Rare disease information processing and consultation system and method based on multi-agent collaboration

Abstract

This invention discloses a rare disease information processing and consultation system and method based on multi-agent collaboration, belonging to the field of medical information processing technology. To address the problems of discontinuous data flow and insufficient interaction security in rare diseases, this invention dynamically maps user requests into a directed acyclic graph workflow through orchestration agents; patient agents and gene agents collaboratively extract clinical phenotype and sequence data and perform variant analysis; inference agents perform auxiliary tasks and generate inference trajectories; auditing agents detect trajectory faults in parallel and trigger correction loops; an active query module identifies missing elements and guides information completion; finally, a response integration module aggregates and verifies the conclusions and outputs a response. This invention enables dynamic orchestration, continuous maintenance, and security auditing of data in rare disease scenarios, providing highly reliable and low-latency technical support for full-cycle rare disease information consultation.

Description

Technical Field

[0001] This invention discloses a rare disease information processing and consultation system and method based on multi-agent collaboration, belonging to the field of medical information processing technology. Background Technology

[0002] Rare diseases are characterized by low incidence, complex pathogenesis, and life-threatening conditions. While individual rare diseases are extremely rare, recent estimates suggest that approximately 7,000–10,000 rare diseases affect more than 300 million people globally, posing a significant challenge to global healthcare systems. Meanwhile, the entire process of rare disease care, including diagnosis, medication advice, and long-term psychosocial support, is fragmented and resource-intensive. Furthermore, the scarcity of expertise further exacerbates the disparity in care levels, making it inadequate even for complex diseases. With advancements in large language modeling technology, multi-agent systems driven by large language models offer a promising paradigm for addressing complex needs and resource-constrained challenges in healthcare. Recent approaches attempt to integrate various medical tools and heterogeneous knowledge into multi-agent systems to enhance their robustness and clinical effectiveness in rare disease scenarios.

[0003] For example, Figure 1 The architecture diagram for RareAgents comes from the arXiv paper "RareAgents: Autonomous Multi-disciplinary Team for Rare Disease Diagnosis and Treatment." RareAgents is the first multi-agent architecture driven by a large model, integrating multidisciplinary agents, a memory system, and various tools, primarily addressing rare disease diagnosis and drug recommendation scenarios. For example, Figure 2 This is an architecture diagram of DeepRare, taken from the arXiv article "An Agentic System for Rare Disease Diagnosis with Traceable Reasoning." RareAgents is a multi-agent system driven by a large model for rare disease differential diagnosis decisions, integrating over 40 tools and state-of-the-art knowledge sources. DeepRare can generate transparent disease diagnosis rankings with verifiable medical evidence based on heterogeneous inputs such as text descriptions, structured human phenotypic ontologies, and genetic testing results.

[0004] It can be seen that existing methods are generally only applicable to isolated rare disease scenarios (e.g., diagnostic tasks, medication recommendations), lacking a comprehensive understanding of the entire care process, such as continuous monitoring of various indicators of a single patient and answering patients' questions at different clinical stages. Furthermore, existing technologies may face various obstacles in practical use, including the possibility of system-level failures and high latency in some modules. Summary of the Invention

[0005] The purpose of this invention is to address the technical problems of existing rare disease information processing solutions, such as isolated application scenarios, lack of full-process data management, incomplete interactive information, and security risks in the execution process. This invention proposes a rare disease information processing and consultation system based on multi-agent collaboration, which achieves dynamic arrangement and management of rare disease data throughout its entire lifecycle, proactive information completion, and real-time security auditing, and significantly improves the efficiency and reliability of complex medical data processing.

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

[0007] A rare disease information processing and consultation system based on multi-agent collaboration, comprising: The orchestration agent receives unstructured user query requests, maps them into structured directed acyclic graph workflows, and outputs task orchestration and scheduling instructions. The patient agent is used to parse heterogeneous medical files, extract clinical phenotype information and raw sequence information, and output standardized clinical data and gene data to be analyzed. A gene intelligence agent is used to receive the gene data to be analyzed, and combine it with the standardized clinical data to perform genomic variation analysis and pathogenicity analysis, and output gene feature text description information. An inference agent is used to perform clinical auxiliary tasks according to the task scheduling instructions, and to aggregate the standardized clinical data and the gene feature text description information, and output intermediate inference trajectories. An auditing agent is used to perform hallucination detection on the intermediate inference trajectory according to preset medical constraints, and output the verified inference conclusion by correcting the intermediate inference trajectory or reconstructing the workflow based on the detection results. The proactive inquiry module is used to analyze the missing key elements in the intermediate reasoning trajectory, generate targeted follow-up questions, and output supplementary guidance information.

[0008] Furthermore, the orchestration agent is specifically used to: parse the intent type of the user query request, match a predefined expert prior directed acyclic graph workflow, instantiate sub-agents and schedule the interactions between sub-agents, map unstructured requests into structured task flows, and output the task orchestration and scheduling instructions.

[0009] Furthermore, the patient intelligent agent is specifically used to: extract information from heterogeneous medical files using a multimodal parsing tool, extract and maintain patient dynamic context information and original sequence information, and output the standardized clinical data and the gene data to be analyzed.

[0010] Furthermore, the gene agent is specifically used to: receive the standardized clinical data and the gene data to be analyzed, perform sequence data parsing and pathogenicity analysis on the original sequence information, convert sequence variation information into text information, and output the gene feature text description information.

[0011] Furthermore, the reasoning agent is specifically used to: receive the task orchestration and scheduling instructions, execute the task nodes in the directed acyclic graph workflow in sequence with the support of the standardized clinical data and the gene feature text description information, generate verifiable intermediate reasoning steps, and output the intermediate reasoning trajectory.

[0012] Furthermore, the auditing agent is specifically used to: extract the intermediate reasoning trajectory, perform hallucination detection by combining preset medical constraints and instruction consistency rules, generate modification suggestions when an error is detected and feed them back to the corresponding agent to trigger a correction loop, and output the verified reasoning conclusion.

[0013] Furthermore, the proactive inquiry module is specifically used to: analyze the logical path in the intermediate reasoning trajectory, identify the pending indicators or phenotypic information that affect the result, generate targeted follow-up question information, and output the supplementary guidance information.

[0014] Furthermore, the orchestration agent, patient agent, gene agent, reasoning agent, and auditing agent all contain a memory cache. The memory cache is used to: store and semantically retrieve the historical input information and historical execution trajectory of each agent, realize the reuse and anchoring of the verified execution trajectory, and use the relevant historical trajectory as a reference example when the retrieval matching degree is insufficient.

[0015] Furthermore, the tools invoked by the orchestration agent, patient agent, gene agent, reasoning agent, and auditing agent include: Heterogeneous data transformation tool for report parsing and standard phenotypic terminology extraction; Knowledge enhancement tools for retrieving information from public knowledge sources and internal databases; Biological tools for assessing sequence variation and pathogenicity; The agent memory tool is used to store and retrieve the memory cache of each agent.

[0016] A rare disease information processing and consultation method based on multi-agent collaboration includes the following steps: The orchestration agent receives unstructured user query requests and maps them into a structured directed acyclic graph workflow, outputting task orchestration and scheduling instructions. By parsing heterogeneous medical files through a patient intelligent agent, clinical phenotype information and raw sequence information are extracted, and standardized clinical data and gene data to be analyzed are output. The gene agent receives the gene data to be analyzed and combines it with the standardized clinical data to perform genomic variation analysis and pathogenicity analysis, and outputs gene feature text description information. The reasoning agent executes clinical auxiliary tasks by following the task scheduling instructions and aggregating the standardized clinical data and the gene feature text description information, and outputs intermediate reasoning trajectories. The auditing agent performs hallucination detection on the intermediate inference trajectory based on preset medical constraints, and outputs the verified inference conclusion by correcting the intermediate inference trajectory or reconstructing the workflow based on the detection results. By analyzing the missing key elements in the intermediate reasoning trajectory through the proactive inquiry module, targeted follow-up questions are generated, and supplementary guidance information is output. The verified reasoning conclusion is concatenated with the supplementary guidance information and then output.

[0017] The present invention has achieved the following beneficial effects.

[0018] 1. This invention adopts a dynamic workflow orchestration mechanism in the form of a directed acyclic graph, which maps unstructured requests into structured task flows based on expert priors and real-time context, and realizes flexible adaptation and continuous data monitoring for rare disease scenarios throughout the entire cycle, such as consultation, advice, and prognostic analysis.

[0019] 2. By introducing an active inquiry module, this invention can analyze logical gaps in the intermediate reasoning trajectory in real time and initiate targeted questions to users, effectively solving the technical problem of incomplete information provided by users in the early stages of rare diseases and improving the system's ability to discover potential key indicators.

[0020] 3. This invention achieves semantic storage and reuse of time-consuming tool call trajectories by setting up a memory cache for patient isolation and combining it with intelligent agent memory tools. While ensuring data privacy, it significantly reduces system computational overhead and improves response efficiency in scenarios requiring continuous information maintenance.

[0021] 4. This invention integrates an online auditing agent and employs a large model evaluation strategy to perform real-time illusion detection and correction on intermediate inference chains and tool outputs. It can dynamically reconstruct workflows or trigger node re-execution based on the severity of errors, thereby curbing the spread of erroneous information at the source and ensuring the security and reliability of clinical auxiliary information output. Attached Figure Description

[0022] Figure 1 This is a diagram of the existing RareAgents technology architecture. Figure 2 This is a diagram of the existing DeepRare architecture. Figure 3 This is an architecture diagram of a rare disease information processing and consultation system based on multi-agent collaboration in the embodiment. Figure 4 This is the medical science popularization consultation workflow in the example; Figure 5 This is a flowchart of the proactive query mechanism in the embodiment; Figure 6 This is a flowchart of the online audit process in the embodiment; Figure 7 This is a flowchart of the memory caching process in the embodiment; Figure 8 This is a flowchart of a rare disease information processing and consultation method based on multi-agent collaboration in an embodiment. Detailed Implementation

[0023] To make the various technical features, advantages, or effects of the present invention more apparent and understandable, detailed descriptions are provided below through embodiments.

[0024] This invention provides a rare disease information processing and consultation system (RareDASH) based on multi-agent collaboration, such as... Figure 3 As shown, the system mainly includes an orchestration agent, a patient agent, a gene agent, a reasoning agent, an auditing agent, and an active inquiry module. For user queries input in natural language, the system breaks down the reasoning and response process into multiple agent calls, with each call's input and output in text format for easy traceability. Each agent has a corresponding set of callable tools.

[0025] 1. Arrange intelligent agents.

[0026] The orchestration agent, as the core module of the system, is responsible for parsing the initial user query request and determining the optimal task execution path. Based on a pre-designed directed acyclic graph (DAG) workflow, the orchestration agent instantiates relevant sub-agents and schedules the interactions between them. By mapping unstructured user requests to structured DAG workflows, such as scenarios involving symptom checks and medication recommendations, the system ensures that the execution process conforms to established medical standards, rather than relying solely on the generation of unstructured and free-form text.

[0027] End-to-end rare disease data processing involves various clinical scenarios, each corresponding to different agent collaboration paradigms. For common service scenarios such as public health consultation, data-assisted analysis, suggestion generation, and prognostic assessment, this invention utilizes expert prior knowledge to construct a predefined Directed Acyclic Graph (DAG) workflow for structured management. In this graph structure, each node represents an instantiated sub-agent, including orchestration agents, patient agents, gene agents, and reasoning agents, while directed edges represent information propagation paths between agents. This approach prevents information exchange between agents from entering infinite loops. For different types of user consultations, the expert-designed workflow can serve as high-quality reference input to provide context for the orchestration agent.

[0028] The following is a workflow example, such as Figure 4 As shown. Let q be the current user's question, i be the user's intent type, df(i) be the text-based workflow, and I be the pre-built intent set. For Figure 4 In the medical science popularization consultation workflow, a dynamic workflow representation for context is df(science popularization consultation) = "from statement understanding to web page search; from statement understanding to knowledge base query; from web page search to report generation; from knowledge base query to report generation". This representation represents... Figure 4 The four directed edges in the lower half. Specifically, for consultation types different from the predefined intent, the context learning capability of the orchestration agent can be utilized to dynamically generate a new workflow based on the predefined workflow, as shown in the following formula: in, This indicates a new workflow dynamically generated by the system based on the current input; Orchestrator represents the orchestration agent that executes the scheduling and generation logic. For the question or query request entered by the current user, Indicates the type of user intent. Indicates a specific type of intent The corresponding predefined workflow in text format, This represents a pre-built set of intents, while This represents the set of known intentions, consisting of intentions and their corresponding workflows, which serve as a context learning reference for the orchestration agent.

[0029] 2. The patient's intelligent agent.

[0030] The patient agent is specifically designed to parse and manage patient-related data. It possesses multimodal parsing capabilities, handling various uploaded medical document types, including clinical reports and electronic health records. Utilizing Optical Character Recognition (OCR) and information extraction tools such as MinerU, the agent parses these heterogeneous documents into text-based reports, which are then provided to downstream agents. Furthermore, the patient agent is responsible for continuously maintaining the dynamic context of the patient's status, encompassing patient identification information, key vital signs and indicators, a list of main records, and laboratory test results.

[0031] 3. Genetic intelligence.

[0032] The gene agent acquires relevant data from the patient agent, aiming to process genomic data and transform complex sequence data into readable text. Variant Call Format (VCF) is a standardized format used in bioinformatics to store gene sequence variations. The gene agent primarily executes VCF-related instructions by calling various gene-related tools, such as pathogenicity analysis and genotype-based auxiliary analysis, as a supplementary method to phenotypic analysis.

[0033] 4. Reasoning agent.

[0034] The reasoning agent is responsible for performing core clinical support operations, including critical tasks such as differential analysis, suggestion generation, and post-intervention evaluation. This agent handles all tasks except task orchestration, patient data parsing, and genomic data parsing. Guided by the orchestration agent, the reasoning agent executes assigned Directed Acyclic Graph (DAG) workflow steps sequentially. Simultaneously, the reasoning agent generates intermediate inference steps in a verifiable text format to ensure the system's interpretability and suitability for human validation.

[0035] 5. Proactive Inquiry Module.

[0036] Relying passively on patient descriptions often leads to incomplete analysis, especially considering the typically incomplete user-side information and the diverse manifestations of rare diseases. To address this issue, this system introduces an active inquiry module, aiming to transform the system from a passive responder to an active investigator. This active inquiry module is an independent functional module that runs in parallel with the main inference process of the multi-agent system. When existing information is insufficient to draw reliable conclusions, this mechanism prompts the agents to dynamically propose targeted follow-up questions and iteratively acquire potential patient information to clarify ambiguous parts. In this module, the system inputs the current inference trajectory into a large language model and uses specific instructions to obtain active inquiry questions. Finally, the system integrates the workflow output of the current inquiry with the generated inquiry questions to synthesize the final system response.

[0037] Figure 5 The process of the proactive inquiry mechanism is demonstrated, emphasizing the integration of proactive inquiry with inference results. The specific steps of this process are as follows: (1) Original input: The process begins with the reasoning trajectory recorded as the thinking of an agent or model.

[0038] (2) Thinking and decomposition: The reasoning trajectory is decomposed into two parallel dimensions through the thinking process of the large model, namely, extracting the important information that affects the result and obtaining the preliminary reasoning result.

[0039] (3) Proactive Inquiry Construction: Proactive inquiry content is generated using the large model generation capability to generate important information on the extracted impact results, aiming to guide and optimize the interaction.

[0040] (4) Information integration: The system integrates the generated proactive inquiries with the previously obtained reasoning results to ensure that the response content is both conclusive and guiding.

[0041] (5) Final response generation: The system generates the final response based on the integrated information flow and uses it as the final output.

[0042] 6. Auditing agent.

[0043] Collaborative chains in multi-agent systems are susceptible to the propagation of hallucinations, where minor errors in early reasoning can lead to serious clinical problems. To address this, an online auditing agent is instantiated to detect faults or hallucinations in the workflow execution of all other agents in real time. This auditing agent runs in parallel with the main workflow process and employs strict clinical standards as auditing rules. The auditing agent uses a Large Language Model (LLM) as the evaluator strategy, checking the correctness of intermediate reasoning steps and tool outputs based on predefined medical constraints and hallucination types related to instruction consistency. Detected errors are intercepted before propagating to the final output. Depending on the severity and type of the error, the auditing agent modifies the Directed Acyclic Graph (DAG) workflow, including re-executing specific nodes and dynamically reconstructing the entire graph, effectively guiding the reasoning process back to the correct trajectory. If a potential error is detected, the auditing agent immediately intervenes to trigger a correction loop and generate modification suggestions, effectively reducing the risk of error propagation and ensuring the safety of the final response.

[0044] Figure 6 The closed-loop process of online auditing and automatic optimization of agent inference trajectories is demonstrated, with the following specific steps: (1) Initial generation: The original reasoning trajectory is generated by the agent to be evaluated; (2) Data input: The generated reasoning trajectory, together with the preset medical constraint hallucination type and instruction consistency hallucination type, are used as input information and incorporated into the subsequent audit process; (3) Online auditing: The online auditing agent receives the above inputs and uses LLM evaluation technology to audit the inference trajectory; (4) Error detection: The system determines whether there are errors in the trajectory based on the audit results; (6) Output of results: If the result is negative, that is, there is no error in the trajectory, the inferred trajectory will be output as the final result directly; (7) Feedback optimization: If the judgment result is yes, that is, there is an error in the trajectory, the system will generate modification suggestions in a targeted manner and feed them back to the agent to be evaluated for regeneration, thereby forming a closed loop of cyclic optimization.

[0045] 7. Callable tools.

[0046] The aforementioned intelligent agent uses tools from the tool pool as needed. In this embodiment, a tool list must first be defined. Taking the Human Phenotype Ontology (HPO) terminology extraction tool as an example, HPO terminology is a vocabulary system used to standardize the description of human phenotypic characteristics, such as symptoms, signs, or abnormal manifestations. For example, manifestations such as epileptic seizures and intellectual disabilities have corresponding standard HPO codes. The function of this tool is to convert non-standardized free text into unified HPO terminology to facilitate subsequent medical research or clinical applications. The specific invocation process is as follows: (1) The large language model identifies the HPO terms that need to be standardized for the current task and decides to call the get_phenotype_information tool; (2) The large language model prepares parameters according to the format defined by the tool and inputs the original text to be processed as a string; (3) After receiving the call request, the external tool runs the internal HPO terminology extraction logic to identify and output standard HPO terms from the input text; (4) The external tool returns the extraction results to the large language model for use in subsequent processing.

[0047] For the input subquery, the agent invokes the large language model and includes a list of callable tools in the context. When the large language model determines that external tool support is required, it returns JSON output containing the tool name and parameters. Upon receiving the tool invocation request, the system runs the tool according to its type and parameters and obtains the execution result. This execution result, along with the input subquery, is fed into the base large language model to generate the final natural language response.

[0048] This system includes the following four types of tools: The first category is heterogeneous data conversion tools, including medical report scanning, PDF parsing, structured table data extraction, and free text HPO terminology extraction tools, etc. The second category is knowledge enhancement tools, including knowledge retrieval tools from public knowledge sources such as PubMed, HPO, the Online Mendelian Inheritance in Man (OMIM) database and search interfaces (APIs), as well as internal databases such as patient case records and knowledge graphs. The third category consists of biological tools, including pathogenicity assessment, transcript matching, nonsense-mediated decay prediction, and amino acid substitution effect assessment tools. The fourth category is intelligent agent memory tools, which include storing and retrieving the memory caches of each intelligent agent.

[0049] 8. Memory cache.

[0050] The computational overhead and time latency associated with multi-agent collaboration can hinder real-time clinical interaction, especially steps involving numerous tool calls, such as... Figure 7 The query for "c.236C>T, what diseases might LP cause?" requires calling time-consuming gene analysis tools. To address this efficiency bottleneck, this invention equips each agent with a memory cache to reuse the agent's past trajectories and experiences. This cache stores the original inputs and inference trajectories of past executions as keys and the execution results as values. Formally, this module saves each successful agent execution to the corresponding structured cache. When encountering repetitive subtasks or similar clinical queries, the system retrieves the semantically most similar key from the cache.

[0051] Figure 7 The decision-making process of a genetic agent using a memory module for efficient retrieval and reasoning is described, with the specific steps as follows: (1) Initiate a query: The system receives new query input, such as a question about a disease that may be triggered by a specific gene locus mutation c.236C>T,LP; (2) Memory retrieval: The retrieval is initiated by the gene agent in the memory module, which stores historical call records, including gene mutation sites as keys, corresponding disease evidence as values, and retrieval scores; (3) Consistency determination: The system extracts the first-ranked item in the search results and evaluates whether its search consistency score is greater than 0.9; (4) Fast path: If the determination result is that the retrieval consistency score is greater than 0.9, the system directly uses the query value stored in memory to replace the agent's regular execution steps in order to achieve a fast response; (5) Enhanced path: If the determination result is that the retrieval consistency score is not greater than 0.9, the system will use the first five relevant key value pairs retrieved as supplementary context demonstration to assist the agent in making more accurate reasoning. (6) Agent output: The final output includes detailed assessment conclusions and sources of evidence, such as mucopolysaccharidosis type I and the expert consensus referenced.

[0052] To maintain data privacy, the memory cache is strictly independent for different patient identities. In summary, this mechanism effectively reduces computational overhead and significantly lowers latency by anchoring agent inference to validated trajectories.

[0053] This invention also provides a method for rare disease information processing and consultation based on multi-agent collaboration, such as... Figure 8 As shown, it includes the following steps: Step S1: Receive unstructured user query requests through the orchestration agent, map them into a structured directed acyclic graph workflow, and output task orchestration and scheduling instructions. Step S2: The patient agent parses heterogeneous medical files to extract clinical phenotype information and raw sequence information, and outputs standardized clinical data and gene data to be analyzed. Step S3: Receive the gene data to be analyzed through the gene agent, and combine it with the standardized clinical data to perform genomic variation analysis and pathogenicity analysis, and output gene feature text description information; Step S4: The reasoning agent executes clinical auxiliary tasks according to the task scheduling instructions, and aggregates the standardized clinical data and the gene feature text description information, and outputs the intermediate reasoning trajectory. Step S5: The auditing agent performs hallucination detection on the intermediate inference trajectory according to the preset medical constraints, and outputs the verified inference conclusion by correcting the intermediate inference trajectory or reconstructing the workflow based on the detection results. Step S6: Analyze the missing key elements in the intermediate reasoning trajectory through the active inquiry module, generate targeted follow-up questions, and output supplementary guidance information; Step S7: The verified reasoning conclusion is concatenated with the supplementary guidance information and then output.

[0054] The following is a specific application example provided by the present invention.

[0055] In this example, the user's question was "How can I treat my illness?" and they uploaded a medical report. The information flow processed by the system is as follows.

[0056] 1. The orchestration agent performs workflow orchestration: its input is the user's question "How can I treat my illness?". The output is a determination of the intent type of the question as "medical consultation", and a corresponding directed acyclic graph (DAG) workflow is constructed accordingly.

[0057] 2. The patient's AI agent executes query comprehension: its input consists of the user's question and the uploaded examination report. By calling the MinerU document information extraction tool, the output is parsed results. Specifically, the patient's actual question is identified as "how to treat otitis media," confirming that the user is seeking comprehensive treatment information, clinical guidelines, and evidence-based recommendations for middle ear infections (otitis media). Key medical terms identified are "otitis media, treatment, antibiotics," and the generated query statements include: "-WebSearch: Otitis Media Treatment Guidelines" and "-PubMed: Otitis Media Treatment."

[0058] 3. The reasoning agent performs a web search: its input is "treatment guidelines for otitis media". It retrieves and outputs the corresponding web search results by calling a web search API tool.

[0059] 4. The reasoning agent performs a knowledge base query: its input is "treatment of otitis media". It retrieves and outputs the corresponding medical literature query results by calling the PubMed API tool.

[0060] 5. The reasoning agent performs report generation: its inputs are the patient's question, web search results, and knowledge base query results. The output reasoning conclusion is: "According to the latest guidelines and research, the treatment of otitis media involves a combination of antibiotics, supportive care, and preventive measures. Treatment methods may include:..."

[0061] 6. The proactive inquiry module performs proactive inquiry: its input is the above reasoning trajectory. The output recognition result is: "Treatment recommendations need to be provided based on the patient's age and severity."

[0062] 7. The response integration module generates the final response: The system integrates the above reasoning conclusions with the generated subsequent questions, and outputs the final response as follows: "According to the latest guidelines and research, the treatment of otitis media involves a combination of antibiotics, supportive care, and preventive measures. Treatment methods may include: ... You can provide your age and current condition, and I will provide you with targeted treatment recommendations." Although the present invention has been disclosed above with reference to embodiments, it is not intended to limit the present invention. Appropriate modifications or equivalent substitutions made by those skilled in the art to the technical solutions of the present invention should be covered within the protection scope of the present invention, which is defined by the claims.

Claims

1. A rare disease information processing and consultation system based on multi-agent collaboration, characterized in that, include: The orchestration agent receives unstructured user query requests, maps them into structured directed acyclic graph workflows, and outputs task orchestration and scheduling instructions. The patient agent is used to parse heterogeneous medical files, extract clinical phenotype information and raw sequence information, and output standardized clinical data and gene data to be analyzed. A gene intelligence agent is used to receive the gene data to be analyzed, and combine it with the standardized clinical data to perform genomic variation analysis and pathogenicity analysis, and output gene feature text description information. An inference agent is used to perform clinical auxiliary tasks according to the task scheduling instructions, and to aggregate the standardized clinical data and the gene feature text description information, and output intermediate inference trajectories. An auditing agent is used to perform hallucination detection on the intermediate inference trajectory according to preset medical constraints, and output the verified inference conclusion by correcting the intermediate inference trajectory or reconstructing the workflow based on the detection results. The proactive inquiry module is used to analyze the missing key elements in the intermediate reasoning trajectory, generate targeted follow-up questions, and output supplementary guidance information.

2. The system of claim 1, wherein, The orchestration agent is specifically used to: parse the intent type of the user query request, match the predefined expert prior directed acyclic graph workflow, instantiate sub-agents and schedule the interaction between sub-agents, map unstructured requests into structured task flows, and output the task orchestration and scheduling instructions.

3. The system of claim 1, wherein, The patient intelligent agent is specifically used to: extract information from heterogeneous medical files using a multimodal parsing tool, extract and maintain patient dynamic context information and original sequence information, and output the standardized clinical data and the gene data to be analyzed.

4. The system of claim 1, wherein, The gene agent is specifically used to: receive the standardized clinical data and the gene data to be analyzed, perform sequence data parsing and pathogenicity analysis on the original sequence information, convert sequence variation information into text information, and output the gene feature text description information.

5. The system of claim 1, wherein, The reasoning agent is specifically used to: receive the task orchestration and scheduling instructions, and with the support of the standardized clinical data and the gene feature text description information, execute the task nodes in the directed acyclic graph workflow in sequence, generate verifiable intermediate reasoning steps, and output the intermediate reasoning trajectory.

6. The system of claim 1, wherein, The auditing agent is specifically used to: extract the intermediate reasoning trajectory, perform hallucination detection by combining preset medical constraints and instruction consistency rules, generate modification suggestions when an error is detected and feed them back to the corresponding agent to trigger a correction loop, and output the verified reasoning conclusion.

7. The system of claim 1, wherein, The proactive inquiry module is specifically used to: analyze the logical path in the intermediate reasoning trajectory, identify the undetermined indicators or phenotypic information that affect the result, generate targeted follow-up question information, and output the supplementary guidance information.

8. The system of claim 1, wherein, The orchestration agent, patient agent, gene agent, reasoning agent, and audit agent all contain a memory cache. The memory cache is used to: store and semantically retrieve the historical input information and historical execution trajectory of each agent, realize the reuse and anchoring of the verified execution trajectory, and use the relevant historical trajectory as a reference example when the retrieval matching degree is insufficient.

9. The system of claim 1 or 8, wherein, The tools invoked by the orchestration agent, patient agent, gene agent, reasoning agent, and auditing agent include: Heterogeneous data transformation tool for report parsing and standard phenotypic terminology extraction; Knowledge enhancement tools for retrieving information from public knowledge sources and internal databases; Biological tools for assessing sequence variation and pathogenicity; The agent memory tool is used to store and retrieve the memory cache of each agent.

10. A method for rare disease information processing and consultation based on multi-agent collaboration, comprising the following steps: The orchestration agent receives unstructured user query requests and maps them into a structured directed acyclic graph workflow, outputting task orchestration and scheduling instructions. By parsing heterogeneous medical files through a patient intelligent agent, clinical phenotype information and raw sequence information are extracted, and standardized clinical data and gene data to be analyzed are output. The gene agent receives the gene data to be analyzed and combines it with the standardized clinical data to perform genomic variation analysis and pathogenicity analysis, and outputs gene feature text description information. The reasoning agent executes clinical auxiliary tasks by following the task scheduling instructions and aggregating the standardized clinical data and the gene feature text description information, and outputs intermediate reasoning trajectories. The auditing agent performs hallucination detection on the intermediate inference trajectory based on preset medical constraints, and outputs the verified inference conclusion by correcting the intermediate inference trajectory or reconstructing the workflow based on the detection results. By analyzing the missing key elements in the intermediate reasoning trajectory through the proactive inquiry module, targeted follow-up questions are generated, and supplementary guidance information is output. The verified reasoning conclusion is concatenated with the supplementary guidance information and then output.

Images