A multi-disciplinary pre-consultation method and system based on multi-agent cooperation

Abstract

This invention discloses a multidisciplinary pre-consultation method and system based on multi-agent collaboration. The method includes: 1) Clinical state initialization: constructing a structured clinical state and dividing it into a case feature information set and a diagnosis and planning set; 2) State-driven agent scheduling: scheduling one or more specialist agents related to the current symptoms based on the current clinical state and historical dialogue information; 3) Parallel analysis and suggestion generation: each scheduled specialist agent independently analyzes the current state and outputs problem suggestions and state update suggestions; 4) Clinical state update: semantically aggregating the outputs of multiple agents, updating the case feature information set, and generating the next round of patient-facing inquiries; 5) Diagnosis generation: when the case feature information set meets the overall integrity condition, generating a diagnostic conclusion and treatment plan based on the set, forming the initial medical record. This invention improves the structuring level, diagnostic accuracy, and clinical interpretability of intelligent consultation without requiring real multidisciplinary consultation resources, and has broad application prospects.

Description

[0001] This invention belongs to the field of medical artificial intelligence technology, specifically relating to a multidisciplinary pre-diagnosis method and system based on multi-agent collaboration. Background Technology

[0002] The initial consultation is a crucial step in the clinical medical process. Within a limited timeframe, doctors must extract symptom clues from the patient's natural language description and, based on professional judgment, structurally record them in the medical record, forming the "Initial Progress Note" (IPN). This is not only the beginning of the medical record but also a vital basis for subsequent diagnostic, testing, and treatment decisions. Therefore, the completeness, logic, and professionalism of the information gathered during the consultation have a decisive impact on the entire treatment pathway.

[0003] However, in traditional outpatient settings, doctors often face time constraints and patient pressure, requiring them to simultaneously record key medical information during real-time conversations with patients. This model relies heavily on personal experience and clinical intuition, making it susceptible to cognitive biases, particularly "anchoring bias," where doctors prematurely focus on a single obvious symptom while neglecting other potential causes. Furthermore, the parallel processing of information gathering and diagnostic thinking leads to fragmented medical records, logical jumps, and disconnects from context, severely impacting the standardization and traceability of documentation.

[0004] To overcome misjudgment in complex cases, some hospitals organize "multidisciplinary teams" (MDTs) in specific situations, bringing together doctors from different specialties to collectively assess a patient's condition. The MDT model has indeed demonstrated superior diagnostic accuracy and comprehensiveness in complex and difficult cases, and has become an important institutional guarantee for large hospitals. However, in reality, fully extending this mechanism to routine outpatient services still faces many challenges, such as high personnel organization costs, slow response times, and limited resource allocation. Therefore, this high-quality treatment model is difficult to replicate on a large scale in ordinary initial consultation scenarios.

[0005] With the development of artificial intelligence technology, an increasing number of studies are attempting to use Large Language Models (LLMs) to assist in initial consultation and medical record generation. While these systems possess a certain level of language understanding and inductive reasoning capabilities, they often employ a single-model, serial processing architecture, lacking multi-perspective and multi-specialty collaborative reasoning mechanisms. This makes them prone to problems such as information redundancy, diagnostic jumps, or generating content out of thin air. Especially when faced with scenarios involving ambiguous patient statements, complex information, and overlapping symptoms, existing AI consultation systems struggle to maintain stable performance, limiting their deployment value in real-world clinical settings.

[0006] Therefore, there is an urgent need for an intelligent pre-diagnosis method and system that can replace the real MDT mechanism and has the ability to perform multi-perspective collaborative reasoning and output structured medical records. This system should be able to adapt to the resource constraints of real-time outpatient scenarios and maintain high-quality, systematic medical logic support during the consultation process, thereby improving the efficiency and quality of initial diagnosis and laying the foundation for future "AI-assisted medical care". Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a multidisciplinary pre-diagnosis method and system based on multi-agent collaboration.

[0008] This invention introduces a structured clinical state management mechanism, a multi-agent-based parallel analysis and dynamic collaboration mechanism, and a state evolution-oriented inquiry and diagnosis generation mechanism. This enables the gradual improvement of information and multidisciplinary collaborative analysis during the pre-diagnosis process, solving problems such as insufficient information collection, lack of constraints on the reasoning process, and difficulty in tracing diagnostic conclusions in existing technologies.

[0009] The first aspect of this invention relates to a multidisciplinary pre-diagnosis method based on multi-agent collaboration, comprising the following steps: Clinical status initialization: Construct a structured clinical status and divide the clinical status into a set of case feature information used to describe the patient's subjective and objective information (i.e., chief complaint, present illness, past medical history, physical examination and auxiliary examinations), and a set of diagnostic and planning information used to describe the diagnostic conclusion and treatment plan (i.e., preliminary diagnosis, discussion of proposed diagnosis, treatment plan).

[0010] State-driven agent scheduling: Based on the clinical status of the current round and historical dialogue information, one or more specialist agents related to the characteristics of the current case are scheduled to participate in the analysis.

[0011] Parallel analysis by specialist agents: Each scheduled specialist agent independently analyzes the current case feature information based on its own medical domain knowledge, and generates query suggestions and status update suggestions to supplement the case feature information.

[0012] Clinical status update: Integrate the suggestions generated by each specialty agent, update the case feature information set to form a new clinical status, and generate the next round of patient-oriented questions based on the updated clinical status.

[0013] Diagnosis generation: After multiple rounds of status updates, when the case feature information set meets the preset integrity conditions, a corresponding diagnostic conclusion and treatment plan are generated based on the case feature information set, and a complete initial medical record is formed.

[0014] Furthermore, the diagnosis generation step is not triggered until the case feature information set is updated and stabilized, so as to ensure that the diagnostic conclusion is generated only based on sufficient case information.

[0015] Furthermore, the clinical status is repeatedly updated in the status update step until the set of case feature information as a whole forms a preliminary medical record information framework that can support subsequent diagnostic reasoning.

[0016] The second aspect of the present invention relates to a multidisciplinary pre-diagnosis system based on multi-agent collaboration, the system comprising multiple functional modules working collaboratively; include: The clinical status management module is used to build, store, and update structured clinical status information; The main control and scheduling module is used to dynamically select and activate the corresponding specialty intelligent agents based on changes in clinical status; The specialist intelligent agent module is used to independently analyze clinical conditions and generate inquiry suggestions or diagnostic suggestions; The clinical status update module is used to integrate the outputs from multiple specialty agents, drive the update of clinical status, and generate patient-oriented inquiry content based on the updated clinical status. The diagnosis generation module is used to generate diagnostic conclusions and treatment plans when the case feature information set is updated. The clinical status management module is used to distinguish between case characteristic information and diagnostic and planning information, and to restrict the generation order of the two.

[0017] The main control scheduling module dynamically adjusts the participation of specialist intelligent agents based on changes in clinical status and dialogue history.

[0018] The specialized intelligent agent modules operate independently during the analysis process to avoid interference between inference results.

[0019] The semantic aggregation module employs a "post-write generation" strategy, which merges the suggested execution content from multiple agents and performs state updates, then generates the natural language query text for the next round of patient interaction.

[0020] The diagnosis generation module processes only the updated set of case feature information during the diagnosis generation stage.

[0021] The system, through the collaboration of the aforementioned modules, forms a pre-diagnosis process that conforms to the logic of multidisciplinary joint diagnosis and treatment.

[0022] This invention constructs a structured clinical state representation and employs a multi-round state-driven agent scheduling mechanism to achieve parallel analysis, information supplementation, and independent reasoning by multiple specialized agents. During the information acquisition phase, the system dynamically evolves the state and performs semantic fusion. Once the case information meets the overall completeness standard, it enters the diagnostic phase, generating a structured preliminary medical record. This solution improves the structuring level, diagnostic accuracy, and clinical interpretability of intelligent consultations without requiring real multidisciplinary consultation resources, and has broad application prospects.

[0023] The working principle of this invention is: This invention is based on the collaborative model of real-world multidisciplinary team (MDT) consultations. It addresses the challenges of routine application in real-world outpatient settings due to high organizational costs, slow response times, and resource scarcity. By implementing this model in a virtual environment through a multi-agent collaboration mechanism, a "virtual MDT" consultation and reasoning process can be implemented in the pre-consultation stage. The system first constructs a structured clinical state in SOAP format, dividing it into a case feature information set and a diagnosis and planning set. The case feature information set carries factual information such as the patient's chief complaint, present illness, past medical history, physical examination, and auxiliary examinations. The diagnosis and planning set carries inferential conclusions such as preliminary diagnosis, discussion of proposed diagnoses, and treatment plans. At the initial stage of consultation, the system only grants write access to the case feature information set and freezes the diagnosis and planning set to ensure that diagnostic outputs are not triggered prematurely. Subsequently, the main control scheduling module continuously reads the current clinical status and historical dialogue context, dynamically matching the required specialist agents based on symptom keywords, risk signals, and missing fields. Multiple specialist agents are then scheduled to run in parallel within their respective independent inference environments, allowing them to analyze the current case's characteristic information from different disciplinary perspectives. This simulates the collaborative mechanism of "independent judgment and complementary coverage" among multiple specialties in a real-world multidisciplinary team (MDT). Each specialist agent outputs suggestions for key questions to be asked the patient in the next round, as well as suggestions for supplementing and revising structured status fields. The semantic aggregation module receives the outputs from each agent and performs unified fusion. It first performs semantic alignment, conflict resolution, and field mapping on the status update suggestions from multiple agents, updating and writing the case characteristic information set. Then, based on the updated structured clinical status, it generates the next round of natural language queries and feeds them back to the patient, forming a "post-write generation" state-driven interaction mode. This ensures that the consultation process is always controlled by the evolution of the clinical status and maintains traceable consistency. The aforementioned scheduling-parallel reasoning-aggregation update-query generation process can be executed in multiple rounds, gradually improving the case feature information set and achieving overall integrity. When the system determines that the case feature information set is sufficient to support diagnostic reasoning, its execution state is frozen, further writing is stopped, and the diagnosis generation phase is initiated. The relevant specialist agents are scheduled again to generate preliminary diagnoses, proposed diagnoses, and treatment plans in parallel based on the frozen unified case feature set. Finally, the semantic aggregation module merges these to generate a set of diagnoses and plans, which, together with the case feature information set, are output as a structured first medical record (IPN). This achieves a pre-consultation effect close to the quality of MDT collaboration without requiring real multidisciplinary physician resources, while ensuring the interpretability and traceability of the diagnostic process.

[0024] The innovation of this invention is: 1. This invention proposes a feasible mechanism for implementing a "virtual MDT" to achieve intelligent migration of real-world clinical collaboration models. Drawing on the organizational logic and diagnostic advantages of multidisciplinary team (MDT) consultations in the real-world healthcare system, this invention realizes the MDT model, which was originally difficult to apply to ordinary outpatient initial consultation scenarios due to limitations in time, manpower, organizational costs, and response efficiency, in a virtual environment through multi-agent collaboration. This invention enables parallel analysis of different specialties without the need for real multidisciplinary physician resources, thereby forming a comprehensive reasoning framework from a multidisciplinary perspective at the initial consultation stage. This achieves "MDT-based outpatient consultation," significantly improving information coverage and diagnostic reliability in complex symptom scenarios, and has significant innovative implications for healthcare workflow paradigms.

[0025] 2. This invention proposes a state-driven dynamic specialist agent scheduling mechanism to achieve adaptive multidisciplinary collaboration oriented towards symptom evolution. Instead of fixedly calling all agents, this invention dynamically identifies the required specialist scope based on the current clinical state and dialogue history, and activates relevant specialist agents to participate in reasoning as needed. This allows the system to adjust the multidisciplinary participation structure as patient information is gradually completed, forming a collaborative mechanism similar to the "specialist participation based on the patient's condition" in real MDT consultations, improving reasoning efficiency and resource utilization. Furthermore, this invention divides the clinical state into a set of case feature information and a set of diagnoses and plans, and controls the timing of diagnosis generation through a freezing mechanism. This ensures that the system only supplements factual information and updates the structure during the information collection phase, avoiding anchoring bias and reasoning drift caused by "asking and diagnosing simultaneously" and "premature conclusions" common in existing technologies, thus enhancing clinical safety and credibility from a mechanistic perspective.

[0026] 3. This invention proposes a pre-consultation driven system based on structured clinical states, providing a computationally computable, evolvable, and traceable data foundation for the consultation process. Using a SOAP structure as a unified information organization framework, this invention continuously maps patient input, dialogue context, and system inference results into structured clinical states. This eliminates the reliance on freely stacked text, instead forming a standardized and manageable evolutionary process of clinical states, significantly improving the completeness and logical consistency of consultation information.

[0027] The beneficial effects of this invention are as follows: Through a stepwise update mechanism of clinical status, the consultation process has a clear information evolution path; by naturally distinguishing between information collection and diagnosis generation, the risk of premature diagnosis is reduced; through multi-agent parallel analysis, the coverage of complex symptom combinations is improved; through structured status management, the interpretability and traceability of diagnostic results are improved; and the pre-consultation effect close to that of multidisciplinary collaboration is achieved without relying on real multidisciplinary consultation resources.

[0028] This invention is applicable to a variety of medical pre-diagnosis and decision support scenarios, and has good practical value and prospects for promotion. Attached Figure Description

[0029] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0030] Figure 1 This is a flowchart illustrating a multidisciplinary pre-diagnosis method based on multi-agent collaboration. Figure 2 This is a schematic diagram of the module structure of a multidisciplinary pre-diagnosis system based on multi-agent collaboration. Detailed Implementation

[0031] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0032] Example 1

[0033] This embodiment proposes a multidisciplinary pre-diagnosis method based on multi-agent collaboration. For example... Figure 1 As shown, this method, centered on structured clinical states, achieves dynamic acquisition of consultation information and decoupled execution of multidisciplinary diagnostic inference through state-driven agent scheduling and semantic aggregation mechanisms. This method maintains the continuity of consultation content and the traceability of diagnostic logic throughout the entire process, thereby significantly improving the structured quality and diagnostic accuracy of initial consultation inquiries.

[0034] This method includes the following steps: Step 1: Clinical status initialization In this invention, the system first initializes a structured clinical state, using a clinically compliant SOAP (Subjective, Objective, Assessment, Plan) format, dividing the state into two parts: a "case feature information set" and a "diagnosis and plan set." The "case feature information set" stores the patient's chief complaint, present illness, past medical history, physical examination, and auxiliary examinations provided during the consultation. The "diagnosis and plan set" is reserved for structured expression in subsequent preliminary diagnosis, preliminary discussion, and treatment planning. In this step, only the case feature information set is initialized to a blank state; the diagnosis and plan set remains unfilled. The entire consultation process is state-driven and logic-controlled based on this structured state.

[0035] Step 2: State-driven agent scheduling

[0036] After completing the initial state construction, the system will analyze the current clinical status and doctor-patient dialogue history in real time to determine which information is missing or unclear in the current case characteristics. Based on the clinical symptom keywords or risk signals involved in the collected information, it will dynamically determine the types of specialist agents to be invoked. The scheduling process is executed by the main control scheduling module, the core of which is a strategy function that combines clinical status and dialogue history for reasoning and judgment. For example, if the current clinical status includes the complaint of "chest pain," the system will match relevant agents such as cardiology and respiratory medicine; if the patient replies "heartburn," a gastroenterology agent will be further involved in the analysis. The result of each round of agent scheduling is a finite set, and the system will activate the selected agents in parallel for the next stage of processing.

[0037] Step 3: Parallel Analysis of Specialized Intelligent Agents

[0038] Activated specialist agents operate within their own independent reasoning environments, performing local analysis of the input clinical state and, combined with their own specialty's common symptom patterns and interaction strategies, proposing targeted supplementary questions and state update suggestions. The agents' outputs primarily include two categories: first, suggestions for further questioning the patient to clarify the diagnostic direction, such as "Does the pain worsen with exercise?" or "Is it accompanied by nausea or acid reflux?", guiding the next round of information collection; second, draft revisions to fields in the existing structured clinical state (i.e., the case feature information set) based on the patient's latest response and current context, such as "It is suggested to update the chief complaint field to 'chest pain with intermittent acid reflux'" or "Add 'History of gastric ulcer three years ago' to the past medical history." These suggestions may involve adding new entries or refining existing fields. All agents do not share intermediate analysis processes with each other, avoiding premature convergence of the reasoning path or the emergence of herd bias, thus ensuring the breadth and diversity of information collection dimensions.

[0039] Step 4: Semantic Aggregation and State Update

[0040] After receiving suggestions from various specialized agents, the system enters the semantic aggregation phase. This phase is executed by the semantic aggregation module, whose main functions include aggregating state revision suggestions and fusing problem suggestions. The system processes the two types of output from different agents separately: On the one hand, for the "structured status revision suggestions" proposed by each agent based on the patient's latest response, the system maps them to specific fields in the structured clinical status, prioritizing suggestions that are logically consistent and context-matched, and comparing and selecting the best version for conflicting fields, or temporarily suspending the writing of ambiguous fields until subsequent clarification, thereby updating the structured clinical status. All updates are performed only on the "case feature information set," while the diagnosis and planning sets remain frozen.

[0041] On the other hand, after completing the above status update, for the "problem suggestions" for patients, the system will aggregate similar content, eliminate redundancy, and merge to generate more general and continuous natural language query statements. For example, when multiple agents suggest "the cause needs to be clarified", the system can uniformly generate the question "Under what circumstances do you think this symptom usually occurs?" and present it to the patient.

[0042] Steps 2 through 4 constitute a complete round of patient information collection. This round can be repeated multiple times as needed to continuously improve the case feature information set. This cyclical process will continue until the current structured clinical status forms a relatively complete preliminary medical record framework at the macro level, that is, the case feature information set has sufficient support for subsequent diagnostic reasoning, and the system determines that the status meets the conditions for transitioning to the diagnostic process.

[0043] Step 5: Diagnosis and Plan Generation

[0044] When the system comprehensively evaluates the current clinical state and deems the case feature information set sufficient to support diagnostic reasoning, it enters the state freeze process and initiates the second phase of diagnosis generation. At this point, the system no longer updates the case feature information set to ensure the diagnostic process is based on a stable and unchanging state. The main control scheduling module reactivates the specialist agents related to the current case. Each specialist agent performs independent diagnostic reasoning based on the frozen case feature information set, outputting content including preliminary diagnosis, proposed diagnosis discussion, and treatment plan. The generated content is then passed to the semantic aggregation module, where it is categorized and merged to ultimately form a structured "diagnosis and plan set."

[0045] After completing the collection, the system integrates the case characteristics information and diagnostic content to generate a complete initial medical record. This record can be used to assist doctors in decision-making, be archived in electronic medical record systems, or serve as the basis for the automatic generation of subsequent detailed medical records.

[0046] Example 2

[0047] like Figure 1This embodiment relates to a multidisciplinary pre-diagnosis system based on multi-agent collaboration. This system is based on the dynamic evolution of structured clinical states, with each module collaboratively performing consultation and diagnosis tasks in a modular manner. The overall system structure adopts a graph structure, where each module is a node to another, and collaboration is driven by control strategies.

[0048] include: The clinical status management module is used to build and maintain structured clinical statuses in SOAP format. The main control and scheduling module is used to dynamically activate the required specialist agents based on the dialogue history and clinical status content; Specialty agent modules are used to simulate roles such as cardiology, gastroenterology, and neurology, and to perform independent reasoning about clinical conditions. The semantic aggregation module is used to integrate the output results of various agents and complete the merging and updating of states; The diagnostic generation module is used to drive the output of diagnostic recommendations and treatment plans after the state has stabilized; The clinical status management module is implemented using an explicit structure, which divides the case feature field and the diagnosis output field, and has status freeze control logic.

[0049] The master control scheduling module selects appropriate intelligent agents through context modeling and domain matching logic, supporting parallel reasoning and resource control.

[0050] The specialized intelligent agent modules are independent of each other and use prompt words to constrain the domain reasoning range to avoid cross-interference.

[0051] The semantic aggregation module employs a "post-write generation" strategy to ensure that all output content to patients is generated by the updated structural state.

[0052] The query generation module supports continuous multi-round dialogue and adaptive content generation to obtain missing feature fields.

[0053] The diagnosis and planning generation module is activated only after the case feature information set is frozen, ensuring that the diagnostic basis is sufficient and traceable.

[0054] In its implementation, the system first constructs a SOAP state framework. The main control scheduling module identifies the specialist roles that need to be activated, and the state content is dynamically improved through multiple rounds of parallel information supplementation. Once the state reaches the preset completeness, the system freezes the case feature set and enters the diagnostic phase, ultimately generating a complete initial medical record.

[0055] Through the above steps, this invention achieves a flexible, efficient, and interpretable multidisciplinary pre-diagnosis mechanism. The system utilizes structured states as core data support, maintaining natural interaction while strictly controlling the boundaries between diagnostic output and information collection, significantly improving consultation quality and document consistency. This system not only enhances the clinical usability of intelligent consultation systems but also has significant application value and promotional significance in areas such as assisted diagnosis and triage decision-making.

[0056] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A multidisciplinary pre-diagnosis method based on multi-agent collaboration, comprising the following steps: 1) Clinical status initialization: Construct a structured clinical status and divide it into a case feature information set and a diagnosis and planning set; 2) State-driven agent scheduling: Based on the current clinical state and historical dialogue information, schedule one or more specialist agents related to the current symptoms; 3) Parallel analysis and suggestion generation: Each scheduled specialist agent independently analyzes the current state and outputs problem suggestions and state update suggestions; 4) Clinical status update: Semantic aggregation of the outputs of multiple agents, updating the case feature information set, and generating the next round of patient-oriented questions; 5) Diagnosis generation: When the set of case feature information meets the overall integrity condition, a diagnosis conclusion and treatment plan are generated based on the set, forming the initial medical record.

2. The multidisciplinary pre-diagnosis method based on multi-agent collaboration as described in claim 1, characterized in that: The structured clinical status is organized using the SOAP structure format, including a set of case characteristic information and a set of diagnoses and plans; The case feature information set includes chief complaint, present illness, past medical history, physical examination, and auxiliary examinations; The diagnostic and planning suite includes preliminary diagnosis, preliminary diagnosis discussion, and treatment plan; The diagnostic and planning set remains unfilled during the state update phase until the state is frozen, at which point the content is generated by a specialist agent.

3. A multidisciplinary pre-diagnosis system based on multi-agent collaboration, characterized in that, The pre-consultation process is built and driven by structured clinical states, including: The clinical status management module is used to build, update, and store structured clinical statuses in SOAP format; The main control and scheduling module is used to dynamically select and activate specialized intelligent agents based on the current state and context; A multi-specialty intelligent agent module is used to independently analyze the current state and generate problem suggestions and state update suggestions; The semantic aggregation module is used to fuse the outputs of multiple agents and update the set of case feature information. The diagnosis generation module is used to generate diagnostic and planning information based on a set of case characteristics.

4. The system according to claim 3, wherein the clinical status management module distinguishes between case characteristic information and diagnosis and planning information, and prohibits further writing after the status is frozen.

5. The system according to claim 3, characterized in that, The master control scheduling module determines the set of specialist intelligent agents to be activated based on the patient's response.

6. The system according to claim 3, characterized in that, The specialized intelligent agent modules operate in their own independent reasoning environments and do not share intermediate reasoning processes.

7. The system according to claim 3, characterized in that, The semantic aggregation module adopts a "write-after-generate" strategy, which merges the suggested execution content from multiple agents and performs state updates, and then generates the natural language query text for the next round of patient interaction.

8. The system according to claim 3, characterized in that, The diagnosis generation module is activated only after the case feature information set is frozen, and is used to output preliminary diagnosis, proposed diagnosis discussion, and treatment plan.

9. The system according to claim 3, characterized in that, The system merges the case feature information set with the diagnosis and planning set and outputs a structured preliminary medical record (IPN) for physician decision support or medical record archiving.

Images