Medical auxiliary diagnosis method and device based on large language model

Abstract

The invention relates to a medical auxiliary diagnosis method and device based on a large language model. The method comprises the following steps: constructing and preprocessing a multi-source medical text data set, and pre-training the constructed large language model by utilizing the multi-source medical text data set, according to the pre-training model, a multi-task instruction data set is constructed in a mode of combining model generation and manual annotation, and the pre-training model continues to be trained according to the constructed multi-task instruction data to obtain a medical large language model; constructing a medical expert system, and inputting related information of the patient into the expert system to inquire corresponding expert knowledge; a medical record retrieval system is constructed, and similar cases and diagnosis and treatment schemes corresponding to the similar cases are retrieved by inputting medical records of a user; and combining patient related information, expert knowledge, similar cases and diagnosis and treatment plans corresponding to the similar cases by designing a Prompt template, and inputting the combined information, expert knowledge, similar cases and diagnosis and treatment plans into a medical big language model to obtain the diagnosis and treatment plans of the medical cases. According to the invention, the precision and efficiency of medical diagnosis are improved.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning, and in particular relates to a medical auxiliary diagnosis method and device based on a large language model. Background Art

[0002] In the current medical field, doctors often need to refer to a large amount of medical literature, expert experience and past medical records when diagnosing patients. However, traditional information retrieval and query methods have problems such as low efficiency and incomplete information, which cannot meet the needs of modern medical care for speed and accuracy. With the rapid development of artificial intelligence technology, especially the widespread application of large language models (LLM), new solutions have been provided for medical auxiliary diagnosis.

[0003] Traditional medical auxiliary diagnosis methods mainly rely on the personal experience and knowledge of doctors, as well as the medical record database within the hospital. However, this method has several obvious flaws. First, the personal experience and knowledge of doctors are limited, and it is difficult to cover all diseases and cases. Second, the medical record database within the hospital is often limited by factors such as region and hospital size, resulting in incomplete and inadequate medical record data. Finally, traditional medical record retrieval methods are usually based on keyword matching, which cannot accurately capture the semantic similarity and correlation between medical records.

[0004] In order to solve these problems, a medical auxiliary diagnosis method based on large language models has emerged in recent years. Large language models have powerful language understanding and generation capabilities and can process and understand massive amounts of text data. By training large language models and giving them knowledge and reasoning capabilities in the medical field, doctors can be provided with more comprehensive, accurate, and personalized auxiliary diagnosis suggestions.

[0005] However, there are still some challenges in the existing medical auxiliary diagnosis methods based on large language models. First, medical text data is diverse and complex, and how to effectively build and preprocess multi-source medical text datasets is a difficult problem. Second, large language models need to be specially trained and optimized for the medical field to improve their performance in medical field tasks. Finally, how to effectively combine the patient's relevant information, expert knowledge and similar medical records to provide doctors with more accurate and practical auxiliary diagnosis suggestions is also a problem that needs to be solved urgently. Summary of the invention

[0006] In order to solve the problems existing in the background technology, one aspect of the present invention provides a medical auxiliary diagnosis method based on a large language model, comprising:

[0007] S1: Build and preprocess a multi-source medical text dataset;

[0008] S2: Use multi-source medical text datasets to pre-train the constructed large language model to obtain a pre-trained model;

[0009] S3: Based on the pre-trained model, a multi-task instruction dataset is constructed by combining model generation with manual annotation. The pre-trained model is further trained based on the constructed multi-task instruction data to obtain a medical large language model.

[0010] S5: Build a medical expert system and input the patient's relevant information into the expert system to query the corresponding expert knowledge;

[0011] S6: Build a medical record retrieval system to retrieve similar cases and corresponding diagnosis and treatment plans of similar cases by inputting the user's medical records;

[0012] S7: By designing a prompt template, patient-related information, expert knowledge, similar cases and corresponding treatment plans for similar cases are combined and input into the medical big language model to obtain the treatment plan for the medical record.

[0013] Another aspect of the present invention provides a medical auxiliary diagnosis device based on a large language model, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the medical auxiliary diagnosis device based on a large language model executes the medical auxiliary diagnosis method based on a large language model.

[0014] Another aspect of the present invention provides a computer-readable storage medium storing a program, characterized in that when the program is executed by a processor, the medical auxiliary diagnosis method based on a large language model is implemented.

[0015] The present invention has at least the following beneficial effects

[0016] The present invention pre-trains the large language model through a multi-source medical text data set, providing rich and accurate training materials for the medical large language model, thereby improving the model's understanding and application capabilities of medical knowledge; by combining a multi-task instruction data set to further train the large language model, the medical large language model can more accurately understand the patient's symptoms, medical history and other information, and give more accurate diagnostic suggestions; by building a medical expert system, the present invention can match the patient's relevant information with expert knowledge and provide personalized medical advice to the patient. This not only helps to improve the level of personalization of medical services, but also alleviates the problem of tight medical resources to a certain extent; the medical record retrieval system can quickly retrieve similar cases and corresponding diagnosis and treatment plans based on the patient's medical record information, providing valuable reference information for the model. This helps the model to consider the patient's specific situation more comprehensively and accurately when generating a treatment plan; by designing a reasonable Prompt template, the present invention can effectively combine key information such as patient-related information, expert knowledge, similar cases and treatment plans corresponding to similar cases, which helps the model make more scientific and reasonable medical decisions based on obtaining comprehensive information, thereby achieving improved accuracy and efficiency of medical diagnosis, personalized and intelligent medical auxiliary diagnosis, optimization of medical decision-making processes, and the development of medical informatization and intelligence. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the method flow of the present invention. DETAILED DESCRIPTION

[0018] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0019] See also Figure 1 The present invention provides a medical auxiliary diagnosis method based on a large language model, comprising:

[0020] S1: Build and preprocess a multi-source medical text dataset;

[0021] Preferably, the constructing and preprocessing of a multi-source medical text dataset comprises:

[0022] S11: Collect medical text data related to medical treatment from public web pages, books, papers and databases;

[0023] S12: Use k-means + large language model method to filter out low-quality text data;

[0024] S13: Based on the large language model, use the few-shot method to generate more text data to build a multi-source medical text dataset.

[0025] In this embodiment, medical-related text data is captured from public web pages such as major medical websites, online medical platforms, and medical forums. For example, discussions and questions and answers between doctors and patients about a certain disease are obtained from medical forums such as "Ding Xiang Yuan". Medical books are obtained from libraries, online bookstores, and other channels, and medical text data is extracted therein. For example, descriptions of disease diagnosis and treatment methods are extracted from classic medical books such as "Internal Medicine". Medical papers are downloaded from medical paper databases such as PubMed and Medline, and medical text data is extracted therein. For example, descriptions of drug efficacy, side effects, etc. are extracted from a clinical research paper on a new drug. Medical text data is extracted from databases such as hospital information systems and electronic medical record systems. For example, the patient's medical history, diagnosis, treatment records, etc. are extracted from electronic medical records.

[0026] The collected medical text data are clustered using the k-means clustering algorithm. The text data is divided into multiple clusters based on the similarity of the text content.

[0027] The large language model is used to evaluate the quality of the text data in each cluster. The large language model can score the text quality based on the grammar, semantics, coherence and other aspects of the text.

[0028] Based on the evaluation results of the large language model, low-quality text data is filtered out. For example, text with grammatical errors, unclear semantics, and poor coherence is deleted.

[0029] Assume that 1,000 medical text data are collected and 10 clusters are obtained after k-means clustering. Then, the large language model is used to evaluate the quality of the text in each cluster and a quality threshold is set. After the evaluation, it is found that the quality of the text in 3 of the clusters is generally low, below the set threshold. Therefore, all the text data in these 3 clusters are filtered out, and finally 700 high-quality medical text data are retained.

[0030] Few-shot learning is a technique that can quickly learn and generalize to new tasks with a small number of samples. Here, the few-shot learning method can be used to generate more medical text data based on the large language model. First, some representative samples are selected from the filtered high-quality medical text data as the input of few-shot learning. Then, the large language model is used to generate more medical text data based on these input samples. For example, texts such as detailed descriptions, diagnostic processes, and treatment methods of a certain disease can be generated. The generated medical text data is merged with the filtered high-quality medical text data to construct a multi-source medical text dataset. This dataset contains medical text data from different sources, in different formats, and of different qualities, and can be used for subsequent medical text analysis, knowledge discovery, and other tasks.

[0031] Assume that 700 high-quality medical text data have been obtained after filtering, and 50 representative samples are selected as input for few-shot learning. Then, 300 new medical text data are generated based on these input samples using a large language model. Finally, these 300 newly generated text data are merged with the original 700 high-quality text data to construct a multi-source medical text dataset containing 1,000 texts.

[0032] Through the above steps and examples, a high-quality, multi-source medical text dataset can be constructed to provide strong support for subsequent medical text analysis, knowledge discovery and other tasks.

[0033] S2: Use multi-source medical text datasets to pre-train the constructed large language model to obtain a pre-trained model;

[0034] Preferably, the pre-training of the constructed large language model using a multi-source medical text dataset includes:

[0035] Based on a multi-source medical text dataset, every 3,000 words are used as a training sample. The full-parameter training method is used to train the constructed large language model, so that the large language model can learn the medical knowledge in the data. After the training, a pre-trained model is obtained, which can generate relevant medical text content based on a small amount of prompt text.

[0036] S3: Based on the pre-trained model, a multi-task instruction dataset is constructed by combining model generation with manual annotation. The pre-trained model is further trained based on the constructed multi-task instruction data to obtain a medical large language model.

[0037] Preferably, the constructed large language model includes: based on the llama large language model, increasing the number of attention layers and mlp layers, and constructing a large language model with a parameter amount greater than 30 billion.

[0038] Preferably, the constructing of multi-task instruction data includes:

[0039] S31: Define the data format of the basic instruction as D={(x i ,y i )}, where x i Indicates input command, y i represents the expected model output;

[0040] S32: designing a large number of basic instructions to enhance the basic capabilities of the large language model, wherein the basic instructions are composed of optional keywords and fixed templates;

[0041] S33: Designing a diagnosis and treatment plan generation instruction to enhance the professional ability of the large language model, wherein the diagnosis and treatment plan generation instruction includes a specific diagnosis and treatment plan generation task;

[0042] S34: The designed instruction x i Input into the large language model to generate the corresponding result y i ;

[0043] S35: All instructions x i And the model output result y i Arranged into standard instruction data format D = {(x i ,y i )}, and conduct manual inspection and modification to ensure the accuracy and practicality of the data;

[0044] S36: Based on the modified instruction data D, the pre-trained large language model is further trained using the full-parameter training method to obtain a medical large language model that can handle tasks in the medical field.

[0045] In this embodiment, x i Represents the instructions input to the large language model, for example: "Please describe the common symptoms of pneumonia."; y iRepresents the expected model output, for example: "Common symptoms of pneumonia include fever, cough, and difficulty breathing." When designing basic instructions, we combined optional keywords and fixed templates. Keywords include: "Disease name", "Symptoms", "Treatment methods", etc. Fixed templates include: "Please list the common symptoms of {disease name}." Or "What are the common treatments for {disease name}?" By combining different keywords and templates, we can generate a large number of basic instructions, such as: "Please list the common symptoms of diabetes.", "What are the common treatments for hypertension?", etc. The diagnosis and treatment plan generation instruction requires the model to generate personalized diagnosis and treatment recommendations based on the patient's specific situation. The instruction may contain information such as the patient's medical history, current symptoms, and physical examination results. For example: "The patient, male, 50 years old, has a history of hypertension and recently developed chest pain and shortness of breath. The electrocardiogram shows myocardial ischemia. Please design a preliminary diagnosis and treatment plan for this patient." Put the designed instruction x i Input into the large language model, the model will generate the corresponding result y i For example, if the input instruction is: "Please describe the early symptoms of lung cancer.", the model may generate the output: "Early symptoms of lung cancer may include coughing, sputum, chest pain, and difficulty breathing." i And the model output result y i Arranged into standard instruction data format D = {(x i ,y i )}. Manually check the data to ensure the accuracy of the output. For example, if the model's output for a certain instruction is found to be inaccurate or incomplete, it can be modified or supplemented.

[0046] Ensure the practicality of the data, that is, the data can truly reflect the needs and challenges in the medical field.

[0047] S5: Build a medical expert system and input the patient's relevant information into the expert system to query the corresponding expert knowledge;

[0048] Preferably, the medical expert system comprises: an expert knowledge base, a recall system and a sorting system; and the construction of the medical expert system comprises:

[0049] Collect expert knowledge, organize expert knowledge into structured text data, encode the structured text data into text vectors through a deep learning model; build an expert knowledge base based on text vectors and text data;

[0050] Build a recall system, process the query text into a query vector, calculate the similarity between the query text and the text data in the expert knowledge base, calculate the similarity between the query vector and the text vector in the expert knowledge base, calculate the average of the two similarities, and recall the text data whose average similarity is greater than a certain value;

[0051] A sorting system is constructed, wherein the sorting system is used to sort the text data recalled by the recall system, score the recalled texts based on the average similarity value, and select the text data with the highest score as the expert knowledge.

[0052] S6: Build a medical record retrieval system to retrieve similar cases and corresponding diagnosis and treatment plans of similar cases by inputting the user's medical records;

[0053] S7: By designing a prompt template, patient-related information, expert knowledge, similar cases and corresponding treatment plans for similar cases are combined and input into the medical big language model to obtain the treatment plan for the medical record.

[0054] In this embodiment, the prompt template is designed as follows, for example, “Please provide a diagnosis and treatment plan based on expert knowledge and patient information, combined with reference cases and their diagnosis and treatment plans.

[0055] Expert knowledge:{}

[0056] Reference case: {}

[0057] Patient information: {}"

[0058] Preferably, the medical record retrieval system includes: a medical record library, a recall system and a sorting system; the construction of the medical record retrieval system includes:

[0059] Collect patient medical records and desensitize them; organize patient medical records into QA question-answer pairs, where Q represents the patient's medical history and examination report, and A represents the diagnosis and treatment plan; encode the QA question-answer text into a text vector using a deep learning model; and build a medical record library based on the text vector and text data;

[0060] Build a recall system, process the query text into a query vector, calculate the similarity between the query text and the text data in the medical record database, calculate the similarity between the query vector and the text vector in the medical record database, calculate the average of the two similarities, and recall the text data whose average similarity is greater than a certain value;

[0061] A sorting system is constructed to sort the text data recalled by the recall system, score the recalled text based on the average similarity value, and select QA question and answer texts with scores higher than a set threshold as similar medical records and corresponding diagnosis and treatment plans.

[0062] Another aspect of the present invention provides a medical auxiliary diagnosis device based on a large language model, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the medical auxiliary diagnosis device based on a large language model executes the medical auxiliary diagnosis method based on a large language model.

[0063] Another aspect of the present invention provides a computer-readable storage medium storing a program, characterized in that when the program is executed by a processor, the medical auxiliary diagnosis method based on a large language model is implemented.

[0064] In summary, the present invention pre-trains the large language model through a multi-source medical text data set, providing rich and accurate training materials for the medical large language model, thereby improving the model's understanding and application capabilities of medical knowledge; by combining a multi-task instruction data set to further train the large language model, the medical large language model can more accurately understand the patient's symptoms, medical history and other information, and give more accurate diagnostic suggestions; by constructing a medical expert system, the present invention can match the patient's relevant information with expert knowledge to provide patients with personalized medical advice. This not only helps to improve the level of personalization of medical services, but also alleviates the problem of tight medical resources to a certain extent; the medical record retrieval system can quickly retrieve similar cases and corresponding diagnosis and treatment plans based on the patient's medical record information, providing valuable reference information for the model. This helps the model to consider the patient's specific situation more comprehensively and accurately when generating a treatment plan; by designing a reasonable Prompt template, the present invention can effectively combine key information such as patient-related information, expert knowledge, similar cases and treatment plans corresponding to similar cases, which helps the model make more scientific and reasonable medical decisions based on obtaining comprehensive information, thereby achieving improved accuracy and efficiency of medical diagnosis, personalized and intelligent medical auxiliary diagnosis, optimization of medical decision-making processes, and the development of medical informatization and intelligence.

[0065] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the program can be stored in a non-volatile computer-readable storage medium. When the program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory may include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0066] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A medical auxiliary diagnosis method based on a large language model, characterized in that: include: S1: Build and preprocess a multi-source medical text dataset; S2: Use multi-source medical text datasets to pre-train the constructed large language model to obtain a pre-trained model; S3: Based on the pre-trained model, a multi-task instruction dataset is constructed by combining model generation with manual annotation. The pre-trained model is further trained based on the constructed multi-task instruction data to obtain a medical large language model. S5: Build a medical expert system and input the patient's relevant information into the expert system to query the corresponding expert knowledge; S6: Build a medical record retrieval system to retrieve similar cases and corresponding diagnosis and treatment plans of similar cases by inputting the user's medical records; S7: By designing a prompt template, patient-related information, expert knowledge, similar cases and corresponding treatment plans for similar cases are combined and input into the medical big language model to obtain the treatment plan for the medical record.

2. A medical auxiliary diagnosis method based on a large language model according to claim 1, characterized in that: The construction and preprocessing of a multi-source medical text dataset includes: S11: Collect medical text data related to medical treatment from public web pages, books, papers and databases; S12: Use k-means + large language model to filter out low-quality text data; S13: Based on the large language model, use the few-shot method to generate more text data to build a multi-source medical text dataset.

3. The medical auxiliary diagnosis method based on a large language model according to claim 1, characterized in that: The constructed large language model includes: based on the llama large language model, increasing the number of attention layers and mlp layers, and constructing a large language model with more than 30 billion parameters.

4. The medical auxiliary diagnosis method based on a large language model according to claim 1, characterized in that: The pre-training of the constructed large language model using a multi-source medical text dataset includes: Based on a multi-source medical text dataset, every 3,000 words are used as a training sample. The full-parameter training method is used to train the constructed large language model, so that the large language model can learn the medical knowledge in the data. After the training, a pre-trained model is obtained, which can generate relevant medical text content based on a small amount of prompt text.

5. The medical auxiliary diagnosis method based on a large language model according to claim 1, characterized in that: The constructing of multi-task instruction data includes: S31: Define the data format of the basic instruction as D={(x i ,y i )}, where x i Indicates input command, y i represents the expected model output; S32: designing a large number of basic instructions to enhance the basic capabilities of the large language model, wherein the basic instructions are composed of optional keywords and fixed templates; S33: Designing a diagnosis and treatment plan generation instruction to enhance the professional ability of the large language model, wherein the diagnosis and treatment plan generation instruction includes a specific diagnosis and treatment plan generation task; S34: The designed instruction x i Input into the large language model to generate the corresponding result y i ; S35: All instructions x i And the model output result y i Arranged into standard instruction data format D = {(x i ,y i )}, and conduct manual inspection and modification to ensure the accuracy and practicality of the data; S36: Based on the modified instruction data D, the pre-trained large language model is further trained using the full-parameter training method to obtain a medical large language model that can handle tasks in the medical field.

6. The medical auxiliary diagnosis method based on a large language model according to claim 1, characterized in that: The medical expert system includes: an expert knowledge base, a recall system and a sorting system; the construction of the medical expert system includes: Collect expert knowledge, organize expert knowledge into structured text data, encode the structured text data into text vectors through a deep learning model; build an expert knowledge base based on text vectors and text data; Build a recall system, process the query text into a query vector, calculate the similarity between the query text and the text data in the expert knowledge base, calculate the similarity between the query vector and the text vector in the expert knowledge base, calculate the average of the two similarities, and recall the text data whose average similarity is greater than a certain value; A sorting system is constructed, wherein the sorting system is used to sort the text data recalled by the recall system, score the recalled texts based on the average similarity value, and select the text data with the highest score as the expert knowledge.

7. The medical auxiliary diagnosis method based on a large language model according to claim 1, characterized in that: The medical record retrieval system includes: a medical record library, a recall system and a sorting system; the construction of the medical record retrieval system includes: Collect patient medical records and desensitize them; organize patient medical records into QA question-answer pairs, where Q represents the patient's medical history and examination report, and A represents the diagnosis and treatment plan; encode the QA question-answer text into a text vector using a deep learning model; and build a medical record library based on the text vector and text data; Build a recall system, process the query text into a query vector, calculate the similarity between the query text and the text data in the medical record database, calculate the similarity between the query vector and the text vector in the medical record database, calculate the average of the two similarities, and recall the text data whose average similarity is greater than a certain value; A sorting system is constructed to sort the text data recalled by the recall system, score the recalled text based on the average similarity value, and select QA question and answer texts with scores higher than a set threshold as similar medical records and corresponding diagnosis and treatment plans.

8. A medical auxiliary diagnosis device based on a large language model, characterized in that: It includes a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the medical auxiliary diagnosis device based on a large language model executes the medical auxiliary diagnosis method based on a large language model described in any one of claims 1 to 7.

9. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, a medical auxiliary diagnosis method based on a large language model as described in any one of claims 1 to 7 is implemented.

Images