A method for text verification of physical examination report recognition results based on a large model

Abstract

The present invention discloses a method for text verification of physical examination report recognition results based on a large model. Relying on the powerful semantic understanding ability and high generalization of the large model, the present invention can quickly, accurately and multi-levelly verify the text of the recognition results of the physical examination report. In addition, by linking external knowledge sources, on the one hand, the present invention supplements the medical background knowledge of the large model, thereby enhancing professionalism and interpretability; on the other hand, the present invention expands the ability of the large model to verify entity relationships under the standards of different physical examination institutions, and ultimately improves the consistency and standardization of medical data.

Description

Technical Field

[0001] The invention relates to a method for performing text verification on a physical examination report recognition result by using a large model, and belongs to the field of medical technology. Background Art

[0002] Physical examination reports are important resources in the medical field, covering a variety of medical data such as basic patient information, various examination and test sheets, imaging reports, questionnaires, etc. Accurately identifying and extracting information from physical examination reports is of great significance. On the one hand, it can help patients understand their health status faster and more clearly; on the other hand, it can also provide reliable data support for subsequent data analysis and management work in various aspects such as disease prevention and health management.

[0003] The recognition result of a physical examination report usually contains the following data: basic patient information, examination department, examination item information, ultrasound image report, health guidance suggestions, and physical examination summary. With the continuous advancement of artificial intelligence technology, the use of deep learning algorithms in the medical field to extract structured information from physical examination reports has become a basic task. However, due to the lack of clarity, diverse formats, and algorithmic limitations of the deep learning model itself, the recognition results may contain text data errors, such as missing text, typos, and semantic deviations.

[0004] In order to ensure the accuracy and consistency of the physical examination report data, it is crucial to strictly verify and proofread the automatically extracted and structured results. With the advantages of massive corpus learning and advanced underlying network architecture, the large model can quickly verify and correct multiple dimensions such as text and semantics, significantly improving the accuracy of the data. In addition, by integrating external knowledge sources, the large model is able to obtain real-time medical knowledge background, which not only further improves the accuracy of the verification, enhances its basis and interpretability, but also enables the large model to more deeply verify the entity relationships in the recognition results, enhancing its verification and generalization capabilities facing the standards of different medical institutions.

[0005] In the patent (application number CN202410483571.9) published on July 9, 2024, a method for text and image correction for digestive endoscopy reports was introduced. This method is based on the CLIP model, which establishes a matching model between endoscopic text and images through training, and uses a similarity algorithm to select the closest correct text in the input. However, this method is subject to some common limitations of small deep learning models, such as high dependence on feature vector encoding algorithms and a large amount of manually annotated text and image data, as well as limitations on network architecture and parameter volume, resulting in insufficient understanding of medical knowledge and lack of interpretability.

[0006] The patent (application number CN201911258459.0) published on February 2, 2024 proposes a method for text recognition and correction of medical physical examination reports. The method first trains the BERT-BiLSTM-CRF neural network to automatically identify medical entities in natural language texts. Subsequently, the identified medical entities are matched with the entity names in the knowledge base, and the medical entities in the original text are corrected by calculating the edit distance and comparing the similarity, and outputting the matching results with the highest similarity. This method relies on the quality of the knowledge base, and because it contains multiple components, it is relatively inflexible. Summary of the invention

[0007] The technical problem to be solved by the present invention is that in the actual application of physical examination report recognition, due to factors such as poor report clarity, diverse and complex styles, and different standards of physical examination institutions, the recognition results often have various errors such as missing words, typos, semantic errors, inconsistent ranges, etc. The current mainstream technology mainly relies on the combination of deep learning small models and knowledge bases, but this method has obvious defects: on the one hand, it is highly dependent on the manually organized knowledge base, which increases the labor cost; on the other hand, it is limited by the semantic understanding and generalization ability of the small model, resulting in insufficient understanding at the semantic level and lack of sufficient interpretability.

[0008] In order to solve the above technical problems, the technical solution of the present invention is to disclose a method for text verification of physical examination report recognition results based on a large model, which is characterized by accurate results, interpretability and generalization ability, including the following steps:

[0009] Step 1: Use a large model to perform semantic verification, including:

[0010] In the semantic verification stage, a targeted text verification guidance prompt is built to guide the large model to perform multi-level text character verification according to predetermined requirements;

[0011] After completing the construction of the text verification guidance prompt, choose Few-shot learning or fine-tuning training to improve the large model's ability to understand specific tasks;

[0012] Step 2: Link external knowledge sources to enhance, including:

[0013] Step 201: Select an external knowledge source;

[0014] Step 202: Background medical knowledge retrieval:

[0015] Before each text verification, the medical background knowledge most relevant to the content to be verified is retrieved from the knowledge source through the designed algorithm;

[0016] Step 203: Knowledge integration:

[0017] The retrieved background medical knowledge is sent into the big model together with the text verification guidance prompt after being concatenated with the prompt using "Known background medical knowledge is:" as a prompt.

[0018] Preferably, in step 1, the text verification guidance prompt includes:

[0019] Clarify the original text of the recognition result to be verified;

[0020] Indicate the general type of recognition result;

[0021] Specify the specific levels at which the large model should perform text verification.

[0022] Preferably, in step 1, in the few-shot learning, 2 to 3 correct text verification processes are selected as samples and added to the historical conversation records of the large model to guide the model to give verification answers based on these samples.

[0023] Preferably, in step 1, in the fine-tuning training, a training set containing more than 100 question-answer pairs is constructed, and the parameter weights of the large model are adjusted by Lora or full-parameter fine-tuning.

[0024] Preferably, in step 201, the external knowledge source includes a traditional knowledge base, a knowledge graph, and a knowledge text collection.

[0025] Preferably, in step 202, the search method adopted is keyword search and / or vector cosine similarity search.

[0026] The present invention relies on the powerful semantic understanding ability and high generalization of the big model, and can quickly, accurately and multi-levelly verify the text of the recognition results of the physical examination report. In addition, by linking external knowledge sources, the present invention supplements the big model with medical background knowledge on the one hand, thereby enhancing professionalism and interpretability; on the other hand, the present invention expands the big model's ability to verify entity relationships under the standards of different physical examination institutions, ultimately improving the consistency and standardization of medical data.

[0027] Compared with the prior art solutions, the present invention has the following beneficial effects:

[0028] 1. Compared with the traditional method based on the combination of a small deep learning model and a knowledge base, the present invention achieves accurate, fast, and multi-level text verification of the physical examination report recognition results by leveraging the powerful semantic understanding ability and high generalization of the large model;

[0029] 2. By linking external knowledge sources, more professional and accurate medical background knowledge is added to the big model, which not only enhances the professionalism, interpretability and credibility of the model, but also expands the ability of the big model to verify entity relationships under the standards of different medical institutions;

[0030] 3. The large model and the framework of multiple external knowledge sources can meet the standards of different physical examination institutions and are more flexible, universal and highly practical. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 It is a flow chart of the present invention. DETAILED DESCRIPTION

[0032] The present invention will be further described below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention. In addition, it should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the appended claims of the application equally.

[0033] In the application scenario of checking the recognition results of physical examination reports, in view of the shortcomings of existing inventions, the present invention proposes a method for text verification based on a large model. The core of this method is to make full use of the advantages of large models in pre-training of massive data and their high generalization ability. By adopting Few-shot learning or fine-tuning training strategies, in-depth analysis and accurate detection of multi-level errors in the recognition results of physical examination reports are performed. These errors may be caused by the low clarity or diverse formats of the reports, resulting in omissions or errors in text in fields such as examination item names and department names.

[0034] At the same time, in order to further improve the accuracy and depth of verification, the present invention also introduces external knowledge sources as supplementary input. By linking external knowledge sources, medical background knowledge closely related to the current content to be verified can be retrieved. This approach not only helps to verify the consistency between entity relationships such as the scope of inspection items and results, but also adds a strong explanation basis for the verification results provided by the large model.

[0035] The technical solution disclosed in the present invention specifically includes the following contents:

[0036] 1. Use a large model for semantic verification

[0037] In the semantic verification stage, the present invention guides the large model to perform multi-level text character verification according to predetermined requirements by constructing targeted text verification guidance prompts. The text verification guidance prompt contains three key parts: first, clarifying the original text of the recognition result to be verified; second, indicating the general type of the recognition result (such as department name, examination item name, etc.); finally, specifying the specific levels at which the large model should perform text verification (such as typos, missing words, semantic understanding, etc.).

[0038] After completing the construction of the text verification guidance prompt, you can choose to use Few-shot learning or fine-tuning training to improve the large model's ability to understand specific tasks. In Few-shot learning, select 2 to 3 correct text verification processes as samples and add them to the historical session records of the large model to guide the model to give verification answers based on these samples. In fine-tuning training, you need to build a training set containing more than 100 question-answer pairs, and adjust the parameter weights of the large model through Lora or full-parameter fine-tuning.

[0039] Taking the text error correction of the examination item name "red blood cell count" as an example, the constructed text verification guidance prompt may be expressed as: "It is known that the recognized text of the physical examination report to be verified is 'red blood cell count', and the type is the examination item name. Please check for typos, missing words and semantics, and output the correct text." The answer of the large model should be "red blood cell count", and the two together constitute a question and answer pair.

[0040] 2. Enhanced links to external knowledge sources

[0041] In order to expand the large model's ability to verify entity relationships under different medical institution standards, the present invention adds a device configuration for linking external knowledge sources.

[0042] (1) Selection of external knowledge sources

[0043] External knowledge sources can include traditional knowledge bases, knowledge graphs, knowledge text collections, and other forms.

[0044] (2) Background medical knowledge retrieval

[0045] Before each text verification, the medical background knowledge most relevant to the content to be verified is retrieved from the knowledge source through the designed algorithm. The retrieval method can be keyword retrieval, vector cosine similarity retrieval, etc., or a combination of multiple methods.

[0046] (3) Knowledge integration

[0047] Finally, the retrieved background medical knowledge is sent to the big model together with the text verification guidance prompt, using "Known background medical knowledge is:" as the prompt word. In this way, the big model can give a more reasonable and explainable verification result based on more comprehensive information.

[0048] The range check is performed using the recognized physical examination report text "The patient's gender is female, the red blood cell count result is 4.2, the normal range is 3 to 4×10^12 / L, so it is abnormal" as an example.

[0049] First, according to the patient's gender (female) and the name of the examination item (red blood cell count), the "normal value of the red blood cell count of normal women is 3.5 to 5×10^12 / L" is queried from the external knowledge source.

[0050] This information is then concatenated with the text verification guidance prompt to form the final input for the large model: " Known background medical knowledge: The normal value of red blood cell count for normal women is 3.5 to 5×10 12 / L. The identification text of the medical examination report to be verified is 'The patient's gender is female, the red blood cell count result is 4.2, and the normal range is 3 to 4×10 12 / L, therefore abnormal', please perform text verification of entity relationship and output the correct text. "

[0051] Based on this background knowledge, the big model can quickly correct the error and give the answer: "The patient's gender is female, and the red blood cell count is 4.2, and the normal range is 3.5 to 5×10 12 / L, so there is nothing abnormal. ”

[0052] Based on the advantages of big models in semantic understanding and generalization, the present invention uses big models to perform text verification on the results of physical examination report recognition. At the same time, by linking external knowledge bases, the professionalism and interpretability of big models in the medical field are further enhanced, and the big model's ability to verify the relationship between entities in recognition results under the standards of different medical institutions is expanded, thereby ensuring the accuracy, consistency and high quality of recognition results from multiple aspects, providing a solid foundation for subsequent medical data analysis and management.

Claims

1. A method for text verification of physical examination report recognition results based on a large model, characterized in that: The following steps are involved: Step 1: Use a large model to perform semantic verification, including: In the semantic verification stage, a targeted text verification guidance prompt is built to guide the large model to perform multi-level text character verification according to predetermined requirements; After completing the construction of the text verification guidance prompt, choose Few-shot learning or fine-tuning training to improve the large model's ability to understand specific tasks; Step 2: Link external knowledge sources to enhance, including: Step 201: Select an external knowledge source; Step 202: Background medical knowledge retrieval: Before each text verification, the medical background knowledge most relevant to the content to be verified is retrieved from the knowledge source through the designed algorithm; Step 203: Knowledge integration: The retrieved background medical knowledge is sent into the big model together with the text verification guidance prompt after being spliced ​​with "Known background medical knowledge is:" as a prompt.

2. A method for text verification of physical examination report recognition results based on a large model as described in claim 1, characterized in that: In step 1, the text verification guidance prompt includes: Clarify the original text of the recognition result to be verified; Indicate the general type of recognition result; Specify the specific levels at which the large model should perform text verification.

3. The method for text verification of physical examination report recognition results based on a large model as claimed in claim 1, characterized in that: In step 1, in the Few-shot learning, 2 to 3 correct text verification processes are selected as samples and added to the historical conversation records of the large model to guide the model to give verification answers based on these samples.

4. A method for text verification of physical examination report recognition results based on a large model as claimed in claim 1, characterized in that: In step 1, in the fine-tuning training, a training set containing more than 100 question-answer pairs is constructed, and the parameter weights of the large model are adjusted by Lora or full parameter fine-tuning.

5. The method for text verification of physical examination report recognition results based on a large model as claimed in claim 1, characterized in that: In step 201, the external knowledge source includes a traditional knowledge base, a knowledge graph, and a knowledge text collection.

6. A method for text verification of physical examination report recognition results based on a large model as claimed in claim 1, characterized in that: In step 202, the search method adopted is keyword search and / or vector cosine similarity search.

Images