Automatic coding method and system for surgical operation records based on large language model

Abstract

The invention provides an automatic coding method and system for a surgical operation record based on a large language model. The method comprises the following steps: inputting the surgical operation record; performing modal classification on the surgical operation records, and uniformly processing the surgical operation records into text modals by adopting different preprocessing modes according to different modals; according to an operation type corresponding to the input operation operation record, a prompt statement for automatically extracting an operation code guided by evidence is constructed; the processed text is sent into an automatic coding big language model of the operation record capable of being associated with the evidence, and the big language model extracts and outputs a standard operation name contained in the text; and according to the established mapping dictionary, obtaining a code corresponding to a standard surgical operation noun generated by the large language model. And by utilizing the characteristic of strong language ability of the large language model, the training cost and the secondary training cost are greatly reduced, the model can reach a practical level only by a small amount of training data, and the efficiency and the accuracy of surgical operation record coding work are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical event recognition, and in particular relates to a method and system for automatically encoding surgical operation records based on a large language model. Background Art

[0002] In the medical field, the standardization of medical terminology is an important demand in the medical field, including the standardization of disease terms, surgical operation terms, and drug names, which is of great significance to medical research, medical insurance payment, and public hospital performance evaluation. However, due to the differences in the terms used by different medical institutions and clinicians, the diagnosis and treatment information in various types of clinical medical data generally has the problem of non-standard and non-uniform terms, making the archiving and classification of data very complicated and difficult, consuming a lot of manpower and material resources, and it is difficult to mine or analyze it with the help of the latest information technologies such as big data and artificial intelligence.

[0003] In order to solve this problem, automatic coding technology has become a new solution in the scenario of medical record digitization. With the development of artificial intelligence technology, more and more people are trying to apply natural language processing technology to automatic coding tasks. However, model training based on traditional deep learning methods requires a large amount of manually labeled data, which is costly and difficult to achieve universality. With the development of large language model technology, people have seen the application prospects of this technology in the medical field. However, there is currently no mature method and system on the market that converts surgical operation records into standard surgical operation codes based on large language models. Summary of the invention

[0004] The purpose of the present invention is to provide a method and system for automatic encoding of surgical operation records based on a large language model, aiming to solve the problems of high training cost and low utilization efficiency in automatic encoding of surgical operation records in the prior art.

[0005] The present invention is implemented as follows: a method for automatically encoding surgical operation records based on a large language model, the method comprising the following steps: Step S1: inputting surgical operation records; Step S2: classify the surgical operation records by modality, adopt different pre-processing methods according to different modalities, and uniformly process them into text modalities; Step S3: constructing a prompt statement for automatic extraction of surgical codes guided by evidence according to the type of surgery corresponding to the input surgical operation record; Step S4: sending the processed text to the surgical record automatic encoding large language model that can be associated with evidence, and the large language model extracts and outputs the standard surgical operation names contained in the text; Step S5: According to the formulated mapping dictionary, the codes corresponding to the standard surgical operation nouns generated by the large language model are obtained.

[0006] A further technical solution of the present invention is: in step S2, if the modality of the surgical operation record is an image modality, image text recognition is performed; if the modality of the surgical operation record is a sound modality, voice text recognition is performed.

[0007] A further technical solution of the present invention is: in step S2, noise cleaning is performed on the surgical operation records in text mode and the surgical operation records in text mode converted from other modes.

[0008] A further technical solution of the present invention is: in step S3, after the surgical operation record is input, the surgical type corresponding to the input surgical operation record is obtained through manual selection or automatic recognition by a large language model, and an evidence-guided surgical coding automatic prompt sentence is constructed according to the surgical type.

[0009] A further technical solution of the present invention is: the training process of the large language model includes data labeling, instruction augmentation and model training; Data annotation: After obtaining the original surgical record text, manual annotation is performed. The annotation content includes annotating the original surgical record fragments containing standard surgical operation nouns and the corresponding standard surgical operation nouns; Instruction augmentation: For the labeled data, perform multiple types of data augmentation to generate data that can be directly used for large language model training; Model training: Use an open source or closed source large language model as the base, and use the above instructions to fine-tune it so that the large language model can acquire the ability to automatically encode surgical operation records.

[0010] A further technical solution of the present invention is that the instruction augmentation includes the following types: Based on the annotated data, create a table of common noun expressions; Based on the original surgical record text, extract the original text fragments containing standard surgical operation nouns; Generate standard surgical operation nouns according to the extracted original text segments containing standard surgical operation nouns; Directly generate standard surgical operation nouns based on the original surgical record text; Generate text fragments in the original text according to the original surgical record text and the standard surgical operation nouns contained in the text; Generate standard surgical operation nouns and their corresponding segments based on the original surgical record text; Generate standard surgical operation terms and their explanations based on the original surgical record text; Determine the type of surgery based on the original surgical record text; Generate standard surgical operation terms based on the original surgical record text and the alternative table corresponding to the surgical type; The standard surgical operation nouns are generated according to the extracted original text segments containing the standard surgical operation nouns and the alternative table corresponding to the surgical type.

[0011] A further technical solution of the present invention is: while using the large language model, the erroneous results that occur during the use process are stored, and these erroneous examples are used to correct the large language data using a retrieval enhanced generation method.

[0012] A further technical solution of the present invention is: in step S5, the mapping dictionary refers to standard surgical operation terms and their corresponding codes issued by national functional departments or professional associations.

[0013] Another object of the present invention is to provide a surgical operation record automatic encoding system based on a large language model, which is used to implement a surgical operation record automatic encoding method based on a large language model, comprising: The annotation module is used to manually annotate a number of original surgical record texts, and properly store the original surgical operation record fragments annotated with standard surgical operation nouns, as well as the corresponding standard surgical operation nouns; An instruction data augmentation module, used to generate an instruction training data set according to the labeled data of the labeling module, and perform data augmentation on the instruction training data set; The training module is used to carry out instruction training on the large language model; The deployment module is used to deploy the large language model and make it accessible through the API. It records new situations that did not appear in the training process or situations where the generation is incorrect, and corrects the large language model in a way that enhances the generation through retrieval. The code mapping module is used to map the results generated by the large language model to obtain standard codes; The user interaction module is used to provide coding services to doctors and coders after the large language model is deployed; The database module is used to provide data support for the annotation module, instruction data augmentation module, training module, deployment module, code mapping module and user interaction module.

[0014] A further technical solution of the present invention is: the user interaction module includes an input module and an output module, the input module is connected to the deployment module and is used by doctors and coders to input surgical operation records, and the output module is connected to the code mapping module to output codes corresponding to surgical operation nouns.

[0015] The beneficial effects of the present invention are as follows: based on natural language processing technology, the present invention performs medical event recognition on original clinical surgical operation records, constructs a large language model instruction training data set, and conducts instruction training on the large language model; in the practical stage, an external knowledge base is also designed to handle new situations that did not appear in the training process. This method takes advantage of the powerful language capabilities of the large language model, greatly reducing the training cost and the cost of secondary training. Only a small amount of training data is needed to make the model reach a practical level, significantly enhancing practicality and improving the efficiency and accuracy of surgical operation record coding. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a flow chart of the automatic encoding method of the present invention; Figure 2 This is an example of the annotation operation of the coronary artery stent implantation operation record text of the present invention; Figure 3 It is a relationship diagram of various modules of the automatic encoding system of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments.

[0018] Embodiment 1: Taking the surgical operation record of “coronary artery stent implantation” as an example, the automatic encoding method of surgical operation records based on a large language model is explained.

[0019] Step S1: Input surgical operation records.

[0020] Step S2: classify the surgical operation records into different modalities, adopt different pre-processing methods according to different modalities, and uniformly process them into text modalities.

[0021] In the pre-processing of the input surgical operation records, image text recognition is performed on the surgical operation records in image mode; speech recognition is performed on the surgical operation records in sound mode; and noise removal is performed on the surgical operation records in text mode and the surgical operation records in text mode converted from other modes.

[0022] Step S3: constructing a prompt statement for automatic extraction of surgical codes guided by evidence according to the type of surgery corresponding to the input surgical operation record; Specifically, after the surgical operation record is input, the type of surgery corresponding to the input surgical operation record can be obtained through manual selection or automatic recognition by the large language model, and an evidence-guided surgical coding automatic prompt sentence can be constructed according to the type of surgery.

[0023] Here is an example of the process of constructing evidence-based surgical coding automatic prompts based on the type of surgery: Template: "The following is a surgical operation record of {surgical type name}. Please generate standard surgical operation nouns and their corresponding segments based on the original surgical record text and the alternative coding table corresponding to the surgery {alternative coding sequence corresponding to the surgical type name}. \n{surgical operation record text}".

[0024] Name of the procedure: Coronary artery stenting; Alternative coding sequences corresponding to the names of the procedures: ['coronary intravascular ultrasound (IVUS)', ..., 'single-catheter coronary angiography', ..., 'cardioversion']; Text of surgical operation record: The patient...punctured the right femoral artery...underwent CAG with a 5F Tig angiography catheter...implanted in the middle segment of the anterior descending branch...returned to the ward.

[0025] Final prompt sentence: The following is a surgical operation record of coronary artery stenting. Please generate standard surgical operation terms and their corresponding fragments based on the original surgical record text and the alternative coding table corresponding to the operation ['coronary intravascular ultrasound (IVUS)',...,'single catheter coronary angiography',...,'cardioversion']. \nThe patient...punctured the right femoral artery...5F Tig angiography catheter for CAG...implanted in the middle of the anterior descending branch...returned to the ward.

[0026] Step S4: The processed text is sent to a large language model for automatic encoding of surgical records that can be associated with evidence. The large language model extracts and outputs the standard surgical operation names contained in the text.

[0027] Specifically, it includes using a specific training method to train a large language model specifically for automatic encoding of surgical operation records, and adding retrieval enhancement generation capabilities when deployed.

[0028] In the large language model encoding, a specific training method is used to train a large language model specifically used for automatic encoding of surgical operation records, specifically including data labeling, instruction augmentation, model training and other parts.

[0029] In the large language model trained by specific training methods for automatic encoding of surgical operation records, including data annotation, after obtaining the original surgical record text, these surgical records are manually annotated. The annotation content includes annotating the original surgical operation record fragments containing standard surgical operation nouns, and the corresponding standard surgical operation nouns.

[0030] Let’s take an example to illustrate the data annotation process. Figure 2As shown in the figure, an excerpt of a coronary artery stent implantation surgical operation record text is shown, and part of the annotation process is demonstrated. The original text is a text describing the surgical process. During the annotation process, segments containing standard surgical operation nouns are selected from the original text. A segment can contain multiple standard surgical operation nouns, and a standard surgical operation noun can be obtained from one or more surgical segments.

[0031] Instruction augmentation: For the labeled data, perform multiple types of data augmentation to generate data that can be directly used for large language model training: An example to illustrate the instruction augmentation process: Text of surgical operation record: The patient...punctured the right femoral artery...underwent CAG with a 5F Tig angiography catheter...implanted in the middle segment of the anterior descending branch...returned to the ward.

[0032] Instruction augmentation includes the following ten types: Type 1: Create a table of common noun expressions based on the labeled data; Example: Cardioversion → This standard surgical term (name) is usually reflected in the surgical record in the form of "electrical cardioversion therapy was given"; Type 2: Extract the original text segments containing standard surgical operation nouns based on the original surgical record text; Example: Patient...puncture right femoral artery...5F Tig angiography catheter for CAG...implanted in the middle of the left anterior descending artery...return to the ward →5F Tig angiography catheter for CAG\n...; Type 3: Generate standard surgical operation nouns based on the extracted original text segments containing standard surgical operation nouns; Example: CAG with 5F Tig angiography catheter → single-catheter coronary angiography; Type 4: Directly generate standard surgical operation terms based on the original surgical record text; Example: Patient...puncture of right femoral artery...5F Tig angiography catheter for CAG...implantation in the middle of the left anterior descending artery...return to the ward → single catheter coronary angiography\n...; Type 5: Generate text fragments from the original surgical record text and the standard surgical operation nouns contained in the text; Example: Text: The patient...punctured the right femoral artery...underwent CAG with a 5F Tig angiography catheter...implanted in the middle of the anterior descending artery...returned to the ward\nCoding result: ['single-catheter coronary angiography',...]→CAG with a 5F Tig angiography catheter\n...; Type 6: Generate standard surgical operation nouns and their corresponding segments based on the original surgical record text; Example: The patient...punctured the right femoral artery...underwent CAG with a 5F Tig catheter...was implanted in the middle of the left anterior descending artery...and returned to the ward → {"single-catheter coronary angiography":["5F Tig catheter for CAG"],...}; Type 7: Generate standard surgical operation terms and their explanations based on the original surgical record text; Example: The patient...punctured the right femoral artery...5F Tig angiography catheter for CAG...implanted in the middle of the left anterior descending artery...returned to the ward → The text contains the standard operation name:\n5F Tig angiography catheter for CAG\n...Among them,\n"5F Tig angiography catheter for CAG", the corresponding standard operation name is: single catheter coronary angiography\n...; Type 8: Determine the type of surgery based on the original surgical record text; Example: The patient...punctured the right femoral artery...underwent CAG with a 5F Tig catheter...implanted in the middle of the anterior descending branch...returned to the ward → coronary artery stenting; Type 9: Generate standard surgical operation terms based on the original surgical record text and the candidate table obtained by pre-statistication corresponding to the surgical type; Example: Text: The patient...punctured the right femoral artery...5F Tig angiography catheter was inserted for CAG...implanted in the middle of the left anterior descending artery...returned to the ward\nAlternative table: ['Coronary intravascular ultrasound (IVUS)',...,'Single-catheter coronary angiography',...,'Cardioversion'] → Single-catheter coronary angiography\n...; Type 10: Generate standard surgical operation nouns based on the extracted original text segments containing standard surgical operation nouns and the candidate list obtained in advance according to the statistics of the surgical type; Example: Text fragment: 5F Tig angiography catheter for CAG\nAlternative table: ['Coronary intravascular ultrasound (IVUS)', …, 'Single-catheter coronary angiography', …, 'Cardioversion'] → Single-catheter coronary angiography\n ….

[0033] Model training refers to using an open source or closed source large language model as a base, and fine-tuning it using the above instructions to enable the large language model to acquire the ability to perform the task. During deployment, a search enhancement generation function is added, and the error results that occur during use are stored. These error examples are used to further enhance the ability of the large language model through search enhancement generation.

[0034] Specifically, after receiving the surgical operation record text, the large language model retrieves words that are similar to the text or text fragments from the database, and sends them to the large language model as part of the large language model prompt words.

[0035] Step S5: Obtain the codes corresponding to the standard surgical operation nouns generated by the large language model according to the mapping dictionary, wherein the mapping dictionary can be manually formulated.

[0036] Specifically, the standard surgical operation nouns generated by the large language model are mapped according to the relevant specification documents to obtain the corresponding codes. The mapping process is illustrated by an example: Single-catheter coronary angiography → 88.5500; Cardioversion → 99.6201.

[0037] Embodiment 2: The present invention also provides a surgical operation record automatic encoding system based on a large language model, which is used to implement a surgical operation record automatic encoding method based on a large language model, comprising: The annotation module is used to manually annotate a number of original surgical record texts, and properly store the original surgical operation record fragments annotated with standard surgical operation nouns, as well as the corresponding standard surgical operation nouns; An instruction data augmentation module, used to generate an instruction training data set according to the labeled data of the labeling module, and perform data augmentation on the instruction training data set; The training module is used to carry out instruction training on the large language model; The deployment module is used to deploy the large language model and make it accessible through the API. It records new situations that did not appear in the training process or situations where the generation is incorrect, and corrects the large language model in a way that enhances the generation through retrieval. The code mapping module is used to map the results generated by the large language model to obtain standard codes; The user interaction module is used to provide coding services to doctors and coders after the large language model is deployed; The database module is used to provide data support for the annotation module, instruction data augmentation module, training module, deployment module, code mapping module and user interaction module.

[0038] The user interaction module includes an input module and an output module. The input module is connected to the deployment module and is used by doctors and coders to input surgical operation records. The output module is connected to the code mapping module and outputs codes corresponding to surgical operation nouns.

[0039] In the system, the user interaction module and the database module are used to implement the operations of step S1 and step S2, the labeling module, instruction data augmentation module, training module, deployment module and database module are used to implement the operations of step S3 and step S4, and the code mapping module and the user interaction module are used to implement the operation of step S5.

[0040] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for automatic encoding of surgical operation records based on a large language model, characterized in that: The method comprises the following steps: Step S1: inputting surgical operation records; Step S2: classify the surgical operation records into different modalities, adopt different pre-processing methods according to different modalities, and uniformly process them into text modalities; Step S3: constructing a prompt statement for automatic extraction of surgical codes guided by evidence according to the type of surgery corresponding to the input surgical operation record; Step S4: sending the processed text to the surgical record automatic encoding large language model that can be associated with evidence, and the large language model extracts and outputs the standard surgical operation names contained in the text; Step S5: According to the formulated mapping dictionary, the codes corresponding to the standard surgical operation nouns generated by the large language model are obtained.

2. The method for automatic encoding of surgical operation records based on a large language model according to claim 1, characterized in that: In step S2, if the modality of the surgical operation record is an image modality, image text recognition is performed; if the modality of the surgical operation record is a sound modality, voice text recognition is performed.

3. The method for automatic encoding of surgical operation records based on a large language model according to claim 2, characterized in that: In step S2, noise cleaning is performed on the surgical operation records in text mode and the surgical operation records in text mode converted from other modes.

4. The method for automatic encoding of surgical operation records based on a large language model according to claim 1, characterized in that: In step S3, after the surgical operation record is input, the surgical type corresponding to the input surgical operation record is obtained through manual selection or automatic recognition by a large language model, and an evidence-guided surgical coding automatic prompt sentence is constructed according to the surgical type.

5. The method for automatic encoding of surgical operation records based on a large language model according to claim 1, characterized in that: The training process of the large language model includes data labeling, instruction augmentation and model training; Data annotation: After obtaining the original surgical record text, manual annotation is performed. The annotation content includes annotating the original surgical operation record fragments containing standard surgical operation nouns and the corresponding standard surgical operation nouns; Instruction augmentation: For the labeled data, perform multiple types of data augmentation to generate data that can be directly used for large language model training; Model training: Use an open source or closed source large language model as the base, and use the above instructions to fine-tune it so that the large language model can acquire the ability to automatically encode surgical operation records.

6. The method for automatic encoding of surgical operation records based on a large language model according to claim 5, characterized in that: Instruction augmentations include the following types: Based on the annotated data, create a table of common noun expressions; Based on the original surgical record text, extract the original text segments containing the nouns of standard surgical operations; Generate standard surgical operation nouns according to the extracted original text segments containing standard surgical operation nouns; Directly generate standard surgical operation nouns based on the original surgical record text; Generate text fragments in the original text according to the original surgical record text and the standard surgical operation nouns contained in the text; Generate standard surgical operation nouns and their corresponding segments based on the original surgical record text; Generate standard surgical operation terms and their explanations based on the original surgical record text; Determine the type of surgery based on the original surgical record text; Generate standard surgical operation terms based on the original surgical record text and the alternative table corresponding to the surgical type; The standard surgical operation nouns are generated according to the extracted original text segments containing the standard surgical operation nouns and the alternative table corresponding to the surgical type.

7. The method for automatic encoding of surgical operation records based on a large language model according to claim 1, characterized in that: While using the large language model, the erroneous results that occur during use are stored, and these erroneous examples are used to correct the large language data using retrieval enhancement generation.

8. The method for automatic encoding of surgical operation records based on a large language model according to claim 1, characterized in that: In step S5, the mapping dictionary refers to standard surgical operation terms and their corresponding codes issued by national functional departments or professional associations.

9. A surgical operation record automatic encoding system based on a large language model, used to implement the surgical operation record automatic encoding method based on a large language model as described in any one of claims 1 to 8, characterized in that: include: The annotation module is used to manually annotate a number of original surgical record texts, and properly store the original surgical operation record fragments annotated with standard surgical operation nouns, as well as the corresponding standard surgical operation nouns; An instruction data augmentation module, used to generate an instruction training data set according to the labeled data of the labeling module, and perform data augmentation on the instruction training data set; The training module is used to carry out instruction training on the large language model; The deployment module is used to deploy the large language model and make it accessible through the API. It records new situations that did not appear in the training process or situations where the generation is incorrect, and corrects the large language model in a way that enhances the generation through retrieval. The code mapping module is used to map the results generated by the large language model to obtain standard codes; The user interaction module is used to provide coding services to doctors and coders after the large language model is deployed; The database module is used to provide data support for the annotation module, instruction data augmentation module, training module, deployment module, code mapping module and user interaction module.

10. The automatic coding system for surgical operation records based on a large language model according to claim 9, characterized in that: The user interaction module includes an input module and an output module. The input module is connected to the deployment module and is used by doctors and coders to input surgical operation records. The output module is connected to the code mapping module and outputs codes corresponding to surgical operation nouns.

Images