Follow-up visit suggestion generation method and system based on structured abnormal label identification
The method and system leverage structured abnormality label recognition and LLMs to enhance the personalization and accuracy of follow-up recommendations, addressing inefficiencies in traditional systems by automating the process and ensuring relevance to individual patient needs.
Patent Information
- Application Number
- CN202410980674.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-22
- Publication Date
- 2024-12-17
AI Technical Summary
Traditional follow-up recommendation systems in healthcare rely heavily on manual operations, leading to inefficiencies, insufficient personalization, and inaccuracies due to the lack of individualized consideration and structured abnormality label recognition, limiting their applicability in precision medicine.
A method and system utilizing structured abnormality label recognition, combining large language models (LLM) with professional medical annotations to automatically identify and generate personalized follow-up recommendations by building a structured abnormality label library and follow-up department suggestion library, incorporating data mining and machine learning to enhance accuracy and efficiency.
Improves the efficiency and accuracy of follow-up recommendations by personalizing them based on individual patient conditions, reducing the workload of healthcare professionals, and ensuring the generated recommendations are scientifically valid and practical.
Smart Images
Figure CN119153018A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of computers, and in particular to a method and system for generating follow-up suggestions based on structured abnormality label recognition. Background Art
[0002] With the development of medical information technology, electronic health records (EHR) have gradually become an important part of medical data management. EHR records the patient's full-cycle data in the hospital, including outpatient, admission, diagnosis, examination, testing, treatment and discharge and other key medical information. These data constitute the whole picture of the patient's medical journey and provide valuable resources for medical decision-making, disease monitoring, health management and scientific research analysis.
[0003] After discharge, patients need to undergo regular follow-up examinations according to the follow-up recommendations in the discharge summary to monitor the recovery of the disease and the results of treatment. Follow-up is an indispensable part of the patient's disease management process. How to use the massive medical data in EHR to extract useful information and formulate discharge follow-up recommendations that meet the individual conditions of patients is a major challenge in the field of medical and health management.
[0004] Traditional follow-up recommendation generation systems are usually based on fixed filling templates and rely heavily on manual operations. They have problems such as low efficiency, insufficient personalization, and easy omission of important information. They are difficult to adapt to the diversity of patients' conditions and cannot meet the needs of precision medicine and personalized medicine.
[0005] In recent years, with the continuous development and maturity of artificial intelligence technology, the medical field has gradually used natural language processing (NLP) technology to extract clinical information from EHR and provide decision support for doctors. For example, by calling a large language model such as ChatGPT and inputting the patient's EHR data, relevant follow-up recommendations can be obtained. However, these answers based on generative AI are not always perfect. They have problems such as incomplete information and hallucinations, which limits their application in clinical practice. In addition, when generating follow-up recommendations, existing AI systems often lack the recognition and consideration of individual differences among patients, resulting in the generated recommendations may not be fully adapted to the patient's specific condition. At the same time, these systems usually do not build a dedicated follow-up knowledge base, which limits the accuracy of follow-up recommendations. Summary of the invention
[0006] The purpose of the present invention is to provide a method and system for generating follow-up suggestions based on structured abnormality label recognition to solve the problems raised in the above background technology.
[0007] To achieve the above-mentioned object of the invention, one aspect of the present invention provides a method for generating follow-up suggestions based on structured abnormality label recognition, comprising the following steps:
[0008] Step S1, processing the test, examination, and medical history reports in the EHR data to build a structured abnormality label library;
[0009] Step S2, using LLM and professional medical annotations, build a follow-up department recommendation library for each atomic abnormality in the abnormality label library;
[0010] Step S3: Generate follow-up recommendations.
[0011] Further, step S1 comprises the following steps:
[0012] Step S101, extracting initial abnormal labels based on LLM;
[0013] Step S102, cleaning the initial abnormal labels;
[0014] Step S103: Establish a structured exception label library.
[0015] Furthermore, the establishment of the structured abnormal label library includes: structuring the abnormal label list, extracting atomic abnormalities and organs from the inspection abnormal label library, extracting inspection names and inspection items from the test abnormal label library, and extracting disease names from the medical history abnormal label library.
[0016] Further, step S2 includes the following steps:
[0017] Step S201, design prompt words and use LLM to generate follow-up departments for each abnormal label;
[0018] Step S202: Correcting errors through professional medical annotation.
[0019] Further, step S3 includes the following steps:
[0020] Step S301, identifying abnormalities in the inspection report;
[0021] Step S302, identifying abnormalities in the inspection report;
[0022] Step S303, identifying abnormal past history;
[0023] Step S304, for the abnormalities extracted in steps S301, S302, and S303, the follow-up department corresponding to the abnormality label is obtained according to the follow-up department suggestion library, and the complete follow-up suggestions are summarized and output.
[0024] Another aspect of the present invention provides a follow-up suggestion generation system based on structured abnormality label recognition, comprising a structured abnormality label module, a follow-up department suggestion module, and a follow-up suggestion generation module, wherein:
[0025] The structured abnormality label module builds a structured abnormality label library by processing the inspection, examination, and past history reports in the EHR data;
[0026] The follow-up department suggestion module uses LLM and professional medical annotations to build a follow-up department suggestion library for each atomic abnormality in the abnormality label library;
[0027] The follow-up suggestion generation module is used to generate follow-up suggestions.
[0028] Furthermore, the structured anomaly label module includes the following components:
[0029] Extraction component, used to extract initial anomaly labels based on LLM;
[0030] A cleaning component, used to clean the initial abnormal labels;
[0031] The library building component is used to build a structured exception tag library.
[0032] Furthermore, the establishment of the structured abnormal label library includes: structuring the abnormal label list, extracting atomic abnormalities and organs from the inspection abnormal label library, extracting inspection names and inspection items from the test abnormal label library, and extracting disease names from the medical history abnormal label library.
[0033] Furthermore, the follow-up department recommendation module includes the following components:
[0034] The LLM component is used to design prompt words and use LLM to generate follow-up departments for each abnormal label;
[0035] Error correction component, used for error correction through professional medical annotation.
[0036] Furthermore, the follow-up suggestion generation module includes the following components:
[0037] An inspection report component, used to identify inspection report anomalies;
[0038] Inspection report component, used to identify inspection report anomalies;
[0039] A past history component, used to identify past history abnormalities;
[0040] The follow-up suggestion component is used to obtain the follow-up department corresponding to the abnormality label according to the follow-up department suggestion library for the abnormalities extracted from the examination report component, the test report component, and the past history component, and summarize and output complete follow-up suggestions.
[0041] Compared with the prior art, the present system and method have the following advantages:
[0042] 1. By using advanced data mining and machine learning technologies, we can automatically identify abnormal labels in the patient's EHR examinations, tests, and medical history, and generate more accurate follow-up recommendations in combination with a professional follow-up recommendation library. This method not only improves the efficiency of follow-up recommendation generation, but also improves the accuracy and clinical applicability of follow-up recommendations through personalized data analysis.
[0043] 2. By building a comprehensive follow-up recommendation library, which contains detailed follow-up recommendations and clinical guidelines for different abnormal labels, it aims to make up for the shortcomings of generative AI technology. In addition, the method of this patent also includes steps to verify and optimize the generated follow-up recommendations to ensure that the generated follow-up recommendations are both scientific and practical.
[0044] 3. By analyzing the patient's EHR data, abnormal labels related to the patient's condition are identified, thereby generating a more personalized follow-up plan. This method can not only improve the pertinence of follow-up, but also adjust the follow-up plan according to the patient's real-time condition to make it more suitable for the patient's actual needs.
[0045] 4. Through automated abnormal label recognition and follow-up suggestion generation, the workload of doctors is reduced and work efficiency is improved.
[0046] 5. The construction of the exception label library and follow-up library provides scalability and flexibility for future follow-up suggestion generation. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1 Flowchart of a method for generating follow-up recommendations based on structured abnormality label identification.
[0048] Figure 2 Schematic diagram of the process of building a structured exception tag library.
[0049] Figure 3 Schematic diagram of the process for building a follow-up department recommendation library.
[0050] Figure 4 A flow chart was created for generating follow-up recommendations. DETAILED DESCRIPTION
[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0052] like Figure 1 Shown is a flow chart of the method of the present invention, and the specific steps are as follows:
[0053] Step S1, by processing the inspection, examination, and medical history reports in the EHR data, a structured abnormality tag library is constructed. In order to speed up the processing flow, an abnormality tag library based on a large language model (LLM) is constructed.
[0054] Step S2, by using LLM and professional medical annotations, a special follow-up department recommendation library is constructed for each atomic abnormality in the abnormality label library.
[0055] Step S3: The follow-up recommendation generation system first identifies the patient's abnormality label, matches the atomic abnormality in the structured abnormality label library, and then maps the follow-up recommendation for each abnormality from the follow-up library to generate the final overall personalized follow-up plan.
[0056] like Figure 2 The figure shows a flow chart of step S2 for constructing a structured anomaly tag library, which includes the following steps:
[0057] Step S101, initial abnormal label extraction based on LLM. Design different prompt words, and use LLM to perform preliminary extraction of abnormalities in inspection, examination, and history reports. The following "" shows an example of setting prompt words related to inspection reports to train LLM models. LLM models can use commercial models such as ChatGPT3.5 and Wenxin Yiyan, or open source models such as LLama and Qwen.
[0058] "You will play the role of a doctor and need to extract all the abnormal physical contents in the patient's examination report and return the result as a python List type. Note that you only need to extract the abnormalities in the body. If it is "no obvious abnormality" or other irrelevant statements, you do not need to extract them.
[0059] In addition, you need to catch all exceptions, not miss any exceptions, and do not split a complete sentence (the same sentence may be on both sides of ',').
[0060] Here is the first example:
[0061] Input: 2023-11-30 Chest CT (thin layer) plain scan: scattered small nodules in both lungs, increased and blurred texture; cystic lesions in the anterior mediastinum. Attached is the mass shadow on the inner side of the right breast. Low-density lesions in the liver. Please follow up in combination with clinical and other related examinations.
[0062] Output: ['Scattered small nodules in both lungs, increased and blurred texture', 'Cystic lesions in the anterior mediastinum', 'A mass shadow is seen on the inner side of the right breast', 'Low-density lesions in the liver']
[0063] Analysis: 'Scattered tiny nodules in both lungs, increased and blurred texture' is a complete sentence that describes both lungs, so there is no need to separate them.
[0064] Here is the second example:
[0065] Input: 2023-11-06 Color Doppler ultrasound of liver, gallbladder, pancreas, spleen, kidney, and portal vein: fatty liver. After cholecystectomy. No obvious abnormalities were found in the pancreas, spleen, and kidney. No obvious abnormalities were found in the portal vein blood flow.
[0066] Output: ['Fatty liver', 'After cholecystectomy']
[0067] Analysis: 'After cholecystectomy' is also an abnormal signal
[0068] Here is a third example:
[0069] Input: 2023-11-04 ECG: Normal ECG.
[0070] Output: []
[0071] Analysis: The patient's examination report showed no abnormalities
[0072] Here is the fourth example:
[0073] Input: 2023-11-23 breast 3D color ultrasound and its drainage lymph nodes, supraclavicular lymph nodes + positioning breast center 3D ultrasound: left breast (9 o'clock) solid nodular lesions: proposed US-BI-RADS 4B category, consider intraductal lesions, has been located, please combine clinical. Bilateral axillary, bilateral
[0074] No obvious abnormally enlarged lymph nodes were found above the clavicle.
[0075] Output: ['Parenchymal nodular lesion in the left breast (9 o'clock): proposed US-BI-RADS 4B category, intraductal lesion considered, localized']
[0076] Analysis: "The left breast..." is a complete sentence and does not need to be separated; "No obvious abnormally enlarged lymph nodes were found in the bilateral axillary and bilateral supraclavicular lymph nodes" indicates that there is no abnormality.
[0077] You need to understand the above four examples and find all the abnormal contents in the patient's examination report (don't miss any sentence or word).
[0078] Below I will give you my input, you need to give me a List type output, no additional irrelevant content is required:"
[0079] Step S102, initial abnormal label cleaning. The abnormalities extracted by the LLM model are not perfect, and some unqualified abnormal labels will be extracted, such as ['No obvious abnormalities were found in the liver, gallbladder, pancreas, spleen and kidney', 'No obvious abnormalities were found in the echocardiogram', 'No obvious abnormal density shadows were found in both lungs']. Such expressions are not abnormal labels and need to be cleaned and filtered to finally obtain an abnormal label list.
[0080] Step S103, establish a structured abnormal label library. Structuring is performed based on the abnormal label list. The inspection abnormal label library extracts atomic abnormalities and organs, the test abnormal label library extracts test names and test items, and the medical history abnormal label library extracts disease names. The structuring here can be generated using an open source medical nomenclature recognition model or an LLM model.
[0081] like Figure 3 The figure shows the process flow of step S2 to build a follow-up department suggestion library. Specifically, it is divided into two steps:
[0082] Step S201, design prompt words, and use LLM to generate follow-up departments for each abnormal label.
[0083] Step S202: Correcting errors through professional medical annotation.
[0084] like Figure 4 The flowchart of step S3 generating follow-up recommendations is shown. Specifically, the following steps are mainly included:
[0085] Step S301, identification of abnormalities in the examination report, firstly, the examination report is divided into sentences by punctuation marks to obtain some short sentences, then these short sentences are filtered, non-abnormal labels such as 'xxx no obvious abnormality' are filtered, and then the examination abnormality label library is used for matching, first matching the atomic abnormality, then matching the organ name, and finally extracting the abnormal label.
[0086] Step S302, test report anomaly identification, is similar to the inspection report. After the sentence is split, it mainly matches the test items and test names in the test anomaly label library, and then determines whether the test value is within the normal range. If it deviates from the range, the abnormal label is extracted.
[0087] Step S303, anomaly identification of medical history, after sentence segmentation, first filter out negative words such as 'deny xxx', then match the disease name in the medical history anomaly tag library, and extract the anomaly tag.
[0088] Step S304, for the abnormal labels extracted in 1, 2, and 3, query the follow-up department suggestion library to obtain the follow-up department corresponding to each abnormal label, and summarize and output complete follow-up suggestions.
[0089] Following the above steps, a complete follow-up recommendation is as follows:
[0090] Other department follow-up recommendations:
[0091] 1. Past medical history shows a history of hypertension for more than 4 years. Regular follow-up with the hypertension department is recommended.
[0092] 2. The frontal chest X-ray showed slightly more texture in both lungs; biochemistry showed carbon dioxide was 34.9↑mmol / L. Regular follow-up in the respiratory department is recommended.
[0093] 3. The chest X-ray shows thoracic scoliosis, and regular orthopedic follow-up is recommended.
[0094] 4. The heart showed mild mitral regurgitation and aortic valve degeneration with mild regurgitation; the frontal chest X-ray showed aortic tortuosity, and regular follow-up by the cardiology department was recommended.
[0095] 5. Biochemistry showed triglyceride level was 1.91↑mmol / L. Regular follow-up with the endocrinology department is recommended.
[0096] 6. 3D color ultrasound of breast and its drainage lymph nodes + positioning shows the right axillary lymph nodes and bilateral breasts.
[0097] The application scenarios of this patent include patient follow-up management in hospitals and clinics, long-term health management of patients with chronic diseases, personalized medical and preventive health care services, and can be used to automatically generate discharge summaries and examination report interpretations. The present invention generates follow-up department recommendations by automatically extracting abnormal labels, which improves the pertinence of follow-up recommendations, and introduces structured abnormal label recognition, which can accurately capture abnormalities in patient reports and reduce the omission rate. The present invention constructs an abnormal label library and a follow-up library based on the LLM model, which improves the construction efficiency.
[0098] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for generating follow-up suggestions based on structured abnormality label recognition, characterized in that: The following steps are involved: Step S1, processing the test, examination, and medical history reports in the EHR data to build a structured abnormality label library; Step S2, using LLM and professional medical annotations, build a follow-up department recommendation library for each atomic abnormality in the abnormality label library; Step S3: Generate follow-up recommendations.
2. The method for generating follow-up suggestions based on structured abnormality label recognition according to claim 1, characterized in that: Step S1 comprises the following steps: Step S101, extracting initial abnormal labels based on LLM; Step S102, cleaning the initial abnormal labels; Step S103: Establish a structured exception label library.
3. The method for generating follow-up suggestions based on structured abnormality label recognition according to claim 2, characterized in that: The establishment of the structured abnormal label library includes: structuring the abnormal label list, extracting atomic abnormalities and organs from the inspection abnormal label library, extracting inspection names and inspection items from the test abnormal label library, and extracting disease names from the medical history abnormal label library.
4. The method for generating follow-up suggestions based on structured abnormality label recognition according to claim 1, characterized in that: Step S2 includes the following steps: Step S201, design prompt words and use LLM to generate follow-up departments for each abnormal label; Step S202: Correcting errors through professional medical annotation.
5. The method for generating follow-up suggestions based on structured abnormality label recognition according to claim 1, characterized in that: Step S3 includes the following steps, Step S301, identifying abnormalities in the inspection report; Step S302, identifying abnormalities in the inspection report; Step S303, identifying abnormal past history; Step S304, for the abnormalities extracted in steps S301, S302, and S303, the follow-up department corresponding to the abnormality label is obtained according to the follow-up department suggestion library, and the complete follow-up suggestions are summarized and output.
6. A follow-up suggestion generation system based on structured abnormality label recognition, characterized in that: It includes a structured abnormality label module, a follow-up department suggestion module, and a follow-up suggestion generation module, among which: The structured abnormality label module builds a structured abnormality label library by processing the inspection, examination, and past history reports in the EHR data; The follow-up department suggestion module uses LLM and professional medical annotations to build a follow-up department suggestion library for each atomic abnormality in the abnormality label library; The follow-up suggestion generation module is used to generate follow-up suggestions.
7. A follow-up suggestion generation system based on structured abnormality label recognition according to claim 6, characterized in that: The structured anomaly label module includes the following components: Extraction component, used to extract initial anomaly labels based on LLM; A cleaning component, used to clean the initial abnormal labels; The library building component is used to build a structured exception tag library.
8. A follow-up suggestion generation system based on structured abnormality label recognition according to claim 7, characterized in that: The establishment of the structured abnormal label library includes: structuring the abnormal label list, extracting atomic abnormalities and organs from the inspection abnormal label library, extracting inspection names and inspection items from the test abnormal label library, and extracting disease names from the medical history abnormal label library.
9. The system for generating follow-up suggestions based on structured abnormality label recognition according to claim 6, characterized in that: The Follow-up Department Recommendation module includes the following components: The LLM component is used to design prompt words and use LLM to generate follow-up departments for each abnormal label; Error correction component, used for error correction through professional medical annotation.
10. The system for generating follow-up suggestions based on structured abnormality label recognition according to claim 6, characterized in that: The follow-up suggestion generation module includes the following components, An inspection report component, used to identify inspection report anomalies; Inspection report component, used to identify inspection report anomalies; A past history component, used to identify past history abnormalities; The follow-up suggestion component is used to obtain the follow-up department corresponding to the abnormality label according to the follow-up department suggestion library for the abnormalities extracted from the examination report component, the test report component, and the past history component, and summarize and output complete follow-up suggestions.