A method and apparatus for AD and MCI multidimensional knowledge discovery

Abstract

The application provides an AD and MCI multidimensional knowledge discovery method and device, comprising: obtaining a natural language question in the AD / MCI field; if the natural language question represents a multidimensional knowledge discovery intention, determining a plurality of information dimensions required for answering the natural language question and data of each information dimension through a plurality of agents; determining a basic answer corresponding to the natural language question according to the data of the information dimension with the highest importance among the data of the plurality of information dimensions; determining a correction answer based on the data of the information dimensions other than the information dimension with the highest importance among the plurality of information dimensions; and determining a final answer corresponding to the natural language question according to the basic answer and the correction answer. The application improves the efficiency and quality of AD / MCI multidimensional knowledge discovery.

Description

Technical Field

[0001] This application relates to the technical field of neuroscience, and more specifically, to a method and apparatus for discovering multidimensional knowledge of AD and MCI. Background Technology

[0002] Currently, Alzheimer's disease (AD) and mild cognitive impairment (MCI), as typical representatives of neurodegenerative diseases, impose a heavy burden on patients, families, and society. However, our understanding of the risk factors, mechanisms of occurrence, and intervention methods for AD / MCI remains significantly insufficient, and the field urgently needs efficient knowledge discovery tools to accelerate related research. In particular, in scenarios requiring the fusion of multimodal data for multidimensional knowledge reasoning, while current large-parameter language models possess strong generalization and expressive capabilities for AD / MCI knowledge discovery, their massive computational overhead leads to high inference latency and resource consumption, making it difficult to meet the efficiency and deployment cost requirements of applications such as AD / MCI multidimensional knowledge discovery. Furthermore, their black-box nature limits the credibility and interpretability of the results; and finally, when applied to specific domains, they also exhibit significant illusion problems, resulting in insufficient accuracy of the results. While lightweight models (including large language models with small parameter sizes) are efficient, they struggle to effectively model the complex semantic relationships between cross-modal data (such as text, images, and structured metrics) and support deep analysis processes involving multi-turn reasoning and multi-dimensional knowledge integration. Therefore, current knowledge discovery processes in the AD / MCI field struggle to balance high efficiency and high quality in multimodal fusion and multi-dimensional knowledge reasoning scenarios. Summary of the Invention

[0003] The purpose of this application is to provide an AD and MCI multidimensional knowledge discovery method and apparatus to improve the efficiency and quality of current AD / MCI multidimensional knowledge discovery.

[0004] This application provides a method for multidimensional knowledge discovery in AD and MCI, comprising: acquiring a natural language question in the AD / MCI domain; if the natural language question represents a multidimensional knowledge discovery intent, determining multiple information dimensions required to answer the natural language question and data for each information dimension through multiple agents; determining a basic answer corresponding to the natural language question based on the data of the information dimension with the highest importance among the multiple information dimensions; determining a corrected answer based on the data of the other information dimensions besides the information dimension with the highest importance among the multiple information dimensions; and determining the final answer corresponding to the natural language question based on the basic answer and the corrected answer.

[0005] Optionally, before determining the multiple information dimensions required to answer the natural language question and the data for each information dimension by multiple agents, the method further includes: analyzing the natural language question by planning agents to determine the required expert agent type and data agent type; constructing the knowledge discovery team based on the expert agent type and the data agent type, the knowledge discovery team being used to collaboratively perform multidimensional knowledge discovery tasks.

[0006] Optionally, the plurality of agents includes a question-generating agent and a process-summarizing agent. The step of determining the multiple information dimensions and data for each information dimension required to answer the natural language question through the plurality of agents includes: determining an initial information dimension question based on the natural language question through the question-generating agent; answering the initial information dimension question based on the natural language question through each agent in the knowledge discovery team, obtaining multiple information dimension answers; summarizing the multiple information dimension answers through the process-summarizing agent, obtaining the initial information dimension and its data; determining a new information dimension data question based on the initial information dimension, its data, and the natural language question through the question-generating agent; obtaining the new information dimension data answer corresponding to the new information dimension data question; and obtaining the final information dimension and its data through the process-summarizing agent based on the initial information dimension, its data, and the new information dimension data answer.

[0007] Optionally, determining the basic answer to the natural language question based on the data of the information dimension with the highest importance among the multiple information dimensions includes: the question-generating agent determining an information dimension ranking question based on the final information dimension and its data; each agent in the knowledge discovery team generating individual ranking results for multiple information dimensions based on the final information dimension, its data, and the ranking question, the individual ranking results including importance scores; the process summarizing agent ranking the multiple information dimensions by importance based on the average importance score of each information dimension, obtaining multiple information dimension ranking results; the information dimension ranking results including: information dimension ranking data and information dimension ranking answers; the information dimension ranking answers including: information dimension ranking and the average importance score of each information dimension; the question-generating agent determining the basic question based on the information dimension ranking results; and the data agent corresponding to the first-order information dimension in the information dimension ranking answering the basic question based on the data of that information dimension, obtaining the basic answer.

[0008] Optionally, determining the corrected answer based on data from information dimensions other than the information dimension with the highest importance among the plurality of information dimensions includes: determining a corrected question using the question-generating agent based on the natural language question, the basic answer, and the information dimension ranking result; determining corrected rules using data agents corresponding to information dimensions other than the data agent corresponding to the first ordered information dimension in the knowledge discovery team based on the natural language question, the basic answer, the corrected question, and the information dimension ranking result; selecting a plurality of second information dimensions other than the information dimension with the highest importance from the information dimension ranking; calculating the weights of the plurality of second information dimensions based on the average importance score of each second information dimension; and weighting and fusing the sub-corrected answers generated based on the data of each second information dimension using the corrected rules and the weights of the information dimensions to obtain the corrected answer.

[0009] Optionally, determining the final answer to the natural language question based on the basic answer and the corrected answer includes: conducting multiple rounds of argumentation on the basic answer and the corrected answer through each Agent in the knowledge discovery team; in each round of argumentation, each Agent performs consistency verification on the basic answer and the corrected answer according to preset argumentation rules; when the basic answer and the corrected answer pass the consistency verification, the weighted fusion result of the basic answer and the corrected answer is determined as the final answer; when the basic answer and the corrected answer fail the consistency verification, the weight of the information dimension is adjusted based on the result of this round of argumentation, and the process of conducting multiple rounds of argumentation on the basic answer and the corrected answer through each Agent in the knowledge discovery team is returned to be executed.

[0010] Optionally, determining the final answer corresponding to the natural language question based on the basic answer and the corrected answer includes: generating an argument question through the question generation agent, the argument question being generated based on the natural language question, the basic answer, the corrected answer, the information dimension ranking result, the weight of the information dimension, and the previous round of argument results; obtaining argument answer content by having each agent in the knowledge discovery team answer the argument question, the argument answer content including: the weight of the information dimension that has not been updated, or the weight of the information dimension that has been updated; and obtaining the weight of the information dimension and the answer corresponding to the natural language question in this round by having the process summary agent perform consensus calculation on the argument answer content, until the weight of the information dimension no longer updates, or the maximum round threshold is reached, to obtain the final answer corresponding to the natural language question.

[0011] Optionally, the method further includes: semantically expanding the basic answer, the corrected answer, and / or the questions generated during the process of determining the final answer using a small-parameter-scale multimodal model of the Agent to obtain extended query questions; extracting preliminary entities from the extended query questions and matching the preliminary entities with the lowest-level entities in the AD / MCI knowledge graph to obtain a preliminary matching entity set; tracing back to the lowest common ancestor entity in the AD / MCI knowledge graph based on the preliminary matching entity set; constructing a subgraph containing entity paths and relationships with the lowest common ancestor entity as the root node and the preliminary matching entities as leaf nodes; querying multiple candidate documents in the AD / MCI knowledge base using the Agent based on the basic answer, the corrected answer, and / or the questions generated during the process of determining the final answer, and selecting relevant candidate documents with relevance scores greater than a threshold from the multiple candidate documents; converting the subgraph and the relevant candidate documents into a structured graph context for the Agent, so that the Agent can make inference decisions based on the structured graph context and obtain inference results; obtaining citation annotations and adding citation annotations to the inference results.

[0012] Optionally, before semantically expanding the base answer, the corrected answer, and / or the questions generated during the determination of the final answer using the Agent's small-parameter-scale multimodal model, the method further includes: acquiring a training dataset and an initial multimodal model, the initial multimodal model comprising: a visual encoder, a multilayer perceptron (MLP) adapter, and a natural language model fine-tuned based on AD / MCI domain knowledge; training the initial multimodal model using the training dataset to update the parameters of the MLP adapter, with the parameters of the visual encoder and the natural language model both frozen; replacing the natural language model in the initial multimodal model with an AD / MCI large language model having reasoning capabilities to obtain a recombined model; acquiring an instruction set with a reasoning process, and training the recombined model using the instruction set with the reasoning process to obtain the Agent's small-parameter-scale multimodal model.

[0013] This application also provides an AD and MCI multidimensional knowledge discovery device, comprising: a question acquisition module for acquiring natural language questions in the AD / MCI domain; an information data determination module for determining, if the natural language question represents a multidimensional knowledge discovery intent, multiple information dimensions required to answer the natural language question and the data for each information dimension through multiple agents; a basic answer determination module for determining a basic answer corresponding to the natural language question based on the data of the information dimension with the highest importance among the multiple information dimensions; a corrected answer determination module for determining a corrected answer based on the data of other information dimensions besides the information dimension with the highest importance among the multiple information dimensions; and a final answer determination module for determining a final answer corresponding to the natural language question based on the basic answer and the corrected answer.

[0014] This application also provides an electronic device, including a processor and a memory, wherein the memory stores machine-readable instructions executable by the processor, and the machine-readable instructions are executed by the processor to perform the methods described above.

[0015] This application also provides a computer-readable storage medium storing a computer program, which is executed by a processor to perform the methods described above.

[0016] This application also provides a computer program product, including: a computer program or computer instructions, which are executed by a processor to perform the method described above.

[0017] The solution provided in this application achieves modular and standardized management of the problem-solving process by decomposing the problem into a combination of expert agent and data agent requirements. This not only improves the overall problem-solving accuracy by enhancing the accuracy of solving sub-problems but also reduces the complexity of unit computation, system maintenance, and updates, providing a structured foundation for subsequent functional expansion and performance optimization. Furthermore, this solution first focuses on the most critical information dimensions to quickly generate reliable basic answers, then utilizes other information dimensions for targeted correction. This preserves the interpretability of the core logic while achieving deep fusion and fine calibration of cross-modal evidence, thus ensuring the accuracy of the answer while controlling the computational complexity of the control unit. It effectively balances the high efficiency and high accuracy of AD and MCI multidimensional knowledge discovery. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 The diagram shows a flowchart of the AD and MCI multidimensional knowledge discovery method provided in the embodiments of this application; Figure 2 The diagram shown is a structural schematic of the AD and MCI multidimensional knowledge discovery device provided in the embodiments of this application; Figure 3 The diagram shows a structural schematic of an electronic device provided in an embodiment of this application. Detailed Implementation

[0020] Currently, research on Alzheimer's disease (AD) and mild cognitive impairment (MCI) faces the challenge of multi-dimensional and multi-modal knowledge fusion. Existing knowledge discovery methods, such as current methods based on large language models, struggle to balance high efficiency and high quality when dealing with such complex scenarios. Large-parameter models have high computational costs, slow response times, and suffer from uninterpretable results and hallucination problems; while lightweight models offer faster responses, their deep reasoning and multi-dimensional knowledge integration capabilities are insufficient, limiting the accuracy and comprehensiveness of the answers.

[0021] To address this, this invention proposes an AD / MCI multi-dimensional knowledge discovery scheme based on multi-agent collaboration and hierarchical answer generation. The inventive concept lies in constructing a multi-agent collaborative system to decompose the complex multi-dimensional knowledge discovery problem into a hierarchical process involving information dimension identification, importance ranking, basic answer generation, and corrected answer fusion. The system first rapidly determines the required information dimensions and their importance order through multiple agents; then, it efficiently generates reliable basic answers based on the most important dimensions; subsequently, it collaboratively integrates information from other dimensions to correct and enhance the basic answers, improving their comprehensiveness and accuracy; finally, it combines the basic and corrected answers to form a high-quality final output. Based on a small-parameter AD / MCI multimodal model, this architecture achieves a balance between high efficiency and high quality in the multi-dimensional knowledge discovery process at the system level through a collaborative mechanism of focusing first and then correcting.

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the accompanying drawings in the embodiments of this application are for illustrative and descriptive purposes only and are not intended to limit the protection scope of the embodiments of this application. Furthermore, it should be understood that the schematic drawings are not drawn to scale. The flowcharts used in the embodiments of this application illustrate operations implemented according to some embodiments of this application. It should be understood that the operations in the flowcharts may not be implemented in sequence, and steps without logical contextual relationships may be reversed or implemented simultaneously. In addition, those skilled in the art, guided by the content of the embodiments of this application, may add one or more other operations to the flowcharts, or remove one or more operations from the flowcharts.

[0023] Furthermore, the described embodiments are merely a part of the embodiments of this application, and not all of them. The components of the embodiments of this application described and illustrated herein can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed embodiments of this application, but merely to illustrate selected embodiments of this application.

[0024] It is understood that the terms "first" and "second" in the embodiments of this application are used to distinguish similar objects. Those skilled in the art will understand that the terms "first" and "second" do not limit the quantity or execution order, and that "first" and "second" do not necessarily imply that they are different. In the description of the embodiments of this application, the term "and / or" is merely a description of the association relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following related objects have an "or" relationship. The term "multiple" refers to two or more (including two), and similarly, "multiple groups" refers to two or more groups (including two groups).

[0025] It should be noted that the AD and MCI multidimensional knowledge discovery methods provided in this application can be executed by electronic devices. Here, electronic devices refer to device terminals or servers with the function of executing computer programs. Device terminals include, for example, smartphones, personal computers, tablets, personal digital assistants, or mobile internet devices. Servers refer to devices that provide computing services through a network. Servers include, for example, x86 servers and non-x86 servers. Non-x86 servers include, for example, mainframes, minicomputers, and UNIX servers.

[0026] Please see Figure 1 This illustrates the flow of the AD and MCI multidimensional knowledge discovery method provided in the embodiments of this application; the method includes the following steps: Step S110: Obtain natural language questions in the AD / MCI domain.

[0027] Alzheimer's disease (AD) is a neurodegenerative disease that primarily affects memory, thinking, and behavior. Mild cognitive impairment (MCI) of AD origin is a transitional stage between normal aging and dementia, and is a precursor to AD.

[0028] Natural language problems refer to problems expressed in everyday human language. In the AD / MCI field, natural language problems can be input by relevant personnel using text or voice.

[0029] Step S120: If the natural language question in the AD / MCI domain represents a multidimensional knowledge discovery intent, then multiple agents are used to determine the multiple information dimensions required to answer the natural language question and the data for each information dimension.

[0030] Multidimensional knowledge discovery intent refers to the complex intent behind a natural language problem that requires comprehensive analysis from multiple information dimensions (such as text, laboratory data, and MRI images) to provide an accurate answer, such as the brain age estimation problem. This application identifies whether a natural language problem involves multiple information dimensions or requires cross-modal data integration by analyzing its semantic content and structural features. For example, if a problem requires comprehensive reasoning from multiple sources of information such as text, MRI images, and laboratory data indicators, then it is determined to represent a multidimensional knowledge discovery intent.

[0031] An Agent refers to an independent intelligent agent in a knowledge discovery team. Each Agent is responsible for processing data or tasks in a specific information dimension. The Agents mentioned above can be implemented through programming (such as using Python or Java), and each Agent has specific functional modules.

[0032] The knowledge discovery team is a collaborative system built from agents selected from an agent library, including expert agents and data agents, to jointly complete multi-dimensional knowledge discovery tasks for complex problems. For example, expert agent types may include neurology expert agents, psychiatry expert agents, gerontology expert agents, nutrition expert agents, etc., while data agent types may include MRI image agents, CT image agents, laboratory data agents, bioinformatics agents, EEG agents, MRI agents, voice agents, video agents, text agents, etc.

[0033] It should be understood that, to better serve the knowledge discovery process, the agent types available in the agent library are not limited to expert agents and data agents alone. They may also include other types of agents such as user agents, planning agents, support agents (e.g., problem generation agents, process summarization agents, result evaluation agents), and AI agents (e.g., inference agents). Each type of agent can implement its functionality by calling a pre-trained AD / MCI multimodal model. Of course, different agents can also call different dedicated lightweight models to implement their functions.

[0034] The method by which multiple agents determine the multiple information dimensions required to answer a natural language question is as follows: Analyze the natural language question, determine the required expert agents and data agents. Based on their professional fields, these agents analyze the question from different dimensions and propose their required information dimensions. After being summarized, they form the multiple information dimensions required to answer the natural language question.

[0035] This application involves many dimensional concepts, some of which have not yet been introduced. A general explanation is provided here: Information dimensions refer to the different data modalities that are needed when analyzing or answering questions in the AD / MCI domain.

[0036] Data in the information dimension refers to specific content, values, or evidence belonging to a particular "information dimension".

[0037] Initial information dimensions refer to the information dimensions identified through preliminary discussions and summaries by the knowledge discovery team at the early stages of the process. These dimensions represent the dimensions that the team deems necessary to focus on at this stage. During this process, it's also necessary to determine the data for these initial information dimensions; it should be understood that some initial information dimensions may have empty data sets.

[0038] The final information dimension refers to the information dimension that is finally determined in the later stages of the process, after interaction and supplementation. This process also involves determining the data for the final information dimension, which may be obtained by adding data from other information dimensions to the initial information dimension data.

[0039] Information dimension ranking results refer to the structured data generated after evaluating the importance of each dimension in the final information dimension. Information dimension ranking results include: information dimension ranking data and information dimension ranking answers; information dimension ranking answers include: information dimension ranking and the average importance score for each information dimension.

[0040] The first-order information dimension refers to the specific information dimension that ranks first among the "multiple information dimensions after sorting," that is, the information dimension that ranks first in importance.

[0041] The second information dimension refers to the information dimensions other than the "first-order information dimension" in the information dimension ranking.

[0042] The weight of an information dimension refers to the importance or influence coefficient assigned to each second information dimension during the argumentation process. The application scenarios for the aforementioned information dimension weights include at least the following: First, in determining the corrected answer, the weight of each information dimension can be calculated based on the average importance score of each second information dimension, and then used to weight and fuse the sub-corrected answers generated from the data of each second information dimension; Second, in determining the final answer, the weight of the information dimension can be adjusted based on the results of this round of argumentation; Third, the information dimension weights can be used by a question-generating agent to generate argumentation questions.

[0043] Step S130: Determine the basic answer to the natural language question based on the data of the information dimension that ranks first in importance among multiple information dimensions.

[0044] Step S140: Determine the corrected answer based on the data of the information dimensions other than the information dimension with the highest importance among multiple information dimensions. The corrected answer is a supplement or correction to the basic answer determined based on the data of the other information dimensions.

[0045] Step S150: Determine the final answer to the natural language question based on the basic answer and the corrected answer.

[0046] Determining the final answer by combining the basic and corrected answers can be implemented in several ways, including but not limited to: First, a weighted fusion of the basic and corrected answers; second, starting with the basic and corrected answers and iteratively optimizing through multiple rounds of argumentation; third, introducing external knowledge to verify, supplement, and merge the basic and corrected answers; and fourth, multiple agents negotiating the basic and corrected answers to reach a consensus answer. It is understood that one, two, or more of these implementation methods can be combined to construct a specific implementation method.

[0047] In implementing the above scheme, a knowledge discovery team composed of multiple agents is employed to determine the multiple information dimensions required to answer natural language questions. Since the collaborative work of multiple agents can simultaneously cover the information needs of different dimensions, multi-dimensional analysis of AD / MCI domain problems can be performed concurrently. This avoids the huge computational overhead and bias of a single model performing end-to-end inference for all modalities, improving processing efficiency and quality. Furthermore, by determining the basic answer based on the data of the most important information dimension among multiple information dimensions, and generating corrective answers based on data from other information dimensions, the basic answer can be corrected and optimized. By first focusing on the most critical information dimension to quickly generate a reliable basic answer, and then using other information dimensions for targeted correction, the interpretability of the core logic is preserved, and deep fusion and fine calibration of cross-modal evidence are achieved. This ensures the accuracy and professionalism of the answer while controlling the computational complexity of the control unit, effectively balancing the high efficiency and high accuracy of multi-dimensional knowledge discovery in AD and MCI.

[0048] In practical applications, natural language questions may not fall under the AD / MCI domain, or they may not represent a multidimensional knowledge discovery intent. For natural language questions that do not represent a multidimensional knowledge discovery intent, a pre-judgment model (such as BERT) can be built to determine the way to answer the question. Based on the answer, AD / MCI knowledge base retrieval and / or interaction with a large language model can be performed to obtain the final answer to the natural language question. For natural language questions that do not fall under the AD / MCI domain, an answer that the user cannot answer can be returned.

[0049] The following is a method for determining whether a natural language problem belongs to the AD / MCI domain and whether it represents a multidimensional knowledge discovery intent: The following formula is used to determine the domain of a natural language problem and the type of knowledge discovery intent: Field yorno senior yorno = LLM admci (query, prompt) yorno ) Among them, Field yorno Representing the domain of natural language problems (including AD / MCI domain and non-AD / MCI domain), senior yorno Indicates the type of knowledge discovery intent (including multidimensional knowledge discovery intent and non-multidimensional knowledge discovery intent), LLM admci This represents a pre-trained AD / MCI multimodal model (the construction process is the same as the Agent's AD / MCI multimodal model), query represents a natural language question, and prompt represents a prompt. yornoThe prompt content can include: identity definition (e.g., you are an expert in domain and multidimensional knowledge discovery intent judgment), work goal (e.g., output judgment results and relevant evidence), problem description (clearly list the user's problem or relevant explanation here), and relevant samples (e.g., give similar problems and judgment results).

[0050] Before step S120, a knowledge discovery team needs to be built. The following describes one way to build a knowledge discovery team, including steps S101-S102: Step S101: Analyze natural language problems by planning agents to determine the required expert agent type and data agent type.

[0051] The core capability of the planning agent is to understand the semantic intent of a natural language question and to perform problem decomposition and resource planning based on a predefined library of expert and data agent capabilities. For example, it can analyze keywords, context, and implicit requirements in the question to determine the relevant professional field (e.g., neurology, nutrition) and necessary data modalities (e.g., MRI images, text, laboratory data), and then map and output a list of the required expert agent and data agent types. This process can be represented as: experts, datas=Agent plan (query, prompt) expert ) Here, "experts" represents one or more expert agent types, and "datas" represents one or more data agent types. plan Indicates planning agent, prompt expertThe prompt content can include identity definition (e.g., you are an AD / MCI multidimensional knowledge discovery planning expert), role description (e.g., analyzing user input questions; determining the type of expert agent or data agent needed to answer user questions), key responsibilities (e.g., in-depth analysis of user input information; providing highly relevant judgment criteria; combining your own extensive experience and abilities when making judgments), work objectives (e.g., outputting the results and relevant evidence determined by the expert agent and data agent), problem description (clearly listing the user's question or related explanation), relevant rules (e.g., determining results must be based on a careful analysis process; the analysis process must be combined with highly relevant evidence), and relevant materials (e.g., relevant reference materials for answering questions). It is understood that the structure of the prompt content involved in this application can include parts such as identity definition, role description, key responsibilities, work objectives, problem description, relevant rules, and relevant materials. Of course, the structure of the prompt content can be selected or adjusted according to specific scenarios. For example, in some scenarios, relevant rules are not needed, so the relevant rules part can be omitted when constructing the prompt content.

[0052] Step S102: Build a knowledge discovery team based on the expert agent type and the data agent type.

[0053] The knowledge discovery team can be represented as an Agent. team (Number of members is N=N) experts +N datas In the implementation of the above solution, by decomposing the problem into a combination of expert agent type and data agent type requirements, modular and standardized management of the problem-solving process is achieved. This not only improves the accuracy of overall problem-solving by improving the accuracy of solving sub-problems, but also reduces the complexity of unit computation, system maintenance and updates, and provides a structured foundation for subsequent functional expansion and performance optimization.

[0054] The following describes one implementation of step S120, including steps S121-S125: Step S121: Generate Agent through Questions. Determine the initial information dimension question based on the natural language question.

[0055] The core capability of the problem-generating agent is to generate targeted questions based on available information such as the current stage of the dialogue, the user's original question, and summarized data. This allows it to guide information gathering, trigger multi-dimensional analysis, or initiate team discussions. The question that determines the initial information dimensions can be represented as: question infoget =Agent ques (query, prompt) ques-infoget ) Among them, question infoget This indicates that the Agent is generated by addressing a question, which determines the initial information dimensions based on the natural language question. ques This indicates that the problem generates an Agent, prompt. ques-infoget The prompt text indicates the question of determining the initial information dimensions.

[0056] Step S122: Each Agent in the knowledge discovery team answers the initial information dimension question based on the natural language question, obtaining multiple information dimension answers.

[0057] This process can be represented as: Answer agent-i-infoget =Agent i (query, question) infoget prompt agent-i-infoget ) Among them, Answer agent-i-infoget Agent represents the i-th information dimension answer among multiple information dimension answers. i Let i represent the i-th agent in a knowledge discovery team composed of multiple agents, and let question be the i-th agent. infoget This indicates that the agent generates an initial information dimension based on a natural language question, prompt. agent-i-infoget The prompts indicate that you can obtain answers from multiple information dimensions.

[0058] Step S123: Through process summary, the Agent summarizes the answers of multiple information dimensions to obtain the initial information dimension and the data of the initial information dimension.

[0059] The core capability of the process summarization agent is to aggregate, deduplicate, resolve contradictions, and structure the scattered answers, ranking opinions, or arguments generated by various agents, outputting high-quality information dimensions, consensus rankings, or final conclusions. The process by which the process summarization agent determines the initial information dimensions and the data for those dimensions can be represented as follows: Data sum-infoget Answer sum-infoget = Agent sum (query, Answer) agent-i-infoget prompt sum -infoget ) Among them, Data sum-infogetThe data representing the initial information dimension (which may include one or more of the preset data types such as MRI images, CT images, laboratory data, bioinformatics, EEG, MRI, speech, video, and text) Answer sum-infoget Indicates the initial information dimension, Agent sum Summary of the process: Agent, prompt sum-infoget The prompt text indicates the information obtained in the initial information dimension and the data in the initial information dimension.

[0060] Step S124: The Agent generates a new information dimension data question based on the initial information dimension, the data of the initial information dimension, and the natural language question. This new information dimension data question can be used for further interaction with the user.

[0061] This process can be represented as: Question new-infoget = Agent ques (query, Data) sum-infoget Answer sum-infoget prompt ques -new-infoget ) Among them, Question new-infoget This indicates a new dimension of information data issue, prompt ques-new-infoget The prompt text indicates the content of the question regarding identifying new information dimensions and data.

[0062] Step S125: Obtain the answer to the new information dimension data question, and summarize the process to obtain the final information dimension and the final information dimension data based on the initial information dimension, the data of the initial information dimension, and the answer to the new information dimension data.

[0063] Specifically, after obtaining the new information dimension data question, it can be sent to the user terminal through the user agent to interact with the real user, thereby receiving the new information dimension data answer corresponding to the user's new information dimension data question. user-new-infoget The process involves the Agent generating the final information dimension and the final information dimension data based on the initial information dimension, the data from the initial information dimension, and the data from the new information dimension. This process can be represented as: Data sum-infoget-final Answer sum-infoget-final = Agent sum (query, Data) sum-infoget Answer sum-infoget Answer user-new-infoget prompt sum-infoget-final) Among them, Data sum-infoget-final Data representing the final information dimension, Answer sum-infoget-final The answer represents the final information dimension. user-new-infoget This refers to the answer to the new information dimension data question sent by the user, prompt. sum-infoget-final The prompt text indicates that the final information dimension and the data of the final information dimension have been obtained.

[0064] In the implementation of the above scheme, the process summary agent summarizes the answers of multiple information dimensions, realizing the effective integration of information. Through the division of labor and cooperation of multiple agents, multi-stage processing of information dimensions is realized. From question generation to answer acquisition and then to answer summary, this process-oriented approach improves the scalability and adaptability of the system and can cope with complex and ever-changing natural language problem scenarios.

[0065] The following describes an optional implementation of step S130, which may include: Step S131: The question-generating agent determines the information dimension ranking question based on the final information dimension and the data of the final information dimension. This process can be represented as: question infoget-rank =Agent ques (query, Data) sum-infoget-final Answer sum-infoget-final prompt ques-infoget-rank ) Among them, question infoget-rank This refers to the information dimension sorting problem, prompt ques-infoget-rank The prompt text indicates the content of the question regarding the sorting of information dimensions.

[0066] Step S132: Each Agent in the knowledge discovery team generates individual ranking results for multiple information dimensions based on the final information dimension and the data and information dimension ranking problem of the final information dimension. The individual ranking results include importance scores. This process can be represented as: Answer agent-i-infoget-rank = Agent i (query, question) infoget-rank Data sum-infoget-final Answer sum-infoget-final prompt agent-i-infoget-rank ) Among them, Answer agent-i-infoget-rank This represents the individual ranking results for information dimensions, including the information dimension, the data for the corresponding information dimension, and the importance score for the corresponding information dimension.agent-i-infoget-rank The prompt text indicates the generation of individual ranking results with multiple information dimensions.

[0067] Step S133: Through process summary, the Agent sorts the importance of multiple information dimensions according to the average importance score of each information dimension, and obtains the information dimension sorting results of multiple information dimensions.

[0068] The average importance score can be considered as the arithmetic mean of the importance scores given by multiple experts / data agents within the knowledge discovery team for the same information dimension. This ranking process can be expressed as: Data sum-infoget-rank Answer sum-infoget-rank = Agent sum (query, Answer) agent-i-infoget-rank prompt sum-infoget-rank ) Among them, Data sum-infoget-rank This represents the information dimension sorting data in the information dimension sorting results. sum -infoget-rank This refers to the information dimension sorting answer in the information dimension sorting results, prompt. sum-infoget-rank The prompt text indicates the information dimension sorting results.

[0069] Step S134: The Agent determines the basic question based on the information dimension sorting results. This process can be represented as: Question ques-basic-answer = Agent ques (query, Data) sum-infoget-rank Answer sum-infoget-rank prompt ques-basic-answer ) Among them, Question ques-basic-answer This indicates a basic question, prompt. ques-basic-answer The prompts indicate the content of the basic questions.

[0070] Step S135: Using the data Agent corresponding to the first ordered information dimension in the information dimension ranking, answer the basic question based on the data of that information dimension to obtain the basic answer. This process can be represented as: Answer basic-answer = Agent first (query, Question) ques-basic-answer Data sum-infoget-rank Answer sum-infoget-rank prompt basic-answer ) Among them, Answer basic-answer This represents the basic answer, Agent. first This represents the data Agent, prompt, corresponding to the first-order information dimension. basic-answer The prompts indicate the basic answer, and the response must be tailored to the question type (e.g., for numeric questions like brain age estimation, the basic answer must be in numerical format; for non-numeric questions, the response must be in a specific format). It's important to note that before submitting the question to the data agent, neuroimaging data must first undergo the standardization process described below; similar steps apply to other data modalities.

[0071] In the implementation of the above scheme, the data agent corresponding to the first sequential information dimension answers the question, ensuring that the generation of the basic answer strictly follows the importance order of the information dimension. This sequentially driven answer generation mechanism can effectively improve the logic and consistency of the answer and avoid the problems of information redundancy or conflict.

[0072] The following describes an optional implementation of step S140, which may include: Step S141: The Agent generates a correction question based on the natural language question, the basic answer, and the information dimension ranking results. This process can be represented as: Question ques-jth-answer =Agent ques (query, Answer) basic-answer Data sum-infoget-rank Answer sum-infoget-rank prompt ques-jth-answer ) Among them, Question ques-jth-answer Let j be the correction problem, and prompt be the prompt. ques-jth-answer This indicates the prompt word content for determining the j-th correction question.

[0073] Step S142: Using the data agents corresponding to other information dimensions besides the first-order information dimension in the knowledge discovery team, determine the correction rules based on the natural language question, basic answer, correction question, and information dimension ranking results. This process can be represented as: Answer jth-answer-rule =Agent jth (query, Answer) basic-answer Question ques-jth-answer Data sum-infoget-rank Answer sum-infoget-rank promptjth-answer-rule ) Among them, Answer jth-answer-rule The Agent represents the j-th correction rule (correction rules can be expressed by regular expressions or natural language). jth This represents the data agent corresponding to the j-th information dimension in the knowledge discovery team, excluding the data agent corresponding to the first-order information dimension. (prompt) jth-answer-rule This indicates the prompt word content for determining the j-th correction rule.

[0074] Step S143: Select multiple second information dimensions from the information dimension ranking, excluding the information dimension ranked first in importance.

[0075] Step S144: Calculate the weights of multiple second information dimensions based on the average importance score of each second information dimension to obtain the weights of the information dimensions.

[0076] For example, through the formula weight info-j = score info-j / ∑(score) info-j ) performs weight calculation, where weight info-j This represents the weight corresponding to the j-th second information dimension (it should be understood that the weights corresponding to multiple second information dimensions together constitute the weight of the information dimension). info ), score info-j This represents the average importance score of the j-th second information dimension.

[0077] Step S145: By using the correction rules and the weights of the information dimensions, the sub-corrected answers generated based on the data of each second information dimension are weighted and fused to obtain the corrected answer.

[0078] The calculation process can be represented as follows: answer added-answer = ∑weight info-j ×answer j-th-answer ; answer j-th-answer = rule j-th (Data) sum-infoget-rank ); Among them, answer j-th-answer The rule represents the sub-corrected answer corresponding to the j-th information dimension. j-th This indicates the above answer jth-answer-rule Determined correction rules, answer added-answer This indicates a corrected answer.

[0079] This solution achieves differentiated processing of information dimensions by selecting the most important information dimension and calculating the weights of other dimensions. This weighting method based on importance scores effectively highlights the impact of key information dimensions while reasonably incorporating the contributions of secondary dimensions. This significantly improves the comprehensiveness, completeness, and interpretability of the answer while ensuring its core accuracy. Furthermore, through the collaborative work of the question-generating agent and the knowledge discovery team, correction questions and rules can be dynamically generated, enabling multi-dimensional intelligent correction of the basic answer. This agent-based distributed processing mechanism not only improves the system's flexibility but also enhances the targeting and accuracy of correction by having each agent focus on processing specific dimensions of information, effectively avoiding information overload and bias in a single processing unit.

[0080] Steps S141 to S145 above mainly address the process of determining the corrected answer for numerical computation problems. For non-numerical computation problems (such as mechanism descriptions), the corrected answer can be determined using a weight-based semantic fusion mechanism, as detailed below: By utilizing data agents for information dimensions other than the first-order information dimension within the knowledge discovery team, semantic reasoning or text generation is performed on the information dimension ranking results according to their corresponding correction rules, yielding text-based sub-corrected answers for each dimension. Then, the reasoning agent merges these text-based sub-corrected answers based on the weights of the information dimensions. Fusion methods include, but are not limited to, weighted text selection, weight-guided text generation, or structured merging. Finally, the fused output is the corrected answer for non-numerical computational questions.

[0081] The above step S150, determining the final answer to the natural language question based on the basic answer and the corrected answer, can be achieved in several ways. Two of them are described below.

[0082] Method 1: Argumentation based on consistency verification The core of this approach lies in the team directly verifying the consistency between the basic and corrected answers through multiple rounds of discussion, thereby driving the generation or iteration of the final answer. This approach includes the following steps: Step S151: Each Agent in the knowledge discovery team conducts multiple rounds of validation of the base answer and the corrected answer.

[0083] First, before the argument is presented, the system calculates an initial final answer based on the base answer and the corrected answer (for example, for numerical calculation problems, this can be done using the formula Answer). all-answer =answer basic-answer +answer added-answerAchieve weighted fusion; for non-numerical computation problems, refer to the weighted semantic fusion mechanism introduced earlier. As the starting point for argumentation, to ensure the reasonableness of the results, the knowledge discovery team needs to demonstrate and evaluate the consistency between this initial final answer and the basic answer used to calculate it, as well as the corrected answer.

[0084] The question-generating agent can generate a question for the k-th round of argumentation based on a natural language question, the basic and corrected answers, the final answer, and the results of previous rounds of argumentation. The value of k can range from 1 to n, where n is the maximum number of argumentations. Then, based on the question generated by the question-generating agent, the knowledge discovery team agent... team Member (Agent) i Each of the following is analyzed independently based on the question posed by the agent, the natural language question, the basic answer and the corrected answer, the final answer, and the results of the previous round of argumentation, and argumentation opinions are provided for each. The argumentation opinions include a judgment on whether the basic answer and the corrected answer are consistent (e.g., "consistent" or "inconsistent," and a consistency score represented by a value between 0 and 1). If the judgment is "inconsistent," a detailed description of the conflict points and preliminary suggestions for weight adjustment are provided.

[0085] Step S152: In each round of argumentation in the multi-round argumentation, each Agent performs consistency verification on the basic answer and the corrected answer according to the preset argumentation rules.

[0086] Each agent reviews the basic and corrected answers according to pre-defined argumentation rules. These pre-defined argumentation rules may include: core fact consistency rules and logical reasoning consistency rules.

[0087] Step S153: When the basic answer and the corrected answer pass the consistency verification, the weighted fusion result of the basic answer and the corrected answer is determined as the final answer.

[0088] After each Agent completes its argumentation, it obtains its own argumentation result answer. This result answer includes a judgment on whether the base answer and the corrected answer are consistent (e.g., "consistent" or "inconsistent," with a consistency score represented by a value between 0 and 1). Then, the process summary Agent performs consensus calculations on all argumentation result answers (e.g., if the percentage of "consistent" among the multiple argumentation result answers from all team members exceeds 2 / 3, then the consistency verification is passed). It can be understood that if the base answer and the corrected answer are determined to have passed the consistency verification, the system determines the final answer to the natural language question as the weighted fusion result of the current base answer and the corrected answer (calculated using a method similar to that described above). all -answer-final ). Step S154: When the basic answer and the corrected answer fail the consistency verification, adjust the weight of the information dimension based on the results of this round of argumentation, and return to execute steps S151 to S153 above.

[0089] If the basic answer and the corrected answer fail the consistency verification, it indicates an inherent conflict in the current answer, requiring correction of the corrected answer. In this case, the system will adjust the relevant factors causing the conflict based on the conflict information and adjustment suggestions provided by each agent during this round of argumentation. This adjustment may involve the weighting of information dimensions. info The redistribution of ).

[0090] After the adjustment is completed, the system will recalculate a final answer to be verified using the adjusted parameters (such as the new weights) and return to step S151 to start a new round of multi-round argumentation. This iterative process continues until the basic answer and the corrected answer pass the consistency verification, or the preset maximum argumentation round threshold is reached (at this point, the weight corresponding to the argumentation round with the highest average consistency score is selected as the weight of the information dimension when the maximum argumentation round threshold is reached; the final answer to be verified corresponding to the corresponding argumentation round is selected as the final answer to the natural language question when the maximum argumentation round threshold is reached). all-answer-final ).

[0091] It should be understood that in the above consistency verification judgment, a certain percentage of members in the knowledge discovery team must agree for the consistency verification to be considered successful. For example, this percentage can be set to 2 / 3.

[0092] Method 2, Argument based on weight iterative update The core of this approach lies in the team's dynamic evaluation and iterative updates of the information dimensions' weights through multiple rounds of argumentation, and the regeneration of answers based on the updated weights, until termination conditions are met. Termination conditions include: reaching the maximum number of rounds, or the weights ceasing to be updated. This approach includes the following steps: Step S155: Generate an argumentation question through the question generation agent. The argumentation question is generated based on the natural language question, the basic answer, the corrected answer, the information dimension ranking result, the information dimension weight, and the previous round of argumentation results.

[0093] First, before proceeding with the argumentation, an initial final answer is calculated based on the basic answer and the corrected answer. all-answer For example, for numerical calculation problems, the formula Answer can be used. all-answer =answer basic -answer +answer added-answerTo achieve weighted fusion; for non-numerical computation problems, the weight-based semantic fusion mechanism introduced earlier can be used as a reference. To ensure the reasonableness of the results, the knowledge discovery team needs to demonstrate and optimize the weight contribution of each dimension.

[0094] Specifically, the problem-generating agent generates a problem for the result argumentation in the k-th round, and this process is represented as follows: Question result-argu-k =Agent ques (query, answer) all-answer , answer basic-answer , answer added-answer Data sum-infoget-rank Answer sum-infoget-rank Weight info , answer j-th-answer , answer sum-result-argu-k-1 prompt ques-result-argu-k ); Among them, Question result-argu-k This represents the problem in the k-th round of argumentation. all-answer This represents the final answer before the start of the current round, weight. info The answer represents the weight of the current information dimension. sum-result-argu-k-1 The prompt represents the answer to the argument in the (k-1)th round. ques-result-argu-k This indicates the content of the argument prompt in round k.

[0095] Step S156: Obtain the argument response content by having each Agent in the knowledge discovery team answer the argument question.

[0096] The agent generates a question related to the problem. result-argu-k Knowledge discovery team agent team Member (Agent) i Answer in order, and the answer process can be represented as follows: Answer agent-i-result-argu-k =Agent i (query, question) result-argu-k , answer all-answer , answer basic-answer , answer added-answer Data sum-infoget-rank Answer sum-infoget-rank Weight info , answer j-th-answer , answer sum-result-argu-k-1 promptagent-i-result-argu-k ) Among them, Answer agent-i-result-argu-k This represents the argument response of the i-th agent in the k-th round.

[0097] The argument response includes: the weight of the information dimension that has not been updated, or the weight of the information dimension that has been updated. The former indicates agreement with the current weight, while the latter proposes a new weight.

[0098] Step S157: Through process summarization, the Agent performs consensus calculation on the argument response content to obtain the information dimension weight and the answer corresponding to the natural language question in this round, until the information dimension weight no longer updates, or the maximum round threshold is reached, to obtain the final answer to the natural language question. This process can be represented as: Answer sum-result-argu-k =Agent sum (query, question) result-argu-k , answer all-answer , answer basic-answer , answer added-answer Data sum-infoget-rank Answer sum-infoget-rank Weight info , answer j-th-answer , answer sum-result-argu-k-1 Answer agent-i-result-argu-k prompt sum-result-argu-k ) Among them, Answer sum-result-argu-k The Agent represents the answer to the argument in the k-th round. sum The process summary is represented by the Agent.

[0099] Process Summary: Agent's Responses to All Team Members agent-i-result-argu-k To summarize, the final information dimension weights for this round (round k) are generated. info-k Correcting the answer added-answer-k and the overall answer all-answer-k The weight adjustment requires the approval of more than a preset percentage (e.g., 2 / 3) of team members. The adjustment value is the average of the new weights provided by all members who agree to the adjustment. Steps S155 to S157 are repeated sequentially until the information dimension weights are no longer updated or the maximum number of rounds is reached, thus obtaining the final Weight. info-final Answer added-answer-final And the Answer as the final answer to the natural language problem (calculated in a similar way to the above). all-answer-final .

[0100] Methods 1 and 2 not only preserve the stability of the basic answer but also incorporate the corrective effect of the corrected answer. This answer acquisition scheme based on a professional-consensus collaboration mechanism improves the overall quality of the answer.

[0101] Therefore, the results (including key processes for obtaining the results, such as the final answer) can be transmitted via the user agent. all-answer-final The process involves providing feedback to the user (including basic answers, weighted final information dimensions, and final corrected answers to ensure interpretability of the results). This process is specifically implemented through a prompt-like mechanism. expert The method for setting the prompt content to generate feedback user (Based on the final answer) all-answer-final The process involves (basic answer, weight of final information dimension, final corrected answer, etc.) followed by calling the corresponding agent to complete the task, which will not be detailed further.

[0102] In one optional embodiment provided in this application, knowledge enhancement and credibility processing of reasoning results can also be implemented, specifically including the following steps: Step S171: Use the Agent's small-parameter-scale multimodal model to semantically expand the basic answer, the corrected answer, and / or the questions generated during the process of determining the final answer to obtain extended query questions. Extract preliminary entities from the extended query questions and match the preliminary entities with the lowest-level entities in the AD / MCI knowledge graph to obtain a preliminary matching entity set.

[0103] A multimodal model is an artificial intelligence model that can simultaneously understand and process multimodal information (such as text and visual information). Based on the original question and its intermediate or final answers generated during the solution process, a multimodal model can analyze its deep semantics to generate a series of semantically similar or related extended query questions, thus covering a more comprehensive range of query intents.

[0104] Small parameter-scale multimodal models refer to neural network models with a pre-built parameter size smaller than a preset threshold and possessing multimodal capabilities. This preset threshold can be on the order of tens of bits, such as 10 bits. For example, a small parameter-scale multimodal model can be built based on the deepseek-r1-distill-qwen2.5-7b model.

[0105] For example, a pre-set prompt template guides a large language model to generate expanded query questions covering synonyms, related terms, and potential implicit intentions, thereby enhancing the semantic coverage of the questions. Subsequently, a Named Entity Recognition (NER) model or a combination of rule-based and dictionary-based methods is used to extract preliminary entities related to the AD / MCI domain from the expanded query questions, such as "β-amyloid protein," "MMSE score," and "mild cognitive impairment." Next, these preliminary entities are compared with the lowest-level entities (i.e., the finest-grained leaf node entities) in the constructed AD / MCI knowledge graph using string matching and semantic similarity calculation (e.g., cosine similarity based on embedding vectors). Entities with a matching degree higher than a set threshold are selected to form a preliminary set of matched entities.

[0106] The aforementioned AD / MCI knowledge graph construction process includes: First, for AD / MCI-related literature, the most basic entity knowledge is constructed using entity extraction methods, and relationships between entities (pre-defined or open) are constructed using relation extraction methods; second, semantic clustering of entities is performed; for each cluster, more abstract entity concepts are generated (specifically, the LLM is prompted to generate the entity concept that best represents the cluster through a hinting method, and the predetermined number of L1 basic entity names most relevant to the entity concept are obtained and recorded); further, for the abstract-level entities, the relation extraction method is used, combined with entity clustering, to construct the most probable relationships (specifically, the LLM is prompted to generate the most probable relationships through a hinting method). (1) The relationship between the clustered entity concepts is best represented, and the names of the L2 basic entities most relevant to the relationship and its associated entity are obtained and recorded. (2) The above two steps are repeated to obtain the hierarchical entity and relationship graph. (3) For the lowest level entities and relationships in the graph, their positional attributes (including document, chapter, section, subsection, paragraph, and sentence) are determined based on the location of the entity and relationship (i.e., the location of the corresponding entity). (4) For abstract level entities and relationships, their positional attributes (including document, chapter, section, subsection, paragraph, and sentence) are determined based on the location of the entity and relationship at the corresponding basic level as determined in the previous steps, and then saved to the AD / MCI knowledge graph.

[0107] Step S172: Based on the initial matching entity set, trace back to the lowest common ancestor entity in the AD / MCI knowledge graph.

[0108] The lowest common ancestor entity refers to the lowest-level common ancestor node on the hierarchical path of a knowledge graph that can cover all entities in the initial matching entity set.

[0109] The process of tracing the lowest common ancestor entity can be as follows: Utilizing the path lookup function of a graph database (such as Neo4j), starting from each initially matched entity, backtrack upwards along hierarchical relationships such as "is-a," "part-of," or "subclass-of," until reaching the root node or a preset maximum depth. Then, calculate the intersection of all backtracking paths and identify the common ancestor node where all paths first intersect; this is the lowest common ancestor entity. This process can be efficiently implemented using graph algorithms (such as the Least Common Ancestor (LCA) algorithm), ensuring that the selected ancestor entity covers the semantic categories of all initially matched entities while maintaining the lowest possible level of abstraction, thus preserving sufficient information granularity.

[0110] Step S173: Construct a subgraph containing entity paths and relationships, with the lowest common ancestor entity as the root node and the initially matched entities as leaf nodes.

[0111] After identifying the lowest common ancestor entity, the system uses this entity as the root node of the subgraph and all initially matched entities as leaf nodes, extracting the complete paths connecting the root node and each leaf node in the knowledge graph. Each path includes intermediate nodes and their semantic relationships (such as "cause," "associate," etc.). Using the subgraph extraction function of the graph database, all paths are merged and deduplicated to construct a connected, directed, hierarchical subgraph. This subgraph not only contains entity nodes and relation edges but also synchronously records the positional attributes of each entity and relation in the original document (such as the document, chapter, paragraph, and character offset) to ensure subsequent traceability. Finally, this subgraph is output in a structured format (such as JSON-LD or RDF triples) as the knowledge context on the graph side.

[0112] Step S174: The Agent queries multiple candidate documents in the AD / MCI knowledge base based on the basic answer, the corrected answer, and / or the questions generated during the process of determining the final answer, and selects relevant candidate documents with a relevance score greater than the threshold from the multiple candidate documents.

[0113] In each user query-related problem-solving process (such as determining information dimensions, result verification, etc.), each agent can retrieve relevant knowledge content from the basic knowledge base (for the retrieved documents, after relevance scoring and sorting, select a predetermined number of relevant document abstracts as input) and incorporate them into the agent context to make more accurate reasoning decisions.

[0114] The AD / MCI knowledge base can be constructed as follows: First, an AD / MCI knowledge base framework needs to be built. The AD / MCI knowledge base framework is designed using a multi-level, multi-link approach. Overall, the first-level categories of the knowledge base framework include subject knowledge, applied knowledge, and cutting-edge knowledge. Further, the first-level categories can be subdivided into second-level categories. For example, subject knowledge can be divided into textbooks, monographs, and papers; applied knowledge can be divided into guidelines, consensus statements, clinical pathways, and drug knowledge; and cutting-edge knowledge includes RCT studies and meta-analyses / systematic reviews. Even further, the second-level categories can be subdivided into tertiary categories, and so on until they are indivisible. To ensure the effectiveness of knowledge application, the knowledge base framework includes a tagging system, such as topic (e.g., AD, MCI), type (e.g., clinical trials, randomized controlled trials), level (e.g., high recommendation, low recommendation), modality (e.g., text txt, pdf), granularity (e.g., document, entity), publication time, file name, and version number. Similarly, to ensure the effectiveness of knowledge application, structural tags, including titles, chapters, sections, subsections, paragraphs, and sentences, are set for the content in the knowledge base. For the obtained AD / MCI textbooks, monographs, and papers, the limitations of retrieval methods can lead to discrepancies with actual needs; that is, unnecessary non-AD / MCI literature may be present in the obtained textbooks, monographs, and papers. Therefore, it is necessary to filter the obtained textbooks, monographs, and papers, specifically targeting the titles, abstracts, keywords, and chapter abstracts of the literature. Machine learning methods, such as few-shot learning, are used to distinguish whether the literature belongs to the AD / MCI field, thereby retaining AD / MCI literature to obtain accurate AD / MCI literature. Subsequently, following the above system, the textbooks, monographs, and papers in the database are processed using methods such as metadata extraction, rule-based (pre-set) processing, few-shot learning model processing, and information extraction model construction. This yields the corresponding tag content for each component of the system, which is then stored in the corresponding fields in the knowledge base (including the relevant literature, chapter, section, subsection, paragraph, sentence, and specific character position, if applicable). Build indexes for the knowledge base, including inverted indexes (including steps such as word segmentation, and index fields include the page number and paragraph number of the text) and vector indexes.

[0115] The Agent's retrieval process in the AD / MCI basic knowledge base is as follows: First, the question is segmented using a word segmenter, and query keywords are generated by combining synonym expansion and stop word filtering. Then, candidate documents are quickly retrieved using an inverted index, and semantic similarity scores between the question and the document abstract or full text are calculated using vector indexes (such as dense retrieval based on Sentence-BERT). All candidate documents are sorted by their comprehensive relevance score (integrating keyword matching score and semantic similarity), and the system selects the top N documents with scores higher than a preset threshold (e.g., 0.75) as relevant candidate documents. Metadata for each document (title, abstract, source, publication date, etc.) and highlighted content are retained for subsequent context building.

[0116] Step S175: Transform the subgraph and related candidate documents into a structured graph context for the Agent, so that the Agent can make inference decisions based on the structured graph context and obtain inference results.

[0117] The structured subgraph constructed in step S173 is integrated with the relevant candidate documents selected in step S174 to construct a unified structured graph context. Specifically, entity-relation triples in the subgraph are aligned with key sentences in the documents according to semantic themes, and positional attribute labels are attached. Subsequently, the above content is organized into a context format that mixes natural language and structured data according to a preset template. After receiving a hint containing this structured graph context (such as relevant information within it), the Agent (or its underlying large language model) can combine it with the original question for reasoning. The model can refer to the logical connections in the subgraph and the evidence in the documents to perform more accurate analysis, judgment, or generation, and finally output a reasoning result with more sufficient knowledge basis and more rigorous logic.

[0118] Step S176: Obtain the reference annotations and add them to the inference results.

[0119] Citation annotation is a type of metadata used to identify the source of information. In this scheme, it refers to indicating which specific relevant candidate literature a key statement, data, or conclusion in the reasoning results originates from.

[0120] The process of adding citation annotations to the reasoning results is as follows: First, in the knowledge base application process for specific stages described above, when the retrieved knowledge content is incorporated into the context of a specific agent to obtain the response results, the agent is prompted to provide a list1 of relevant references corresponding to the question and response results (the agent needs to re-evaluate the correctness of the generated preliminary results, and the content that is correctly evaluated is used as the final list1). Second, after obtaining the response results, the agent is prompted to generate search terms based on the question input and the agent's response results, and then search the basic knowledge base to obtain a list2 of relevant documents. Third, based on the entities and relationships in the subgraphs obtained in the above graph retrieval process, as well as the corresponding positional attributes, a list3 of documents is obtained. Fourth, the above three lists are integrated to obtain list = list1 + list2 + list3, and for each document, the agent is prompted to score the relevance of each sentence in the response results. Documents with scores exceeding a pre-set threshold are used as references for the corresponding sentences, thus obtaining high-quality citation annotations. Finally, these high-quality citation annotations are added to the above reasoning results, thereby improving the reference basis and interpretability of the relevant response content.

[0121] The aforementioned agent's small-parameter-scale multimodal model can be formed by continuously training and fine-tuning a basic model (a small-parameter-scale language model) in general medicine, fine-tuning in the AD / MCI domain, and enhancing multimodal capabilities. Considering that building an AD / MCI multimodal model with reasoning capabilities requires constructing a large cross-modal training dataset, and also to ensure the inference efficiency of the inference model as much as possible, a highly efficient multimodal large language model construction process is designed. The following section uses an MRI neuroimaging multimodal model as an example to introduce one training method for the agent's small-parameter-scale multimodal model, including the following steps: Step S181: Obtain the training dataset and the initial multimodal model, which includes: a visual encoder, a multilayer perceptron (MLP) adapter, and a natural language model fine-tuned based on AD / MCI domain knowledge.

[0122] Building a training dataset requires standardizing the neuroimaging data. Taking a structural image as an example, the necessary processing includes format conversion (e.g., converting DICOM to NIFTI), skull stripping (removing non-brain structures like the skull from the image), image registration (using affine transformations and other methods to register the image using standard template spaces such as MNI and Chinese2020), data stratification (slicing the data into specific numbers of layers according to cross-sections, coronal planes, and sagittal planes (e.g., 60 layers), feature calculation (summing the results of each layer for each dimension, then averaging according to the number of layers; for other modalities such as functional images, this also includes averaging at each time point to obtain results x, y, and z), and result calculation (applying the results of the three dimensions according to predetermined weights w). x w y w z Weighted processing is performed to obtain the processing result R for this image.

[0123] The standardized image data after processing can be combined with corresponding text descriptions to construct image-text pairs for training. Specifically, each processed image is paired with collected or manually labeled text content (such as "What is your brain age?" and its answer "Brain age 55 years old") to form a training data set.

[0124] To effectively utilize the aforementioned training data, this embodiment employs a phased model construction and training strategy. For example: a visual encoder (e.g., vit), an AD / MCI large language model with reasoning capabilities (e.g., a natural language model fine-tuned based on deepseek-r1-distill-qwen2.5-7b), and an alternative model (e.g., a model fine-tuned based on qwen2.5-7b-instruct) with reasoning capabilities are selected. An initial multimodal model (model) is constructed by building an MLP adapter (theta). initial = vis*theta*subs.

[0125] Step S182: Train the initial multimodal model using the training dataset to update the parameters of the MLP adapter, and replace the natural language model in the initial multimodal model with an AD / MCI large language model with inference capabilities to obtain a recombined model.

[0126] Based on the constructed initial multimodal model initial= vis * theta * subs, freezing the parameters of vis and subs, and optimizing only the parameters of theta. The optimization process is carried out based on existing collected visual-text (e.g., MRI-text) alignment instructions. After reaching a specified number of training epochs, the training process is stopped to obtain the theta model. Model recombination refers to replacing the natural language model in the initial multimodal model with an AD / MCI large language model with inference capabilities. In this scheme, subs is replaced with reas in the above multimodal model to obtain the recombinated model. new = vis*theta*reas.

[0127] Step S183: Obtain the instruction set with the inference process, and use the instruction set with the inference process to train the recombined model to obtain a small-parameter-scale multimodal model of the Agent.

[0128] In order to make the new model new It possesses multimodal reasoning capabilities, and by constructing a small set of instructions with reasoning processes, it can perform multimodal reasoning on the model. new Training is then performed. The process of constructing the instruction set is as follows: First, text content and visual image (brain MRI images) pairs are collected (selected from the dataset obtained in the above dataset construction steps). The text content includes text questions and answers, forming the initial dataset D. inital The process involves inputting text content and visual images into an existing multimodal model, such as GPT-4O. The task is to describe multiple thought processes leading to an answer based on the images and text, and to score the difficulty of the questions, thus creating multiple samples. After manual review, samples with correct reasoning processes are manually labeled and scored based on the length of their reasoning (shorter reasoning processes receive higher scores). This results in a dataset D comprising (text question, text answer, image, reasoning process, difficulty score, and quality score). new .

[0129] The training process is as follows: First, based on dataset D... initail For model new Fine-tuning training is performed to obtain the model. new1 Secondly, select data set D. new Samples with a quality score greater than a certain threshold constitute D new1 For model new1 Further fine-tuning yields the model new2 Select data set D new1 The samples that were incorrectly inferred during the previous fine-tuning process constitute a new dataset D. new2 For model new2 Further fine-tuning yields the modelnew3 Select data set D new The sample with the highest quality score constitutes D. new3 The GRPO method is used to process the model. new3 Enhanced training is performed to obtain the multimodal model of the Agent. new4 .

[0130] In implementing the above scheme, multi-dimensional hierarchical processing and weighted calculation of neuroimaging data (such as hierarchical and weighted averaging of transverse, coronal, and sagittal dimensions) can extract more comprehensive and representative image features, thereby significantly improving the model's ability to understand complex medical images. Furthermore, by adopting a two-stage strategy of initial model construction and model recombination (first freezing the visual encoder and optimizing the adapter with alternative model parameters, then replacing them with the inference model), and by training the recombinated model after constructing a small set of instructions with inference processes, the workload of constructing the training dataset and the complexity of multimodal model training are effectively reduced, while ensuring the model has good multimodal inference capabilities. This significantly reduces the consumption of training resources while maintaining model performance.

[0131] Please see Figure 2 The diagram shown is a structural schematic of the AD and MCI multidimensional knowledge discovery device provided in the embodiments of this application; the embodiments of this application provide an AD and MCI multidimensional knowledge discovery device 200, including: The problem acquisition module 210 is used to acquire natural language problems in the AD / MCI domain.

[0132] The information data determination module 220 is used to determine the multiple information dimensions required to answer the natural language question if the natural language question represents a multidimensional knowledge discovery intention.

[0133] The basic answer determination module 230 is used to determine the basic answer to a natural language question based on the data of the information dimension that ranks first in importance among multiple information dimensions.

[0134] The correct answer determination module 240 is used to determine the correct answer based on data from multiple information dimensions other than the information dimension with the highest importance.

[0135] The final answer determination module 250 is used to determine the final answer to the natural language question based on the basic answer and the corrected answer.

[0136] It should be understood that this device corresponds to the above-described embodiments of the AD and MCI multidimensional knowledge discovery methods and is capable of performing the various steps involved in the above method embodiments. The specific functions of this device can be found in the description above, and detailed descriptions are appropriately omitted here. The device includes at least one software functional module that can be stored in memory or embedded in the device's operating system (OS) in the form of software or firmware.

[0137] Please see Figure 3 The diagram shows a structural schematic of an electronic device provided in an embodiment of this application. An electronic device 300 provided in this application includes a processor 310 and a memory 320. The memory 320 stores machine-readable instructions executable by the processor 310. When the machine-readable instructions are executed by the processor 310, the method described above is performed.

[0138] This application embodiment also provides a computer-readable storage medium 330, on which a computer program is stored. This computer program is executed by a processor 310 to perform the methods described above. The computer-readable storage medium 330 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Read-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.

[0139] This application also provides a computer program product, including: a computer program or computer instructions, which are executed by a processor to perform the method described above.

[0140] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For apparatus embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0141] It should be understood that the disclosed apparatus and methods can also be implemented in other ways, as provided in the embodiments of this application. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in the flowchart or block diagram may represent a module, segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the accompanying drawings. For example, two consecutive blocks may actually be executed substantially in parallel, or they may sometimes be executed in reverse order, depending primarily on the functions involved.

[0142] Furthermore, the functional modules of each embodiment in this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part. In addition, in the description of this specification, the reference to terms such as "one embodiment," "some embodiments," "example," "specific example," "some examples," etc., means that the specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or more embodiments or examples. Furthermore, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0143] The above description is only an optional implementation of the embodiments of this application, but the protection scope of the embodiments of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the embodiments of this application should be covered within the protection scope of the embodiments of this application.

Claims

1. A multidimensional knowledge discovery method using AD and MCI, characterized in that, include: Acquiring natural language problems in the AD / MCI domain; If the natural language question represents a multidimensional knowledge discovery intent, then multiple agents are used to determine the multiple information dimensions required to answer the natural language question and the data for each information dimension; The basic answer to the natural language question is determined based on the information dimension that ranks first in importance among the multiple information dimensions. The corrected answer is determined based on the data from the information dimensions other than the information dimension that is ranked first in importance. The final answer to the natural language question is determined based on the basic answer and the corrected answer. The plurality of agents includes a question-generating agent and a process-summarizing agent. The process of determining the multiple information dimensions required to answer the natural language question and the data for each information dimension through these agents includes: The agent is generated by the aforementioned question, and the initial information dimension question is determined based on the natural language question. Each agent in the knowledge discovery team answers the initial information dimension question based on the natural language question, thereby obtaining multiple information dimension answers; the knowledge discovery team is a collaborative system built by expert agent type and data agent type agents, used to collaboratively execute multi-dimensional knowledge discovery tasks; The process is summarized by the Agent to summarize the answers of the multiple information dimensions, and to obtain the initial information dimensions and the data of the initial information dimensions. The Agent generates a new information dimension data question based on the initial information dimension, the data of the initial information dimension, and the natural language question. Obtain the answer to the new information dimension data question, and summarize the process by which the Agent obtains the final information dimension and the data of the final information dimension based on the initial information dimension, the data of the initial information dimension, and the answer to the new information dimension data.

2. The method according to claim 1, characterized in that, Before determining the multiple information dimensions required to answer the natural language question and the data for each information dimension through multiple agents, the method further includes: By planning the agent analysis of the natural language problem, the required expert agent type and data agent type are determined; The knowledge discovery team is constructed based on the expert agent type and the data agent type, and the knowledge discovery team is used to collaboratively execute multidimensional knowledge discovery tasks.

3. The method according to claim 2, characterized in that, The process of determining the basic answer to the natural language question based on the information dimension with the highest importance among the multiple information dimensions includes: The Agent generates an information dimension ranking problem based on the final information dimension and the data of the final information dimension. Each Agent in the knowledge discovery team generates individual ranking results for multiple information dimensions based on the final information dimension, the data of the final information dimension, and the information dimension ranking problem. The individual ranking results include importance scores. The process is summarized as follows: the Agent ranks the multiple information dimensions based on the average importance score of each information dimension, obtaining the information dimension ranking results; the information dimension ranking results include: information dimension ranking data and information dimension ranking answers; the information dimension ranking answers include: information dimension ranking and the average importance score of each information dimension. The Agent generated by the aforementioned question determines the basic question based on the sorting results of the aforementioned information dimensions. By using the data Agent corresponding to the first ordered information dimension in the information dimension sorting, the basic question is answered based on the data of that information dimension to obtain the basic answer.

4. The method according to claim 3, characterized in that, The process of determining the corrected answer based on data from information dimensions other than the information dimension with the highest importance among the multiple information dimensions includes: The Agent generated by the question determines the correction question based on the natural language question, the basic answer, and the information dimension sorting result. The correction rules are determined by using the data agents corresponding to other information dimensions besides the first sequential information dimension in the knowledge discovery team, based on the natural language question, the basic answer, the correction question, and the information dimension sorting results. From the information dimension ranking, select multiple second information dimensions other than the information dimension ranked first in importance; The weights of the information dimensions are calculated by weighting the plurality of second information dimensions according to the average importance score of each second information dimension. By using the correction rules and the weights of the information dimensions, the sub-corrected answers generated based on the data of each second information dimension are weighted and fused to obtain the corrected answer.

5. The method according to any one of claims 2-4, characterized in that, Determining the final answer to the natural language question based on the basic answer and the corrected answer includes: Each Agent in the knowledge discovery team conducts multiple rounds of verification on the base answer and the corrected answer. In each round of the multi-round argumentation, each Agent verifies the consistency between the base answer and the corrected answer according to the preset argumentation rules; When the basic answer and the corrected answer pass the consistency verification, the weighted fusion result of the basic answer and the corrected answer is determined as the final answer; When the base answer and the corrected answer fail the consistency verification, the weight of the information dimension is adjusted based on the result of this round of argumentation, and the process is returned to perform multiple rounds of argumentation on the base answer and the corrected answer by each Agent in the knowledge discovery team.

6. The method according to any one of claims 3-4, characterized in that, Determining the final answer to the natural language question based on the basic answer and the corrected answer includes: The agent generates an argumentation question based on the natural language question, the base answer, the corrected answer, the information dimension ranking result, the weight of the information dimension, and the previous round of argumentation results. Each Agent in the knowledge discovery team answers the argument question to obtain the argument answer content, which includes: the weight of the information dimension that has not been updated, or the weight of the information dimension that has been updated. By summarizing the process, the Agent performs consensus calculations on the argument response content to obtain the weight of the information dimension and the answer to the natural language question in this round, until the weight of the information dimension is no longer updated, or the maximum round threshold is reached, and the final answer to the natural language question is obtained.

7. The method according to any one of claims 2-4, characterized in that, Also includes: The basic answer, the corrected answer, and / or the questions generated during the process of determining the final answer are semantically extended using the small-parameter-scale multimodal model of each agent in the knowledge discovery team to obtain extended query questions. Preliminary entities are extracted from the extended query questions and matched with the lowest-level entities in the AD / MCI knowledge graph to obtain a preliminary set of matched entities. Based on the initial matching entity set, trace back to the lowest common ancestor entity in the AD / MCI knowledge graph; Using the lowest common ancestor entity as the root node and the initially matched entity as the leaf node, construct a subgraph containing entity paths and relationships; The Agent queries multiple candidate documents in the AD / MCI knowledge base based on the basic answer, the corrected answer, and / or questions generated during the process of determining the final answer, and selects relevant candidate documents with a relevance score greater than a threshold from the multiple candidate documents. The subgraph and the relevant candidate documents are transformed into a structured graph context for the Agent, so that the Agent can make inference decisions based on the structured graph context and obtain inference results. Obtain the reference annotations and add them to the inference results.

8. The method according to claim 7, characterized in that, Before semantically expanding the base answer, the corrected answer, and / or the questions generated during the determination of the final answer using a small-parameter-scale multimodal model of each agent in the knowledge discovery team, the method further includes: Acquire a training dataset and an initial multimodal model, which includes: a visual encoder, a multilayer perceptron (MLP) adapter, and a natural language model fine-tuned based on AD / MCI domain knowledge; With the parameters of the visual encoder and the natural language model both frozen, the initial multimodal model is trained using the training dataset to update the parameters of the MLP adapter. The natural language model in the initial multimodal model is replaced by an AD / MCI large language model with reasoning capabilities to obtain a recombined model; Obtain an instruction set with an inference process, and use the instruction set with the inference process to train the recombined model to obtain a small-parameter-scale multimodal model of the Agent.

9. A multidimensional knowledge discovery device for AD and MCI, characterized in that, include: The question acquisition module is used to acquire natural language questions in the AD / MCI domain; The information data determination module is used to determine, through multiple agents, multiple information dimensions required to answer the natural language question and the data for each information dimension if the natural language question represents a multidimensional knowledge discovery intention. The basic answer determination module is used to determine the basic answer to the natural language question based on the information dimension that ranks first in importance among the multiple information dimensions. The corrected answer determination module is used to determine the corrected answer based on the data of the other information dimensions besides the information dimension with the highest importance among the multiple information dimensions; The final answer determination module is used to determine the final answer corresponding to the natural language question based on the basic answer and the corrected answer; The plurality of agents includes a question-generating agent and a process-summarizing agent. The process of determining the multiple information dimensions required to answer the natural language question and the data for each information dimension through these agents includes: The agent is generated by the aforementioned question, and the initial information dimension question is determined based on the natural language question. Each agent in the knowledge discovery team answers the initial information dimension question based on the natural language question, thereby obtaining multiple information dimension answers; the knowledge discovery team is a collaborative system built by expert agent type and data agent type agents, used to collaboratively execute multi-dimensional knowledge discovery tasks; The process is summarized by the Agent to summarize the answers of the multiple information dimensions, and to obtain the initial information dimensions and the data of the initial information dimensions. The Agent generates a new information dimension data question based on the initial information dimension, the data of the initial information dimension, and the natural language question. Obtain the answer to the new information dimension data question, and summarize the process by which the Agent obtains the final information dimension and the data of the final information dimension based on the initial information dimension, the data of the initial information dimension, and the answer to the new information dimension data.

Images