Multi-agent large model disease diagnosis knowledge reasoning system based on data double driving

Abstract

The application discloses a multi-agent large model disease diagnosis knowledge reasoning system based on data double driving, relates to the technical field of artificial intelligence assisted medical diagnosis, and through multi-source collection of patient symptom follow-up records, laboratory test results, observation diagnosis probability and expert diagnosis recommendation results, the system extracts time sequence evolution characteristics and calculates a time sequence diagnosis sensitivity coefficient, realizes early identification of disease risk; combined with counterfactual simulation and statistical reasoning, a causal consistency coefficient is obtained, which is used for verifying the causal rationality of observation diagnosis and control results; based on agent group consensus analysis, a game consistency coefficient is calculated, which is used for judging the credibility of the diagnosis conclusion; abnormal reasoning steps and knowledge fragments are positioned and verified in multiple levels, so that the reliability and safety of the results are guaranteed; through log analysis and knowledge backflow mechanism, the diagnosis model is continuously optimized; the application can significantly improve the accuracy, explainability and safety of disease diagnosis.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted medical diagnosis technology, specifically to a disease diagnosis knowledge reasoning system based on a dual-drive approach of real-world clinical data and a medical knowledge base, integrating a multi-agent large model, causal reasoning, and reasoning game / multi-expert consultation mechanism. Background Technology

[0002] Clinical diagnosis typically relies on a comprehensive assessment of multimodal information, including patient complaints and medical history, physical examination, laboratory tests, imaging, and pathology. With the development of medical informatics and artificial intelligence technologies, computer-aided diagnostic (CAD) systems based on rule bases, statistical learning, or single-model inference have been applied in some specialties. However, in complex real-world cases and interdisciplinary collaborative scenarios, existing technologies still have insurmountable limitations, specifically:

[0003] Multimodal data fragmentation and semantic inconsistency. Clinical data is scattered across heterogeneous platforms such as electronic medical records, laboratory / examination systems, follow-up and nursing records, with inconsistent data structures, coding systems, and quality, lacking a unified semantic alignment and standardized integration mechanism. This makes it difficult to compare and integrate evidence in the same representation space, and key clues are easily missed or repeated, affecting the completeness and consistency of diagnosis.

[0004] The current clinical decision-making process relies heavily on individual physician experience and limited guidelines. For complex, rare, or multi-disease cases, significant differences in conclusions exist among different experts. There is a lack of systematic multi-expert consultation organizations and a "proposal-support-refutation-convergence" argumentation process. Furthermore, there is a lack of quantitative indicators and thresholds to transform diverse opinions into collective consensus, making it difficult to objectively assess the stability and credibility of conclusions.

[0005] Insufficient causal reasoning ability and lack of counterfactual analysis. Most systems rely primarily on correlation modeling, focusing on the statistical association between features and disease labels, lacking structured causal representation and testing of pathophysiological mechanisms and treatment pathways. They also lack the ability to perform counterfactual simulations using causal graphs or structured causal models (SCM), making it difficult to assess the robustness and transferability of diagnostic conclusions "if evidence changes / interventions differ".

[0006] The system suffers from weak security verification and traceability capabilities. Existing systems offer limited support for the traceability of reasoning and evidence chains, lacking automatic location and multi-layered verification mechanisms for anomalous knowledge fragments, conflicting evidence, and high-impact nodes. High-risk conclusions are difficult to trigger timely alerts and remediation processes, failing to meet the stringent requirements of reliability, interpretability, and compliance in medical settings.

[0007] Knowledge updates are lagging, and there is a lack of closed-loop evolution. Diagnostic knowledge bases and models rely heavily on manual maintenance and offline updates, making it difficult to promptly absorb new evidence from clinical logs, consultation processes, validation feedback, and counterexample libraries. The lack of automated channels for structured accumulation of practical experience and its feedback to the knowledge and model layers hinders the system's ability to achieve continuous optimization and adaptive evolution.

[0008] The human-machine collaboration and governance framework is incomplete. In complex diagnostic processes involving large models and multiple agents, governance elements such as role responsibilities, authority boundaries, compliance of external knowledge citations, and attribution of responsibility lack clear specifications, which limits the large-scale implementation of the system in real medical institutions and the approval rate of regulatory reviews.

[0009] In summary, there is an urgent need for an intelligent diagnostic system that integrates multi-intelligence diagnosis and game-like reasoning, causal reasoning and counterfactual simulation within a unified data and knowledge framework, and possesses traceable security verification and knowledge closed-loop evolution capabilities, in order to improve the accuracy, robustness and auditability in complex disease scenarios. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this invention provides a multi-agent large-scale model disease diagnosis knowledge reasoning system based on dual data-driven approaches, in order to solve the problems mentioned in the background art.

[0011] To achieve the above objectives, the present invention provides the following technical solution: a multi-agent large-scale disease diagnosis knowledge reasoning system based on data-driven dual-engineering, comprising:

[0012] include:

[0013] The data acquisition module is used to collect patient information from multiple sources, including patient complaints and follow-up records, physical examinations, multimodal clinical data from imaging, laboratory tests, and genes, observational diagnostic probability data, and diagnostic suggestions and interpretive evidence from various expert agents in the multi-agent platform; the "dual drive" refers to the synergistic drive of real-world clinical data and medical knowledge base / guidelines / case database;

[0014] The representation layer modeling module is used to construct patient representation vectors and evidence saliency distributions based on multimodal data, and generate interpretable labels for key evidence items; without relying on temporal features, it completes the extraction of associations and conflict resolution between symptoms, signs and tests, forming standardized cases and evidence for reasoning.

[0015] The knowledge hypergraph construction module is used to construct a knowledge hypergraph of disease-evidence-treatment triples based on clinical data and knowledge base, and to perform counterfactual simulation under the structured causal model (SCM). It obtains the observed diagnosis probability Pobs and the control diagnosis probability Pcf through statistical and counterfactual reasoning, calculates the causal consistency coefficient CIC, and compares it with the causal consistency threshold Cth to determine whether the causal consistency between the observed diagnosis and the counterfactual control results is qualified. If it is qualified, the causal chain results are passed to the reasoning game module; if it is not qualified, the corresponding strategy is given.

[0016] The reasoning and game theory module sets up a group of role-based experts. Each agent provides a diagnostic claim, a set of supporting evidence, and counterexamples (identification diagnosis). It constructs an argument graph of claim-support-refutation and executes multi-round collaboration of "proposal-questioning-defense-convergence". It allows the use of external authoritative knowledge fragments for corroboration. It obtains the similarity Rxs of the group consensus by calculating the information entropy threshold, calculates the game consistency coefficient GCI, and compares it with the game consistency threshold Gth to determine whether the diagnostic conclusion is credible. If it is not credible, it gives an appropriate strategy.

[0017] The diagnostic traceability safety verification module, based on the causal consistency coefficient (CIC) and game-theoretic consistency coefficient (GCI), performs difference analysis between diagnostic results and follow-up or historical data, automatically locating abnormal knowledge fragments and reasoning steps; the safety verification unit uses multi-level verification for abnormal nodes to check the reliability of diagnostic results, and triggers repair and generates safety warnings when risks are detected, so that the final diagnostic results are both consistent with clinical data and have passed safety verification.

[0018] The knowledge evolution and update module performs structured parsing and knowledge extraction on the log data generated during the diagnosis and verification process, generates update candidates, and after expert review, feeds them back to the knowledge layer and diagnostic model to achieve continuous iteration and optimization of system knowledge.

[0019] Preferably, the data acquisition module is used to collect multimodal information such as patient complaints and follow-up records, physical examinations, and imaging / laboratory / pathological / genetic data through the electronic medical record system and follow-up terminal; to collect continuous results of the patient's blood, urine, and biochemical test indicators through laboratory testing equipment; to collect observational diagnostic probability data by calling the diagnostic knowledge hypergraph and historical case database; and to collect diagnostic recommendation results from clinicians, imaging experts, laboratory experts, and guideline experts through a virtual multi-agent interaction platform.

[0020] Preferably, the representation layer modeling module includes a temporal feature unit, a first calculation unit, and a first analysis unit;

[0021] The temporal feature unit is used to calculate the patient symptom follow-up records through temporal modeling and differential algorithms to obtain the time change rate Tsym of the symptom evolution trajectory; and to process the laboratory test result sequence through dynamic statistical analysis and fluctuation detection methods to obtain the dynamic fluctuation value Vlab of key test indicators.

[0022] It is used to structure and semantically align clinical data from different sources and formats to form patient representation vectors and evidence packages; the first computing unit is used to calculate and obtain the evidence significance weight and evidence consistency coefficient ECT based on evidence quality, source reliability and matching degree with disease prior, and generate interpretable labels.

[0023] The first analysis unit is used to evaluate the evidence package based on a preset evidence sufficiency threshold and conflict degree threshold, and obtain the first evaluation result including:

[0024] When the evidence is sufficient and the degree of conflict is low, the process proceeds to the normal knowledge reasoning process; when the evidence is insufficient or the degree of conflict is high, the first warning instruction is triggered, and the first strategy is generated: a priority supplementary sampling list and key examination recommendations are given for differential diagnosis, and the individual is marked as a "key individual" in the clinical auxiliary diagnostic system.

[0025] Preferably, the knowledge hypergraph construction module includes a simulation reasoning unit, a second calculation unit, and a second analysis unit;

[0026] The simulation reasoning unit is used to couple the established time series map with clinical data, and combine the disease, phenotype, gene, molecular component, physiological process, pathway ontology knowledge and causal relationship of the observed diagnostic probability data to perform counterfactual simulation; through statistical reasoning methods, the results of the knowledge hypergraph and historical cases are normalized to obtain the observed diagnostic probability Pobs; through the counterfactual reasoning algorithm, the results under the comparison scenario are calculated to obtain the comparison diagnostic probability Pcf.

[0027] The second calculation unit is used to calculate the causal consistency coefficient CIC by obtaining the observed diagnostic probability Pobs and the control diagnostic probability Pcf, after dimensionless processing.

[0028] Preferably, the second analysis unit is used to preset a causal consistency threshold Cth, and compare the causal consistency coefficient CIC with the causal consistency threshold Cth to obtain a second evaluation result, including:

[0029] When the causal consistency coefficient CIC ≥ causal consistency threshold Cth, it indicates that the causal consistency between the observation diagnosis and the counterfactual comparison results is qualified, the reasoning result is credible, and it enters the regular knowledge graph reasoning process and transmits the causal chain result to the reasoning game module.

[0030] When the causal consistency coefficient CIC < causal consistency threshold Cth, it indicates that the causal consistency between the observed diagnosis and the counterfactual control results is unqualified, the inference result is unreliable, triggering a second warning instruction and generating a second strategy: generating a test optimization and differential diagnosis scheme, prompting the clinical auxiliary diagnostic system to further verify.

[0031] Preferably, the reasoning game module includes a game negotiation unit, a third calculation unit, and a third analysis unit;

[0032] The game negotiation unit is used to analyze the diagnostic recommendation results of clinicians, imaging experts, laboratory experts and guideline experts based on the causal chain results and through similarity measurement methods to obtain the similarity Rxs between the agent and the group consensus;

[0033] The third computing unit is used to calculate the game consensus coefficient GCI by obtaining the similarity Rxs between the agent and the group consensus, after dimensionless processing. GCI is mainly composed of the knowledge graph connectivity KG. connect Consistency between authoritative database disease records (DB) and phenotypic matching degree (P) match Gene matching degree G match Literature support (Lit) and the weights of each piece of evidence (W) evid constitute.

[0034] Preferably, the third calculation unit is used to preset the game consistency threshold Gth, and compare and analyze the game consistency coefficient GCI with the game consistency threshold Gth to obtain the third evaluation result, including:

[0035] When the game consistency coefficient GCI ≥ game consistency threshold Gth, it means that there is no disagreement among the experts in their diagnosis, the diagnosis is credible, and the process of unified knowledge reasoning and decision output begins.

[0036] When the game consistency coefficient GCI < game consistency threshold Gth, it indicates that there is a disagreement among experts in their diagnostic results, and the diagnostic conclusion is unreliable. This triggers a third warning instruction and generates a third strategy: outputting expert disagreement prompts and generating differentiated diagnostic schemes and additional testing suggestions for further confirmation by the clinical auxiliary diagnostic system.

[0037] Preferably, the diagnostic traceability security verification module includes a traceability analysis unit and a security verification unit;

[0038] The source tracing analysis unit is used to perform difference analysis between diagnostic results and actual follow-up records or historical case data based on the time-series diagnostic sensitivity coefficient (SDC), causal consistency coefficient (CIC), and game consistency coefficient (GCI). When the diagnostic results are found to be inconsistent with the follow-up results or there is an abnormal deviation, the unit automatically identifies the relevant knowledge fragments, reasoning steps, and causal chains that are abnormal and marks and locates the abnormal nodes.

[0039] The security verification unit is used to perform comprehensive verification of the diagnostic results by using hierarchical verification methods at the individual, cross-modal, and guideline levels through the located abnormal nodes, so that the inference output conforms to the medical safety boundary.

[0040] Preferably, the knowledge evolution and update module is used to perform structured parsing and knowledge extraction on the log data generated during the diagnosis, verification and tracing process, and form knowledge update candidates. Combined with the human expert review mechanism, the data is fed back to the knowledge layer and the diagnostic model to achieve continuous knowledge evolution.

[0041] This invention provides a multi-agent large-scale model disease diagnosis knowledge reasoning system based on dual data-driven approaches. It offers the following advantages:

[0042] (1) The multi-agent large model disease diagnosis knowledge reasoning system based on dual data drive comprehensively collects patient symptoms, test indicators, diagnostic probabilities and expert recommendation results through the data acquisition module, and calculates the temporal diagnostic sensitivity coefficient SDC in combination with the representation layer modeling module, so as to realize early identification and accurate warning of abnormal evolution trend of patients and improve the pertinence of clinical follow-up and intervention.

[0043] (2) The multi-agent large model disease diagnosis knowledge reasoning system based on dual data driving combines counterfactual simulation and statistical reasoning methods in the knowledge hypergraph construction module to calculate the causal consistency coefficient CIC, verify the observation diagnosis and control diagnosis results, effectively improve the causal credibility of the diagnosis results and reduce the risk of misdiagnosis.

[0044] (3) The multi-agent large model disease diagnosis knowledge reasoning system based on dual knowledge data driving mechanism uses the reasoning game module to calculate the game consistency coefficient (GCI), identify diagnostic disagreements among experts and generate differentiated diagnostic strategies, thereby achieving collaborative optimization of agent and group consensus and improving the scientificity and reliability of multi-expert decision-making.

[0045] (4) The multi-agent large model disease diagnosis knowledge reasoning system based on dual knowledge and data drive has a diagnosis traceability security verification module that combines multi-level verification mechanism rules / guidelines hard constraints, counterfactual re-simulation, and adversarial questioning to ensure that the diagnosis results meet the medical safety boundaries; the knowledge evolution and update module parses and extracts knowledge from log data, and combines expert review with the return to the model to realize the continuous iteration of diagnostic knowledge and the intelligent upgrade of the system. Attached Figure Description

[0046] Figure 1 This is a flowchart illustrating the multi-agent large-scale disease diagnosis knowledge reasoning system based on dual data-driven approaches of the present invention.

[0047] Figure 2This is a flowchart illustrating the steps involved in establishing a diagnostic model. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] Example 1

[0050] Please see Figure 1 This invention provides a multi-agent large-scale model disease diagnosis knowledge reasoning system based on data-driven dual-engineering, comprising:

[0051] The data acquisition module is used to collect patient information from multiple sources, including patient complaints and follow-up records, physical examinations, multimodal clinical data from imaging, laboratory tests, and genes, observational diagnostic probability data, and diagnostic suggestions and interpretive evidence from various expert agents in the multi-agent platform; the "dual drive" refers to the synergistic drive of real-world clinical data and medical knowledge base / guidelines / case database;

[0052] The representation layer modeling module is used to construct patient representation vectors and evidence saliency distributions based on multimodal data, and generate interpretable labels for key evidence items; without relying on temporal features, it completes the extraction of associations and conflict resolution between symptoms, signs and tests, forming standardized cases and evidence for reasoning.

[0053] The knowledge hypergraph construction module is used to construct a knowledge hypergraph of disease-evidence-treatment triples based on clinical data and knowledge base, and to perform counterfactual simulation under the structured causal model (SCM). It obtains the observed diagnosis probability Pobs and the control diagnosis probability Pcf through statistical and counterfactual reasoning, calculates the causal consistency coefficient CIC, and compares it with the causal consistency threshold Cth to determine whether the causal consistency between the observed diagnosis and the counterfactual control results is qualified. If it is qualified, the causal chain results are passed to the reasoning game module; if it is not qualified, the corresponding strategy is given.

[0054] The reasoning and game theory module sets up a group of role-based experts. Each agent provides a diagnostic claim, a set of supporting evidence, and counterexamples (identification diagnosis). It constructs an argument graph of claim-support-refutation and executes multi-round collaboration of "proposal-questioning-defense-convergence". It allows the use of external authoritative knowledge fragments for corroboration. It obtains the similarity Rxs of the group consensus by calculating the information entropy threshold, calculates the game consistency coefficient GCI, and compares it with the game consistency threshold Gth to determine whether the diagnostic conclusion is credible. If it is not credible, it gives an appropriate strategy.

[0055] The diagnostic traceability safety verification module, based on the causal consistency coefficient (CIC) and game-theoretic consistency coefficient (GCI), performs difference analysis between diagnostic results and follow-up or historical data, automatically locating abnormal knowledge fragments and reasoning steps; the safety verification unit uses multi-level verification for abnormal nodes to check the reliability of diagnostic results, and triggers repair and generates safety warnings when risks are detected, so that the final diagnostic results are both consistent with clinical data and have passed safety verification.

[0056] The knowledge evolution and update module performs structured parsing and knowledge extraction on the log data generated during the diagnosis and verification process, generates update candidates, and after expert review, feeds them back to the knowledge layer and diagnostic model to achieve continuous iteration and optimization of system knowledge.

[0057] In this embodiment, by constructing a multi-module collaborative system covering data acquisition, representation modeling, knowledge hypergraph construction, reasoning game, diagnostic tracing and verification, and knowledge evolution and updating, comprehensive analysis and intelligent reasoning of multi-source clinical information of patients can be achieved. This not only enables early identification of abnormal evolution trends, improves diagnostic credibility and consistency of multi-expert decision-making, but also ensures that the final diagnostic results are consistent with clinical data and have passed security verification through multi-level security verification and continuous knowledge iteration, thereby significantly improving the accuracy, reliability and intelligence level of disease diagnosis.

[0058] Example 2

[0059] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the data acquisition module is used to collect multimodal information such as patient complaints and follow-up records, physical examinations, and imaging / laboratory / pathological / genetic data through the electronic medical record system and follow-up terminal; to collect continuous results of patients' blood, urine, and biochemical test indicators through laboratory testing equipment; to collect observational diagnostic probability data by calling the diagnostic knowledge hypergraph and historical case database; and to collect diagnostic recommendation results from clinicians, imaging experts, laboratory experts, and guideline experts through a virtual multi-agent interaction platform.

[0060] In this embodiment, the data acquisition module comprehensively collects multi-source information such as patient symptom follow-up records, laboratory test results, diagnostic probability data, and multi-agent expert diagnostic recommendations, providing a complete and rich data foundation for subsequent time series modeling, knowledge reasoning, and multi-expert collaborative decision-making, thereby improving the comprehensiveness and accuracy of disease diagnosis.

[0061] Example 3

[0062] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the representation layer modeling module includes a temporal feature unit, a first calculation unit, and a first analysis unit;

[0063] The temporal feature unit is used to calculate the patient symptom follow-up records through temporal modeling and differential algorithms to obtain the time change rate Tsym of the symptom evolution trajectory; and to process the laboratory test result sequence through dynamic statistical analysis and fluctuation detection methods to obtain the dynamic fluctuation value Vlab of key test indicators.

[0064] The first calculation unit is used to calculate the time-series diagnostic sensitivity coefficient SDC by using the acquired symptom evolution trajectory's time change rate Tsym and the dynamic fluctuation value Vlab of key test indicators, after dimensionless processing. The formula is as follows:

[0065]

[0066] In the formula, a1 and a2 represent weighting coefficients.

[0067] The time change rate (Tsym), which characterizes the trajectory of symptom evolution, has a high weighting on the sensitivity coefficient of temporal diagnosis. It is a key indicator that directly reflects the dynamic evolution trend of patient symptoms over time and makes a core contribution to early disease risk identification.

[0068] The dynamic fluctuation value Vlab, which characterizes the key test indicators, has the second highest weighting on the sensitivity coefficient of time-series diagnosis. It reflects the abnormal fluctuation of the patient's laboratory test parameters over a continuous period of time and is an important basis for helping to reveal potential pathological changes.

[0069] By constructing a temporal diagnostic sensitivity coefficient (SDC) that is a weighted fusion of the rate of change of symptom evolution trajectory and the dynamic fluctuation value of test indicators, the temporal evolution sensitivity of individuals under multimodal clinical data can be quantified, providing a scientific basis for disease risk warning and diagnostic decision-making.

[0070] It is used to structure and semantically align clinical data from different sources and formats to form patient representation vectors and evidence packages; the first computing unit is used to calculate and obtain the evidence significance weight and evidence consistency coefficient ECT based on the matching degree of evidence quality, source reliability and disease prior, and generate interpretable labels.

[0071] In this embodiment, the presentation layer modeling module is used to quantitatively analyze the evolution trajectory of patient symptoms and the dynamic fluctuations of key test indicators, and calculate the temporal diagnostic sensitivity coefficient (SDC) to achieve early identification of the patient's disease evolution trend. This provides a scientific basis for timely intervention and personalized treatment, and improves the timeliness and accuracy of diagnosis.

[0072] Example 4

[0073] This embodiment is an explanation based on Embodiment 3. Please refer to it. Figure 1 Specifically, the first analysis unit is used to preset a sensitivity threshold Sth, and compare the time-series diagnostic sensitivity coefficient SDC with the sensitivity threshold Sth to obtain a first evaluation result, including:

[0074] When the time-series diagnostic sensitivity coefficient SDC < sensitivity threshold Sth, it indicates that the current patient does not show an abnormal evolution trend and enters a routine follow-up monitoring state, where multimodal clinical data are collected regularly and a time-series atlas is established.

[0075] When the sequential diagnostic sensitivity coefficient SDC ≥ sensitivity threshold Sth, the current patient shows an abnormal evolution trend and has an early disease risk, triggering the first warning instruction and generating the first strategy: it is recommended to prioritize key laboratory tests and imaging re-examinations, and focus on monitoring abnormal indicators; and mark the patient as a "key individual" in the clinical auxiliary diagnostic system.

[0076] When the evidence is sufficient and the degree of conflict is low, the process proceeds to the normal knowledge reasoning process; when the evidence is insufficient or the degree of conflict is high, the first warning instruction is triggered, and the first strategy is generated: a priority supplementary sampling list and key examination recommendations are given for differential diagnosis, and the individual is marked as a "key individual" in the clinical auxiliary diagnostic system.

[0077] The sensitivity threshold Sth is obtained by statistically analyzing a large amount of disturbance response data during the operation of target systems, extracting the distribution range of the system sensitivity index under different environmental conditions, load levels, and key parameter fluctuations, and combining the experience judgment of domain experts on system stability and robustness to determine a reasonable sensitivity threshold. With reference to relevant industry technical specifications and historical operation case data, this threshold is used to effectively distinguish whether the system is in a state of high sensitivity to external disturbances, thereby ensuring the scientific nature and preventiveness of control measures.

[0078] In this embodiment, a sensitivity threshold Sth is set by the first analysis unit, and dynamic comparative analysis is performed in combination with the time-series diagnostic sensitivity coefficient SDC. This can automatically distinguish between normal and abnormal risk states in the early stages of a patient's disease progression. This not only ensures routine follow-up monitoring for stable patients, but also triggers early warnings for high-risk individuals and generates personalized treatment strategies, thereby improving the effectiveness of early disease intervention and clinical decision support.

[0079] Example 5

[0080] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the knowledge hypergraph construction module includes a simulation reasoning unit, a second computing unit, and a second analysis unit;

[0081] The simulation reasoning unit is used to couple the established time series map with clinical data, and combine the disease, phenotype, gene, molecular component, physiological process, pathway ontology knowledge and causal relationship of the observed diagnostic probability data to perform counterfactual simulation; through statistical reasoning methods, the results of the knowledge hypergraph and historical cases are normalized to obtain the observed diagnostic probability Pobs; through the counterfactual reasoning algorithm, the results under the comparison scenario are calculated to obtain the comparison diagnostic probability Pcf.

[0082] The second calculation unit is used to calculate the causal consistency coefficient CIC by obtaining the observed diagnostic probability Pobs and the control diagnostic probability Pcf, after dimensionless processing, as follows:

[0083]

[0084] In the formula, It is represented as a positive constant.

[0085] In this embodiment, through the synergistic effect of simulated reasoning and the second computing unit in the knowledge hypergraph construction module, the observed diagnostic probability Pobs and the control diagnostic probability Pcf are normalized and causal consistency is calculated. This can effectively assess the consistency of diagnostic conclusions in real and counterfactual scenarios, thereby improving the causal credibility and clinical interpretability of disease diagnosis reasoning results.

[0086] Example 6

[0087] This embodiment is an explanation based on Embodiment 5. Please refer to it. Figure 1 Specifically, the second analysis unit is used to preset the causal consistency threshold Cth, and compare the causal consistency coefficient CIC with the causal consistency threshold Cth to obtain the second evaluation result, including:

[0088] When the causal consistency coefficient CIC ≥ causal consistency threshold Cth, it indicates that the causal consistency between the observation diagnosis and the counterfactual comparison results is qualified, the reasoning result is credible, and it enters the regular knowledge graph reasoning process and transmits the causal chain result to the reasoning game module.

[0089] When the causal consistency coefficient CIC < the causal consistency threshold Cth, it indicates that the causal consistency between the observed diagnosis and the counterfactual control results is unqualified, the inference result is unreliable, triggering a second warning instruction and generating a second strategy: generating a test and differential diagnosis scheme, prompting the clinical auxiliary diagnostic system to further verify.

[0090] The causal consistency threshold Cth is obtained by statistically analyzing a large number of multi-source observation samples and experimental data to extract the distribution range of causal consistency coefficients under different variable influence strengths, conditional dependencies, and temporal characteristics. Combined with the experience judgment of causal reasoning experts on the logical rationality between variables and the causal direction, a reasonable consistency threshold is determined. Referring to the technical specifications for causal reasoning and the verification results of existing cases, this threshold is used to effectively distinguish whether causal relationships have stable and credible explanatory power, ensuring the scientific nature of causal reasoning and model construction.

[0091] In this embodiment, the second analysis unit sets a causal consistency threshold Cth and judges the causal consistency coefficient CIC to achieve causal credibility classification of diagnostic reasoning results. When the consistency is insufficient, it can automatically trigger an early warning and generate differentiated diagnosis and testing plans, thereby improving the reliability of the diagnostic process and the ability to prevent and control clinical risks.

[0092] Example 7

[0093] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the reasoning game module includes a game negotiation unit, a third calculation unit, and a third analysis unit;

[0094] The game negotiation unit is used to analyze the diagnostic recommendation results of clinicians, imaging experts, laboratory experts and guideline experts based on the causal chain results and through similarity measurement methods to obtain the similarity Rxs between the agent and the group consensus;

[0095] The third computing unit is used to calculate the game consensus coefficient GCI by obtaining the similarity Rxs between the agent and the group consensus, after dimensionless processing, as follows:

[0096]

[0097] In the formula, N represents the number of agents. This represents the similarity between the result of the i-th agent and the group consensus. The weight coefficients, or Game Consistency Coefficient (GCI), are used to represent the game consistency coefficients. GCI can also be derived from the knowledge graph connectivity (KG). connect Consistency between authoritative database disease records (DB) and phenotypic matching degree (P) match Gene matching degree G match Literature support (Lit) and the weights of each piece of evidence (W) evid The formula is constructed by weighted summation, and the weights are determined through multi-objective optimization to maximize diagnostic accuracy and interpretability.

[0098] The acquisition method is as follows: the diagnostic credibility weight of the i-th agent is used to measure the relative importance of the agent's recommendation result in the group consensus; the credibility weight of different agents is determined by statistical analysis of the accuracy, stability and consistency with expert opinions of the agents in historical tasks, and normalization processing is performed in combination with the model training results.

[0099] Weighting coefficient The introduction of GCI enables GCI to highlight the influence of high-credibility agents and suppress the interference of noisy agents when calculating group consensus, thereby improving the reliability and scientific nature of game consensus determination.

[0100] In this embodiment, a game negotiation and similarity measurement method is introduced through a reasoning game module to quantify the matching degree between multi-agent diagnostic recommendations and group consensus, and to calculate the game consistency coefficient (GCI), thereby realizing the fusion of expert opinions and the identification of disagreements, effectively improving the collective credibility and clinical application value of diagnostic conclusions.

[0101] Example 8

[0102] This embodiment is an explanation based on Embodiment 7. Please refer to it. Figure 1 Specifically, the third calculation unit is used to preset the game consistency threshold Gth, and compare and analyze the game consistency coefficient GCI with the game consistency threshold Gth to obtain the third evaluation result, including:

[0103] When the game consistency coefficient GCI ≥ game consistency threshold Gth, it means that there is no disagreement among the experts in their diagnosis, the diagnosis is credible, and the process of unified knowledge reasoning and decision output begins.

[0104] When the game consistency coefficient GCI < game consistency threshold Gth, it indicates that there is a disagreement among experts in their diagnostic results, and the diagnostic conclusion is unreliable. This triggers a third warning instruction and generates a third strategy: outputting expert disagreement prompts and generating differentiated diagnostic schemes and additional testing suggestions for further confirmation by the clinical auxiliary diagnostic system.

[0105] The consistency threshold Gth is obtained by statistically analyzing the evolution of player strategies and payoff feedback data in a large number of game scenarios, extracting the distribution range of game consistency coefficients under different game types, strategy combinations, and environmental constraints, and combining the empirical judgments of game theorists on the rationality of equilibrium states and the convergence of strategies to determine a reasonable consistency threshold. Referring to game analysis standards and historical game experiment data, this threshold is used to effectively distinguish whether the player strategies have reached a stable equilibrium, thereby ensuring the reliability and predictability of game results.

[0106] In this embodiment, by setting a game consensus threshold Gth and comparing it with the game consensus coefficient GCI, a reliable diagnostic conclusion can be quickly output when experts agree, and an early warning can be automatically triggered and differentiated diagnostic plans and additional test suggestions can be generated when there is a disagreement, thereby ensuring the reliability of the diagnostic results and the robustness of clinical decision-making.

[0107] Example 9

[0108] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the diagnostic traceability security verification module includes a traceability analysis unit and a security verification unit;

[0109] The source tracing analysis unit is used to perform difference analysis between diagnostic results and actual follow-up records or historical case data based on the time-series diagnostic sensitivity coefficient (SDC), causal consistency coefficient (CIC), and game consistency coefficient (GCI). When the diagnostic results are found to be inconsistent with the follow-up results or there is an abnormal deviation, the unit automatically identifies the relevant knowledge fragments, reasoning steps, and causal chains that are abnormal and marks and locates the abnormal nodes.

[0110] The security verification unit is used to perform comprehensive verification of the diagnostic results by using hierarchical verification methods at the individual, cross-modal, and guideline levels through the located abnormal nodes, so that the inference output conforms to the medical safety boundary.

[0111] This unit assesses the reliability of diagnostic results at different verification levels based on SDC, CIC, and game-theoretic consistency information. When potential risks or result deviations are detected, a source tracing and remediation process is automatically triggered, generating security alerts and correcting the relevant knowledge fragments or reasoning steps, thereby ensuring that the final diagnostic results are both consistent with clinical data and have undergone security verification.

[0112] In this embodiment, the diagnostic traceability and security verification module enables multi-dimensional difference analysis and hierarchical verification. When abnormalities or deviations occur in the diagnostic results, the source of the problem can be automatically located and the repair process can be triggered. This ensures that the final diagnostic conclusion is consistent with both clinical data and medical safety boundaries, significantly improving the credibility and security of the diagnostic results.

[0113] Example 10

[0114] This embodiment is an explanation based on Embodiment 1. Please refer to it. Figure 1 Specifically, the knowledge evolution and update module is used to perform structured parsing and knowledge extraction on the log data generated during the diagnosis, verification and tracing process, and form knowledge update candidates. Combined with the human expert review mechanism, the data is fed back to the knowledge layer and the diagnostic model to achieve continuous knowledge evolution.

[0115] In this embodiment, the log data is structured and parsed and knowledge is extracted through the knowledge evolution update module. Under the review mechanism of human experts, the knowledge is fed back to the knowledge layer and the diagnostic model, which can realize the dynamic supplementation and continuous optimization of diagnostic knowledge, so that the system can continuously improve the accuracy and adaptability of identifying complex diseases in long-term operation.

[0116] The steps for building the model are as follows, please refer to them. Figure 2 :

[0117] Input layer and data base:

[0118] New patient case entry: Obtain patient's chief complaint and physical examination description, laboratory test results (blood, urine, biochemistry, etc.), imaging / pathology reports, genetic / genetic locus information, and current medication history from electronic medical records and follow-up terminals. This forms a standardized evidence package (Ep).

[0119] Previous patient cohort: Access the historical case database, which includes "phenotype-disease" pairings, distribution of test results, typical / atypical manifestations, and follow-up outcomes of confirmed cases, for use in case similarity retrieval and prior statistics.

[0120] Knowledge Graph / Knowledge Hypergraph: Aggregates authoritative knowledge sources (guidelines, databases such as MedGen / Orphanet / OMIM, ontology and terminology system), with nodes covering relationships such as "disease-phenotype-gene-pathway-drug", used for reasoning and counterfactual simulation.

[0121] S1. Initial diagnosis of the large candidate generation model:

[0122] Driven by both knowledge and data, the system's large model infers from the input evidence package Ep and outputs:

[0123] Candidate diagnosis list {d1,…,dk} and its observed diagnosis probability Pobs(di);

[0124] Key evidence chains corresponding to each candidate (supporting / refuting evidence, tracing the source);

[0125] Recommendations for supplemental evidence for differential diagnosis (such as additional gene, imaging sequence, or specific laboratory tests required).

[0126] The above results are recorded as "Diagnosis A" (which can be Top-1 or Top-N) and proceed to the verification stage.

[0127] S2, Multi-source verification stack (interpretable evidence determination)

[0128] For diagnosis A, automated verification and quantitative integration were carried out from four dimensions of evidence:

[0129] 1. [gene] Evidence: Connectivity of patient variant sites to the disease in the map, pathogenicity grade, familial segregation, and corresponding gene matching degree G. match ;

[0130] 2. [lit] Literature Evidence: Authoritative literature, case series, and reviews matching case characteristics; extracting supporting / counterexamples; corresponding literature support (Lit);

[0131] 3. [db] Database evidence: This includes entries and evidence levels from authoritative databases such as MedGen / Orphanet / ClinVar; and the consistency of disease records with the corresponding authoritative database.

[0132] 4. [Phenotype] Evidence: Coverage and conflict between the patient's actual phenotype and the phenotypes in the knowledge inventory, combined with the distribution of similar historical cases, corresponding to the phenotype match degree P. match ;

[0133] 5. Spectral graph data, corresponding to the connectivity KG of the knowledge graph. connect ;

[0134] S3. To avoid subjective fabrication, a weight vector W=[wgene,wlit,wdb,wpheno,wzhi] is introduced to weight and fuse the four types of evidence, resulting in:

[0135] Evidence consistency coefficient ECI = f(W;gene,lit,db,pheno);

[0136] Counterfactual simulations were conducted under the structured causal model (SCM) to calculate the causal consistency coefficient CIC=g(Pobs,Pcf).

[0137] Verification and judgment:

[0138] When ECI≥Eth and CIC≥Cth, the verification is successful, and diagnosis A and its evidence basis are output.

[0139] If any threshold is not reached, the verification fails, and the "knowledge × data dual-drive + expert consultation" process is automatically initiated.

[0140] S4, Multi-Agent Diagnosis and Reasoning Game

[0141] When validation is insufficient or a high-confidence conclusion is required, the system enters the multi-agent reasoning game module. This module consists of role-based agents {Ai} such as clinicians, imaging experts, laboratory experts, and guideline / pharmacy experts, conducting structured argumentation around diagnosis A and several alternative diagnoses.

[0142] 1) Argumentation rounds: Conduct round-based collaboration according to the "claim - support - refute - convergence" approach, and cite knowledge bases / databases / literature fragments as evidence;

[0143] 2) Similarity and consensus: Based on case similarity retrieval and evidence quality, calculate the similarity Rxs(Ai) between individual claims and group consensus;

[0144] 3) Game consistency coefficient: Aggregate multiple claims using methods such as weighted voting, Borda scoring, or D–S evidence synthesis to obtain GCI;

[0145] 4) Disagreement handling: When GCI < Gth, locate the key disagreement points and the minimum controversial evidence set, and generate a differential diagnosis and supplementary collection list.

[0146] S5, Final judgment and output

[0147] When ECI ≥ Eth, CIC ≥ Cth, and GCI ≥ Gth, the system outputs the final diagnosis conclusion (single or graded candidate), and gives:

[0148] a) Key causal chain and evidence source;

[0149] b) Diagnosis and treatment suggestions / follow-up points and potential risk warnings;

[0150] c) Explanation of uncertainty (if there are concurrent conclusions).

[0151] If the threshold is still not met, output a disagreement report and review / supplementary inspection suggestions, and mark it as a key follow-up case.

[0152] S6, Traceability and security verification;

[0153] The system audits and traces the complete reasoning and consultation process:

[0154] Automatically locate abnormal knowledge fragments, conflicting evidence, and vulnerable reasoning steps;

[0155] Initiate multi-level security checks (hard constraints of rules / guidelines, counterfactual re-simulation, adversarial interrogation), and give an alarm when the risk threshold is triggered and require strong evidence pairs (positive + exclusive evidence) before releasing;

[0156] Generate traceable logs to meet medical compliance and regulatory reviews.

[0157] S7, Knowledge evolution and closed-loop update

[0158] Form structured logs and argumentation graphs from initial diagnosis, verification, consultation, review, and follow-up feedback; automatically extract new evidence and counterexamples to form knowledge update candidates, which are fed back to the knowledge hypergraph and model parameters after expert review; continuously adaptively update the weights W, thresholds E th / C th / G th and aggregation strategies to achieve long-term iterative optimization of system performance.

[0159] Appendix: Explanation of symbols and terms is shown in Table 1 below.

[0160] Table 1: Explanation of Symbols and Terms

[0161]

[0162] The threshold is set to facilitate comparison. The size of the threshold depends on the amount of sample data and the number of bases set by those skilled in the art for each set of sample data; as long as it does not affect the ratio between the parameter and the quantized value, it is acceptable.

[0163] The above formulas are all derived from software simulation using a large amount of data, and are selected to be close to the actual values. The coefficients in the formulas are set by those skilled in the art based on the actual situation. The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or changes made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A multi-agent large-scale model disease diagnosis knowledge reasoning system based on data-driven dual-engineering, characterized in that: include: The data acquisition module is used to collect patient information from multiple sources, including patient complaints and follow-up records, physical examinations, multimodal clinical data from imaging, laboratory tests, and genes, observational diagnostic probability data, and diagnostic suggestions and interpretative evidence from various expert agents in the multi-agent platform; "dual drive" refers to the synergistic drive of real-world clinical data and medical knowledge base / guidelines / case database; The representation layer modeling module is used to construct patient representation vectors and evidence saliency distributions based on multimodal data, and generate interpretable labels for key evidence items; without relying on temporal features, it completes the extraction of associations and conflict resolution between symptoms, signs and tests, forming standardized cases and evidence for reasoning. The knowledge hypergraph construction module is used to construct a knowledge hypergraph of disease-evidence-treatment triples based on clinical data and knowledge base, and to perform counterfactual simulation under the structured causal model (SCM). It obtains the observed diagnosis probability (Pobs) and the control diagnosis probability (Pcf) through statistical and counterfactual reasoning, calculates the causal consistency coefficient (CIC), and compares it with the causal consistency threshold (Cth) to determine whether the causal consistency between the observed diagnosis and the counterfactual control results is qualified. If qualified, the causal chain results are passed to the reasoning game module; if not qualified, an appropriate strategy is given. The reasoning and game theory module sets up a group of role-based experts. Each agent provides a diagnostic claim, a set of supporting evidence, and counterexamples. It constructs an argument graph of claim-support-rebuttal and executes a multi-round collaborative process of "proposal-questioning-defense-convergence". It allows the use of external authoritative knowledge fragments for corroboration. It obtains the similarity Rxs of the group consensus by calculating the information entropy threshold, calculates the game consistency coefficient GCI, and compares it with the game consistency threshold Gth to determine whether the diagnostic conclusion is credible. If it is not credible, it gives an appropriate strategy. The diagnostic traceability safety verification module, based on the causal consistency coefficient (CIC) and game-theoretic consistency coefficient (GCI), performs difference analysis between diagnostic results and follow-up or historical data, and automatically locates abnormal knowledge fragments and reasoning steps; the safety verification unit uses multi-level verification for abnormal nodes to check the reliability of diagnostic results, and triggers repair and generates safety warnings when risks are detected, so that the final diagnostic results are both consistent with clinical data and have passed safety verification. The knowledge evolution and update module performs structured parsing and knowledge extraction on the log data generated during the diagnosis and verification process, generates update candidates, and after expert review, feeds them back to the knowledge layer and diagnostic model to achieve continuous iteration and optimization of system knowledge.

2. The data-driven, dual-agent, large-scale model disease diagnosis knowledge reasoning system according to claim 1, characterized in that, The data acquisition module is used to collect multimodal information such as patient complaints and follow-up records, physical examinations, and imaging / laboratory / pathological / genetic data through the electronic medical record system and follow-up terminal. Continuous results of patients' blood, urine, and biochemical test indicators are collected using laboratory testing equipment. By accessing the diagnostic knowledge hypergraph and historical case database, observational diagnostic probability data is collected; and through a virtual multi-agent interaction platform, diagnostic recommendation results from clinicians, imaging experts, laboratory experts, and guideline experts are collected.

3. The data-driven, dual-agent, large-scale disease diagnosis knowledge reasoning system based on a multi-agent model, as described in claim 1, is characterized in that... The representation layer modeling module includes a temporal feature unit, a first calculation unit, and a first analysis unit; The temporal feature unit is used to calculate the time change rate Tsym of the symptom evolution trajectory by using temporal modeling and differential algorithms on the patient symptom follow-up records. By using dynamic statistical analysis and fluctuation detection methods, the laboratory test result sequence is processed to obtain the dynamic fluctuation value Vlab of key test indicators; The first calculation unit is used to calculate and obtain the time-series diagnostic sensitivity coefficient SDC by using the time change rate Tsym of the acquired symptom evolution trajectory and the dynamic fluctuation value Vlab of the key test index, after dimensionless processing. It is also used to structure and semantically align clinical data from different sources and formats to form patient representation vectors and evidence packages; The first calculation unit is used to calculate the evidence significance weight and evidence consistency coefficient (ECT) based on the matching degree of evidence quality, source reliability and disease prior, and generate interpretable labels.

4. The data-driven, dual-agent, large-scale model disease diagnosis knowledge reasoning system according to claim 3, characterized in that, The first analysis unit is used to preset a sensitivity threshold Sth, and compare and analyze the time-series diagnostic sensitivity coefficient SDC with the sensitivity threshold Sth to obtain the first evaluation result, including: When the time-series diagnostic sensitivity coefficient SDC < sensitivity threshold Sth, it indicates that the current patient does not show an abnormal evolution trend and enters a routine follow-up monitoring state, where multimodal clinical data are collected regularly and a time-series atlas is established. When the sequential diagnostic sensitivity coefficient SDC is greater than or equal to the sensitivity threshold Sth, the current patient shows an abnormal evolution trend and has an early disease risk, triggering the first warning instruction and generating the first strategy: it is recommended to prioritize key laboratory tests and imaging re-examinations, and focus on monitoring abnormal indicators; and mark the patient as a "key individual" in the clinical auxiliary diagnostic system.

5. The data-driven, dual-agent, large-scale model disease diagnosis knowledge reasoning system according to claim 1, characterized in that, The knowledge hypergraph construction module includes a simulation reasoning unit, a second calculation unit, and a second analysis unit; The simulation reasoning unit is used to couple the established time series map with clinical data, and combine the disease, phenotype, gene, molecular component, physiological process, pathway ontology knowledge and causal relationship of the observed diagnostic probability data to perform counterfactual simulation; through statistical reasoning methods, the results of the knowledge hypergraph and historical cases are normalized to obtain the observed diagnostic probability Pobs; through the counterfactual reasoning algorithm, the results in the control scenario are calculated to obtain the control diagnostic probability Pcf. The second calculation unit is used to calculate the causal consistency coefficient CIC by obtaining the observed diagnostic probability Pobs and the control diagnostic probability Pcf, after dimensionless processing.

6. The data-driven, dual-agent, large-scale model disease diagnosis knowledge reasoning system according to claim 5, characterized in that, The second analysis unit is used to preset the causal consistency threshold Cth, and compare the causal consistency coefficient CIC with the causal consistency threshold Cth to obtain the second evaluation result, including: When the causal consistency coefficient CIC ≥ causal consistency threshold Cth, it indicates that the causal consistency between the observation diagnosis and the counterfactual comparison results is qualified, the reasoning result is credible, and it enters the regular knowledge graph reasoning process and transmits the causal chain result to the reasoning game module. When the causal consistency coefficient CIC < the causal consistency threshold Cth, it indicates that the causal consistency between the observed diagnosis and the counterfactual control results is unqualified, the inference result is unreliable, triggering a second warning instruction and generating a second strategy: generating a test and differential diagnosis scheme, prompting the clinical auxiliary diagnostic system to further verify.

7. The data-driven, dual-agent, large-scale model disease diagnosis knowledge reasoning system according to claim 1, characterized in that, The reasoning game module includes a game negotiation unit, a third calculation unit, and a third analysis unit; The game negotiation unit is used to analyze the diagnostic recommendation results of clinicians, imaging experts, laboratory experts and guideline experts based on the causal chain results and through similarity measurement methods to obtain the similarity Rxs between the agent and the group consensus; The third computing unit is used to calculate the game consensus coefficient GCI by obtaining the similarity Rxs between the agent and the group consensus, after dimensionless processing; GCI is mainly composed of the knowledge graph connectivity KG. connect Consistency between authoritative database disease records (DB) and phenotypic matching degree (P) match Gene matching degree G match Literature support (Lit) and the weights of each piece of evidence (W) evid The system is structured and weights are determined through multi-objective optimization to maximize diagnostic accuracy and interpretability.

8. The data-driven, dual-agent, large-scale model disease diagnosis knowledge reasoning system according to claim 7, characterized in that, The third calculation unit is used to preset the game consistency threshold Gth, and compare and analyze the game consistency coefficient GCI with the game consistency threshold Gth to obtain the third evaluation result, including: When the game consistency coefficient GCI ≥ game consistency threshold Gth, it means that there is no disagreement among the experts in their diagnostic results, the diagnostic conclusion is credible, and the process of unified knowledge reasoning and decision output begins. When the game consistency coefficient GCI < game consistency threshold Gth, it indicates that there is a disagreement among experts in their diagnostic results, and the diagnostic conclusion is unreliable. This triggers a third warning instruction and generates a third strategy: outputting expert disagreement prompts and generating differentiated diagnostic schemes and additional testing suggestions for further confirmation by the clinical auxiliary diagnostic system.

9. The data-driven, dual-agent, large-scale model disease diagnosis knowledge reasoning system according to claim 1, characterized in that, The diagnostic traceability and security verification module includes a traceability analysis unit and a security verification unit; The source tracing analysis unit is used to perform difference analysis between diagnostic results and actual follow-up records or historical case data based on the time-series diagnostic sensitivity coefficient (SDC), causal consistency coefficient (CIC), and game consistency coefficient (GCI). When the diagnostic results are found to be inconsistent with the follow-up results or there is an abnormal deviation, the unit automatically identifies the relevant knowledge fragments, reasoning steps, and causal chains that are abnormal and marks and locates the abnormal nodes. The security verification unit is used to perform comprehensive verification of the diagnostic results by using hierarchical verification methods at the individual, cross-modal, and guideline levels through the located abnormal nodes, so that the inference output conforms to the medical safety boundary.

10. The data-driven, dual-agent, large-scale model disease diagnosis knowledge reasoning system according to claim 1, characterized in that, The knowledge evolution and update module is used to perform structured parsing and knowledge extraction on the log data generated during diagnosis, verification and tracing, and form knowledge update candidates. Combined with the human expert review mechanism, the data is fed back to the knowledge layer and diagnostic model to achieve continuous knowledge evolution.

Images