A clinical medicine review generation method and system based on a knowledge graph and multi-agent collaboration

Abstract

The present application relates to the cross field of artificial intelligence and medical technology, in particular to a clinical medicine review generation method and system based on knowledge graph and multi-agent cooperation, comprising the following steps: S1: task decomposition and concept prior retrieval receiving a user's complex clinical problem, decomposing it into several sub-tasks through planning agent and generating a report outline; for each sub-task, retrieve relevant core concept nodes in the graph node library, starting from the core node, preferentially traversing high-weight edges using the weighted sub-graph extraction algorithm, extracting knowledge sub-graphs related to the problem logic, and locking and recalling the original literature through the source literature ID recorded in the edge attribute of the knowledge graph, in the present application, the "illusion" risk is significantly reduced: through the innovative "decoupling verification" mechanism, the path of model fabrication reference is fundamentally cut off, and each sentence of discussion must be verified through independent evidence matching, ensuring the rigor of the medical review.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and medical technology, and in particular to a method and system for generating clinical medical reviews based on knowledge graphs and multi-agent collaboration. Background Technology

[0002] As large language models (LLMs) are increasingly applied in the medical field, doctors and researchers urgently need to leverage AI to rapidly process massive amounts of medical literature and answer complex clinical questions. However, general-purpose LLMs suffer from a serious "illusion" problem and struggle to handle long-tail, specialized medical knowledge. Furthermore, clinical decision-making requires strict evidence-based logic and cannot be based solely on probability.

[0003] Current mainstream solutions typically employ a standard RAG (Retrieval Augmentation and Generation) technical architecture, which includes: a data preprocessing stage where medical literature is segmented into text blocks, vectors are calculated, and stored in a vector database; a retrieval stage where the user's input natural language question is transformed into a query vector, and the top-K most similar text blocks are retrieved from the database; and a generation stage where the retrieved text blocks are directly concatenated as context and input into an LLM to generate an answer. This type of approach is widely found in basic tutorials for open-source frameworks such as LangChain and LlamaIndex, as well as in related technical papers such as Microsoft GraphRAG.

[0004] However, existing technologies have the following drawbacks: Limited retrieval coverage: Traditional RAGs, based on semantic similarity, are prone to getting bogged down in localized information, leading to incomplete or biased reviews; Attribution illusion (citation errors): When generating long reviews, LLMs often fabricate citations for the sake of fluency, or the cited literature may not actually support the generated arguments; Lack of evidence-based logic and guideline constraints: Existing systems struggle to differentiate levels of evidence, are unable to engage in "critical thinking," and often ignore the constraints of the latest clinical guidelines; Semantic gap: There are differences between users' natural language queries and the terminology of professional literature, making it difficult for simple vector retrieval to uncover deep logical connections. Summary of the Invention

[0005] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for generating clinical medical reviews based on knowledge graphs and multi-agent collaboration.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: a method for generating clinical medical reviews based on knowledge graphs and multi-agent collaboration, comprising the following steps: S1: Task decomposition and concept-first retrieval. The system receives complex clinical questions from users, decomposes them into several sub-tasks through a planning agent, and generates a report outline. For each sub-task, the system retrieves relevant core concept nodes from the graph node library. Starting from the core nodes, the system uses a weighted subgraph extraction algorithm to prioritize traversing high-weight edges and extracts knowledge subgraphs related to the question logic. The system then uses the source document ID recorded in the edge attributes of the knowledge graph to reverse-lock and recall the original documents. S2: Based on the breadth and completeness check analysis of the extracted knowledge subgraph topology, identify the neighboring nodes closely connected to the core concepts and generate the expected knowledge list; scan the content of the currently retrieved documents to check their coverage of the key points of the expected knowledge list and generate a document breadth review report; if it is found that the documents do not cover the key nodes, a second search loop is automatically triggered to construct supplementary query instructions to fill in the missing information. S3: Dialectical Criticism and Guideline Compliance Check: The challenger agent reads the extensive review report and the initial draft, generates challenging questions based on evidence-based medicine principles, automatically searches the clinical guideline database to check whether the generated draft violates standard treatment guidelines, and performs targeted supplementary searches and corrections for challenging questions and guideline conflicts. S4: Decoupled Generation and Atomic-Level Attribution Validation Stage A: The generation model writes a review based on the retrieved content, outputting only plain text content without any citation markers; Stage B: Natural language processing tools are used to decompose the plain text into several independent fact statement units; Stage C: For each fact statement, a local semantic similarity search is performed in the retrieved literature pool to construct a micro-evidence context specific to that statement; Stage D: The validation model is called to determine whether the micro-context supports the statement. If it does, the source literature ID is accurately inserted; if it does not support it or there are errors, the statement text is automatically corrected or the statement is deleted; Stage E: The validated statements are recombined to generate a final clinical medical review containing accurate citations; The edge attributes of the knowledge graph include the number of evidence and the total number of citations, which are used to quantify the credibility of the knowledge; the high-weight edges are determined based on the number of evidence and the total number of citations.

[0007] Preferably, in step S1, the core concept nodes include entities extracted from medical literature such as diseases, drugs, and MeSH terms.

[0008] Preferably, in step S2, the expected knowledge list includes key topics related to the core concept, such as adverse reactions, contraindications, and efficacy evaluation.

[0009] Preferably, in step S3, the standard treatment guidelines include contraindications, recommended dosages, and other clinical guidelines.

[0010] Preferably, in step S4, the source document ID is marked in the format [CITE_PMID:...].

[0011] A clinical medical review generation system based on knowledge graph and multi-agent collaboration includes a data storage module for maintaining a document vector library, a graph node library, and a knowledge graph. The document vector library stores vector representations of the titles, abstracts, and metadata of medical documents. The graph node library stores vector representations of entities extracted from medical documents and their relationships. The knowledge graph stores the topological structure between entities, with edge attributes including the number of evidence and the total number of citations. The multi-agent module includes a planning agent and a challenger agent; the planning agent is used to decompose complex clinical problems into sub-tasks and generate report outlines; the challenger agent is used to generate challenging questions based on evidence-based medicine principles. The retrieval module is used to perform concept-first retrieval and secondary retrieval, including retrieving core concept nodes from the knowledge graph node library, recalling literature based on knowledge graph traversal, and targeted completion retrieval. The breadth-of-knowledge review module is used to analyze the topology of the knowledge subgraph, generate a list of expected knowledge, calculate the literature coverage, and generate a breadth-of-knowledge review report. The compliance check module is used to search the clinical guideline database and check the consistency between drafts and standard treatment guidelines; The decoupled generation and verification module is used to perform plain text generation, fact statement decomposition, micro-evidence context construction, statement verification and correction, and final summary assembly.

[0012] Preferably, in the data storage module, the knowledge graph is implemented using NetworkX, JSON, Neo4j, NebulaGraph, or TigerGraph.

[0013] Preferably, the verification model in the decoupled generation and verification module adopts a large language model or a BERT or RoBERTa classification model fine-tuned based on natural language inference tasks.

[0014] Preferably, the system employs a state machine architecture to manage the reasoning process, and the state machine architecture is a task scheduler based on a directed acyclic graph.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows: In this invention, the risk of "illusion" is significantly reduced: through an innovative "decoupled verification" mechanism, the path of model fabrication and citation is fundamentally cut off, and every statement must be independently verified by evidence matching, ensuring the rigor of the medical review; It greatly improves the comprehensiveness of the answers: by introducing knowledge graphs as a "God's-eye view" to conduct a breadth check, the system can automatically discover information blind spots and actively fill in the gaps, avoiding the one-sidedness of traditional RAGs that "say whatever is found"; It conforms to the strict logic of evidence-based medicine: the system simulates the evidence-based thinking of human experts by calculating edge weights (number of evidence + citation impact) and constraining the guideline library, giving priority to adopting high-level evidence to ensure the scientific nature and safety of the recommendations; Improved recall of long-tail knowledge: The "concept-first" graph retrieval strategy can discover documents that do not match keywords but are logically closely related, overcoming the limitations of traditional vector retrieval in matching professional terms. Attached Figure Description

[0016] Figure 1 This invention proposes an adaptive RAG panoramic data flow architecture diagram for a clinical medical review generation method and system based on knowledge graph and multi-agent collaboration. Figure 2 This invention proposes a dual-processing verification and cleaning logic diagram for a clinical medical review generation method and system based on knowledge graph and multi-agent collaboration. Figure 3 This invention proposes a graph-driven breadth-of-access inspection and reverse constraint mechanism graph for generating clinical medical reviews based on knowledge graphs and multi-agent collaboration. Detailed Implementation

[0017] 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.

[0018] Please see Figure 1-3 This invention provides a technical solution: a method for generating clinical medical reviews based on knowledge graphs and multi-agent collaboration, comprising the following steps: S1: Task decomposition and concept-first retrieval. The system receives complex clinical questions from users, decomposes them into several sub-tasks through a planning agent, and generates a report outline. For each sub-task, the system retrieves relevant core concept nodes from the graph node library. Starting from the core nodes, the system uses a weighted subgraph extraction algorithm to prioritize traversing high-weight edges and extracts knowledge subgraphs related to the question logic. The system then uses the source document ID recorded in the edge attributes of the knowledge graph to reverse-lock and recall the original documents. S2: Based on the breadth and completeness check analysis of the extracted knowledge subgraph topology, identify the neighboring nodes closely connected to the core concepts and generate the expected knowledge list; scan the content of the currently retrieved documents to check their coverage of the key points of the expected knowledge list and generate a document breadth review report; if it is found that the documents do not cover the key nodes, a second search loop is automatically triggered to construct supplementary query instructions to fill in the missing information. S3: Dialectical Criticism and Guideline Compliance Check: The challenger agent reads the extensive review report and the initial draft, generates challenging questions based on evidence-based medicine principles, automatically searches the clinical guideline database to check whether the generated draft violates standard treatment guidelines, and performs targeted supplementary searches and corrections for challenging questions and guideline conflicts. S4: Decoupled Generation and Atomic-Level Attribution Validation Stage A: The generation model writes a review based on the retrieved content, outputting only plain text content without any citation markers; Stage B: Natural language processing tools are used to decompose the plain text into several independent fact statement units; Stage C: For each fact statement, a local semantic similarity search is performed in the retrieved literature pool to construct a micro-evidence context specific to that statement; Stage D: The validation model is called to determine whether the micro-context supports the statement. If it does, the source literature ID is accurately inserted; if it does not support it or there are errors, the statement text is automatically corrected or the statement is deleted; Stage E: The validated statements are recombined to generate a final clinical medical review containing accurate citations; The edge attributes of the knowledge graph include the number of evidence and the total number of citations, which are used to quantify the credibility of the knowledge; the high-weight edges are determined based on the number of evidence and the total number of citations.

[0019] In step S1, the core concept nodes include entities extracted from medical literature, such as diseases, drugs, and MeSH terms.

[0020] In step S2, the expected knowledge list includes key topics related to the core concepts, such as adverse reactions, contraindications, and efficacy evaluation.

[0021] In step S3, the standard treatment guidelines include contraindications, recommended dosages, and other clinical guidelines.

[0022] In step S4, the source document ID is marked in the format [CITE_PMID:...].

[0023] A clinical medical review generation system based on knowledge graph and multi-agent collaboration includes a data storage module for maintaining a document vector library, a graph node library, and a knowledge graph. The document vector library stores vector representations of the titles, abstracts, and metadata of medical documents. The graph node library stores vector representations of entities extracted from medical documents and their relationships. The knowledge graph stores the topological structure between entities, with edge attributes including the number of evidence and the total number of citations. The multi-agent module includes a planning agent and a challenger agent; the planning agent is used to decompose complex clinical problems into sub-tasks and generate report outlines; the challenger agent is used to generate challenging questions based on evidence-based medicine principles. The retrieval module is used to perform concept-first retrieval and secondary retrieval, including retrieving core concept nodes from the knowledge graph node library, recalling literature based on knowledge graph traversal, and targeted completion retrieval. The breadth-of-knowledge review module is used to analyze the topology of the knowledge subgraph, generate a list of expected knowledge, calculate the literature coverage, and generate a breadth-of-knowledge review report. The compliance check module is used to search the clinical guideline database and check the consistency between drafts and standard treatment guidelines; The decoupled generation and verification module is used to perform plain text generation, fact statement decomposition, micro-evidence context construction, statement verification and correction, and final summary assembly.

[0024] In the data storage module, the knowledge graph is implemented using NetworkX, JSON, Neo4j, NebulaGraph, or TigerGraph.

[0025] The verification model in the decoupled generation and verification module adopts a large language model or a BERT or RoBERTa classification model fine-tuned based on natural language inference tasks.

[0026] The system employs a state machine architecture to manage the reasoning process, and the state machine architecture is a task scheduler based on a directed acyclic graph.

[0027] This embodiment presents a clinical medical review generation system based on knowledge graphs and multi-agent collaboration. The data storage module uses NetworkX to implement a lightweight knowledge graph structure, and the literature vector library and graph node library are stored using the FAISS vector database. The generation model in the decoupled generation and verification module uses GPT-4, and the verification model uses the RoBERTa model based on NLI task fine-tuning. The system uses a task scheduler based on a directed acyclic graph to manage state transitions. In actual operation, when a user inputs the clinical question "Analyze the efficacy and safety of drug A in patients with advanced lung cancer," the system executes the following process: The planning agent decomposes the question into three sub-tasks: "Efficacy data of drug A," "Safety data of drug A," and "Characteristics of the applicable population for patients with advanced lung cancer," generating corresponding report outlines; The retrieval module searches the knowledge graph node library for core concept nodes of "drug A" and "advanced lung cancer," traverses the knowledge graph from the core nodes, prioritizes edges with ≥5 pieces of evidence and ≥100 total citations, extracts knowledge subgraphs containing nodes such as "efficacy," "adverse reactions," "contraindications," and "applicable population," and recalls 20 relevant original documents through the source_pmid attribute of the edges; The breadth inspection module analyzes the knowledge subgraph, generates an expected knowledge list, including key points such as "objective response rate," "progression-free survival," "incidence of serious adverse reactions," and "contraindications for patients with hepatic or renal insufficiency." After scanning the 20 retrieved documents, it finds that the content related to "contraindications for patients with hepatic or renal insufficiency" is missing, triggering a secondary search to complete 3 relevant documents. After the challenger agent reads the initial draft, it generates challenging questions such as "Is the difference in efficacy among patients with advanced lung cancer of different pathological types clear?" and "Are there literature supports measures for managing adverse reactions?" The compliance check module searches the latest lung cancer diagnosis and treatment guidelines and finds that the recommended dosage of drug A in the draft is consistent with the guidelines and there is no conflict. The decoupled generation and verification module first generates a plain text review without citations, then breaks it down into 32 fact statements, constructs a micro-evidence context for each statement, and after verification, 29 statements are supported by valid citations, 2 statements are corrected, and 1 statement is deleted due to lack of evidence support. Finally, the modules are combined to generate a clinical medical review containing accurate [CITE_PMID:...] citations.

[0028] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for generating clinical medical reviews based on knowledge graphs and multi-agent collaboration, characterized in that, Includes the following steps: S1: Task decomposition and concept-first retrieval: Receives complex clinical questions from users, decomposes them into several sub-tasks through a planning agent, and generates a report outline; For each subtask, relevant core concept nodes are retrieved in the graph node library. Starting from the core nodes, the weighted subgraph extraction algorithm is used to traverse high-weight edges first to extract knowledge subgraphs related to the problem logic. The original documents are then retrieved by using the source document IDs recorded in the edge attributes of the knowledge graph. S2: Based on the breadth and completeness check analysis of the extracted knowledge subgraph topology, identify the neighboring nodes closely connected to the core concepts and generate the expected knowledge list; scan the content of the currently retrieved documents to check their coverage of the key points of the expected knowledge list and generate a document breadth review report; if it is found that the documents do not cover the key nodes, a second search loop is automatically triggered to construct supplementary query instructions to fill in the missing information. S3: Dialectical Criticism and Guideline Compliance Check: The challenger agent reads the extensive review report and the initial draft, generates challenging questions based on evidence-based medicine principles, automatically searches the clinical guideline database to check whether the generated draft violates standard treatment guidelines, and performs targeted supplementary searches and corrections for challenging questions and guideline conflicts. S4: Decoupled Generation and Atomic-Level Attribution Validation Stage A: The generation model writes a review based on the retrieved content, outputting only plain text content without any citation markers; Stage B: Natural language processing tools are used to decompose the plain text into several independent fact statement units; Stage C: For each fact statement, a local semantic similarity search is performed in the retrieved literature pool to construct a micro-evidence context specific to that statement; Stage D: The validation model is called to determine whether the micro-context supports the statement. If it does, the source literature ID is accurately inserted; if it does not support it or there are errors, the statement text is automatically corrected or the statement is deleted; Stage E: The validated statements are recombined to generate a final clinical medical review containing accurate citations; The edge attributes of the knowledge graph include the number of evidence and the total number of citations, which are used to quantify the credibility of the knowledge; the high-weight edges are determined based on the number of evidence and the total number of citations.

2. The clinical medical review generation method based on knowledge graph and multi-agent collaboration according to claim 1, characterized in that: In step S1, the core concept nodes include entities extracted from medical literature, such as diseases, drugs, and MeSH terms.

3. The clinical medical review generation method based on knowledge graph and multi-agent collaboration according to claim 1, characterized in that: In step S2, the expected knowledge list includes key topics related to the core concepts, such as adverse reactions, contraindications, and efficacy evaluation.

4. The method for generating clinical medical reviews based on knowledge graphs and multi-agent collaboration according to claim 1, characterized in that: In step S3, the standard treatment guidelines include contraindications, recommended dosages, and other clinical guidelines.

5. The clinical medical review generation method based on knowledge graph and multi-agent collaboration according to claim 1, characterized in that: In step S4, the source document ID is marked in the format [CITE_PMID:...].

6. A clinical medical review generation system based on knowledge graphs and multi-agent collaboration, characterized in that: It includes a data storage module for maintaining a document vector library, a graph node library, and a knowledge graph; the document vector library stores vector representations of the titles, abstracts, and metadata of medical documents; the graph node library stores vector representations of entities and their relationships extracted from medical documents; and the knowledge graph stores the topological structure between entities, with edge attributes including the number of evidence and the total number of citations. A multi-agent module includes a planning agent and a challenger agent; the planning agent is used to decompose complex clinical problems into sub-tasks and generate report outlines; The challenger agent is used to generate challenging questions based on evidence-based medicine principles; The retrieval module is used to perform concept-first retrieval and secondary retrieval, including retrieving core concept nodes from the knowledge graph node library, recalling literature based on knowledge graph traversal, and targeted completion retrieval. The breadth-of-knowledge review module is used to analyze the topology of the knowledge subgraph, generate a list of expected knowledge, calculate the literature coverage, and generate a breadth-of-knowledge review report. The compliance check module is used to search the clinical guideline database and check the consistency between drafts and standard treatment guidelines; The decoupled generation and verification module is used to perform plain text generation, fact statement decomposition, micro-evidence context construction, statement verification and correction, and final summary assembly.

7. The clinical medical review generation system based on knowledge graph and multi-agent collaboration according to claim 6, characterized in that: In the data storage module, the knowledge graph is implemented using NetworkX, JSON, Neo4j, NebulaGraph, or TigerGraph.

8. The method and system for generating clinical medical reviews based on knowledge graphs and multi-agent collaboration as described in claim 6, characterized in that: The verification model in the decoupled generation and verification module adopts a large language model or a BERT or RoBERTa classification model fine-tuned based on natural language inference tasks.

9. The method and system for generating clinical medical reviews based on knowledge graphs and multi-agent collaboration as described in claim 6, characterized in that: The system employs a state machine architecture to manage the reasoning process, and the state machine architecture is a task scheduler based on a directed acyclic graph.

Images