Gene report interpretation system based on large language model

Abstract

The invention discloses a gene report interpretation system based on a large language model, relates to a gene report interpretation system, and belongs to the technical field of computers. The objective of the invention is to solve the problems of information inaccuracy, insufficient correlation or information imagination and the like possibly occurring in the interpretation process of the existing gene report interpretation method based on a large language model. The system comprises a user interface initialization module used for setting a user interface as a message input box, a file uploading option, a dialogue clearing option and an interface display style option; the large language model initialization module is used for selecting a universal open source large language model; the gene detection report processing module is used for analyzing the content of a PDF file uploaded by a user to obtain a character string; questioning the character string, and carrying out standardization processing on the questions to obtain the questions after standardization processing; the question reply module is used for replying the questions after standardization processing; and the interface display module is used for processing the reply to obtain background information and presenting the background information on a user interface.

Description

Technical Field

[0001] The invention relates to a gene report interpretation system and belongs to the technical field of computers. Background Art

[0002] In recent years, the development of genomics has brought unprecedented opportunities to the medical field, especially in personalized medicine and precision treatment. With the rapid popularization of gene sequencing technology, doctors and patients can obtain a large amount of genomic data. However, how to effectively interpret and use these complex gene reports remains an urgent problem to be solved. Traditional gene reports often contain a lot of professional terms and complex data, which are difficult for ordinary patients and non-professionals to understand, thus affecting patients' cognition and decision-making about their own health status.

[0003] The rapid development of large language models (LLMs) in the field of natural language processing has provided new solutions for the interpretation of genetic reports. These models, especially those based on the Transformer architecture, are able to extract rich language features and knowledge from large amounts of text data by pre-training. Therefore, systems based on large language models have the potential to convert complex genetic reports into easy-to-understand language, improving patients' understanding and medical decision-making capabilities.

[0004] However, current gene report interpretation methods based on large language models still face some challenges, such as inaccurate information, insufficient relevance, or fictitious information that may occur during the interpretation process. These problems may arise from the model's lack of knowledge in a specific medical field or the bias of the training data set. Therefore, in medical applications, it is crucial to ensure the accuracy and credibility of the generated information, especially in key areas that affect patient health decisions. Summary of the invention

[0005] The purpose of the present invention is to solve the problems of inaccurate information, insufficient relevance or fictitious information that may occur in the interpretation process of the existing gene report interpretation method based on a large language model, and to propose a gene report interpretation system based on a large language model.

[0006] The gene report interpretation system based on the large language model includes:

[0007] User interface initialization module, large language model initialization module, gene test report processing module, question response module, interface display module;

[0008] The user interface initialization module is used to set the user interface to a message input box, a file upload option, a clear dialogue option, and an interface display style option;

[0009] The large language model initialization module is used to select an open source large language model in a general field;

[0010] The gene detection report processing module is used to parse the content of the PDF file uploaded by the user to obtain a character string; ask questions to the character string, perform standardization on the questions, and obtain standardized questions;

[0011] The question answering module is used to input the standardized questions into the large language model, and answer the standardized questions as needed or without the need for additional information sources;

[0012] The interface display module is used to process the reply to obtain background information, put the background information in a floating window, and present it on the user interface.

[0013] The beneficial effects of the present invention are:

[0014] The present invention significantly improves the intelligent level of gene report interpretation by introducing large language model technology and multi-source knowledge fusion. First, the system combines gene knowledge base and network resources, and can provide richer and more comprehensive gene interpretation information through the accurate analysis of user input by large model. Users can not only obtain report interpretation based on existing gene knowledge, but also obtain the latest scientific research results and clinical research dynamics through network enhancement module to ensure the timeliness and scientificity of interpretation results. In addition, the system uses context injection technology to deeply understand the context of user input, effectively eliminate ambiguity, and make the interpretation results more targeted and personalized. This interpretation system has particularly significant advantages in pharmacogenomics, genetic susceptibility analysis, etc., which can help users quickly understand complex gene mutation information and provide personalized medication advice, avoiding the problem that the information in traditional gene report interpretation is too professional or obscure. Users can obtain accurate and easy-to-understand answers through concise input, which improves the efficiency and accuracy of report interpretation. In addition, the system has strong scalability, and can continuously optimize the interpretation effect with the continuous updating of knowledge base and network resources, providing users with cutting-edge and authoritative gene interpretation services. This paper solves the problems of inaccurate information, insufficient relevance or fictitious information that may occur in the interpretation process of existing gene report interpretation methods based on large language models. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0016] Specific implementation method 1: The gene report interpretation system based on the large language model in this implementation method includes:

[0017] User interface initialization module, large language model initialization module, gene test report processing module, question response module, interface display module;

[0018] The user interface initialization module is used to set the user interface to a message input box, a file upload option, a clear dialogue option, and an interface display style option;

[0019] The large language model initialization module is used to select an open source large language model in a general field;

[0020] The gene detection report processing module is used to parse the content of the PDF file uploaded by the user to obtain a character string; ask questions to the character string, perform standardization on the questions, and obtain standardized questions;

[0021] The question answering module is used to input the standardized questions into the large language model, and answer the standardized questions as needed or without the need for additional information sources;

[0022] The interface display module is used to process the reply to obtain background information, put the background information in a floating window, and present it on the user interface.

[0023] Specific implementation method 2: This implementation method is different from specific implementation method 1 in that: the message input box is used to receive questions from users;

[0024] The file upload option is used to support users to upload genetic test reports in PDF format;

[0025] The clear conversation option is used to help users clean up previous conversation records;

[0026] The interface display style option is used for the user to switch between different theme styles (such as light mode or dark mode).

[0027] The other steps and parameters are the same as those in the first embodiment.

[0028] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that: the large language model initialization module is used to select an open source large language model in a general field; the specific process is:

[0029] The open-source large language models in general fields are Wenxin Yiyan, Tongyi Qianwen or Harbin Institute of Technology movable type large model;

[0030] The Wenxin Yiyan, Tongyi Qianwen or Harbin Institute of Technology movable type large model is run on the local server;

[0031] The Wenxin Yiyan, Tongyi Qianwen or Harbin Institute of Technology movable type large model uses the local port to monitor and receive all input information of the user.

[0032] The other steps and parameters are the same as those in the first or second embodiment.

[0033] Specific implementation method 4: This implementation method is different from the specific implementation methods 1 to 3 in that: the gene detection report processing module is used to parse the content of the PDF file uploaded by the user to obtain a string; ask questions to the stored string, perform standardization on the questions, and obtain standardized questions;

[0034] The specific process is:

[0035] Call the PyMuPDF library to extract all the text information in the PDF uploaded by the user, convert all the text information in the PDF into strings, and segment the strings according to the page numbers of the PDF. The content of each page is stored separately in buffer 1;

[0036] A question is asked to the character string stored in buffer 1, and the question is standardized to obtain a standardized question.

[0037] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0038] Specific implementation method 5: This implementation method is different from the specific implementation methods 1 to 4 in that the process of the standardization process is:

[0039] The questions are input into the big prediction model for standardized processing to obtain the standardized questions.

[0040] The other steps and parameters are the same as those in Specific Implementation 1 to 4-1.

[0041] Specific implementation method 6: This implementation method is different from the specific implementation methods 1 to 5 in that: the question reply module is used to input the standardized questions into the large language model, and reply to the standardized questions as needed or without the need for additional information sources;

[0042] The specific process is:

[0043] If no additional information sources are needed, respond directly to the standardized questions;

[0044] If additional sources of information are needed, they are classified into the following situations:

[0045] Case 1: Extract keywords from the standardized questions based on the large language model, search for keywords through the Bing search engine, and return the top 10 search results;

[0046] Use a crawler tool (such as Requests or Selenium) to crawl the first 10 search results returned and extract the web page content;

[0047] The standardized question and the extracted webpage content are input into the large language model, and the output content of the large language model is stored separately in the buffer 2;

[0048] Case 2: Extract keywords from standardized questions based on a large language model;

[0049] Based on the string matching algorithm, the content matching the keywords is extracted from the local gene knowledge base as the first batch of search results;

[0050] Based on the vector retrieval model, the extracted keywords are converted into vector representations as the second batch of retrieval results;

[0051] Input the first batch of search results and the second batch of search results into the large language model, and store the output content of the large language model separately in buffer 3;

[0052] Case 3: Extract keywords from standardized questions based on a large language model;

[0053] Extract the content matching the keyword from the string in buffer 1 based on the string matching algorithm and store it separately in buffer 4;

[0054] Based on the vector retrieval model, the contents in buffers 2, 3, and 4 are vectorized to obtain vectorized results; similarity is used to calculate the similarity of any two vectorized strings in buffers 2, 3, and 4. If the similarity of the two strings is greater than a threshold, one of the two strings is retained; if the similarity of the two strings is less than or equal to the threshold, both strings are retained;

[0055] The retained string is input into the big prediction model, and the big language model outputs the response to the corresponding question.

[0056] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 5-1.

[0057] Specific implementation method 7: This implementation method is different from any one of specific implementation methods 1 to 6 in that: the interface display module is used to process the reply to obtain background information, and put the background information in a floating window to present it on the user interface; the specific process is:

[0058] Sentence the generated response and split the entire answer into multiple independent sentences;

[0059] Vectorize each sentence using a vector retrieval model;

[0060] The similarity between each vectorized sentence and any string in the strings retained in buffers 2, 3, and 4 is calculated by similarity, and the string with the highest similarity is selected as the background information of each sentence;

[0061] Put the background information in a floating window and present it on the user interface.

[0062] The other steps and parameters are the same as those in Specific Implementations 1 to 6-1.

[0063] Specific implementation method 8: This implementation method is different from the specific implementation methods 1 to 7 in that: the generated reply is processed into sentences; the specific process is:

[0064] The generated response is processed using the nltk toolkit, splitting the entire answer into multiple independent sentences.

[0065] The other steps and parameters are the same as those in Specific Implementation 1 to 7-1.

[0066] Specific implementation method 9: This implementation method is different from any one of specific implementation methods 1 to 8 in that: the vector retrieval model is a BGE-v1 vector retrieval model.

[0067] The other steps and parameters are the same as those in Specific Embodiments 1 to 8.

[0068] Specific implementation method ten: This implementation method is different from any one of specific implementation methods one to nine in that: the similarity is cosine similarity.

[0069] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 9.

[0070] The following examples are used to verify the beneficial effects of the present invention:

[0071] Embodiment 1:

[0072] Step 1: When the system is initialized, the user interface is designed as an intuitive, user-friendly interactive platform to provide users with a clear operating experience. The interface contains multiple interactive elements: the message input box is used to receive user text input, the file upload option supports users to upload genetic testing reports in PDF format, the clear conversation option helps users clean up previous conversation records, and the interface display style option allows users to switch between different themes (such as light mode or dark mode). These elements ensure that users can submit questions and reports in a convenient way, and the system can efficiently receive and process user input.

[0073] At the same time, the system background will initialize a large language model, which can be the Wenxin Yiyan, Tongyi Qianwen or Huozi large language model selected by the user according to their needs. These models can run on the local server, using the local port to listen to and receive all user input information. The server's operating environment must be adjusted according to the needs of the model, such as computing resources, memory allocation, etc., to ensure the efficient operation of the model. In addition, the system will also start the BGE-v1 vector retrieval model, which is specifically used for vector retrieval of genetic knowledge bases, user reports, and user questions. It uses vectorized representation to better match user input with the content of the knowledge base.

[0074] Step 2: When the user uploads a genetic test report in PDF format, the system will call the PyMuPDF library to parse the content of the PDF file. PyMuPDF can quickly and accurately extract all text information in PDF. The system first converts the PDF file into plain text paragraphs. The extracted text will be segmented according to the page numbers of the PDF, and the content of each page will be stored separately in the system. This process is very important because genetic test reports usually contain a lot of detailed information, such as gene mutation descriptions, drug reactions, genetic risks, etc. By storing each page independently, it can ensure that the system can accurately locate the relevant text paragraphs during subsequent retrieval. At the same time, the parsed content will be marked with page numbers and paragraph positions, which is particularly critical for the subsequent generation of report fragments related to user questions. Through this structured storage method, the system can also process and retrieve relevant information with high efficiency when faced with complex reports.

[0075] Step 3: When the user asks a question, the system will first send the user's question to the large language model for processing. The large language model will perform a series of natural language processing tasks, the first of which is to standardize the user's question and convert it into a language expression in a unified format. This process includes operations such as word segmentation, part-of-speech tagging, and semantic analysis. Standardized questions are easier for the model to understand and ensure that subsequent operations can accurately capture the user's core intent. On this basis, the large language model will also perform intent analysis on the user's question, extract important keywords, and further infer whether the user's needs are for the explanation of gene mutations, the prediction of drug reactions, or the analysis of genetic diseases. This analysis step of the large language model will also determine whether additional information sources need to be called. If the system detects that the user's question content can be directly answered through the model's internal knowledge base, the system will skip the call of the external information source and directly rely on the model's internal knowledge to generate answers.

[0076] Step 4: If the system determines after analysis that it needs to call an external information source, such as an Internet search engine or a local genetic knowledge base, it will proceed to the next step. First, when the system decides to call Internet resources, it will search through the Bing search engine. Based on the standardized user questions, the system will automatically construct a search query and return the top 10 search results. Subsequently, the system will crawl these web pages through crawler tools (such as Requests or Selenium), extract the web page content and perform a preliminary semantic analysis on it. These contents will be further sent to the large language model. The model will combine the user's questions and the content extracted from the web page, select the most relevant parts for formatting, and return the final content to the user. Throughout the process, the system not only simply obtains the web page content, but also performs structured processing on it, such as removing redundant information, extracting core paragraphs, etc., to ensure that the content returned to the user is concise and useful.

[0077] For calls to the local genetic knowledge base, the system will first search for relevant content in the knowledge base through a keyword matching algorithm to obtain the first batch of search results. Then, the BGE-v1 vector retrieval model will be started to further search the content in the genetic knowledge base based on the vectorized representation to obtain the second batch of search results. These results will be merged together and sent to the large language model, which will generate answers related to the user's questions based on these results. In this process, the role of the vector retrieval model is extremely critical because it can efficiently filter out the most relevant content from a large amount of knowledge base data based on the semantic similarity of the content.

[0078] If the genetic test report uploaded by the user needs to be searched, the system will first search the report by keywords to obtain preliminary search results. At the same time, the system will also conduct an in-depth analysis of the entire report through the large language model, divide the report into 8192-byte context paragraphs, and send them to the large language model for semantic analysis one by one, returning more accurate search results. After merging the preliminary results with the results returned by the large language model, the system will organize the content and finally generate a complete answer.

[0079] Step 5. After all information source calls and retrieval operations are completed, the system enters a key processing link, which is to summarize and refine the results from different information sources. In this process, the system first needs to perform a comprehensive deduplication operation on all content obtained from the Internet, genetic knowledge bases, and genetic test reports uploaded by users. Since multiple information sources may provide similar answers or contain the same gene-related information, deduplication is not just a simple text matching, but also involves semantic analysis. The system will identify the repetition of content through vector similarity calculations to ensure that the information finally retained is accurate, non-repetitive and most relevant. This deduplication step can avoid information redundancy, so that users will not feel repetitive or confused when reading replies.

[0080] After removing duplicate content, the system will organize the information structure according to preset templates. The design of the template can be customized according to specific application scenarios, such as the interpretation of gene mutations, risk assessment of genetic diseases, personalized recommendations for drug responses, etc. The system will select appropriate templates based on different types of user input and integrate all information into a logically clear and well-organized output. These templates will divide the content into different modules according to user needs, such as "Gene Mutation Analysis", "Drug Response Prediction", "Genetic Risk Interpretation", etc. Each module will cover specific areas of concern to users. The template not only standardizes the output format of the information, but also improves the user's understanding and readability of the reply content.

[0081] Next, the system will send all the organized information to the large language model, requiring the model to conduct a comprehensive analysis based on knowledge from different sources. The large language model has powerful natural language processing and reasoning capabilities, and can generate a complete and detailed answer based on the integration of information from multiple sources. The model will evaluate the credibility and relevance of each information source, and give priority to citing highly authoritative and up-to-date resources to ensure the accuracy and scientificity of the answer. The large language model will also convert lengthy and complex technical terms into popular language that users can understand, so that the answer is not only detailed, but also clear and easy to understand. This comprehensive analysis stage can effectively improve the user experience, so that the final answer not only contains rich professional knowledge, but also intuitively answers the user's questions.

[0082] Step 6. After the large language model generates a complete answer, the system will not send it to the user immediately. Instead, the system will use the nltk toolkit to process the generated answer into sentences, splitting the entire answer into multiple independent sentences. The purpose of this step is to facilitate the system to further perform detailed semantic matching and association operations on each sentence. For each sentence, the system will use the BGE-v1 vector retrieval model to perform vector matching with each previously called information source (including genetic knowledge base, Internet content, and genetic test reports) in turn. Through vectorized retrieval, the system can find the most relevant background information for each sentence. This operation not only ensures the accuracy of the answer, but also provides users with a transparent reference basis. The vector matching results of each sentence will display the top 10 most matching references related to the sentence.

[0083] The matching results will not be displayed directly in the body of the answer, but will be presented in the user interface in the form of a floating window. Each sentence will be followed by a hyperlink, and users can click on the link to open the floating window to view detailed reference information matching the sentence. The information sources in the floating window can include web links, genetic knowledge base entries, or relevant paragraphs in genetic test reports. In this way, users can better trace the basis behind the answer and ensure the credibility and interpretability of the response content. In addition, this design also gives users a higher degree of interactive freedom. They can choose which information to view and deeply understand every detail related to their questions.

[0084] The final system output is an optimized answer with detailed background information. This answer not only answers the user's core question, but also provides a way for the user to further explore related information. By associating each sentence with the reference in the floating window, users can immediately obtain the evidence support related to each sentence when reading, so as to have a deeper understanding of the genetic test report and its potential medical and clinical impact. This design greatly improves the readability and professionalism of the information, and at the same time increases the user's trust in the answer through transparent reference information.

[0085] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A gene report interpretation system based on a large language model, characterized by: The system comprises: User interface initialization module, large language model initialization module, gene test report processing module, question response module, interface display module; The user interface initialization module is used to set the user interface to a message input box, a file upload option, a clear dialogue option, and an interface display style option; The large language model initialization module is used to select an open source large language model in a general field; The gene detection report processing module is used to parse the content of the PDF file uploaded by the user to obtain a character string; ask questions to the character string, perform standardization on the questions, and obtain standardized questions; The question answering module is used to input the standardized questions into the large language model, and answer the standardized questions as needed or without the need for additional information sources; The interface display module is used to process the reply to obtain background information, put the background information in a floating window, and present it on the user interface.

2. The gene report interpretation system based on a large language model according to claim 1, characterized in that: The message input box is used to receive questions from users; The file upload option is used to support users to upload genetic test reports in PDF format; The clear conversation option is used to help users clean up previous conversation records; The interface display style option is used for the user to switch between different theme styles.

3. The gene report interpretation system based on a large language model according to claim 2, characterized in that: The large language model initialization module is used to select an open source large language model in a general field; the specific process is: The open-source large language models in general fields are Wenxin Yiyan, Tongyi Qianwen or Harbin Institute of Technology movable type large model; The Wenxin Yiyan, Tongyi Qianwen or Harbin Institute of Technology movable type large model is run on the local server; The Wenxin Yiyan, Tongyi Qianwen or Harbin Institute of Technology movable type large model uses the local port to monitor and receive all input information of the user.

4. The gene report interpretation system based on a large language model according to claim 3, characterized in that: The gene detection report processing module is used to parse the content of the PDF file uploaded by the user to obtain a character string; ask questions to the stored character string, perform standardized processing on the questions, and obtain standardized questions; The specific process is: Call the PyMuPDF library to extract all the text information in the PDF uploaded by the user, convert all the text information in the PDF into strings, and segment the strings according to the page numbers of the PDF. The content of each page is stored separately in buffer 1; A question is asked to the character string stored in buffer 1, and the question is standardized to obtain a standardized question.

5. The gene report interpretation system based on a large language model according to claim 4, characterized in that: The process of the standardization process is as follows: The questions are input into the big prediction model for standardized processing to obtain the standardized questions.

6. The gene report interpretation system based on a large language model according to claim 5, characterized in that: The question answering module is used to input the standardized questions into the large language model, and answer the standardized questions as needed or without the need for additional information sources; the specific process is: If no additional information sources are needed, respond directly to the standardized questions; If additional sources of information are needed, they are classified into the following situations: Case 1: Extract keywords from standardized questions based on a large language model; Perform keyword search through Bing search engine and return the first 10 search results; Use crawler tools to crawl the first 10 search results returned and extract the web page content; The standardized question and the extracted webpage content are input into the large language model, and the output content of the large language model is stored separately in the buffer 2; Case 2: Extract keywords from standardized questions based on a large language model; Based on the string matching algorithm, the content matching the keywords is extracted from the local gene knowledge base as the first batch of search results; Based on the vector retrieval model, the extracted keywords are converted into vector representations as the second batch of retrieval results; Input the first batch of search results and the second batch of search results into the large language model, and store the output content of the large language model separately in buffer 3; Case 3: Extract keywords from standardized questions based on a large language model; Extract the content matching the keyword from the string in buffer 1 based on the string matching algorithm and store it separately in buffer 4; Based on the vector retrieval model, the contents in buffers 2, 3, and 4 are vectorized to obtain vectorized results; similarity is used to calculate the similarity of any two vectorized strings in buffers 2, 3, and 4. If the similarity of the two strings is greater than a threshold, one of the two strings is retained; if the similarity of the two strings is less than or equal to the threshold, both strings are retained; The retained string is input into the big prediction model, and the big language model outputs the response to the corresponding question.

7. The gene report interpretation system based on a large language model according to claim 6, characterized in that: The interface display module is used to process the reply to obtain background information, put the background information in a floating window, and present it on the user interface; the specific process is: Process the generated responses into sentences; Vectorize each sentence using a vector retrieval model; The similarity between each vectorized sentence and any string in the strings retained in buffers 2, 3, and 4 is calculated by similarity, and the string with the highest similarity is selected as the background information of each sentence; Put the background information in a floating window and present it on the user interface.

8. The gene report interpretation system based on a large language model according to claim 7, characterized in that: The generated reply is processed by sentence division; the specific process is: The generated responses are processed by using the nltk toolkit.

9. The gene report interpretation system based on a large language model according to claim 8, characterized in that: The vector retrieval model is the BGE-v1 vector retrieval model.

10. The gene report interpretation system based on a large language model according to claim 9, characterized in that: The similarity is cosine similarity.

Images