphRAG, causal graphs, and meta-reinforcement learning in isolation. For example, some patents use causal graphs for insurance fraud detection, but do not combine them with claims intent recognition; some patents use meta-reinforcement learning for robot control, but do not involve policy adaptation scenarios.

[0006] In summary, existing technologies cannot simultaneously resolve the four core contradictions of "dynamic policy adaptation, data privacy protection, interpretable reasoning, and efficient decision-making," thus failing to achieve the precision, efficiency, and transparency of identifying and allocating claims for the Huiminbao insurance program. Summary of the Invention

[0007] The purpose of this application is to provide a causal federated meta-reinforcement learning intelligent decision-making method and system for Huiminbao claims, which can simultaneously enhance the accuracy of user intent recognition and rule interpretability, improve the response and adaptation efficiency of dynamic policy updates, protect privacy without leaving the data domain, achieve small sample learning, thereby optimizing the efficiency of Huiminbao claims processing and customer satisfaction, and reducing operating costs.

[0008] To achieve the above objectives, this application provides the following solution.

[0009] Firstly, this application provides a causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance. The method includes: constructing a three-dimensional entity node set based on the policy text using a causal knowledge graph engine; detecting and pruning loops through association filtering, temporal correction, and counterfactual verification to construct a loop-free knowledge graph; and performing regional adaptation on the loop-free knowledge graph through entity mapping to obtain a regionally adapted causal knowledge graph. The policy text includes: disease type, claim type, material requirements, policy terms, and effective date. The three-dimensional causal entity set includes: entity nodes, time-sensitive nodes, and causal relationship nodes. A meta-policy network is constructed and trained; the trained meta-policy network is used to initialize local adaptation networks for each region; and the local adaptation networks are used to perform regional adaptation based on small sample data under the condition that the data does not leave the domain, through federated learning. The system fine-tunes encryption and adapts the local network to policy updates, resulting in a local intent recognition model updated with the new policy. It acquires and parses user claim consultation texts, employs neural symbolic dual-track reasoning to identify intent within these texts, and traverses the regionally adapted causal knowledge graph for causal chain matching and conclusion integration, outputting intent recognition results. A decision tree is trained based on relevant features from historical claim cases, and triage decision results are obtained from the trained causal decision tree. Based on the intent recognition results, triage decision results, and the regionally adapted causal knowledge graph, a causal path graph is constructed using CausalML counterfactual reasoning, generating a visual report. The relevant features from historical claim cases specifically include: intent label, claim amount, material completeness, and causal chain complexity. Claims decisions are made for the Huiminbao claim based on the visual report.

[0010] Secondly, this application provides a computer system, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the above-described causal federated meta-reinforcement learning intelligent decision-making method for Huiminbao claims settlement.

[0011] Based on the specific embodiments provided in this application, the following technical effects are disclosed.

[0012] This application constructs and adapts region-specific causal knowledge graphs based on multi-dimensional policy texts, providing an interpretable, structured knowledge foundation for subsequent reasoning. Secondly, it employs a federated meta-reinforcement learning framework, initializing regional models using a pre-trained meta-policy network and achieving rapid policy adaptation without data leaving the region through encrypted federated fine-tuning. Updates can be completed with only a few new cases, solving the challenges of sample sparsity and data privacy. In actual decision-making, the system uses neural symbolic dual-track reasoning for intent recognition of user claim consultation texts, significantly improving the accuracy of intent recognition. It also combines the knowledge graph with causal chain matching to improve the accuracy of rule interpretation. Finally, by combining the intent recognition results with the decision tree triage results trained based on historical case features, a clear causal path graph is constructed using CausalML counterfactual reasoning, generating a visual report, and making claims decisions based on this report for the Huiminbao (a type of social welfare insurance) claim.

[0013] This application improves the intelligence level of Huiminbao claims processing by deeply integrating causal reasoning, federated learning, and meta-reinforcement learning. It not only significantly improves the efficiency of manual decision-making and reduces the claims dispute rate, but also achieves key breakthroughs in multiple dimensions: in user intent recognition, accuracy and rule interpretability are simultaneously enhanced; in policy dynamic response, update and adaptation efficiency are significantly improved; and in data security and sample utilization, efficient small-sample learning is achieved while strictly protecting data privacy. These comprehensive advantages ultimately translate into overall optimization of claims processing efficiency and customer satisfaction, and effectively reduce operating costs, thus providing efficient, accurate, interpretable, and secure intelligent decision support for Huiminbao claims processing. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 A flowchart illustrating a causal federated meta-reinforcement learning intelligent decision-making method for Huiminbao claims provided in this application embodiment. Figure 1 .

[0016] Figure 2 A flowchart illustrating a causal federated meta-reinforcement learning intelligent decision-making method for Huiminbao claims provided in this application embodiment. Figure 2 .

[0017] Figure 3 A schematic diagram of the three-dimensional structure of the causal knowledge graph provided in the embodiments of this application.

[0018] Figure 4 This is a schematic diagram of the causal loop detection and pruning algorithm provided in an embodiment of this application.

[0019] Figure 5 This is a schematic diagram of the training of a federated meta-reinforcement learning network provided in an embodiment of this application.

[0020] Figure 6 The flowchart of the neural symbol dual-track inference provided for the embodiments of this application.

[0021] Figure 7 This is a schematic diagram of the visual decision report interface provided in an embodiment of this application.

[0022] Figure 8 This is an internal structure diagram of the computer system provided in this application embodiment. Detailed Implementation

[0023] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0024] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0025] Example 1, such as Figures 1-2 As shown in the figure, this embodiment provides a causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance, the method including the following steps.

[0026] S1. As Figures 3-4 As shown, a three-dimensional entity node set is constructed using a causal knowledge graph engine based on the policy text. Loop pruning is performed through association filtering, time-series correction, and counterfactual verification to construct a loop-free knowledge graph. The loop-free knowledge graph is then adapted to a specific region through entity mapping to obtain a region-adapted causal knowledge graph. The policy text includes: disease type, claim type, material requirements, policy terms, and effective date. The three-dimensional causal entity set includes: entity nodes, time-series nodes, and causal relationship nodes.

[0027] Optionally, the entity nodes should include at least: disease type, claim type, policy terms, required materials, and constraints.

[0028] In practical applications, examples of entity nodes are as follows: Entity "Cancer" (disease type), "Special Drug Reimbursement" (claim type), "Ninghuibao

[2025] No. 6 Clause" (policy clause), "Genetic Testing Report" (required material).

[0029] Optionally, the effective date should include at least the policy effective date and the policy expiration date; an example of an effective date is as follows: "Effective on September 16, 2025".

[0030] Optionally, causal relationship nodes may include at least: necessary conditions, sufficient conditions, and exclusionary conditions.

[0031] In practical applications, causal relationship nodes are assigned causal strength values ​​(necessary condition ≥ 0.9, sufficient condition ≥ 0.85, exclusion condition ≤ -0.8) by analyzing historical claims data and policy texts. An example of a causal relationship node is as follows: "Special drug reimbursement → Necessary condition (strength 0.92) → Genetic testing report".

[0032] Furthermore, step S1 specifically includes the following steps.

[0033] S11. Input the policy text into the causal knowledge graph engine to construct a three-dimensional causal entity set.

[0034] S12. Based on the partial correlation coefficient, a preliminary association screening of the three-dimensional causal entity set is performed to obtain the initial causal entity set.

[0035] In practical applications, correlation screening is based on calculating the original correlation between entities using the partial correlation coefficient, and initially screening entity pairs with a correlation of ≥0.7.

[0036] S13. The temporal correlation of the initial causal entity set is corrected by using the temporal effect to obtain the candidate causal chain; the temporal effect is the temporal matching between the policy implementation time and the occurrence time of the claim case.

[0037] In practical applications, the introduction of the time-series intervention effect is achieved by matching the time of policy implementation with the time of claims occurrence, and calculating the time-series correlation degree = original correlation degree × time-series matching coefficient (value 0-1).

[0038] S14. Simulate the removal of pre-existing condition nodes, track the changes in the accuracy of claim denial intent identification, perform counterfactual reasoning verification on candidate causal chains, and obtain the change in accuracy.

[0039] S15. Prune the candidate causal chain using the change in accuracy to obtain the pruned causal chain.

[0040] In practical applications, the causal relationship determination process during pruning is as follows: if the accuracy decreases by ≥10%, it is determined to be a true causal relationship, and the causal chain is retained; if the accuracy decreases by <10%, it is determined to be a false association, and pruning is performed.

[0041] S16. Calculate the confidence of the causal chain after pruning, and filter the causal chains with high confidence thresholds to construct a loop-free knowledge graph.

[0042] Optionally, the confidence level may include at least one of the following: overall causal strength, time series matching degree, and data support.

[0043] Optionally, the confidence threshold is 0.8.

[0044] Optionally, a high confidence threshold causal chain is a causal chain with a confidence level exceeding the confidence threshold.

[0045] Optionally, after calculating the confidence of the causal chain after pruning and filtering out causal chains with high confidence thresholds, low-confidence chains should also be removed.

[0046] S17. Based on a loop-free knowledge graph and regionally differentiated policy provisions, a basic causal model and a regional extension model are constructed in a hierarchical manner.

[0047] S18. Establish entity mapping relationships between the basic causal model and the regional extension model using entity mapping tables. By adapting to different regions, obtain a regionally adapted causal knowledge graph.

[0048] In practical applications, the basic causal model is used to integrate nationally applicable causal rules (such as "special drug reimbursement → special drug access determination"); the regional extension model is used to configure regionally differentiated causal chains through ontological mapping; for example, for the Nanjing region, new adaptive causal chains are added such as "special drug reimbursement → necessary conditions → prescription from a secondary or higher-level hospital in Nanjing" and "C Nanjing RT drug reimbursement → necessary conditions → diagnosis certificate from a designated medical institution in Nanjing". In addition, the adaptation time for adding new regions is ≤8 hours.

[0049] S2. For example Figure 5 As shown, a meta-policy network is constructed and trained. The trained meta-policy network is used to initialize the local adaptation network for each region. The local adaptation network is then fine-tuned through federated learning based on small sample data under the condition that the data does not leave the domain. The local adaptation network is then used to update and adapt the policy, resulting in a local intent recognition model after the policy update.

[0050] Furthermore, step S2 specifically includes the following steps.

[0051] S21. Construct and train the meta-policy network.

[0052] S22. Based on common policy data and local compensation data, a general intent recognition strategy is used to pre-train the meta-policy network to obtain a trained meta-policy network.

[0053] In practical application, common policy data includes: common data on the people-benefiting insurance policies of 31 provinces across the country (5,000+ policy clauses, 100,000+ historical claims cases); local claims data includes: 5-10 newly added policy cases.

[0054] S23. Initialize the local adaptation network using the general policy parameters output by the trained meta-policy network.

[0055] Optionally, the network structure of the local adaptation network includes: a 2-layer lightweight neural network (input layer dimension = 768, hidden layer dimension = 256, output layer dimension = local intent recognition model parameter dimension).

[0056] S24. Based on the newly added policy cases in each region, the parameters of the local adaptation network are fine-tuned using the Model-Agnostic Meta-Learning (MAML) algorithm.

[0057] In practical applications, the local intent recognition accuracy rate after fine-tuning is ≥95%.

[0058] S25. The gradient parameters of the local adaptation network are encrypted and uploaded using the Paillier homomorphic encryption algorithm.

[0059] S26. The federated learning coordination node receives the encrypted gradient parameters of each region and performs aggregation calculation. The coordination node then feeds back the aggregated gradient parameters to the local adaptation of each region to obtain the intent recognition model adapted to each region.

[0060] In practical applications, the adaptation time for the intent recognition model to different regions is ≤24 hours.

[0061] S27. Collect local compensation data and gradient parameters for each region.

[0062] S28. Store the local geographical claims data on a local node, and use the Paillier homomorphic encryption algorithm to encrypt the gradient parameters of each region. Only share the encrypted gradient, and perform differential privacy noise enhancement on the encrypted gradient to obtain the encrypted gradient parameters.

[0063] In practical applications, data storage is handled as follows: data from each region is stored on local nodes and is not transmitted across regions; gradient encryption uses homomorphic encryption algorithms to encrypt gradient parameters, and only the encrypted gradient is shared; differential privacy enhancement involves adding noise (noise intensity ≤ 0.01) to prevent gradients from being used to infer the original data. The encrypted gradient parameters comply with the requirements of the Personal Information Protection Law.

[0064] S29. Using the intent recognition accuracy and causal consistency score, the parameters of the local adaptation network are optimized through a causal-driven reward function to obtain the optimized local adaptation network parameters.

[0065] In practical applications, step S29 proceeds as follows.

[0066] 1) Input: Intent recognition accuracy (A), causal consistency score (C).

[0067] 2) Calculation of reward function: R=α×A+(1-α)×C, where α=0.6, C=number of causal chain matches / total number of intent recognition cases.

[0068] 3) Parameter adjustment trigger: If the identification result violates the causal rule (such as being judged as having the intention to reimburse special drugs even though the special drug access conditions are not met), the C value will decrease, the reward R will decrease, and the local adaptation network parameter adjustment will be triggered.

[0069] 4) Output: Optimized local adaptation network parameters to ensure that the recognition results conform to causal logic.

[0070] S210. Obtain policy update announcements and newly added policy documents (PDF format) uploaded manually from the official website of the Medical Insurance Bureau.

[0071] S211. A policy document parsing engine is used to extract policy features from policy update announcements on the official website of the Medical Insurance Bureau and newly uploaded policy documents to obtain policy update feature vectors.

[0072] S212. Use the policy update feature vector to trigger the fine-tuning process of the local adaptation network, update the local intent recognition model, and obtain the local intent recognition model after the policy update.

[0073] Furthermore, the meta-policy network is a 3-layer fully connected neural network; the input layer dimension = policy feature vector dimension + intent label vector dimension; the hidden layer dimension = 512; and the output layer dimension = general intent recognition policy parameter dimension.

[0074] S3. For example Figure 6 As shown, the system acquires and parses user claim consultation text, uses neural symbolic dual-track reasoning to identify the intent of the user claim consultation text, and traverses the regional adaptive causal knowledge graph to perform causal chain matching and conclusion integration, outputting the intent identification results.

[0075] Furthermore, step S3 specifically includes the following steps.

[0076] S31. Obtain and parse user claim inquiry text (e.g., "What materials are needed for cancer patients in Nanjing to claim special drugs?").

[0077] In practical applications, the semantic parsing process of user claims consultation text is as follows: the consultation text is semantically parsed using the ERNIE3.0 model to extract a 768-dimensional text feature vector.

[0078] S32. Input the user's claim inquiry text into the local adaptation network, predict the user's intent through neural track inference, and use the confidence threshold to filter out user intents with low confidence.

[0079] In practical applications, the confidence level is first determined: if the confidence level is ≥0.8, the result is output directly; if the confidence level is <0.8, the inference process proceeds to the symbolic track.

[0080] S33. Use a first-order predicate logic rule base to decompose low-confidence user intents into core intents and sub-intents.

[0081] For example, consultation on special drug reimbursement materials → [Special Drug Access Judgment Sub-intention] ∧ [Materials Requirements in Nanjing Area Sub-intention] ∧ [Policy Timeliness Verification Sub-intention].

[0082] S34. Traverse the region-adapted causal knowledge graph, match the causal chains corresponding to each sub-intention, and obtain the causal chain matching results.

[0083] For example: Special drug access determination intention → matching 'Special drug reimbursement' → necessary conditions → gene testing report.

[0084] S35. Integrate the conclusions based on the core intent and the causal chain matching results, and output the intent recognition results.

[0085] Optional intent identification results include: policy basis and necessary conditions.

[0086] For example, consultation on materials for special drug reimbursement requires a gene testing report and a prescription from a secondary or higher-level hospital in Nanjing; the policy basis is Article 6 of Ninghuibao

[2025] (effective September 16, 2025), and the scope of application includes the entire Nanjing city area.

[0087] Furthermore, the first-order predicate logic rule base is constructed by transforming business logic into definition expressions using predicates, quantifiers, and connectives.

[0088] In practical applications, the construction process of the first-order predicate logic rule base is as follows.

[0089] 1) Input: 50+ categories of subdivided intentions for benefiting the people and causal knowledge graph rules.

[0090] 2) Rule definition: Define the predicate logic rules.

[0091] 2) Output: First-order predicate logic rule base (containing 50+ rules).

[0092] Optionally, the logical rules of the first-order predicate logic rule base are as follows.

[0093] Rule 1: Intent to reimburse special drugs (X) → Special drug access determination (X) ∧ Material list matching (X) ∧ Policy validity verification (X).

[0094] Rule 2: Material list matching (X) → ∃Y (Required materials (Y) ∧ causal relationship (X, Y, necessary condition)).

[0095] Rule 3: Policy validity verification (X) → Policy clause (Z) ∧ Validity node (T) ∧ T≤ Current time.

[0096] S4. Train a decision tree based on the relevant features of historical claims cases, and obtain the diversion decision results from the trained causal decision tree. Based on the intent recognition results, diversion decision results and regionally adapted causal knowledge graph, construct a causal path graph through CausalML counterfactual reasoning to generate a visualization report. The relevant features of historical claims cases specifically include: intent label, claim amount, material completeness and causal chain complexity.

[0097] Furthermore, step S4 specifically includes the following steps.

[0098] S41. Obtain relevant features of historical claims cases.

[0099] S42. Construct a decision tree.

[0100] Furthermore, the process of constructing a decision tree is as follows.

[0101] S421. Take the causal chain complexity as the root node; the causal chain complexity = causal chain length × average causal strength.

[0102] S422. Start from the root node to perform the first division and generate the first branch node: If the causal chain complexity is greater than the complexity threshold, it is judged as a high-risk complex case and directly assigned to the manual review branch without further judgment.

[0103] S423. Proceed to the next level of the decision tree and generate the second branch node: If the complexity of the causal chain is less than the complexity threshold and the amount of a single claim is greater than the claim amount threshold, it will still be classified into the manual review branch without further judgment.

[0104] S424. Enter the next level of the decision tree and generate the third branch node: when the causal chain complexity is less than the complexity threshold and the single claim amount is greater than the claim amount threshold, it is classified into the automatic claim branch.

[0105] S425. Integrate the partitioning rules of the root node, the first branch node, the second branch node, and the third branch node to construct a decision tree.

[0106] S43. Train a decision tree using relevant features from historical claims cases.

[0107] S44. Input the intent recognition results, diversion decision results, and regional adaptation causal knowledge graph into CausalML counterfactual reasoning to generate counterfactual conclusions.

[0108] For example, if a gene testing report is provided (missing material), the causal chain "special drug reimbursement → gene testing report" is met, and the claim can be transferred to the automatic claims channel, reducing the processing time from 48 hours to 8 hours.

[0109] S45. Based on counterfactual conclusions, use D3.js to construct an interactive causal path graph.

[0110] In practical applications, key nodes of "cause → effect" are marked (such as "2025 Nanjing Ninghuibao Phase VI Special Drug Catalog → Special Drug Access → Gene Testing Report → Automatic Claims"), and manual drag-and-drop of nodes is supported to view detailed rules.

[0111] S46. For example Figure 7 As shown, a visual decision report is generated based on the interactive causal path diagram.

[0112] Furthermore, after step S4, the method also includes: adaptively adjusting the causal knowledge graph and the local adaptation network based on the feedback results from the manual review platform.

[0113] In practical applications, the process of adaptively adjusting the causal knowledge graph and the local adaptation network based on the feedback from the manual review platform is as follows.

[0114] 1) Input: "Decision bias" annotations provided by the manual review platform (e.g., "Causal chain 'pre-existing condition → claim denial' does not consider cases where pre-existing condition and claim disease are unrelated").

[0115] 2) Generation of counterexample samples: Transform feedback information into "counterexample samples" (including deviation scenario features and correct rule labels).

[0116] 3) Model and graph optimization: Reduce the causal strength weight of unreasonable causal chains (e.g., adjust the strength of "pre-existing condition → claim rejection" from 0.9 to 0.6); retrain the local adaptation network: fine-tune parameters based on counterexamples.

[0117] 4) Output: Optimized causal knowledge graph and local adaptation network, with an optimization period of ≤24 hours.

[0118] S5. Make claims decisions for the Huiminbao insurance based on the visualization report.

[0119] The technical effects of this application are as follows.

[0120] This application constructs and adapts region-specific causal knowledge graphs based on multi-dimensional policy texts, providing an interpretable, structured knowledge foundation for subsequent reasoning. Secondly, it employs a federated meta-reinforcement learning framework, initializing regional models using a pre-trained meta-policy network and achieving rapid policy adaptation without data leaving the region through encrypted federated fine-tuning. Updates can be completed with only a few new cases, solving the challenges of sample sparsity and data privacy. In actual decision-making, the system uses neural symbolic dual-track reasoning for intent recognition of user claim consultation texts, significantly improving the accuracy of intent recognition. It also combines the knowledge graph with causal chain matching to improve the accuracy of rule interpretation. Finally, by combining the intent recognition results with the decision tree triage results trained based on historical case features, a clear causal path graph is constructed using CausalML counterfactual reasoning, generating a visual report, and making claims decisions based on this report for the Huiminbao (a type of social welfare insurance) claim.

[0121] This application improves the intelligence level of Huiminbao claims processing by deeply integrating causal reasoning, federated learning, and meta-reinforcement learning. It not only significantly improves the efficiency of manual decision-making and reduces the claims dispute rate, but also achieves key breakthroughs in multiple dimensions: in user intent recognition, accuracy and rule interpretability are simultaneously enhanced; in policy dynamic response, update and adaptation efficiency are significantly improved; and in data security and sample utilization, efficient small-sample learning is achieved while strictly protecting data privacy. These comprehensive advantages ultimately translate into overall optimization of claims processing efficiency and customer satisfaction, and effectively reduce operating costs, thus providing efficient, accurate, interpretable, and secure intelligent decision support for Huiminbao claims processing.

[0122] Dual improvement in intent recognition accuracy and interpretability: Through dual-track reasoning of causal knowledge graph and neural symbols, the intent recognition accuracy reaches 97.2%, which is 0.7% higher than the traditional GraphRAG solution and 35% higher than the existing rule engine; the causal rule interpretation accuracy reaches 95%, completely solving the "black box reasoning" problem and providing clear logical support for manual review.

[0123] Policy adaptation efficiency and flexibility have been significantly enhanced: Federal meta-reinforcement learning can complete policy adaptation with only 5-10 new cases, reducing the policy update response time from 48 hours to 24 hours, and the adaptation time for multiple regions is ≤8 hours, fully meeting the dynamic iteration needs of the people's welfare insurance policy.

[0124] Simultaneous solutions to data privacy protection and sample sparsity issues: Horizontal federated learning enables data to be "locally stored and encrypted for sharing," avoiding the risk of cross-regional data leakage; Meta-reinforcement learning's few-sample learning capability solves the sample sparsity problem for newly added regions or niche policies, improving the model's generalization ability by 40%.

[0125] Claims processing efficiency and customer satisfaction have been greatly improved: cause-effect visualization decision-making has increased the efficiency of manual decision-making by 60% (the manual processing time has been reduced from an average of 2 hours to 48 minutes), reduced the claims dispute rate by 42% (from 18% to 10.4%), increased the automatic claims approval rate to 75%, and significantly reduced operating costs.

[0126] High technological innovation barriers and avoidance of existing patents: This invention is the first to deeply integrate "causal knowledge graph + federated meta-reinforcement learning + neural symbol dual-track reasoning + counterfactual visualization". No similar patents were found in the search, completely avoiding the scope of existing patent protection for simple causal graphs and meta-reinforcement learning, highlighting its inventiveness and novelty.

[0127] As an optional implementation method, this application takes Nanjing Ninghuibao (Nanjing Customized Huiminbao) as an example to illustrate the specific process of identifying and triaging special drug reimbursement intentions as follows.

[0128] System Deployment and Initialization: Input: Basic causal model (200+ general causal rules), differentiated policy terms in Nanjing (including the basic / upgraded version of "Nanjing Ninghuibao" Phase VI coverage rules, CAR-T drug reimbursement policy, etc.), and 5000+ historical claims cases in Nanjing.

[0129] Processing flow: 1) Deploy local data nodes and federated learning coordination nodes in various regions, and the system supports 1000+ concurrent claims inquiries.

[0130] 2) The causal knowledge graph engine loads the basic causal model and configures the Nanjing area extended model through the entity mapping table (adding differentiated causal chains such as "Special drug reimbursement → Necessary conditions → Prescription from secondary or above hospitals in Nanjing", "Special drug catalog reference '2025 Nanjing Ninghuibao Phase VI Special Drug Catalog'", and "CAR-T drug reimbursement → Necessary conditions → Diagnostic certificate from designated medical institutions in Nanjing").

[0131] 3) Initialization of the federated meta-reinforcement learning module: The meta-policy network loads pre-trained parameters, and the local adaptation network accesses historical claims cases in Nanjing to complete initial fine-tuning.

[0132] Output: The system deployed in Nanjing achieved an initial intent recognition accuracy of 96.8%.

[0133] 4) Policy update and adaptation process.

[0134] Input: The "Announcement on the Update of the 6th Phase of the 'Nanjing Ninghuibao' Insurance Plan in 2025" issued by the Nanjing Municipal Medical Insurance Bureau on September 16, 2025, and 8 claims cases for newly added CAR-T drugs.

[0135] Processing flow: a. The policy document parsing engine automatically captures and parses the announcement, extracts the entity information and corresponding causal rules of the four newly added CAR-T drugs, and clarifies that "CAR-T drug reimbursement → necessary conditions → gene testing report + prescription from a designated medical institution in Nanjing" and "CAR-T drug reimbursement → time limit → effective from September 16, 2025".

[0136] b. Causal knowledge graph engine update: Add new CAR-T drug entity nodes to the basic model, add corresponding causal chains to the extended model in Nanjing area, and prune the causal chains of special drugs that have been discontinued.

[0137] c. Adaptation of the Federated Meta-Reinforcement Learning Module: Eight new cases of local network access adaptation in Nanjing were added. The parameters were fine-tuned through the MAML algorithm, and the adaptation time was 12 hours.

[0138] Output: A causal knowledge graph adapted to the new policy and a local intent recognition model, with an accuracy rate of 97.5% in recognizing intent for special drug reimbursement.

[0139] 5) Example of user intent recognition and traffic triage.

[0140] Input: User inquiry text "I am a cancer patient in Nanjing and I want to claim reimbursement for the newly added XXCAR-T drug. What materials do I need to prepare?"

[0141] Processing flow: a. Neural track reasoning: The ERNIE3.0 model parses semantic features and outputs the initial intent label "Special drug reimbursement material consultation" with a confidence level of 0.78 (<0.8), triggering symbolic track reasoning.

[0142] b. Symbolic orbital reasoning: Intent decomposition: Decomposed into "Special Drug Access Determination ∧ Material List Matching ∧ Nanjing Area Policy Verification"; Causal chain matching: Traverse the causal knowledge graph adapted to multiple regions to accurately match the causal chain corresponding to the sub-intent, namely "Special drug access judgment sub-intent → match 'Special drug reimbursement → necessary condition → gene testing report'", "Nanjing area material requirements sub-intent → match 'Special drug reimbursement → necessary condition → prescription from secondary or above hospitals in Nanjing'", "Policy validity verification sub-intent → match 'Special drug reimbursement → validity constraint → effective from September 16, 2025'".

[0143] Conclusion integration: Generate interpretable intent conclusions.

[0144] c. Triage Decision: The causal decision tree determines that "the causal chain complexity is low (length 3) + the claim amount is not clear (default < 50,000 yuan) + the material list needs to be supplemented", and then the case is transferred to the "material supplementation reminder channel".

[0145] d. Visual report generation: Generate a cause-and-effect path diagram using D3.js, with a counterfactual suggestion of "If the above materials are supplemented, automatic claims are expected to be completed within 1 business day", and push it to the user and manual review platform.

[0146] Output: Interpretable intent recognition results, material replenishment and diversion instructions, and visual decision reports.

[0147] 6) Human feedback and rule optimization.

[0148] Input: The manual review platform reported that "some out-of-town medical treatment users who hold prescriptions from secondary or higher-level hospitals outside the province are also eligible for reimbursement." The reason for the decision-making bias was noted: "The causal chain of the Nanjing extended model, 'Special drug reimbursement → Necessary conditions → Prescriptions from secondary or higher-level hospitals in Nanjing,' does not cover out-of-town medical treatment scenarios and does not consider the requirement for new non-Nanjing residents to provide proof of work / life in Nanjing."

[0149] Processing flow: a. The system automatically converts feedback into negative examples (including deviation scenarios such as "medical treatment in other places", "prescriptions from secondary or above hospitals outside the province", and "new citizens without Nanjing household registration").

[0150] b. Causal knowledge graph optimization: For the Nanjing area, the extended model iteration is adjusted to "Special drug reimbursement → Necessary conditions → (Prescription from a secondary or above hospital in Nanjing) ∨ (Out-of-town medical treatment registration materials + prescription from a secondary or above hospital outside the province + proof of work / life in Nanjing)", and at the same time, the associated causal chain "Out-of-town medical treatment registration materials → Related policies → 'Implementation Rules for Direct Settlement Management of Cross-Provincial Out-of-Town Medical Treatment in Nanjing'" is added.

[0151] c. Local adaptation network readjustment: update parameters based on counterexample samples.

[0152] Output: The optimized causal knowledge graph and local adaptation network improved the intent recognition accuracy to 97.8%, and eliminated all biased cases.

[0153] As demonstrated by the above embodiments, this invention can accurately identify the intent of claims under the "Ninghuibao" (Ningminbao) program, quickly adapt to regional policy iterations, and significantly improve claims processing efficiency and decision-making transparency while strictly protecting data privacy. Its core advantage lies in its perfect adaptation to the Nanjing "Ninghuibao" city-level pooling, the upgrade of the sixth phase of the protection plan, the dynamic expansion of the special drug catalog, and support for non-local new residents' participation and claims. It effectively solves the core pain points of adapting to differentiated policies across multiple regions and modeling with limited sample data, possessing outstanding technological advancement, practical application value, and industrialization prospects.

[0154] As demonstrated by the above examples, this embodiment can accurately identify the intent of claims under the "Huiminbao" (a type of social welfare insurance) program, quickly adapt to regional policy iterations, and improve claims processing efficiency and decision-making transparency while strictly protecting data privacy. Its core advantage lies in its perfect adaptation to the Nanjing "Ninghuibao" city-level pooling, the upgrade of the sixth phase of the protection plan, the dynamic expansion of the special drug list, and support for non-local new residents' participation and claims. It solves the core pain points of adapting to differentiated policies in multiple regions and modeling with limited sample data, possessing outstanding technological advancement, practical application value, and industrialization prospects.

[0155] Example 2: This example provides a computer system, which can be a server or a terminal, and its internal structure diagram can be as follows. Figure 8 As shown, the computer system includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs stored in the non-volatile storage media. The database stores forced oscillation samples and sub / supersynchronous oscillation samples. The I / O interfaces are used for information exchange between the processor and external devices. The communication interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements the aforementioned method for rapid prediction and identification of the dominant frequency of sub / supersynchronous oscillations in new energy power systems based on transfer learning.

[0156] Those skilled in the art will understand that Figure 8 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer system to which the present application is applied. A specific computer system may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0157] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0158] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0159] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0160] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance, characterized in that, The method includes: Based on the policy text, a three-dimensional entity node set is constructed using a causal knowledge graph engine. Loop pruning is performed through association filtering, temporal correction, and counterfactual verification to construct a loop-free knowledge graph. This loop-free knowledge graph is then regionally adapted using entity mapping to obtain a regionally adapted causal knowledge graph. The policy text includes: disease type, claim type, required materials, policy terms, and effective date. The three-dimensional causal entity set includes: entity nodes, time-sensitive nodes, and causal relationship nodes. Construct and train a meta-policy network, initialize local adaptation networks for each region using the trained meta-policy network, and fine-tune the local adaptation networks using federated learning based on small sample data under the condition that the data does not leave the domain, and perform policy update adaptation on the local adaptation networks to obtain the local intent recognition model after policy update. The system acquires and parses user claim consultation texts, uses neural symbolic dual-track reasoning to identify intent in the texts, and traverses regional adaptive causal knowledge graphs to match causal chains and integrate conclusions, outputting intent identification results. A decision tree is trained based on the relevant features of historical claims cases, and the triage decision results are obtained from the trained causal decision tree. Based on the intent recognition results, triage decision results, and regionally adapted causal knowledge graph, a causal path graph is constructed through CausalML counterfactual reasoning to generate a visual report. The relevant features of historical claims cases specifically include: intent label, claim amount, material completeness, and causal chain complexity. Make claims decisions based on the visualization report.

2. The causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance as described in claim 1, characterized in that, After training a decision tree based on the relevant features of historical claims cases and obtaining the diversion decision results from the trained causal decision tree, and constructing a causal path graph through CausalML counterfactual reasoning based on the intent recognition results, diversion decision results and regionally adapted causal knowledge graph, and generating a visualization report, the process also includes: adaptively adjusting the causal knowledge graph and local adaptation network based on the feedback results from the manual review platform.

3. The causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance as described in claim 1, characterized in that, Based on the policy text, a three-dimensional entity node set is constructed using a causal knowledge graph engine. Loop pruning is performed through association filtering, temporal correction, and counterfactual verification to construct a loop-free knowledge graph. Finally, the loop-free knowledge graph is region-adapted through entity mapping to obtain a region-adapted causal knowledge graph, which specifically includes: Input the policy text into the causal knowledge graph engine to construct a three-dimensional causal entity set; A preliminary association screening of the three-dimensional causal entity set is performed based on the partial correlation coefficient to obtain the initial causal entity set; The temporal correlation of the initial causal entity set is corrected by utilizing the temporal effect to obtain candidate causal chains; the temporal effect refers to the temporal matching between the policy implementation time and the occurrence time of the claim case. Simulate the removal of pre-existing condition nodes, track the changes in the accuracy of claim denial intent identification, verify candidate causal chains through counterfactual reasoning, and obtain the change in accuracy. The candidate causal chain is pruned using the change in accuracy to obtain the pruned causal chain. Calculate the confidence of the causal chain after pruning, and filter the causal chains with high confidence thresholds to construct a loop-free knowledge graph. Based on a loop-free knowledge graph and differentiated policy provisions in different regions, a basic causal model and a regional extension model are constructed in a hierarchical manner. Entity mapping tables are used to establish entity mapping relationships between the basic causal model and the regional extension model. By adapting to different regions, a regionally adapted causal knowledge graph is obtained.

4. The causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance as described in claim 1, characterized in that, A meta-policy network is constructed and trained. The trained meta-policy network is then used to initialize local adaptation networks for each region. These local adaptation networks are then fine-tuned using federated learning based on small sample data, provided the data does not leave the region. Policy updates are then applied to these local adaptation networks to obtain the updated local intent recognition model. Specifically, this includes: Construct and train the meta-policy network; Based on common policy data and local claims data, a general intent recognition strategy is used to pre-train the meta-policy network to obtain a trained meta-policy network. The local adaptation network is initialized using the general policy parameters output by the trained meta-policy network. Based on the newly added policy cases in various regions, the parameters of the local adaptation network are fine-tuned using the model-independent meta-learning algorithm; The gradient parameters of the local adaptation network are encrypted and uploaded using the Paillier homomorphic encryption algorithm. The federated learning coordination node receives encrypted gradient parameters from each region and performs aggregation calculations. The coordination node then feeds back the aggregated gradient parameters to the local adaptation of each region, resulting in an intent recognition model adapted to each region. Collect local compensation data and gradient parameters for each region; The local claims data is stored on a local node, and the gradient parameters of each region are encrypted using the Paillier homomorphic encryption algorithm. Only the encrypted gradient is shared, and differential privacy noise enhancement is applied to the encrypted gradient to obtain the encrypted gradient parameters. By leveraging intent recognition accuracy and causal consistency scores, the parameters of the local adaptation network are optimized using a causal-driven reward function, resulting in optimized local adaptation network parameters. Obtain policy update announcements and newly uploaded policy documents from the official website of the Medical Insurance Bureau; A policy document parsing engine was used to extract policy features from policy update announcements on the official website of the Medical Insurance Bureau and newly uploaded policy documents, resulting in a policy update feature vector. The local adaptation network is fine-tuned by using the policy update feature vector to update the local intent recognition model, resulting in the local intent recognition model after the policy update.

5. The causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance as described in claim 4, characterized in that, The meta-policy network is a 3-layer fully connected neural network; the input layer dimension = policy feature vector dimension + intent label vector dimension; the hidden layer dimension = 512; the output layer dimension = general intent recognition policy parameter dimension.

6. The causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance as described in claim 1, characterized in that, The system acquires and parses user claim inquiry texts, employs neural symbolic dual-track reasoning to identify the intent within these texts, and traverses a region-adapted causal knowledge graph for causal chain matching and conclusion integration. The resulting intent identification output includes: Obtain and parse the user's claim inquiry text; The user's claim inquiry text is input into the local adaptation network, and the user's intent is predicted through neural track inference. The user intent with low confidence is then filtered out using a confidence threshold. The low-confidence user intent is decomposed into core intent and sub-intent using a first-order predicate logic rule base; Traverse the region-adaptive causal knowledge graph, match the causal chains corresponding to each sub-intention, and obtain the causal chain matching results; The conclusions are integrated based on the core intent and the causal chain matching results, and the intent recognition results are output.

7. The causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance as described in claim 6, characterized in that, The first-order predicate logic rule base is constructed by transforming business logic into definition expressions using predicates, quantifiers, and connectives.

8. The causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance as described in claim 1, characterized in that, A decision tree is trained based on the relevant features of historical claims cases, and the triage decision results are obtained from the trained causal decision tree. Based on the intent recognition results, triage decision results, and regionally adapted causal knowledge graph, a causal path graph is constructed using CausalML counterfactual reasoning, generating a visual report, specifically including: Obtain relevant characteristics of historical claims cases; Construct a decision tree; Train a decision tree using relevant features from historical claims cases; The intent recognition results, triage decision results, and region-adaptive causal knowledge graph are input into CausalML counterfactual reasoning to generate counterfactual conclusions. Based on counterfactual conclusions, an interactive causal path graph is constructed using D3.js; Generate a visual decision report based on the interactive causal path diagram.

9. The causal federated meta-reinforcement learning intelligent decision-making method for claims settlement of public welfare insurance as described in claim 8, characterized in that, The process of constructing a decision tree is as follows: The causal chain complexity is taken as the root node; the causal chain complexity = causal chain length × average causal strength; Starting from the root node, the first division is performed to generate the first branch node: if the causal chain complexity is greater than the complexity threshold, it is judged as a high-risk complex case and directly assigned to the manual review branch without further judgment; Enter the next level of the decision tree and generate a second branch node: if the complexity of the causal chain is less than the complexity threshold and the amount of a single claim is greater than the claim amount threshold, it will still be classified into the manual review branch without further judgment. Enter the next level of the decision tree and generate the third branch node: when the causal chain complexity is less than the complexity threshold and the single claim amount is greater than the claim amount threshold, it is classified into the automatic claim branch; By integrating the rules for dividing the root node, the first branch node, the second branch node, and the third branch node, a decision tree is constructed.

10. A computer system, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the causal federalized meta-reinforcement learning intelligent decision-making method for Huiminbao claims as described in any one of claims 1-9.

Images