CN-121583511-B - Multi-agent large model disease diagnosis knowledge reasoning system based on data dual driving

CN121583511BCN 121583511 BCN121583511 BCN 121583511BCN-121583511-B

Abstract

The invention discloses a multi-agent large model disease diagnosis knowledge reasoning system based on data dual driving, which relates to the technical field of artificial intelligence auxiliary medical diagnosis, and the system extracts time sequence evolution characteristics and calculates time sequence diagnosis sensitivity coefficients through multi-source acquisition of patient symptom follow-up records, laboratory test results, observation diagnosis probability and expert diagnosis recommendation results, so as to realize early identification of disease risks; the method and the system combine the anti-facts simulation and the statistical reasoning to obtain a causal consistency coefficient for verifying causal rationality of observation diagnosis and comparison results, calculate a game consistency coefficient based on the intelligent body group consensus analysis for judging the credibility of diagnosis conclusion, locate and multi-level verify abnormal reasoning steps and knowledge fragments to ensure the reliability and safety of the results, realize continuous optimization of a diagnosis model through a log analysis and knowledge backflow mechanism, and remarkably improve the accuracy, the interpretability and the safety of disease diagnosis.

Inventors

GUO GUILING
LI QIYUAN
WU JINZHUN
ZHOU YULIN
SHAO ZHENYU
Luo Heying

Assignees

厦门大学
厦门市妇幼保健院(厦门市优生优育服务中心、厦门大学附属妇女儿童医院、厦门市林巧稚妇女儿童医院)

Dates

Publication Date: 20260512
Application Date: 20260126

Claims (10)

1. The utility model provides a multi-agent large model disease diagnosis knowledge reasoning system based on data dual drive which characterized in that includes: The data acquisition module is used for acquiring patient information in a multi-source manner, and comprises patient complaints and follow-up records, physical examination, image/inspection/gene multi-mode clinical data, observation diagnosis probability data and diagnosis suggestions and explanatory basis of various expert intelligent agents in the multi-intelligent agent platform, wherein 'dual driving' refers to the cooperative driving of clinical real world data and a medical knowledge base/guide/case base; The system comprises a representation layer modeling module, a data processing module and a data processing module, wherein the representation layer modeling module is used for constructing a patient characterization vector and evidence significance distribution based on multi-mode data and generating an interpretable label for key evidence items; The knowledge hypergraph construction module is used for constructing a knowledge hypergraph of a disease-evidence-treatment triplet based on clinical data and a knowledge base, executing anti-fact simulation under a structured causal model SCM, acquiring observation diagnosis probability Pobs and control diagnosis probability Pcf through statistics and anti-fact reasoning, calculating and acquiring causal consistency coefficient CIC, carrying out comparative analysis with a causal consistency threshold Cth, judging whether causal consistency of observation diagnosis and anti-fact control results is qualified or not, transmitting causal chain results to the reasoning game module if the causal consistency coefficient CIC is qualified, and giving corresponding strategies if the causal consistency coefficient CIC is not qualified; The reasoning game module is used for setting a role expert group, each agent gives out diagnosis claims, evidence support sets and counterexamples, constructing a arguments graph of claims, supports and counterexamples, executing multi-round collaboration of 'giving-inquiring-dialectical protection and convergence', allowing to call an external authoritative knowledge segment for evidence, acquiring similarity Rxs of group consensus through calculating an information entropy threshold value, calculating and acquiring a game consistency coefficient GCI, carrying out comparison analysis with a game consistency threshold value Gth, judging whether a diagnosis conclusion is credible, and giving a corresponding policy if the diagnosis conclusion is not credible; the diagnosis tracing safety verification module is used for carrying out difference analysis on a diagnosis result and follow-up or historical data based on a causal consistency coefficient CIC and a game consistency coefficient GCI and automatically positioning abnormal knowledge fragments and reasoning steps; and the knowledge evolution updating module performs structural analysis and knowledge extraction on log data generated in the diagnosis and verification process to generate updating candidates, and the updating candidates are returned to the knowledge layer and the diagnosis model after being checked by an expert, so that continuous iteration and optimization of system knowledge are realized.
2. The multi-agent large model disease diagnosis knowledge reasoning system based on data dual driving according to claim 1, wherein the data acquisition module is used for acquiring multi-mode information of patient complaints and follow-up records, physical examination, images/examination/pathology/genes through the electronic medical record system and the follow-up terminal, acquiring continuous results of blood, urine and biochemical examination indexes of a patient through laboratory examination equipment, acquiring observation diagnosis probability data through calling a diagnosis knowledge hypergraph and a historical case database, and acquiring diagnosis recommendation results of clinicians, image experts, examination experts and guideline experts through the virtual multi-agent interaction platform.
3. The data dual-drive-based multi-agent large model disease diagnosis knowledge reasoning system of claim 1, wherein the representation layer modeling module comprises a timing sequence feature unit, a first calculation unit and a first analysis unit; the time sequence characteristic unit is used for calculating the patient symptom follow-up record through a time sequence modeling and difference algorithm to obtain the time change rate Tsym of a symptom evolution track, and processing a laboratory test result sequence through a dynamic statistical analysis and fluctuation detection method to obtain a dynamic fluctuation value Vlab of a key test index; the first calculating unit is used for calculating and obtaining a time sequence diagnosis sensitivity coefficient SDC after dimensionless processing through the time change rate Tsym of the obtained symptom evolution track and the dynamic fluctuation value Vlab of the key inspection index; the first computing unit is used for computing and acquiring evidence significance weights and evidence consistency coefficients ECT based on the matching degree of evidence quality, source reliability and disease type priori, and generating interpretable labels.
4. The system for reasoning knowledge about disease diagnosis of a multi-agent large model based on data dual driving of claim 3, wherein the first analysis unit is configured to preset a sensitivity threshold Sth, compare and analyze a time sequence diagnosis sensitivity coefficient SDC with the sensitivity threshold Sth, and obtain a first evaluation result includes: When the sensitivity coefficient SDC of the time sequence diagnosis is less than the sensitivity threshold Sth, the current patient does not show abnormal evolution trend, the current patient enters a conventional follow-up monitoring state, and multi-mode clinical data are collected regularly and a time sequence map is established; When the sensitivity coefficient SDC of the time sequence diagnosis is more than or equal to the sensitivity threshold Sth, the current patient presents an abnormal evolution trend, early disease risks exist, a first early warning instruction is triggered, a first strategy is generated, the key laboratory inspection and the imaging review are preferably arranged, the abnormal index is monitored in a key way, and the key laboratory inspection and the imaging review are marked as 'focus on individuals' in a clinical auxiliary diagnosis system.
5. The knowledge reasoning system for diagnosing the disease of the multi-agent large model based on the data dual driving of claim 1, wherein the knowledge hypergraph construction module comprises a simulation reasoning unit, a second calculation unit and a second analysis unit; The simulation reasoning unit is used for coupling with clinical data based on the established time sequence map, and executing inverse fact simulation by combining disease, phenotype, gene, molecular composition, physiological process, path ontology knowledge and causal relation of observation and diagnosis probability data, normalizing the results of the knowledge hypergraph and the historical cases through a statistical reasoning method to obtain observation and diagnosis probability Pobs; The second calculation unit is configured to calculate and obtain a causal consistency coefficient CIC after dimensionless processing according to the obtained observed diagnosis probability Pobs and the comparison diagnosis probability Pcf.
6. The system for reasoning knowledge about diagnosis of a multi-agent large model disease based on data dual driving of claim 5, wherein the second analysis unit is configured to preset a causal consistency threshold Cth, and perform a comparative analysis on a causal consistency coefficient CIC and the causal consistency threshold Cth, and obtain a second evaluation result includes: When the cause and effect consistency coefficient CIC is more than or equal to a cause and effect consistency threshold Cth, the cause and effect consistency of observation diagnosis and a counter fact comparison result is qualified, the reasoning result is reliable, a conventional knowledge graph reasoning flow is entered, and a cause and effect chain result is transmitted to a reasoning game module; When the causal consistency coefficient CIC is smaller than the causal consistency threshold Cth, the causal consistency of the observation diagnosis and the anti-facts contrast result is not qualified, the reasoning result is not credible, a second early warning instruction is triggered, a second strategy is generated, and a checking and difference diagnosis scheme is generated to prompt a clinical auxiliary diagnosis system to further verify.
7. The data dual-drive-based multi-agent large model disease diagnosis knowledge reasoning system of claim 1, wherein the reasoning game module comprises a game negotiation unit, a third calculation unit and a third analysis unit; the game negotiation unit is used for analyzing diagnosis recommendation results of clinicians, image experts, inspection experts and guide experts through a similarity measurement method based on a causal chain result to obtain similarity Rxs of the intelligent agent and group consensus; The third calculation unit is used for calculating and obtaining a game consistency coefficient GCI after dimensionless processing through the obtained similarity Rxs of the intelligent agent and the community consensus, wherein the GCI mainly comprises knowledge graph connectivity KG connect , authority database disease record consistency DB, phenotype matching degree P match , gene matching degree G match , document support degree Lit and each evidence quality weight W evid , and the weights are determined through multi-objective optimization to maximize diagnosis accuracy and interpretability.
8. The system for reasoning the diagnosis knowledge of the multi-agent large model disease based on the data dual driving of claim 7, wherein the third calculation unit is configured to preset a game consistency threshold value Gth, and perform a comparative analysis on the game consistency coefficient GCI and the game consistency threshold value Gth, and obtain a third evaluation result includes: When the game consistency coefficient GCI is more than or equal to the game consistency threshold value Gth, the diagnosis results among the experts are not separated, the diagnosis conclusion is reliable, and a unified knowledge reasoning and decision output flow is entered; When the game consistency coefficient GCI is smaller than the game consistency threshold value Gth, the fact that the diagnosis results among the experts are divergent is indicated, the diagnosis conclusion is not reliable, a third early warning instruction is triggered, a third strategy is generated, namely, an expert divergence prompt is output, and a differential diagnosis scheme and an additional inspection suggestion are generated for further confirmation of a clinical auxiliary diagnosis system.
9. The multi-agent large model disease diagnosis knowledge reasoning system based on data dual driving of claim 1, wherein the diagnosis traceability safety verification module comprises a traceability analysis unit and a safety verification unit; The tracing analysis unit is used for carrying out difference analysis on the diagnosis result and the actual follow-up record or the historical case data based on the time sequence diagnosis sensitivity coefficient SDC, the causal consistency coefficient CIC and the game consistency coefficient GCI; The security verification unit is used for carrying out omnibearing verification on the diagnosis result by adopting a layering verification method of individual dimension, cross-mode and guide level through the positioned abnormal node, so that the reasoning output accords with the medical security boundary.
10. The knowledge reasoning system for diagnosing the disease of the multi-agent large model based on the data dual driving of claim 1, wherein the knowledge evolution updating module is used for carrying out structural analysis and knowledge extraction on log data generated in the diagnosis, verification and tracing processes, forming knowledge updating candidates, and realizing the continuous evolution of knowledge by combining a manual expert rechecking mechanism, returning to a knowledge layer and a diagnosis model.

Description

Multi-agent large model disease diagnosis knowledge reasoning system based on data dual driving Technical Field The invention relates to the technical field of artificial intelligence auxiliary medical diagnosis, in particular to a disease diagnosis knowledge reasoning system based on dual driving of real world clinical data and a medical knowledge base, fusion of a large multi-agent model and causal reasoning and reasoning game/multi-expert consultation mechanism. Background Clinical diagnosis generally relies on patient complaints and medical history, physical examination, laboratory tests, imaging, pathology, and other multimodal information to make comprehensive decisions. With the development of medical informatization and artificial intelligence technology, computer-aided diagnosis (CAD) systems based on rule bases, statistical learning or single model inference have been applied in some specialized departments. However, in real-world complex cases and interdisciplinary collaboration scenarios, there are still limitations that are difficult to overcome in the prior art, particularly in: Multimodal data fragmentation is not unified with semantics. Clinical data are scattered in heterogeneous platforms such as electronic medical records, inspection/examination systems, follow-up visit and nursing records, and the like, and the data structure, the coding system and the quality are uneven, so that a unified semantic alignment and standardization integration mechanism is lacked. Therefore, the evidence is difficult to compare and fuse in the same representation space, key clues are easy to miss or repeat, and the integrity and consistency of diagnosis are affected. A single expert relies on and lacks a mechanism for quantitative consistency. The existing clinical decision mainly depends on the experience of an individual doctor and limited guideline rules, and the conclusion difference among different specialists is obvious for complicated, rare or concurrent cases. The system multi-expert consultation organization and the demonstration flow of 'claim-support-refute-convergence' are lacking, the quantitative index and the threshold for converting the multi-opinion into the group consensus are further lacking, and the stability and the credibility of the conclusion are difficult to objectively evaluate. The causal reasoning capability is insufficient, and the anti-facts analyze the deficiency. Most systems are mainly modeled by correlation, and the statistical correlation between the characteristics and the disease labels is emphasized, so that the structural causal representation and inspection of pathophysiological mechanisms and diagnosis and treatment paths are lacked. Lacking the ability to conduct counterfactual simulations under causal graphs or Structured Causal Models (SCMs), it is difficult to evaluate the robustness and portability of diagnostic conclusions when the evidence changes/interventions are different. Security verification and traceability are weak. The traceability of the reasoning links and the evidence links of the existing system is limited, and an automatic positioning and multi-layer checking mechanism for abnormal knowledge fragments, conflict evidence and high-influence nodes is lacked. It is difficult to trigger alarm and repair processes in time for high risk conclusions, and it is difficult to meet the strict requirements of medical scenes on reliability, interpretability and compliance. Knowledge update lags and lacks closed loop evolution. The diagnosis knowledge base and the model are manually maintained and updated off-line, so that new evidence from clinical logs, consultation processes and verification feedback and counterexample bases are difficult to absorb in time. The lack of automated channels to structurally precipitate and reflow the practical experience to the knowledge and model layers makes continuous optimization and adaptive evolution difficult to achieve for the system. The human-computer cooperation and treatment framework is imperfect. In the complex diagnosis flow with large models and multiple intelligent agents involved, the management elements such as role responsibilities, authority boundaries, external knowledge reference compliance, responsibility attribution and the like lack of clear specifications, so that the large-scale landing and supervision examination passing rate of the system in a real medical institution is limited. In summary, there is an urgent need for an intelligent diagnosis system that combines multi-agent consultation with game reasoning, causal reasoning and anti-facts simulation, and has traceable security verification and knowledge closed-loop evolution capability under a unified data and knowledge framework, so as to improve accuracy, robustness and auditability in complex disease scenes. Disclosure of Invention Aiming at the defects of the prior art, the invention provides a multi-agent large model disease diagnosis knowledg