A large model knowledge dynamic iterative updating method and system

Abstract

The application provides a large model knowledge dynamic iteration updating method and system, relates to the technical field of large models, collects multiple types of dynamic knowledge sources in the field of children's tumors and is attached with relevant information, and constructs a knowledge evolution life cycle network based on this; the knowledge content nodes are subjected to multi-source credibility cross verification and dynamic conflict reconciliation processing to obtain a target knowledge set; a parameter fine tuning and memory reservation collaborative mechanism is called to convert into a knowledge updating parameter package; the knowledge updating parameter package is input into a large model to perform collaborative updating operation and generate a knowledge updating whole-process traceability document. The application can comprehensively and accurately process multiple types of knowledge sources, effectively improve the knowledge accuracy and reliability, and realize model collaborative updating.

Description

Technical Field

[0001] This invention relates to the field of large model technology, and more specifically, to a method and system for dynamic iterative updating of knowledge in large models. Background Technology

[0002] In the field of pediatric oncology education, continuous updating and iteration of knowledge is crucial for improving teaching quality and cultivating professionals. Pediatric oncology is a constantly evolving discipline, with new research findings, treatment methods, and diagnostic techniques emerging frequently. Currently, knowledge sources in the field of pediatric oncology are extensive and diverse, including the latest medical guideline texts, full-text clinical research papers, medical conference minutes, and multicenter clinical case reports. However, existing methods for updating pediatric oncology teaching knowledge have several problems.

[0003] On the one hand, existing knowledge updating methods often lack the ability to comprehensively process multiple types of knowledge sources. Different types of knowledge sources differ in format, content, and credibility, and existing methods struggle to effectively integrate these sources, resulting in incomplete and inaccurate knowledge updates. For example, medical guideline texts and clinical research papers differ in their content focus and update frequency, and existing methods cannot process and integrate them in a targeted manner according to their characteristics.

[0004] On the other hand, there is a lack of effective mechanisms for handling knowledge credibility and conflicts during the knowledge updating process. Knowledge from different sources may contradict each other, and existing methods struggle to reconcile these conflicts, thus affecting the accuracy and reliability of the knowledge. Furthermore, existing knowledge updating methods often neglect the model's memory retention when integrating new knowledge into the teaching model, leading to the loss of some original knowledge during the updating process and affecting the coherence and stability of teaching. Summary of the Invention

[0005] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for dynamic iterative updating of knowledge in a large model, the method comprising:

[0006] Collect multiple types of dynamic knowledge sources in the field of pediatric oncology. These sources include the latest medical guidelines, full texts of clinical research papers, medical conference minutes, and multicenter clinical case summary reports. Each knowledge source is accompanied by information on the knowledge release time and the identifier of the knowledge source institution.

[0007] The various types of dynamic knowledge sources and their associated information are correlated to construct a knowledge evolution lifecycle network in the field of pediatric oncology. The knowledge evolution lifecycle network includes knowledge content nodes, knowledge state labels corresponding to the nodes, and temporal and logical association edges between the nodes.

[0008] Based on the knowledge evolution lifecycle network, multi-source credibility cross-validation and dynamic conflict reconciliation are performed on each knowledge content node to obtain a set of target knowledge that has passed the verification and is conflict-free.

[0009] The pre-built parameter fine-tuning and memory retention collaborative mechanism is invoked to convert the target knowledge set into a knowledge update parameter package adapted to the large-scale pediatric oncology teaching model. The knowledge update parameter package includes module parameter adjustment schemes, knowledge association constraint rules, and model memory anchoring parameters.

[0010] The knowledge update parameter package is input into the pediatric oncology teaching model, and a collaborative update operation is performed on the knowledge storage module, reasoning calculation module, and memory retention module of the pediatric oncology teaching model to generate a knowledge update full-process traceability document. The knowledge update full-process traceability document includes the updated content, the updated module, the corresponding knowledge node information, and the memory anchoring record.

[0011] Furthermore, embodiments of the present invention also provide a large-scale model knowledge dynamic iterative update system, comprising:

[0012] A processor; a machine-readable storage medium for storing machine-executable instructions of the processor; wherein the processor is configured to execute the above-described large model knowledge dynamic iterative update method by executing the machine-executable instructions.

[0013] Based on the above, this study collects multiple types of dynamic knowledge sources in the field of pediatric oncology, along with their publication dates and source institution identifiers. Then, it constructs a knowledge evolution lifecycle network by associating these sources with their associated information. This network structure presents the content, state, temporal relationships, and logical connections of knowledge, facilitating a deeper understanding of the knowledge's evolution and development. Based on this knowledge evolution lifecycle network, multi-source credibility cross-validation and dynamic conflict reconciliation are performed to obtain a set of validated and conflict-free target knowledge, effectively improving the accuracy and reliability of the knowledge and avoiding teaching misguidance caused by knowledge conflicts. A pre-built parameter fine-tuning and memory retention collaborative mechanism is invoked to convert the target knowledge set into a knowledge update parameter package adapted to the large-scale pediatric oncology teaching model. This parameter package comprehensively considers module parameter adjustment, knowledge association constraints, and model memory anchoring, achieving collaborative optimization of knowledge updates and model memory retention. The knowledge update parameter package is input into the large-scale model to perform collaborative update operations and generate a full-process traceability document for knowledge updates. This not only ensures the effectiveness of knowledge updates in the large-scale model but also provides detailed records for the traceability and review of knowledge updates, improving the transparency and manageability of knowledge updates. This significantly improved the quality of knowledge updates and the effectiveness of teaching applications in the large-scale pediatric oncology teaching model. Attached Figure Description

[0014] Figure 1This is a schematic diagram of the execution flow of the large-scale model knowledge dynamic iterative update method for pediatric oncology teaching provided in an embodiment of the present invention.

[0015] Figure 2 This is a schematic diagram of the execution process provided by the embodiments of the present invention, which determines the knowledge status of all initial knowledge units and historical knowledge units. If the knowledge content appears for the first time and has no related historical knowledge, it is marked as the knowledge budding state; if the knowledge content is published by at least two different institutions and there is no conflict, it is marked as the knowledge maturity stage; if the knowledge content modifies or supplements the historical knowledge and the publication time is later than the publication time of all related historical knowledge units, it is marked as the knowledge iteration state.

[0016] Figure 3 This is a schematic diagram of the execution process provided by an embodiment of the present invention, which organizes all knowledge units, knowledge status tags and related edges in a hierarchical structure according to the time dimension and the logical dimension, sorting them according to the publication time and classifying them according to the attributes of the related edges.

[0017] Figure 4 This is a schematic diagram of exemplary hardware and software components of a large-scale model knowledge dynamic iterative update system for pediatric oncology teaching provided in an embodiment of the present invention. Detailed Implementation

[0018] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for dynamically iteratively updating large-scale model knowledge for pediatric oncology teaching, provided by an embodiment of the present invention. The following is a detailed description of this method for dynamically iteratively updating large-scale model knowledge for pediatric oncology teaching.

[0019] Step S110: Collect multiple types of dynamic knowledge sources in the field of pediatric oncology. These multiple types of dynamic knowledge sources include the latest medical guideline texts, full texts of clinical research papers, medical conference minutes, and multi-center clinical case summary reports. Each knowledge source is accompanied by information on the knowledge release time and the identifier of the knowledge source institution.

[0020] In this embodiment, the application scenario throughout the text is the updating of knowledge on the diagnosis and treatment of childhood acute lymphoblastic leukemia (ALL). The knowledge acquisition module obtains relevant knowledge sources by establishing compliant data interfaces with medical guideline publishing platforms, academic databases, academic organizations, and multi-center clinical research institutions. For the latest medical guideline text, the updated ALL diagnosis and treatment guidelines are regularly retrieved from the medical guideline publishing platform designated by the National Health Commission. These guidelines include knowledge release time information and the knowledge source institution identification of the National Children's Medical Center Hematology and Oncology Specialty Alliance. For full-text clinical research papers, authorized interfaces are accessed through academic databases such as PubMed and Elsevier to retrieve papers published within the last six months whose titles contain keywords such as "childhood acute lymphoblastic leukemia," "diagnosis and treatment," and "treatment." Each paper includes knowledge release time information and the knowledge source institution identification of the corresponding university-affiliated children's hospital's hematology department. The medical conference minutes documents are obtained through collaboration with academic organizations such as the Hematology Group of the Pediatrics Branch of the Chinese Medical Association, summarizing the symposiums on childhood acute lymphoblastic leukemia held at their annual academic conferences. These minutes include information on the publication date and the identifier of the Hematology Group of the Pediatrics Branch of the Chinese Medical Association as the source organization for the knowledge. The multicenter clinical case summary reports are provided by medical institutions participating in the multicenter clinical studies, including the identifier of the source organization jointly identified by multiple participating institutions and information on the publication date.

[0021] The data collection process strictly adheres to relevant laws and regulations. For medical guideline texts, full-text clinical research papers, and medical conference minutes, the knowledge acquisition module establishes compliant data interfaces with the medical guideline publishing platform designated by the National Health Commission, academic databases such as PubMed and Elsevier, and academic organizations such as the Hematology Group of the Pediatrics Branch of the Chinese Medical Association. Data usage authorization agreements are signed, clearly defining the scope of data acquisition, usage restrictions, and intellectual property ownership, ensuring that all publicly available knowledge sources collected are legally authorized. For data that may involve patient privacy, such as multi-center clinical case summary reports, in addition to signing data cooperation agreements with participating medical institutions, each participating center is required to complete data anonymization locally, removing direct identifiers such as patient names, ID numbers, and hospital numbers. Indirect identifiers such as birth dates and consultation dates are processed using intervalization (e.g., only the year of birth and the year of consultation are retained). At the same time, a federated learning technology framework is adopted, with each center only performing model training-related parameter calculations locally. Multi-center data collaborative utilization is achieved through encrypted parameter aggregation, without involving cross-institutional transmission of original case data throughout the process. Before data collection, the medical institution's ethics committee reviews the data collection plan and issues an ethics approval document, clarifying that the purpose of data use is limited to knowledge updates for the pediatric oncology teaching model, and that the data processing process must comply with relevant regulations on data security management in medical and health institutions. The collection of the knowledge source institution's identification and knowledge release time information is also done through authorized interfaces to ensure data traceability and avoid the use of unauthorized third-party data or anonymous data.

[0022] Step S120: Perform association processing on the multi-type dynamic knowledge sources and accompanying information to construct a knowledge evolution lifecycle network in the field of pediatric oncology. The knowledge evolution lifecycle network includes knowledge content nodes, knowledge state labels corresponding to the nodes, and time association edges and logical association edges between nodes.

[0023] Step S121: Extract core knowledge statements from each multi-type dynamic knowledge source. The core knowledge statements are complete statements containing medical conclusions, treatment suggestions, or case analysis conclusions. Each core knowledge statement must cover at least one medical concept specific to the field of pediatric oncology.

[0024] For the collected dynamic knowledge sources related to the diagnosis and treatment of childhood acute lymphoblastic leukemia (ALL), a BERT-based medical text extraction model was used to process each knowledge source. This model was trained using a large amount of labeled data in the field of pediatric oncology and can identify key statements such as medical conclusions, treatment recommendations, and case analysis conclusions. For example, from the latest medical guidelines, the model can extract the treatment recommendation, "For standard-risk childhood ALL patients, induction remission therapy can be performed using vincristine combined with prednisone," which includes pediatric oncology-specific medical concepts such as "standard-risk childhood ALL," "induction remission therapy," "vincristine," and "prednisone." From full-text clinical research papers, medical conclusions such as "A certain novel chemotherapy drug combination regimen improves the complete remission rate in the treatment of childhood ALL compared to traditional regimens" are extracted, containing specific medical concepts such as "novel chemotherapy drug combination regimen" and "complete remission rate." From medical conference minutes, a treatment recommendation was extracted stating that "minimal residual disease (MRD) testing in children with acute lymphoblastic leukemia should be used as an important indicator for efficacy evaluation," involving concepts such as "minimal residual disease testing" and "efficacy evaluation." From multicenter clinical case reports, a case analysis conclusion was extracted stating that "event-free survival improved after individualized treatment regimens were adopted in children with acute lymphoblastic leukemia of a certain age group," encompassing specific medical concepts such as "individualized treatment regimens" and "event-free survival."

[0025] Step S122: Add the knowledge source organization identifier and knowledge release time information to each core knowledge representation to generate an initial knowledge unit containing content, organization, and time attributes.

[0026] After extracting the core knowledge statements, each core knowledge statement is linked to its corresponding source institution identifier and publication date. For example, a core knowledge statement extracted from a medical guideline published by the National Children's Medical Center Hematology and Oncology Specialty Alliance is linked with the institution identifier and the guideline's publication date; a core knowledge statement extracted from a clinical research paper published by the hematology department of a university-affiliated children's hospital is linked with the institution identifier of that department and the paper's publication date; a core knowledge statement extracted from the minutes of a meeting of the Hematology Group of the Pediatrics Branch of the Chinese Medical Association is linked with the academic organization identifier and the meeting's date; and a core knowledge statement extracted from a multi-center clinical case summary report is linked with the identifiers of the participating institutions and the report's completion date. Through this process, each core knowledge statement becomes an initial knowledge unit containing content, institution, and time attributes.

[0027] Step S123: Retrieve the historical knowledge base stored within the pediatric oncology teaching model, extract the historical core expression, historical source institution identifier, and historical release time information corresponding to each historical knowledge from the historical knowledge base, and generate a historical knowledge unit containing content, institution, and time attributes.

[0028] The historical knowledge base of the children's oncology teaching large model stores the previous diagnosis and treatment knowledge of childhood acute lymphoblastic leukemia. The knowledge retrieval module obtains the data in the historical knowledge base by accessing the storage interface of the model. For each historical knowledge, its historical core expression is extracted, such as "the five-year survival rate data of traditional chemotherapy regimens in the treatment of childhood acute lymphoblastic leukemia", and at the same time, the corresponding historical source institution identifier, such as the identifier of the medical institution or academic organization that released the knowledge in the past, and the historical release time information, such as the time when the knowledge was first entered into the model or the original release time, are obtained. The above information is combined to generate a historical knowledge unit with the same structure as the initial knowledge unit, including content, institution, and time attributes.

[0029] Step S124: Determine the knowledge status of all initial knowledge units and historical knowledge units. If the knowledge content appears for the first time and there is no associated historical knowledge, it is marked as the knowledge germination state; if the knowledge content is released by at least two different institutions and there is no conflict, it is marked as the knowledge maturity state; if the knowledge content revises or supplements historical knowledge and the release time is later than the release time of all associated historical knowledge units, it is marked as the knowledge iteration state.

[0030] Specific reference Figure 2 In step S1241, a knowledge content comparison library for the field of children's oncology is constructed. The knowledge content comparison library contains all historical core expressions in the historical knowledge base and the core expressions of the initial knowledge units with marked status.

[0031] In the construction process of the knowledge content comparison library, first, all historical core expressions in the historical knowledge base are collected. These expressions cover various previous medical conclusions, diagnosis and treatment suggestions, and case analysis conclusions in the field of childhood acute lymphoblastic leukemia diagnosis and treatment. At the same time, the core expressions of the initial knowledge units that have completed status marking are also included to form a comprehensive knowledge content set. To improve the efficiency and accuracy of subsequent semantic similarity retrieval, the above core expressions are preprocessed, including word segmentation, removal of stop words (such as meaningless words like "of", "in", "for", etc.), and standardization of medical terms (unifying different expressions of the same medical concept into standard terms). Then, the processed core expressions are converted into vector representations using the Word2Vec model and stored in the knowledge content comparison library for fast semantic similarity calculation.

[0032] Step S1242: Input the core knowledge expression of each knowledge unit to be determined into the knowledge content comparison library and perform semantic similarity retrieval, with the retrieval scope covering all stored core expressions.

[0033] For each knowledge unit requiring state determination (including initial knowledge units and historical knowledge units without labeled states), its core knowledge representation is input into the knowledge content comparison database. The comparison database uses a cosine similarity algorithm to calculate the semantic similarity between the core knowledge representation to be determined and all stored core representations in the database. During the calculation, the core knowledge representation to be determined is first preprocessed using the same methods as when constructing the comparison database, then converted into a vector representation, and finally cosine similarity is calculated with all vectors in the database to obtain the similarity value between each stored core representation and the core knowledge representation to be determined.

[0034] Step S1243: If the search results show that there are no stored core expressions with semantic similarity higher than the preset threshold, then it is determined that the knowledge unit has no associated historical knowledge and is marked as knowledge in its nascent stage.

[0035] The preset semantic similarity threshold is set based on the characteristics of knowledge in the field of pediatric oncology and expert experience, for example, 0.3 (this is just an example; in actual applications, it will be determined based on data statistics and experiments). When the semantic similarity value between the core knowledge statement of the knowledge unit to be judged and all stored core statements in the knowledge content comparison database is lower than this threshold, it indicates that there is no similar knowledge in the historical knowledge base, and it is the first appearance of this knowledge content. At this time, the knowledge unit is marked as a knowledge nascent stage. For example, if a core knowledge statement is a description of a brand-new gene detection technology for childhood acute lymphoblastic leukemia, and no similar statement can be found in the comparison database, it is marked as a knowledge nascent stage.

[0036] Step S1244: If the search results show that there are at least two core statements published by different institutions, and the semantic similarity between the core statements and the core knowledge statements of the knowledge unit to be judged is higher than a preset threshold, and the content of the statements is not contradictory, then retrieve the institution identifiers corresponding to the core statements, confirm that the institution identifiers are different from each other, determine that the knowledge unit meets the multi-institution consensus condition, and mark it as the knowledge maturity stage.

[0037] When at least two stored core expressions in the search results have a semantic similarity higher than a preset threshold with the core knowledge expression to be determined, it is necessary to further check whether the source institution identifiers of these stored core expressions are different and whether there are contradictions in the expression content. For example, if the core knowledge expression to be determined is "the recommended dosage range of a certain chemotherapy drug", and core expressions from institution A and institution B are found, both with semantic similarity higher than the threshold and both recommending similar dosage ranges without content conflict, then the identifiers of institution A and institution B are retrieved to confirm that they are different institutions, such as one being a university-affiliated children's hospital and the other a local children's medical center. In this case, the knowledge unit to be determined is determined to meet the multi-institution consensus condition and is marked as knowledge maturity.

[0038] Step S1245: If the search results show that the core expression of the historical knowledge unit has a semantic relationship with the core knowledge expression of the knowledge unit to be judged, and the core knowledge expression of the knowledge unit to be judged contains a modified or supplementary statement of the historical core expression, then extract the knowledge publication time information of the knowledge unit to be judged and the historical publication time information of the historical knowledge unit.

[0039] If the semantic similarity between the core knowledge statement to be determined and the historical core statement is higher than a preset threshold, but not completely identical, and the core knowledge statement to be determined contains words indicating correction or supplementation such as "revised to," "supplementary explanation," or "updated to," it indicates that the knowledge content may be an update of historical knowledge. For example, if the historical core statement is "The incidence rate of a certain complication of childhood acute lymphoblastic leukemia is X," and the core knowledge statement to be determined is "The incidence rate of a certain complication of childhood acute lymphoblastic leukemia is revised to Y (Y≠X)," then a correction relationship exists. In this case, the knowledge publication time information of the knowledge unit to be determined and the historical publication time information of the corresponding historical knowledge unit are extracted.

[0040] Step S1246: Compare the knowledge release time information of the knowledge unit to be judged with the historical release time information of the historical knowledge units. If the release time of the knowledge unit to be judged is later than the release time of all related historical knowledge units, then the knowledge unit is judged to be the latest updated knowledge and marked as the knowledge iteration state.

[0041] The publication time of the knowledge unit to be judged is compared with the publication times of all related historical knowledge units. If the publication time of the knowledge unit to be judged is after all related historical knowledge units, it means that it is the latest update based on historical knowledge, and is therefore marked as a knowledge iteration state. For example, if the historical knowledge unit was published in 2020, the knowledge unit to be judged was published in 2023, and the historical knowledge has been revised, then it is marked as a knowledge iteration state.

[0042] Step S1247: For each marked knowledge unit, record the search results, organization identifiers and time comparison results involved in the judgment process, generate a knowledge status judgment basis document, and associate and bind the judgment basis document with the knowledge unit.

[0043] After completing the status labeling of knowledge units, key information from the judgment process is recorded in detail. For knowledge units in the nascent stage, all retrieved similarity values ​​and thresholds are recorded; for knowledge units in the mature stage, the core similar expressions, corresponding institutional identifiers, and similarity values ​​are recorded; for knowledge units in the iterative stage, the core expressions of associated historical knowledge units, comparison results of publication times, and specific statements of corrections or additions are recorded. This information is compiled into a document outlining the criteria for knowledge status judgment, and this document is linked to the corresponding knowledge unit using a unique identifier and stored in a database for subsequent querying and tracing.

[0044] Step S125: Classify each knowledge unit according to its knowledge state label to form a set of knowledge nascent stage units, a set of knowledge mature stage units, and a set of knowledge iterative stage units.

[0045] Based on the knowledge state labels of knowledge units (knowledge nascent stage, knowledge maturity stage, and knowledge iteration stage), all knowledge units are categorized into three different sets. For example, all knowledge units marked as nascent stage are integrated into the nascent stage knowledge unit set, and knowledge units in the knowledge maturity and iteration stages are similarly processed to form corresponding sets. This classification facilitates targeted processing and analysis of knowledge units in different stages.

[0046] Step S126: For each unit in the knowledge budding unit set, query whether there are any related knowledge units that will be published later. If so, establish a time-related edge from the budding unit to the subsequent unit, and label the edge attribute as time continuation.

[0047] For each knowledge unit in the set of nascent knowledge units, a query is performed among all knowledge units (including initial and historical knowledge units) to determine if there exists a knowledge unit whose publication time is later than that of the nascent knowledge unit and which has a semantic connection with it (semantic similarity higher than a preset threshold), i.e., a subsequently published related knowledge unit. If such a knowledge unit exists, a directed edge is established in the knowledge evolution lifecycle network from the nascent knowledge unit to the subsequent related knowledge unit. This edge is called a temporal association edge, and its attribute is labeled "temporal continuity," indicating the continuity of knowledge over time. For example, if a nascent knowledge unit about a novel gene detection technology is subsequently published, and a knowledge unit researching the application effects of this technology is published, a temporal association edge is established from the former to the latter.

[0048] Step S127: For each unit in the knowledge maturity unit set, query whether there are other organizations that have published the same core expression unit. If so, establish logical connection edges between units and mark the edge attributes as consensus verification.

[0049] Within the knowledge maturity unit set, for each knowledge unit, a query is performed to determine if any other knowledge units (i.e., units with the same core expression) published by different institutions exist in the knowledge content comparison database, provided that their core expression has a semantic similarity to the core expression of this unit exceeding a preset threshold and their content is consistent. If such units exist, logical connections are established in the network linking these units, with the edge attribute labeled "consensus verification," representing the consensus relationship between different institutions regarding the same knowledge content. For example, if two knowledge units published by different institutions regarding the efficacy of the same chemotherapy regimen have similar core expressions and are consistent, a logical connection for consensus verification is established between them.

[0050] Step S128: For each unit in the knowledge iteration state unit set, query its corrected or supplemented historical knowledge units, establish logical connection edges from historical units to iteration state units, and label the edge attributes as knowledge update.

[0051] For each knowledge unit in the set of iterative knowledge units, based on the correction or supplementation relationship determined during its knowledge state determination process, the corresponding historical knowledge unit that was corrected or supplemented is found. Then, a directed logical connection edge is established in the knowledge evolution lifecycle network from the historical knowledge unit to the iterative unit, with the edge attribute labeled "knowledge update," representing the knowledge update iteration relationship. For example, if an iterative knowledge unit corrects the dosage of a treatment plan in a historical knowledge unit, then a knowledge update logical connection edge is established from that historical knowledge unit to the iterative unit.

[0052] Step S129: Organize all knowledge units, knowledge status tags, and related edges in a hierarchical structure according to the time dimension and the logical dimension. The time dimension is sorted according to the publication time, and the logical dimension is classified according to the attributes of the related edges.

[0053] After establishing the state labels and associated edges of knowledge units, it is necessary to organize all knowledge units, knowledge state labels, and associated edges in a hierarchical and structured manner in order to construct a clear knowledge evolution lifecycle network.

[0054] For details, please refer to the following: Figure 3 Step S1291: Create a time-dimensional hierarchical framework. The time-dimensional hierarchical framework is divided into multiple time layers according to time intervals. Each time layer corresponds to a fixed time interval. The length of the time interval is determined based on the average knowledge update cycle in the field of pediatric oncology.

[0055] The average knowledge update cycle in the field of pediatric oncology is obtained through statistical analysis of historical knowledge release time data. For example, the average time interval between important knowledge updates in the field of childhood acute lymphoblastic leukemia over the past ten years is calculated as the average knowledge update cycle. Assuming this average cycle is 2 years (this is just an example), the time-dimensional hierarchical framework is divided into multiple time layers based on 2-year time intervals. Each time layer corresponds to a fixed 2-year time interval, such as 2018-2020, 2021-2023, etc. (the start and end years of the time interval are determined according to the actual situation).

[0056] Step S1292: Assign each knowledge unit to the corresponding time layer according to its knowledge release time information or historical release time information. The knowledge units in the same time layer are arranged in chronological order of release time.

[0057] Based on the knowledge release time information (for initial knowledge units) or historical release time information (for historical knowledge units), the time interval to which the knowledge unit belongs is determined, and then it is assigned to the corresponding time layer. For example, a knowledge unit released in 2022 is assigned to the 2021-2023 time layer. Within the same time layer, all knowledge units are arranged in chronological order according to their specific release time (accurate to the month or day) to clearly show the release order of knowledge within that time layer.

[0058] Step S1293: Within each time layer, establish a knowledge unit index table, which includes a knowledge unit identifier, a core knowledge description summary, a knowledge status label, and a knowledge source organization identifier.

[0059] A knowledge unit index table is created for each time layer to facilitate quick querying and locating of knowledge units within that time layer. The knowledge unit identifier is a unique identifier assigned to each knowledge unit, such as a string consisting of letters and numbers. The core knowledge statement summary is a concise overview of the core knowledge statement, extracting key information to facilitate quick understanding of the knowledge content. The knowledge status label indicates whether the knowledge unit is in its "knowledge nascent stage," "knowledge maturity stage," or "knowledge iteration stage." The knowledge source institution identifier indicates the institution from which the knowledge originates. This information is organized into an index table and stored in the metadata of the time layer.

[0060] Step S1294: Create a logical dimension classification framework, which is divided into three logical categories according to the associated edge attributes: time continuity, consensus verification, and knowledge update.

[0061] The logical dimension classification framework is constructed based on the attribute types of associated edges. Since associated edges have three attributes—"time continuity," "consensus verification," and "knowledge update"—the logical dimension is divided into three corresponding logical categories: time continuity, consensus verification, and knowledge update. Each logical category is used to centrally manage associated edges with the corresponding attributes.

[0062] Step S1295: All time-related edges and logically related edges are classified into their corresponding logical categories according to their attributes. A list of related edges is established under each logical category. The list of related edges includes edge identifier, starting knowledge unit identifier, ending knowledge unit identifier, and edge attribute description.

[0063] All connected edges in the network (including temporally connected edges and logically connected edges) are categorized into corresponding logical categories within the logical dimension classification framework based on their attributes (time continuity, consensus verification, knowledge update). Under each logical category, a list of connected edges is created, recording detailed information for all connected edges within that category. The edge identifier is a unique identifier for the connected edge; the starting and ending knowledge unit identifiers are the identifiers of the two knowledge units connected by the connected edge, respectively; for directed edges, the starting and ending directions are explicitly defined; the edge attribute description is the attribute of the connected edge (e.g., "time continuity"). For example, all connected edges with the attribute "consensus verification" are categorized into the "consensus verification" logical category, and the edge identifier, starting and ending knowledge unit identifiers, and the "consensus verification" attribute description for each edge are recorded in the connected edge list under this category.

[0064] Step S1296: Establish a mapping relationship between the time-dimensional hierarchical framework and the logical-dimensional classification framework, and record the relationship between the position of each knowledge unit in the time layer and the affiliation of the associated edge of the knowledge unit in the logical category.

[0065] To organically combine the time and logical dimensions, a mapping relationship needs to be established between the two frameworks. For each knowledge unit, its time layer position (i.e., which time interval it belongs to) and all related edges in which it participates as a starting or ending node are recorded. These related edges are then assigned to which logical category within the logical dimension classification framework. For example, a knowledge unit located in the 2021-2023 time layer participates as a starting node in a related edge belonging to the consensus verification category and as an ending node in a related edge belonging to the knowledge update category. This information needs to be linked and recorded to form a cross-index, allowing the relevant related edges to be found through the time layer, and the corresponding time layer knowledge units to be found through the related edges.

[0066] Step S1297: Construct a graph retrieval interface, which supports three retrieval methods: querying knowledge units by time layer, querying related edges by logical category, and filtering knowledge units by knowledge status label.

[0067] The graph retrieval interface is the interface through which users or systems interact with the knowledge evolution lifecycle network. The function to query knowledge units by time layer allows users to input a time layer identifier (such as a time interval) to retrieve information on all knowledge units within that time layer, including the knowledge unit identifier, a summary of the core knowledge description, etc., and further view the detailed content and associated edges of the knowledge unit. The function to query associated edges by logical category allows users to select a logical category (time continuation, consensus verification, knowledge update) to retrieve a list of associated edges for all associated edges under that category, and view information such as edge identifiers, starting and ending knowledge unit identifiers, etc. The function to filter knowledge units by knowledge state label allows users to select a knowledge state label (knowledge nascent stage, knowledge maturity stage, knowledge iteration stage) to retrieve all knowledge units with that label, and further filter by time layer or other conditions.

[0068] Step S1298: Integrate the time-dimensional hierarchical framework, the logical-dimensional classification framework, the mapping relationship, and the graph retrieval interface to form a structured organization result of the knowledge evolution life cycle network.

[0069] The aforementioned time-dimensional hierarchical framework (including indexes of each time layer and knowledge unit), logical-dimensional classification framework (including lists of logical categories and associated edges), mapping relationships between the time and logical dimensions, and graph retrieval interface are systematically integrated to form a complete structured organization of a knowledge evolution lifecycle network. This structured organization is stored in a graph database (such as Neo4j), with knowledge units as nodes and associated edges as edges. The attributes of nodes and edges store the attributes of knowledge units (content, structure, time, status labels, etc.) and the attributes of associated edges (edge ​​attribute descriptions, etc.), respectively. The mapping relationships are implemented through the graph database's index and relation queries, and the graph retrieval interface implements the corresponding retrieval functions through the graph database's query language (such as Cypher).

[0070] Step S130: Based on the knowledge evolution lifecycle network, perform multi-source credibility cross-validation and dynamic conflict reconciliation processing on each knowledge content node to obtain a set of target knowledge that has passed the verification and is conflict-free.

[0071] Step S131: Extract the knowledge source institution identifier corresponding to each knowledge content node from the knowledge evolution lifecycle network, and retrieve the institution qualification description from the institution professional archive based on the knowledge source institution identifier. The institution qualification description includes the institution's scope of diagnosis and treatment in the field of pediatric oncology, research project experience and industry certification status.

[0072] Each knowledge content node in the knowledge evolution lifecycle network is associated with a knowledge source institution identifier. This identifier accesses the institution's professional archive, which stores detailed qualification information for various medical institutions in the field of pediatric oncology. For each knowledge content node, the qualification description of its knowledge source institution is retrieved, including the institution's scope of practice in pediatric oncology (e.g., whether it has a dedicated pediatric hematology-oncology department, and the types of pediatric tumors it can treat), research project experience (e.g., national and provincial-level research projects on pediatric oncology undertaken in the past five years, project names, start and end dates, and research content), and industry certifications (e.g., whether it has passed the National Key Clinical Specialty Construction Project certification, and whether it has obtained quality certifications related to pediatric oncology diagnosis and treatment). For example, a knowledge content node with a knowledge source institution identifier of a university-affiliated children's hospital retrieves the hospital's pediatric hematology-oncology department's scope of practice (treating various hematological malignancies such as childhood acute lymphoblastic leukemia), research project experience, and industry certifications (core unit of the National Children's Medical Center, National Key Clinical Specialty (Pediatrics), etc.) from the institution's professional archive.

[0073] Step S132: Extract the knowledge state label corresponding to each knowledge content node from the knowledge evolution life cycle network. If the label is in the knowledge budding stage, the research project experience of the knowledge source institution will be checked. If the label is in the knowledge maturity stage, the industry certification of the institution will be checked. If the label is in the knowledge iteration stage, the scope of diagnosis and treatment and research project experience of the institution will be checked at the same time.

[0074] Based on the knowledge status tags of knowledge content nodes (knowledge in its nascent stage, knowledge in its mature stage, and knowledge in its iterative stage), the key points for verifying the qualifications of the institutions that provide the knowledge are determined. For knowledge content nodes in the nascent stage, since the knowledge content may be cutting-edge and exploratory, the focus is on verifying the institution's research project experience to see if the institution has an in-depth research foundation in the relevant field and whether it has the research capabilities to generate this nascent knowledge. For knowledge content nodes in the mature stage, whose knowledge content has been agreed upon by multiple institutions, the focus is on verifying the institution's industry certification status to see the institution's authority and recognition in the industry, to ensure the reliability of the consensus knowledge it publishes. For knowledge content nodes in the iterative stage, which involve the correction or supplementation of historical knowledge, the institution needs to have both rich clinical experience (demonstrated through the scope of practice) and solid research capabilities (demonstrated through research project experience). Therefore, both the institution's scope of practice and research project experience are verified simultaneously.

[0075] Step S133: Extract supporting information citation identifiers from the core knowledge descriptions corresponding to the knowledge content nodes. The supporting information citation identifiers are clinical case numbers, experimental research numbers, or literature index numbers.

[0076] Core knowledge statements typically cite supporting information to support their conclusions, such as clinical cases, experimental research data, or references. Natural language processing techniques (such as regular expression matching and named entity recognition) are used to extract the citation identifiers of this supporting information from the core knowledge statement. Clinical case numbers are unique identifiers for specific cases in multi-center clinical case summary reports; experimental research numbers are the numbers of experimental research projects initiated internally or externally; and literature index numbers are unique identifiers (such as PMID numbers) of cited academic literature in databases such as PubMed. For example, the core knowledge statement "A certain treatment regimen achieved a remission rate of X% in children with acute lymphoblastic leukemia" might cite "Experimental Research Number: SY2022-015" or "Literature Index Number: PMID: 35678901" as supporting information citation identifiers.

[0077] Step S134: Based on the supporting information citation identifier, retrieve the corresponding supporting information original text from the multi-center clinical case database, experimental research results database, and medical literature database. The supporting information original text includes clinical case diagnosis and treatment records, experimental data collection process, and core conclusions of the literature.

[0078] Based on the extracted supporting information citation identifiers, the original supporting information is retrieved from the corresponding databases. For clinical case numbers, the multi-center clinical case database is accessed to retrieve the corresponding clinical case diagnosis and treatment records, including the patient's basic information (age, gender, medical history, etc.), tumor diagnosis information (diagnostic basis, type, stage, etc.), treatment process (treatment plan, drug dosage, treatment cycle, etc.), and efficacy evaluation results (changes in laboratory test indicators, imaging examination results, remission status, etc.). For experimental research numbers, the experimental research results database is accessed to retrieve the corresponding experimental data collection process, including the experimental design (research objective, inclusion and exclusion criteria for research subjects, sample size, grouping method, intervention measures, observation indicators, etc.), sample collection and... Processing methods (such as blood sample collection time, storage conditions, and measurement methods for detection indicators), data measurement methods (model of testing instrument used, reagent brand, operating procedures, etc.), and result statistics methods (statistical software, statistical methods, P-value settings, etc.). For literature index numbers, access medical literature databases (such as PubMed), retrieve the full text of the literature corresponding to that number, and extract the core conclusions of the literature, including research objectives, research methods (similar to the experimental design scheme in the experimental data collection process, but with a greater emphasis on the overall research methods of the literature), results (statistical results of the main research indicators), and the process of derivation of conclusions (how to draw the final conclusions based on the research results).

[0079] Step S135: Perform a completeness analysis on the supporting information in the original text. Clinical case diagnosis and treatment records should include basic patient information, tumor type, treatment plan and efficacy evaluation results; experimental data collection process should include experimental design, sample size, data measurement methods and result statistics; core conclusions of the literature should include research objectives, research methods and conclusion derivation process.

[0080] The retrieved supporting information should be thoroughly checked for completeness to ensure it contains all necessary elements. For clinical case records, this includes checking for patient basic information (age, gender, etc., anonymized), tumor type (clearly diagnosed as childhood acute lymphoblastic leukemia and its subtype), treatment plan (specific drug names, dosages, usage, treatment duration, etc.), and efficacy evaluation results (e.g., descriptions of complete remission, partial remission, no remission, etc., and relevant laboratory test data such as complete blood count and bone marrow morphology). For experimental data collection, this includes checking for the following: experimental design (clearly describing the research hypothesis, intervention measures, and control settings), sample size (clearly specifying the number of subjects included, such as the number of cell samples, animal models, or clinical patients), data measurement methods (detailed explanation of how to measure observation indicators, such as the specific steps for using flow cytometry to detect minimal residual disease), and statistical methods (explaining how to perform statistical analysis, such as using t-tests to compare differences between groups, and using Kaplan-Negative cytometry). (Meier method for survival analysis, etc.); for the core conclusions of the literature, check whether it includes the research objective (clearly stating the problem the literature aims to solve), research methods (briefly describing the research design, research subjects, interventions, etc.), and the conclusion derivation process (the logical reasoning process from the research results to the conclusions, whether potential confounding factors were considered, and whether the conclusions are statistically and clinically significant, etc.). If the supporting information in the original text lacks any of the above necessary elements, it is judged to be incomplete.

[0081] Step S136: Associate the institutional qualification description with the results of the integrity analysis of the original supporting information to generate an initial credibility assessment report for each knowledge content node.

[0082] Based on the analysis results of the integrity of the original text of the comprehensive institutional qualification description and supporting information, an initial credibility assessment is conducted on the knowledge content nodes.

[0083] Step S1361: Construct an institutional qualification assessment index system. The institutional qualification assessment index system includes three indicators: matching degree of diagnosis and treatment scope, relevance of scientific research projects, and industry certification level. Each indicator corresponds to multiple assessment levels, and the assessment level is determined based on the specific content in the institutional qualification description.

[0084] The institutional qualification assessment index system is used to quantitatively evaluate the qualifications of institutions providing knowledge. The scope of treatment matching index assesses the degree of matching between the institution's scope of treatment and the types of tumors covered in the knowledge content, categorized into three levels: complete match, partial match, and no match. Complete match indicates that the institution's scope of treatment explicitly includes the pediatric tumor types covered in the knowledge content; partial match indicates that the institution's scope of treatment includes similar types, but not completely identical; no match indicates that the institution's scope of treatment does not cover that type. The research project relevance index assesses the degree of relevance between the institution's research projects and the research direction of the knowledge content, categorized into four levels: highly relevant, moderately relevant, lowly relevant, and irrelevant. Highly relevant indicates that the institution has undertaken research projects completely consistent with the research direction of the knowledge content within the past five years; moderately relevant indicates that the institution has undertaken related but not completely identical projects; lowly relevant indicates that the projects are somewhat related but weakly related; and irrelevant indicates no related projects. The industry certification level index assesses the authority of the institution's industry certification, determined based on the certification level obtained by the institution.

[0085] Step S1362: For each knowledge content node, the institution's qualification description is evaluated according to the institution qualification assessment index system. The treatment scope matching index determines whether the institution's treatment scope is consistent with the tumor types involved in the knowledge content. The scientific research project relevance index determines whether the institution's past scientific research projects are the same as the research direction of the knowledge content. The industry certification level index determines whether the industry certification obtained by the institution covers the relevant fields of the knowledge content.

[0086] Based on the institutional qualification assessment index system, the qualification description of the knowledge source institution for each knowledge content node is assessed using index levels. Regarding the matching degree of treatment scope, the pediatric tumor types involved in the knowledge content (e.g., childhood acute lymphoblastic leukemia) are compared with the institution's treatment scope. If the institution explicitly treats this type, it is considered a complete match; if the institution's treatment scope is pediatric hematologic diseases, without explicitly mentioning this subtype but including acute leukemia, it is considered a partial match; if the institution primarily treats solid tumors and does not involve hematologic malignancies, it is considered a mismatch. Regarding the relevance of research projects, the research direction of the knowledge content (e.g., targeted therapy for childhood acute lymphoblastic leukemia) is compared with the research content of projects in the institution's research project experience. If a project is entirely focused on this research direction, it is highly relevant; if the project's research direction is the treatment of childhood leukemia, including targeted therapy but not limited to this subtype, it is moderately relevant; if the project only involves basic research on pediatric tumors and is not closely related to targeted therapy, it is poorly relevant; if there are no relevant projects, it is irrelevant. To determine the industry certification level, check the industry certifications obtained by the organization. If it has obtained the National Key Clinical Specialty (Pediatric Hematology) certification, it is a national-level certification; if it has obtained the Provincial Key Clinical Specialty certification, it is a provincial-level certification; if it has obtained the Municipal Key Specialty certification, it is a municipal-level certification; if there is no relevant certification, it is no certification.

[0087] Step S13633: Construct a completeness assessment index system for supporting information. The completeness assessment index system for supporting information includes three indicators: completeness of clinical case elements, completeness of experimental data records, and completeness of literature conclusion derivation. Each indicator corresponds to multiple assessment levels, and the assessment level is determined based on the content coverage of the original supporting information.

[0088] The supporting information completeness assessment index system is used to evaluate the completeness of the original supporting information. The clinical case element completeness index assesses the completeness of clinical case records containing necessary elements, categorized into three levels: complete, relatively complete, and incomplete. Complete indicates the inclusion of all elements including basic patient information, tumor type, treatment plan, and efficacy evaluation results; relatively complete indicates the absence of one element; and incomplete indicates the absence of two or more elements. The experimental data record completeness index assesses the completeness of experimental data collection processes containing necessary elements, categorized into three levels: complete, relatively complete, and incomplete. Complete indicates the inclusion of all elements including experimental design, sample size, data measurement methods, and result statistical methods; relatively complete indicates the absence of one element; and incomplete indicates the absence of two or more elements. The literature conclusion derivation completeness index assesses the completeness of the core conclusions of a literature, categorized into three levels: complete, relatively complete, and incomplete. Complete indicates the inclusion of all elements including research objectives, research methods, and the conclusion derivation process; relatively complete indicates the absence of one element; and incomplete indicates the absence of two or more elements.

[0089] Step S1364: For the supporting information original text corresponding to each knowledge content node, determine the indicator level one by one according to the supporting information completeness assessment indicator system. The clinical case element completeness indicator determines whether the case record contains all necessary elements. The experimental data record completeness indicator determines whether the experimental process record covers key steps. The literature conclusion derivation completeness indicator determines whether the literature conclusion has a derivation process to support it.

[0090] Based on the supporting information completeness assessment index system, the index level of the original supporting information for each knowledge content node is determined. For the completeness of clinical case elements, the system checks whether the clinical case diagnosis and treatment record includes basic patient information, tumor type, treatment plan, and efficacy evaluation results. If all four elements are included, it is considered complete; if one is missing (e.g., a specific indicator in the efficacy evaluation results is missing), it is considered relatively complete; if two are missing (e.g., both tumor type and treatment plan are missing), it is considered incomplete. For the completeness of experimental data records, the system checks whether the experimental data collection process includes the experimental design, sample size, data measurement methods, and result statistical methods. If all four elements are included, it is considered complete; if one is missing, it is considered relatively complete; if two or more are missing, it is considered incomplete. For the completeness of literature conclusion derivation, the system checks whether the core conclusion of the literature includes the research objective, research methods, and conclusion derivation process. If all three elements are included, it is considered complete; if one is missing, it is considered relatively complete; if two or more are missing, it is considered incomplete.

[0091] Step S1365: Based on the evaluation levels of each indicator in the institutional qualification assessment indicator system and the evaluation levels of each indicator in the supporting information integrity assessment indicator system, and according to the preset credibility assessment rules, generate the initial credibility assessment conclusion for each knowledge content node. The credibility assessment conclusion is either in line with the preset basic credibility standard or does not meet the preset basic credibility standard.

[0092] The pre-defined credibility assessment rules are a set of judgment criteria developed based on expert experience and domain knowledge. For example, the rules stipulate that: if the knowledge status label is "knowledge maturity stage," the industry certification level in the institutional qualification assessment is national or provincial / ministerial level, and at least one indicator in the supporting information integrity assessment is complete, then it is judged to meet the pre-defined basic credibility standard; if the knowledge status label is "knowledge iteration stage," the matching degree of the scope of diagnosis and treatment in the institutional qualification assessment is complete, and the relevance of the research project is highly or moderately relevant, and the completeness of experimental data records or the completeness of literature conclusion derivation in the supporting information integrity assessment is complete, then it is judged to meet the pre-defined basic credibility standard; if the knowledge status label is "knowledge nascent stage," the relevance of the research project in the institutional qualification assessment is highly relevant, and the completeness of experimental data records in the supporting information integrity assessment is complete or relatively complete, then it is judged to meet the pre-defined basic credibility standard. For knowledge content nodes that do not meet the above corresponding rules, they are judged to not meet the pre-defined basic credibility standard. Based on these rules, combined with the institutional qualification assessment indicator level and the supporting information integrity assessment indicator level, an initial credibility assessment conclusion is generated for each knowledge content node.

[0093] Step S1366: Integrate the institutional qualification description, the results of the analysis of the completeness of the original supporting information, the evaluation level of each indicator, and the initial credibility assessment conclusion generated according to the credibility assessment rules to form an initial credibility assessment report for each knowledge content node.

[0094] The following information is integrated to form a complete initial credibility assessment report, which serves as the preliminary basis for assessing the credibility of the knowledge content node: the description of the institution's qualifications (scope of treatment, research project experience, industry certification status), the results of the analysis of the completeness of the original supporting information (type of supporting information, whether it contains necessary elements, etc.), the assessment levels of each indicator of the institution's qualifications (matching degree of scope of treatment, relevance of research projects, industry certification level), the assessment levels of each indicator of the completeness of supporting information (completeness level of clinical case elements, completeness level of experimental data records, completeness level of literature conclusion derivation), and the initial credibility assessment conclusion (whether it meets or does not meet the preset basic credibility standards).

[0095] Step S137: Select knowledge content nodes that meet the preset basic credibility standards in the initial credibility assessment report as candidate credibility nodes.

[0096] Based on the conclusions of the initial credibility assessment report, all knowledge content nodes that meet the preset basic credibility standards are selected to form a candidate credibility node set. These nodes are knowledge content nodes that have undergone preliminary screening and have a certain degree of credibility, and will proceed to the next step of cross-source cross-validation.

[0097] Step S138: Perform cross-source verification on candidate trusted nodes. If a candidate trusted node is in the knowledge maturity stage or knowledge iteration stage, retrieve at least one related knowledge node published by another source institution and compare the core knowledge descriptions and supporting information of the two. If the descriptions are consistent and the supporting information types are complementary, then improve the credibility level of the related knowledge node.

[0098] For knowledge-mature and knowledge-iterative nodes in the candidate trusted node set, cross-source cross-validation is required to further confirm their trustworthiness.

[0099] Step S1381: Classify the knowledge maturity nodes and knowledge iteration nodes among the candidate trusted nodes to form a cross-source verification node set.

[0100] From the candidate trusted nodes, nodes in the knowledge maturity stage and nodes in the knowledge iteration stage are separated according to their knowledge state labels to form a cross-source verification node set. Nodes in the knowledge nascent stage, due to the cutting-edge and exploratory nature of their knowledge content, usually lack sufficient multi-source related knowledge, and therefore will not be subject to cross-source verification at this time.

[0101] Step S1382: For each node in the cross-source verification node set, use the graph retrieval interface of the knowledge evolution lifecycle network to retrieve related knowledge nodes published by other source institutions based on the semantic similarity of the core knowledge representation. The retrieval results shall contain at least one knowledge node that is different from the source institution of the current node.

[0102] Using the graph retrieval interface of the knowledge evolution lifecycle network, for each node in the cross-source verification node set, its core knowledge representation is used as the retrieval condition. A semantic similarity threshold is set (which can be the same as or slightly lower than the threshold used for knowledge state determination, such as 0.25). All knowledge nodes in the network whose source institution identifiers are different from the current node are retrieved. The retrieval results are required to return at least one of the aforementioned knowledge nodes, i.e., related knowledge nodes. For example, if the current node is a knowledge maturity node from institution A, and its core knowledge representation is "the standard dose of a certain chemotherapy regimen," then by searching, knowledge nodes from institution B with semantically similar core knowledge representations can be found.

[0103] Step S1383: Retrieve the knowledge source organization identifier, core knowledge description and supporting information original text for each related knowledge node, and generate verification materials for the related nodes.

[0104] For each retrieved related knowledge node, obtain its knowledge source institution identifier from the knowledge evolution lifecycle network, extract the core knowledge expression from the knowledge unit, retrieve the original text of supporting information from the supporting information database (similar to step S134), and organize the above information into related node verification materials for comparison with the current node.

[0105] Step S1384: Compare the core knowledge description of the current node with the core knowledge description of the related knowledge nodes sentence by sentence, and extract the identical and different descriptions. If the difference is only in the way it is expressed but the core meaning is the same, then the descriptions are considered to be consistent.

[0106] A detailed text comparison is performed on the core knowledge statements of the current node and related knowledge nodes. Sentence similarity calculation methods from natural language processing (such as cosine similarity of sentence embedding vectors based on the BERT model) are used to calculate similarity sentence by sentence. Identical statements are directly marked; for differing statements, the reasons for the differences are analyzed. If the difference is only in the way it is expressed (e.g., "remission rate" and "complete remission rate" have the same meaning in a specific context, or synonyms or different sentence structures are used but express the same meaning), and the core meaning (e.g., treatment plan, efficacy index values, conclusions, etc.) is consistent, then the two statements are considered consistent. For example, if the current node states "the treatment cycle is 28 days" and the related node states "each treatment course lasts 28 days," then the statements are considered consistent.

[0107] Step S1385: Analyze the supporting information type of the current node and the supporting information type of the associated knowledge node. The supporting information type includes clinical case data, experimental research data, and literature citation data. If the supporting information type of the current node is different from the supporting information type of the associated knowledge node, it is determined that the supporting information types are complementary.

[0108] Supporting information is categorized into clinical case data (derived from clinical case diagnosis and treatment records), experimental research data (derived from the experimental data collection process), and literature citation data (derived from the core conclusions of the literature). Analyzing the supporting information types of the current node and related knowledge nodes, if the types differ, their supporting information is considered complementary, supporting the core knowledge statement from different perspectives. For example, if the supporting information type of the current node is experimental research data (verifying the effectiveness of the treatment plan through animal experiments), and the supporting information type of the related node is clinical case data (observing the efficacy of the treatment plan through clinical cases), then the supporting information types are considered complementary.

[0109] Step S1386: If the core knowledge descriptions of the current node and the related knowledge nodes are consistent and the supporting information types are complementary, then retrieve the initial credibility assessment reports of both nodes, raise the credibility level of the current node by one level, and record the level increase in the supplementary explanation column of the assessment report.

[0110] If the core knowledge statements of the current node are consistent with those of related knowledge nodes, and the supporting information types are complementary, it indicates that the knowledge content of the current node is supported by evidence from different sources and of different types, further enhancing its credibility. At this point, retrieve the initial credibility assessment reports for the current node and related knowledge nodes, and upgrade the credibility level of the current node by one level based on the initial assessment. Assuming the initial credibility assessment levels are divided into generally credible, reasonably credible, and highly credible, if the current node's initial assessment is reasonably credible, then it is upgraded to highly credible. Record the reason for the upgrade (e.g., "consistent with the statements of related knowledge nodes of institution B and complementary supporting information types") in the supplementary explanation section of the current node's initial credibility assessment report.

[0111] Step S1387: If there are multiple related knowledge nodes in the search results, compare their descriptions with the current node and analyze the types of supporting information. If all related knowledge nodes meet the requirements of consistent descriptions and complementary types of supporting information, then increase the credibility level of the current node by two levels.

[0112] If multiple (e.g., two or more) related knowledge nodes are retrieved through the graph retrieval interface, and each related knowledge node is consistent with the core knowledge description of the current node, supporting complementary information types, then the credibility of the current node receives stronger multi-source verification, raising its credibility level by two levels. For example, if the initial assessment is generally credible, it can be upgraded to highly credible; if the initial assessment is relatively credible, an even higher level can be set after the upgrade, such as "extremely highly credible." Similarly, the information of all related knowledge nodes and the reasons for the upgrade are recorded in the supplementary explanation section of the assessment report.

[0113] Step S1388: If there is a contradiction between the core knowledge expression of the related knowledge node and the current node that is not a difference in expression method, then record the contradiction point and the related knowledge node information as a reference for subsequent conflict reconciliation.

[0114] During the comparison of statements, if a contradiction is found between the core knowledge statement of the related knowledge node and the current node, and the contradiction is not caused by the difference in the expression method, but by the conflict in the core meaning (e.g., the current node states that the relief rate of treatment plan A is 80%, while the related node states it is 60%, and there is no significant difference in statistical methods, sample size, etc. leading to reasonable error), then the contradiction points (e.g., "difference in relief rate values: 80% vs 60%)" as well as the identifier, source institution, supporting information, etc. of the related knowledge node are recorded in detail as a reference for subsequent dynamic conflict reconciliation.

[0115] Step S1389: For each candidate trusted node that has completed cross-source cross-validation, generate a cross-source verification report. The cross-source verification report includes information on associated knowledge nodes, description comparison results, analysis results of supporting information types, and trust level adjustments. The cross-source verification report is merged with the initial trust assessment report to form trust assessment materials.

[0116] The cross-source verification report summarizes the process and results of cross-source verification, including the identifier of the knowledge source institution of the related knowledge nodes, a summary of the core knowledge description, and the type of supporting information; the description comparison results (consistency or contradictions); the analysis results of the supporting information types (complementary or identical); and the current node's credibility level adjustment (level upgrade, reasons for upgrade). This cross-source verification report is combined with the initial credibility assessment report to form a complete credibility assessment material.

[0117] Step S139: Identify the node combinations with knowledge update-related edges among the candidate trusted nodes. If the core knowledge statements of two nodes in the combination are contradictory, extract the knowledge release time information of both and retain the nodes whose release time is later than the release time of all related historical knowledge units.

[0118] In the knowledge evolution lifecycle network, knowledge update-related edges connect historical knowledge units and iterative knowledge units. For candidate trusted nodes with the aforementioned related edges (i.e., historical knowledge units and their corresponding iterative knowledge units), check whether their core knowledge statements are contradictory (same as the contradiction definition in step S1388). If a contradiction exists, it is processed according to the knowledge publication time information. Extract the knowledge publication time of the iterative knowledge unit and the historical publication time of the historical knowledge unit. Since the definition of the iterative knowledge unit requires that its publication time is later than the publication time of all associated historical knowledge units (step S1246), the iterative knowledge unit is retained, and the historical knowledge unit is discarded or marked as pending update. For example, a historical knowledge unit published in 2020 states "the recommended dose of a certain drug is X mg / m²", while an iterative knowledge unit published in 2023 states "the recommended dose of the drug is revised to Y mg / m²". There is a contradiction between the two, so the iterative knowledge unit is retained.

[0119] Step S1310: For the retained nodes, extract the update basis description from the original supporting information and associate and bind the update basis description with the core knowledge expression.

[0120] For the nodes retained in step S139 (usually iterative knowledge units), the update basis description is extracted from their supporting information. The update basis description is key evidence supporting the updating of knowledge content, such as a clinical study finding that the original dosage group had an excessively high incidence of adverse reactions, or new pharmacokinetic studies showing that different dosages are more effective, or adjustments to the treatment plan based on new genotyping results. The above update basis description is associated and bound to the core knowledge representation of the node using a unique identifier and stored in the attributes of the knowledge unit so that the basis for knowledge updates is clearly defined during subsequent model updates.

[0121] Step S1311: Summarize the core knowledge descriptions and associated supporting information and update basis descriptions corresponding to all knowledge content nodes that have passed cross-source cross-validation and completed conflict reconciliation, and generate a set of target knowledge that has passed verification and is conflict-free.

[0122] Collect all knowledge content nodes that have undergone cross-source verification (for knowledge maturity and iterative nodes) and meet the required credibility level (e.g., reaching high or very high credibility), and have completed conflict reconciliation (retaining the latest iterative nodes and handling contradictory nodes). Combine the core knowledge descriptions of these nodes, the associated supporting original information (clinical case records, experimental data collection processes, core literature conclusions, etc.), and the description of the update basis (for iterative nodes) to form a target knowledge set. All knowledge content in this set has passed multi-source credibility verification and is internally consistent, making it suitable for knowledge updates in the large-scale pediatric oncology teaching model.

[0123] Step S140: Invoke the pre-built parameter fine-tuning and memory retention collaborative mechanism to convert the target knowledge set into a knowledge update parameter package adapted to the pediatric oncology teaching model. The knowledge update parameter package includes module parameter adjustment schemes, knowledge association constraint rules, and model memory anchoring parameters.

[0124] The target knowledge set needs to be converted into a parameter form that can be understood and applied by the large-scale pediatric oncology teaching model, namely, a knowledge update parameter package.

[0125] Step S141: For each knowledge content in the target knowledge set, classify it into topics according to the knowledge storage module division criteria of the pediatric oncology teaching model. The knowledge storage module division criteria are determined based on the knowledge system in the field of pediatric oncology.

[0126] The knowledge storage modules of the large-scale pediatric oncology teaching model are divided according to the knowledge system of the field of pediatric oncology. For example, they are divided into modules such as basic theory (including etiology, pathogenesis, and pathophysiology of pediatric tumors), diagnosis (including clinical manifestations, laboratory tests, imaging examinations, and differential diagnosis), treatment (including knowledge of different treatment methods such as chemotherapy, radiotherapy, surgery, targeted therapy, and immunotherapy), prognosis and follow-up (including efficacy evaluation criteria, prognostic factor analysis, and follow-up protocols), and case analysis (including discussions of typical and difficult cases). Each piece of knowledge in the target knowledge set is assigned to a corresponding knowledge storage module based on its core theme. For example, knowledge about chromosomal karyotype analysis in childhood acute lymphoblastic leukemia is assigned to the diagnosis module; knowledge about the use of novel targeted drugs is assigned to the treatment module; and knowledge about factors affecting five-year survival is assigned to the prognosis and follow-up module.

[0127] Step S142: Summarize the knowledge content under the same topic category to generate a target knowledge subset corresponding to each knowledge storage module. Each target knowledge subset contains the knowledge content that needs to be added or updated in that knowledge storage module.

[0128] Knowledge content belonging to the same knowledge storage module is aggregated to form the target knowledge subset of that module. The knowledge content in the target knowledge subset may be new knowledge that has not been stored in the model before (that needs to be added), or it may be an update of the old knowledge already stored in the model (that needs to be updated). For example, the target knowledge subset of the treatment module may contain newly added knowledge about a certain targeted drug, as well as updated knowledge about adjusting the dosage of the original chemotherapy regimen.

[0129] Step S143: Retrieve the current parameter configuration file for each knowledge storage module. The parameter configuration file contains the weight parameter matrix, knowledge association parameter list, and parameter adjustment history of the module.

[0130] Each knowledge storage module in the pediatric oncology teaching model has a corresponding parameter configuration file, stored in the model's parameter server. The current parameter configuration file for each knowledge storage module is retrieved by accessing the parameter server's interface. The weight parameter matrix represents the connection weights between neural network layers within the module, determining the strength and correlation of knowledge representation. The knowledge association parameter list records the correlation strength parameters between different knowledge units within the module, such as the correlation coefficient between two medical concepts. The parameter adjustment history records the parameter adjustments made during each past knowledge update for this module, including the adjustment time, the adjusted parameter location, and the values ​​before and after the adjustment, used for tracing and analyzing parameter changes.

[0131] Step S144: The feature parsing module of the input parameter fine-tuning and memory retention synergy mechanism for each target knowledge subset extracts medical concept features and logical relationship features from the knowledge content, and generates a feature vector matrix. The dimension of the feature vector matrix is ​​consistent with the dimension of the weight parameter matrix of the corresponding knowledge storage module.

[0132] The parameter fine-tuning and memory retention collaborative mechanism includes a feature parsing module, which uses a pre-trained medical BERT model (pre-trained on a large-scale medical text corpus, such as PubMed summarization) to extract features from the knowledge content of the target knowledge subset. First, the knowledge content is segmented and tokenized, then input into the medical BERT model, and the model's last hidden layer outputs the embedding vector for each token. Then, for medical concept features, the embedding vectors of tokens belonging to the same medical concept are averaged to obtain the feature vector of the medical concept. For logical relationship features (such as the "used for treatment" relationship in "treatment plan A is used to treat disease B"), a relationship extraction model (such as a BERT-based relationship classification model) is used to identify entity pairs and relationship types in the knowledge content, converting the relationship types into one-hot vectors or embedding vectors. Finally, all medical concept feature vectors and logical relationship feature vectors are arranged in a specific order to form a feature vector matrix. The number of rows and columns (dimensions) of this matrix are exactly the same as the dimensions of the weight parameter matrix of the corresponding knowledge storage module, facilitating subsequent parameter comparison and adjustment.

[0133] Step S145: Compare the feature vector matrix and the weight parameter matrix, calculate the difference between them in each parameter dimension, determine the position and direction of the parameters to be adjusted based on the difference, and form a preliminary parameter adjustment plan.

[0134] The feature vector matrix (the feature representation of new knowledge) is compared element-by-element with the current weight parameter matrix (the representation of old knowledge) of the knowledge storage module. For each element (parameter dimension) in the matrix, the difference between the element value of the feature vector matrix and the element value of the weight parameter matrix is ​​calculated, i.e., the difference value. The larger the absolute value of the difference value, the greater the difference between the new and old knowledge in that parameter dimension, and the more adjustment is needed. A difference value threshold is set (determined based on model training experience). When the absolute value of the difference value is greater than this threshold, the position of that parameter is considered to need adjustment. The adjustment direction is determined by the sign of the difference value. If the difference value is positive, the element value of the weight parameter matrix needs to be increased; if it is negative, it needs to be decreased. All parameter positions that need adjustment, their corresponding difference values, and adjustment directions are recorded to form a preliminary parameter adjustment plan.

[0135] Step S146: Based on the preliminary parameter adjustment scheme, introduce model memory retention constraints, retrieve historical key parameters from the knowledge storage module. The historical key parameters are parameters that have a significant impact on the core reasoning ability of the model. During the adjustment process, reduce the adjustment range of the adjustment items involving the position of historical key parameters to a range close to zero, so that the values ​​of historical key parameters remain stable. Only non-key parameters are adjusted according to the preliminary scheme to form the final module parameter adjustment scheme.

[0136] To prevent the model from forgetting its original core reasoning abilities due to new knowledge updates, it is necessary to introduce model memory retention constraints.

[0137] Step S1461: Construct a historical key parameter identification model. The historical key parameter identification model is generated based on the historical inference performance data of the pediatric oncology teaching model. It is used to determine whether a parameter is a key parameter based on the degree of influence of the parameter on the inference result.

[0138] The historical key parameter identification model is a supervised learning model. Its training data includes multiple historical parameter configuration versions from the knowledge storage module and corresponding model inference performance metrics (such as accuracy, recall, F1 score, and accuracy in analyzing complex cases). The model's input is the numerical changes of each parameter in the parameter configuration file (such as the difference before and after parameter adjustment, the percentage change, etc.), and the output is a score of the parameter's impact on inference performance. Through training, the model learns which parameter changes have a significant impact on inference performance. Parameters with an impact score higher than a preset threshold are identified as historical key parameters. This model can employ a simple linear regression model (using parameter changes as features and inference performance changes as labels) or a more complex tree model (such as random forest).

[0139] Step S1462: Input the weight parameter matrix in the current parameter configuration file of the knowledge storage module into the historical key parameter identification model. The model outputs the keyness score of each parameter. The higher the keyness score, the greater the impact of the parameter on the core reasoning ability.

[0140] The current weight parameter matrix of the knowledge storage module (i.e., the matrix in the parameter configuration file retrieved in step S143) is input into the historical key parameter identification model. The model analyzes each parameter in the matrix, combines it with historical inference performance data, and calculates and outputs a keyness score for each parameter. For example, parameters related to the core knowledge of "diagnostic criteria for childhood acute lymphoblastic leukemia" have a higher keyness score; while parameters related to some marginal or rare knowledge have a lower score.

[0141] Step S1463: Mark parameters with criticality scores higher than the critical parameter threshold as historical critical parameters and generate a historical critical parameter list, which includes the parameter position, current value and criticality score.

[0142] Set a key parameter threshold (determined based on the model's requirements for core inference capabilities and the output distribution of the historical key parameter identification model). Filter out parameters with a keyness score higher than this threshold and mark them as historical key parameters. Generate a list of historical key parameters, recording the specific position (row and column indices), current value, and keyness score of each historical key parameter in the weight parameter matrix, so as to identify and protect them during subsequent parameter adjustments.

[0143] Step S1464: In the preliminary parameter adjustment scheme, select adjustment items involving the position of historical key parameters, perform numerical correction on the adjustment items, reduce the adjustment range to near zero, and keep the values ​​of historical key parameters stable.

[0144] Review all adjustment items in the initial parameter adjustment plan and check if the parameter position of each adjustment item is in the historical key parameter list. If so, revise the adjustment magnitude of that item. The adjustment magnitude was originally determined based on the difference value (e.g., if the difference value is d, then the adjustment magnitude is k*d, where k is the adjustment coefficient). Now, reduce the adjustment coefficient k to a very small value (e.g., 0.001), making the adjustment magnitude k*d close to zero. For example, if the original adjustment magnitude was an increase of 0.5, the revised adjustment magnitude might be an increase of 0.0005, which hardly changes the current value of the historical key parameter, thus maintaining its stability.

[0145] Step S1465: For adjustment items in the preliminary parameter adjustment plan that do not involve the location of historical key parameters, retain the original adjustment direction and adjustment range to form a revised set of parameter adjustment items.

[0146] For adjustment items in the initial parameter adjustment plan whose parameter positions are not in the historical key parameter list, it is considered that their impact on the core reasoning ability of the model is small and no correction is required. The original adjustment direction and adjustment magnitude (the magnitude calculated based on the difference value) are retained. All adjusted items (involving historical key parameters) and uncorrected items (not involving historical key parameters) are combined to form the corrected parameter adjustment item set.

[0147] Step S1466: Sort the corrected set of parameter adjustment items according to the parameter position to generate a parameter adjustment order table. The adjustment order is determined based on the functional priority of the parameters in the knowledge storage module, and parameters with higher functional priority are adjusted first.

[0148] Parameters in the knowledge storage module are prioritized according to their function. For example, parameters related to the module's core functions (such as the main decision parameters for treatment plan recommendations) have high priority, while parameters related to supplementary explanations and background knowledge have low priority. The adjustment items in the revised parameter adjustment item set are sorted according to their functional priority, with adjustment items corresponding to parameters with higher functional priority listed first, generating a parameter adjustment order table. During actual parameter adjustments, this order is followed to ensure that important parameters are adjusted first, guaranteeing the correctness of the module's function.

[0149] Step S1467: Add an adjustment basis description for each parameter adjustment item. The adjustment basis description includes the feature vector difference value of the corresponding knowledge content, the parameter criticality score, and the calculation process of the adjustment range.

[0150] Each parameter adjustment item needs a clear basis for adjustment. The explanation of the basis should include: the difference value of the eigenvectors of the knowledge content corresponding to the parameter adjustment in the eigenvector matrix (i.e., the difference value calculated in step S145); the criticality score of the parameter (output in step S1462; a higher score if it is a historical critical parameter); and the calculation process of the adjustment magnitude (for non-historical critical parameters, adjustment magnitude = adjustment coefficient * difference value; for historical critical parameters, adjustment magnitude = corrected adjustment coefficient * difference value). Adding the above information to each parameter adjustment item facilitates subsequent tracking and verification of the rationality of the parameter adjustments.

[0151] Step S1468: Integrate the parameter adjustment sequence table, all parameter adjustment items and corresponding adjustment basis descriptions to form the final module parameter adjustment scheme.

[0152] The parameter adjustment sequence table, the revised set of parameter adjustment items (including parameter position, adjustment direction, and adjustment range), and the explanation of the adjustment basis for each adjustment item are integrated to form the final module parameter adjustment scheme for each knowledge storage module. This scheme clarifies which parameters in the module need to be adjusted, how to adjust them (direction and range), the order of adjustment, and the basis for adjustment, serving as a specific operational guide for updating parameters in the knowledge storage module.

[0153] Step S147: Analyze the logical relationships between knowledge content within each target knowledge subset, and construct knowledge association constraint rules based on the logical relationships. The knowledge association constraint rules clarify the calling order and association priority of knowledge content in the reasoning process. The logical relationships include causal relationships, step sequence relationships, and subordinate relationships.

[0154] The knowledge content within the target knowledge subset does not exist in isolation; various logical relationships exist between them. Natural language processing techniques (such as dependency parsing, semantic role labeling, and logical relation extraction models) are used to analyze the logical relationships between the knowledge content within each target knowledge subset. Causal relationships refer to one knowledge content being the cause or result of another, such as "gene mutation leads to childhood acute lymphoblastic leukemia" (gene mutation is the cause, and leukemia is the effect); sequential relationships refer to the chronological or procedural order of the operations or events described by the knowledge content, such as "chemotherapy is divided into induction remission, consolidation and intensification, and maintenance therapy" (these three stages have a sequential order); and subordinate relationships refer to one knowledge content belonging to another broader knowledge content, such as "childhood acute lymphoblastic leukemia is a type of childhood leukemia" (childhood acute lymphoblastic leukemia belongs to childhood leukemia). Based on these logical relationships, knowledge association constraint rules are constructed. The rules clearly define the order of knowledge content retrieval (e.g., in a step sequence relationship, knowledge should be retrieval in order) and the priority of association (e.g., in a causal relationship, the priority of causal knowledge may be higher than that of result knowledge, and in a subordinate relationship, the priority of superior knowledge may be higher than that of subordinate knowledge).

[0155] Step S148: Construct model memory anchoring parameters. The model memory anchoring parameters are used to mark the parameter positions corresponding to the core knowledge that needs to be retained in the knowledge storage module for a long time. In subsequent updates, the adjustment range of the anchoring parameter positions is controlled within a preset range.

[0156] Model memory anchoring parameters are another mechanism to protect core knowledge from being forgotten. This is achieved by marking the position of the parameters corresponding to the core knowledge and limiting their adjustment range.

[0157] For example, step S1481: Extract core knowledge content from the target knowledge subset corresponding to the knowledge storage module. The core knowledge content includes basic theoretical knowledge, standard diagnosis and treatment guidelines, and high-frequency case analysis conclusions in the field of pediatric oncology.

[0158] The core knowledge content comprises the key knowledge that the large-scale pediatric oncology teaching model must master consistently over the long term. This includes fundamental theoretical knowledge (such as the basic classification and common causes of pediatric tumors), standard diagnostic and treatment guidelines (such as standardized diagnostic procedures and first-line treatment plans for various types of pediatric tumors), and conclusions from high-frequency case analyses (such as the typical clinical manifestations and treatment pathways for childhood acute lymphoblastic leukemia). From the target knowledge subset of the knowledge storage module, these core knowledge contents were jointly selected by domain experts and model developers based on their importance, frequency of use, and pedagogical necessity.

[0159] Step S1482: Input the core knowledge content into the knowledge storage module, simulate the knowledge storage process, and record the parameter activation area corresponding to the core knowledge content in the module. The parameter activation area is the set of parameter positions where the value changes significantly when storing the core knowledge.

[0160] Through simulation experiments, core knowledge content was input into the knowledge storage module, and the changes in the module's weight parameter matrix were observed. During the knowledge storage process, the values ​​of parameters related to the core knowledge content changed significantly (compared to other parameter positions). These parameter positions with significant changes were recorded, forming parameter activation regions. For example, when the core knowledge content "MICM classification criteria for childhood acute lymphoblastic leukemia" was input, the values ​​of parameters related to concepts such as "MICM classification," "morphology," "immunology," "cytogenetics," and "molecular biology" in the weight parameter matrix changed significantly; these positions constitute the parameter activation regions.

[0161] Step S1483: For each parameter position in the parameter activation region, calculate the frequency and magnitude of parameter changes in that position during multiple historical knowledge updates.

[0162] Retrieve the parameter adjustment history of the knowledge storage module (included in the parameter configuration file in step S143). For each parameter position in the parameter activation area, count the number of times the parameter position was adjusted (frequency of change) and the magnitude of each adjustment (average and maximum magnitude of change) during the past N (e.g., 5) knowledge updates. Parameter positions with low frequency of change and small magnitude of change indicate that the corresponding knowledge content is relatively stable and is more likely to be core knowledge.

[0163] Step S1484: Determine the intersection of the parameter activation region and the long-term stable parameter location as the key parameter location corresponding to the core knowledge, and generate a list of key parameter locations.

[0164] Long-term stable parameter locations refer to the set of parameter locations whose change frequency is lower than a frequency threshold and whose change amplitude is lower than an amplitude threshold during multiple historical knowledge updates. Taking the intersection of the parameter activation region obtained in step S1482 and the long-term stable parameter locations yields the key parameter locations corresponding to the core knowledge. The parameters at these locations are related to core knowledge storage and possess long-term stability, requiring focused protection. A list of key parameter locations is generated, recording the index information of these parameter locations.

[0165] Step S1485: Assign anchor tags to each key parameter location. Anchor tags include knowledge type codes and anchor priorities. Knowledge type codes correspond to the topic categories of core knowledge content, and anchor priorities are determined based on the importance of core knowledge content.

[0166] Knowledge type coding is a pre-defined coding system used to represent the topic category of core knowledge content. For example, "basic theory" is coded as "BT", "standard treatment guidelines" as "ST", and "high-frequency case analysis" as "CA". Anchoring priority is divided into three levels based on the importance of the core knowledge content, with level one being the most important and level three being relatively important. For example, "diagnostic criteria for childhood acute lymphoblastic leukemia" belongs to standard treatment guidelines and is of extremely high importance, assigned the anchoring identifier "ST - Level One"; "epidemiological characteristics of childhood tumors" belongs to basic theory and is of relatively high importance, assigned the anchoring identifier "BT - Level Two". A corresponding anchoring identifier is assigned to each parameter position in the list of key parameter positions.

[0167] Step S1486: Set the upper limit of the adjustment range of the anchoring parameters. Different anchoring priorities correspond to different upper limits of the adjustment range. The higher the anchoring priority, the lower the upper limit of the adjustment range.

[0168] Different adjustment range limits are set based on anchoring priority. The lowest adjustment range limit is for parameter positions in Level 1 anchoring priority (e.g., 0.0001), followed by Level 2 (e.g., 0.001), and then Level 3 (e.g., 0.01). The adjustment range limit refers to the absolute value of the adjustment range of a parameter position that cannot exceed this limit during any knowledge update. The higher the priority of core knowledge, the stricter the adjustment restrictions on its corresponding parameter positions, in order to maximize its stability.

[0169] Step S1487: Integrate the list of key parameter locations, the anchoring identifier for each location, and the corresponding upper limit of adjustment range to form a model memory anchoring parameter table.

[0170] The indexes of key parameter locations, assigned anchoring identifiers (knowledge type encoding and anchoring priority), and corresponding adjustment limits are organized into a table, known as the model memory anchoring parameter table. This table clearly shows which parameter locations need to be anchored for protection, as well as the level of protection and adjustment restrictions.

[0171] Step S1488: Add parameter monitoring rules to the model memory anchoring parameter table. The monitoring rules require that, in the subsequent knowledge update process, before each parameter adjustment, the adjustment position is checked to see if it is the anchoring parameter position. If it is the anchoring parameter position, the adjustment range shall not exceed the corresponding upper limit.

[0172] Add parameter monitoring rules to the model memory anchoring parameter table as constraints that must be followed during subsequent knowledge updates. The rules stipulate that before performing any parameter adjustment operation, the model memory anchoring parameter table must be queried to check whether the position of the parameter to be adjusted is in the list of key parameter positions (i.e., whether it is an anchoring parameter position). If so, the absolute value of the adjustment must be less than or equal to the upper limit of the adjustment range corresponding to that position; if the adjustment exceeds the upper limit, it will be automatically truncated to the upper limit value (keeping the direction unchanged).

[0173] Step S1489: Associate the model memory anchoring parameter table with the parameter configuration file of the knowledge storage module to generate an anchoring parameter mapping relationship.

[0174] By using parameter position indexing, the positions of key parameters in the model memory anchor parameter table are mapped one-to-one with the positions of weight parameter matrices in the knowledge storage module parameter configuration file, generating an anchor parameter mapping relationship. This mapping relationship allows for quick querying of whether the current parameter position is an anchor parameter position and its corresponding adjustment range limit during parameter adjustment, facilitating the monitoring of rule execution.

[0175] Step S149: Summarize the module parameter adjustment schemes corresponding to all knowledge storage modules, the knowledge association constraint rules corresponding to all knowledge topics, and the model memory anchoring parameters of the entire model. Add parameter package version identifiers and generation time information to generate a knowledge update parameter package adapted to the large-scale pediatric oncology teaching model.

[0176] Collect the module parameter adjustment schemes for each knowledge storage module, the knowledge association constraint rules for all knowledge topics (corresponding to different knowledge storage modules or across modules), and the model memory anchoring parameters for the entire model (model memory anchoring parameter table for each knowledge storage module). Add a unified parameter package version identifier (e.g., V202312) and generation time information (year, month, day, hour, minute, second) to these contents for version management and traceability. Package all this information to generate a knowledge update parameter package.

[0177] Step S150: Input the knowledge update parameter package into the pediatric oncology teaching model, perform a collaborative update operation on the knowledge storage module, reasoning calculation module and memory retention module of the pediatric oncology teaching model, and generate a knowledge update full-process traceability document. The knowledge update full-process traceability document includes the updated content, the updated module, the corresponding knowledge node information and the memory anchoring record.

[0178] After the knowledge update parameter package is generated, it is input into the large-scale pediatric oncology teaching model to initiate the collaborative update operation of the model, while recording the update process and generating a traceability document.

[0179] For example, step S151: parse the module parameter adjustment scheme in the knowledge update parameter package, and obtain the exclusive parameter adjustment scheme for each module according to the knowledge storage module. The exclusive parameter adjustment scheme includes the parameter adjustment position, the value before adjustment, the value after adjustment, and the adjustment basis.

[0180] The module parameter adjustment schemes in the knowledge update parameter package are organized by knowledge storage module. The parameter package is parsed to extract a unique parameter adjustment scheme for each knowledge storage module. This scheme details the parameter positions (row and column indices) within the module that need adjustment, the current value of each parameter position before adjustment, the adjusted value calculated based on the adjustment range (previous value + adjustment range), and the adjustment basis (from the adjustment basis description in step S1467). For example, the unique parameter adjustment scheme for the treatment module includes: "Parameter position (102, 235): previous value 0.45, adjusted value 0.52, adjustment basis: difference value 0.07, adjustment coefficient 1.0, non-historical key parameter."

[0181] Step S152: Input the exclusive parameter adjustment scheme for each knowledge storage module into the corresponding knowledge storage module, start the module parameter update process, and modify the weight parameter matrix inside the module in the order of the parameter positions in the adjustment scheme. After each parameter position adjustment is completed, record the adjustment time and the adjusted value.

[0182] Each knowledge storage module's proprietary parameter adjustment scheme is input into the corresponding knowledge storage module through the model's internal interface. The module parameter update process modifies the parameter positions in the weight parameter matrix sequentially according to the parameter adjustment sequence table in the scheme (step S1466), updating the parameter values ​​to the adjusted values. During the update process, after each parameter position adjustment is completed, the adjustment time (accurate to milliseconds) and the adjusted value are immediately recorded, forming a parameter update log for subsequent parameter change report generation and problem troubleshooting.

[0183] Step S153: After all knowledge storage module parameters are updated, retrieve the knowledge association constraint rules from the knowledge update parameter package and load the rules into the rule parsing unit of the inference calculation module.

[0184] After the knowledge storage module parameters are updated, the inference computation module is updated to enable it to correctly utilize the new knowledge for reasoning. Knowledge association constraint rules are extracted from the knowledge update parameter package; these rules describe the logical relationships and calling order between knowledge content. These rules are then loaded into the rule parsing unit of the inference computation module. The rule parsing unit is responsible for converting the rules, whether described in natural language or structured format, into an internal representation executable by the inference computation module (such as logical expressions or rule formats recognizable by the rule engine).

[0185] Step S154: Update the knowledge retrieval logic of the reasoning and calculation module so that when the reasoning and calculation module processes teaching knowledge query requests, it determines the order of knowledge content retrieval based on knowledge association constraint rules.

[0186] The knowledge retrieval logic of the reasoning and computation module determines how to organize and use the knowledge in the knowledge storage module to answer query requests, and needs to be updated according to the new knowledge association constraint rules.

[0187] Step S1541: Parse the logical relationship description and association priority setting in the knowledge association constraint rules, and construct a knowledge call priority list. The knowledge call priority list includes knowledge content identifier, association priority level and corresponding logical relationship type.

[0188] The knowledge association constraint rules include descriptions of logical relationships (e.g., "Treatment plan A should be invoked after diagnosis B," indicating a step-by-step sequence) and association priority settings (e.g., "Core treatment principles have higher priority than auxiliary treatment recommendations"). These rules are parsed to extract the association priority level (e.g., high, medium, low) and corresponding logical relationship type (causal, step-by-step sequence, subordinate) for each knowledge content (represented by a knowledge content identifier). This information is then sorted from highest to lowest association priority level to construct a knowledge invocation priority list. For example, knowledge content identifier K1001 (diagnostic criteria) has a high priority and its logical relationship type is a preceding step in the step-by-step sequence; in the priority of K1002 (treatment plan), its logical relationship type is a following step in the step-by-step sequence.

[0189] Step S1542: Add a rule matching unit to the reasoning calculation module. The rule matching unit is used to receive teaching knowledge query requests, extract the core query terms in the teaching knowledge query requests, and match the core query terms with the knowledge content identifiers in the knowledge call priority list.

[0190] The rule matching unit is a newly added functional unit integrated into the front-end processing flow of the inference and computation module. When a teaching knowledge query request (such as "How to diagnose acute lymphoblastic leukemia in children?") is input into the inference and computation module, the rule matching unit first performs natural language understanding on the request to extract core query terms (such as "diagnosis" and "acute lymphoblastic leukemia in children"). Then, it performs semantic matching between the core query terms and the knowledge content identifiers in the knowledge call priority list to find the set of knowledge content identifiers related to the query.

[0191] Step S1543: If multiple knowledge content identifiers are matched, sort them from high to low according to their association priority level and generate a knowledge call order suggestion.

[0192] If the rule matching unit matches multiple knowledge content identifiers (such as simultaneously matching multiple related knowledge contents such as diagnostic criteria, differential diagnosis, and laboratory tests), it sorts them according to the association priority level of each knowledge content identifier in the knowledge call priority list, with knowledge content identifiers with higher priority levels listed first, generating a preliminary knowledge call order suggestion. For example, if K1001 (diagnostic criteria, high priority), K1003 (laboratory tests, medium priority), and K1004 (differential diagnosis, medium priority) are matched, the order would be K1001, K1003, and K1004.

[0193] Step S1544: Add a logic verification unit to the reasoning calculation module. The logic verification unit can check whether the suggested knowledge calling order meets the logical relationship requirements in the knowledge association constraint rules. If the calling order violates the causal relationship or the step order relationship, the order is readjusted.

[0194] The logic verification unit performs a secondary check on the suggested knowledge retrieval order to ensure it conforms to the logical relationships in the knowledge association constraint rules. For example, if the rule stipulates that "laboratory tests (K1003) must be performed before a diagnostic criterion (K1001) can be made" (step order relationship), and the initial suggested knowledge retrieval order is K1001, K1003, then the logic verification unit identifies a violation of the step order relationship and automatically adjusts the order to K1003, K1001. For causal relationships, such as "cause A leads to disease B," then when retrieving knowledge about disease B, it may be necessary to first retrieve knowledge about cause A as background. The logic verification unit corrects the order by checking the logical relationship description in the rules.

[0195] Step S1545: Add a call log recording unit to the inference calculation module. The log recording unit records the knowledge content identifier, call time, associated priority level and logical verification result of each knowledge call, and generates a knowledge call log.

[0196] The call logging unit is activated when the inference calculation module calls knowledge. It records detailed information for each call: the identifier of the called knowledge content, the timestamp of the call, the associated priority level of the knowledge content, and the verification result (pass / adjust / fail) of the logic verification unit. The above information is stored in chronological order to form a knowledge call log, which is used to analyze the validity of the knowledge call logic and to troubleshoot problems.

[0197] Step S1546: If the knowledge call log shows that the call frequency of any knowledge content with an associated priority level is lower than the call logic monitoring threshold, then analyze the association constraint rules of that knowledge content and correct any rules that are unreasonable.

[0198] Set monitoring thresholds for knowledge call logic. For example, the average call frequency of knowledge content with a certain priority level should not be less than X times / day over a period of time (e.g., one week). Regularly check the knowledge call logs and statistically analyze the call frequency of knowledge content with different priority levels. If the call frequency of a certain level of knowledge content is found to be lower than the monitoring threshold, it may be due to unreasonable knowledge association constraint rules (e.g., rules that are too strict in restricting its calls, or priority settings that are too low). In this case, organize domain experts and model developers to analyze the association constraint rules of the knowledge content, identify the problem, and make corrections, such as adjusting the priority level and modifying the logical relationship description, to improve the rationality and comprehensiveness of knowledge calls.

[0199] Step S1547: Integrate the rule matching unit, logic verification unit, and call log recording unit into the knowledge call process of the inference calculation module, replace the original call logic, and complete the update of the knowledge call logic of the inference calculation module.

[0200] The newly added rule matching unit, logic verification unit, and call logging unit are integrated into the knowledge call flow of the inference and computation module according to the processing order (rule matching → generation of order suggestions → logic verification → call execution → log recording), replacing the original knowledge call logic in the module (which may not have considered the new knowledge association constraint rules). Through the reconfiguration of the module's internal interfaces and data flow, it is ensured that the new knowledge call logic can correctly receive the new knowledge in the knowledge storage module and perform inference and response according to the knowledge association constraint rules.

[0201] Step S155: Retrieve the model memory anchoring parameters from the knowledge update parameter package, load the parameters into the anchoring management unit of the memory retention module, and the memory retention module monitors all subsequent parameter adjustment operations based on the anchoring parameters to ensure that the adjustment range of the anchoring parameter position does not exceed the preset upper limit.

[0202] The memory retention module is a dedicated module in the large-scale pediatric oncology teaching model responsible for protecting the memory of core knowledge. It includes an anchoring management unit. Model memory anchoring parameters (model memory anchoring parameter table) are extracted from the knowledge update parameter package and loaded into the anchoring management unit of the memory retention module. The anchoring management unit monitors all parameter adjustment operations of the knowledge storage module in real time (not only for the current update, but also for future updates) based on the anchoring parameter mapping relationship in the model memory anchoring parameter table. Before each parameter adjustment, it checks whether the adjustment position is an anchoring parameter position. If so, it calculates the adjustment range to ensure it does not exceed the preset adjustment range upper limit (step S1486). If the adjustment range exceeds the upper limit, it automatically truncates the adjustment range, allowing adjustments only within the specified range.

[0203] Step S156: Create a memory anchoring record in the memory retention module. The memory anchoring record includes the anchoring parameter position, the corresponding core knowledge content identifier, the anchoring priority, and the upper limit of the adjustment range.

[0204] After loading the model's anchoring parameters, the memory retention module creates a memory anchoring record for each anchoring parameter position. The record includes: the anchoring parameter position (row and column indices of the weight parameter matrix), the identifier of the core knowledge content associated with the anchoring parameter mapping relationship (e.g., the identifier of "diagnostic criteria for childhood acute lymphoblastic leukemia"), the anchoring priority of that parameter position (level 1 / 2 / 3), and the upper limit of the adjustment range. These records are stored in the memory retention module's database for easy querying and monitoring by the anchoring management unit.

[0205] Step S157: Extract the parameter configuration files of each knowledge storage module before and after the update, compare and generate a parameter change report. The parameter change report includes the parameter location, the difference in values ​​before and after the adjustment, and the knowledge content update information corresponding to the parameter.

[0206] After the knowledge storage module parameters are updated (step S152), the updated parameter configuration file for each module is extracted and compared with the original parameter configuration file (the file retrieved in step S143). For each module, all changed parameter locations are identified, and the numerical difference before and after adjustment is calculated (adjusted value - original value). Based on the adjustment criteria in the module parameter adjustment plan, the corresponding knowledge content update information (such as newly added knowledge content identifiers and core knowledge description summaries) is associated with that parameter location. The above information is compiled into a parameter change report, clearly showing the parameter update status of each knowledge storage module.

[0207] Step S158: Extract the knowledge call logic description before and after the update of the reasoning calculation module, and generate a logic update report. The logic update report includes the logic modification location, the logic before modification and the logic after modification, and the corresponding knowledge association constraint rules.

[0208] For the reasoning and computation module, extract the logical description before the knowledge call logic update (e.g., the call order based on old knowledge associations) and the logical description after the update (e.g., the call order based on new knowledge association constraint rules, i.e., the logic updated in step S154). Compare the two to determine the specific location of the logical modification (e.g., newly added rule matching units, modified order sorting algorithms, etc.), the logical content before the modification, the logical content after the modification, and the knowledge association constraint rules (rule number, rule description) supporting this modification. Summarize the above information to generate a logical update report.

[0209] Step S159: Retrieve the knowledge node identifier, knowledge status label, and associated edge information corresponding to each updated knowledge content from the knowledge evolution lifecycle network, and generate a knowledge node association report.

[0210] Each updated knowledge content corresponds to a knowledge node (knowledge content node) in the knowledge evolution lifecycle network. By associating knowledge content with identifiers, the network retrieves the identifiers (unique identifiers of knowledge units), knowledge state labels (knowledge nascent stage / mature stage / iterative stage), and all associated edge information (edge ​​identifiers, associated node identifiers, and edge attribute descriptions) of these knowledge nodes. This information is grouped by knowledge node identifiers to generate a knowledge node association report, displaying the position and relationships of the updated knowledge within the entire knowledge evolution network.

[0211] Step S1510: Integrate the parameter change report, logic update report, knowledge node association report, and memory anchoring record, organize them according to the update process sequence, add the update operation initiation time, executor identifier, model version number, and parameter package version identifier, and generate a knowledge update full-process traceability document containing update content, update module, corresponding knowledge node information, and memory anchoring record.

[0212] The parameter change reports (knowledge storage module update), logic update reports (inference calculation module update), knowledge node association reports (knowledge source and evolution), and memory anchoring records (memory retention module update) are organized according to the knowledge update process sequence (parameter update → logic update → memory anchoring). At the beginning of the document, add the update operation initiation time (the time the knowledge update parameter package was input into the model), the executor identifier (the administrator or system identifier responsible for this update), the current version number of the pediatric oncology teaching model, and the version identifier of the knowledge update parameter package used. Integrate all this information to form a complete knowledge update process traceability document.

[0213] Step S1511: Store the knowledge update full-process traceability document in the model update archive and establish the association mapping between the knowledge update full-process traceability document and the model version after this update.

[0214] The model update archive is a database specifically designed to store documents related to model updates. The generated knowledge update process traceability documents are stored in this archive, and each document is assigned a unique document identifier. Simultaneously, a mapping is established within the archive between document identifiers and the version number of the pediatric oncology teaching model after this knowledge update. This allows for quick retrieval of the corresponding knowledge update process traceability documents using the model version number, achieving bidirectional traceability of model versions and update history.

[0215] In one exemplary embodiment, a large-scale model knowledge dynamic iterative update system is provided. This system, applied to pediatric oncology teaching, can be a terminal, server, etc., and its internal structure diagram can be as follows: Figure 4 As shown, this large-scale model knowledge dynamic iterative update system for pediatric oncology teaching includes a processor, memory, input / output interface, communication interface, display unit, and input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface is used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, near-field communication, or other technologies. When the computer program is executed by the processor, it implements a method for dynamic iterative update of large-scale model knowledge for pediatric oncology teaching. The display unit is used to form a visually visible image and can be a display screen, projection device, or virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device can be a touch layer covering the display screen, or it can be a button, trackball, or touchpad set on the shell of a large model knowledge dynamic iteration update system for pediatric oncology teaching, or it can be an external keyboard, touchpad, or mouse, etc.

[0216] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for dynamic iterative updating of knowledge in a large model, characterized in that, The method includes: Collect multiple types of dynamic knowledge sources in the field of pediatric oncology. These sources include the latest medical guidelines, full texts of clinical research papers, medical conference minutes, and multicenter clinical case summary reports. Each knowledge source is accompanied by information on the knowledge release time and the identifier of the knowledge source institution. The various types of dynamic knowledge sources and their associated information are correlated to construct a knowledge evolution lifecycle network in the field of pediatric oncology. The knowledge evolution lifecycle network includes knowledge content nodes, knowledge state labels corresponding to the nodes, and temporal and logical association edges between the nodes. Based on the knowledge evolution lifecycle network, multi-source credibility cross-validation and dynamic conflict reconciliation are performed on each knowledge content node to obtain a target knowledge set that passes the verification and is conflict-free. The pre-built parameter fine-tuning and memory retention collaborative mechanism is invoked to convert the target knowledge set into a knowledge update parameter package adapted to the large-scale pediatric oncology teaching model. The knowledge update parameter package includes module parameter adjustment schemes, knowledge association constraint rules, and model memory anchoring parameters. The knowledge update parameter package is input into the pediatric oncology teaching model, and a collaborative update operation is performed on the knowledge storage module, reasoning calculation module, and memory retention module of the pediatric oncology teaching model to generate a knowledge update full-process traceability document. The knowledge update full-process traceability document includes the updated content, the updated module, the corresponding knowledge node information, and the memory anchoring record.

2. The method for dynamic iterative updating of large model knowledge according to claim 1, characterized in that, The process of associating and processing the various types of dynamic knowledge sources and their accompanying information to construct a knowledge evolution lifecycle network in the field of pediatric oncology includes: Extract core knowledge statements from each multi-type dynamic knowledge source. The core knowledge statements are complete statements containing medical conclusions, treatment suggestions, or case analysis conclusions. Each core knowledge statement must cover at least one medical concept specific to the field of pediatric oncology. Add the knowledge source organization identifier and knowledge publication time information to each core knowledge expression to generate an initial knowledge unit containing content, organization, and time attributes; The historical knowledge base stored within the large-scale pediatric oncology teaching model is retrieved, and the historical core expression, historical source institution identifier, and historical release time information corresponding to each historical knowledge are extracted from the historical knowledge base to generate a historical knowledge unit containing content, institution, and time attributes. All initial knowledge units and historical knowledge units are assessed for their knowledge status. If the knowledge content appears for the first time and is not related to historical knowledge, it is marked as the nascent stage of knowledge. If the knowledge content is published by at least two different institutions and there is no conflict, it is marked as the mature stage of knowledge. If the knowledge content modifies or supplements historical knowledge and is published later than the publication time of all related historical knowledge units, it is marked as the iterative stage of knowledge. Each knowledge unit is categorized according to its knowledge state label, forming a set of knowledge nascent units, a set of knowledge mature units, and a set of knowledge iterative units. For each unit in the knowledge nascent stage unit set, query whether there are any related knowledge units that are subsequently published. If so, establish a time-related edge from the nascent stage unit to the subsequent unit, and label the edge attribute as time continuation. For each unit in the knowledge maturity stage unit set, query whether there is a core expression unit published by other institutions. If so, establish a logical connection edge between the units, and mark the edge attribute as consensus verification. For each unit in the knowledge iteration state unit set, query its corrected or supplemented historical knowledge units, establish logical connection edges from historical units to iteration state units, and label the edge attributes as knowledge update; All knowledge units, knowledge status labels, and related edges are organized hierarchically and structurally according to the time dimension and the logical dimension. The time dimension is sorted according to the publication time, and the logical dimension is classified according to the attributes of the related edges, forming a knowledge evolution life cycle network that includes knowledge content nodes, knowledge status labels corresponding to nodes, and time-related edges and logical-related edges between nodes.

3. The method for dynamic iterative updating of large model knowledge according to claim 2, characterized in that, The knowledge status is determined for all initial knowledge units and historical knowledge units. If the knowledge content appears for the first time and has no related historical knowledge, it is marked as the knowledge budding state. If knowledge content is published by at least two different institutions without conflict, it is marked as knowledge maturity. If the knowledge content modifies or supplements historical knowledge and its release time is later than the release time of all related historical knowledge units, it is marked as a knowledge iteration state, including: A knowledge content comparison database for the field of pediatric oncology is constructed, which includes all historical core expressions in the historical knowledge base and the initial knowledge unit core expressions with marked status. The core knowledge representation of each knowledge unit to be judged is input into the knowledge content comparison database, and semantic similarity retrieval is performed, with the retrieval scope covering all stored core representations; If the search results show that there are no stored core expressions with semantic similarity higher than the preset threshold, then the knowledge unit is determined to have no related historical knowledge and is marked as a knowledge budding stage. If the search results show that there are at least two core statements published by different institutions, and the semantic similarity between the core statements and the core knowledge statements of the knowledge unit to be judged is higher than a preset threshold, and the content of the statements is not contradictory, then the institution identifiers corresponding to the core statements are retrieved, and it is confirmed that the institution identifiers are different from each other. The knowledge unit is then judged to meet the multi-institution consensus condition and marked as knowledge maturity. If the search results show that the core expression of a historical knowledge unit is semantically related to the core knowledge expression of the knowledge unit to be judged, and the core knowledge expression of the knowledge unit to be judged contains a correction statement or supplementary statement to the historical core expression, then the knowledge publication time information of the knowledge unit to be judged and the historical publication time information of the historical knowledge unit are extracted. Compare the knowledge release time information of the knowledge unit to be judged with the historical release time information of historical knowledge units. If the release time of the knowledge unit to be judged is later than the release time of all related historical knowledge units, then the knowledge unit is judged to be the latest updated knowledge and marked as the knowledge iteration state. For each marked knowledge unit, record the search results, organization identifiers, and time comparison results involved in the judgment process, generate a knowledge status judgment basis document, and associate and bind the judgment basis document with the knowledge unit.

4. The method for dynamic iterative updating of large model knowledge according to claim 2, characterized in that, The document describes a hierarchical and structured organization of all knowledge units, knowledge status tags, and related edges according to time and logical dimensions. The time dimension is sorted by publication date, and the logical dimension is categorized by related edge attributes, including: A time-dimensional hierarchical framework is created, which is divided into multiple time layers according to time intervals. Each time layer corresponds to a fixed time interval, and the length of the time interval is determined based on the average knowledge update cycle in the field of pediatric oncology. Each knowledge unit is assigned to a corresponding time layer based on its knowledge release time information or historical release time information. Knowledge units in the same time layer are arranged in chronological order of release time. Within each time layer, a knowledge unit index table is established, which includes a knowledge unit identifier, a core knowledge description summary, a knowledge status label, and a knowledge source organization identifier. Create a logical dimension classification framework, which is divided into three logical categories based on the attributes of the associated edges: time continuity, consensus verification, and knowledge update. All time-related edges and logically related edges are classified into corresponding logical categories according to their attributes. A list of related edges is established under each logical category. The list of related edges includes edge identifier, starting knowledge unit identifier, ending knowledge unit identifier and edge attribute description. A mapping relationship is established between the time-dimensional hierarchical framework and the logical-dimensional classification framework. The position of each knowledge unit in the time layer is associated with the logical category of the related edges that the knowledge unit participates in. A graph retrieval interface is constructed, which supports three retrieval methods: querying knowledge units by time layer, querying related edges by logical category, and filtering knowledge units by knowledge status label. By integrating the time-dimensional hierarchical framework, the logical-dimensional classification framework, the mapping relationship, and the graph retrieval interface, a structured organizational result of the knowledge evolution lifecycle network is formed.

5. The method for dynamic iterative updating of large model knowledge according to any one of claims 1-2, characterized in that, Based on the knowledge evolution lifecycle network, multi-source credibility cross-validation and dynamic conflict reconciliation are performed on each knowledge content node to obtain a set of target knowledge that passes verification and is conflict-free, including: Extract the knowledge source institution identifier corresponding to each knowledge content node from the knowledge evolution life cycle network, and retrieve the institution qualification description from the institution professional archive based on the knowledge source institution identifier. The institution qualification description includes the institution's scope of diagnosis and treatment in the field of pediatric oncology, research project experience and industry certification status. Extract the knowledge state label corresponding to each knowledge content node from the knowledge evolution life cycle network. If the label is in the knowledge budding stage, focus on verifying the research project experience of the knowledge source institution; if the label is in the knowledge maturity stage, focus on verifying the institution's industry certification status; if the label is in the knowledge iteration stage, simultaneously verify the institution's scope of diagnosis and treatment and research project experience. Extract supporting information citation identifiers from the core knowledge statements corresponding to the knowledge content nodes. The supporting information citation identifiers are clinical case numbers, experimental study numbers, or literature index numbers. Based on the supporting information citation identifier, the corresponding supporting information original text is retrieved from the multi-center clinical case database, experimental research results database and medical literature database. The supporting information original text includes clinical case diagnosis and treatment records, experimental data collection process and core conclusions of the literature. A completeness analysis of the supporting information in the original text is required. Clinical case records must include basic patient information, tumor type, treatment plan, and efficacy evaluation results. The experimental data collection process must include the experimental design, sample size, data measurement methods, and statistical methods. The core conclusions of the literature must include the research objective, research methods, and the process of derivation of the conclusions. The system links the description of the institution's qualifications with the results of the analysis of the completeness of the original supporting information to generate an initial credibility assessment report for each knowledge content node. Knowledge content nodes that meet the preset basic credibility standards in the initial credibility assessment report are selected as candidate credible nodes; Cross-source cross-validation is performed on candidate trusted nodes. If a candidate trusted node is in the knowledge maturity stage or knowledge iteration stage, at least one related knowledge node published by another source institution is retrieved, and the core knowledge descriptions and supporting information of the two are compared. If the descriptions are consistent and the supporting information types are complementary, the credibility level of the related knowledge node is improved. Identify node combinations with knowledge update-related edges among candidate trusted nodes. If the core knowledge statements of two nodes in the combination are contradictory, extract the knowledge publication time information of both nodes and retain nodes whose publication time is later than the publication time of all related historical knowledge units. For the retained nodes, extract the update basis description from the original text of their supporting information, and associate and bind the update basis description with the core knowledge expression; The core knowledge descriptions, associated supporting information, and update basis descriptions corresponding to all knowledge content nodes that have passed cross-source cross-validation and completed conflict reconciliation are summarized to generate a set of target knowledge that has passed verification and is conflict-free.

6. The method for dynamic iterative updating of large model knowledge according to claim 5, characterized in that, The process of linking the institutional qualification description with the results of the integrity analysis of the original supporting information to generate an initial credibility assessment report for each knowledge content node includes: An institutional qualification assessment index system is constructed, which includes three indicators: the matching degree of the scope of diagnosis and treatment, the relevance of scientific research projects, and the industry certification level. Each indicator corresponds to multiple assessment levels, and the assessment level is determined based on the specific content in the institutional qualification description. For each knowledge content node, the institution's qualification description is evaluated according to the institution qualification assessment index system. The scope of treatment matching index determines whether the institution's scope of treatment is consistent with the tumor types involved in the knowledge content. The research project relevance index determines whether the institution's past research projects are the same as the research direction of the knowledge content. The industry certification level index determines whether the industry certification obtained by the institution covers the relevant fields of the knowledge content. A supporting information integrity assessment index system is constructed, which includes three indicators: completeness of clinical case elements, completeness of experimental data records, and completeness of literature conclusion derivation. Each indicator corresponds to multiple assessment levels, and the assessment level is determined based on the content coverage of the original supporting information. For each knowledge content node, the supporting information text is used to determine the level of each indicator according to the supporting information integrity assessment index system. The clinical case element integrity index determines whether the case record contains all the necessary elements. The experimental data record integrity index determines whether the experimental process record covers the key steps. The literature conclusion derivation integrity index determines whether the literature conclusion has a derivation process to support it. Based on the evaluation levels of each indicator in the institutional qualification assessment indicator system and the evaluation levels of each indicator in the supporting information integrity assessment indicator system, and according to the preset credibility assessment rules, an initial credibility assessment conclusion is generated for each knowledge content node, wherein the credibility assessment conclusion is either in line with the basic credibility standard or not in line with the basic credibility standard. The initial credibility assessment report for each knowledge content node is formed by integrating the description of the institution's qualifications, the results of the analysis of the completeness of the original supporting information, the evaluation level of each indicator, and the initial credibility assessment conclusion generated according to the credibility assessment rules.

7. The method for dynamic iterative updating of large model knowledge according to claim 5, characterized in that, The cross-source verification of candidate trusted nodes involves, if the candidate trusted node is in a knowledge maturity or knowledge iteration stage, retrieving at least one related knowledge node published by another source institution, comparing the core knowledge statements and supporting information of the two, and if the statements are consistent and the supporting information types are complementary, then the credibility level of the related knowledge node is increased, including: The knowledge maturity stage nodes and knowledge iteration stage nodes among the candidate trusted nodes are classified to form a cross-source verification node set; For each node in the cross-source verification node set, the graph retrieval interface of the knowledge evolution life cycle network is used to retrieve related knowledge nodes published by other source institutions according to the semantic similarity of the core knowledge representation. The retrieval results must contain at least one knowledge node that is different from the source institution of the current node. Retrieve the knowledge source organization identifier, core knowledge description, and supporting information of each related knowledge node to generate verification materials for the related nodes; The core knowledge description of the current node is compared sentence by sentence with the core knowledge description of the related knowledge nodes. The identical and different descriptions are extracted. If the difference is only in the way it is expressed but the core meaning is the same, then the descriptions are considered to be the same. Analyze the supporting information types of the current node and the supporting information types of related knowledge nodes. Supporting information types include clinical case data, experimental research data, and literature citation data. If the supporting information types of the current node and the related knowledge nodes are different, it is determined that the supporting information types are complementary. If the core knowledge descriptions of the current node and the related knowledge nodes are consistent and the supporting information types are complementary, then the initial credibility assessment reports of both nodes are retrieved, and the credibility level of the current node is raised by one level. The level increase is recorded in the supplementary explanation column of the assessment report. If there are multiple related knowledge nodes in the search results, their descriptions are compared with the current node and their supporting information types are analyzed. If all related knowledge nodes meet the requirements of consistent descriptions and complementary supporting information types, the credibility level of the current node is increased by two levels. If there is a contradiction between the core knowledge expression of the related knowledge node and the current node (not a difference in expression method), then record the contradiction point and the related knowledge node information as a reference for subsequent conflict reconciliation. For each candidate trusted node that has completed cross-source cross-validation, a cross-source verification report is generated. The cross-source verification report includes information on associated knowledge nodes, description comparison results, analysis results of supporting information types, and adjustments to the trust level. The cross-source verification report is merged with the initial trust assessment report to form trust assessment materials.

8. The method for dynamic iterative updating of large model knowledge according to claim 1, characterized in that, The aforementioned invocation of a pre-built parameter fine-tuning and memory retention collaborative mechanism transforms the target knowledge set into a knowledge update parameter package adapted to the large-scale pediatric oncology teaching model, including: For each knowledge content in the target knowledge set, the topic is classified according to the knowledge storage module division standard of the pediatric oncology teaching model. The knowledge storage module division standard is determined based on the knowledge system in the field of pediatric oncology. The knowledge content under the same topic category is aggregated to generate a target knowledge subset for each knowledge storage module. Each target knowledge subset contains the knowledge content that needs to be added or updated for that knowledge storage module. Retrieve the current parameter configuration file for each knowledge storage module. The parameter configuration file contains the weight parameter matrix, knowledge association parameter list, and parameter adjustment history of the module. The feature parsing module of the input parameter fine-tuning and memory retention synergy mechanism for each target knowledge subset extracts medical concept features and logical relationship features from the knowledge content to generate a feature vector matrix. The dimension of the feature vector matrix is ​​consistent with the dimension of the weight parameter matrix of the corresponding knowledge storage module. By comparing the feature vector matrix and the weight parameter matrix, the difference between the two in each parameter dimension is calculated. Based on the difference, the position and direction of the parameters that need to be adjusted are determined, and a preliminary parameter adjustment plan is formed. Based on the preliminary parameter adjustment scheme, a model memory retention constraint is introduced to retrieve historical key parameters from the knowledge storage module. These historical key parameters are parameters that have a significant impact on the core reasoning ability of the model. During the adjustment process, the adjustment range of the adjustment items involving the position of historical key parameters is reduced to a range close to zero, so that the values ​​of historical key parameters remain stable. Only non-key parameters are adjusted according to the preliminary scheme to form the final module parameter adjustment scheme. Analyze the logical relationships between knowledge content within each target knowledge subset, and construct knowledge association constraint rules based on the logical relationships. The knowledge association constraint rules clarify the calling order and association priority of knowledge content in the reasoning process. The logical relationships include causal relationships, step sequence relationships, and subordinate relationships. Construct model memory anchoring parameters, which are used to mark the parameter positions corresponding to the core knowledge that needs to be retained in the knowledge storage module for a long time. In subsequent updates, the parameter adjustment range of the anchoring parameter positions is controlled within a preset range. The module parameter adjustment schemes for all knowledge storage modules, the knowledge association constraint rules for all knowledge topics, and the model memory anchoring parameters for the entire model are summarized. The parameter package version identifier and generation time information are added to generate a knowledge update parameter package adapted to the large-scale pediatric oncology teaching model.

9. The method for dynamic iterative updating of large model knowledge according to claim 8, characterized in that, Based on the initial parameter adjustment scheme, a model memory retention constraint is introduced. Historical key parameters are retrieved from the knowledge storage module. These historical key parameters are those that significantly influence the model's core reasoning ability. During the adjustment process, the adjustment magnitude of items involving the location of historical key parameters is reduced to near zero, ensuring the stability of their values. Only non-key parameters are adjusted according to the initial scheme, forming the final module parameter adjustment scheme, including: A historical key parameter identification model is constructed. The historical key parameter identification model is trained and generated based on the historical inference performance data of the large-scale pediatric oncology teaching model. It is used to determine whether a parameter is a key parameter based on the degree of influence of the parameter on the inference result. Input the weight parameter matrix from the current parameter configuration file of the knowledge storage module into the historical key parameter identification model. The model outputs a keyness score for each parameter. The higher the keyness score, the greater the impact of the parameter on the core reasoning ability. Parameters whose criticality scores are higher than the critical parameter threshold are marked as historical critical parameters, and a list of historical critical parameters is generated, which includes the parameter position, current value and criticality score. In the initial parameter adjustment plan, adjustment items involving the location of historical key parameters are selected, and the values ​​of these adjustment items are numerically corrected to reduce the adjustment range to near 0, so that the values ​​of historical key parameters remain stable. For adjustment items in the preliminary parameter adjustment plan that do not involve the location of historical key parameters, the original adjustment direction and adjustment range are retained to form a revised set of parameter adjustment items; The revised set of parameter adjustment items is sorted according to the parameter position to generate a parameter adjustment order table. The adjustment order is determined based on the functional priority of the parameters in the knowledge storage module, with parameters with higher functional priority being adjusted first. Add an adjustment basis description for each parameter adjustment item. The adjustment basis description includes the feature vector difference value of the corresponding knowledge content, the parameter criticality score, and the calculation process of the adjustment range. The parameter adjustment sequence table, all parameter adjustment items, and corresponding adjustment basis descriptions are integrated to form the final module parameter adjustment scheme.

10. A large-scale model knowledge dynamic iterative update system, characterized in that, include: processor; A machine-readable storage medium for storing machine-executable instructions of the processor; The processor is configured to execute the large model knowledge dynamic iterative update method according to any one of claims 1 to 9 by executing the machine-executable instructions.

Images