Tumor knowledge base construction system and method based on large model and prompt word optimization

Abstract

The present invention relates to the field of medical big data technology in artificial intelligence, and in particular to a tumor knowledge base construction system and method based on a large model and prompt word optimization, the system comprising: a data acquisition module, a prompt word generation and optimization module, an entity recognition and relationship extraction module, a knowledge graph construction module, a clinical decision support module, and a knowledge base update and maintenance module. The present invention uses prompt words that are initially designed and continuously optimized to ensure that the subsequent large language model LLM accurately and efficiently extracts required information, and uses the large language model LLM to quickly and accurately extract related entities and the association relationship between related entities from target text information, so as to construct an accurate hierarchical tumor knowledge graph database. When the hierarchical tumor knowledge graph database is subsequently applied, a personalized clinical report on recommended treatment plans can be generated by inputting the clinical data of the target patient, thereby providing better decision support for medical personnel.

Description

Technical Field

[0001] The present invention relates to the field of medical big data technology in artificial intelligence, and in particular to a tumor knowledge base construction system and method based on a large model and prompt word optimization. Background Art

[0002] Cancer is a major global health problem and poses a serious threat to human health. Cancer treatment has entered the era of precision medicine. Precision cancer medicine integrates genomics, molecular biology, clinical data and other information to tailor the most appropriate treatment plan for each cancer patient. The tumor knowledge base improves the accuracy and effectiveness of cancer treatment by providing detailed genomic and clinical data to support individualized treatment, targeted therapy, immunotherapy and prognosis assessment. For example, the OncoKB database is a precision oncology knowledge base that contains a large amount of information on tumor gene mutations and drug responses. It is used to assist doctors in developing individualized treatment plans, promote the transformation of cancer treatment from the traditional "one prescription for thousands of people" to "tailor-made", and improve patients' treatment effects and quality of life.

[0003] However, with the continuous deepening of medical research and the continuous advancement of medical technology, more and more tumor-related data are generated and accumulated, including scientific literature, professional guidelines and recommendations, treatment strategies, multi-dimensional omics information, etc. These data are scattered in different databases and literatures, and the data volume is huge and complex. Traditional manual collation and analysis methods have problems such as inconsistent data sources, low data collation efficiency, and inaccurate knowledge extraction, which makes it difficult to obtain and apply tumor-related information, and thus cannot meet the needs of rapid acquisition and application of tumor knowledge. In view of this, how to quickly and accurately extract relevant information from a large number of literature, databases and clinical guidelines, and build a dynamically updated and clinically applicable tumor knowledge base, has become a technical problem that needs to be solved in the field of tumor research and clinical adjuvant therapy.

[0004] Therefore, a tumor knowledge base construction system and method based on large model and prompt word optimization were proposed. Summary of the invention

[0005] Based on this, it is necessary to provide a tumor knowledge base construction system and method based on a large model and prompt word optimization to address the above technical issues.

[0006] According to a first aspect of the present invention, a tumor knowledge base construction system based on a large model and prompt word optimization is provided, including: a data acquisition module, used to obtain data related to tumor diseases from multiple data sources; a prompt word generation and optimization module, used to interact with a large language model LLM through prompt words that are initially designed and continuously optimized; an entity recognition and relationship extraction module, used to extract relevant entities and the association relationships between relevant entities from data related to tumor diseases using the large language model LLM; a knowledge graph construction module, used to construct a hierarchical tumor knowledge graph database based on the relevant entities and the association relationships between the relevant entities; a clinical decision support module, used to generate a personalized clinical report on recommended treatment plans based on the clinical data of the target patient and the constructed hierarchical tumor knowledge graph database; a knowledge base update and maintenance module, used to regularly collect the latest data related to tumor diseases according to a preset retrieval cycle, automatically extract new relevant entities and the association relationships between new relevant entities, and update the hierarchical tumor knowledge graph database.

[0007] Optionally, the data acquisition module includes: a literature data acquisition submodule, which is used to obtain original literature data related to tumor diseases from PubMed database, Google Scholar database, Web of Sciencee database, Scopus database, and CNKI database; a clinical guideline acquisition submodule, which is used to obtain original tumor conference paper data or original clinical guideline data related to tumor diseases from ASCO website, ESMO website, CSCO website, FDA website, EMA website, and NMPA website; a biomedical database acquisition submodule, which is used to obtain attribute information of related entities from NCBI biomedical database and DrugBank biomedical database.

[0008] Optionally, the data acquisition module also includes: a data preprocessing submodule, which is used to perform data deduplication processing, data format unification processing and redundant data elimination processing on the acquired original literature data, original tumor conference paper data, and original clinical guideline data to obtain target text information, wherein the redundant data elimination processing on the acquired original literature data, original tumor conference paper data, and original clinical guideline data includes: if N original literature data and / or original tumor conference paper data and / or original clinical guideline data all correspond to the same conclusion, N is an integer greater than a preset value, according to a preset priority screening rule, the priority screening rule includes at least one of the number of citations, experimental design type, sample size and journal impact factor arranged in descending order of priority, and the experimental design type includes at least one of randomized controlled trial research, cohort study, case control study, case series study, case study, animal experimental study, cell line experimental study and experimental deviation study arranged in descending order of priority, and the highest priority literature data or tumor conference paper data or clinical guideline data are screened out from the N original literature data and / or original tumor conference paper data and / or original clinical guideline data in order of priority from high to low.

[0009] Optionally, the prompt word generation and optimization module includes: a prompt word generation submodule, which is used to obtain a target processing task, identify the prompt word category to which the target processing task belongs according to a user-defined prompt word category, and generate an Mth prompt word based on the information range of the corresponding prompt word category, where M is an integer greater than 0; the prompt word category includes a tumor targeted drug category, a tumor immunotherapy category, a tumor diagnosis category, and a tumor prognosis category; a large language model calling submodule, which is used to input the Mth prompt word into a large language model LLM and output the Mth model generation result; an output result evaluation submodule, which is used to compare the Mth model generation result with the expected answer result in the standard data set based on a preset evaluation standard to generate the Mth output model evaluation result; a feedback mechanism submodule, which is used to generate the Mth feedback information according to the Mth output model evaluation result; and a feedback and prompt word optimization submodule, which is used to optimize the Mth prompt word according to the Mth feedback information to obtain the M+1th prompt word.

[0010] Optionally, the feedback and prompt word optimization submodule also includes: a multi-round interactive optimization submodule, which is used to input the M+1th prompt word into the large language model LLM, output the M+1th prompt word model generation result, compare the M+1th prompt word model generation result with the expected answer result in the standard data set based on a preset evaluation standard, generate the M+1th output model evaluation result, generate the M+1th feedback information according to the M+1th output model evaluation result, optimize the M+1th prompt word according to the M+1th feedback information, and obtain the M+2th prompt word.

[0011] Optionally, the entity recognition and relationship extraction module includes: an entity recognition submodule, which is used to parse the target text information according to the optimized prompt words using a large language model LLM, and identify related entities contained in the target text information, wherein the related entities include genes, gene mutations, disease types and drugs; a relationship extraction submodule, which is used to extract the association relationship between related entities according to the semantic relationship in the target text information, wherein the association relationship is a causal relationship or a drug response relationship.

[0012] Optionally, the knowledge graph construction module includes: a node and edge construction submodule, which is used to create each related entity as a parent node, determine whether there is an association relationship between two related entities, and if there is an association relationship between the two related entities, create a first-class edge with directionality between the two related entities, wherein the first-class edge with directionality is determined by the causal relationship or the drug response relationship, and construct an initial tumor knowledge graph database based on the created parent node and the first-class edge; a graph hierarchy and classification submodule, which is used to create a child node under the parent node in the constructed initial tumor knowledge graph database according to the classification attribute relationship of the drug or the subtype attribute relationship of the disease, and create a second-class edge between the parent node and the child node according to the classification attribute relationship of the drug or the subtype attribute relationship of the disease between the parent node and the child node, and construct a hierarchical tumor knowledge graph database based on the created child node and the second-class edge; a data fusion and expansion submodule, which is used to store the attribute information of the related entities obtained by the biomedical database acquisition submodule into the hierarchical tumor knowledge graph database; a knowledge graph visualization submodule, which is used to visualize and query the hierarchical tumor knowledge graph database.

[0013] Optionally, the clinical decision support module includes: a personalized clinical decision support submodule, which is used to input the target patient's genome sequencing data and / or diagnostic information and / or clinical history data into a hierarchical tumor knowledge graph database for query processing, and output a first category of recommended solutions including gene mutations and drugs; a gene and drug matching submodule, which is used to input the target patient's gene mutation information into a hierarchical tumor knowledge graph database for query processing, output a second category of recommended solutions including drugs, and obtain a drug reliability assessment result based on the evidence level based on the target text information corresponding to the drug in the second category of recommended solutions; a clinical report generation submodule, which is used to input the target patient's clinical data into a hierarchical tumor knowledge graph database for query processing, and output a personalized clinical report on recommended treatment solutions, wherein the clinical report on recommended treatment solutions includes information on gene mutations, recommended treatment solutions related to gene mutations, the evidence level of the recommended treatment solutions, recommended drugs, and the target text information to which the recommended drugs belong, and the recommended treatment solutions are generated based on related entities including genes, gene mutations, disease types, and drugs, and the association relationship between the related entities; the evidence level of the recommended treatment solution is determined based on the treatment solution approved on the FDA website or the NMPA website or in the clinical research stage.

[0014] Optionally, the knowledge base update and maintenance module includes: a knowledge base dynamic update submodule, which is used to regularly collect the latest data related to tumor diseases according to a preset retrieval cycle, automatically extract new related entities and the association relationships between new related entities, and update the hierarchical tumor knowledge graph database; a database version control submodule, which is used to perform version management on the hierarchical tumor knowledge graph database.

[0015] According to a second aspect of the present invention, a method for constructing a tumor knowledge base based on a large model and prompt word optimization is provided, comprising: obtaining data related to tumor diseases from multiple data sources through a data acquisition module; interacting with a large language model LLM through prompt words initially designed and continuously optimized by a prompt word generation and optimization module; extracting related entities and the association relationships between related entities from data related to tumor diseases using a large language model LLM through an entity recognition and relationship extraction module; constructing a hierarchical tumor knowledge graph database based on the association relationships between related entities and related entities through a knowledge graph construction module; generating a personalized clinical report on recommended treatment plans based on the constructed hierarchical tumor knowledge graph database according to the clinical data of target patients through a clinical decision support module; and regularly collecting the latest data related to tumor diseases according to a preset retrieval cycle through a knowledge base update and maintenance module, automatically extracting new related entities and the association relationships between new related entities, and updating the hierarchical tumor knowledge graph database.

[0016] The advantages and beneficial effects of the present invention are as follows: the present invention provides a tumor knowledge base construction system and method based on a large model and prompt word optimization. The system obtains data related to tumor diseases from multiple data sources, and performs data deduplication processing, data format unification processing and redundant data elimination processing on the obtained data, thereby ensuring the accuracy and reliability of the data for accurate data analysis; through the initial design and continuous optimization of prompt words, it is ensured that the subsequent large language model LLM accurately and efficiently extracts the required information, and then the large language model LLM is used to quickly and accurately extract related entities and the association relationship between related entities from the target text information, so as to construct an accurate hierarchical tumor knowledge graph database. When the hierarchical tumor knowledge graph database is subsequently applied, the clinical data of the target patient can be input to generate a personalized clinical report of recommended treatment plans, thereby providing better decision support for medical personnel; in addition, by setting a retrieval cycle, the latest data related to tumor diseases can be regularly collected, and the hierarchical tumor knowledge graph database can be automatically updated to ensure its latestness and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 This is an exemplary data sending and receiving scenario diagram to which the present application can be applied.

[0018] Figure 2 It is a structural diagram of an embodiment of a tumor knowledge base construction system based on a large model and prompt word optimization according to the present application.

[0019] Figure 3 It is a flowchart of an embodiment of a method for constructing a tumor knowledge base based on a large model and prompt word optimization according to the present application. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below through specific implementation methods in conjunction with the accompanying drawings. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0021] In view of the existing technology, data related to tumor diseases are scattered in different databases and documents, and the data volume is huge and complex. Traditional manual collation and analysis methods have problems such as inconsistent data sources, low data collation efficiency, and inaccurate knowledge extraction. This application proposes a tumor knowledge base construction system based on large models and prompt word optimization. The system continuously optimizes the prompt word content to improve the accuracy and efficiency of the large language model LLM in extracting required information, thereby realizing rapid and accurate extraction of relevant information from a large number of documents, databases and clinical guidelines, and constructing a dynamically updated, clinically applicable tumor knowledge base to meet the needs of clinical scientific research data mining and assist doctors and researchers in making more accurate decisions.

[0022] Figure 1 The diagram schematically shows a data transmission and reception scenario diagram between a terminal device and a server when a user uses a large model on a terminal device according to an embodiment of the present disclosure.

[0023] like Figure 1 As shown, the network 104 is used to provide a medium for communication links between the terminal devices 101, 102, 103 and the server 105. The network 104 may include various connection types, such as wired, wireless communication links or optical fiber cables, etc.

[0024] Users can use terminal devices 101, 102, 103 to interact with server 105 through network 104 to receive or send messages, etc. Terminal devices 101, 102, 103 can be installed with various communication client applications, such as web browser applications, search applications, etc. (only examples).

[0025] The terminal devices 101, 102, 103 may be various electronic devices with display screens and supporting web browsing, including but not limited to smart phones, tablet computers, laptop computers, and desktop computers. The server 105 may be a server that provides various services, such as a background management server (only as an example) that provides support for the large model website browsed by the user using the terminal devices 101, 102, 103. The background management server may analyze and process the received data such as the user's question request, and feed back the processing results (such as obtaining the most matching large model according to the user's question, or generating relevant information or data of the user's first question or subsequent questions, etc.) to the terminal devices 101, 102, 103.

[0026] It should be noted that the tumor knowledge base construction system based on the large model and prompt word optimization provided in the embodiment of the present disclosure can generally be set in the server 105. Accordingly, the tumor knowledge base construction method based on the large model and prompt word optimization provided in the embodiment of the present disclosure can be executed by the server 105. The tumor knowledge base construction system based on the large model and prompt word optimization provided in the embodiment of the present disclosure can also be set in a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105. Accordingly, the tumor knowledge base construction method based on the large model and prompt word optimization provided in the embodiment of the present disclosure can also be executed by a server or server cluster that is different from the server 105 and can communicate with the terminal devices 101, 102, 103 and / or the server 105.

[0027] It should be understood that Figure 1 The number of terminal devices, networks and servers in the embodiment is only for illustration. Any number of terminal devices, networks and servers may be provided according to implementation requirements.

[0028] The following will be based on Figure 1 The scene described by Figure 2 The tumor knowledge base construction system based on a large model and prompt word optimization of the disclosed embodiment is described in detail.

[0029] like Figure 2 The structure diagram of an embodiment of a tumor knowledge base construction system based on a large model and prompt word optimization according to the present application is shown, and the system can be specifically applied to various electronic devices. The tumor knowledge base construction system 300 based on a large model and prompt word optimization described in this embodiment includes: a data acquisition module 310, a prompt word generation and optimization module 320, an entity recognition and relationship extraction module 330, a knowledge graph construction module 340, a clinical decision support module 350 and a knowledge base update and maintenance module 360.

[0030] The data acquisition module 310 is used to acquire data related to tumor diseases from multiple data sources.

[0031] According to one embodiment of the present application, the data acquisition module 310 includes a literature data acquisition submodule, a clinical guideline acquisition submodule, a biomedical database acquisition submodule and a data preprocessing submodule.

[0032] The literature data collection submodule is used to obtain original literature data related to tumor diseases from PubMed database, Google Scholar database, Web of Science database, Scopus database, and CNKI database. Furthermore, this application can capture the latest literature data related to specific genes, tumor types, and treatment plans through a preset search cycle, such as automatic search once a day, to ensure that the latest research results are covered.

[0033] The clinical guideline collection submodule is used to obtain original tumor conference paper data or original clinical guideline data related to tumor diseases from the ASCO website, ESMO website, CSCO website, FDA website, EMA website, and NMPA website. Furthermore, this application can obtain the latest original tumor conference paper data or original clinical guideline data related to tumor diseases from the ASCO website, ESMO website, CSCO website, FDA website, EMA website, and NMPA website through API or crawler technology, especially information about drug efficacy, gene mutation and disease association.

[0034] The biomedical database acquisition submodule is used to obtain attribute information of genes, drugs, diseases and other related entities from the NCBI biomedical database and the DrugBank biomedical database, for example, gene-related omics information and drug-related attribute information, etc. These attribute information are an important basis for the subsequent construction of the knowledge graph.

[0035] The data preprocessing submodule is used to perform data deduplication processing, data format unification processing, and redundant data elimination processing on the acquired original literature data, original tumor conference paper data, and original clinical guideline data to obtain the target text information.

[0036] Furthermore, in order to ensure the accuracy and reliability of the data and to conduct accurate data analysis, the present application performs redundant data elimination processing on the acquired original literature data, original tumor conference paper data, and original clinical guideline data. The specific implementation steps of the redundant data elimination processing include: if N original literature data and / or original tumor conference paper data and / or original clinical guideline data all correspond to the same conclusion, N is an integer greater than a preset value, according to a preset priority screening rule, the priority screening rule includes at least one of the number of citations, experimental design type, sample size, and journal impact factor arranged in descending order of priority, and the experimental design type includes at least one of randomized controlled trial research, cohort study, case control study, case series study, case study, animal experimental study, cell line experimental study, and experimental deviation study arranged in descending order of priority, and the highest priority literature data or tumor conference paper data or clinical guideline data are screened out from the N original literature data and / or original tumor conference paper data and / or original clinical guideline data in order of priority from high to low.

[0037] For example, when the same conclusion is supported by five original literature data, three original tumor conference paper data, and three original clinical guideline data, the selection can be made according to the following priority screening rules: number of citations > experimental design type > sample size > journal impact factor. The priority screening rule for experimental design type is: randomized controlled trial study > cohort study > case-control study > case series study > case study > animal experimental study > cell line experimental study > experimental deviation study. The five original literature data, three original tumor conference paper data, and three original clinical guideline data are sorted from high to low priority, and then the literature data or tumor conference paper data or clinical guideline data with the highest priority are screened out as the target text information to be retained.

[0038] The prompt word generation and optimization module 320 is used to interact with the large language model LLM through the prompt words that are initially designed and continuously optimized.

[0039] According to an embodiment of the present application, the prompt word generation and optimization module 320 includes a prompt word generation submodule, a large language model calling submodule, an output result evaluation submodule, a feedback mechanism submodule and a feedback and prompt word optimization submodule.

[0040] The prompt word generation submodule is used to obtain the target processing task, identify the prompt word category to which the target processing task belongs according to the customized prompt word category, and generate the Mth prompt word based on the information range of the corresponding prompt word category. M is an integer greater than 0. The prompt word categories include tumor targeted drug category, tumor immunotherapy category, tumor diagnosis category and tumor prognosis category, so as to ensure the subsequent targeted extraction of required information.

[0041] It can be understood that when M is 1, that is, the initial prompt word, the prompt word category to which the target processing task belongs is identified according to the customized prompt word category based on the target processing task, and the initial prompt word is generated based on the information range of the corresponding prompt word category.

[0042] Furthermore, the information scope of the prompt word category is based on the tumor targeted drug category, tumor immunotherapy category, tumor diagnosis category and tumor prognosis category, and the entity and entity relationship information related to the prompt word category to which the target processing task belongs is extracted from the target text information to guide the large language model LLM to output accurate information. For example, when processing tasks related to "tumor targeted drug", it is preferred to identify the prompt word category to which the target processing task belongs as the tumor targeted drug category according to the custom prompt word category. Based on the tumor targeted drug category, the information scope of the corresponding prompt word category is determined to be the entity related to tumor targeted drug in the literature that needs to be extracted. The initial prompt word can be designed as "Please extract and list six entities related to tumor treatment from the text: gene, variant, disease, drug, sensitivity and evidence level. These entities will be used to generate annotations to describe whether a certain tumor patient carrying a specific gene variant is sensitive or resistant to a specific drug. Extract entities from the text and list them as sextuples. Each sextuple should consist of six related entities, which together represent a treatment plan." That is, the initial prompt word clarifies the required scope of information, aiming to allow the large language model LLM to generate information about whether a certain tumor patient carrying a specific gene variant is sensitive or resistant to a specific drug.

[0043] The large language model calling submodule is used to input the Mth prompt word into the large language model LLM and output the Mth model generation result.

[0044] Furthermore, the large language model LLM used in this application is obtained after training based on the data processed by the data preprocessing submodule, and can be used to understand natural language and generate corresponding answers; by inputting the initial prompt word, the large language model LLM will return the extraction result, which is the initial model generation result, so as to realize the interaction process between the prompt word and the large language model LLM.

[0045] The output result evaluation submodule is used to compare the M-th model generation result with the expected answer result in the standard data set based on the preset evaluation criteria, and generate the M-th output model evaluation result.

[0046] Furthermore, after the model generates results, it is compared with the expected answer results in the standard data set based on the preset evaluation criteria, the deviation of the model output is recorded, and the corresponding output model evaluation results are output so that the corresponding feedback information can be generated later.

[0047] Furthermore, the evaluation criteria include the following four parts.

[0048] (1) Evaluation criteria for key entity recognition: automatic recognition of key entities is performed on the model generation results to determine whether the large language model (LLM) correctly extracts the core content of the prompt word. Key entities include but are not limited to genes, variants, diseases, drugs, sensitivity, and evidence levels. For example, for the model generation results, the system automatically identifies entities such as "EFGR", "Exon 19 Deletion", and "Erlotinib" and compares them with the expected entities in the standard data set to determine whether there is a deviation in the model output. If there is a deviation, the output model evaluation result is that the generated model generation result does not correctly extract the core content of the prompt word.

[0049] (2) Evaluation criteria for relationship extraction: Automatically extract the association between key entities in the model generation results to ensure that the model generation results meet the requirements in the prompt words. For example, in "EFGR Exon 19Deletion mutation in non-small cell lung cancer and the response to erlotinib", it is necessary to determine whether the large language model LLM correctly establishes the drug response relationship "Sensitive" between the gene mutation and the drug. If the drug response relationship between the gene mutation and the drug is not correctly established as "Sensitive", the output model evaluation result is that the generated model generation results did not correctly extract the association between key entities.

[0050] (3) Evaluation criteria for relevance: determine whether the text content of the model-generated results is consistent with the context of the prompt word to ensure that the output content of the model-generated results does not deviate from the topic. For example, if the prompt word focuses on "tumor targeted medication", determine whether the output content of the model-generated results contains other irrelevant medication information (such as "tumor chemotherapy information"). If it contains other irrelevant medication information, the output model evaluation result is that the generated model-generated results are not relevant to the prompt word.

[0051] (4) Evaluation criteria for completeness and accuracy: to determine whether the text content of the model-generated results fully answers the question in the prompt word, or whether it omits key entities or relationships, or whether it generates answers that are too brief or not detailed. For example, if the output content of the model-generated results does not output complete key entities and relationships or the output is inaccurate, then the output model evaluation result is that the output of the generated model-generated results is incomplete or inaccurate.

[0052] The feedback mechanism submodule is used to generate the Mth feedback information according to the Mth output model evaluation result.

[0053] Furthermore, the present application generates corresponding feedback information based on the above output model evaluation results, and the feedback information includes positive feedback information and negative feedback information; for example, for positive feedback information: if the output model evaluation result is that the model generation result fully meets the prompt word requirements, the system will record this successful prompt word design for use in subsequent similar tasks; for negative feedback information: if the output model evaluation result is that the model generation result is inaccurate, incomplete or irrelevant, the system will generate negative feedback information, indicating that the prompt word needs to be modified or the model processing method needs to be further optimized.

[0054] The feedback and prompt word optimization submodule is used to optimize the Mth prompt word according to the Mth feedback information to obtain the M+1th prompt word.

[0055] Furthermore, the present application can optimize the prompt words automatically or manually according to the feedback information. For example, in terms of modifying the prompt words, the vague prompt words can be adjusted to be clearer, more background information can be provided, or more specific professional terms can be used, such as the prompt words can be adjusted from "KRAS gene mutation" to "efficacy data of patients with KRAS G12C mutation" to ensure accurate capture of the context.

[0056] The multi-round interactive optimization submodule is used to input the M+1th prompt word into the large language model LLM, output the M+1th prompt word model generation result, compare the M+1th prompt word model generation result with the expected answer result in the standard data set based on the preset evaluation criteria, generate the M+1th output model evaluation result, generate the M+1th feedback information based on the M+1th output model evaluation result, optimize the M+1th prompt word based on the M+1th feedback information, and obtain the M+2th prompt word.

[0057] It can be understood that M+1 represents the next round of prompt words based on M, that is, when M is 1, it is the initial prompt word. After completing the prompt word generation, inputting the large language model LLM and outputting the model generation result, outputting the result evaluation, generating feedback information and the prompt word optimization steps, when an optimized prompt word is detected, the steps of inputting the large language model LLM and outputting the model generation result, outputting the result evaluation and generating feedback information can be further repeated. If the generated feedback information is positive feedback information, there is no need to optimize the prompt word again, and it can be used in the entity recognition and relationship extraction module. If the generated feedback information is negative feedback information, it is still necessary to further perform the prompt word optimization step and repeat the steps of inputting the large language model LLM and outputting the model generation result, outputting the result evaluation and generating feedback information until the generated feedback information is positive feedback information. The present application continuously generates feedback and adjusts the prompt words through multiple rounds of interaction to improve the accuracy and quality of the model output.

[0058] The entity recognition and relationship extraction module 330 is used to extract relevant entities and the association relationships between relevant entities from data related to tumor diseases using a large language model LLM.

[0059] According to one embodiment of the present application, the entity recognition and relationship extraction module 330 includes: an entity recognition submodule and a relationship extraction submodule.

[0060] The entity recognition submodule is used to parse the target text information using the large language model LLM according to the optimized prompt words, and identify the relevant entities contained in the target text information, wherein the relevant entities include genes, gene mutations, disease types and drugs.

[0061] Furthermore, the target text information is parsed using the large language model (LLM) through the optimized prompt words to automatically identify relevant entities in the target text information. The identified related entities include genes (such as "EGFR"), gene mutations (such as "Exon 19 Deletion"), disease types (such as "lung cancer"), and drugs (such as "erlotinib").

[0062] The relationship extraction submodule is used to extract the association relationship between related entities according to the semantic relationship in the target text information, and the association relationship is a causal relationship or a drug reaction relationship.

[0063] Furthermore, based on the semantic relationship in the target text information, the association relationship between related entities is extracted. For example, the association relationship between genes and gene mutations, drugs and disease types, for example, the extracted association relationship may be "The association relationship between lung cancer patients with EGFR Exon 19Deletion mutation and erlotinib treatment is: Sensitive.

[0064] The knowledge graph construction module 340 is used to construct a hierarchical tumor knowledge graph database based on related entities and the association relationships between related entities.

[0065] According to one embodiment of the present application, the knowledge graph construction module 340 includes: a node and edge construction sub-module, a graph hierarchy and classification sub-module, a data fusion and expansion sub-module and a knowledge graph visualization sub-module.

[0066] The node and edge construction submodule is used to create each related entity as a parent node, determine whether there is an association relationship between two related entities, and if there is an association relationship between the two related entities, create a directed first-class edge between the two related entities. The directed first-class edge is determined by the causal relationship or drug response relationship, and based on the created parent node and the first-class edge, construct an initial tumor knowledge graph database.

[0067] Furthermore, based on the extracted related entities and associations, an initial tumor knowledge graph database is constructed. Each gene, gene mutation, drug, and disease type is regarded as a node, and the causal relationship or drug response relationship between nodes (such as "mutation-drug relationship: Sensitive / Resistant", "drug-disease relationship: RELATED") constitutes the edges in the initial tumor knowledge graph database.

[0068] The graph hierarchy and classification submodule is used to create child nodes under the parent node in the constructed initial tumor knowledge graph database according to the classification attribute relationship of the drug or the subtype attribute relationship of the disease, and to create two types of edges between the parent node and the child node according to the classification attribute relationship of the drug or the subtype attribute relationship of the disease between the parent node and the child node. Based on the created child nodes and two types of edges, a hierarchical tumor knowledge graph database is constructed.

[0069] Furthermore, the present application can also construct a hierarchical tumor knowledge graph database through the classification attribute relationship of drugs or the subtype attribute relationship of diseases, so as to support multi-level and multi-dimensional complex queries. For example, lung cancer, as a parent node, can be divided into multiple subtypes, and each subtype can be further subdivided into gene mutations and corresponding drug treatments.

[0070] The data fusion and expansion submodule is used to store the attribute information of related entities obtained by the biomedical database acquisition submodule into the hierarchical tumor knowledge graph database.

[0071] Furthermore, after constructing the hierarchical tumor knowledge graph database, the present application automatically supplements the attribute information of related entities obtained from the biomedical database (i.e., from the NCBI biomedical database and the DrugBank biomedical database) into the hierarchical tumor knowledge graph database to enrich the constructed graph hierarchy.

[0072] The knowledge graph visualization submodule is used to visualize and query the hierarchical tumor knowledge graph database.

[0073] Furthermore, the present application also provides a visual interactive graph interface, through which users can interact with the hierarchical tumor knowledge graph database, and click on a node to view its detailed information (such as attribute information, target text information support, etc.). At the same time, they can also explore the complex association network of genes, diseases and drugs through the hierarchical tumor knowledge graph database.

[0074] The clinical decision support module 350 is used to generate a personalized clinical report of recommended treatment plans based on the clinical data of the target patient and the constructed hierarchical tumor knowledge graph database.

[0075] According to one embodiment of the present application, the clinical decision support module 350 includes a personalized clinical decision support submodule, a gene and drug matching submodule, and a clinical report generation submodule.

[0076] The personalized clinical decision support submodule is used to input the target patient's genome sequencing data and / or diagnostic information and / or clinical history data into the hierarchical tumor knowledge graph database for query processing, and output the first-class recommendation plan including gene mutations and drugs.

[0077] Furthermore, by inputting the target patient's genome sequencing data and / or diagnostic information and / or clinical history data and querying relevant information in the hierarchical tumor knowledge graph database, it is possible to automatically match related gene mutations and recommended drugs, that is, output a first-category recommendation plan that includes gene mutations and drugs.

[0078] The gene and drug matching submodule is used to input the gene mutation information of the target patient into the hierarchical tumor knowledge graph database for query processing, output the second category recommendation scheme containing the drug, and obtain the reliability evaluation result of the drug based on the evidence level based on the target text information corresponding to the drug in the second category recommendation scheme.

[0079] Furthermore, by inputting the target patient's gene mutation information (such as KRAS G12C mutation), retrieve the efficacy data of related drugs from the hierarchical tumor knowledge graph database, and provide specific recommended drugs, that is, output the second category recommendation scheme containing drugs. At the same time, the reliability of the recommended drugs can be evaluated based on the recommended drugs and the evidence level of the target text information to which the drugs belong, and the reliability evaluation results of the drugs can be obtained. The evidence level of the target text information to which the drugs belong is determined by grading the reliability of the clinical research and evidence of the drugs in the target text information; at the same time, it is similar to the AD-level evidence in the genetic test report, and the evidence level of the drugs is divided into four levels from high to low: A, B, C, and D. Among them, A-level evidence means that the drug is approved for marketing by the FDA or comes from professional clinical guidelines (the professional clinical guidelines can be obtained from the ASCO website, ESMO website, CSCO website, FDA website, EMA website, and NMPA website); B-level evidence means that it is confirmed based on large-scale clinical studies and has been agreed upon by clinical experts; C-level evidence means that the drug is used as A-level evidence in other cancers, or as a screening inclusion criterion in clinical trials, or supported by multiple small studies; D-level evidence means that it is based on preclinical studies, case reports or other lower-level evidence.

[0080] The clinical report generation submodule is used to input the clinical data of the target patient into the hierarchical tumor knowledge graph database for query processing, and output a personalized clinical report on recommended treatment plans. The clinical report on recommended treatment plans includes information on gene mutations, recommended treatment plans related to gene mutations, the level of evidence for the recommended treatment plans, recommended drugs, and target text information to which the recommended drugs belong. The recommended treatment plans are generated based on related entities including genes, gene mutations, disease types, and drugs, and the associations between related entities; the level of evidence for the recommended treatment plans is determined based on treatment plans approved on the FDA website or the NMPA website or in the clinical research stage.

[0081] Furthermore, the present application can also generate a personalized clinical report on recommended treatment plans based on the personalized clinical data of the target patients. The report content includes but is not limited to information on gene mutations, recommended treatment plans related to gene mutations, the level of evidence for the recommended treatment plans, recommended drugs and target text information to which the recommended drugs belong, etc. It may also include information on the mechanism of action, indications, contraindications, etc. of the recommended drugs.

[0082] The knowledge base updating and maintenance module 360 ​​is used to regularly collect the latest data related to tumor diseases according to a preset retrieval cycle, automatically extract new related entities and the association relationship between new related entities, and update the hierarchical tumor knowledge graph database.

[0083] According to an embodiment of the present application, the knowledge base update and maintenance module 360 ​​includes: a knowledge base dynamic update submodule and a database version control submodule.

[0084] The knowledge base dynamic update submodule is used to regularly collect the latest data related to tumor diseases according to a preset retrieval cycle, automatically extract new related entities and the association relationships between new related entities, and update the hierarchical tumor knowledge graph database.

[0085] Furthermore, the present application can also regularly monitor the release of new literature or tumor conference papers or clinical guidelines. After the release of new literature or tumor conference papers or clinical guidelines, the new related entities and the association relationship between the new related entities can be automatically extracted, and the relevant nodes and edges in the hierarchical tumor knowledge graph database can be updated. For example, the results of new drug clinical trials may cause changes in the relationship between certain drugs and gene mutations, and the system will automatically update these association information.

[0086] The database version control submodule is used to perform version management on the hierarchical tumor knowledge graph database.

[0087] Furthermore, the present application may also retain historical versions of the hierarchical tumor knowledge graph database so that users can trace the hierarchical tumor knowledge graph database and corresponding decision support information at a certain point in time, thereby ensuring the repeatability of clinical research.

[0088] According to an embodiment of the present disclosure, any multiple modules of the data acquisition module 310, the prompt word generation and optimization module 320, the entity recognition and relationship extraction module 330, the knowledge graph construction module 340, the clinical decision support module 350 and the knowledge base update and maintenance module 360 ​​can be combined into one module for implementation, or any one of the modules can be split into multiple modules. Alternatively, at least part of the functions of one or more of these modules can be combined with at least part of the functions of other modules and implemented in one module. According to an embodiment of the present disclosure, at least one of the data acquisition module 310, the prompt word generation and optimization module 320, the entity recognition and relationship extraction module 330, the knowledge graph construction module 340, the clinical decision support module 350 and the knowledge base update and maintenance module 360 ​​can be at least partially implemented as a hardware circuit, such as a field programmable gate array (FPGA), a programmable logic array (PLA), a system on a chip, a system on a substrate, a system on a package, an application-specific integrated circuit (ASIC), or can be implemented by hardware or firmware such as any other reasonable way of integrating or packaging the circuit, or by any one of the three implementation methods of software, hardware and firmware or by an appropriate combination of any of them. Alternatively, at least one of the data acquisition module 310, the prompt word generation and optimization module 320, the entity recognition and relationship extraction module 330, the knowledge graph construction module 340, the clinical decision support module 350 and the knowledge base update and maintenance module 360 ​​can be at least partially implemented as a computer program module, and when the computer program module is run, the corresponding function can be performed.

[0089] The electronic device according to an embodiment of the present application includes a processor, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) or a program loaded from a storage portion into a random access memory (RAM). The processor may include, for example, a general-purpose microprocessor (e.g., a CPU), an instruction set processor and / or a related chipset and / or a dedicated microprocessor (e.g., an application-specific integrated circuit (ASIC)), etc. The processor may also include an onboard memory for caching purposes. The processor may include a single processing unit or multiple processing units for performing different actions of the method flow according to an embodiment of the present application.

[0090] In the RAM, various programs and data required for the operation of the electronic device are stored. The processor, ROM, and RAM are connected to each other through a bus. The processor performs various operations of the method flow according to the embodiment of the present application by executing the program in the ROM and / or RAM. It should be noted that the program can also be stored in one or more memories other than ROM and RAM. The processor can also perform various operations of the method flow according to the embodiment of the present application by executing the program stored in the one or more memories.

[0091] According to an embodiment of the present application, the electronic device may further include an input / output (I / O) interface, which is also connected to the bus. The electronic device may also include one or more of the following components connected to the I / O interface: an input part including a keyboard, a mouse, etc.; an output part including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker; a storage part including a hard disk, etc.; and a communication part including a network interface card such as a LAN card, a modem, etc. The communication part performs communication processing via a network such as the Internet. The drive is also connected to the I / O interface as needed. Removable media, such as magnetic disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on the drive as needed so that the computer program read therefrom is installed into the storage part as needed.

[0092] The present application also provides a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or may exist independently without being assembled into the device / apparatus / system. The above computer-readable storage medium carries one or more programs, and when the above one or more programs are executed, the method according to the embodiment of the present application is implemented.

[0093] According to an embodiment of the present application, a computer-readable storage medium may be a non-volatile computer-readable storage medium, for example, it may include but is not limited to: a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof. In the present application, a computer-readable storage medium may be any tangible medium containing or storing a program, which may be used by or in combination with an instruction execution system, an apparatus or a device. For example, according to an embodiment of the present application, a computer-readable storage medium may include the ROM and / or RAM described above and / or one or more memories other than ROM and RAM.

[0094] The embodiment of the present application also includes a computer program product, which includes a computer program, and the computer program contains program code for executing the method shown in the flowchart. When the computer program product is run in a computer system, the program code is used to enable the computer system to implement the item recommendation method provided in the embodiment of the present application.

[0095] When the computer program is executed by the processor, the above functions defined in the system / device of the embodiment of the present application are performed. According to the embodiment of the present application, the system, device, module, unit, etc. described above can be implemented by a computer program module.

[0096] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device, a magnetic storage device, etc. In another embodiment, the computer program may also be transmitted and distributed in the form of a signal on a network medium, and downloaded and installed through a communication part, and / or installed from a removable medium. The program code contained in the computer program may be transmitted using any appropriate network medium, including but not limited to: wireless, wired, etc., or any suitable combination of the above.

[0097] In such an embodiment, the computer program can be downloaded and installed from the network through the communication part, and / or installed from a removable medium. When the computer program is executed by the processor, the above-mentioned functions defined in the system of the embodiment of the present application are executed. According to the embodiment of the present application, the system, equipment, device, module, unit, etc. described above can be implemented by a computer program module.

[0098] According to an embodiment of the present application, the program code for executing the computer program provided by the embodiment of the present disclosure can be written in any combination of one or more programming languages. Specifically, these computing programs can be implemented using high-level process and / or object-oriented programming languages, and / or assembly / machine languages. Programming languages ​​include, but are not limited to, Java, C++, python, "C" language or similar programming languages. The program code can be executed entirely on the user computing device, partially on the user device, partially on the remote computing device, or entirely on the remote computing device or server. In the case of a remote computing device, the remote computing device can be connected to the user computing device through any type of network, including a local area network (LAN) or a wide area network (WAN), or can be connected to an external computing device (e.g., using an Internet service provider to connect through the Internet).

[0099] Further references Figure 3 , as a response to the above Figure 2 The present application provides an embodiment of a method for constructing a tumor knowledge base based on a large model and prompt word optimization. Figure 2 The system embodiment shown corresponds to the embodiment shown.

[0100] like Figure 3 As shown, the method for constructing a tumor knowledge base based on a large model and prompt word optimization described in this embodiment includes the following steps.

[0101] S1. Obtain data related to tumor diseases from multiple data sources through a data acquisition module.

[0102] S2. The prompt words initially designed and continuously optimized by the prompt word generation and optimization module interact with the large language model LLM.

[0103] S3. Through the entity recognition and relationship extraction module, the large language model LLM is used to extract relevant entities and the association relationships between related entities from data related to tumor diseases.

[0104] S4. A hierarchical tumor knowledge graph database is constructed through the knowledge graph construction module according to the related entities and the association relationships between the related entities.

[0105] S5. Generate a personalized clinical report on recommended treatment plans based on the clinical data of the target patient and the constructed hierarchical tumor knowledge graph database through the clinical decision support module.

[0106] S6. Through the knowledge base update and maintenance module, the latest data related to tumor diseases are regularly collected according to the preset retrieval cycle, new related entities and the association relationship between new related entities are automatically extracted, and the hierarchical tumor knowledge graph database is updated.

[0107] Obviously, the embodiments described above are only some embodiments of the present application, rather than all embodiments. The preferred embodiments of the present application are given in the accompanying drawings, but they do not limit the patent scope of the present application. The present application can be implemented in many different forms. On the contrary, the purpose of providing these embodiments is to make the understanding of the disclosure of the present application more thorough and comprehensive. Although the present application is described in detail with reference to the aforementioned embodiments, for those skilled in the art, it is still possible to modify the technical solutions recorded in the aforementioned specific implementation methods, or to perform equivalent replacement of some of the technical features therein. Any equivalent structure made using the contents of the specification and drawings of this application, directly or indirectly used in other related technical fields, is similarly within the scope of patent protection of this application.

Claims

1. A tumor knowledge base construction system based on a large model and optimized prompt words, characterized in that: include: A data acquisition module, used for acquiring data related to tumor diseases from multiple data sources; The prompt word generation and optimization module is used to interact with the large language model LLM through the prompt words that are initially designed and continuously optimized; The entity recognition and relationship extraction module is used to extract relevant entities and the relationship between relevant entities from data related to tumor diseases using a large language model (LLM); A knowledge graph construction module is used to construct a hierarchical tumor knowledge graph database based on related entities and the association relationships between related entities; The clinical decision support module is used to generate personalized clinical reports on recommended treatment plans based on the clinical data of target patients and the constructed hierarchical tumor knowledge graph database; The knowledge base update and maintenance module is used to regularly collect the latest data related to tumor diseases according to the preset retrieval cycle, automatically extract new related entities and the association relationship between new related entities, and update the hierarchical tumor knowledge graph database; The prompt word generation and optimization module includes: A prompt word generation submodule is used to obtain a target processing task, identify the prompt word category to which the target processing task belongs according to a custom prompt word category, and generate an Mth prompt word based on the information range of the corresponding prompt word category, where M is an integer greater than 0; the prompt word categories include tumor targeted medication category, tumor immunotherapy category, tumor diagnosis category, and tumor prognosis category; The large language model calling submodule is used to input the Mth prompt word into the large language model LLM and output the Mth model generation result; The output result evaluation submodule is used to compare the M-th model generation result with the expected answer result in the standard data set based on the preset evaluation criteria to generate the M-th output model evaluation result; A feedback mechanism submodule is used to generate the Mth feedback information according to the Mth output model evaluation result; The feedback and prompt word optimization submodule is used to optimize the Mth prompt word according to the Mth feedback information to obtain the M+1th prompt word; The feedback and prompt word optimization submodule also includes: The multi-round interactive optimization submodule is used to input the M+1th prompt word into the large language model LLM, output the M+1th prompt word model generation result, compare the M+1th prompt word model generation result with the expected answer result in the standard data set based on the preset evaluation criteria, generate the M+1th output model evaluation result, generate the M+1th feedback information based on the M+1th output model evaluation result, optimize the M+1th prompt word based on the M+1th feedback information, and obtain the M+2th prompt word.

2. The tumor knowledge base construction system based on large model and prompt word optimization according to claim 1, characterized in that: The data acquisition module comprises: The literature data collection submodule is used to obtain original literature data related to tumor diseases from PubMed database, Google Scholar database, Web of Science database, Scopus database, and CNKI database; The clinical guideline collection submodule is used to obtain original tumor conference paper data or original clinical guideline data related to tumor diseases from the ASCO website, ESMO website, CSCO website, FDA website, EMA website, and NMPA website; The biomedical database acquisition submodule is used to obtain attribute information of related entities from the NCBI biomedical database and the DrugBank biomedical database.

3. The tumor knowledge base construction system based on large model and prompt word optimization according to claim 2, characterized in that: The data acquisition module also includes: The data preprocessing submodule is used to perform data deduplication processing, data format unification processing and redundant data elimination processing on the acquired original literature data, original tumor conference paper data and original clinical guideline data to obtain target text information, wherein the redundant data elimination processing is performed on the acquired original literature data, original tumor conference paper data and original clinical guideline data, including: If N original literature data and / or original tumor conference paper data and / or original clinical guideline data all correspond to the same conclusion, N is an integer greater than a preset value, according to a preset priority screening rule, the priority screening rule includes at least one of the number of citations, experimental design type, sample size and journal impact factor arranged in descending order of priority, and the experimental design type includes at least one of randomized controlled trial research, cohort study, case control study, case series study, case study, animal experimental study, cell line experimental study and experimental deviation study arranged in descending order of priority, the highest priority literature data or tumor conference paper data or clinical guideline data are screened out from the N original literature data and / or original tumor conference paper data and / or original clinical guideline data in order of priority from high to low.

4. The tumor knowledge base construction system based on large model and prompt word optimization according to claim 1, characterized in that: The entity recognition and relationship extraction module includes: The entity recognition submodule is used to parse the target text information using the large language model LLM according to the optimized prompt words, and identify the relevant entities contained in the target text information, wherein the relevant entities include genes, gene mutations, disease types and drugs; The relationship extraction submodule is used to extract the association relationship between related entities according to the semantic relationship in the target text information, and the association relationship is a causal relationship or a drug reaction relationship.

5. The tumor knowledge base construction system based on large model and prompt word optimization according to claim 4, characterized in that: The knowledge graph construction module includes: A node and edge construction submodule, which is used to create each related entity as a parent node, determine whether there is an association relationship between two related entities, and if there is an association relationship between the two related entities, create a directed first-class edge between the two related entities, wherein the directed first-class edge is determined by the causal relationship or drug response relationship, and construct an initial tumor knowledge graph database based on the created parent node and the first-class edge; The graph hierarchy and classification submodule is used to create child nodes under the parent node according to the classification attribute relationship of the drug or the subtype attribute relationship of the disease in the constructed initial tumor knowledge graph database, and to create a second-class edge between the parent node and the child node according to the classification attribute relationship of the drug or the subtype attribute relationship of the disease between the parent node and the child node, and to construct a hierarchical tumor knowledge graph database based on the created child nodes and second-class edges; The data fusion and expansion submodule is used to store the attribute information of the related entities obtained by the biomedical database acquisition submodule into the hierarchical tumor knowledge graph database; The knowledge graph visualization submodule is used to visualize and query the hierarchical tumor knowledge graph database.

6. The tumor knowledge base construction system based on large model and prompt word optimization according to claim 5, characterized in that: The clinical decision support module includes: A personalized clinical decision support submodule is used to input the target patient's genome sequencing data and / or diagnostic information and / or clinical history data into the hierarchical tumor knowledge graph database for query processing, and output a first-class recommendation plan including gene mutations and drugs; The gene and drug matching submodule is used to input the gene mutation information of the target patient into the hierarchical tumor knowledge graph database for query processing, output the second category recommendation scheme containing the drug, and obtain the reliability evaluation result of the drug based on the evidence level based on the target text information corresponding to the drug in the second category recommendation scheme; The clinical report generation submodule is used to input the clinical data of the target patient into the hierarchical tumor knowledge graph database for query processing, and output a personalized clinical report on recommended treatment plans. The clinical report on recommended treatment plans includes information on gene mutations, recommended treatment plans related to gene mutations, the level of evidence for the recommended treatment plans, recommended drugs, and target text information to which the recommended drugs belong. The recommended treatment plans are generated based on related entities including genes, gene mutations, disease types, and drugs, and the associations between related entities; the level of evidence for the recommended treatment plans is determined based on treatment plans approved on the FDA website or the NMPA website or in the clinical research stage.

7. The tumor knowledge base construction system based on large model and prompt word optimization according to claim 6, characterized in that: The knowledge base update and maintenance module includes: The knowledge base dynamic update submodule is used to regularly collect the latest data related to tumor diseases according to the preset retrieval cycle, automatically extract new related entities and the association relationship between new related entities, and update the hierarchical tumor knowledge graph database; The database version control submodule is used to perform version management on the hierarchical tumor knowledge graph database.

8. A method for constructing a tumor knowledge base based on a large model and optimized prompt words, characterized in that: include: Acquire data related to tumor diseases from multiple data sources through a data acquisition module; The prompt words initially designed and continuously optimized by the prompt word generation and optimization module interact with the large language model LLM; Through the entity recognition and relationship extraction module, the large language model LLM is used to extract relevant entities and the association relationships between related entities from the data related to tumor diseases; Through the knowledge graph construction module, a hierarchical tumor knowledge graph database is constructed according to the related entities and the association relationships between related entities; The clinical decision support module generates personalized clinical reports on recommended treatment plans based on the clinical data of target patients and the constructed hierarchical tumor knowledge graph database; Through the knowledge base update and maintenance module, the latest data related to tumor diseases are regularly collected according to the preset retrieval cycle, new related entities and the association relationship between new related entities are automatically extracted, and the hierarchical tumor knowledge graph database is updated; The prompt word generation and optimization module includes: A prompt word generation submodule is used to obtain a target processing task, identify the prompt word category to which the target processing task belongs according to a custom prompt word category, and generate an Mth prompt word based on the information range of the corresponding prompt word category, where M is an integer greater than 0; the prompt word categories include tumor targeted medication category, tumor immunotherapy category, tumor diagnosis category, and tumor prognosis category; The large language model calling submodule is used to input the Mth prompt word into the large language model LLM and output the Mth model generation result; The output result evaluation submodule is used to compare the M-th model generation result with the expected answer result in the standard data set based on the preset evaluation criteria to generate the M-th output model evaluation result; A feedback mechanism submodule is used to generate the Mth feedback information according to the Mth output model evaluation result; The feedback and prompt word optimization submodule is used to optimize the Mth prompt word according to the Mth feedback information to obtain the M+1th prompt word; The feedback and prompt word optimization submodule also includes: The multi-round interactive optimization submodule is used to input the M+1th prompt word into the large language model LLM, output the M+1th prompt word model generation result, compare the M+1th prompt word model generation result with the expected answer result in the standard data set based on the preset evaluation criteria, generate the M+1th output model evaluation result, generate the M+1th feedback information based on the M+1th output model evaluation result, optimize the M+1th prompt word based on the M+1th feedback information, and obtain the M+2th prompt word.

Images