Knowledge graph and multi-agent game based diagnosis task analysis method and system

Abstract

The application discloses a kind of based on knowledge graph and multi-agent game's diagnostic task analysis method and system, method includes: obtaining patient multimodal medical data, constructs dynamic patient state graph and extracts embedding vector;From character library, match the exclusive medical intelligent agent consultation team of agent, carry out multi-round reasoning by confidence weighted Markov game, output optimal diagnostic hypothesis and evidence chain after meeting convergence condition;Abnormal risk events are encountered in reasoning, trigger human expert arbitration, continue reasoning after modifying reasoning path, simultaneously, arbitration operation is encoded as sparse reward, with the help of reinforcement learning iterative optimization agent parameter.The unstructured medical data is converted into structured diagnostic directed acyclic graph in the application, the defect that traditional search enhancement generates technical knowledge fragmentation is solved;Through game reasoning, the illusion of large model and group delusion are inhibited, and the diagnostic accuracy is improved;Through dynamic formation consultation team, the problem of fixed frame rigidity is broken.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and computer-aided diagnosis technology, and in particular to a diagnostic task parsing method and system based on knowledge graphs and multi-agent game theory. Background Technology

[0002] With the rapid development of large language model technology, its application potential in the medical field is becoming increasingly prominent, especially in key scenarios such as medical question answering, medical record summary generation, and assisted diagnosis. However, when applying general large language models directly to rigorous clinical diagnostic tasks, existing technical solutions still have many technical defects that need to be addressed, which seriously restrict the reliability, accuracy, and clinical applicability of medical AI.

[0003] In existing technologies, direct question-and-answer or thought chain reasoning schemes based on single large models are prone to medical factual errors and logical illusions due to the lack of effective knowledge constraint mechanisms. They also tend to blindly follow misleading prompts from users and lack self-correction capabilities, leading to diagnostic biases and error accumulation. While retrieval-enhanced auxiliary diagnostic technologies alleviate the illusion problem by introducing external medical document retrieval, they are limited by the mismatch between retrieval granularity and tasks, making it difficult to construct a complete diagnostic logic chain. Furthermore, the static nature of the knowledge base prevents it from dynamically adapting to the diagnostic process. Although some multi-agent collaborative frameworks based on fixed roles improve diagnostic complexity through role division, they suffer from rigid roles and resource waste, lack rigorous decision-making mechanisms, are prone to groupthink, and are unable to critically examine marginal situations.

[0004] Furthermore, existing medical AI generally adopts an end-to-end "black box" design, where human doctors can only obtain the final diagnosis results but cannot intervene in the reasoning process, resulting in low system trust. Moreover, when the model gets stuck in a reasoning deadlock or makes a directional error, there is a lack of effective mechanisms to allow experts to intervene in real time and feed back corrective signals to the model, which restricts the clinical practical value and continuous evolution capability of medical AI. Summary of the Invention

[0005] In view of the shortcomings of the prior art, the present invention provides a diagnostic task parsing method and system based on knowledge graph and multi-agent game theory, in order to solve the problems of insufficient diagnostic accuracy, poor flexibility of multi-agent collaboration framework and lack of effective human-machine collaboration mechanism in existing medical AI.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] In a first aspect, the present invention provides a diagnostic task parsing method based on knowledge graphs and multi-agent game theory, comprising:

[0008] Acquire patients' multimodal medical data;

[0009] The multimodal medical data is parsed to construct a dynamic patient state graph, and the embedding vector set of the dynamic patient state graph is extracted.

[0010] The dynamic patient state graph includes medical evidence nodes, diagnostic hypothesis nodes, and edges representing the logical relationships between the nodes.

[0011] Based on the aforementioned set of embedded vectors, using meta-agents, multiple agents with specific medical roles are identified from a predefined role library to establish an intelligent medical diagnosis team.

[0012] Using a confidence-weighted Markov game reasoning algorithm, each agent in the intelligent diagnosis team can perform multiple rounds of game reasoning in a shared environment based on the embedded vector set until the game convergence condition is met, and then output the optimal diagnostic hypothesis and the corresponding evidence chain as the final diagnostic conclusion.

[0013] If an abnormal risk event occurs during the game reasoning process, the current game reasoning is paused, and arbitration by human experts is requested and accepted. After the current game reasoning path is corrected according to the arbitration, the game reasoning is returned and continues.

[0014] In an optional implementation, the method further includes:

[0015] The arbitration operations of the human experts are encoded as sparse reward values, and the parameters of the corresponding agents are optimized using the sparse reward values ​​based on a reinforcement learning algorithm.

[0016] In an optional implementation, the step of parsing the multimodal medical data to construct a dynamic patient state map and extracting the embedding vector set of the dynamic patient state map specifically includes the following steps:

[0017] Based on a pre-trained medical entity recognition model, the multimodal medical data is parsed to identify and extract medical entities;

[0018] The medical entities include symptom entities, examination entities, drug entities, and disease entities;

[0019] Based on the attribute enhancement function, the attribute vector of the corresponding medical entity is generated;

[0020] Based on a pre-trained medical knowledge graph, the symptom entities and examination result entities are used as medical evidence nodes, and disease nodes associated with the evidence nodes are retrieved as diagnostic hypothesis nodes.

[0021] Based on the relevance filtering mechanism, the weights of the edges between entity nodes are calculated to form a directed acyclic graph that reflects the current diagnostic task status, which serves as a dynamic patient status graph.

[0022] Based on a graph attention network, the dynamic patient state graph is encoded to generate a set of embedding vectors containing graph topological features and semantic information.

[0023] In an optional implementation, the step of determining multiple agents with specific medical roles from a predefined role library based on the embedded vector set, using meta-agents, to establish an intelligent medical diagnosis team, specifically includes the following steps:

[0024] By using a meta-agent to perform attention-based graph pooling on the set of embedded vectors of the state graph, a global task vector is generated.

[0025] Based on the global task vector, the meta-agent predicts the medical specialty field required for the current diagnostic task.

[0026] Based on the capability-demand matching algorithm, the matching degree score between the embedding vector of the capability in the predefined role library and the embedding vector of the diagnostic hypothesis node is calculated;

[0027] To maximize coverage of the diagnostic hypothesis nodes while minimizing team size and role redundancy, a multi-objective optimization function is constructed.

[0028] Solve the multi-objective optimization function to determine the optimal set of intelligent agent roles from the role library, so as to form the optimal intelligent agent diagnosis team;

[0029] Generate corresponding personalized prompts for each intelligent agent in the consultation team;

[0030] The prompts include basic instructions, role definitions, contextual perspectives based on graph attention weight allocation, and relevant external tool specifications.

[0031] In an optional implementation, the confidence-weighted Markov game reasoning algorithm enables each agent in the intelligent diagnostic team to perform multiple rounds of game reasoning in a shared environment based on the embedded vector set until the game convergence condition is met, at which point the optimal diagnostic hypothesis and the corresponding evidence chain are output as the final diagnostic conclusion. Specifically, this includes the following steps:

[0032] A1. Based on the speaking intention function, calculate the speaking intention score of each agent, and determine the agent that obtains the right to speak based on the speaking intention score;

[0033] A2. The agent that gains the right to speak selects an action from the predefined action space to execute and outputs the corresponding argument triples;

[0034] The argument triad includes the target assumption, the type of action, and supporting evidence;

[0035] A3. The unspeaking agent receives the argument triples, evaluates the credibility of the argument triples using a two-layer Bayesian update algorithm, and then updates its own confidence distribution for each diagnostic hypothesis.

[0036] The two-layer Bayesian update algorithm includes an argument quality evaluation function and a hypothesis posterior probability update function;

[0037] A4. When there is a diagnostic hypothesis with a confidence level exceeding a preset threshold, the AI ​​agent in the diagnostic team, which is given the role of a critic, generates attack arguments based on counterfactual reasoning to minimize the overall confidence level of the diagnostic hypothesis.

[0038] A5. Repeat steps A1 to A4 until the game convergence condition is met. Then stop the game and output the diagnostic hypothesis with the highest confidence and the corresponding chain of evidence as the final diagnostic conclusion.

[0039] In an optional implementation, if an abnormal risk event occurs during the game reasoning process, the current game reasoning is paused, arbitration by a human expert is requested and accepted, the current game reasoning path is corrected according to the arbitration operation, and then the game reasoning is returned to and continues. Specifically, this includes the following steps:

[0040] The game reasoning process is monitored in real time, and a risk assessment score for the current game reasoning state is calculated based on cognitive uncertainty and decision risk.

[0041] When the risk assessment score exceeds a preset threshold, it is determined that the abnormal risk event has occurred, the game reasoning is suspended, and human experts are requested to intervene.

[0042] The system receives arbitration input from human experts through an interactive interface, corrects the current game reasoning path based on the arbitration input, and then returns to continue the game reasoning.

[0043] The cognitive uncertainty is quantified by the information entropy of the confidence distribution of all expert agents on the optimal diagnostic hypothesis;

[0044] The decision risk is quantified by multiplying the average confidence level of the current diagnostic hypothesis by a predefined risk weight.

[0045] In an optional implementation, encoding the human expert's arbitration operation as a sparse reward value and optimizing the parameters of the corresponding agent using the sparse reward value based on a reinforcement learning algorithm specifically includes the following steps:

[0046] The arbitration operations of the human experts are encoded as sparse reward values;

[0047] By using a credit allocation mechanism based on attention weights, sparse reward values ​​are allocated to relevant actions in the game reasoning history that lead to the arbitration result.

[0048] Using reinforcement learning algorithms, the policy network parameters of relevant agents are updated based on the allocated sparse reward values.

[0049] The step of encoding the arbitration operation of the human expert into a sparse reward value specifically includes:

[0050] If the agent's proposal is consistent with the final decision of the human expert, it is determined to be a positive reward value;

[0051] If the agent's proposal is rejected by the forced pruning operation, it is determined to be a negative reward value;

[0052] If the evidence cited by the agent is adopted by the human expert, it is determined as a procedural reward value.

[0053] In one optional implementation, the arbitration operation by the human expert includes at least one of the following operations:

[0054] Forced pruning is used to mark a specific diagnostic hypothesis as incorrect and reset the confidence of that incorrect diagnostic hypothesis to zero across all agents.

[0055] The evidence injection operation is used to add new evidence nodes to the dynamic patient state graph and update the graph structure.

[0056] The weight reset operation is used to adjust the power weight of a specific agent in the game process.

[0057] Secondly, the present invention provides a diagnostic task analysis system based on knowledge graphs and multi-agent game theory, comprising:

[0058] The data acquisition module is used to acquire patients' multimodal medical data;

[0059] The graph construction module is used to parse the multimodal medical data to construct a dynamic patient state graph and extract the embedding vector set of the dynamic patient state graph.

[0060] The dynamic patient state graph includes medical evidence nodes, diagnostic hypothesis nodes, and edges representing the logical relationships between the nodes.

[0061] The agent generation module is used to determine agents for multiple specific medical roles from a predefined role library based on the set of embedded vectors and using meta-agents, so as to establish an intelligent medical diagnosis team.

[0062] The game reasoning module is used to utilize a confidence-weighted Markov game reasoning algorithm to enable each agent in the intelligent diagnosis team to perform multiple rounds of game reasoning in a shared environment based on the embedded vector set until the game convergence condition is met, and then output the optimal diagnostic hypothesis vector as the final diagnostic conclusion.

[0063] The arbitration operation module is used to pause the current game reasoning if an abnormal risk event occurs during the game reasoning process, request and accept arbitration operation from human experts, correct the current game reasoning path according to the arbitration operation, and then return to the game reasoning module.

[0064] In an optional implementation, the system further includes:

[0065] The agent optimization module encodes the arbitration operation of the human expert into sparse reward values ​​and optimizes the parameters of the corresponding agent based on the sparse reward values ​​using a reinforcement learning algorithm.

[0066] The beneficial effects of the embodiments provided by the present invention include:

[0067] This invention constructs a dynamic patient state graph and performs graph encoding, transforming unstructured medical data into a directed acyclic graph for structured diagnostic tasks. This provides a deeply correlated knowledge foundation for reasoning and overcomes the knowledge fragmentation defect of traditional retrieval enhancement generation technology.

[0068] This invention establishes a confidence-weighted Markov game reasoning mechanism through adversarial debate, a two-layer Bayesian update algorithm, and critical review. This forces diagnostic hypotheses to undergo rigorous logical testing in dynamic games, effectively suppressing the illusion problem in large models and the group myth phenomenon of multi-agent systems, thereby significantly improving the accuracy of diagnostic conclusions. At the same time, this invention can provide an explicit debate process and evidence chain graph as an interpretable report, overcoming the trust barrier of traditional black-box models.

[0069] This invention solves the rigidity problem of fixed role frameworks by adaptively matching and generating optimal expert roles from a role library to dynamically assemble an intelligent diagnostic team. This invention also constructs a human expert intervention mechanism based on risk monitoring, which allows arbitration operations at key decision points to correct reasoning paths. Furthermore, the arbitration operations are encoded as sparse rewards and fed back to the agent's policy through reinforcement learning, enabling the AI ​​to continuously absorb human experience. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other embodiments can be obtained based on these drawings.

[0071] Figure 1 A flowchart illustrating the diagnostic task parsing method based on knowledge graphs and multi-agent game theory in the embodiments of this specification is shown.

[0072] Figure 2 A schematic diagram of the structure of the diagnostic task parsing system based on knowledge graphs and multi-agent game theory in the embodiments of this specification is shown. Detailed Implementation

[0073] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. In the following detailed description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that the invention may be practiced without requiring some of these specific details. The following description of embodiments is merely intended to provide a better understanding of the invention by illustrating examples of the invention.

[0074] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appear in a specific order. However, it should be clearly understood that these operations may not be performed in the order they appear herein, or they may be performed in parallel. The operation numbers, such as S1, S2, etc., are merely used to distinguish different operations and do not themselves represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be performed sequentially or in parallel.

[0075] Example 1

[0076] like Figure 1 As shown, this embodiment provides a diagnostic task parsing method based on knowledge graphs and multi-agent game theory, including:

[0077] S1. Acquire patients' multimodal medical data;

[0078] Multimodal medical data includes, but is not limited to, unstructured patient complaints, electronic medical record history, laboratory test reports, and medical image descriptions.

[0079] S2. Analyze the multimodal medical data to construct a dynamic patient state map and extract the embedding vector set of the dynamic patient state map;

[0080] In some embodiments, the dynamic patient state graph includes medical evidence nodes, diagnostic hypothesis nodes, and edges representing the logical relationships between the nodes;

[0081] Specifically, medical evidence nodes are objective clinical facts identified and extracted from multimodal medical data, such as symptoms and examination results; diagnostic hypothesis nodes are potential diseases or health conditions inferred based on medical evidence nodes; edges connect medical evidence nodes and diagnostic hypothesis nodes respectively, used to represent diagnostic reasoning dependencies between nodes;

[0082] The edges are constructed based on the dependency relationships under the diagnostic reasoning logic, rather than generalized medical knowledge associations.

[0083] For example, step S2 specifically includes the following steps:

[0084] S201. Based on a pre-trained medical entity recognition model, the multimodal medical data is parsed to identify and extract medical entities;

[0085] The medical entity recognition model includes a pre-trained encoder fine-tuned from a medical corpus, a linear classification layer, and a conditional random field-based decoder; the encoder can be either BioBERT or ClinicalBERT.

[0086] Specifically, the encoder extracts features from the input multimodal medical data to obtain a text sequence. Subsequently, the text sequence X is input into the linear classification layer for further sequence labeling to obtain the entity type prediction label corresponding to each character or word and generate the label sequence Y. Finally, the label sequence Y is decoded by the decoder to determine the entity boundaries, thereby identifying and extracting a set of medical entities containing entities such as symptoms, examination results, drugs, and diseases.

[0087] Among them, medical entity collection The definition of is:

[0088] ;

[0089] In the formula, Symptoms; For the inspection results; For medicine; For disease.

[0090] S202. Based on the attribute enhancement function, generate the attribute vector of the corresponding medical entity;

[0091] Specifically, in order to accurately handle negative expressions and degree modifications in the text descriptions of medical data, this embodiment introduces an attribute enhancement function to generate a structured attribute vector for each extracted entity based on the entity's context.

[0092] The expression for the attribute enhancement function is as follows:

[0093] ;

[0094] In the formula, Let i be the attribute vector of medical entity i; For property enhancement functions; For the i-th medical entity; A text sequence; For medical entity i, the polarity attribute ∈{+1,−1,0}, used to distinguish whether the symptom is present (+1), confirmed to be absent (-1), or suspected (0), in order to prevent the large language model from generating hallucinations; The duration step of the symptom corresponding to the medical entity is used to construct the temporal dependency; Quantified values ​​for medical entities;

[0095] S203. Based on the pre-trained medical knowledge graph, the symptom entities and examination result entities are used as medical evidence nodes, and disease nodes associated with the evidence nodes are retrieved as diagnostic hypothesis nodes.

[0096] Specifically, in this embodiment, entity mentions in the text sequence are first mapped to corresponding canonical concepts in the standard medical ontology using entity linking technology. Then, a pre-trained large-scale medical knowledge graph is used as the background knowledge base, and confirmed entities such as symptoms and examination results are used as medical evidence nodes with fixed states. For each medical evidence node, related disease candidate nodes in its k-hop neighborhood are retrieved in the background knowledge base, and these disease concepts are used as diagnostic hypothesis nodes to be verified.

[0097] Among them, the pre-trained large-scale medical knowledge graph is trained from massive medical literature, clinical guidelines and structured databases, and encodes a variety of medical relationships between concepts such as diseases, symptoms, examinations and drugs;

[0098] Entity linking is achieved by calculating the similarity score between entity mentions and candidate concepts in the standard medical ontology library; entity mentions include specific expressions of symptoms and examination names;

[0099] Specifically, the formula for calculating link similarity is:

[0100] ;

[0101] In the formula, represents link similarity; m represents entity mentions; c represents candidate concepts; The embedding vector for the entity reference; Let P(c|m) be the embedding vector of the candidate concept; P(c|m) is the prior probability. These are adjustable weight hyperparameters;

[0102] S204. Based on the relevance filtering mechanism, calculate the weight of the edges between entity nodes to form a directed acyclic graph that reflects the current diagnostic task status, which serves as a dynamic patient status graph.

[0103] Specifically, this embodiment quantifies the necessity of the inference relationship by defining the weights of the edges between nodes, and uses a relevance filtering mechanism to prune the retrieved initial associations, retaining only strongly related paths that conform to causal logic, so as to construct a directed acyclic graph reflecting the current diagnostic reasoning task based on entity nodes, hypothesis nodes, and edges between nodes.

[0104] The functional expression for a directed acyclic graph is:

[0105] ;

[0106] In the formula, It is a directed acyclic graph; These are entity nodes, including medical evidence nodes and diagnostic hypothesis nodes; The edges between entity nodes;

[0107] The weight of the edge between entity nodes is calculated using the following formula:

[0108] ;

[0109] In the formula, For entity nodes With entity nodes The weight of the edges; For entity nodes Inter-point mutual information; I( ) is a causal indicator function, which indicates that there exists a causal relationship between the causal relationship and the ... arrive The value is 1 when there is a medical causal or inferential relationship, and 0 otherwise.

[0110] The directed acyclic graph structure of the task generated in this embodiment can clearly express the dependencies of diagnostic reasoning, for example:

[0111] For example, a directed acyclic graph is:

[0112] ;

[0113] That is, the directed acyclic graph can clearly express that, in order to verify the diagnostic hypothesis node of "myocardial infarction", it is necessary to pay attention to the medical evidence node of "chest pain" and the node to be verified of "troponin test".

[0114] S205. Based on a graph attention network, the dynamic patient state graph is encoded to generate a set of embedding vectors containing graph topological features and semantic information.

[0115] Specifically, in order to enable the subsequent intelligent agent to understand the complex structure of the graph, this embodiment uses a graph attention network to encode the dynamic patient state graph and generate a context-aware set of embedding vectors.

[0116] For any node in the dynamic patient state graph, the feature vector of that node is updated by aggregating the features of its neighboring nodes.

[0117] The embedding vector of any node in the dynamic patient state graph has the following functional expression:

[0118] ;

[0119] In the formula, Let i be the feature vector of node i in the (l+1)th layer; It is a non-linear activation function; Let i be the set of neighbors of node i; The weight matrix is ​​a learnable matrix; Let be the attention coefficient of node j to node i;

[0120] The function expression for the attention coefficient is:

[0121] ;

[0122] In the formula, This is a vector concatenation operation. This is a learnable attention vector; For activation functions; Let i be the feature vector of node i;

[0123] It should be noted that this embodiment encodes the dynamic patient state graph through a graph attention network, transforming unstructured multimodal medical data into a set of high-dimensional embedding vectors. This set of embedding vectors not only integrates the semantic information of each node, but also encodes the diagnostic logic and topological structure carried by the graph, which can be directly used as input for the subsequent generation of intelligent diagnostic teams. Among them, the attention coefficients in the graph attention network dynamically represent the relative importance of adjacent nodes in medical differential diagnosis, realizing the quantitative modeling of diagnostic dependencies.

[0124] This embodiment uses the above process to parse and map the original medical data into a structured dynamic patient state graph, constructing a directed acyclic graph that reflects the internal logic of the diagnostic task, thereby clearly expressing the causal dependence and temporal relationship between symptoms, evidence, and hypotheses.

[0125] S3. Based on the embedded vector set, using meta-agents, determine multiple agents with specific medical roles from a predefined role library to establish an intelligent medical diagnosis team.

[0126] For example, step S3 specifically includes the following steps:

[0127] S301. The embedding vector set of the state graph is subjected to attention-based graph pooling operation by the meta-agent to generate a global task vector.

[0128] Specifically, the meta-agent is a high-level control model that has been fine-tuned by instructions. It receives a set of embedded vectors from a dynamic patient state graph. In order to understand the full picture of the current diagnostic task, the meta-agent first performs an attention-based graph pooling operation, which aggregates all node vectors in the set of embedded vectors into a global task vector that focuses on the core contradiction.

[0129] In this embodiment, the attention mechanism can automatically identify key nodes in the graph, so that the global task vector can focus on the core contradiction of the disease.

[0130] The function expression for the global task vector is:

[0131] ;

[0132] In the formula, This is the global task vector; It is the set of all nodes in the dynamic graph; A learnable attention weight vector is used to assign higher weights to key nodes;

[0133] S302. Based on the global task vector, predict the medical specialty field required for the current diagnostic task through the meta-agent;

[0134] Specifically, the meta-agent predicts the medical specialty involved in the current case based on the global task vector through a multi-label classification head; this embodiment ensures specialized treatment by providing prior departmental constraints for subsequent role matching. If the value of a certain dimension in the prediction vector exceeds a preset threshold, it is determined that the corresponding medical specialty needs intervention.

[0135] The function expression for the multi-label classification head is as follows:

[0136] ;

[0137] In the formula, For the predicted medical specialty field vector, It is dimensional vector; For activation functions; This is the weight matrix for the classification heads; For the bias term of the classification head; The total number of predefined medical specialties.

[0138] S303. Based on the capability-demand matching algorithm, calculate the matching score between the embedding vector of the role capability in the predefined role library and the embedding vector of the diagnostic hypothesis node;

[0139] Among them, the predefined role library stores multiple intelligent agent roles with specific medical specialties and personality archetypes;

[0140] Specifically, for each diagnostic hypothesis node in the dynamic patient status graph, its feature vector is regarded as a specific diagnostic demand vector, and the matching degree score between each role in the role library and the node is calculated by the capability-demand matching algorithm.

[0141] The function expression for the capability-demand matching algorithm is as follows:

[0142] ;

[0143] In the formula, The j-th character in the character library; For diagnostic hypothesis nodes; The department to which the diagnostic hypothesis node belongs; Score the match rate; For the role Capability embedding vector; This is the projection matrix used to align the role space with the dynamic patient state space; The feature vector of the hypothetical node for diagnosis; Departmental consistency reward coefficient; I( ) is an indicator function;

[0144] S304. To maximize the coverage of the diagnostic hypothesis nodes and minimize team size and role redundancy, construct a multi-objective optimization function.

[0145] Specifically, the optimal team composition is modeled as a multi-objective optimization problem with a set coverage, and the function expression of the multi-objective optimization function is as follows:

[0146] ;

[0147] In the formula, The optimal set of agent roles; A predefined role library; N represents the set of nodes for diagnosing the hypothesis; N represents the set of candidate agent roles. Penalty for team size; This is a penalty for character redundancy. and This is the penalty coefficient;

[0148] In this embodiment, the team size penalty term refers to a linear penalty on the number of team members, forcing the meta-agent to streamline personnel while still being able to solve the problem; the role redundancy penalty term is used to prevent the selection of two roles with highly overlapping functions.

[0149] The redundancy of the candidate agent role set is expressed as a function:

[0150] ;

[0151] In the formula, The redundancy of the candidate agent role set.

[0152] S305. Solve the multi-objective optimization function to determine the optimal set of intelligent agent roles from the role library, so as to form the optimal intelligent agent diagnosis team;

[0153] Specifically, this embodiment uses a greedy algorithm to approximate the solution of the multi-objective optimization function, selecting the role that brings the greatest marginal coverage gain to join the team at each step, until the gain is lower than the threshold or the number of people reaches the upper limit.

[0154] S306. Generate corresponding personalized prompts for each intelligent agent in the consultation team;

[0155] Specifically, the expression for personalized prompt words is:

[0156] ;

[0157] In the formula, Personalized prompts; Basic instructions; Define the role for the i-th character in the character library; The context view of the i-th character in the character library; Let i be the relevant external tool specification for the i-th character in the character library.

[0158] The role definition includes the role's personality settings, which describe the agent's medical expertise and mission objectives;

[0159] Among them, context vision refers to assigning substructure information of the dynamic patient state graph related to its role as context to each agent according to graph attention weight, thereby achieving sparse input of information and improving the signal-to-noise ratio of inference;

[0160] Among them, the relevant external tool specification refers to the ability of the meta-agent to call specific external tools according to the professional needs of the role;

[0161] Preferably, external tools include general-purpose tools and special-purpose tools; general-purpose tools include, but are not limited to, medical literature retrieval and medical calculators; special-purpose tools include, but are not limited to, radiology agents, image recognition APIs, pharmacology agents, and drug interaction database query interfaces.

[0162] In this embodiment, the meta-agent, as a high-level control model fine-tuned by instructions, is capable of semantic understanding, task planning, and decision scheduling. Specifically, the meta-agent, based on a dynamically constructed patient state graph, utilizes attention-based graph pooling operations to grasp the global characteristics of the task and predict the required specialty domain. Subsequently, through capability-demand matching and multi-objective optimization, the optimal set of agent roles is dynamically determined from a predefined role library. This enables on-demand adaptive generation of heterogeneous intelligent diagnostic teams, overcoming the shortcomings of fixed role presets and rigid resource allocation in existing technologies. It can dynamically assemble a digital expert team with optimal role configuration, clear division of labor, and corresponding professional tools based on the semantic and logical complexity of specific diagnostic tasks, providing a suitable foundation for subsequent efficient and collaborative diagnostic reasoning.

[0163] S4. Using a confidence-weighted Markov game reasoning algorithm, each agent in the intelligent diagnosis team performs multiple rounds of game reasoning in a shared environment based on the embedded vector set until the game convergence condition is met, and outputs the optimal diagnostic hypothesis and the corresponding evidence chain as the final diagnostic conclusion.

[0164] Specifically, the diagnostic reasoning process of the agent team is formalized as a partially observable Markov game, the core elements of which are defined as follows:

[0165] State space: In any round of reasoning, the global state of a Markov game process is defined as follows:

[0166] ;

[0167] in, Let be the global state of the Markov game process in the t-th round of reasoning. ; Let be the current dynamic patient state graph in the t-th round of reasoning, where the nodes and edges in the dynamic patient state graph are dynamically updated as reasoning progresses; Let be the set of diagnostic hypotheses held by all agents in the current round of reasoning (round t). ; Let be the set of confirmed evidence in the t-th round of reasoning, wherein all diagnostic hypotheses must be logically compatible with the evidence set;

[0168] Each agent in the intelligent agent diagnostic team is defined as a tuple:

[0169] ;

[0170] ;

[0171] In the formula, For the intelligent medical diagnosis team; For the i-th intelligent agent; Let be the policy function of the i-th agent, which determines its action probability in a specific state; A specialized knowledge base unique to the i-th intelligent agent; Use the role prompt words for the i-th agent to constrain its language style and thought patterns;

[0172] Action Space: To standardize the interaction between agents, this embodiment defines a discrete action space, in which any agent... Actions that can be executed at any time Belongs to the following set of actions:

[0173] ;

[0174] In the formula, For action space; To propose a new diagnostic hypothesis h; To utilize evidence Refutation of the hypothesis ; To utilize evidence Supporting hypothesis ; To raise a question To obtain more information, such as triggering external tool calls or asking humans; To test hypotheses during the decision-making stage Cast a vote in favor.

[0175] This embodiment introduces adversarial payoffs to prevent groupthink, where each agent aims to maximize its cumulative payoff.

[0176] The function expression for the agent's cumulative reward is as follows:

[0177] ;

[0178] In the formula, The cumulative benefit for the intelligent agent; The maximum information gain reward derived from the action; Rewards are given for consistency within one's own professional knowledge base; Punishment items that contradict confirmed evidence; The penalty coefficient is used to adjust the agent's tolerance for errors; is the discount factor for the reward obtained at time t.

[0179] For example, step S4 specifically includes the following steps:

[0180] S401. Based on the speaking intention function, calculate the speaking intention score of each agent, and determine the agent that obtains the right to speak based on the speaking intention score;

[0181] Specifically, this embodiment adopts a preemptive speaking protocol based on contribution, which dynamically determines the speaking agent in each inference round to replace the fixed sequential polling, thereby improving the efficiency and focus of the debate.

[0182] The contribution-based preemptive speaking protocol has the following function expression:

[0183] ;

[0184] In the formula, The score represents the willingness of the i-th agent to speak in the t-th round. The ability to change the current diagnostic distribution based on the information held by the agent; The degree of conflict between the agent's viewpoint and the current mainstream consensus; The conflict incentive coefficient;

[0185] S402. The agent that gains the right to speak selects an action from the predefined action space to execute and outputs the corresponding argument triples.

[0186] Specifically, the agent that gains the right to speak selects an action from the action space to execute and formalizes its speech into an argument triple;

[0187] The functional expression of the argument triple is:

[0188] ;

[0189] In the formula, For the argument triple of the i-th agent in the t-th round; This is the diagnostic hypothesis that the argument targets; For the action type, type∈{Support,Refute}; The supporting evidence cited for this argument.

[0190] S403. The unspoken agent receives the argument triples, evaluates the credibility of the argument triples using a two-layer Bayesian update algorithm, and then updates its own confidence distribution for each diagnostic hypothesis.

[0191] Specifically, after receiving the argument triples from the speaking agent, the unspeaking agent updates its confidence in each diagnostic hypothesis using a two-layer Bayesian update algorithm; the two-layer Bayesian update algorithm includes an argument quality evaluation function and a hypothesis posterior probability update function.

[0192] The expression for the argument quality assessment function is as follows:

[0193] ;

[0194] In the formula, Provide the credibility of the argument triples for the speaking agent; The learnable weight vector for the argument quality assessment function; The semantic vector of the argument triple; For speaker role vectors; A verification function to validate the authenticity of evidence;

[0195] The expression for the posterior probability update function is assumed to be:

[0196] ;

[0197] In the formula, For the i-th agent when receiving the arguments Then, regarding the diagnostic hypothesis The posterior confidence level; Given a diagnostic hypothesis Under the condition that it is true, observe the evidence. The likelihood probability; For the i-th agent, the diagnostic hypothesis The prior confidence level; H is the set of all possible diagnostic hypotheses;

[0198] To quantify the strength of the argument's support for or refutation of the hypothesis, this embodiment models the likelihood function as the logical implication relationship between the argument and the hypothesis. If the argument refutes the hypothesis, the likelihood function value is lower; if the argument supports the hypothesis, the likelihood function value is higher.

[0199] The expression for the likelihood function is:

[0200] ;

[0201] In the formula, It is a symbolic function; The credibility assessment value for the argument triple; For diagnostic hypothesis With arguments Semantic similarity between them;

[0202] S404. When there is a diagnostic hypothesis with a confidence level exceeding a preset threshold, the intelligent agent in the diagnostic team, which is given the role of a critic, generates attack arguments based on counterfactual reasoning to minimize the overall confidence level of the diagnostic hypothesis.

[0203] Specifically, when the confidence level of a diagnostic hypothesis exceeds a first preset threshold, adversarial criticism is triggered, temporarily designating one or more agents to play the role of critics, whose goal is to maximize the logical uncertainty of the target hypothesis. The critics attempt to retrieve paths or missing evidence that contradict the target hypothesis in the dynamic patient state graph based on counterfactual reasoning in order to generate offensive arguments.

[0204] The functional expression for the aggressive argument is:

[0205] ;

[0206] In the formula, As an offensive argument; This is the current high-confidence target hypothesis; The expected posterior confidence of the target hypothesis after the j-th agent performs an action;

[0207] S405. Repeat steps S401 to S404 until the game convergence condition is met. Then stop the game and output the diagnostic hypothesis with the highest confidence and the corresponding chain of evidence as the final diagnostic conclusion.

[0208] Specifically, when the game reasoning satisfies the Nash equilibrium or reaches the preset maximum number of reasoning rounds, the game reasoning stops, and the diagnostic hypothesis with the highest global confidence at this time is output along with its complete supporting and refuting evidence chain.

[0209] Preferably, the Nash equilibrium is either belief variance convergence or policy stability;

[0210] Wherein, belief variance convergence means that the variance of the confidence of all agents in the current optimal hypothesis is lower than the pre-judgment confidence threshold, that is, the disagreement within the agent's diagnostic team has been eliminated, and the diagnosis of the patient and the factual basis upon which it is based are unquestionable. The functional expression for belief variance convergence is:

[0211] ;

[0212] In the formula, The size of the agent team; Let be the confidence level of agent i regarding the current optimal hypothesis; The average confidence level of the intelligent diagnostic team regarding the current optimal hypothesis; This is the preset variance convergence threshold;

[0213] Policy stability refers to the inability of any agent to gain a significant information gain by unilaterally proposing new arguments; that is, no agent can gain a higher return by unilaterally changing its viewpoint or proposing new arguments. The functional expression for policy stability is:

[0214] ;

[0215] In the formula, The change in the confidence distribution of the entire team toward the set of diagnostic hypotheses H after any agent performs action a in the current state; This is a preset policy stability threshold;

[0216] The functional expression for the diagnostic hypothesis with the highest global confidence is:

[0217] ;

[0218] In the formula, The diagnostic hypothesis with the highest global confidence. The average confidence level of the diagnostic hypothesis h in the intelligent diagnostic team;

[0219] This embodiment constructs a structured multi-agent collaborative reasoning mechanism with self-questioning and correction capabilities. Through a contribution-based speaking mechanism, two-layer Bayesian belief updates, and proactive adversarial criticism, the reasoning process is forced to approach the true diagnosis in a dynamic game, effectively improving the accuracy, robustness, and interpretability of the conclusions.

[0220] S5. If an abnormal risk event occurs during the game reasoning process, the current game reasoning is paused, and arbitration by human experts is requested and accepted. After correcting the current game reasoning path according to the arbitration operation, the game reasoning is returned and continued.

[0221] For example, step S5 specifically includes the following steps:

[0222] S501. Monitor the game reasoning process in real time, and calculate the risk assessment score of the current game reasoning state based on cognitive uncertainty and decision risk.

[0223] Specifically, the functional expression for the risk assessment score of the current game reasoning state is:

[0224] ;

[0225] In the formula, The risk assessment score for the current game reasoning state; To understand uncertainty; To assess the risk level of the decision; These are weighting coefficients;

[0226] Cognitive uncertainty is used to quantify the degree of disagreement within the agent team. It is calculated using the information entropy of the confidence distribution of all agents regarding the current optimal hypothesis, and its functional expression is:

[0227] ;

[0228] In the formula, Let be the confidence level of agent i regarding the current optimal hypothesis;

[0229] Among them, the decision risk level is used to measure the potential severity of the current diagnostic conclusion. It is calculated by multiplying the confidence level of the candidate diagnostic hypothesis with the highest average confidence level by its predefined risk weight. Its functional expression is as follows:

[0230] ;

[0231] In the formula, The average confidence level of the diagnostic hypothesis h across all agents; Predefined risk weights for diagnostic hypothesis h;

[0232] S502. When the risk assessment score exceeds the preset threshold, it is determined that the abnormal risk event has occurred, the game reasoning is suspended, and human experts are requested to intervene.

[0233] S503. Receive arbitration operation input by human expert through interactive interface, correct the current game reasoning path according to the arbitration operation, return and continue the game reasoning.

[0234] Specifically, human experts intervene through an interactive arbitration interface that visualizes the game logic and evidence chain in the background. Experts can perform one or more arbitration operations to correct the current game reasoning state, and after correction, return to continue executing the confidence-weighted Markov game reasoning process until the convergence condition is met or human intervention is triggered again.

[0235] In some embodiments, the arbitration process of human experts includes at least one of the following:

[0236] Forced pruning is an operation used by experts to directly mark a specific diagnostic hypothesis as incorrect and reset the confidence level of that incorrect diagnostic hypothesis to zero across all agents.

[0237] Evidence injection is an operation in which experts input key information that the agent may ignore, in order to add new evidence nodes to the dynamic patient state graph and update the graph structure.

[0238] The weight reset operation is used by experts to adjust the power weight of a specific agent in the game process.

[0239] S6. Encode the arbitration operation of the human expert into a sparse reward value, and optimize the parameters of the corresponding agent using the sparse reward value based on the reinforcement learning algorithm.

[0240] For example, step S6 specifically includes the following steps:

[0241] S601. Encode the arbitration operation as a sparse reward value;

[0242] The functional expression for the sparse reward value is as follows:

[0243] ;

[0244] In the formula, For agent i in state Next action Total rewards received; Rewards for the game environment; Rewards are based on feedback from human experts;

[0245] The rewards based on feedback from human experts include the following:

[0246] Positive reward: If the final decision of the human expert aligns with that of the intelligent agent... The proposal was consistent, and a positive reward value was assigned;

[0247] Negative reward: If the agent's proposal is pruned by a human using a forced pruning operation, a strong negative reward value is assigned;

[0248] Process reward: If a human adopts a piece of evidence cited by the agent, even if the final diagnosis is different, a small positive reward is given to encourage the correct reasoning process.

[0249] S602. By using a credit allocation mechanism based on attention weights, sparse reward values ​​are allocated to the historical action sequence that leads to the arbitration result during the game reasoning process.

[0250] Specifically, to address the credit allocation problem caused by sparse and delayed human feedback, an attention-weight-based backtracking mechanism is adopted to identify the action with the highest semantic relevance to the final intervention point of the human expert in the historical action sequence and assign it a greater weight.

[0251] For a historical action sequence {a0, a1, ..., aM}, the decaying reward obtained from the m-th action is calculated as follows:

[0252] ;

[0253] In the formula, The decaying reward obtained in the m-th step; As a discount factor, ∈[0, 1), used for exponential decay of future rewards; M is the total number of time steps; This is the attention weight vector; For action semantic feature vectors;

[0254] S603. Using reinforcement learning algorithms, update the policy network parameters of the relevant agents based on the allocated reward values.

[0255] Specifically, using the proximal policy optimization algorithm, the policy network parameters of the agent and meta-agent are updated by maximizing the expected cumulative reward. The gradient calculation formula is as follows:

[0256] ;

[0257] In the formula, The gradient of the policy network parameters; For mathematical expectation operators; Importance sampling weights; This is the estimated value of the dominance function at time step m; This is the clipping function; These are the hyperparameters for the near-end policy optimization algorithm.

[0258] In this embodiment, a closed-loop human-in-the-loop optimization mechanism is constructed through steps S5 and S6. This mechanism not only ensures the safety of the diagnostic process through risk-triggered human arbitration, but also empowers the system with the ability to continuously learn and self-evolve by encoding human feedback into sparse rewards and optimizing the agent's strategy. This allows the system to gradually absorb the tacit knowledge of human experts and improve overall diagnostic performance.

[0259] Example 2

[0260] like Figure 2 As shown, this embodiment provides a diagnostic task parsing system 100 based on knowledge graphs and multi-agent game theory, including:

[0261] Data acquisition module 101 is used to acquire patients' multimodal medical data;

[0262] The atlas construction module 102 is connected to the data acquisition module 101 to parse the multimodal medical data in order to construct a dynamic patient state atlas and extract the embedding vector set of the dynamic patient state atlas.

[0263] The dynamic patient state graph includes medical evidence nodes, diagnostic hypothesis nodes, and edges representing the logical relationships between the nodes.

[0264] The agent generation module 103 is connected to the graph construction module 102 and is used to determine agents for multiple specific medical roles from a predefined role library based on the embedded vector set and using meta-agents, so as to establish an intelligent medical diagnosis team.

[0265] The game reasoning module 104 is connected to the agent generation module 103. It is used to utilize the confidence-weighted Markov game reasoning algorithm to enable each agent in the intelligent diagnosis team to perform multiple rounds of game reasoning in a shared environment based on the embedded vector set until the game convergence condition is met, and then output the optimal diagnostic hypothesis vector as the final diagnostic conclusion.

[0266] Arbitration operation module 105 is connected to game reasoning module 104. It is used to pause the current game reasoning if an abnormal risk event occurs during the game reasoning process, request and accept arbitration operation from human experts, correct the current game reasoning path according to the arbitration operation, and then return to the game reasoning module.

[0267] The agent optimization module 106 is connected to the arbitration operation module 105 to encode the arbitration operation of the human expert into sparse reward values, and optimize the parameters of the corresponding agent based on the sparse reward values ​​using a reinforcement learning algorithm.

[0268] In some embodiments, the system achieves efficient collaboration and data flow between modules through the following interface protocol:

[0269] The graph-agent interface is located between the graph construction module 102 and the agent generation module 103. It is used to transmit the semantic embedding vector of the dynamic patient state graph generated by the graph construction module to the agent generation module. Based on the received embedding vector, the agent generation module parses the graph structure to identify key and unresolved nodes, and dynamically assembles an agent diagnosis team accordingly.

[0270] The agent-game interface is located between the agent generation module 103 and the game reasoning module 104. It is used to receive the instantiated agent team and its role configuration information output by the agent generation module and import the team into the shared environment of the game reasoning module.

[0271] The agent-game interface maintains a shared context buffer and implements message distribution based on an attention mechanism, so that each agent can only access the dialogue history and global key evidence related to its role, thereby optimizing long context processing.

[0272] The game-human interface is located between the game reasoning module 104 and the arbitration operation module 105. This interface has a two-way communication function: when the game reasoning module determines that human intervention is required, it sends an intervention request containing the current reasoning state, hypothesis set and debate context to the arbitration operation module through this interface.

[0273] Among them, the interpreter integrated into the game-human interface can translate the technical arguments and actions in the game process into natural language explanations and diagnostic suggestions that are readable by human doctors.

[0274] This embodiment, through the aforementioned modular design and interface protocol, realizes a complete process from multimodal data perception, diagnostic task graph analysis, heterogeneous intelligent agent dynamic orchestration, adversarial collaborative reasoning to human-computer interaction closed-loop optimization, forming an implementable, evolvable, safe and controllable diagnostic task analysis system.

[0275] In one specific embodiment, the complete implementation process of the diagnostic task parsing system in a specific and complex clinical application scenario—cardiovascular disease diagnosis—is provided to clearly and concretely illustrate the collaborative working mechanism and final effect of each technical module;

[0276] In this embodiment, the system is deployed in a high-performance computing server cluster equipped with multiple NVIDIA A100 graphics processors to support real-time computation of large-scale knowledge graph inference and deep neural network models. The following example illustrates the diagnostic process of a patient suspected of having acute coronary syndrome:

[0277] The data acquisition module connects to the hospital information system and reads the electronic medical record data of patient Zhang.

[0278] Among them, the electronic medical record data of patient Zhang includes: male, 65 years old, chief complaint of chest tightness for 3 days, present medical history of hypertension for 10 years, and test result of troponin T: 0.08 ng / mL (slightly high).

[0279] The graph construction module extracts key entities from the electronic medical record data through a pre-trained medical entity recognition model, and then constructs an initial dynamic patient status graph based on the pre-trained medical knowledge graph.

[0280] The key entities mentioned include chest tightness, high blood pressure, and slightly elevated troponin levels.

[0281] Among them, the evidence node for elevated troponin is connected to the diagnostic hypothesis node for myocardial infarction and the diagnostic hypothesis node for myocarditis, respectively, forming a directed acyclic graph of the task.

[0282] The meta-agent in the agent generation module analyzes the directed acyclic graph of the task, identifies that the task involves the fields of cardiology and emergency medicine, and instantiates the following three agents from the predefined role library through capability matching and multi-objective optimization, and generates prompt words for each agent that integrate role definition, personalized context vision and special tool specification.

[0283] Among them, the intelligent agent roles include: a cardiology expert role with conservative strategies, an emergency department expert role that focuses on vital signs, and a medical quality control officer role responsible for logical error correction and quality control.

[0284] In the game reasoning module, each agent engages in multiple rounds of game reasoning within a shared environment:

[0285] In the first round of reasoning, the cardiology expert proposed a diagnostic hypothesis of non-ST-segment elevation myocardial infarction based on elevated troponin levels, and gave an initial confidence level of 0.8.

[0286] In the second round of reasoning, the medical quality control officer refuted the cardiology expert's viewpoint, pointing out that the patient lacked typical chest pain characteristics and that troponin was only slightly elevated, suggesting the need to differentiate between false positives caused by heart failure or renal insufficiency.

[0287] In the third round of reasoning: The cardiology expert initiated a query, calling the system-integrated medical calculator tool to query the patient's glomerular filtration rate. After the tool returned that the glomerular filtration rate was normal, the cardiology expert ruled out the interference of kidney disease and maintained the hypothesis of myocardial infarction, but reduced the confidence level of the myocardial infarction hypothesis to 0.65.

[0288] In the multi-round game reasoning process, all agents dynamically adjust their confidence distribution for different diagnostic hypotheses based on their statements through a two-layer Bayesian update algorithm.

[0289] The arbitration module triggered human intervention because the confidence variance of the myocardial infarction hypothesis was lower than the preset confidence threshold and involved high-risk decision-making. The human doctor viewed the visualized debate process and evidence chain through the arbitration interface, added a piece of information that the agent had ignored, indicating that the patient had recently had flu-like symptoms, and performed an evidence injection operation. At the same time, the system dynamically updated the patient status graph based on the new evidence, and the association weight and probability of the myocarditis hypothesis node increased significantly.

[0290] The agent optimization module encodes the arbitration operation and final ruling of the human expert into sparse reward values. Through a credit allocation mechanism based on attention weights, the rewards are back-allocated to historical reasoning steps. The system also uses a proximal policy optimization algorithm to update the policy network parameters of the relevant agents, enabling the system to proactively prompt for the retrieval of infection history in subsequent similar clinical scenarios, thus achieving continuous learning.

[0291] Finally, after manual correction, the system re-performed the game-theoretic reasoning and converged to the diagnostic hypothesis of acute viral myocarditis. It output the diagnostic hypothesis and its complete chain of supporting and refuting evidence, and recommended that the patient undergo a cardiac MRI for further verification.

[0292] Example 3

[0293] This embodiment provides an electronic device, including at least one control processor and a memory for communicatively connecting to the at least one control processor;

[0294] Memory, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs and non-transitory computer-executable programs. Furthermore, memory may include high-speed random access memory, and may also include non-transitory memory, such as at least one disk storage device, flash memory device, or other non-transitory solid-state storage device. In some embodiments, memory may optionally include memory remotely located relative to the processor, and these remote memories can be connected to the processor via a network. Examples of such networks include, but are not limited to, the Internet, intranets, local area networks, mobile communication networks, and combinations thereof.

[0295] The non-transient software program and instructions required to implement the diagnostic task parsing method based on knowledge graphs and multi-agent game theory in the above embodiments are stored in memory. When executed by the processor, the diagnostic task parsing method based on knowledge graphs and multi-agent game theory in the above embodiments is executed. For example, the method described above is executed. Figure 1 Method steps S1 to S6;

[0296] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0297] Example 4

[0298] This embodiment provides a computer-readable storage medium storing computer-executable instructions for causing a computer to execute a diagnostic task parsing method based on knowledge graphs and multi-agent game theory as described in Embodiment 1.

[0299] It should be noted that the computer-readable storage medium in this embodiment may be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. The computer-readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof.

[0300] More specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0301] In this embodiment, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this embodiment, the computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying a computer-readable program. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof.

[0302] The computer-readable signal medium may also be any computer-readable storage medium other than a computer-readable storage medium, which can send, propagate or transmit a program for use by or in connection with an instruction execution system, apparatus or device.

[0303] The computer program contained on the computer-readable storage medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof. The aforementioned computer-readable storage medium can be used to write a computer program for performing this embodiment in one or more programming languages ​​or combinations thereof. These programming languages ​​include object-oriented programming languages—such as Python, Java, and C++—and conventional procedural programming languages—such as C or similar programming languages. The program can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving a remote computer, the remote computer can be connected to the user's computer via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0304] The above description is merely a preferred embodiment of the present invention. It will be apparent to those skilled in the art that the present invention is not limited to the details of the above exemplary embodiments, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects. The scope of the present invention is defined by the appended claims rather than the foregoing description, and thus all variations falling within the meaning and scope of the equivalents of the claims are intended to be included within the present invention.

Claims

1. A diagnostic task parsing method based on knowledge graphs and multi-agent game theory, characterized in that, include: Acquire patients' multimodal medical data; The multimodal medical data is parsed to construct a dynamic patient state graph, and the embedding vector set of the dynamic patient state graph is extracted. The dynamic patient state graph includes medical evidence nodes, diagnostic hypothesis nodes, and edges representing the logical relationships between the nodes. Based on the aforementioned set of embedded vectors, using meta-agents, multiple agents with specific medical roles are identified from a predefined role library to establish an intelligent medical diagnosis team. Using a confidence-weighted Markov game reasoning algorithm, each agent in the intelligent diagnosis team can perform multiple rounds of game reasoning in a shared environment based on the embedded vector set until the preset game convergence condition is met, and then output the optimal diagnostic hypothesis and the corresponding evidence chain as the final diagnostic conclusion. If an abnormal risk event occurs during the game reasoning process, the current game reasoning is paused, and arbitration by human experts is requested and accepted. After the current game reasoning path is corrected according to the arbitration, the game reasoning is returned and continues. The process of parsing the multimodal medical data to construct a dynamic patient state map and extracting the embedding vector set of the dynamic patient state map specifically includes the following steps: Based on a pre-trained medical entity recognition model, the multimodal medical data is parsed to identify and extract medical entities; The medical entities include symptom entities, examination result entities, drug entities, and disease entities; Based on the attribute enhancement function, the attribute vector of the corresponding medical entity is generated; Based on a pre-trained medical knowledge graph, the symptom entities and examination result entities are used as medical evidence nodes, and disease nodes associated with the evidence nodes are retrieved as diagnostic hypothesis nodes. Based on the relevance filtering mechanism, the weights of the edges between entity nodes are calculated to form a directed acyclic graph that reflects the current diagnostic task status, which serves as a dynamic patient status graph. Based on a graph attention network, the dynamic patient state graph is encoded to generate a set of embedding vectors containing graph topological features and semantic information.

2. The method according to claim 1, characterized in that, The method further includes: The arbitration operations of the human experts are encoded as sparse reward values, and the parameters of the corresponding agents are optimized using the sparse reward values ​​based on a reinforcement learning algorithm.

3. The method according to claim 1, characterized in that, Based on the embedded vector set, multiple agents with specific medical roles are determined from a predefined role library using meta-agents to establish an intelligent medical diagnosis team. This process specifically includes the following steps: By using a meta-agent to perform attention-based graph pooling on the set of embedded vectors of the state graph, a global task vector is generated. Based on the global task vector, the meta-agent predicts the medical specialty field required for the current diagnostic task. Based on the capability-demand matching algorithm, the matching degree score between the embedding vector of the capability in the predefined role library and the embedding vector of the diagnostic hypothesis node is calculated; To maximize coverage of the diagnostic hypothesis nodes while minimizing team size and role redundancy, a multi-objective optimization function is constructed. Solve the multi-objective optimization function to determine the optimal set of intelligent agent roles from the role library, so as to form the optimal intelligent agent diagnosis team; Generate corresponding personalized prompts for each intelligent agent in the consultation team; The prompts include basic instructions, role definitions, contextual perspectives based on graph attention weight allocation, and relevant external tool specifications.

4. The method according to claim 1, characterized in that, The confidence-weighted Markov game reasoning algorithm enables each agent in the intelligent diagnostic team to conduct multiple rounds of game reasoning in a shared environment based on the embedded vector set until a preset game convergence condition is met. Then, it outputs the optimal diagnostic hypothesis and the corresponding evidence chain as the final diagnostic conclusion. Specifically, it includes the following steps: A1. Based on the speaking intention function, calculate the speaking intention score of each agent, and determine the agent that obtains the right to speak based on the speaking intention score; A2. The agent that gains the right to speak selects an action from the predefined action space to execute and outputs the corresponding argument triples; The argument triad includes the target assumption, the type of action, and supporting evidence; A3. The unspeaking agent receives the argument triples, evaluates the credibility of the argument triples using a two-layer Bayesian update algorithm, and then updates its own confidence distribution for each diagnostic hypothesis. The two-layer Bayesian update algorithm includes an argument quality evaluation function and a hypothesis posterior probability update function; A4. When there is a diagnostic hypothesis with a confidence level exceeding a preset threshold, the AI ​​agent in the diagnostic team, which is given the role of a critic, generates attack arguments based on counterfactual reasoning to minimize the overall confidence level of the diagnostic hypothesis. A5. Repeat steps A1 to A4 until the game convergence condition is met. Then stop the game and output the diagnostic hypothesis with the highest confidence and the corresponding chain of evidence as the final diagnostic conclusion.

5. The method according to claim 1, characterized in that, If an abnormal risk event occurs during the game reasoning process, the current game reasoning is paused, and arbitration by a human expert is requested and accepted. After correcting the current game reasoning path according to the arbitration, the game reasoning is returned to and continues. Specifically, this includes the following steps: The game reasoning process is monitored in real time, and a risk assessment score for the current game reasoning state is calculated based on cognitive uncertainty and decision risk. When the risk assessment score exceeds a preset threshold, it is determined that the abnormal risk event has occurred, the game reasoning is suspended, and human experts are requested to intervene. The system receives arbitration input from human experts through an interactive interface, corrects the current game reasoning path based on the arbitration input, and then returns to continue the game reasoning. The cognitive uncertainty is quantified by the information entropy of the confidence distribution of all expert agents on the optimal diagnostic hypothesis; The decision risk is quantified by multiplying the average confidence level of the current diagnostic hypothesis by a predefined risk weight.

6. The method according to claim 2, characterized in that, The arbitration operation of the human expert is encoded as a sparse reward value, and the parameters of the corresponding agent are optimized using the sparse reward value based on a reinforcement learning algorithm. Specifically, this includes the following steps: The arbitration operations of the human experts are encoded as sparse reward values; By using a credit allocation mechanism based on attention weights, sparse reward values ​​are allocated to relevant actions in the game reasoning history that lead to the arbitration result. Using reinforcement learning algorithms, the policy network parameters of relevant agents are updated based on the allocated sparse reward values. The step of encoding the arbitration operation of the human expert into a sparse reward value specifically includes: If the agent's proposal is consistent with the final decision of the human expert, it is determined to be a positive reward value; If the agent's proposal is rejected by the human expert's arbitration operation, it is determined to be a negative reward value; If the evidence cited by the agent is adopted by the human expert, it is determined as a procedural reward value.

7. The method according to claim 5 or 6, characterized in that, The arbitration procedure includes at least one of the following: Forced pruning is used to mark a specific diagnostic hypothesis as incorrect and reset the confidence of that incorrect diagnostic hypothesis to zero across all agents. The evidence injection operation is used to add new evidence nodes to the dynamic patient state graph and update the dynamic patient state graph. The weight reset operation is used to adjust the power weight of a specific agent in the game process.

8. A diagnostic task parsing system based on knowledge graphs and multi-agent game theory, characterized in that, include: The data acquisition module is used to acquire patients' multimodal medical data; The graph construction module is used to parse the multimodal medical data to construct a dynamic patient state graph and extract the embedding vector set of the dynamic patient state graph. The process of parsing the multimodal medical data to construct a dynamic patient state map and extracting the embedding vector set of the dynamic patient state map specifically includes the following steps: Based on a pre-trained medical entity recognition model, the multimodal medical data is parsed to identify and extract medical entities; The medical entities include symptom entities, examination result entities, drug entities, and disease entities; Based on the attribute enhancement function, the attribute vector of the corresponding medical entity is generated; Based on a pre-trained medical knowledge graph, the symptom entities and examination result entities are used as medical evidence nodes, and disease nodes associated with the evidence nodes are retrieved as diagnostic hypothesis nodes. Based on the relevance filtering mechanism, the weights of the edges between entity nodes are calculated to form a directed acyclic graph that reflects the current diagnostic task status, which serves as a dynamic patient status graph. Based on a graph attention network, the dynamic patient state graph is encoded to generate a set of embedding vectors containing graph topological features and semantic information. The dynamic patient state graph includes medical evidence nodes, diagnostic hypothesis nodes, and edges representing the logical relationships between the nodes. The agent generation module is used to determine agents for multiple specific medical roles from a predefined role library based on the set of embedded vectors and using meta-agents, so as to establish an intelligent medical diagnosis team. The game reasoning module is used to utilize a confidence-weighted Markov game reasoning algorithm to enable each agent in the intelligent diagnosis team to perform multiple rounds of game reasoning in a shared environment based on the embedded vector set until the preset game convergence condition is met, and then output the optimal diagnostic hypothesis vector as the final diagnostic conclusion. The arbitration module is used to pause the current game reasoning process if an abnormal risk event occurs, request and accept arbitration from human experts, correct the current game reasoning path according to the arbitration operation, and then return to the game reasoning module.

9. The system according to claim 8, characterized in that, The system also includes: The agent optimization module encodes the arbitration operation of the human expert into sparse reward values ​​and optimizes the parameters of the corresponding agent based on the sparse reward values ​​using a reinforcement learning algorithm.

Images