Medical information processing method and system based on multi-agent confrontation debate

Abstract

The invention discloses a medical information processing method and system based on multi-agent confrontation debate. According to the method, reasoning, confrontation, judgment and risk control agents are instantiated in a heterogeneous execution environment. The reasoning agent generates an initial hypothesis; an adversarial agent generates an anti-fact challenge vector by using a feature vector mask or inversion, and initiates logic questions; and the system executes multiple rounds of iterative debate and hypothesis correction, and the risk control agent blocks illegal contents in real time. And finally, calculating the path consistency by using the Kendall harmony coefficient and the semantic entropy, and outputting a result when a threshold value is met. According to the method, by simulating an adversarial game mechanism of expert consultation and introducing an anti-factual reasoning and semantic clustering technology, the problems of frequent illusion, misconfirmation, lack of interpretability and the like of a single model are effectively solved, and the accuracy and robustness of medical data analysis are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence, natural language processing, and medical data analysis, and in particular to a medical information processing method and system that improves the accuracy, robustness, and interpretability of medical text analysis by simulating the thinking process of multi-expert consultation (MDT) and utilizing adversarial interaction, counterfactual reasoning, and semantic entropy control among multiple agents. Background Technology

[0002] With the rapid development of deep learning technology, especially the emergence of Large Language Models (LLMs) based on the Transformer architecture, medical artificial intelligence is undergoing a profound paradigm shift from traditional "perceptual intelligence" (such as medical image recognition) to "cognitive intelligence" (such as Clinical Decision Support Systems, CDSS). However, when applying these advanced technologies to high-risk areas such as clinical decision support, existing technical architectures face the well-known "medical AI impossible triangle" challenge, namely, the difficulty in simultaneously achieving high accuracy, strong interpretability, and generalized robustness.

[0003] The existing technology suffers from the following core defects, which severely restrict its practical application in clinical settings:

[0004] 1. The Hallucination Hazard of Generative Models: A single generative model is essentially a probability-based next-token predictor. When dealing with long-tailed cases or complex complications outside the distribution of its training data, the model often relies on statistical correlation rather than causal logic, confidently generating seemingly reasonable but medically completely wrong recommendations. For example, to cater to a high-probability diagnosis (such as "myocardial infarction"), the model might fabricate symptoms never mentioned by the patient (such as "chest pain") or cite non-existent medical literature and clinical guidelines. This phenomenon, known as "serious nonsense," is fatal in the extremely low-tolerance medical field.

[0005] 2. Lack of logical consistency and transparent interpretability: Current end-to-end neural network models are often considered "black boxes." If a model provides a diagnostic suggestion, it is difficult for doctors to trace its specific reasoning path. This lack of interpretability makes it difficult for doctors to trust AI-assisted suggestions. Furthermore, single models are prone to "confirmation bias," meaning that once a false hypothesis is formed in the early stages of reasoning, subsequent reasoning processes tend to seek evidence supporting that false hypothesis, automatically ignoring or underestimating evidence that contradicts it.

[0006] 3. Mediocre Consensus and Hallucination in Multi-Model Fusion: To address the problems of single-model approaches, existing technologies attempt to introduce multi-model fusion, typically by running a rule-based model and a data model in parallel, and then drawing conclusions through weighted averaging or simple voting mechanisms. However, this fusion strategy has a deep flaw: if multiple underlying models produce the same erroneous judgment (homogenization error), a simple voting mechanism can actually reinforce this error, creating a hallucination of consensus.

[0007] 4. Rigidity in Dialogue Management: In real-world consultation scenarios, patients' statements are often fragmented, disjointed, and even contradictory. Traditional decision tree-based dialogue management systems are extremely rigid and unable to handle non-linear information input; while end-to-end LLM-based dialogue systems, although fluent, are prone to losing context in multi-turn conversations.

[0008] To address the aforementioned pain points, there is an urgent need in this field for a new paradigm for medical information processing that can break through the limitations of the "black box," proactively expose and fix logical vulnerabilities by introducing internal adversarial mechanisms, and possess strict compliance control capabilities. Summary of the Invention

[0009] This invention aims to solve the aforementioned technical problems by providing a medical information processing method and system based on multi-agent adversarial debate. The core of this technical solution lies in constructing a distributed cognitive system based on the "Chiyou Intelligent Core" architecture. By simulating the "multidisciplinary team" (MDT) process of a medical expert team, and utilizing the "adversarial-debate-convergence" mechanism between heterogeneous agents, it achieves in-depth processing of medical information.

[0010] The main technical innovations of this invention are as follows:

[0011] 1. Heterogeneous Multi-Agent Architecture

[0012] This invention does not rely on a single "omnipotent model," but rather constructs an execution environment containing functionally heterogeneous intelligent agents. These agents each assume different cognitive roles:

[0013] ●Proponent (positive / reasoning agent): As the "attending physician", it is responsible for generating positive diagnostic hypotheses based on knowledge graphs and probabilistic models.

[0014] ●Opponent (Adversarial Agent): As the "Devil's Spokesperson" or "Red Team", it is responsible for logically auditing and attacking the assumptions of the affirmative side, and looking for gaps in the chain of evidence.

[0015] ●Judge (Referee / Judgment Agent): Acts as a "neutral observer" and is responsible for calculating the convergence of the debate.

[0016] ●Safety (Risk Control Agent): As an "ethics committee", it is responsible for blocking high-risk recommendations based on hard rules.

[0017] 2. Counterfactual Adversarial Debate Mechanism

[0018] This invention introduces a counterfactual reasoning technique. The Opponent agent does not merely point out errors, but actively generates counterfactual challenge vectors (e.g., "If the patient did not have a fever, would this diagnosis still be valid?"). This mechanism forces the system to perform second-order reflection, simulating different disease progression paths to verify the robustness of the current hypothesis, thereby effectively breaking down confirmation bias and the illusion of commonality.

[0019] 3. Mathematical Consensus Convergence

[0020] To transform the abstract debate process into computable technical metrics, this invention introduces Kendall's concordance coefficient (W) and semantic entropy. The system no longer relies on subjective judgment but instead calculates in real-time the consistency of the hypothetical paths generated by each agent in the semantic space. This provides a rigorous mathematical basis for the system's "halting problem."

[0021] 4. Active Hallucination Cleaning & Ontological Alignment

[0022] This invention embeds an illusion-cleansing sub-step within the debate loop. The system utilizes a medical knowledge graph as an "ontology anchor," requiring all reasoning entities to have a mapping within the graph. The Opponent AI scans the reasoning path in real time, and upon detecting "word splicing illusions" or logical contradictions, immediately generates a high-priority blocking signal.

[0023] 5. Computational optimization via Semantic Distance Clustering

[0024] To improve the computational efficiency of multi-round debates, this invention introduces a semantic distance clustering technique. When the reasoning agent generates a large number of semantically similar hypotheses, the system calculates their vector similarity and clusters them. The Opponent attacks only the "centroid" of each cluster. This avoids repeated debates on subtle variations. Attached Figure Description

[0025] Figure 1 This is an overall logical architecture diagram of a medical information processing system based on multi-agent adversarial debate provided in an embodiment of the present invention. The diagram clearly shows that the Blackboard storage module serves as the data hub, connecting the first inference module, the second adversarial module, the consensus calculation module, and the security risk control module, forming a star topology.

[0026] Figure 2 This is the main flowchart of the multi-agent adversarial debate method provided in this embodiment of the invention. The diagram details the complete closed-loop process from step S1 (data vectorization) to step S7 (result output), with the logical branch nodes of "counterfact challenge generation" and "illusion cleansing" specifically marked.

[0027] Figure 3 This is a schematic diagram of the algorithm logic for generating counterfactual challenge vectors in an embodiment of the present invention. The diagram illustrates how the adversarial agent uses an attention mechanism to identify weak evidence nodes and generates challenge vectors through feature masking and inversion operations.

[0028] Figure 4 This is a consensus convergence determination logic diagram based on Kendall's concordance coefficient (W) in an embodiment of the present invention. The diagram, in the form of a curve, shows the fluctuation trend of the W value and the convergence threshold determination mechanism as the number of debate rounds increases, vividly illustrating how the system moves from "disagreement" to "consensus".

[0029] Figure 5 This is an interaction sequence diagram of semantic distance clustering and illusion cleansing in an embodiment of the present invention. The diagram illustrates the micro-processing flow of how the system first clusters candidate hypotheses, then performs ontology alignment checks, and finally generates blocking signals or rebuttal vectors in a certain round of debate.

[0030] Figure 6 This is a schematic diagram of the layered design of the blackboard storage structure provided in an embodiment of the present invention. The diagram illustrates the data organization of the evidence wall layer, hypothesis space layer, debate record layer, and control state layer, as well as the division of read and write permissions for each agent. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0032] Example 1: System Architecture Setup and Heterogeneous Environment Configuration

[0033] This embodiment constructs a medical information processing system based on the "Chiyou Intelligent Core" architecture. The system runs on a high-performance computing cluster and is designed to handle complex clinical diagnostic reasoning tasks.

[0034] 1. Hardware infrastructure

[0035] like Figure 1 As shown, the system's physical layer is deployed based on a cloud-native architecture:

[0036] ● Computing Units: Employs NVIDIA A100 or H100 GPU clusters, with a minimum of 512GB of GPU memory per node. This is to support the simultaneous running of multiple large-parameter language models (LLMs). For example, an inference agent may need to load a model with 70+ parameters to ensure knowledge coverage, while an adversarial agent may need to load a specialized model focused on logical reasoning (such as DeepSeek-R1).

[0037] ● Storage unit: Adopts a hierarchical storage architecture.

[0038] ○Hot Data Layer: Uses a Redis cluster as the physical carrier of the "Blackboard" and leverages its in-memory key-value storage feature to achieve millisecond-level state read and write, supporting high-frequency agent interaction.

[0039] ○ Cold Data and Knowledge Layer: A static medical knowledge graph (MKG) is stored using the Neo4j graph database, containing millions of medical entities (such as diseases, symptoms, and drugs) and their semantic relationships (such as "cause," "treatment," and "contraindication"). This provides the ontological foundation for "illusion cleansing."

[0040] 2. Software Architecture and Intelligent Agent Role Definition

[0041] The system adopts a microservices architecture, utilizing frameworks such as LangGraph or AutoGen for multi-agent orchestration. The core agents and their roles are defined as follows:

[0042] Agent Name Corresponding model type (example) Functional roles Core Tasks and Prompt Strategies ProponentAgent (positive side) General knowledge-enhanced LLM (such as Qwen-Max) Knowledge base / main promoter Task: Generate the most comprehensive diagnostic hypothesis possible based on the evidence on the blackboard. Prompt: "As an experienced internist, list all possible diseases, prioritize high-risk conditions, and cite the chain of evidence from the knowledge graph." OpponentAgent (the opposing side) Logical reasoning enhanced LLMs (such as DeepSeek-R1) Logic Auditing / Red Team Task: Ignore existing conclusions and find logical fallacies. Perform counterfactual reasoning. Prompt: "Ignore the above conclusions. Look for missing evidence. If key symptom X is absent, derive an alternative conclusion. Perform logical consistency checks." JudgeAgent Statistical Analysis Models / Lightweight LLM Consensus computation Task: Calculate the vector distance and Kendall's W. Prompt: "Based solely on the strength of the evidence chains provided by both parties, calculate semantic consistency and determine whether a Nash equilibrium has been reached." SafetyAgent (Risk Control) Rules engine + risk control model (such as AntModel) Trust layer / Blue team Task: Intercept risks based on regulatory databases. Prompt: "Check for contraindications, overtreatment, or recommendations that violate clinical guidelines. Veto power." Interaction Agent Anthropomorphic LLM (e.g., Doubao) Patient simulation / interface Task: Semantic normalization and patient intent simulation. Prompt: "Translate spoken language into standard terminology. Simulate potential questions a patient may have about the diagnosis."

[0043] 3. Design of Blackboard Storage Structure

[0044] In this embodiment, the blackboard is not merely a database; it is a structured dynamic graph, strictly divided into four regions:

[0045] ● Evidence Wall: Stores "atomic facts" cleaned by the Interaction Agent. This area is read-only / append-only to prevent illusions during the reasoning process from contaminating the original data.

[0046] ● Hypothesis Space: This is the main battleground of the debate. It stores the "Argumentation Graph" generated by the Proponent, where nodes represent hypotheses and edges represent supporting relationships. This area is mutable and can be read and written.

[0047] ●Debate Floor: Stores the counterfactual challenge vector and blocking signal generated by Opponent.

[0048] ●Control Zone: Stores the current iteration round, semantic entropy value, current focus, and system state machine flags.

[0049] Example 2: In-depth analysis of the adversarial debate handling process

[0050] This example uses a typical complex case of "differential diagnosis of chest pain" to illustrate in detail... Figure 2 The illustrated process demonstrates how the system addresses the problems of "hallucination" and "misdiagnosis" through adversarial and collaborative approaches.

[0051] Step S1: Data Acquisition and Semantic Normalization

[0052] The user entered: "Male, 55 years old, had a chest pain like a heavy stone pressing on it for two hours, broke out in a cold sweat, and also had shoulder pain. Has high blood pressure."

[0053] The Interaction Agent (Doubao) intervenes first to perform semantic normalization:

[0054] ● "Feeling like a heavy stone is pressing on my chest" -> Mapped to the standard term Symptom: Chest Pain (Quality: Crushing / Oppressive)

[0055] ● "Break out in a cold sweat" -> Mapped to Symptom: Diaphoresis

[0056] ●“My shoulder also hurts” -> Mapped to Symptom: Radiation Pain (Location: Shoulder)

[0057] ● "Hypertension" -> Mapped to History: Hypertension. These standardized feature vectors $V_{init}$ are written into the evidence wall on the blackboard.

[0058] Step S2: Initial Hypothesis Generation

[0059] The Proponent Agent (Qwen) reads the evidence wall and retrieves information from the Neo4j knowledge graph.

[0060] It generates the first hypothesis path $H_1$:

[0061] ●Hypothesis A: Acute myocardial infarction (AMI). Confidence level: 0.85. Reasoning: Crushing chest pain, radiating pain, and profuse sweating are all typical symptoms of AMI, and hypertension is a high-risk factor.

[0062] ●Hypothesis B: Angina pectoris. Confidence level: 0.10.

[0063] ●Hypothesis C: Aortic dissection. Confidence level: 0.05.

[0064] Step S3: Counterfactual Challenge Vector Generation (Opponent)

[0065] The Opponent Agent (DeepSeek) is activated. Instead of blindly trusting $H_1$, it performs logical auditing.

[0066] It found a hidden danger in the chain of evidence: although the symptoms were highly consistent with AMI, the consequences of missing the diagnosis of "aortic dissection" were disastrous, and "hypertension" was also a strong correlation factor for dissection.

[0067] Opponent constructs the counterfactual challenge vector $C_1$:

[0068] ● Masking operation: Temporarily mask the "radiating pain" feature.

[0069] ●Logical perturbation: Introduce the assumption that "the nature of the pain is tearing rather than squeezing".

[0070] ● Generate challenge text: "Although current evidence supports AMI, the patient has a history of hypertension. If radiating pain is ignored, can the current evidence completely rule out aortic dissection? Note that misdiagnosing a dissection as a myocardial infarction (such as using anticoagulants) will lead to the patient's death. Is there evidence of asymmetric blood pressure in both upper limbs?" This process embodies "red team" thinking, probing the robustness of the diagnosis through counterfactual assumptions.

[0071] Step S4: Iterative Debate and Illusion Cleansing

[0072] The Proponent received $C_1$.

[0073] ● Hallucination cleansing trigger: Suppose that the Proponent attempts to fabricate evidence in the response: "The patient's ECG shows ST segment elevation."

[0074] ●Blocking Mechanism: The Safety Agent (Ant) scans the blackboard in real time and finds no "ECG" related data on the evidence wall. The Safety Agent immediately generates a logic blocking signal, marking the response as "Hallucination," forcing the Proponent to withdraw the assertion.

[0075] ● Revised Response: Proponent acknowledges that the current evidence is insufficient to completely rule out the mezzanine. It updates the hypothesis path $H_2$:

[0076] The confidence level for AMI has been lowered to 0.70.

[0077] ○ The confidence level for aortic dissection was increased to 0.25.

[0078] ○ Generate dynamic query: Request "Body limb blood pressure measurement results" and "D-dimer data".

[0079] Step S5: Consensus Convergence Judgment (Judge)

[0080] The Judge Agent calculates the semantic consistency between $H_1$ and $H_2$.

[0081] Due to a significant shift in the confidence distribution (from 0.85 / 0.05 to 0.70 / 0.25), the semantic entropy is in a state of high-level oscillation.

[0082] Judge's decision: Nash equilibrium has not been reached.

[0083] The system outputs an intermediate result suggesting immediate targeted examinations (blood pressure, CT scan) rather than providing a definitive diagnosis. This reflects the system's safety and prudence.

[0084] Example 3: Mathematical Implementation of the Core Algorithm

[0085] This embodiment details the key algorithms involved in the claims.

[0086] 1. Kendall's Concordance Factor (W) Calculation

[0087] To quantify "consensus", the system constructs an m × n sorting matrix $R$.

[0088] Suppose there are 3 agents (Proponent, Opponent, Safety) ranking 4 diseases (AMI, aortic dissection, pulmonary embolism, gastritis).

[0089] In round $t$, the sorting is as follows:

[0090] ●P:

[0091] ●O:

[0092] ●S: Calculate the sum of rankings for each disease $R_j$: AMI=3, aortic dissection=7, pulmonary embolism=9, gastritis=11. Average ranking $\bar{R} = 7.5$. Sum of squared deviations $S = \sum (R_j - \bar{R})^2 = 35$. $$W = \frac{12 \times 35}{3^2(4^3-4)} \approx 0.77$$ Since $0.77 > 0.7$ (threshold), the system determines that a consensus has been reached on the "primary diagnostic direction".

[0093] 2. Semantic Distance Clustering

[0094] In complex cases, Proponent may generate numerous subtle hypotheses (such as "influenza A," "influenza B," "viral upper respiratory infection"). To avoid Proponent wasting computational resources refuting each one:

[0095] ● The system will map the hypothesis to vectors $V_{h1}, V_{h2},...$.

[0096] ● Calculate the Euclidean distance $D(V_{h1}, V_{h2}) = ||V_{h1} - V_{h2}||$.

[0097] ● If $D < \theta$ (threshold), then they are classified into one category.

[0098] ●Opponent generates challenge vectors only for the centroid of the cluster. This significantly improves the system's response speed and computational efficiency.

[0099] Example 4: Security and Risk Control Blocking Mechanism

[0100] This embodiment demonstrates the specific working method of the Safety Agent (Ant Model).

[0101] Scenario: Proponent recommends prescribing ribavirin to a pregnant patient.

[0102] 1. Rule matching: The Safety Agent scan suggested the entity "ribavirin" and the patient characteristic "pregnant woman".

[0103] 2. Knowledge base query: In the rule base, ribavirin is "FDA Category X" (absolutely contraindicated) for pregnant women.

[0104] 3. Forced blocking: Although the logic is sound from the perspective of treating viral infection (Proponent's reasoning), the Safety Agent triggers a veto.

[0105] 4. Output Correction: The system not only deletes the suggestion, but also highlights a warning in the output report: "Contraindication risk detected: Ribavirin is contraindicated in pregnant women." This mechanism ensures that the AI ​​system is not only "smart," but also "safe."

[0106] In summary, this invention, by constructing a heterogeneous multi-agent system based on the "Chiyou Intelligent Core," and utilizing counterfactual adversarial debate, rigorous mathematical consensus determination, and proactive illusion cleansing mechanisms, successfully overcomes the bottlenecks in accuracy, interpretability, and robustness of single models, providing a safe, reliable, and efficient technical solution for medical auxiliary diagnosis.

Claims

1. A medical information processing method based on multi-agent adversarial debate, characterized in that, The method operates in a distributed system containing heterogeneous computing units, configured to perform deep reasoning tasks based on the "Chiyou Intelligent Core" architecture. The method includes the following steps: Step S1: Data Acquisition and Multimodal Feature Vectorization. Unstructured medical text data of the target object is acquired, including chief complaint text, present medical history, past medical history, and auxiliary examination results. A pre-trained natural language processing model is used to extract entities and encode features from the text data, generating an initial feature vector containing symptom features, physical sign features, negation features, and time dimension features. The initial feature vector is written into a shared blackboard structure. Step S2: Instantiation and Role Assignment of the Multi-Agent Execution Environment. An execution environment containing multiple heterogeneous functional agents is instantiated in the computing device. The environment is configured with an event-driven scheduler. The agents include at least: a first reasoning agent (Proponent Agent), configured to load a medical knowledge reasoning model to generate a forward reasoning path; a second adversarial agent (Opponent Agent); and a third adversarial agent (Opponent Agent). The first agent is configured to load a logical critique model to generate reverse questioning paths and counterfactual challenges; the consensus judgment agent is configured to load a statistical evaluation model to evaluate the semantic consistency between paths; the safety agent is configured to load an ethics and regulatory rule base to execute compliance blocking; Step S3: Generation of positive hypotheses based on the knowledge graph. The first reasoning agent reads the initial feature vector in the blackboard and retrieves the preset medical knowledge graph. Through a combination of graph path search and probability prediction, a first set of hypothetical paths and its corresponding initial confidence distribution are generated. The first set of hypothetical paths contains at least one candidate medical conclusion and its complete evidence chain index; Step S4: Construction and generation of counterfactual challenge vectors. The second adversarial agent performs logical auditing on the first set of hypothetical paths. Counterfactual scenarios are constructed by masking, inverting, or perturbating key feature nodes in the initial feature vector. At least one counterfactual challenge vector is generated based on the counterfactual scenarios. Vector), which contains logical negation queries or alternative hypothesis hints for a specific node in the chain of evidence;Step S5: Iterative Adversarial Debate and Dynamic Correction The process involves multiple rounds of iterative debate. In each iteration: the first reasoning agent receives the counterfactual challenge vector, reassesses the completeness of the evidence chain based on the perturbation information in the vector, retrieves supplementary evidence or corrects the reasoning logic, and updates the first hypothesis path and confidence distribution; the second adversarial agent, based on the updated first hypothesis path, retrieves potential contradictory evidence or logical gaps again and generates a new challenge vector; the security risk control agent monitors the hypothesis paths generated in each round in real time, and immediately generates a high-priority blocking signal once a path violating preset clinical guidelines or containing contraindication features is detected; Step S6: Consensus Convergence Determination Based on Kendall's Concordance Coefficient The consensus decision agent calculates the semantic consistency index between hypothesis paths in each round in real time; when the semantic consistency index meets the preset convergence threshold and the security risk control agent does not output a blocking signal, the debate is terminated; Step S7: Structured Result Output The finally converged hypothesis path, the corresponding confidence distribution, and the debate process summary are converted into structured medical information processing results and output.

2. The method according to claim 1, characterized in that, The specific process of generating the counterfactual challenge vector in step S4 includes: calculating the contribution weight of each feature node in the first hypothesis path to the conclusion using an attention mechanism; identifying weak evidence feature nodes whose weights are below a first threshold but are highly sensitive to conclusion reversal; constructing a "hypothesis-negation" logical chain for the weak evidence feature nodes to form a counterfactual prompt; inputting the counterfactual prompt into the second adversarial agent to generate a set of candidate challenges that are logically mutually exclusive with the first hypothesis path; and selecting candidate challenges that conform to physiological and pathological logic as the counterfactual challenge vector based on the preset ontological constraints in the medical knowledge graph.

3. The method according to claim 1, characterized in that, The semantic consistency index in step S6 is calculated using Kendall's Coefficient of Concordance (W), which is calculated as follows: $$W= \frac{12 S}{m^2 (n^3 - n)}$$ where $S$ is the sum of squared deviations of the hypothetical paths generated in each round at the candidate ranking level, $m$ is the number of agents participating in the debate, and $n$ is the number of candidate conclusions; the preset convergence threshold is set to $W > 0.7$; when the value of $W$ is lower than the threshold, the system triggers the next round of debate iteration; if the number of iterations reaches the preset upper limit and convergence is still not achieved, an intermediate state report containing multiple ambiguities is output and manual intervention is prompted.

4. The method according to claim 1, characterized in that, The method further includes a computational optimization step based on semantic distance clustering: in step S3 or S5, if the first reasoning agent generates multiple candidate hypothesis paths, the cosine similarity or Euclidean distance between each path vector is calculated; paths with similarity higher than a preset clustering threshold are merged into the same semantic cluster; when the second adversarial agent generates challenge vectors, it only generates counterfactual challenge vectors for the centroid of each semantic cluster to reduce redundant consumption of computational resources.

5. The method according to claim 1, characterized in that, The iterative adversarial debate process also includes a hallucination cleaning and ontology alignment sub-step: In each iteration, the second adversarial agent performs a logical consistency check, mapping key entities in the hypothesis path generated by the first reasoning agent to a pre-set medical knowledge graph; if the hypothesis path contains entities that cannot be mapped in the knowledge graph (word-splicing hallucination), or contains assertions that contradict the negative features in the initial feature vector (consistency error), a high-priority logical blocking vector is generated, forcing the first reasoning agent to remove the relevant assertions in the next update.

6. The method according to claim 1, characterized in that, The blackboard storage structure adopts a layered design, including at least the following: Evidence Wall: stores the original fact data after de-privacy and standardization processing, set to read-only or append-only mode; Hypothesis Space: stores the dynamic hypotheses, inference chains, and confidence data generated by each agent, supporting read and write operations; Debate Floor: stores the counterfactual challenge vector, blocking signals, and interaction logs for each round; Control Zone: stores the current iteration round, semantic entropy value, and system state machine flags.

7. The method according to claim 1, characterized in that, The heterogeneous agents in the multi-agent execution environment are loaded with the following types of models: the first reasoning agent is loaded with a general large language model pre-trained on massive medical literature and an external medical literature retrieval database, focusing on breadth-first search; the second adversarial agent is loaded with a logic reasoning-specific model fine-tuned by reinforcement learning, focusing on depth-first logic chain auditing and vulnerability discovery; the security risk control agent is loaded with a rule engine based on a financial-grade risk control architecture, focusing on compliance filtering based on hard rules.

8. A medical information processing system, characterized in that, include: The data acquisition and preprocessing module is used to receive unstructured medical text and vital sign data of the target object, convert them into a standard medical term sequence through semantic normalization, and generate an initial feature vector. The blackboard storage module is configured as a shared storage structure based on an in-memory database to store feature vectors, hypothesis paths, challenge vectors, and debate history, and supports asynchronous read and write based on a publish-subscribe pattern. The multi-agent adversarial reasoning engine includes: a first reasoning module (Proponent Module), configured to generate positive diagnostic hypotheses based on feature vectors; a second adversarial module (Opponent Module), configured to generate challenge vectors and logical vulnerability hints based on counterfactual reasoning algorithms; a safety risk control module (Safety Module), configured to execute real-time blocking based on a rule base; a consensus calculation and scheduling module, configured to load a statistical analysis model for calculating semantic consistency indicators according to the formula described in claim 3, and scheduling the execution sequence of each agent according to the calculation results; and a result output module, used to convert medical recommendations that meet the convergence conditions into structured reports and output them.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method described in any one of claims 1 to 7.

Images