Medical record structuring system based on large model

Abstract

The invention discloses a medical record structuring system based on a large model. The system comprises a hardware acquisition device, a data processing module, a large model structuring module, a rule engine module and a patient information integration module, the method for processing the paper medical record comprises the following steps of: S1, scanning the paper medical record by adopting a hardware acquisition device, and acquiring image data; s2, in a data processing module, converting the image data acquired in the S1 into medical record text data; s3, in a large model structured module, taking the text data obtained in the S2 as input of a large model, and generating structured data of a single medical record; s4, in a rule engine module, performing standardization processing on the structured data of the single medical record obtained in the S3; s5, in a patient information integration module, performing information integration on multiple medical records of the same patient to form complete patient diagnosis and examination information; in the step S3, a pre-trained medical base model is firstly obtained, then medical data structured training is carried out on the medical model, and a medical record structured model is obtained.

Description

Technical Field

[0001] The invention relates to a device and a system for realizing structuring of medical records, and belongs to the field of computer artificial intelligence application. Background Art

[0002] Medical records are important documents in the medical system that record the patient's medical history, diagnosis and treatment process, and doctor's orders. They contain rich and complex information, such as medication records, diagnosis results, and laboratory tests. There are many types and forms of medical record files, including but not limited to admission and discharge records, examination reports, genetic reports, outpatient medical records, etc. Because medical records may come from any hospital or other medical institution, and the data of medical institutions cannot be interconnected. Therefore, medical record information outside this hospital can only exist in paper form. For paper medical records, you can use the scanning camera device attached to the doctor's computer terminal or an external scanning device to scan and enter the system through OCR (optical character recognition).

[0003] Structured medical records refer to the arrangement and classification of various information in medical records according to certain specifications and formats, so that they are machine-readable and operable. Through the structuring of medical records, medical data can be automatically parsed and analyzed by computer systems, thereby improving the utilization efficiency and accuracy of medical data. There are many structuring methods, such as template-based structuring: by defining a specific template, the information in the medical record is filled in according to the requirements of the template, such as recording the patient's personal information, medical history, examination results, etc. according to the format of the template.

[0004] Structured medical records help medical staff read and analyze data more efficiently, facilitating clinical decision-making and medical research. However, existing methods for structuring medical record data mainly rely on rule engines and traditional machine learning algorithms, which are susceptible to format differences and data noise when processing multi-source data, resulting in inaccurate and incomplete information extraction. At the same time, existing technologies are difficult to cope with complex, multi-dimensional information extraction needs and lack versatility. Therefore, medical record structuring systems need to have both processing accuracy and efficiency.

[0005] Existing large models have some defects in medical data processing, such as being unable to fully cope with the complex data structure and high standards of semantic analysis in the medical field. In addition, large models may produce hallucination information, that is, generate content that does not conform to actual conditions or medical facts, thereby affecting clinical decision-making. The medical data annotation process usually requires a lot of expertise and manual intervention, making the annotation process complex and time-consuming, which also affects the training and effect optimization of the model. These problems hinder the widespread application and promotion of large model technology in medical scenarios. Summary of the invention

[0006] The present invention provides a medical record structuring device and system that combines hardware and a large model, which can efficiently realize a complete processing flow from scanning and entering paper medical records to extracting designated medical record information and outputting it in a structured manner.

[0007] The present invention provides a hardware modification and paper medical record report input system to improve the quality of data input. Specifically:

[0008] Step 1: Modify the existing doctor terminal computer or scanning equipment, such as installing a scanning camera under the monitor or using an external scanning device to achieve efficient collection and entry of paper medical records;

[0009] Step 2: The scanned paper medical records are subjected to OCR (optical character recognition) and preprocessing operations, and (optionally) privacy protection methods such as anonymization are used to obtain the medical record text;

[0010] Step 3: Use a large model specially trained for medical record structuring tasks to process the medical record text into blocks according to examination date or single examination to generate structured medical record data;

[0011] Step 4: The rule engine processes a single medical record, maps various medical entity information to the standard terminology library, and forms a standardized mapping relationship;

[0012] Step 5: Multiple reports of a patient can be merged according to predefined rules to ensure the integrity and consistency of information integration. In this process, the rule engine assists the merging operation to obtain a summary of all the patient's diagnosis and examination information.

[0013] The modification of the doctor's terminal computer or scanning equipment in step 1 specifically includes but is not limited to:

[0014] Install a scanning camera at a suitable position below the doctor's monitor to quickly and efficiently collect paper medical records.

[0015] Provide optional external scanning devices or terminal scanning devices to adapt to a variety of working environments (offline storage function is required for non-network environments) to ensure the flexibility of medical record collection.

[0016] Ensure that the device is compatible with existing medical systems and connects through common interfaces.

[0017] The device should have high-resolution image acquisition capabilities to improve the accuracy of OCR processing.

[0018] The supporting software system is used to accurately transmit the images collected by the scanning equipment to the current medical system, realizing efficient data entry and information transmission.

[0019] In this solution, the scanned paper medical records in step 2 are processed by OCR (optical character recognition), and the text is preprocessed to generate the preprocessed medical record text. This process includes:

[0020] Perform OCR (optical character recognition) conversion on the scanned paper medical records, convert the image data into text, and retain the location information of entities or sentences in the text for tracing in subsequent processing steps;

[0021] Apply preset pre-processing rules, including text structure adjustment, stop word removal, and text pre-structuring, and then optionally apply text anonymization and de-identification, and optionally use differential privacy and other technologies to further protect patient data;

[0022] In step 3 of this solution, a large model specially trained for medical record structuring tasks is used to process the medical record text into blocks according to the examination date or a single examination to generate structured medical record data. The specific steps are as follows:

[0023] Obtain medical record text data;

[0024] Obtain a preset extraction template according to the extracted report type to extract different medical entities;

[0025] Perform text data preprocessing according to the medical record text data, clean and preprocess the input data, clean and pre-structure the input data, and obtain the preprocessed medical record text;

[0026] The pre-processed (to be identified) medical record text is input into the preset text classification model to determine the report category based on the report features;

[0027] According to different report categories, select the corresponding preset processing template;

[0028] Input the medical record text to be identified into the medical record structured model deployed locally or in the cloud to perform information structured identification, including but not limited to diseases, pathological examinations, genetic testing, medication status, etc., to obtain the preliminary identification results of a single report;

[0029] The preliminary recognition results are checked for preset output structure to detect whether their format conforms to the specifications of the target format (such as the key-value pair structure of JSON, the row and column alignment of tables, the tag nesting relationship of XML, etc.) to ensure that their format is convenient for subsequent rule engine processing and integration into the system.

[0030] The training steps of the large model for the medical record structuring task used in step 3 include but are not limited to the following key steps:

[0031] Get the pre-trained medical base model (Base Model);

[0032] Extract and annotate structured data from historical medical records, including report categories, disease information, and medical entity information, to obtain a preset amount of training data;

[0033] Import the obtained training data into the medical model for medical data structured training to obtain an efficient medical case record structured model;

[0034] The entity recall and precision of the medical record structured model are calculated based on the actual medical records. If the recall or precision is less than the preset value, repeated training is performed until the recall and precision reach the preset value.

[0035] The model that reaches the preset value is exported and stored to obtain a large model for processing medical record structured tasks.

[0036] In step 4 of this solution, the rule engine processes a single medical record by mapping various medical entity information to the standard library to form a standardized mapping relationship, specifically:

[0037] Access to a library of standard medical terms;

[0038] Organize the preliminary recognition results of the single report in step 3 into a structured information set, which contains multiple tuples with medical entity categories as keys and medical record entity information as values, including but not limited to medication records, diagnosis results, laboratory tests, etc.;

[0039] Use key as index to obtain the candidate items related to key in the standard medical terminology library, map the value to the medical standard expression term through synonyms or normalized large models, and obtain the mapped medical data.

[0040] In step 5 of this solution, multiple reports of the patient can be merged according to predefined rules to ensure the integrity and consistency of information integration. In this process, the rule engine assists the merging operation to obtain a summary of all the diagnosis and examination information of the patient, specifically:

[0041] Get preset patient result merging rules;

[0042] The patient's medical records are sequentially processed in step 4 to obtain a set of all single image results for the patient;

[0043] The single image result sets are merged according to the preset patient result merging rules to obtain the patient's final structured data.

[0044] The following beneficial effects can be achieved through this scheme:

[0045] 1. By transforming the terminal equipment for report entry, the rapid collection and high-quality entry of paper medical records can be achieved, and combined with OCR de-identification and optional differential privacy protection measures, the efficiency of data entry and privacy compliance can be ensured.

[0046] 2. Use a combination of big models and rule engines to achieve standardization and structured processing of medical record data, including synonym mapping and medical terminology standardization, to ensure data accuracy and consistency, and support the effective integration of multiple reports, making the information extraction process accurate and efficient, and suitable for multi-dimensional data analysis.

[0047] 3. The solution supports the efficient collection of electronic and paper medical records, can flexibly adapt to a variety of working environments, and the generated structured data can be easily integrated into other medical information systems. It has good compatibility and scalability and is suitable for a variety of clinical applications and medical research needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is the overall flow chart of the medical record structured processing in the present invention.

[0049] Figure 2 Schematic diagram of the display modified in the present invention

[0050] Figure 3a This is the structured large model training flow chart in the present invention

[0051] Figure 3b It is a flow chart of the large model structured extraction module in the present invention. DETAILED DESCRIPTION

[0052] The specific implementations described below are only used to explain the present invention and are not intended to limit the present invention.

[0053] 1. Overview

[0054] The present invention relates to a system for structuring medical records, and in particular to an automatic generation of structured medical records in the medical field. The present invention realizes efficient collection and digital entry of paper medical records by modifying the computer terminal on the doctor's side (for example, installing a scanning camera under the display or using an external scanning device, but this modification is not a necessary operation), and combines optical character recognition (OCR) and privacy protection technology to ensure the efficiency and privacy protection of data entry. The system adopts a method combining a large language model with a rule engine to perform structured processing on medical record data, extract information such as diseases, pathological examinations, genetic testing, and medication conditions, respectively, and realize accurate analysis of medical information through a medically unified terminology library and model, ensuring the consistency and accuracy of data from different sources. In addition, the present invention supports the integration of multiple medical record reports, merges relevant information through a rule engine, ensures the integrity and consistency of medical record data, and thus provides a reliable data basis for clinical decision support and medical research. The present invention reduces the burden of medical workers in the collation of medical record data, improves the efficiency of data utilization, and has good compatibility and extensibility, and is suitable for any medical information system.

[0055] 2. Technical Solution

[0056] The system for structuring medical records based on a large model of the present invention comprises:

[0057] A hardware acquisition device, used for efficiently acquiring paper medical records, the hardware acquisition device comprising a scanning camera or an external scanning device installed below the doctor's display, and the hardware acquisition device has high-resolution image acquisition capabilities;

[0058] The data processing module is used to convert the collected paper medical record images into text data through OCR (optical character recognition) technology, and retain the location information of entities or sentences in the text to facilitate subsequent tracing;

[0059] A large model structuring module is used to input text data into a trained structured model deployed locally or in the cloud, process the medical record text by examination date or single examination, and generate structured medical record data, wherein the large model is trained based on preset training data of medical records;

[0060] If the fields extracted by the large model structuring module are missing or the output format does not reach the predetermined output format, the large model is restructured until all predetermined fields are extracted.

[0061] In the rule engine module, the structured data of a single medical record is standardized using preset rules or normalized rule models, and the extracted medical entities are mapped to the standard medical terminology library to form a standardized mapping relationship;

[0062] The patient information integration module is used to integrate multiple medical records of patients according to predefined merging rules to form complete patient diagnosis and examination information.

[0063] The hardware acquisition device includes a high-resolution scanning camera installed under the display or built into the display, and an optional external scanning device; the hardware acquisition device is connected to the medical information system through a standard interface, and the standard interface includes but is not limited to a USB interface, an Ethernet interface, a wireless interface, and a new interface (such as Thunderbolt or USB-C); the hardware acquisition device has a plug-and-play function and is compatible with existing medical information systems to achieve rapid and accurate transmission of collected data.

[0064] The acquisition device has a built-in data preprocessing module, which is used to optimize the acquired image data in real time to improve the accuracy of optical character recognition (OCR) processing and the stability of data transmission. The acquisition device supports multiple acquisition modes, including manual triggering, automatic triggering and continuous scanning; the acquisition device has an automatic gain control function to adapt to low light conditions; the acquisition device has image enhancement functions, including noise removal, edge enhancement, and automatic brightness and contrast adjustment.

[0065] The training of the structured large model is as follows: First, the text medical records obtained by the data processing module are annotated with their report classification, disease, and medical entity to obtain the annotated data of a single medical record. Then, the annotated data of multiple medical records are divided into training sets and test sets to train the structured model.

[0066] The training of the normalization rule model is as follows: obtain standard medical terms and their synonyms and non-standard expressions, and construct a medical knowledge graph; based on the knowledge graph, train the model's term normalization capability so that the model can convert non-standard medical terms into corresponding standard expressions; in the normalization process, through candidate scoring and selection, the model can select the most appropriate standardized term from the candidates according to the context.

[0067] The processing flow of the large model structured module is as follows: first, load the medical record text and perform data preprocessing on it; second, use the tokenizer in the large model to segment and vectorize the medical record text, and convert the text into a token sequence; then, classify the text features converted by the tokenizer through the preset text classification model to automatically identify different types of medical records; finally, select the corresponding processing template according to the classification results, and extract the preset entity information in the medical record based on the processing template.

[0068] The rule engine module performs synonym normalization on the preliminary recognition results to ensure that the structured data is consistent with standardized medical terminology, including:

[0069] Preset rules: The preset rules include standardized options for specific medical entities, ensuring consistent mapping of medical terms in specific fields (such as genetic testing), for example, directly mapping "EGFR '19 exon deletion mutation'" to the standardized term "19del".

[0070] Normalization rule model: The normalization rule model is built based on the knowledge graph of standard medical terms. It can identify and process synonyms and complex non-standard expressions in the initial recognition results, and map different forms of medical terms to standardized terms through semantic matching. For example, when "EGFR exon 19 deletion" or "EGFR exon19deletion" appears in the text, the normalization rule model can automatically normalize it to the standardized term "19del" to cope with more complex and diverse term expressions. This flexibility is difficult to fully cover by preset rules.

[0071] The patient information integration module first obtains the patient information merging rules defined by medical personnel and demanders, and the rules are loaded into the system in an externally configurable manner; secondly, the patient's multiple medical records are analyzed according to the obtained merging rules to identify key information items that need to be integrated; finally, the patient's medical records are integrated according to the merging rules.

[0072] In the present invention:

[0073] The “big model” can adopt language models such as Qwen2.5-7B, internlm2.5-7B, etc. that have been trained for the medical record structuring task suitable for this system.

[0074] Medical entity information includes but is not limited to pathological type, medication name, immunoassay TPS score, respiratory system diseases, digestive system diseases, etc.

[0075] The extraction template determined by the report type means that different report types need to extract different medical entities and different types of dates. For example, the inspection record should extract the inspection date, report date and related medical entities of the inspection results during the inspection process, but should not extract discharge and admission dates, medication information, etc.

[0076] The text preprocessing rules are as follows: for the medical record text converted from the image data, the sentences are divided according to the date identifier and the missing dates are completed. The date is recognized by regular expression, and the separator [Seg] is inserted before the recognized date. For example, the original medical record text is "2024-05-10 took amoxicillin 2024-05-14 diagnosed with lung inflammation", which becomes "[Seg] 2024-05-10 took amoxicillin [Seg] 2024-05-14 diagnosed with lung inflammation" after processing. If the date format is not YYYY%mm%dd%, NA is used as a placeholder when the year, month, and day parts are missing, such as 2024-12 will be replaced with 2024-12-NA, thus obtaining the prestructured report text.

[0077] The construction of the normalized rule model is based on the pre-trained medical base model. During the construction process, medical professionals carry out annotation work based on the normalized medical terminology library for the commonly used expressions in the collected medical records, and then train with these annotated data to finally generate a normalized rule model.

[0078] The normalized rule model is used to normalize different medical expressions. For example, the input of the normalized rule model is coronary vascular heart disease, and the output of the normalized rule model is coronary heart disease.

[0079] Patient information merging rules include, but are not limited to: removing duplicate examination results, merging information on the same date, and retaining only the maximum value of certain indicators such as immune detection tumor proportion score TPS.

[0080] In the present invention, unless otherwise specified, the term "medical report" is equivalent to "medical record". 3. Specific embodiments

[0082] Specific embodiments of the present invention combine the synergy of hardware devices and trained large language models.

[0083] The large language model is implemented by a high-performance computing server with sufficient storage and computing resources to support efficient reasoning and processing of the model. The model service can be deployed on a terminal device or a cloud server and requires at least one GPU with a video memory of more than 20G to support it.

[0084] In the task of medical record structuring, terminal devices and cloud servers each have their own advantages and disadvantages. The advantage of terminal device deployment is that data privacy protection is better, and data does not need to leave the local environment, thereby reducing the risk of sensitive information leakage. In addition, the processing of terminal deployment has lower latency and is more suitable for application scenarios with high real-time requirements. However, its disadvantages are that it requires sufficient hardware resources, the hardware cost is high, and it may face bottlenecks in storage and computing power when processing large-scale data.

[0085] In contrast, cloud servers have strong scalability and can dynamically increase computing resources to process large-scale medical record data. However, cloud processing requires data to be transmitted over the network, which has potential latency issues and may face challenges in data privacy and security. Therefore, in practical applications, terminal deployment is more suitable for scenarios with high requirements for data security and real-time processing, while cloud deployment is suitable for large-scale parallel computing and centralized management environments.

[0086] Comparison Dimensions Local deployment Cloud deployment Data Privacy Data is stored locally, with good privacy protection Data needs to be uploaded to the cloud, which poses a risk of privacy leakage Scalability Hardware limitations and poor scalability Resources can be added dynamically, with good scalability Delay Low latency, suitable for real-time processing Depends on the network, delays may occur Hardware Cost Need to purchase and maintain local servers, which is costly No need to purchase additional hardware, pay as you go, strong flexibility Data processing capabilities Limited processing power due to local hardware limitations Utilizes cloud resources, suitable for large-scale data processing Security Local control and high physical security Need to address cyberattacks and remote data security risks Regulatory Compliance Local control to meet strict data compliance requirements The legal requirements for cross-border data transmission and storage must be met

[0087] In this embodiment, the process of using this system is as follows Figure 1 As shown, the specific steps are as follows:

[0088] S1. Scan paper medical records with high resolution to collect high-quality image data.

[0089] In the specific implementation process, the scanning device uses high-resolution image acquisition technology to automatically optimize the shooting quality, obtain high-quality images and transmit them to the medical information system through a standard interface. The schematic diagram of the scanning device installed on the left or right side below the monitor is as follows Figure 2 The acquisition operation of the S21 scanning device can be controlled by software, supporting manual triggering and automatic triggering modes, and the installation position is not limited to the lower right corner of the display S22 is the scanning area for paper medical records.

[0090] S2. The collected paper medical records are converted into text data through OCR technology. During the OCR process, the text data retains the location information of the medical entity to facilitate traceability in subsequent steps. After the text data conversion is completed, the text is first preprocessed through the preset text preprocessing rules, including text pre-structuring, invalid character filtering, and patient identity information de-identification.

[0091] S3, large model training flow chart as follows Figure 3a shown.

[0092] The large model training process of the present invention is divided into multiple steps to ensure that the generated medical record structured data meets the accuracy and consistency requirements.

[0093] S31. The system collects medical record data from various sources, including various types of medical record reports, such as admission records, genetic testing reports, diagnosis reports, etc.

[0094] S32. Entity annotation and data enhancement are performed on the collected medical record data. The annotation content covers report categories, disease types, and medical entities, etc., to ensure that the model can correctly identify key information in medical texts. At the same time, through data enhancement technology, the diversity of training data is increased and the generalization ability of the model is improved.

[0095] S33, the labeled data is divided into training set and test set for training and verification of the large model. The training set is used to train the large model, while the test set is used to evaluate the performance of the model.

[0096] S34. Use the divided training set to train the pre-trained medical model on the medical record structured task. The model processes the medical record text in blocks, and uses the examination date or single examination as a unit to identify and extract various information in the medical record (such as diseases, examination results, genetic testing information, medication status, etc.).

[0097] S35. After training, the model will be validated and evaluated using the test set to determine the accuracy and performance of the model. The evaluation results are used to determine whether the model has achieved the expected performance standards.

[0098] S36. If the model does not meet the predetermined performance criteria, the model parameters will be adjusted and trained and evaluated again. This process will be iterated until the model meets the set performance criteria.

[0099] After the model training is completed, the medical record text data is input into the structured model (LLM, Large Language Model). The processing flow of a single medical record text is as follows: Figure 3b shown.

[0100] S31', determine the disease to which the text belongs, and obtain a corresponding list of entities to be extracted according to the disease, the entity list containing all entity information contained in the disease;

[0101] S32', determine the report type, and obtain an extraction template according to the report type, the extraction template including what content needs to be extracted under the report type;

[0102] S33', extracting medical entities in sequence according to the extraction templates determined in S31' and S32', including identifying and extracting information such as medication status and laboratory tests, and generating structured data;

[0103] S34', finally performing a predetermined format check and a medical entity integrity check on the generated structured data to facilitate subsequent processing and system integration.

[0104] S4. For the preliminary recognition result of a single medical record, the present invention introduces a rule engine to perform standardized processing.

[0105] S41. Obtain a standard medical terminology library based on disease classification, and organize the preliminary recognition results into a key-value pair format, where the key represents the medical entity category and the value represents the specific medical entity information;

[0106] S42. Use standard medical terminology to standardize the mapping of these entity information, including the normalization of synonyms, to ensure that the extracted data is consistent with the standardized medical terminology;

[0107] S4', In particular, for entities that cannot be matched using the rule engine, a normalization model combined with knowledge graph training is introduced to normalize different medical expressions. For example, when "EGFR exon 19 deletion" or "EGFRexon 19deletion" appears in the text, the normalization rule model can automatically normalize it to the standardized term "19del" to cope with more complex and diverse term expressions. This flexibility is difficult to fully cover by the preset rules.

[0108] S5. In order to ensure the integrity and consistency of the information integration of multiple examination reports of patients, the present invention introduces predefined merging rules and uses a rule engine to assist in merging multiple medical records of patients.

[0109] S51. Summarize the processing results of a single medical record;

[0110] Then, multiple medical records are merged according to the preset patient merging rules to form the patient's complete diagnosis and examination information;

[0111] S52. The generated structured data has good format and consistency and can be easily integrated into other medical systems for clinical decision support and medical research.

[0112] S6. Obtain the final structured data and connect the structured data to the existing medical system.

[0113] The above are only some embodiments of the present invention. For those skilled in the art, various modifications, improvements and refinements can be made without departing from the spirit and scope of the present invention. These equivalent changes and modifications should also be included in the protection scope of the present invention.

Claims

1. A system for structuring medical records based on a large model, characterized in that: include: Hardware acquisition device, data processing module, large model structuring module, rule engine module and patient information integration module; The steps for processing paper medical records include: S1. Scan paper medical records using a hardware acquisition device to collect image data; S2, in the data processing module, converting the image data collected by S1 into medical record text data; S3, in the large model structuring module, the text data obtained in S2 is used as the input of the large model to generate structured data of a single medical record; S4, in the rule engine module, standardize the structured data of the single medical record obtained in S3; S5. In the patient information integration module, multiple medical records of the same patient are integrated to form complete patient diagnosis and examination information; In S3, first obtain the pre-trained medical base model Base Model, then perform medical data structured training on the medical model to obtain a medical record structured model; the training process includes: S31, collecting medical record text data; these data include various types of medical record reports; S32, performing entity labeling on the collected data; S33, dividing the labeled data into a training set and a test set; S34. Use the training set to train the pre-trained medical model on the medical record structuring task; The medical model processes case text data in blocks: based on the examination date or a single examination, various information in the medical record is identified and structured, i.e., classified, extracted; the information in the medical record includes diseases, examination results, genetic testing information, and medication status; S35. Use the test set to verify and evaluate the model. The evaluation results are used to determine whether the model has achieved the expected performance standards. S36: If the model does not reach the predetermined performance index, the parameters of the model are adjusted, and the process returns to step S34 for training again; Repeat S34 to S36 until the model reaches the set performance standard, save the model parameters, and obtain the medical record structured model; In S3, the process of processing a single medical record using the medical record structured model includes: S31', the medical case text data obtained in S2 is used as the input of the model, and the model is used to determine the disease to which the case text data belongs; then the corresponding list of medical entity information to be extracted is obtained according to the disease; the medical entity information list contains the medical entity information corresponding to the disease; S32', using the model to determine the type of medical record report, and determining the corresponding extraction template according to the type of medical record report; the extraction template contains the content to be extracted under the corresponding medical record report type; S33', extracting corresponding medical entity information from the medical entity information list according to the extraction template, and generating structured data; S34', verify the generated structured data; if the verification fails, return to S33'.

2. The system for structuring medical records based on a large model according to claim 1, characterized in that: The hardware acquisition device is a device with high-resolution image acquisition capability.

3. The system for structuring medical records based on a large model according to claim 1, characterized in that: In the data processing module, the collected paper medical record images are first converted into text data through the optical character recognition (OCR) method; during the OCR processing, the text data retains the location information of the medical entity; The text is then preprocessed using text preprocessing rules, which includes text pre-structuring, invalid character filtering, and patient identity information de-identification.

4. The system for structuring medical records based on a large model according to claim 1, characterized in that: In S4, the structured data of the single medical record obtained in S3 is standardized using preset rules or a normalized rule model, and the extracted medical entities are mapped to a standard medical terminology library to form a standardized mapping relationship; wherein: For structured data that is subject to preset rules, the processing steps include: S41, obtaining a standard medical terminology library according to the disease classification, and organizing the structured data of the single medical record obtained in S3, i.e., the preliminary recognition result, into a key-value pair format, wherein the key represents the medical entity category and the value represents the specific medical entity information; S42, using a standard medical terminology database to perform standardized mapping on the medical entity information obtained in S41, so that the data is consistent with the standardized medical terminology; For structured data that is not applicable to preset rules, a normalization rule model combined with knowledge graph training is used to normalize different medical expressions; The preset rules include standardization options for specific medical entities, ensuring consistent mapping of medical terms in a specific field.

5. The system for structuring medical records based on a large model according to claim 1, characterized in that: In S5, first, the processing results of each individual medical record of the patient in S1 to S4 are summarized; Then, the multiple medical records of the patient are analyzed according to the patient information merging rules to identify the key information items that need to be integrated; Finally, the patient medical record data is integrated according to the merging rules.

6. The system for structuring medical records based on a large model according to claim 1, characterized in that: In S32, the labeled data is enhanced and then processed by S33; In S34', the verification includes whether the structured data complies with the predetermined format and whether it complies with the medical entity integrity.

Images