Hybrid expert model-driven medical large model discharge settlement generation system and method

Abstract

The invention discloses a mixed expert model-driven medical large model discharge settlement generation method and system, and the system comprises an acquisition module, a labeling module, a model training module, and a generation module, and the acquisition module is used for data acquisition and preprocessing; through rule and expert model review, the consistency and high quality of a data labeling result are ensured; the model training module is used for training a Chinese medical large model, and integrates a plurality of field expert networks in a large-scale pre-training language model by adopting a hybrid expert model technology; and the generation module is used for generating a discharge nodule, and comprises the following steps: the system receives the input of the patient, processes the input data of the Chinese medical large model in the model training module, and generates a personalized discharge nodule. By adopting the method and the system, the generation efficiency and the quality of the hospital discharge notes are remarkably improved, the personalized hospital discharge notes can be generated according to the specific condition of each patient, and nursing suggestions and follow-up visit plans which better meet the requirements are provided.

Description

Technical Field

[0001] The present invention relates to the field of machine learning, and in particular to a system and method for generating a discharge summary of a large medical model driven by a hybrid expert model. Background Art

[0002] In the modern healthcare system, the discharge summary is a key medical document provided by the patient when he or she is discharged from the hospital. It contains the patient's diagnosis, treatment, nursing advice, and follow-up plan. The discharge summary plays an important role in the patient's recovery process. It not only provides doctors with information about the patient's condition and treatment, but also helps patients understand their recovery and care plans. In addition, the discharge summary also includes post-discharge nursing advice and follow-up plans, which helps patients to perform proper self-care at home and reduce the risk of relapse and re-hospitalization.

[0003] In recent years, the computer field, especially large model technology, has made significant progress. Large-scale pre-trained language models represented by ChatGPT have powerful natural language processing capabilities through deep learning of massive text data. These models can not only understand and generate complex texts, but also perform a variety of tasks such as language translation, question-answering systems, and text summarization. The Mixture of Experts (MoE) in large models can efficiently switch between different tasks by integrating multiple expert networks in the model, thereby improving the processing power and efficiency of the model.

[0004] In the medical field, the application of AI technology has also made important progress. For example, AI can identify abnormalities and improve the accuracy of diagnosis by analyzing medical images (such as X-rays, CT scans, and MRIs). In addition, AI-assisted clinical decision support systems can provide real-time advice to medical staff, helping them make more accurate treatment decisions, thereby improving treatment effectiveness and efficiency. For example, AI can assist in the diagnosis of complex diseases, optimize treatment plans, and provide patient education and support through virtual assistants.

[0005] However, despite the great potential of AI in the medical field, existing methods for generating discharge summaries still have some obvious shortcomings:

[0006] 1. Time-consuming: Writing a discharge summary takes a long time, which increases the workload of doctors and affects the operating efficiency of the hospital.

[0007] 2. Inconsistent quality: The quality of discharge summaries written by different doctors varies greatly and lacks standardization, which can easily lead to missing or inaccurate information.

[0008] 3. Low degree of personalization: Traditional template-based methods make it difficult to provide personalized discharge summaries based on the specific circumstances of each patient and cannot fully meet the individual needs of patients.

[0009] 4. Lack of patient-friendliness: Current discharge summaries are often written in professional terms, difficult to understand, and cannot effectively convey post-discharge care information to patients.

[0010] In the existing methods of generating discharge summaries, problems such as low efficiency, inconsistent information, and complex data processing have always existed. These problems mainly stem from the over-reliance on manual operations of medical staff and the system's inability to extract in-depth information about patients. This not only affects the quality of discharge summaries, but also reduces the work efficiency of medical staff and the satisfaction of patients. Summary of the invention

[0011] The purpose of the present invention is to provide a system and method for generating a discharge summary of a large medical model driven by a hybrid expert model to solve the problems raised in the above-mentioned background technology.

[0012] To achieve the above-mentioned purpose, one aspect of the present invention provides a medical large model discharge summary generation system driven by a hybrid expert model, comprising a collection module, a labeling module, a model training module, and a generation module, wherein:

[0013] The acquisition module is used for data acquisition and preprocessing, including: obtaining basic patient information, medical records, imaging data and follow-up records from hospital information systems, electronic medical record systems and image storage systems; achieving real-time and accurate data acquisition through API interfaces and data exchange protocols; cleaning, standardizing, labeling and integrating data to ensure data consistency and high quality; data is desensitized and encrypted to protect privacy; and data quality checks are performed to ensure that input data is reliable.

[0014] The annotation module is used to efficiently obtain high-quality annotated data, including: combining multiple large language models to generate soft label candidate results; ensuring the consistency and high quality of data annotation results through rule and expert model review;

[0015] The model training module is used for Chinese medical large model training. It uses hybrid expert model technology to integrate multiple domain expert networks in a large-scale pre-trained language model. Each expert network focuses on different domain tasks. Routing experts dynamically assign tasks to domain experts, and train based on high-quality annotated data to generate a model suitable for discharge summary generation.

[0016] The generation module is used to generate a discharge summary, including: the system receives the patient's input, uses the Chinese medical large model input data in the model training module to process, and generates a personalized discharge summary.

[0017] Furthermore, the annotation module includes the following components:

[0018] The candidate component is used to generate soft label candidate results. It uses a large language model to generate candidate data. ChatGPT is used to generate preliminary text and provide basic language processing capabilities. SFT LLM is used to generate data using pre-trained models in specific fields to increase professionalism. Med LLM is used to specifically process medical data and generate more accurate medical texts.

[0019] The evaluation component is used for data quality assessment and correction. The quality of candidate data is evaluated and corrected through the following methods: ChatGPT is used to evaluate text quality and consistency and check whether the language expression is clear; Med LLM is used to evaluate in combination with medical knowledge to ensure the accuracy of medical terminology and content; rule verification is used to apply predefined rules to correct data to ensure that the data complies with medical specifications and standards;

[0020] The annotation output component is used to output high-quality annotation data that has been evaluated and corrected for subsequent use.

[0021] Furthermore, the model training module includes the following components:

[0022] The input layer is used to process the input text through the tokenizer, split it into multiple tokens, and map each token into a word vector of fixed size for model processing;

[0023] The decoder encoding layer is used to process the input high-dimensional vector layer by layer through the decoder encoding layer, extract deep-level features, and provide rich feature representation for the output layer. The decoder encoding layer consists of multiple decoders;

[0024] The output layer is used to predict the most likely token based on the semantic features obtained by the decoder encoding layer and output it.

[0025] Furthermore, the decoder includes a routing expert layer, a mixed expert layer, a shared self-attention layer, a residual and normalization layer, wherein:

[0026] The routing expert layer is responsible for dynamically selecting the most appropriate expert model based on the input data;

[0027] The mixed expert layer consists of multiple expert models, each focusing on a specific task;

[0028] Shared self-attention layer, used to provide global context information and enhance the representation ability of input data;

[0029] The residual and normalization layers are used to weight and normalize the output of the expert model to ensure the consistency and stability of the output.

[0030] Furthermore, the contents of the discharge summary include diagnosis results, treatment plan and process, post-discharge care recommendations, and follow-up plan; the treatment plan and process include drug therapy, surgery, and physical therapy.

[0031] Another aspect of the present invention provides a method for generating a discharge summary of a medical model driven by a hybrid expert model, comprising the following steps:

[0032] Step S1, data collection and preprocessing, including: obtaining the patient's basic information, medical records, imaging data and follow-up records from the hospital information system, electronic medical record system and image storage system; realizing real-time and accurate data collection through API interface and data exchange protocol; cleaning, standardizing, labeling and integrating data to ensure data consistency and high quality; data is desensitized and encrypted to protect privacy; data quality checks are performed to ensure that input data is reliable;

[0033] Step S2, efficient acquisition of high-quality annotated data, including: combining multiple large language models to generate soft label candidate results; ensuring the consistency and high quality of data annotation results through rule and expert model review;

[0034] Step S3, training of a large Chinese medical model, including: using a hybrid expert model technology to integrate multiple domain expert networks in a large-scale pre-trained language model, each expert network focusing on different domain tasks; dynamically assigning tasks to domain experts through routing experts, training based on high-quality annotated data, and generating a model suitable for generating discharge summaries;

[0035] Step S4, generating a discharge summary, includes: the system receives the patient's input, uses the Chinese medical model input data in step S3 to process, and generates a personalized discharge summary.

[0036] Further, step S2 includes the following steps:

[0037] Step S201, generating soft label candidate results, is to use a large language model to generate candidate data, where ChatGPT is used to generate preliminary text and provide basic language processing capabilities; SFT LLM is used to generate data using a pre-trained model in a specific field to increase professionalism; Med LLM is used to specifically process medical data and generate more accurate medical text;

[0038] Step S202, data quality assessment and correction, the quality of candidate data is assessed and corrected by the following methods, wherein ChatGPT is used to assess text quality and consistency and check whether the language expression is clear; Med LLM is used to evaluate in combination with medical knowledge to ensure the accuracy of medical terminology and content; rule verification is used to apply predefined rules to correct the data to ensure that the data complies with medical specifications and standards;

[0039] Step S203: output the evaluated and corrected high-quality labeled data for subsequent use.

[0040] Further, step S3 includes the following steps:

[0041] Step S301, the input text is processed by a tokenizer, divided into multiple tokens, and each token is mapped to a word vector of a fixed size dimension for model processing;

[0042] Step S302, through the decoder encoding layer, the input high-dimensional vector is processed layer by layer, deep-level features are extracted, and rich feature representation is provided for the output layer. The decoder encoding layer is composed of multiple decoders;

[0043] Step S303: predict the most likely token based on the semantic features obtained by the decoder encoding layer and output it.

[0044] Furthermore, the decoder includes a routing expert layer, a mixed expert layer, a shared self-attention layer, a residual and normalization layer, wherein:

[0045] The routing expert layer is responsible for dynamically selecting the most appropriate expert model based on the input data;

[0046] The mixed expert layer consists of multiple expert models, each focusing on a specific task;

[0047] Shared self-attention layer, used to provide global context information and enhance the representation ability of input data;

[0048] The residual and normalization layers are used to weight and normalize the output of the expert model to ensure the consistency and stability of the output.

[0049] Furthermore, the contents of the discharge summary include diagnosis results, treatment plan and process, post-discharge care recommendations, and follow-up plan; the treatment plan and process include drug therapy, surgery, and physical therapy.

[0050] Compared with the prior art, the present system and method have the following advantages:

[0051] 1. This invention adopts a large language model based on a hybrid expert model, integrating multiple expert networks in a large-scale pre-trained language model, and improving the model's ability and efficiency in processing different tasks. Compared with the single model of the traditional method, the hybrid expert model significantly improves processing efficiency and accuracy by dynamically assigning tasks to different expert networks.

[0052] 2. The present invention introduces routing experts, which dynamically allocate domain experts according to the importance or difficulty of each token. Routing experts not only improve the training efficiency of the model, but also optimize task allocation and ensure load balancing.

[0053] 3. Through deep information extraction and understanding, the present invention can generate personalized discharge summaries based on the specific circumstances of each patient, provide nursing advice and follow-up plans that are more tailored to the needs, and improve patient compliance and satisfaction; ensure the consistency and standardization of discharge summaries to avoid information omissions and inaccuracies; generate clear and accurate content, and present it in a way that is easy for patients to understand, solving the problem of difficult-to-understand professional terms;

[0054] 4. The present invention adopts a variety of large language models to generate candidate data, and ensures data quality through automatic evaluation and correction of large language models and rules, and finally outputs high-quality labeled data, thereby achieving efficient and low-cost acquisition of high-quality data.

[0055] 5. Since each expert model is relatively independent, the model decision process of the present invention is easier to explain and understand, provides higher explainability, and enhances the trust of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 Flowchart of the method for generating discharge summaries for a large medical model driven by a hybrid expert model.

[0057] Figure 2 The flowchart for obtaining high-quality labeled data in step S2.

[0058] Figure 3 This is the architecture diagram of the Chinese medical big model driven by a hybrid expert model.

[0059] Figure 4 Schematic diagram of the entire process of generating discharge summaries based on a Chinese medical big model driven by a hybrid expert model. DETAILED DESCRIPTION

[0060] 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.

[0061] like Figure 1 The method flow chart of the embodiment of the present invention is shown, and the specific steps are as follows:

[0062] Step S1, data collection and preprocessing. Obtain the patient's basic information, medical records, imaging data and follow-up records from the hospital information system, electronic medical record system and image storage system. Realize real-time and accurate data collection through API interface and data exchange protocol. Data preprocessing includes data cleaning, standardization, labeling and integration to ensure data consistency and high quality. In order to protect privacy, the data is desensitized and encrypted. Data quality checks cover completeness, consistency and timeliness to ensure the reliability of input data.

[0063] Step S2, efficient acquisition of high-quality annotated data. After data collection and preprocessing, soft label candidate results are generated by combining multiple large language models (such as domain SFT LLM, open source LLM, etc.). Through rule and expert model review, the consistency and high quality of data annotation results are ensured. Finally, high-quality data is output as annotated data set.

[0064] Step S3, training of the Chinese medical large model. Using the hybrid expert model technology, multiple domain expert networks are integrated into the large-scale pre-trained language model, and each expert network focuses on different domain tasks; and tasks are dynamically assigned to domain experts through routing experts, and training is performed based on high-quality annotated data to generate a model suitable for discharge summary generation.

[0065] Step S4, discharge summary generation. The system receives the patient's medical history data, diagnosis information, treatment records and other inputs, processes the input data using the hybrid expert model, and generates a personalized discharge summary, including diagnosis results, treatment plans, nursing recommendations and follow-up plans. Ensure that the generated discharge summary is clear and accurate, and is presented in a way that is easy for patients to understand, thereby improving patient compliance and satisfaction.

[0066] like Figure 2 The flowchart of obtaining high-quality labeled data in step S2 of the above embodiment is shown, which includes the following steps:

[0067] Step S201, obtain the patient's basic information, medical records, treatment records and other data from the hospital information system, and generate candidate results after data cleaning.

[0068] Data cleaning is to clean and standardize the collected data to ensure the consistency and high quality of the data, including deleting duplicate records, processing missing values, and standardizing data formats.

[0069] Generating candidate results is to generate candidate data using multiple language models, among which:

[0070] ChatGPT is used to generate preliminary text and provide basic language processing capabilities;

[0071] SFT LLM (Supervised Fine-Tuning Large Language Model) uses pre-trained models in specific fields to generate data and increase professionalism;

[0072] Med LLM (Medical Large Language Model) is a language model designed specifically for medical data to generate more accurate medical texts.

[0073] Step S202: evaluate and calibrate the data quality. The quality of the candidate data is evaluated and calibrated by the following method:

[0074] ChatGPT is used to evaluate text quality and consistency and check whether the language expression is clear;

[0075] Med LLM is used to combine medical knowledge for assessment and ensure the accuracy of medical terminology and content;

[0076] Rule verification is the process of applying predefined rules to correct data to ensure that the data complies with medical specifications and standards.

[0077] Step S203, output high-quality labeled data. Output the evaluated and corrected high-quality labeled data for subsequent use. These data will be used to train and optimize the hybrid expert large model to improve the accuracy and reliability of the model.

[0078] like Figure 3 The figure shows the architecture of the Chinese Medicine Large Model (Med-LLM-MoE) driven by a hybrid expert model. It is mainly divided into the input layer, the decoder encoding layer and the output layer. The task of this model is to achieve efficient and accurate automatic generation of discharge summaries by training Med-LLM-MoE. This includes improving processing efficiency, enhancing generation accuracy, achieving personalized output, and reducing the paperwork burden of doctors, thereby improving the overall quality of medical services and patient satisfaction.

[0079] The expected input of this model is large-scale medical literature and teaching materials, clinical medical data, expert experience text data, and high-quality annotated data. The input layer is used to process the input text through a word segmenter, split it into multiple tokens, and map each token to a fixed-size word vector for model processing. The decoder encoding layer consists of multiple decoders, each of which contains a router expert layer, a shared self-attention layer, a hybrid expert layer, and a residual and normalization layer. Through these decoder layers, the input high-dimensional vector is processed layer by layer, deep features are extracted, and rich feature representations are provided for the output layer. The expected output of the model is a hybrid expert model with high accuracy, personalization, high consistency, and high efficiency. The model can quickly generate accurate diagnosis results, treatment plans, nursing recommendations, and follow-up plans, provide personalized discharge summaries, and ensure that the content complies with medical standards and clinical practices. The output layer predicts the most likely token based on the semantic features obtained by the decoder encoding layer and uses it as output.

[0080] The routing expert layer is responsible for dynamically selecting the most appropriate expert model (Expert) based on the input data. The hybrid expert layer includes multiple expert models, each of which focuses on a specific task. The routing expert selects appropriate experts to participate in the calculation based on the characteristics of the input data. The shared self-attention layer provides global context information and enhances the representation ability of the input data. The residual and normalization layers weight and normalize the output of the expert model to ensure the consistency and stability of the output.

[0081] like Figure 4 As shown in the figure, the schematic diagram of the whole process of generating discharge summaries based on the Chinese medical big model driven by the hybrid expert model is shown in the figure. The specific description is as follows:

[0082] First, obtain patient data. This system first obtains medical data such as the patient's basic information, medical records, and treatment records from the hospital information system. The patient's basic information includes the patient's name, age, gender, and other basic personal information. The medical record should cover the patient's detailed medical history information, including previous diseases, allergies, surgical history, etc. The treatment record records all the diagnosis and treatment processes of the patient during his stay in the hospital, including examination results, diagnosis information, treatment measures, etc.

[0083] The data is then fed into the Chinese Medicine Big Model for processing. The model uses hybrid expert model technology to integrate multiple expert networks, each of which focuses on a specific medical task, thereby improving processing efficiency and accuracy.

[0084] The model generates a discharge summary that first lists the diagnostic results during the implementation period in detail to ensure the accuracy and completeness of the information. Secondly, it describes the treatment plan and process used, including medication, surgery, and physical therapy. At the same time, it provides post-discharge care recommendations to help users properly perform self-care and reduce the risk of relapse and re-admission. Finally, a detailed follow-up plan is developed to continuously monitor the user's recovery and adjust the treatment plan in a timely manner as needed.

[0085] The present invention significantly improves the generation efficiency and quality of discharge summaries through a hybrid expert model and a dynamic task allocation mechanism, reducing the workload of doctors; it can generate personalized discharge summaries based on the specific circumstances of each patient, provide nursing advice and follow-up plans that are more in line with the needs, and improve patient compliance and satisfaction; ensure the consistency and standardization of discharge summaries to avoid information omissions and inaccuracies; generate clear and accurate content, and present it in a way that is easy for patients to understand, solving the problem of difficult-to-understand professional terms; and improve the overall performance of the system through efficient acquisition of high-quality annotated data and deep information extraction. Through the above technical scheme, the discharge summary generation method of the present invention can significantly improve the generation efficiency, quality and personalization of discharge summaries, reduce the workload of doctors, improve patient satisfaction and compliance, and promote the development of intelligent medical care.

[0086] The user usage scenarios of the present invention include but are not limited to the following:

[0087] 1. Hospitals and medical institutions. Doctors and medical staff can use the system to quickly generate high-quality discharge summaries, reduce paperwork, and improve work efficiency.

[0088] 2. Family doctors and community medical services. Family doctors and community medical service personnel can use the system to provide patients with personalized care advice and follow-up plans to improve service quality.

[0089] 3. Insurance companies. Insurance companies can use the system to review discharge summaries and ensure that medical information during the claims process is accurate and consistent.

[0090] 4. Patients and their families. Patients and their families can use the system to obtain easy-to-understand discharge summaries, learn about nursing advice and follow-up plans during the rehabilitation process, and improve their self-care capabilities.

[0091] The present invention can also be applied to the following products:

[0092] 1. Electronic medical record system (EMR). The electronic medical record system integrating this patented technology can automatically generate high-quality discharge summaries and improve the intelligence level of the system.

[0093] 2. Clinical Decision Support System (CDSS). This technology can be applied to the clinical decision support system to provide medical staff with real-time discharge summary generation and nursing recommendations, thus optimizing the clinical decision-making process.

[0094] 3. Patient management system. A patient management system that integrates this technology can provide patients with personalized discharge summaries and nursing recommendations, improving the effectiveness of patient management.

[0095] 4. Insurance claims system: Insurance companies can apply this technology in the claims system to automatically review discharge summaries and improve the accuracy and efficiency of claims review.

[0096] 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 medical large model discharge summary generation system driven by a hybrid expert model, characterized by: It includes acquisition module, annotation module, model training module and generation module, among which: The acquisition module is used for data acquisition and preprocessing, including: obtaining basic patient information, medical records, imaging data and follow-up records from hospital information systems, electronic medical record systems and image storage systems; achieving real-time and accurate data acquisition through API interfaces and data exchange protocols; cleaning, standardizing, labeling and integrating data to ensure data consistency and high quality; data is desensitized and encrypted to protect privacy; and data quality checks are performed to ensure that input data is reliable. The annotation module is used to efficiently obtain high-quality annotated data, including: combining multiple large language models to generate soft label candidate results; ensuring the consistency and high quality of data annotation results through rule and expert model review; The model training module is used for Chinese medical large model training. It uses hybrid expert model technology to integrate multiple domain expert networks in a large-scale pre-trained language model. Each expert network focuses on different domain tasks. Routing experts dynamically assign tasks to domain experts, and train based on high-quality annotated data to generate a model suitable for discharge summary generation. The generation module is used to generate a discharge summary, including: the system receives the patient's input, uses the Chinese medical large model input data in the model training module to process, and generates a personalized discharge summary.

2. The medical large model discharge summary generation system driven by a hybrid expert model according to claim 1 is characterized in that: The annotation module includes the following components: The candidate component is used to generate soft label candidate results. It uses a large language model to generate candidate data. ChatGPT is used to generate preliminary text and provide basic language processing capabilities. SFT LLM is used to generate data using pre-trained models in specific fields to increase professionalism. Med LLM is used to specifically process medical data and generate more accurate medical texts. The evaluation component is used for data quality assessment and correction. The quality of candidate data is evaluated and corrected through the following methods. ChatGPT is used to evaluate text quality and consistency and check whether the language expression is clear; Med LLM is used to evaluate in combination with medical knowledge to ensure the accuracy of medical terminology and content; Rule verification is used to apply predefined rules to correct data to ensure that the data complies with medical regulations and standards; The annotation output component is used to output high-quality annotation data that has been evaluated and corrected for subsequent use.

3. The medical large model discharge summary generation system driven by a hybrid expert model according to claim 1 is characterized in that: The model training module includes the following components: The input layer is used to process the input text through the tokenizer, split it into multiple tokens, and map each token into a word vector of fixed size for model processing; The decoder encoding layer is used to process the input high-dimensional vector layer by layer through the decoder encoding layer, extract deep-level features, and provide rich feature representation for the output layer. The decoder encoding layer consists of multiple decoders; The output layer is used to predict the most likely token based on the semantic features obtained by the decoder encoding layer and output it.

4. The medical large model discharge summary generation system driven by a hybrid expert model according to claim 3 is characterized in that: The decoder includes a routing expert layer, a mixing expert layer, a shared self-attention layer, a residual and a normalization layer, where: The routing expert layer is responsible for dynamically selecting the most appropriate expert model based on the input data; The mixed expert layer consists of multiple expert models, each focusing on a specific task; Shared self-attention layer, used to provide global context information and enhance the representation ability of input data; The residual and normalization layers are used to weight and normalize the output of the expert model to ensure the consistency and stability of the output.

5. The medical model discharge summary generation system driven by a hybrid expert model according to claim 1, characterized in that: The contents of the discharge summary include the diagnosis results, treatment plan and process, post-discharge care recommendations, and follow-up plan; the treatment plan and process include drug therapy, surgery, and physical therapy.

6. A method for generating discharge summaries of a large medical model driven by a hybrid expert model, characterized in that: The following steps are involved: Step S1, data collection and preprocessing, including: obtaining the patient's basic information, medical records, imaging data and follow-up records from the hospital information system, electronic medical record system and image storage system; realizing real-time and accurate data collection through API interface and data exchange protocol; cleaning, standardizing, labeling and integrating data to ensure data consistency and high quality; data is desensitized and encrypted to protect privacy; data quality checks are performed to ensure that input data is reliable; Step S2, efficient acquisition of high-quality annotated data, including: combining multiple large language models to generate soft label candidate results; ensuring the consistency and high quality of data annotation results through rule and expert model review; Step S3, training of a large Chinese medical model, including: using a hybrid expert model technology to integrate multiple domain expert networks in a large-scale pre-trained language model, each expert network focusing on different domain tasks; dynamically assigning tasks to domain experts through routing experts, training based on high-quality annotated data, and generating a model suitable for generating discharge summaries; Step S4, generating a discharge summary, includes: the system receives the patient's input, uses the Chinese medical model input data in step S3 to process, and generates a personalized discharge summary.

7. The method for generating a discharge summary of a medical model driven by a hybrid expert model according to claim 6, characterized in that: Step S2 includes the following steps: Step S201, generating soft label candidate results, is to use a large language model to generate candidate data, where ChatGPT is used to generate preliminary text and provide basic language processing capabilities; SFT LLM is used to generate data using a pre-trained model in a specific field to increase professionalism; Med LLM is used to specifically process medical data and generate more accurate medical text; Step S202, data quality assessment and correction, the quality of candidate data is assessed and corrected by the following methods, wherein ChatGPT is used to assess text quality and consistency and check whether the language expression is clear; Med LLM is used to evaluate in combination with medical knowledge to ensure the accuracy of medical terminology and content; rule verification is used to apply predefined rules to correct the data to ensure that the data complies with medical specifications and standards; Step S203: output the evaluated and corrected high-quality labeled data for subsequent use.

8. The method for generating a discharge summary of a medical model driven by a hybrid expert model according to claim 6, characterized in that: In step S3 The following steps are included: Step S301, the input text is processed by a tokenizer, divided into multiple tokens, and each token is mapped to a word vector of a fixed size dimension for model processing; Step S302, through the decoder encoding layer, the input high-dimensional vector is processed layer by layer, deep-level features are extracted, and rich feature representation is provided for the output layer. The decoder encoding layer is composed of multiple decoders; Step S303: predict the most likely token based on the semantic features obtained by the decoder encoding layer and output it.

9. The method for generating a discharge summary of a medical model driven by a hybrid expert model according to claim 8, characterized in that: The decoder includes a routing expert layer, a mixing expert layer, a shared self-attention layer, a residual and a normalization layer, where: The routing expert layer is responsible for dynamically selecting the most appropriate expert model based on the input data; The mixed expert layer consists of multiple expert models, each focusing on a specific task; Shared self-attention layer, used to provide global context information and enhance the representation ability of input data; The residual and normalization layers are used to weight and normalize the output of the expert model to ensure the consistency and stability of the output.

10. The method for generating a discharge summary of a medical model driven by a hybrid expert model according to claim 6, characterized in that: The contents of the discharge summary include the diagnosis results, treatment plan and process, post-discharge care recommendations, and follow-up plan; the treatment plan and process include drug therapy, surgery, and physical therapy.

Images