Medical information response method, computer device and readable storage medium
By using multi-model cross-validation and thought chain technology, and combining medical databases to construct vectorized information, the problems of insufficient professionalism and difficulty in multi-step reasoning of large language models in medical question answering are solved, and the generated responses are more representative and credible.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHIYING MATRIX (XIONGAN) MEDICAL TECHNOLOGY CO LTD
- Filing Date
- 2026-04-10
- Publication Date
- 2026-07-31
AI Technical Summary
Existing large language models in the medical field suffer from insufficient professionalism, difficulties in multimodal information processing, difficulties in multi-step reasoning, and a lack of question type identification and result arbitration, resulting in low-quality answers.
The method employs multi-model cross-validation, using the first major model for initial classification, the second major model to search the medical database and correct or supplement the first response, and the logical steps to be generated are guided by the thinking chain technique. Combined with the medical database, vectorized information is constructed to generate the third response.
It improves the professionalism and accuracy of medical Q&A, and the generated responses are more in line with user needs, possessing high credibility and representativeness.
Smart Images
Figure CN122489690A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of smart healthcare, and in particular to medical information response methods, computer devices, and readable storage media. Background Technology
[0002] Currently, mainstream large language models (MLMs), such as ChatGPT, GPT-4, Claude, and Wenxin Yiyan, can perform generalized question answering, but they lack the support of specialized domain knowledge. Retrieval-augmented generation (RAG) technology has been initially applied in some open-source frameworks, for example, by retrieving relevant paragraphs after word embedding from a specific literature database before generating the model. However, the following problems still exist: (1) Surgical data is distributed in various formats (such as PDF, PPT, forum posts, pictures, etc.), making it difficult to collect and integrate this data; (2) Multimodal medical information (text and image hybrid) poses challenges to the processing of existing text models; (3) User questions often involve multiple steps of reasoning and judgment, while conventional RAG systems can only provide superficial answers based on similar paragraphs; (4) Lacking a multi-agent structure such as problem type identification, task routing, and result arbitration, it is easy to generate low-quality content without evidence support. Summary of the Invention
[0003] The purpose of this application is to provide a medical information response method, computer device, and readable storage medium, which aims to solve the technical problem in the prior art that conventional RAG systems can only provide superficial answers based on similar paragraphs.
[0004] To achieve the above objectives, this application provides a medical information response method, comprising: Obtain input information from the client; The input information is classified using the first large model, and a first response is generated based on the input information, wherein the categories of the input information include general and professional categories; If the input information is of a professional type, then based on the input information and the first response, the second major model is used to search the medical database to obtain the first search result; Using the second major model, a second response is generated based on the input information, the first response, and the first search result; A third response is generated based on the input information, the first response, and the second response; The third response is output to the client.
[0005] Optionally, generating the first response based on the input information includes: The second search result is obtained by retrieving a general database based on the input information; Using the first large model, the first response is generated based on the input information and the second search result.
[0006] Optionally, it further includes: if the input information is confirmed to be of the normal type, then outputting the first response to the client.
[0007] Optionally, the first large model can be guided step by step to generate the first response using thought chain technology.
[0008] Optionally, the second major model can be guided step-by-step to generate the second response using mind chain technology.
[0009] Optionally, when generating the second response using the second large model, the second large model is also used to label the information sources corresponding to each item of the generated second response.
[0010] Optionally, the third response is generated by a third major model, and when generating the third response, the information source corresponding to each item of the third response is marked.
[0011] Optionally, the method further includes the following step before outputting the third response to the client: Confirm whether all items in the third response have been marked with their corresponding information sources. If not, delete the items in the third response that do not have information sources. Confirm whether any of the contents of the third response contain any prohibited words. If so, replace the prohibited words with standard words.
[0012] Optionally, the construction of the medical database includes: Collect professional information from various medical websites and vectorize the professional information; The vectorized professional information is stored in the medical database.
[0013] Optionally, the professional information includes character information and non-character information; The process of vectorizing the professional information includes: The first character segment is extracted from the non-character information using a character extraction module, and the character information and the first character segment are vectorized. Optionally, vectorizing the professional information includes: The non-character information is converted into a second character segment using an image understanding module, and the character information and the second character segment are vectorized.
[0014] To achieve the above objectives, this application also provides a computer device, including a memory and a processor; The memory is connected to the processor. The memory is used to store computer programs, and the processor is used to call the computer programs so that the computer device executes the above-described medical information response method.
[0015] To achieve the above objectives, this application also provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and to execute the medical information response method of claim 1.
[0016] This application utilizes a second response generated by a second major model to correct or supplement the first response generated by a first major model, and then organizes the second and first responses to form a third response which is output to the client. In this process, it is equivalent to simulating a thought-provoking process involving at least two people on a single issue using the first and second major models, making the output third response more tailored to the user's needs. Furthermore, since the second response is generated based on search results from a medical database, the aforementioned thought-provoking process is more professional, and the output third response is more representative. Attached Figure Description
[0017] Figure 1 This is a flowchart illustrating the medical information response method according to an embodiment of this application.
[0018] Figures 2 to 4 This is a partial flowchart illustrating the medical information response method according to an embodiment of this application.
[0019] Figure 5 This is a schematic diagram of the structure of a computer device according to an embodiment of this application. Detailed Implementation
[0020] To explain in detail the technical content, structural features, objectives and effects of this application, the following description is provided in conjunction with the embodiments and accompanying drawings.
[0021] Existing large-scale models mostly retrieve input information, generate a hypothesis, and then output an answer based on that hypothesis. However, if the hypothesis itself is incorrect, the final output will not meet the user's expectations. For example, a user who is a doctor needs to understand the term "PE" in the medical field. The user asks the large-scale model, "What is PE?" There are multiple possibilities for this question. For example, "PE" could refer to polyethylene (PE) in the chemical field, physical education in the educational context, or pulmonary embolism (a life-threatening emergency) in the medical field. However, due to the lack of specific contextual description, the large-scale model might generate an answer based on only one possibility (e.g., only based on polyethylene in the chemical field), or it might provide a general explanation based on all three possibilities. Therefore, large-scale models may provide incorrect answers or answers with insufficient depth. Furthermore, for some newer medical knowledge, such as information on new drugs that is currently lacking in general databases (e.g., the mechanism of action of herbal decoctions used during the COVID-19 pandemic), the lack of corresponding search results prevents large-scale models from providing sufficiently in-depth answers. Therefore, existing technologies suffer from insufficient professionalism in the answers provided by large-scale models. To address this, this application discloses the following embodiments based on cross-validation by multiple models (CVM) technology: Example 1 Please see Figure 1 This application discloses a medical information response method, including: S1, obtain the client's input information.
[0022] S2 utilizes the first large model to classify the input information and generate at least one first response based on the input information. The input information categories include professional and general categories. Professional categories require a more specialized answer, such as "What should be done if the portal vein bleeds during surgery?" or "Please provide specific methods for stopping bleeding at the portal vein bifurcation during surgery." These are questions requiring specialized medical knowledge to answer. General categories are questions that do not require particularly specialized medical knowledge to answer, such as "What fruits and vegetables can lower blood pressure?". Furthermore, as can be seen from the above examples, the input information is not limited by its specific form; it can be a question posed to the large model or a requirement for the model's output, but it is not limited to these.
[0023] S3. If the input information is confirmed to be of a professional category, then based on the input information and the first response, the second major model is used to search the medical database to obtain the first search result. S4. Using the second major model, a second response is generated based on the input information, the first response, and the first search result.
[0024] It should be explained that, based on the search results from the medical database, if it is confirmed that the content of the first response does not correspond to the input information, the second model can point out the problems in the first response and correct them in the generated second response. Alternatively, if it is confirmed that the content of the first response cannot fully meet the response requirements corresponding to the input information, the second model can supplement the first response in the generated second response. The medical database can be built based on medical websites such as the FDA database (US Food and Drug Administration), PubMed, and Guidelines.
[0025] For example, if the input information is "What is PE?", after sending the input information and the first response generated by the first model to the second model, the second response output by the second model can correct or supplement the content of the first response. For example, "The first response answered PE plastic, but the user is a doctor or a researcher in the medical field, and the required answer should be about specific information on pulmonary embolism. After searching the medical database, the corresponding specific information is..." or "The first answer is a general explanation of pulmonary embolism, but considering that the user is a doctor, they may also be interested in the causes of pulmonary embolism, related treatments, or drugs. After searching the medical database, the corresponding information is as follows...". Thus, the second model can generate a second response based on the search results, where the second response includes corrections or supplements to the first response, achieving an effect similar to two people debating a problem.
[0026] S5 generates a third response based on the input information, the first response, and the second response. The third response can be generated by other major models besides the first and second major models, by processing the first and second responses according to the input information. Alternatively, it can be generated by the first major model itself. This can be adjusted according to the actual situation and is not subject to specific restrictions.
[0027] S6 outputs the third response to the client.
[0028] This application utilizes a second response generated by a second major model to correct or supplement the first response generated by a first major model, and then organizes the input information, the first response, and the second response to form a third response, which is output to the client. In this process, each major model answers the input information, and the content of the answers is further summarized, organized, and supplemented. This is equivalent to using the first and second major models to simulate a thought-provoking process involving at least two people on a single issue, making the output third response more tailored to the user's needs. Furthermore, since the second response is generated based on search results from a medical database, the aforementioned thought-provoking process is more professional, and the output third response is more representative.
[0029] Specifically, there can be multiple second-largest models. By utilizing multiple second-largest models to generate multiple second responses, the generation of the third response can be based on these multiple second responses. Since the second-largest models search medical databases, it's equivalent to referencing the suggestions of multiple large-scale medical experts when generating the third response, making the final generated third response more reliable. Of course, this is not limited to this; a single second-largest model can also output multiple times to obtain multiple second responses.
[0030] More specifically, when organizing the input information, the first response, and each of the second responses, complementary content in each response can be merged, while contradictory content is scored, discarding content with lower scores and retaining only content with the highest scores. Then, a third response is generated based on the above content.
[0031] Please see Figure 2 Specifically, step S2 includes: S21, retrieve the second search result from the general database based on the input information.
[0032] S22, using the first major model, generates the first response based on the input information and the second search results.
[0033] More specifically, step S2 also includes: S23. If the input information is confirmed to be of the normal type, the first response will be output to the client without performing any subsequent steps.
[0034] For example, if the input information is "What is aspirin?", since aspirin has been around for a long time, existing general databases contain a large amount of information about aspirin. The first major model can be used to confirm that this input information is of a general type, eliminating the need to call the second major model for an answer. Steps S21 to S22 can be directly executed to generate a first response, which is then output to the client. This helps reduce unnecessary calculations and saves computational resources. However, if the input information is more specialized, such as "What should I do if there is bleeding from a bifurcation of the portal vein during surgery?", and existing general databases lack relevant detailed information, the first major model will confirm that this input information is of a specialized type. In this case, the second major model will be used to provide a further response based on the input information, the first search result, and the first response generated by the first major model.
[0035] Existing large-scale models may produce "hallucinations" in medical question answering, meaning they generate answers that sound reasonable but lack medical evidence. To address this, this application discloses the following solution to remove these "hallucinations" from large-scale models: Specifically, in step S3, the second large model is guided step-by-step to generate a second response using chain-of-thought (CoT) technology. CoT involves explicitly writing out the logical steps of the reasoning process of the large model, rather than directly providing the final answer. This facilitates cross-validation, error checking, and enhanced interpretability. The large model is required not only to output the answer but also to provide the reasoning chain, such as: input question → analyze keywords → search the database → generate the answer by combining literature conclusions. Throughout the process, it is ensured that each logical step corresponds to the search results in the medical database, making the output of the large model traceable and allowing doctors and researchers to quickly verify the source.
[0036] More specifically, when generating the second response using the second major model, the second major model is also used to identify the corresponding information sources for each item in the generated second response.
[0037] This application adopts the RAG framework, which restricts the second-largest model from generating a second response based on the information source by forcing the second-largest model to specify the corresponding information source when generating the second response. This prevents the second-largest model from generating unfounded response content and makes the content generated by the second-largest model more credible.
[0038] Similarly, the first and third responses can also be generated using the RAG framework: Specifically, when generating the first response using the first major model, the first major model is also used to identify the corresponding information sources for each item in the generated first response.
[0039] Specifically, when generating the third response, the large model for generating the third response is used to identify the corresponding information sources for each item in the generated third response.
[0040] Medical databases can be built based on authoritative medical websites such as the FDA database (US Food and Drug Administration), PubMed, and Guidelines.
[0041] Please see Figure 3 Specifically, the construction of medical databases includes: S31. Collect professional information from various medical websites and vectorize it. This professional information can include literature, journals, drug instructions, etc., from these websites.
[0042] S32 stores the vectorized professional information into a medical database.
[0043] More specifically, the professional information includes both character and non-character information. Vectorizing the professional information involves: using a character extraction module to extract the first character segment from the non-character information, and then vectorizing both the character information and the first character segment. The character extraction module can be implemented using common OCR (Optical Character Recognition) technologies, such as deep learning models used to extract characters from images.
[0044] More specifically, the professional information includes both character and non-character information. Vectorizing the professional information involves: using an image understanding module to convert the non-character information into a second character segment, and then vectorizing the character information and the second character segment. The image understanding module can be implemented using a deep learning model.
[0045] Since large models need to vectorize information before processing it, if the medical database stores vectorized information, then there is no need to vectorize the search results again during retrieval, which helps to save computing resources.
[0046] Specifically, the third response is generated by the third major model. When generating the third response, the third major model is guided step by step to generate the third response using the thinking chain technology, and the information sources corresponding to each item of the third response are marked.
[0047] Please see Figure 4 More specifically, the procedure before step S6 includes: S61. Confirm whether all items in the third response have been marked with their corresponding information sources. If not, delete the content in the third response that does not have a corresponding information source marked. Of course, the content in the third response that does not have a corresponding information source marked can also be regenerated. The information source can be a hyperlink, or it can be a specific book title and page number, etc., without any restrictions on the specific format.
[0048] S62, confirm whether any content in the third response contains prohibited words. If so, replace the prohibited words with standard medical terminology (which can be obtained by searching medical databases or authoritative medical websites). Alternatively, the content containing prohibited words can be regenerated. Prohibited words can be terms that do not conform to standard terminology, such as ethical keywords like "experimental" and "off-label." When prohibited words are detected, the content containing the prohibited words can be replaced based on relevant content from authoritative medical websites such as the FDA database, PubMed, and Guidelines.
[0049] Optionally, after step S6, the method further includes: S7 scores the credibility of the third response and outputs the score to the client. For example, after outputting the third response, several fourth models review the third response and combine the scores from all the fourth models to obtain a credibility level of A for the third response. Then, a "Review passed [credibility level A]" mark is output to the client to help the user judge the credibility of the content of the third response. Of course, the third response input can also be scored for credibility by the aforementioned third models, which will not be elaborated further.
[0050] It is understandable that steps S61 and S62 have no specific order; for example, step S62 can be executed first, followed by step S61.
[0051] Optionally, the generation of corresponding response content by each model can be restricted according to the response framework. For example, each model can be restricted to describe the hemostasis process of portal vein bifurcation bleeding according to a step-by-step response framework, or when explaining pulmonary embolism, it can be described in points according to the response framework of cause, main symptoms, treatment methods and treatment drugs. By restricting the generation of response content by the large model according to a fixed response framework, the content generated by the large model can be more organized and easier to understand.
[0052] For example, in response to the input "What should I do if the portal vein bleeds during surgery?", the following chain of prompts can be generated step by step. Each prompt restricts the second-largest model to provide specific references or explanations for each step, guiding the second-largest model to respond: Please analyze the input information following the four steps below, and provide your detailed analysis results for each step: Step 1: Analyze the cause of bleeding (anatomical locations must be cited); "Step 2: Select a hemostasis technique (compression / electrocoagulation / suture) and explain the rationale;" "Step 3: Determine the conditions for conversion to laparotomy (quantitative indicators);" "Step 4: Based on the answers in Steps 1 to 3, please organize and output the complete processing flow." The second model generates the following content as a second response based on each prompt statement: "1. The cause of bleeding from the portal vein bifurcation is: cystic artery injury (source: video VL2023 02:15)". "2. The first-line method of hemostasis is compression to stop the bleeding at the portal vein bifurcation (according to guideline p23). If hemostasis is ineffective, the bleeding site should be sutured." "3. If the bleeding volume from the portal vein bifurcation is >500ml or excessive bleeding from the portal vein bifurcation causes loss of vision, the patient should be immediately transferred to an open abdominal surgery (based on page 1 of the PubMed paper XXX)." "4. In summary, the complete procedure for managing portal vein bifurcation bleeding is as follows: compression → suction → suturing → transfer decision."
[0053] The information source is the content within parentheses at the end of each step. If no corresponding information source is detected for each content before the second response is output, the second model can regenerate the second response.
[0054] Similarly, the mind chain technique can be used to gradually guide the corresponding large model when generating the first or third response.
[0055] Example 2 Figure 5 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Figure 5 As shown, the computer device may include a processor 201 and a memory 202. The memory 202 is connected to the processor 201 and is used to store computer programs. The processor 201 is used to invoke the computer programs to cause the computer device to execute the aforementioned medical information response method. Furthermore, the computer device may also include at least one communication bus 203. The communication bus 203 is used to enable communication between components. The memory 202 may be a high-speed RAM or non-volatile memory, such as at least one disk storage device.
[0056] Example 3 This application also provides a computer-readable storage medium storing a computer program adapted to be loaded by a processor and executed by the above-described medical information response method.
[0057] Example 4 This application also provides a computer program product or computer program that includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the electronic device to perform the aforementioned medical information response method.
[0058] It should be understood that, in the embodiments of this application, the processor may be a central processing unit (CPU), but it may also be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.
[0059] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be accomplished by hardware related to computer program instructions. The program can be stored in a computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. The storage medium can be a magnetic disk, optical disk, read-only memory (ROM), or random access memory (RAM), etc.
[0060] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.
[0061] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the scope of this application shall still fall within the scope of this application.
Claims
1. A medical information response method, characterized in that, include: Obtain input information from the client; The input information is classified using the first large model, and a first response is generated based on the input information, wherein the categories of the input information include general and professional categories; If the input information is of a professional type, then based on the input information and the first response, the second major model is used to search the medical database to obtain the first search result; Using the second major model, a second response is generated based on the input information, the first response, and the first search result; A third response is generated based on the input information, the first response, and the second response; The third response is output to the client.
2. The medical information response method as described in claim 1, characterized in that, The step of generating the first response based on the input information includes: The second search result is obtained by retrieving a general database based on the input information; Using the first large model, the first response is generated based on the input information and the second search result.
3. The medical information response method according to claim 2, wherein Also includes: If the input information is confirmed to be of the normal type, the first response is output to the client.
4. The medical information response method as described in claim 1, characterized in that, Using thought chain technology, the first large model is gradually guided to generate the first response; and / or, The second major model is gradually guided to generate the second response using the mind chain technique.
5. The medical information response method as described in claim 1, characterized in that, When generating the second response using the second major model, the second major model is also used to label the information sources corresponding to each item in the generated second response.
6. The medical information response method as described in claim 1 or 5, characterized in that, The third response is generated by the third major model. When generating the third response, the information source corresponding to each item in the third response is marked.
7. The medical information response method of claim 6, wherein, The process also includes the following steps before outputting the third response to the client: Confirm whether all items in the third response have been marked with their corresponding information sources. If not, delete the items in the third response that do not have information sources. Confirm whether any of the contents of the third response contain any prohibited words. If so, replace the prohibited words with standard words.
8. The medical information response method of claim 1, wherein, The construction of the medical database includes: Collect professional information from various medical websites and vectorize the professional information; The vectorized professional information is stored in the medical database.
9. The medical information response method as described in claim 8, characterized in that, The professional information includes both character information and non-character information; The process of vectorizing the professional information includes: The first character segment is extracted from the non-character information using a character extraction module, and the character information and the first character segment are vectorized. And / or, The process of vectorizing the professional information includes: The non-character information is converted into a second character segment using an image understanding module, and the character information and the second character segment are vectorized.
10. A computer device, comprising: Including memory and processor; The memory is connected to the processor, the memory is used to store computer programs, and the processor is used to invoke the computer programs so that the computer device executes the medical information response method according to any one of claims 1 to 9.
11. A computer readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted to be loaded by a processor and executed as described in any one of claims 1 to 9.