Meta-cognition-based multi-agent medical insurance fund risk control system and method

Abstract

The application relates to the technical field of medical insurance fund supervision, and discloses a multi-agent medical insurance fund risk control system and method based on meta-cognition, wherein the system comprises a multi-source data acquisition and privacy calculation processing module, a dynamic heterogeneous information network construction module, a multi-modal feature coding and cross-modal fusion module, a space-time heterogeneous multi-modal graph neural network reasoning module, a cognitive enhancement large language model deep cognitive reasoning module, a multi-agent collaborative reasoning system collaborative execution module, a cognitive collaborative hierarchical early warning conduction and differentiated intervention module and a cross-institutional federated collaborative training module.

Description

Technical Field

[0001] This invention relates to the field of medical insurance fund supervision technology, and more specifically, to a multi-agent medical insurance fund risk control system and method based on metacognition. Background Technology

[0002] In the practice of medical insurance fund supervision, after obtaining settlement data from the medical insurance information platform, third-party risk control agencies usually rely on rule base matching and manual review to screen for anomalies.

[0003] Existing improvement solutions introduce knowledge graphs or single deep learning models to analyze medical insurance data. Some solutions attempt to model the medical treatment behavior of insured persons or the settlement records of medical institutions in order to improve the ability to identify anomalies.

[0004] However, the above methods have the following drawbacks: First, risk identification lags behind fund disbursement, resulting in a long recovery period; second, existing methods lack the overall modeling capability for the complex relationships between insured individuals, medical institutions, physicians, drugs, and treatment items, and cannot identify organized and interconnected fraudulent insurance activities; third, existing solutions fail to achieve integrated modeling of multi-source heterogeneous data, and also fail to solve the data privacy and collaborative computing issues between third-party regulatory agencies and medical insurance agencies, resulting in limited coverage of regulatory data and difficulty in forming a proactive full-process regulatory capability. Summary of the Invention

[0005] This invention provides a multi-agent medical insurance fund risk control system and method based on metacognition, which solves the technical problems in related technologies such as the strong concealment of medical insurance fund fraud, insufficient generalization ability of risk control models due to cross-institutional data silos, lack of interpretability of single model reasoning, and low efficiency of multi-party collaborative execution.

[0006] This invention discloses a multi-agent medical insurance fund risk control system based on metacognition, comprising: The multi-source data acquisition and privacy computing processing module is configured to acquire raw data from medical insurance information platforms, medical institution systems, pharmacy inventory management systems, drug traceability code platforms, and geographic information systems. Based on the federated learning framework, it performs distributed feature encoding on the data of each participant. The raw data of each participant does not leave the local machine, but only uploads the encrypted feature representation to generate an encrypted feature representation that can be used collaboratively. The dynamic heterogeneous information network construction module is configured to model seven types of entities, including insured persons, medical institutions, physicians, drugs, diagnosis and treatment items, medical insurance agencies, and medical insurance policies and rules, and their relationships as a heterogeneous graph with timestamps, generating a structured graph representation that supports dynamic evolution. The multimodal feature encoding and cross-modal fusion module is configured to extract semantic representations from seven modalities of data carried by each node and edge, namely structured, temporal, text, image, knowledge graph, geospatial and social network data, calculate the interaction weights of each modality through an attention mechanism and sum them up to generate a fused node representation vector. The spatiotemporal heterogeneous multimodal graph neural network inference module is configured to perform multi-level graph neural network inference on the fused node representation vectors, and output node-level risk scores, edge-level anomaly scores, and gang fraud subgraph risk scores. The cognitive enhancement big language model deep cognitive reasoning module is configured to receive the structured risk representation output by the graph neural network, perform deep reasoning in combination with the cognitive knowledge graph, and output a structured cognitive decision graph that includes risk level judgment, suspected violation type, fund loss estimate, disposal suggestions, evidence chain sorting, interpretability explanation and cognitive confidence score. The multi-agent collaborative reasoning system's collaborative execution module is configured to transform the structured cognitive decision graph into multi-agent collaborative execution instructions, which are then executed by six specialized intelligent agents: risk assessment, audit scheduling, settlement interception, policy interpretation, judicial collaboration, and federal collaboration. The execution logs are written to the blockchain. The cognitive collaborative hierarchical early warning transmission and differentiated intervention module is configured to calculate and predict risk deviations based on node-level risk scores and risk thresholds, generate four-level early warnings (blue, yellow, orange, and red), and implement three differentiated intervention measures: pre-event reminders, in-event interception, and post-event auditing. The cross-agency federated collaborative training module is configured to perform federated collaborative training based on the local data of each participant, following a closed-loop process of local training, secure uploading, global aggregation, and model distribution, to generate a global cognitive collaborative model. The cognitive regulatory effectiveness evaluation module is configured to periodically calculate model prediction bias, fund recovery rate, false alarm rate, false alarm rate, regulatory response timeliness, and cognitive decision accuracy. When any indicator exceeds a preset threshold, it triggers a model retraining or rule base update process.

[0007] Furthermore, the dynamic heterogeneous information network construction module models the medical insurance ecosystem as a heterogeneous graph. ,in It is a set of seven types of entity nodes. It is a set of six categories: medical treatment relationship, prescription relationship, practice relationship, drug circulation relationship, settlement relationship, and policy application relationship. It is a set of multimodal features of nodes and edges. A set of timestamps The total number of timestamps supports dynamic modeling of graph structures evolving over time.

[0008] Furthermore, the multimodal feature encoding and cross-modal fusion module calculates the interaction weights of each modality through the attention mechanism as follows: after concatenating the semantic representations of the seven modalities, a learnable weight matrix is ​​linearly transformed and a bias term is added, and then Softmax normalization is performed to obtain the attention weights of each modality. The fusion node representation vector is the weighted sum of the semantic representations of each modality and the corresponding attention weights. The semantic information of the seven modalities is used as the initial feature input of the nodes of the subsequent graph neural network layer.

[0009] Furthermore, the spatiotemporal heterogeneous multimodal graph neural network inference module is composed of four sub-modules in sequence: a relation-aware heterogeneous graph attention network sub-module, a temporal graph evolution sub-module, a hypergraph neural network sub-module, and a contrastive learning enhancement sub-module. The relation-aware heterogeneous graph attention network sub-module performs relation-aware heterogeneous graph message passing on the fused node representation vectors and outputs heterogeneous graph relation representations. The temporal graph evolution sub-module introduces a time decay function to perform time-aware graph evolution modeling on the heterogeneous graph relation representations and outputs temporal evolution features. The hypergraph neural network sub-module performs high-order relation modeling on the temporal evolution features and outputs high-order hypergraph features. The contrastive learning enhancement sub-module performs contrastive learning enhancement on the high-order hypergraph features and outputs risk representations.

[0010] Furthermore, the timing graph evolution submodule introduces a time decay function. ,in It is an exponential function. The difference between the current moment and the historical interaction moment. As the decay coefficient, it captures the gradual changes in node features over time by weighted aggregation of graph snapshots at each moment in the timestamp set.

[0011] Furthermore, the contrastive learning enhancement submodule uses the representations of the same node at different times as positive sample pairs and the randomly sampled non-fraud node representations as negative samples. It uses a contrastive loss function to bring the positive sample pairs closer together and push the negative sample pairs further apart, and outputs a risk representation.

[0012] Furthermore, the cognitive enhancement big language model deep cognitive reasoning module is based on an open-source big language model and is constructed through four stages: cognitive pre-training in the medical insurance field, cognitive instruction fine-tuning, cognitive knowledge graph enhancement, and expert cognitive feedback reinforcement learning. Among them, the cognitive knowledge graph enhancement stage superimposes a mask matrix generated based on the cognitive knowledge graph into the standard self-attention calculation. accomplish, Each element is determined by the relationship path score between two corresponding tokens in the knowledge graph. Token pairs that are directly or indirectly related in the knowledge graph are assigned a positive bias, otherwise it is zero. In the expert cognitive feedback reinforcement learning stage, medical insurance regulatory experts, clinicians, pharmacists, and legal experts give multi-dimensional scores to the model output, and the direct preference optimization algorithm is used to optimize the model parameters based on expert preferences.

[0013] Furthermore, the collaborative planning in the collaborative execution module of the multi-agent collaborative reasoning system generates a sequence of execution steps through the CoReAct framework, following a cyclical iterative process of collective thinking, division of labor, and shared observation. The toolset of the division of labor execution module includes settlement interception tools, audit dispatch tools, early warning notification tools, data query tools, report generation tools, knowledge retrieval tools, federated collaboration tools, and judicial collaboration tools. The collective feedback module feeds back the execution results of each agent to the cognitively enhanced large language model, which iteratively corrects the structured cognitive decision graph.

[0014] Furthermore, in the cross-agency federated collaborative training module, each participant trains a local model on local data, calculates the gradient, and then uploads it to the global aggregation server using a secure multi-party computation protocol. The global aggregation server aggregates the gradients of each party through a secure aggregation protocol and performs differential privacy protection and gradient anomaly detection. The generated global cognitive collaborative model is then distributed to each participant for the next round of training. The participant admission and gradient aggregation process is recorded on the blockchain.

[0015] This invention discloses a metacognitive-based multi-agent medical insurance fund risk control method, used to execute the aforementioned metacognitive-based multi-agent medical insurance fund risk control system, comprising: Step 1: Obtain raw data from the medical insurance information platform, medical institution system, pharmacy inventory management system, drug traceability code platform and geographic information system, and perform distributed feature encoding on the data of each participant based on the federated learning framework to generate encrypted feature representations that can be used collaboratively. Step 2: Model the seven types of entities and their six types of relationships as heterogeneous graphs with timestamps to generate a structured graph representation that supports dynamic evolution; Step 3: Extract semantic representations from the seven modal data carried by each node and edge, calculate the interaction weights of each modality through an attention mechanism and sum them up to generate a fused node representation vector; Step 4: Perform relation-aware heterogeneous graph message passing, time-aware graph evolution modeling, higher-order relation modeling and contrastive learning enhancement on the fused node representation vectors in sequence, and output node-level risk scores, edge-level anomaly scores and gang fraud subgraph risk scores. Step 5: Input the structured risk representation output by the graph neural network into the cognitive enhancement big language model, combine it with the cognitive knowledge graph to perform deep reasoning, and output a structured cognitive decision graph. Step 6: Transform the structured cognitive decision graph into multi-agent collaborative execution instructions. Six types of specialized intelligent agents are assigned to perform processing operations such as settlement interception, audit dispatch, and judicial collaboration. The execution log is written to the blockchain. Step 7: Calculate the predicted risk deviation based on node-level risk scores and risk thresholds, generate four-level early warnings, and implement three differentiated intervention measures: pre-event reminders, in-event interception, and post-event audits. Step 8: Perform federated collaborative training based on the local data of each participant to generate a global cognitive collaborative model and distribute it to each participant; Step 9: Periodically calculate evaluation metrics. When any metric exceeds a preset threshold, trigger the model retraining or rule base update process.

[0016] This invention utilizes a spatiotemporal heterogeneous multimodal graph neural network to spatiotemporally model the dynamic heterogeneous relationships of seven types of entities in the medical insurance ecosystem. This addresses the technical problems of traditional rule engines and single machine learning models failing to capture the evolutionary characteristics of multi-entity relationships and identifying organized fraudulent behavior, achieving the technical effect of explicitly expressing organized fraudulent behavior at the graph structure level. Furthermore, by combining structured risk representation with cognitive knowledge graphs through a cognitively enhanced large language model for deep reasoning, it solves the technical problems of general large models lacking cognitive capabilities in the medical insurance field and risk assessment conclusions lacking policy basis and evidence chain support, achieving the technical effect of improving decision interpretability and legal applicability. Finally, by using a multi-agent collaborative reasoning system to integrate cognitive decision-making... The policy is automatically transformed into specific execution instructions such as settlement interception, audit dispatch, and judicial collaboration, solving the technical problem of delayed response caused by excessively long manual decision-making and execution links, and achieving the technical effect of shortening the handling response time; through the federated learning framework, collaborative modeling of all participants is achieved without data leaving the local area, solving the technical problem of data silos between third-party regulatory agencies and medical insurance agencies, and achieving the technical effect of expanding the coverage of regulatory data; through the linkage of three differentiated intervention modes of pre-event reminder, in-event interception, and post-event audit, the technical problem of traditional regulatory methods being mainly based on post-event audit and unable to achieve full-process proactive supervision is solved, and the technical effect of reducing fund losses from the source is achieved. Attached Figure Description

[0017] Figure 1 This is a flowchart of the cognitive enhancement large language model construction method provided in the embodiments of the present invention. Detailed Implementation

[0018] Example 1 In the practice of medical insurance fund supervision, after obtaining settlement data from the medical insurance information platform, third-party risk control agencies typically rely on rule base matching and manual review for anomaly screening. This approach has two fundamental flaws: first, risk identification lags behind fund disbursement, resulting in collection cycles lasting several months; second, existing methods lack the ability to comprehensively model the complex relationships between insured individuals, medical institutions, physicians, drugs, and treatment items, making it impossible to identify organized and interconnected fraudulent activities. Some improvement solutions introduce knowledge graphs or single deep learning models, but fail to achieve fusion modeling of multi-source heterogeneous data, nor do they address the issues of data privacy and collaborative computing between third-party regulatory agencies and medical insurance management agencies.

[0019] like Figure 1 As shown, this embodiment proposes a multi-agent medical insurance fund risk control method based on metacognition, including the following steps: Step 1: Obtain raw data from the medical insurance information platform, medical institution system, pharmacy inventory management system, drug traceability code platform and geographic information system, and perform distributed feature encoding on the data of each participant based on the federated learning framework to generate encrypted feature representations that can be used collaboratively. Step 2: Model the seven types of entities and their six types of relationships as heterogeneous graphs with timestamps to generate a structured graph representation that supports dynamic evolution; Step 3: Extract semantic representations from the seven modal data carried by each node and edge, calculate the interaction weights of each modality through an attention mechanism and sum them up to generate a fused node representation vector; Step 4: Perform relation-aware heterogeneous graph message passing, time-aware graph evolution modeling, higher-order relation modeling and contrastive learning enhancement on the fused node representation vectors in sequence, and output node-level risk scores, edge-level anomaly scores and gang fraud subgraph risk scores. Step 5: Input the structured risk representation output by the graph neural network into the cognitive enhancement big language model, combine it with the cognitive knowledge graph to perform deep reasoning, and output a structured cognitive decision graph. Step 6: Transform the structured cognitive decision graph into multi-agent collaborative execution instructions. Six types of specialized intelligent agents are assigned to perform processing operations such as settlement interception, audit dispatch, and judicial collaboration. The execution log is written to the blockchain. Step 7: Calculate the predicted risk deviation based on node-level risk scores and risk thresholds, generate four-level early warnings, and implement three differentiated intervention measures: pre-event reminders, in-event interception, and post-event audits. Step 8: Perform federated collaborative training based on the local data of each participant to generate a global cognitive collaborative model and distribute it to each participant; Step 9: Periodically calculate evaluation metrics. When any metric exceeds a preset threshold, trigger the model retraining or rule base update process.

[0020] The corresponding risk control method in this embodiment also proposes the following system, including: The multi-source data acquisition and privacy-preserving computation processing module is configured to acquire raw data from multiple data sources, perform privacy-preserving processing, and generate encrypted feature representations that can be used collaboratively. Raw data is obtained in real time from medical insurance information platforms, medical institution HIS systems, pharmacy inventory management systems, drug traceability code platforms, and geographic information systems. The data types cover insured persons' medical behavior sequences, medical records, drug inventory management flow, physician practice information, drug traceability code circulation information, medical insurance policy and rule database, historical audit case database, geospatial data, clinical pathway standard database, medical imaging data, and medical record text data.

[0021] The raw data is subjected to de-identification, outlier cleaning, missing value imputation, and multimodal alignment to unify timestamps and entity identifiers from different sources. Furthermore, based on a federated learning framework, distributed feature encoding is performed on the data of each participant (third-party risk control company, medical insurance agency, designated medical institution, and designated retail pharmacy). Each party's original data remains local; only encrypted feature representations are uploaded, thus enabling collaborative utilization of multi-institutional data while protecting data privacy.

[0022] The dynamic heterogeneous information network construction module is configured to model various entities and relationships in the medical insurance ecosystem as heterogeneous graphs, generating structured graph representations that support dynamic evolution. Modeling various entities in the medical insurance ecosystem as heterogeneous graphs ,in: It is a set of nodes, corresponding to seven types of entities: insured persons, medical institutions, physicians, drugs, medical treatment items, medical insurance agencies, and medical insurance policies and rules. The set of edges corresponds to six types of relationships: medical treatment relationship, prescription relationship, practice relationship, drug circulation relationship, settlement relationship, and policy application relationship. It is a multimodal feature set of nodes and edges, including structured features, temporal features, text features, image features, knowledge graph features, geospatial features, and social network features; A set of timestamps The total number of timestamps supports dynamic modeling of graph structures evolving over time.

[0023] Through the above modeling, the relationship between insured individuals and medical institutions, the relationship between physicians and drugs, and the traceability relationship of drugs in the circulation process are all explicitly expressed in the form of heterogeneous edges with attributes, providing structured input for subsequent relational reasoning in graph neural networks.

[0024] The multimodal feature encoding and cross-modal fusion module is configured to process heterogeneous graphs. Semantic representations are extracted from the seven modalities carried by each node and edge, and cross-modal fusion is performed to generate a fused node representation vector. For heterogeneous graphs The seven types of modal data carried by each node and edge are used to extract semantic representations using modality-specific encoders: Structured feature encoding: Multilayer perceptron (MLP) is used to embed numerical and categorical features such as age, cost, and frequency of visits. The input is the structured attribute vector of each node, and the output is a structured representation. During training, the node type classification cross-entropy loss is used as the supervision signal, and the Adam optimization algorithm is used to update the parameters. Temporal feature encoding: A bidirectional LSTM is used to encode the time series of medical visits, with the input being a set of timestamps. The sequence of patient visits is arranged to capture the periodicity, suddenness, and abnormal patterns of patient behavior, and output a time-series representation. ; Text feature encoding: A pre-trained language model (BERT or ERNIE-Health) is used to semantically encode medical records and diagnostic descriptions. The input is a sequence of tokens representing the medical record text and the diagnostic description. The hidden state of the location is projected onto the text representation after passing through a linear mapping layer. ; Image Feature Encoding: ResNet is used to extract visual features from prescription scans and invoice images. The input is a normalized image tensor, and the feature vector output from the global average pooling layer is used as the image representation. ; Knowledge Graph Feature Encoding: The RotatE model is used to embed knowledge graphs such as medical insurance database rules, drug indications, and treatment guidelines. The input is the index of entities and relations, and the output is a knowledge graph representation. During training, a translation distance loss based on negative sampling is used; Geospatial Feature Coding: Geohash coding combined with a spatial attention mechanism is used to encode the insured person's address coordinates and the location of the medical institution, outputting a geospatial representation. ; Social network feature encoding: Node2Vec is used to learn the embeddings of the physician-patient-drug recommendation relationship network. The input is a sequence of random walks of nodes on the graph, and the output is a social network representation. ; Furthermore, an attention mechanism is employed to calculate the interaction weights of each modality, thereby achieving cross-modal fusion: ; in, For modal indexing, traverse the seven modalities; For modality Attention weights; The weight matrix is ​​a learnable matrix; For bias terms; For modality The corresponding semantic representation; This is the fused node representation vector. The semantic information of seven modalities is integrated and used as the initial feature input for the nodes of the subsequent graph neural network layers.

[0025] The spatiotemporal heterogeneous multimodal graph neural network inference module is configured to process the fusion node representation vectors. Perform multi-level graph neural network inference and output node-level, edge-level, and subgraph-level risk representations. The Spatiotemporal Heterogeneous Multimodal Graph Neural Network (ST-MM-GNN) consists of four sub-modules: Relation-Aware Heterogeneous Graph Attention Network (R-HGAT), Temporal Graph Evolution Module (TGNN), Hypergraph Neural Network Module (HGNN), and Contrastive Learning Reinforcement Module (GCL). Each sub-module processes and fuses the node representation vectors. Features are passed down level by level, and finally the risk representations at the node level, edge level, and subgraph level are output. ST-MM-GNN uses node binary classification (fraud / non-fraud) as the training objective, the output layer is a fully connected layer, the activation function is sigmoid, the training loss is weighted binary cross-entropy loss, and the Adam optimization algorithm is used to update the parameters.

[0026] The R-HGAT submodule is configured to represent the fusion node vector. Perform relation-aware heterogeneous graph message passing and output a heterogeneous graph relation representation. : For edge set For each type of relation edge, a relation-specific message passing function is defined. A multi-head attention mechanism is used to calculate the influence weights of neighbor nodes on the target node under different relation types. Neighbor information from multiple relation types is aggregated, and a node representation capturing the semantics of heterogeneous relations is output. R-HGAT enables the distinction and processing of the consultation relationship between insured individuals and medical institutions, and the prescription relationship between physicians and pharmaceutical companies, in message transmission, avoiding confusion of features of different semantic relationships.

[0027] The TGNN submodule is configured to represent heterogeneous graph relationships. Perform time-aware graph evolution modeling and output temporal evolution features. : Introducing a time decay function ,in Represents an exponential function. The difference between the current moment and the historical interaction moment. As a decay coefficient, recent interactions are given higher weight. This is achieved by adjusting the set of timestamps. By weighted aggregation of graph snapshots at various time points, TGNN identifies the gradual changes in node features over time, thereby capturing the fraud evolution pattern of "card nurturing period → probing period → outbreak period" and outputting temporal evolution features. .

[0028] The HGNN submodule is configured to process temporal evolution features. Perform high-order relation modeling and output high-order features of the hypergraph. : To address the high-order collaborative patterns that transcend binary relationships among "physicians, patients, medications, and treatment items," a hyperedge is introduced to explicitly model complex associations involving multiple entities. By weighting the contributions of each node within the hyperedge using hyperedge attention weights, the project identifies ternary and multi-level collusion fraud patterns such as "fake patients + fake prescriptions + fake examinations," and outputs high-order features of the hypergraph. .

[0029] The GCL submodule is configured to handle higher-order features of the hypergraph. Perform contrastive learning reinforcement to output the final risk representation. : Constructing positive and negative sample pairs: Using the representations of the same node at different times as positive sample pairs, and randomly sampled non-fraudulent node representations as negative samples, the model enhances its ability to distinguish hidden fraud patterns by comparing the loss function to bring positive sample pairs closer together and widen the distance between negative sample pairs, thus outputting a risk representation. .

[0030] Furthermore, ST-MM-GNN ultimately represents the risk through a fully connected output layer. Mapped to node-level risk scoring Edge-level anomaly scoring and gang fraud subgraph risk scoring, along with fused node representation vectors Temporal evolution characteristics Hypergraph higher-order features It is then fed into the cognitive enhancement large language model layer.

[0031] The cognitive enhancement large language model's deep cognitive reasoning module is configured to receive structured risk representations and combine them with cognitive knowledge graphs to perform deep reasoning, generating a structured cognitive decision graph. The Cognitive Enhanced Large Language Model (CE-LLM) is built upon open-source large language models (DeepSeek-R1 or Qwen2.5) through the following four stages: Step 1, Medical Insurance Domain Cognitive Pre-training: Incremental pre-training is performed on medical insurance policy texts, clinical diagnosis and treatment guidelines, drug instructions, medical insurance audit cases, medical literature and judicial precedents to enable the base model to acquire background knowledge of the medical insurance vertical domain; Step 2, Cognitive Instruction Fine-tuning: Construct an instruction dataset containing six task types: "risk assessment", "rule interpretation", "case reasoning", "decision generation", "evidence chain construction" and "law application", and perform supervised fine-tuning of the model; Step 3, Cognitive Knowledge Graph Enhancement: Knowledge graphs such as medical insurance database rules, clinical pathways, drug indications, DRG grouping schemes, and judicial precedents are injected into the model in the form of structured cognitive prompts. Attention calculation is then guided by the cognitive knowledge graph. in, , , These are the query matrix, key matrix, and value matrix, respectively. Indicates transpose; This is a mask matrix generated based on a cognitive knowledge graph, used to enhance semantic associations related to medical insurance rules during attention computation; The dimension of the key vector. Each element From the token in the knowledge graph with token The relationship path score is determined. If there is a direct or indirect relationship between the two in the knowledge graph, a positive bias is assigned; otherwise, it is zero. Thus, the prior association information of the knowledge graph is superimposed on the standard self-attention calculation. Step 4, Expert Cognitive Feedback Reinforcement Learning: Medical insurance regulatory experts, clinicians, pharmacists and legal experts give multi-dimensional scores to the model output, and the DPO algorithm is used to optimize the model parameters based on expert preferences.

[0032] Cognitive Enhanced Large Language Model (CE-LLM) receives fused node representation vectors Temporal evolution characteristics Hypergraph higher-order features and node-level risk scoring As input for structured cognitive prompts, deep reasoning is performed by combining the context of the cognitive knowledge graph, historical cases, and judicial precedents to output a structured cognitive decision graph (CDG). The structured cognitive decision graph (CDG) includes the following seven dimensions: risk level determination. Suspected violation types, estimated fund losses, disposal recommendations, evidence chain analysis, explanatory explanation, and cognitive confidence score.

[0033] Furthermore, when ST-MM-GNN detects a new fraud pattern, the Cognitive Enhanced Large Language Model (CE-LLM) automatically analyzes the abnormal features of the pattern, combines the cognitive knowledge graph to generate regulatory rule text, and converts the regulatory rule text into executable code for the rule engine to achieve dynamic expansion of regulatory rules; at the same time, it continuously monitors policy changes and automatically parses and updates the cognitive knowledge graph and rule base.

[0034] The collaborative execution module of the multi-agent collaborative reasoning system is configured to transform structured cognitive decision graphs into multi-agent collaborative execution instructions and complete the processing, generating execution logs and processing reports. The Multi-Agent Collaborative Reasoning System (MACRS) consists of a distributed perception module, a collaborative planning module, a division of labor execution module, and a collective feedback module. The modules communicate with each other through a distributed message bus and adopt a distributed architecture of "coordinating agent overall planning + specialized agent division of labor + collaborative agent arbitration".

[0035] The distributed perception module is configured to receive the structured cognitive decision graph (CDG) and perceive the state of the external environment, generating the perception input for each agent: Six types of specialized intelligent agents (risk assessment agent, audit scheduling agent, settlement interception agent, policy interpretation agent, judicial collaboration agent, and federal collaboration agent) each receive content corresponding to their responsibilities from the structured cognitive decision graph (CDG), and at the same time obtain environmental information such as the status of the settlement system, the status of the audit task queue, and the response status of medical institutions in real time, forming the perception input of each intelligent agent.

[0036] The collaborative planning module is configured to generate a sequence of collaborative execution steps based on cognitive decision goals and perceptual inputs in a structured cognitive decision graph (CDG) using the CoReAct framework. Multi-agent teams jointly plan execution paths through a negotiation mechanism. The CoReAct framework iterates in a loop of “Collective Thought → Division Action → Shared Observation” until a sequence of execution steps that meets the decision objectives is generated, thus determining the division of labor boundaries and execution order of each agent.

[0037] The task allocation module is configured to allow each specialized intelligent agent to call the corresponding tools in the toolset to complete the assigned tasks based on the sequence of execution steps. The toolset includes settlement interception tools, audit and dispatch tools, early warning and notification tools, data query tools, report generation tools, knowledge retrieval tools, federal collaboration tools, and judicial collaboration tools.

[0038] Taking the scenario of "fraudulent hospitalization gang testing" as an example, after the risk assessment intelligent agent confirms the red alert, the settlement interception intelligent agent calls the settlement interception tool to freeze the settlement interface, the audit scheduling intelligent agent calls the audit dispatch tool to dispatch special audit tasks, the data query intelligent agent calls the data query tool to query historical records, the judicial collaboration intelligent agent calls the judicial collaboration tool to transfer the case to the relevant authorities, and the report generation intelligent agent calls the report generation tool to generate a complete handling report. Each step is executed sequentially or in parallel according to the collaborative planning sequence.

[0039] The collective feedback module is configured to aggregate the execution results of each agent and feed them back to the Cognitive Enhanced Large Language Model (CE-LLM), and write the execution log to the blockchain. The execution results include whether the interception was successful, the status of the audit task reception, and the confirmation of judicial transfer. The cognitive enhancement big language model (CE-LLM) iteratively corrects the structured cognitive decision graph (CDG) based on the feedback results to optimize subsequent cognitive decision-making strategies. The execution log is stored on the blockchain to ensure the immutability and traceability of the handling records.

[0040] The cognitive collaborative hierarchical early warning transmission and differentiated intervention module is configured to calculate the predicted risk deviation based on node-level risk scoring and generate a risk level determination, and then implement differentiated intervention measures. Calculate the predicted risk deviation ,in For the node-level risk score predicted by ST-MM-GNN, For nodes The current risk threshold. Cognitive Enhanced Large Language Model (CE-LLM) is based on the prediction risk bias. It uses cognitive knowledge graph context, historical cases, and judicial precedents to conduct in-depth reasoning and generate risk level determinations. The corresponding blue / yellow / orange / red warning levels are assigned to the relevant medical insurance agencies or regulatory personnel.

[0041] Differentiated intervention strategies are implemented in the following three modes: Advance warning: The risk assessment intelligent agent dynamically compares the insured person's historical medical behavior with the current appointment information, and combines the real-time bed availability and physician scheduling status of medical institutions. It generates warning text through cognitive enhanced large language model (CE-LLM) and pushes it to the HIS system of the designated medical institution before the diagnosis and treatment behavior occurs to prevent potential violations such as duplicate prescriptions, over-prescription, and admission under low indications. In-process interception: The settlement interception intelligent agent will perform fusion scoring calculation by combining real-time settlement data stream with rule engine scoring, graph neural network risk scoring, and cognitive augmented big language model (CE-LLM) deep reasoning scoring. When the fusion score exceeds the interception threshold, settlement interception will be automatically triggered, and the document will be marked as "pending manual review". Post-event audit: The audit scheduling intelligent agent automatically summarizes intercepted documents, early warning records and risk assessment reports, generates audit task packages, and dispatches them to local medical insurance agencies or third-party audit teams to track audit progress and results, forming a closed-loop management system.

[0042] The cross-agency federated collaborative training module is configured to perform federated collaborative training based on the local data of each participant to generate a global cognitive collaborative model.

[0043] Based on local data from various participants (third-party risk control companies, medical insurance agencies, designated medical institutions, and designated retail pharmacies), federated collaborative training is performed according to a closed-loop process of "local training → secure upload → global aggregation → model distribution": Each participant trains a local model on local data, calculates the gradient, and then uploads it to the global aggregation server in encrypted form using the secure multi-party computation (MPC) protocol. The global aggregation server aggregates gradients from all parties through a secure aggregation protocol to generate a global cognitive collaborative model, while simultaneously performing differential privacy protection and gradient anomaly detection. The global cognitive collaboration model is distributed to all participating parties, and each party updates its local model before entering the next round of training.

[0044] Security measures such as participant access control, gradient aggregation, privacy protection, model heterogeneity compatibility, and blockchain notarization are implemented throughout the entire federated training process. The federated collaborative intelligent agent is responsible for coordinating the communication scheduling and anomaly handling of all participants.

[0045] The cognitive regulatory effectiveness assessment module is configured to periodically calculate assessment indicators and trigger model update processes to ensure the continuous effectiveness of the risk control system. The following evaluation metrics are calculated periodically: model prediction bias, fund recovery rate, false positive rate, false negative rate, regulatory response timeliness, specific scenario identification rate, and cognitive decision-making accuracy. When any metric exceeds a preset threshold, a model retraining or rule base update process is triggered to ensure the continuous effectiveness of the risk control system.

[0046] This implementation uses ST-MM-GNN to spatiotemporally model the dynamic heterogeneous relationships of seven types of entities in the medical insurance ecosystem: "insured persons, medical institutions, physicians, drugs, treatment items, medical insurance agencies, and medical insurance policies and rules." This overcomes the limitations of traditional rule engines and single machine learning models in capturing the evolutionary characteristics of multi-entity associations, enabling organized fraudulent behavior to be explicitly expressed at the graph structure level. By combining the structured risk representation output by the graph neural network with the cognitive knowledge graph through the cognitive enhanced big language model (CE-LLM), deep reasoning is performed, overcoming the deficiency of general big models in lacking cognitive ability in the medical insurance field. This enables risk assessment conclusions to have policy basis and evidence chain support, and improves the interpretability and legal applicability of decisions. By using the Multi-Agent Collaborative Reasoning System (MACRS), cognitive decisions are automatically transformed into specific execution instructions such as settlement interception, audit dispatch, and judicial collaboration, overcoming the problem of delayed response caused by excessively long manual decision-making and execution links, and compressing the handling response from the day level to the minute level. By using a federated learning framework, collaborative modeling among all participants is achieved without leaving their local data sites, overcoming the data silo problem between third-party regulatory agencies and medical insurance management agencies, and expanding the coverage of regulatory data. By linking three differentiated intervention models—pre-event reminders, in-event interception, and post-event audits—a shift from "post-event audits" to "proactive full-process supervision" has been achieved, reducing fund losses at the source.

[0047] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A multi-agent medical insurance fund risk control system based on metacognition, characterized in that, include: The multi-source data acquisition and privacy computing processing module is configured to acquire raw data from medical insurance information platforms, medical institution systems, pharmacy inventory management systems, drug traceability code platforms, and geographic information systems. Based on the federated learning framework, it performs distributed feature encoding on the data of each participant. The raw data of each participant does not leave the local machine, but only uploads the encrypted feature representation to generate an encrypted feature representation that can be used collaboratively. The dynamic heterogeneous information network construction module is configured to model seven types of entities, including insured persons, medical institutions, physicians, drugs, diagnosis and treatment items, medical insurance agencies, and medical insurance policies and rules, and their relationships as a heterogeneous graph with timestamps, generating a structured graph representation that supports dynamic evolution. The multimodal feature encoding and cross-modal fusion module is configured to extract semantic representations from seven modalities of data carried by each node and edge, namely structured, temporal, text, image, knowledge graph, geospatial and social network data, calculate the interaction weights of each modality through an attention mechanism and sum them up to generate a fused node representation vector. The spatiotemporal heterogeneous multimodal graph neural network inference module is configured to perform multi-level graph neural network inference on the fused node representation vectors, and output node-level risk scores, edge-level anomaly scores, and gang fraud subgraph risk scores. The cognitive enhancement big language model deep cognitive reasoning module is configured to receive the structured risk representation output by the graph neural network, perform deep reasoning in combination with the cognitive knowledge graph, and output a structured cognitive decision graph that includes risk level judgment, suspected violation type, fund loss estimate, disposal suggestions, evidence chain sorting, interpretability explanation and cognitive confidence score. The multi-agent collaborative reasoning system's collaborative execution module is configured to transform the structured cognitive decision graph into multi-agent collaborative execution instructions, which are then executed by six specialized intelligent agents: risk assessment, audit scheduling, settlement interception, policy interpretation, judicial collaboration, and federal collaboration. The execution logs are written to the blockchain. The cognitive collaborative hierarchical early warning transmission and differentiated intervention module is configured to calculate and predict risk deviations based on node-level risk scores and risk thresholds, generate four-level early warnings (blue, yellow, orange, and red), and implement three differentiated intervention measures: pre-event reminders, in-event interception, and post-event auditing. The cross-agency federated collaborative training module is configured to perform federated collaborative training based on the local data of each participant, following a closed-loop process of local training, secure uploading, global aggregation, and model distribution, to generate a global cognitive collaborative model. The cognitive regulatory effectiveness evaluation module is configured to periodically calculate model prediction bias, fund recovery rate, false alarm rate, false alarm rate, regulatory response timeliness, and cognitive decision accuracy. When any indicator exceeds a preset threshold, it triggers a model retraining or rule base update process.

2. The multi-agent medical insurance fund risk control system based on metacognition according to claim 1, characterized in that, The dynamic heterogeneous information network construction module models the medical insurance ecosystem as a heterogeneous graph. ,in It is a set of seven types of entity nodes. It is a set of six categories: medical treatment relationship, prescription relationship, practice relationship, drug circulation relationship, settlement relationship, and policy application relationship. It is a set of multimodal features of nodes and edges. A set of timestamps The total number of timestamps supports dynamic modeling of graph structures evolving over time.

3. The multi-agent medical insurance fund risk control system based on metacognition according to claim 1, characterized in that, The multimodal feature encoding and cross-modal fusion module calculates the interaction weights of each modality through an attention mechanism as follows: after concatenating the semantic representations of the seven modalities, a learnable weight matrix is ​​linearly transformed and a bias term is added. Then, the attention weights of each modality are obtained by Softmax normalization. The fusion node representation vector is the weighted sum of the semantic representations of each modality and the corresponding attention weights. The semantic information of the seven modalities is used as the initial feature input of the nodes of the subsequent graph neural network layers.

4. The multi-agent medical insurance fund risk control system based on metacognition according to claim 1, characterized in that, The spatiotemporal heterogeneous multimodal graph neural network inference module consists of four sub-modules in sequence: a relation-aware heterogeneous graph attention network sub-module, a temporal graph evolution sub-module, a hypergraph neural network sub-module, and a contrastive learning enhancement sub-module. The relation-aware heterogeneous graph attention network sub-module performs relation-aware heterogeneous graph message passing on the fused node representation vectors and outputs heterogeneous graph relation representations. The temporal graph evolution sub-module introduces a time decay function to perform time-aware graph evolution modeling on the heterogeneous graph relation representations and outputs temporal evolution features. The hypergraph neural network submodule performs high-order relation modeling on temporal evolution features and outputs high-order hypergraph features; The contrastive learning enhancement submodule performs contrastive learning enhancement on the high-order features of the hypergraph and outputs a risk representation.

5. The multi-agent medical insurance fund risk control system based on metacognition according to claim 4, characterized in that, The timing graph evolution submodule introduces a time decay function. ,in It is an exponential function. The difference between the current moment and the historical interaction moment. As the decay coefficient, it captures the gradual changes in node features over time by weighted aggregation of graph snapshots at each moment in the timestamp set.

6. The multi-agent medical insurance fund risk control system based on metacognition according to claim 4, characterized in that, The contrastive learning enhancement submodule uses the representations of the same node at different times as positive sample pairs and the representations of randomly sampled non-fraudulent nodes as negative samples. It uses a contrastive loss function to bring the positive sample pairs closer and push the negative sample pairs further apart, and outputs a risk representation.

7. The multi-agent medical insurance fund risk control system based on metacognition according to claim 1, characterized in that, The cognitive enhancement big language model deep cognitive reasoning module is based on an open-source big language model and is constructed through four stages: cognitive pre-training in the medical insurance field, cognitive instruction fine-tuning, cognitive knowledge graph enhancement, and expert cognitive feedback reinforcement learning. The cognitive knowledge graph enhancement stage involves superimposing a mask matrix generated based on the cognitive knowledge graph into the standard self-attention calculation. accomplish, Each element is determined by the relationship path score between two corresponding tokens in the knowledge graph. Token pairs that are directly or indirectly related in the knowledge graph are assigned a positive bias, otherwise it is zero. In the expert cognitive feedback reinforcement learning stage, medical insurance regulatory experts, clinicians, pharmacists, and legal experts give multi-dimensional scores to the model output, and the direct preference optimization algorithm is used to optimize the model parameters based on expert preferences.

8. The multi-agent medical insurance fund risk control system based on metacognition according to claim 1, characterized in that, The collaborative planning in the collaborative execution module of the multi-agent collaborative reasoning system generates a sequence of execution steps through the CoReAct framework, following a cyclical iterative process of collective thinking, division of labor, and shared observation. The toolset of the division of labor execution module includes settlement interception tools, audit dispatch tools, early warning notification tools, data query tools, report generation tools, knowledge retrieval tools, federated collaboration tools, and judicial collaboration tools. The collective feedback module feeds back the execution results of each agent to the cognitively enhanced big language model, which then iteratively corrects the structured cognitive decision graph.

9. The multi-agent medical insurance fund risk control system based on metacognition according to claim 1, characterized in that, In the cross-agency federated collaborative training module, each participant trains a local model on local data, calculates gradients, and then uploads them to the global aggregation server using a secure multi-party computation protocol. The global aggregation server aggregates the gradients from each party through a secure aggregation protocol and performs differential privacy protection and gradient anomaly detection. The generated global cognitive collaborative model is then distributed to each participant for the next round of training. The participant admission and gradient aggregation process is recorded on the blockchain.

10. A multi-agent medical insurance fund risk control method based on metacognition, used to execute the multi-agent medical insurance fund risk control system based on metacognition as described in any one of claims 1 to 9, characterized in that, include: Step 1: Obtain raw data from the medical insurance information platform, medical institution system, pharmacy inventory management system, drug traceability code platform and geographic information system, and perform distributed feature encoding on the data of each participant based on the federated learning framework to generate encrypted feature representations that can be used collaboratively. Step 2: Model the seven types of entities and their six types of relationships as heterogeneous graphs with timestamps to generate a structured graph representation that supports dynamic evolution; Step 3: Extract semantic representations from the seven modal data carried by each node and edge, calculate the interaction weights of each modality through an attention mechanism and sum them up to generate a fused node representation vector; Step 4: Perform relation-aware heterogeneous graph message passing, time-aware graph evolution modeling, higher-order relation modeling and contrastive learning enhancement on the fused node representation vectors in sequence, and output node-level risk scores, edge-level anomaly scores and gang fraud subgraph risk scores. Step 5: Input the structured risk representation output by the graph neural network into the cognitive enhancement big language model, combine it with the cognitive knowledge graph to perform deep reasoning, and output a structured cognitive decision graph. Step 6: Transform the structured cognitive decision graph into multi-agent collaborative execution instructions. Six types of specialized intelligent agents are assigned to perform processing operations such as settlement interception, audit dispatch, and judicial collaboration. The execution log is written to the blockchain. Step 7: Calculate the predicted risk deviation based on node-level risk scores and risk thresholds, generate four-level early warnings, and implement three differentiated intervention measures: pre-event reminders, in-event interception, and post-event audits. Step 8: Perform federated collaborative training based on the local data of each participant to generate a global cognitive collaborative model and distribute it to each participant; Step 9: Periodically calculate evaluation metrics. When any metric exceeds a preset threshold, trigger the model retraining or rule base update process.

Images