Electronic medical record structuring method based on large language model

Abstract

An electronic medical record structuring method based on a large language model comprises the steps of 1, obtaining a to-be-structured electronic medical record, and converting the to-be-structured electronic medical record into first question and answer dialogue data according to meta-model attribute tags; 2, according to each attribute tag of the meta-model, searching a first threshold number of similar electronic medical records and corresponding second question and answer dialogue data from a training database as context information; 3, inputting the first question and answer dialogue data and the context information into a large language model of medical record structured question and answer to obtain third question and answer dialogue data, and taking the third question and answer dialogue data as a medical record structured question and answer result corresponding to the input to-be-structured electronic medical record; and 4, according to the output template, converting the third question and answer dialogue data into an output text consistent with the structured standard. According to the method, the medical record is converted into a question and answer form, related medical record texts are searched to serve as context information to enhance the answer accuracy, output texts consistent in standard are output, and the medical record document structuring performance is improved.

Description

Technical Field

[0001] The present invention relates to the fields of text information extraction, data structuring and artificial intelligence technology, and in particular to an electronic medical record structuring method based on a large language model. Background Art

[0002] The input of the electronic medical record structuring process is an electronic medical record, which includes multiple electronic documents such as admission records, discharge records, examination reports and medical orders, etc. The output of the electronic medical record structuring is a set of key-value pairs, where the key values ​​are attribute words in the medical field such as body temperature, past medical history, long-term medication, etc. The goal of the electronic medical record structuring process is to extract a set of key-value pairs related to the medical field from the electronic medical record, and then store them in an xls file or database table. The structuring of electronic medical records helps to standardize and integrate medical data, promote information sharing between different medical institutions, and help reduce the situation where patients need to repeatedly provide the same information during the medical treatment process, thereby improving medical efficiency.

[0003] At present, the structured electronic medical records mainly use named entity recognition algorithms or information extraction algorithms in natural language processing technology to extract key data such as symptoms, diagnoses, and treatment plans from electronic medical records. Commonly used algorithms include rule-based methods, machine learning-based methods, and deep learning-based methods. However, the above methods require manual editing or a large amount of labeled data to complete the model, so conventional methods of constructing structured electronic medical records take a lot of time.

[0004] Therefore, due to the defects of the prior art, further improvement and development are needed. Summary of the invention

[0005] (I) Purpose of the invention: In order to solve the problems existing in the above-mentioned prior art, the purpose of the present invention is to provide an electronic medical record structuring method based on a large language model, to obtain a more accurate electronic medical record structuring effect, and to improve the efficiency of electronic medical record structuring.

[0006] (II) Technical solution: In order to solve the above technical problems, this technical solution provides an electronic medical record structuring method based on a large language model, which converts medical record documents into structured electronic medical records, including the following steps:

[0007] Step 1: obtaining an electronic medical record to be structured, and converting the electronic medical record to be structured into first question-answering dialogue data according to a metamodel attribute label;

[0008] Step 2: According to each attribute label of the meta-model, a first threshold number of similar electronic medical records and second question-answering dialogue data corresponding to the first threshold number of similar electronic medical records are searched from the training database as context information;

[0009] Step 3: Input the first question-answering dialogue data and context information into the large language model of medical record structured question-answering to obtain third question-answering dialogue data, and use the third question-answering dialogue data as the medical record structured question-answering result corresponding to the input electronic medical record to be structured;

[0010] Step 4: According to the output template, the third question-answer dialogue data is converted into an output text that conforms to the structured standard.

[0011] Among them, model training is carried out before the electronic medical record is structured, including the following steps:

[0012] Step 1: Establish a metamodel of structured targets for disease types; establish a correspondence between the attribute items of the metamodel and the medical record documents according to the attribute labels of the metamodel;

[0013] Step 2: preparing training data in a meta-model format, and converting the training data in the meta-model format into fourth question-answering dialogue data according to attribute labels of the meta-model;

[0014] Step 3: Use the large language model fine-tuned by the dialogue data as the large language model for medical record structured question and answer, input the training data of the fourth question and answer dialogue data in the meta-model format into the large language model for medical record structured question and answer, and output it as the second question and answer dialogue data.

[0015] The metamodel attribute tag is a unique identifier of the metamodel, and the attribute tag is an attribute word in the medical field.

[0016] The attribute items of the meta-model include attribute labels, questions corresponding to the attribute labels, all option values ​​corresponding to the attribute labels, and whether multiple selection is allowed;

[0017] A question-and-answer template is constructed based on all the attribute item information of the meta-model. The multiple-choice information of the attribute item determines whether it is a single-choice or multiple-choice question-and-answer template. The question information corresponding to the attribute label constitutes the question stem of the question-and-answer template, and all the option values ​​corresponding to the attribute label constitute the options of the question-and-answer template.

[0018] The question and answer template is: Please answer the multiple-choice question based on the above content - attribute item, which of the following options is correct? - option list.

[0019] The training database of the meta-model includes the training data in the meta-model format, the fourth question-answering dialogue data and the second question-answering dialogue data.

[0020] Among them, the format of the first question and answer dialogue data in step one is: attribute item-medical record document content-question and answer stem-options.

[0021] Wherein, the output template in step 4 is a preset template;

[0022] The format of the output template is: question answer of attribute item - answer conversion template: "Therefore, the correct option is: ".

[0023] The electronic medical record includes all diagnostic records of a single user's single medical course.

[0024] Among them, when the electronic medical record to be structured is a new disease, the model training is first performed to obtain the large language model and training data for medical record structured questions and answers, and then the medical record is structured; when the electronic medical record to be structured is an existing disease, the medical record is directly structured.

[0025] (III) Beneficial effects: The present invention provides an electronic medical record structuring method based on a large language model, which converts each medical record into a question-and-answer dialogue for the electronic medical record structuring goal, and transmits it to the large language model of medical record structured question and answer for question and answer. The most relevant first threshold medical record texts are searched for a question and answer item as context information to enhance the answer accuracy of the large language model, and finally outputs an output text consistent with the structuring standard, which effectively improves the performance of medical record document structuring. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 It is a flow chart of an electronic medical record structuring method based on a large language model of the present invention;

[0027] Figure 2 It is a flow chart of a method for training an electronic medical record structured model based on a large language model of the present invention;

[0028] Figure 3 It is a result display box of the second dialogue question and answer data of the large language model of the electronic medical record structured question and answer of the degree of coronary artery stenosis in an embodiment of the present invention;

[0029] Figure 4 It is a result display box of the third question-and-answer dialogue data of the large language model for structured question-and-answer of electronic medical records on the degree of coronary artery stenosis in an embodiment of the present invention;

[0030] Figure 5 It is an output text display box of the large language model of the electronic medical record structured question and answer of the degree of coronary artery stenosis in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] The present invention is further described in detail below in conjunction with preferred embodiments. More details are elaborated in the following description to facilitate a full understanding of the present invention. However, the present invention can obviously be implemented in a variety of other ways different from the description. Those skilled in the art can make similar generalizations and deductions based on actual application situations without violating the connotation of the present invention. Therefore, the protection scope of the present invention should not be limited by the content of this specific embodiment.

[0032] The accompanying drawings are schematic diagrams of embodiments of the present invention. It should be noted that the drawings are only examples and are not drawn to scale, and should not be used to limit the actual protection scope of the present invention.

[0033] Before introducing the solution of the present invention, the English words involved in this article are first explained:

[0034] Large Language Model: (LLM) refers to a generative deep neural network model based on the Transformer structure, which has super semantic understanding and multi-round dialogue capabilities. Common large language models include GPT4 and large language models developed by other companies.

[0035] In-Context Learning (ICL) refers to the fact that in unsupervised pre-training, the language model has acquired a generalization and pattern recognition ability, which can quickly adapt to or recognize the task to be done. There is no need to fine-tune the model weights to keep them fixed. Just show the pre-trained model some input-output examples and it will learn to do something completely new. The most important things are the input distribution, output space, and prompt format.

[0036] Chain-of-thought (CoT) is an improved prompting strategy for improving the performance of LLM in complex reasoning tasks, such as arithmetic reasoning, common sense reasoning, and symbolic reasoning. Instead of simply constructing prompts with input-output pairs like ICL, CoT incorporates intermediate reasoning steps that can introduce the final output into the prompt. Compared with traditional contextual learning, Chain-of-thought has more intermediate derivation prompts.

[0037] The present invention discloses an electronic medical record structuring method based on a large language model, which is used to convert the electronic medical records of the same patient in a course of illness into structured data in the form of questions and answers, wherein the electronic medical record includes all diagnostic records of a single user and a single course of illness, including admission records, discharge records, examination reports, medical orders, etc. The electronic medical record has a serial number. It is worth noting that the same course of illness uses the same serial number. The present invention can determine an electronic medical record for the same course of illness in a medical database by the serial number of the electronic medical record.

[0038] The present invention provides a method for structuring electronic medical records based on a large language model, such as Figure 1 As shown, the specific steps include:

[0039] Step 1: Obtain the electronic medical record to be structured, and convert the electronic medical record to be structured into first question-and-answer dialogue data according to the meta-model attribute label.

[0040] The attribute tag is a unique identifier of the meta-model, and the attribute tag indicates the type of disease corresponding to the electronic medical record to be structured. The first question-and-answer dialogue data corresponds the attribute items of the disease to the electronic medical record to be structured, and completely includes all the attribute items of the disease. The format of the first question-and-answer dialogue data is: attribute item-medical record document content-question and answer stem-option.

[0041] Step 2: According to each attribute label of the meta-model, a first threshold number of similar electronic medical records and second question-answering dialogue data corresponding to the similar electronic medical records are searched from the training database as context information.

[0042] The similar electronic medical records refer to the use of representation learning methods to convert the attribute items corresponding to all medical record documents into vector values ​​for representation, and the similarity between medical records is judged by calculating the distance between the vectors. The smaller the distance between the vectors, the higher the similarity between the medical records; otherwise, the lower the similarity between the medical records. Therefore, the smaller the distance between the vectors of the similar electronic medical records, the closer the vectors are, and the more similar the medical record documents are. Therefore, the similar electronic medical record of the present invention refers to an electronic medical record in the training database that is similar to the medical record document when the vector value between the medical record document to be structured is less than or equal to the second threshold.

[0043] The representation learning method includes but is not limited to CNN, LSTM, and BERT algorithms, as long as it can convert medical records into vector values.

[0044] Step three: input the first question-and-answer dialogue data and context information into the large language model of medical record structured question-and-answer to obtain third question-and-answer dialogue data, and the third question-and-answer dialogue data is used as the medical record structured question-and-answer result corresponding to the input electronic medical record to be structured.

[0045] The large language model for structured medical record questions and answers needs to be trained before medical record structuring. Therefore, the electronic medical record structuring method based on the large language model of the present invention also includes a model training method, such as Figure 2 As shown, the specific steps include:

[0046] Step 1: Establish a meta-model of structured targets for disease types. The meta-model includes a set of key-value pairs, where the key is an attribute item and the value is a numerical value corresponding to the attribute item.

[0047] The attribute items of the metamodel described in the present invention must have the following information: attribute label, the question corresponding to the attribute label, all option values ​​corresponding to the attribute label, and whether it is multiple choice. Based on all the attribute item information of the metamodel, a question and answer template is constructed, wherein the multiple choice information in the attribute item determines whether it is a single choice or multiple choice question and answer template; the attribute label is the unique identifier of the metamodel corresponding to the disease type; the question information corresponding to the attribute label constitutes the question stem information of the question and answer template, and all option values ​​corresponding to the attribute label constitute the option part of the question and answer template. Attribute labels are attribute words in the medical field, such as body temperature, past medical history, long-term medication, etc.

[0048] The format of the question and answer template of a preferred embodiment of the present invention is: Please answer the single-choice question based on the above content. Which of the following options is correct? - List of options.

[0049] The correspondence between the attribute items of the metamodel and the medical record documents is established according to the attribute labels of the metamodel.

[0050] Step 2: prepare training data in a meta-model format, and convert the training data in the meta-model format into fourth question-answering dialogue data according to the attribute labels of the meta-model.

[0051] The format of the fourth question-and-answer dialogue data is: medical record document content - question-and-answer stem - option - answer.

[0052] Step 3: Use the large language model fine-tuned by the dialogue data as the large language model for structured question and answer of medical records, input the training data of the fourth question and answer dialogue data in the meta-model format into the large language model for structured question and answer of medical records, and output the training data as the second question and answer dialogue data. The training data in the meta-model format, the fourth question and answer dialogue data, and the second question and answer dialogue data constitute the training database of the meta-model.

[0053] The large language model of the structured question and answer of medical records of the present invention is a multi-layer neural network model. The training of the large language model is to obtain the weight value of each layer of the neural network, and the large language model can be ChatGPT, GLM, PaLM, LLama, etc., without specific limitation. These large language models all provide a fine-tuning (Finetune) interface. The developer inputs the input and output data of the question and answer dialogue as parameters to call the fine-tuning interface. The fine-tuning will adjust the parameters of a few layers of neural networks in the model, and the parameters of other layers remain unchanged. In this way, after fine-tuning the limited field task-related data, the large language model can be adjusted to a state that is better at structured question and answer of medical records.

[0054] More specifically, in step 3, Chain-of-thoughts is used to improve the accuracy of large language model question answering. By inserting CoT, each context information<input,output> Expand to<input,CoT,output> .

[0055] The present invention first uses zero-shot thinking chain (Zero-shot CoT) to add a text "Let's think step by step." after the question of the large language model of the medical record structured question and answer input, so as to stimulate the large language model of the medical record structured question and answer to think about the ideas of question and answer questions; then uses few-shot thinking chain (Few-shot CoT) to search the training database for the first threshold second question and answer dialogue data that is most similar to the current medical record for each attribute item question, and add them to the back of each attribute label question as context information to stimulate the large language model of the medical record structured question and answer to think and output the correct answer.

[0056] The Few-shot Chain of Thought (Few-shot CoT) is a special case of In Context Learning, which integrates the CoT reasoning steps as context and gives LLM a set of demonstration examples in context learning.<input,output> Add a test input to the end of the examples and allow the LLM to make predictions by conditioning on only the demonstration examples.

[0057] Few-shot refers to giving a small number of samples to the model during inference, but not allowing weight updates. For a training data set, Few-shot has context and examples. The way Few-shot works is to provide multiple samples and then expect the model to generate corresponding results. Usually the number of samples is set in the range of 2 to 64. The main advantage of Few-shot is that it greatly reduces the demand for task-specific data.

[0058] The (input, output> refers to an <input, output> pair.

[0059] Said<input,CoT,output> It refers to inserting a thinking chain into the <input, output> pair to trigger LLM thinking.

[0060] The zero-shot thought chain (Zero-Shot) does not allow samples to be provided, only natural language instructions describing the task are given. The zero-shot thought chain method provides the greatest convenience, robustness, and the possibility of avoiding false correlations, but it is also the most challenging setting. However, at least in some cases, Zero-shot is the closest to how humans perform tasks, and may know what to do based on text instructions alone.

[0061] In the step three, each question is converted into: Few-shot CoT-medical record document content-question and answer stem-Zero-Shot CoT-option according to the first question and answer dialogue data and context information.

[0062] Step 4: According to the output template, the third question-answer dialogue data is converted into an output text that conforms to the structured standard.

[0063] The output template is a preset template, which is set in the electronic medical record structuring method based on a large language model. The format of the output template is: attribute item question answer-answer conversion template: "Therefore, the correct option is: (output the label corresponding to the option)". By inputting the third question and answer dialogue data as the question and answer result into the large language model of the electronic medical record structured question and answer, the output template converts the third question and answer dialogue data into an electronic medical record structured answer consistent with the structured standard.

[0064] The following is described in conjunction with a specific embodiment. First, a meta-model is established based on the existing coronary artery stenosis degree classification as an attribute label, and the questions corresponding to the coronary artery stenosis degree classification and the option values ​​corresponding to the coronary artery stenosis degree classification are respectively input into the meta-model in the form of key-value pairs, for example, <key body temperature, value below 37.4 degrees>, <key coronary artery stenosis degree classification, value grade III>, and a one-to-one correspondence is established between the attribute items of the meta-model and the medical record documents to be structured.

[0065] According to all the input attribute information, a question-answering template is constructed: "Please answer the multiple-choice question based on the above content. Which of the following options is correct for the question corresponding to the attribute label? All option values ​​corresponding to the attribute label." Thus, all the attribute items of the meta-model are converted into a multiple-choice question, and the electronic medical record to be structured is converted into the fourth question-answering dialogue data. The meta-model forms a meta-model of the electronic medical record structure of the degree of coronary artery stenosis.

[0066] Then, the existing medical records of the degree of coronary artery stenosis are used to form training data, and the training data is converted into fourth question-answering dialogue data, and the fourth question-answering dialogue data is input into the electronic medical record structured meta-model of the degree of coronary artery stenosis, and the electronic medical record structured meta-model of the degree of coronary artery stenosis is fine-tuned through multiple fourth question-answering dialogue data to form a large language model of structured questions and answers of electronic medical records of the degree of coronary artery stenosis.

[0067] The electronic medical record to be structured is converted into the first question-answering dialogue data in the form of the fourth question-answering dialogue data, and then the first question-answering dialogue data is input into the large language model of the structured question-answering of the electronic medical record of the degree of coronary artery stenosis, such as Figure 3As shown, the second question-answering dialogue data is output:

[0068] Based on the above content, please answer the multiple-choice question "LAD left anterior descending branch - coronary artery stenosis grade". Which of the following options is correct?

[0069] a: Grade I lesions: the lumen area is reduced by 1%-25%;

[0070] b: Grade II lesions: the lumen area is reduced by 26%-50%;

[0071] c: Grade III lesions: the lumen area is reduced by 51%-75%;

[0072] d: Grade IV lesions: the lumen area is reduced by 76%-100%;

[0073] e: Unable to determine or no record.

[0074] The electronic medical record document to be structured is converted into the form of the second question-answering dialogue data, and according to the attribute label of each attribute item, the first threshold number of similar electronic medical records of coronary artery stenosis and the second question-answering dialogue data corresponding to the first threshold number of similar electronic medical records of coronary artery stenosis are found from the training database composed of the meta-model training data, the fourth question-answering dialogue data and the second question-answering dialogue data. The first threshold number of similar electronic medical records of coronary artery stenosis and the corresponding second question-answering dialogue data are input into the large language model of structured question-answering of the electronic medical record of coronary artery stenosis, such as Figure 4 As shown, the results are expressed as:

[0075] Similar medical record 1 document content

[0076] question

[0077] Medical Records 1 Answer

[0078] Similar medical record 2 document content

[0079] question

[0080] Medical Records 2 Answers

[0081] …

[0082] Current medical record document content

[0083] Based on the above, please answer the multiple-choice question "LAD left anterior descending branch - coronary artery stenosis grade". Which of the following options is correct? Let's think about it step by step.

[0084] a: Grade I lesions: the lumen area is reduced by 1%-25%;

[0085] b: Grade II lesions: the lumen area is reduced by 26%-50%;

[0086] c: Grade III lesions: the lumen area is reduced by 51%-75%;

[0087] d: Grade IV lesions: the lumen area is reduced by 76% to 100%;

[0088] e: Unable to determine or no record.

[0089] The large language model of the electronic medical record structured question and answer on the degree of coronary artery stenosis outputs the third question and answer dialogue data: "LAD left anterior descending branch - coronary artery stenosis degree grading": "Grade II lesions: lumen area is reduced by 26%-50%;".

[0090] The third question-and-answer dialogue data is re-input into the large language model of the medical record structured question and answer, and the large language model of the medical record structured question and answer converts the third question-and-answer dialogue data into a structured and consistent standard output text according to a preset output template, such as Figure 5 As shown: According to the given interventional treatment results, the stenosis of the left anterior descending branch of LAD is 30%-50%, which belongs to the grade II lesion with a 26%-50% reduction in lumen area. Therefore, a, c, d, and e are wrong answers, and the correct answer is b.

[0091] At this point, the electronic medical record of the coronary artery stenosis degree to be structured is converted into a structured and consistent standard output text, and the text is obtained from the selection and spliced ​​into a key-value pair with the attribute item, thereby completing the entire electronic medical record structuring process, and the answer can be easily parsed. In addition, a large structured model of coronary artery stenosis degree decomposition is trained, which is convenient for direct input and use next time, and improves the efficiency of structuring.

[0092] It should be noted that when the electronic medical record to be structured is a new disease, the model training is first performed according to steps 1 to 3 to obtain a large language model and training database for structured medical record questions and answers, and then the medical record is structured through steps 1 to 4. When the electronic medical record to be structured is an existing disease, the medical record is directly structured through steps 1 to 4 to obtain a standard output text with consistent structure.

[0093] The present invention constructs a medical record structuring method based on a large language model, so that each electronic medical record is converted into a question-and-answer dialogue for the medical record structuring goal, and is passed into the large language model for question-and-answering. The most relevant medical record texts are searched for a question-and-answer item as context information to enhance the accuracy of the answer of the large model. Finally, the answer of the large model is converted into a consistent expression method using an output template, which effectively improves the performance of medical record document structuring. The structuring of electronic medical records helps to standardize and integrate medical data, promote information sharing between different medical institutions, and help reduce the situation where patients need to repeatedly provide the same information during the medical treatment process, thereby improving medical efficiency.

[0094] The above content is an explanation of the preferred embodiments of the present invention, which can help those skilled in the art to more fully understand the technical solution of the present invention. However, these embodiments are merely illustrative, and it cannot be determined that the specific implementation methods of the present invention are limited to the description of these embodiments. For ordinary technicians in the technical field to which the present invention belongs, without departing from the concept of the present invention, several simple deductions and transformations can be made, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A method for structuring electronic medical records based on a large language model, which converts medical record documents into structured electronic medical records, characterized in that: The following steps are included: Step 1: obtaining an electronic medical record to be structured, and converting the electronic medical record to be structured into first question-answering dialogue data according to a metamodel attribute label; Step 2: According to each attribute label of the meta-model, a first threshold number of similar electronic medical records and second question-answering dialogue data corresponding to the first threshold number of similar electronic medical records are searched from the training database as context information; Step 3: Input the first question-answering dialogue data and context information into the large language model of medical record structured question-answering to obtain third question-answering dialogue data, and use the third question-answering dialogue data as the medical record structured question-answering result corresponding to the input electronic medical record to be structured; Step 4: According to the output template, the third question-answer dialogue data is converted into an output text that conforms to the structured standard.

2. The electronic medical record structuring method based on a large language model according to claim 1, characterized in that: Model training is performed before the electronic medical record is structured, including the following steps: Step 1: Establish a metamodel of structured targets for disease types; establish a correspondence between the attribute items of the metamodel and the medical record documents according to the attribute labels of the metamodel; Step 2: preparing training data in a meta-model format, and converting the training data in the meta-model format into fourth question-answering dialogue data according to attribute labels of the meta-model; Step 3: Use the large language model fine-tuned by the dialogue data as the large language model for medical record structured question and answer, input the training data of the fourth question and answer dialogue data in the meta-model format into the large language model for medical record structured question and answer, and output it as the second question and answer dialogue data.

3. The method for structuring electronic medical records based on a large language model according to claim 2, characterized in that: The metamodel attribute tag is a unique identifier of the metamodel, and the attribute tag is an attribute word in the medical field.

4. The method for structuring electronic medical records based on a large language model according to claim 2, characterized in that: The attribute items of the meta-model include attribute labels, questions corresponding to the attribute labels, all option values ​​corresponding to the attribute labels, and whether multiple selection is allowed; A question-and-answer template is constructed based on all the attribute item information of the meta-model. The multiple-choice information of the attribute item determines whether it is a single-choice or multiple-choice question-and-answer template. The question information corresponding to the attribute label constitutes the question stem of the question-and-answer template, and all the option values ​​corresponding to the attribute label constitute the options of the question-and-answer template.

5. The method for structuring electronic medical records based on a large language model according to claim 4, characterized in that: The question and answer template is: Please answer the multiple-choice question based on the above content - attribute item, which of the following options is correct? - option list.

6. The method for structuring electronic medical records based on a large language model according to claim 2, characterized in that: The training database of the meta-model includes the training data in the meta-model format, the fourth question-answering dialogue data and the second question-answering dialogue data.

7. The method for structuring electronic medical records based on a large language model according to claim 1, characterized in that: The format of the first question-and-answer dialogue data in step one is: attribute item-medical record document content-question-and-answer stem-options.

8. The method for structuring electronic medical records based on a large language model according to claim 1, characterized in that: The output template in step 4 is a preset template; The format of the output template is: question answer of attribute item - answer conversion template: "Therefore, the correct option is: ".

9. The method for structuring electronic medical records based on a large language model according to claim 1, characterized in that: The electronic medical record includes all diagnostic records of a single user's single medical course.

10. The electronic medical record structuring method based on a large language model according to claim 1, characterized in that: When the electronic medical record to be structured is a new disease, model training is first performed to obtain a large language model and training data for structured questions and answers of medical records, and then the medical records are structured; when the electronic medical record to be structured is an existing disease, the medical records are directly structured.

Images