An intelligent medical literature query system based on an AI large model and a working method thereof

Abstract

The application discloses an intelligent medical literature query system based on an AI large model and a working method thereof, and belongs to the medical application field of artificial intelligence information technology. The system comprises a problem collection module, an intention recognition module, a large model semantic analysis module, a similarity sorting module and a medical literature acquisition module. The working method comprises the following steps: S1, medical problem collection and labeling; S2, intention recognition model building and training; S3, large language model semantic analysis; S4, medical literature relevance reordering; S5, medical literature acquisition. The application constructs an intelligent medical literature query system based on an AI large model and a working method thereof, the system can intelligently provide medical literature related to the input medical problem, reduces the labor cost, and promotes the automation and intelligentization of the medical field.

Description

Technical Field

[0001] This invention relates to the fields of computer science and technology, linguistics, and natural language processing, and particularly to the medical applications of artificial intelligence information technology. Specifically, it relates to an intelligent medical literature retrieval system based on a large AI model and its working method. Background Technology

[0002] The volume of medical literature is growing exponentially, with countless new medical papers, clinical trial reports, and other research findings published daily. This information deluge often leaves medical professionals struggling to quickly find accurate and effective information when conducting literature searches. Especially when faced with such a massive amount of literature, the lack of clear screening criteria and tools forces medical professionals to spend considerable time sifting through irrelevant or low-quality literature, hindering information retrieval efficiency. This not only impacts the work efficiency of medical professionals but can also delay clinical decisions and patient treatment processes, and in some cases, may even lead to incorrect medical judgments. Furthermore, the highly specialized and complex nature of the medical field often presents obstacles to collaboration and information sharing between different disciplines. This situation makes it difficult for many advanced research findings to flow and be applied across disciplines, and researchers often face difficulties such as disciplinary barriers and poor information flow when conducting interdisciplinary collaborations. Therefore, a flexible and intelligent medical literature retrieval solution is urgently needed. Summary of the Invention

[0003] To overcome the shortcomings of existing technologies, the present invention aims to provide an intelligent medical literature retrieval system based on an AI large-scale model and its working method. This system combines the latest large-scale model technology to extract keywords needed for medical literature retrieval, employs a vectorized local knowledge base to expand the large-scale model's professional knowledge in the medical field, integrates the API of a third-party literature retrieval system for literature searching, and applies a re-ranking model to improve the quality of the retrieved literature. To achieve the above objectives, the present invention provides the following technical solution, comprising the following steps:

[0004] Step S1: Collect and label medical questions.

[0005] Step S2: Build the intent recognition module, train the intent recognition model, and verify the accuracy after training is completed.

[0006] Step S3: Semantic Analysis Using a Large Language Model. The large language model is used to extract literature information from the user's question, such as title, DOI, and keywords. After extracting the literature information, a literature search API is used to perform a literature search. After completing the search, the large model is applied to filter relevant literature.

[0007] Step S4: Re-ranking of medical literature relevance. Construct a re-ranking model and re-rank the literature based on relevance.

[0008] Step S5: Medical Literature Retrieval. Users input a question and apply the modules from the previous steps to obtain the final literature results.

[0009] This invention presents an intelligent medical literature retrieval system based on an AI-powered large-scale model and its operating method. Based on the user's input questions, the system recommends relevant medical literature, avoiding the complexity of manual screening and improving the flexibility and convenience of learning within hospitals. Furthermore, this invention provides a new solution for medical literature retrieval, enabling cross-disciplinary flow of medical knowledge and promoting automation and intelligence in the medical field. Attached Figure Description

[0010] Figure 1 This is a flowchart illustrating the system operation in an embodiment of the present invention; Detailed Implementation

[0011] Some knowledge not explained in this specification belongs to prior art known to those skilled in the art and is easy to understand. The technical solutions in the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. This invention relates to an intelligent medical literature retrieval system based on an AI large model and its working method, specifically including the following steps:

[0012] Step S1: Collect and label medical questions.

[0013] Step S1-1: Collect and organize medical literature search questions, including online Q&A within hospitals and publicly available medical Q&A datasets. After organization, filter the data to ensure that each question has at least one corresponding document (i.e., each medical literature search question has a corresponding answer).

[0014] Step S1-2: Label the medical literature search questions that have been screened, and divide them into two categories according to the search intent: find literature titles and find related literature. After labeling, divide the training set and test set according to the 8:2 principle.

[0015] Step S2: Build the intent recognition module, train the intent recognition model, and verify the accuracy after training is completed.

[0016] Step S2-1: Use the Erlangshen model (Erlangshen-DeBERTa-v2-97M-Chinese) from (https: / / github.com / IDEA-CCNL / Fengshenbang-LM) as the base model. Add a fully connected layer at the end of the base model so that the output of the base model is mapped to 2, corresponding to the two intent classifications in step S1-2.

[0017] Step S2-2: Train the intent recognition model. Using the model built in Step S2-1, input the training set from Step S1-2 and apply Cross Entropy Loss as the loss function during training. Carefully adjust parameters such as the learning rate and training epochs to train the intent recognition model. Save the model weights after training is complete.

[0018] Step S2-3: Initialize the intent recognition model using the weights saved in step S2-2. Input the test set from step S2-1 to verify accuracy. If the verification accuracy does not meet the requirements, adjust the relevant training parameters and repeat step S2-2 until the accuracy meets the requirements.

[0019] Step S3: Semantic Analysis Using a Large Language Model. The large language model is used to extract literature information from the user's question, such as title, DOI, and keywords. After extracting the literature information, a literature search API is used to perform a literature search. After completing the search, the large model is applied to filter relevant literature.

[0020] Step S3-1: Call the API of the large model open platform, such as the DeepSeek model (https: / / platform.deepseek.com). Input prompt words to extract the title, DOI, and keywords of the literature in the user's question. To reduce model illusion, the large model is set up as two: llm1 (title and DOI extraction) and llm2 (keyword extraction). For keyword extraction, Chinese keywords and English keywords need to be extracted, corresponding to the Chinese literature query API and the English literature query API, respectively. To ensure the accuracy of the translation of medical terms such as drug names and disease names, translation information needs to be injected into the prompt words of the large model. The translation information injection process is as follows: First, prepare the knowledge base corresponding to the medical terms. The knowledge base contains the brand name, generic name, medical terminology, and corresponding English translation of the drug name. Then, select bge-zh-lager (https: / / huggingface.co / BAAI / bge-large-zh) as the embedding model, introduce the BM25 algorithm for dual-path recall, and apply the Reciprocal Ranking Fusion (RRF) algorithm to process the results of dual-path recall. Finally, the most relevant knowledge from the knowledge base to the user's question is retrieved. To ensure accuracy, this knowledge is then input into the prompts of the large model for reference during translation.

[0021] Step S3-2: Use the literature search API to search for Chinese and English literature. For Chinese literature search, select Wanfang Medical (https: / / med.wanfangdata.com.cn), and for English literature search, select Infox-med (https: / / www.infox-med.com). First, determine if the title was extracted from the large model in Step S3-1. If the title was extracted, prioritize searching the corresponding API, i.e., determine whether the title is in Chinese or English; use Wanfang Medical for Chinese titles and Infox-med for English titles. Then, if no title was extracted, check if there is a DOI. If so, send the results to Wanfang Medical and Infox-med for literature retrieval respectively. Finally, if there is no title or DOI, enter Chinese keywords and English keywords respectively to obtain the medical literature (Chinese medical literature and English medical literature) retrieved from the two APIs.

[0022] Step S3-3: Use the DeepSeek model to filter the literature from Step S3-2. The model input includes the literature title, abstract, and user question. The DeepSeek model is used to filter Chinese and English medical literature separately, allowing the model to remove literature that is clearly not strongly relevant to the user question based on the title and abstract. To ensure better accuracy, the filtering of Chinese and English medical literature is performed separately; that is, llm3 is set for filtering Chinese medical literature, and llm4 for filtering English medical literature.

[0023] Step S4: Re-ranking of medical literature relevance. Construct a re-ranking model and re-rank the literature selected in steps S3-3 according to their relevance.

[0024] Step S4-1: Construct a medical literature reranking model. Select and apply bge-reranker-lager (https: / / huggingface.co / BAAI / bge-reranker-large) as the reranking model.

[0025] Step S4-2: Input the Chinese and English medical literature and keywords from Step S3-3 into the reordering model. To improve the accuracy of the reordering, the Chinese and English medical literature are reordered separately. For Chinese medical literature, the input to the reordering model is Chinese keywords, and for English medical literature, it is English keywords. After sorting, the literature will be sorted from high to low relevance to the user's query, with each literature corresponding to a relevance threshold (value between 0 and 1). Since the Chinese and English medical literature are reordered separately, a final mixed sort will be performed based on the relevance threshold from high to low to obtain the final literature list.

[0026] Step S5: Medical Literature Retrieval. Users input a question and apply the modules from the previous steps to obtain the final literature results.

[0027] Step S5-1: Input the user's question, apply the intent recognition model from step S2 to classify the intent, and determine whether the user wants to find a specific document title or related documents.

[0028] Step S5-2: After obtaining the intent, call the corresponding llm1 (find document title) or llm2 (find related documents) to call the API to perform document search.

[0029] Step S5-3: Input the user's question and the literature found, and call llm3 and llm4 to filter the literature.

[0030] Step S5-4: Sort the documents by similarity and obtain the documents with the highest similarity.

Claims

1. An intelligent medical literature retrieval system based on an AI large model and its working method, characterized in that... The system includes a problem collection module, an intent recognition module, a large-scale model semantic analysis module, a similarity ranking module, and a medical literature acquisition module. The problem collection module collects and annotates data from hospitals or public datasets. The aforementioned speech processing module uses the Erlangshen-DeBERTa-v2-97M-Chinese model combined with a fully connected layer for intent recognition and classification; The large-scale semantic analysis module first uses the DeepSeek model to analyze the semantics of user questions and extract the titles, DOIs, and keywords of the documents in the user's query; then it uses a third-party document query API to retrieve documents; finally, it uses the DeepSeek model to filter the retrieved documents. The similarity ranking module ranks documents based on the semantic similarity between the document titles and user questions, and obtains the ranked documents through the bge-reranker-large model. The medical literature acquisition module sequentially performs intent recognition, large-scale semantic analysis, and literature similarity ranking on the user's questions to obtain the final literature results.

2. An intelligent medical literature retrieval system based on an AI large-scale model and its working method, characterized in that... The method includes the following steps: Step S1: Collect and label medical questions. Step S2: Build the intent recognition module, train the intent recognition model, and verify the accuracy after training is completed. Step S3: Semantic Analysis Using a Large Language Model. The large language model is used to extract literature information from the user's question, such as title, DOI, and keywords. After extracting the literature information, a literature search API is used to perform a literature search. After completing the search, the large model is applied to filter relevant literature. Step S4: Re-ranking of medical literature relevance. Construct a re-ranking model and re-rank the literature based on relevance. Step S5: Medical Literature Retrieval. Users input a question and apply the modules from the previous steps to obtain the final literature results.

3. The intelligent medical literature retrieval system based on an AI large model and its working method according to claim 2, characterized in that... The specific method for collecting and labeling medical issues in step S1 is as follows: Step S1-1: Collect and organize medical literature search questions, including online Q&A within hospitals and publicly available medical Q&A datasets. After organization, filter the data to ensure that each question has at least one corresponding document (i.e., each medical literature search question has a corresponding answer). Step S1-2: Label the medical literature search questions that have been screened, and divide them into two categories according to the search intent: find literature titles and find related literature. After labeling, divide the training set and test set according to the 8:2 principle.

4. The intelligent medical literature retrieval system based on an AI large model and its working method according to claim 2, characterized in that... The intent recognition module setup, training, and validation in step S2 are as follows: Step S2-1: Use the Erlangshen model (Erlangshen-DeBERTa-v2-97M-Chinese) from (https: / / github.com / IDEA-CCNL / Fengshenbang-LM) as the base model. Add a fully connected layer at the end of the base model so that the output of the base model is mapped to 2, corresponding to the two intent classifications in step S1-2. Step S2-2: Train the intent recognition model. Using the model built in Step S2-1, input the training set from Step S1-2 and apply Cross Entropy Loss as the loss function during training. Carefully adjust parameters such as the learning rate and training epochs to train the intent recognition model. Save the model weights after training is complete. Step S2-3: Initialize the intent recognition model using the weights saved in step S2-2. Input the test set from step S2-1 to verify accuracy. If the verification accuracy does not meet the requirements, adjust the relevant training parameters and repeat step S2-2 until the accuracy meets the requirements.

5. The intelligent medical literature retrieval system based on an AI large model and its working method according to claim 2, characterized in that... The semantic parsing method of the large language model in step S3 is as follows: Step S3-1: Call the API of the large model open platform, such as the DeepSeek model (https: / / platform.deepseek.com). Input prompt words to extract the title, DOI, and keywords of the literature in the user's question. To reduce model illusion, the large model is set up as two: llm1 (title and DOI extraction) and llm2 (keyword extraction). For keyword extraction, Chinese keywords and English keywords need to be extracted, corresponding to the Chinese literature query API and the English literature query API, respectively. To ensure the accuracy of the translation of medical terms such as drug names and disease names, translation information needs to be injected into the prompt words of the large model. The translation information injection process is as follows: First, prepare the knowledge base corresponding to the medical terms. The knowledge base contains the brand name, generic name, medical terminology, and corresponding English translation of the drug name. Then, select bge-zh-lager (https: / / huggingface.co / BAAI / bge-large-zh) as the embedding model, introduce the BM25 algorithm for dual-path recall, and apply the Reciprocal Ranking Fusion (RRF) algorithm to process the results of dual-path recall. Finally, the most relevant knowledge from the knowledge base to the user's question is retrieved. To ensure accuracy, this knowledge is then input into the prompts of the large model for reference during translation. Step S3-2: Use the literature search API to search for Chinese and English literature. For Chinese literature search, select Wanfang Medical (https: / / med.wanfangdata.com.cn), and for English literature search, select Infox-med (https: / / www.infox-med.com). First, determine if the title was extracted from the large model in Step S3-1. If the title was extracted, prioritize searching the corresponding API, i.e., determine whether the title is in Chinese or English; use Wanfang Medical for Chinese titles and Infox-med for English titles. Then, if no title was extracted, check if there is a DOI. If so, send the results to Wanfang Medical and Infox-med for literature retrieval respectively. Finally, if there is no title or DOI, enter Chinese keywords and English keywords respectively to obtain the medical literature (Chinese medical literature and English medical literature) retrieved from the two APIs. Step S3-3: Use the DeepSeek model to filter the literature from Step S3-2. The model input includes the literature title, abstract, and user question. The DeepSeek model is used to filter Chinese and English medical literature separately, allowing the model to remove literature that is clearly not strongly relevant to the user question based on the title and abstract. To ensure better accuracy, the filtering of Chinese and English medical literature is performed separately; that is, llm3 is set for filtering Chinese medical literature, and llm4 for filtering English medical literature.

6. The intelligent medical literature retrieval system based on an AI large model and its working method according to claim 2, characterized in that... The specific method for re-ranking medical literature relevance in step S4 is as follows: Step S4-1: Construct a medical literature reranking model. Select and apply bge-reranker-lager (https: / / huggingface.co / BAAI / bge-reranker-large) as the reranking model. Step S4-2: Input the Chinese and English medical literature and keywords from Step S3-3 into the reordering model. To improve the accuracy of the reordering, the Chinese and English medical literature are reordered separately. For Chinese medical literature, the input to the reordering model is Chinese keywords, and for English medical literature, it is English keywords. After sorting, the literature will be sorted from high to low relevance to the user's query, with each literature corresponding to a relevance threshold (value between 0 and 1). Since the Chinese and English medical literature are reordered separately, a final mixed sort will be performed based on the relevance threshold from high to low to obtain the final literature list.

7. The intelligent medical literature retrieval system based on an AI large model and its working method according to claim 2, characterized in that... The specific method for obtaining medical literature in step S5 is as follows: Step S5-1: Input the user's question, apply the intent recognition model from step S2 to classify the intent, and determine whether the user wants to find a specific document title or related documents. Step S5-2: After obtaining the intent, call the corresponding llm1 (find document title) or llm2 (find related documents) to call the API to perform document search. Step S5-3: Input the user's question and the literature found, and call llm3 and llm4 to filter the literature. Step S5-4: Sort the documents by similarity and obtain the documents with the highest similarity.

Images