A copd structured diagnosis and treatment auxiliary method based on multi-llm cooperation

Abstract

This invention discloses a structured diagnostic and treatment assistance method for COPD based on multi-LLM collaboration, belonging to the field of computer-aided medical diagnostic technology. The method includes: constructing a specialized dataset with task separation; fine-tuning a knowledge-based question-answering model and an auxiliary diagnostic model; performing semantic analysis on user input using a general model to generate routing control parameters; scheduling corresponding specialized models based on a preset structured diagnostic and treatment process; utilizing retrieval enhancement generation technology to introduce authoritative medical guidelines for retrieving and constraining external knowledge; and finally outputting diagnostic and treatment assistance information through an interactive interface. This invention reduces the risk of knowledge confusion and improves the accuracy, standardization, and evidence-based nature of COPD assisted diagnosis and treatment through multi-model collaboration and process control, making it suitable for respiratory chronic disease management in primary healthcare settings.

Description

Technical Field

[0001] This invention relates to the field of computer-aided medical diagnostic technology, and in particular to a structured diagnostic and treatment assistance method for COPD based on multi-LLM collaboration. Background Technology

[0002] Chronic obstructive pulmonary disease (COPD) is a common, preventable, and treatable chronic respiratory disease with high prevalence and mortality rates worldwide. Especially in areas with relatively scarce medical resources, primary healthcare institutions generally face shortages of specialist physicians and insufficient clinical experience, leading to low early identification rates and inadequate standardized treatment for COPD and other chronic respiratory diseases.

[0003] While existing technologies have attempted to utilize artificial intelligence for disease diagnosis, most rely on a single model to handle all tasks. This can easily lead to knowledge confusion between different tasks, making it difficult to simultaneously ensure the accuracy of theoretical knowledge and the complexity of clinical practice. Furthermore, traditional intelligent consultation systems often depend on hard-coded rules, resulting in rigid processes that cannot flexibly respond to diverse user input and complex clinical scenarios. Therefore, how to build an intelligent auxiliary tool that possesses both deep professional knowledge and the ability to flexibly adhere to rigorous clinical logic is a pressing technical challenge in the field of smart healthcare. Summary of the Invention

[0004] To address the limitations of single-model capabilities and knowledge confusion in conventional COPD diagnosis and treatment, as well as the rigidity and inflexibility of traditional intelligent consultation systems, this invention proposes a structured COPD diagnosis and treatment assistance method based on multi-LLM collaboration. This method balances the accuracy of professional knowledge, the rigor of clinical logic, and the flexibility of human-computer interaction, overcoming the limitations of single models in multi-task scenarios and freeing the diagnosis and treatment process from the constraints of hard-coded rules. This effectively improves the standardized diagnosis and treatment level of COPD and other chronic respiratory diseases, thus solving the aforementioned problems.

[0005] This application discloses a structured diagnostic and treatment assistance method for COPD based on multi-LLM collaboration, including the following steps: S1. Acquire medical theoretical knowledge data and clinical case data, and through preprocessing, construct a task-separated professional dataset for multi-model fine-tuning; S2. Select an open-source base model and fine-tune the instructions using theoretical knowledge datasets and clinical case datasets respectively to obtain a knowledge question-answering model and an auxiliary diagnosis model with separated responsibilities. S3. Deploy the knowledge question answering model and the auxiliary diagnosis model, and set up a general model to perform semantic analysis on user input to output routing control parameters. Based on the control logic of the structured diagnosis and treatment process, use the routing control parameters to call and coordinate the knowledge question answering model and the auxiliary diagnosis model in an orderly manner. S4. Use authoritative medical guidelines as an external knowledge base, and enhance the model output by retrieval and generation; S5. Receive user input through a graphical user interface and, under the constraints of a structured diagnosis and treatment process, call the corresponding model to output diagnostic and treatment auxiliary information.

[0006] Preferably, the medical theoretical knowledge data includes Chinese and English medical review literature, medical textbooks, and authoritative medical guidelines; The clinical case data mentioned are real clinical case data after removing duplicates and null values.

[0007] Preferably, the preprocessing includes: For medical theoretical knowledge data, the original theoretical knowledge data is converted into structured Markdown text and split into text blocks according to semantics. For each text block, question-answer pairs are automatically constructed through a large language model. The question-answer pairs are then organized into the first JSON instruction dataset, which serves as the theoretical knowledge dataset. For clinical case data, the desensitized clinical case data is cleaned and screened, and the chief complaint and present medical history information in the case are extracted as input, and the diagnosis and medical order information are extracted as output. The input and output are organized into a second JSON instruction dataset, which serves as the clinical case dataset.

[0008] Preferably, the open-source base model is Qwen2.5 7B, and the instruction fine-tuning uses Lora.

[0009] Preferably, the knowledge question-answering model is obtained by fine-tuning the theoretical knowledge dataset with instructions, and is used to answer theoretical questions related to COPD; The auxiliary diagnostic model is obtained by fine-tuning the clinical case dataset and is used to generate preliminary diagnostic and treatment assistance results based on the patient's chief complaint and present medical history.

[0010] Preferably, the routing control parameters are a set of parameters generated by semantic analysis of user input through a general model to characterize the current diagnosis and treatment stage and status information; The parameter set includes at least: stage parameters, symptom parameters, and examination parameters.

[0011] Preferably, the control logic of the structured diagnosis and treatment process includes: When the routing control parameters indicate that the user has not made a statement, the knowledge question-answering model is invoked. When routing control parameters instruct the user to make a chief complaint, ask the user whether they have undergone any pulmonary function tests that are strongly correlated with a COPD diagnosis: If the user has not had a lung function test, guide the user to answer the COPD-SQ scale and determine whether there is a risk of COPD. If there is a risk of COPD, call the auxiliary diagnostic model. If the user has undergone pulmonary function testing, obtain the measured values ​​of FEV1%FVC and the predicted / actual values ​​of FEV1, which are relevant to the diagnosis of COPD: when and At that time, the reply to the user may be to retain the ratio of impaired lung function as a diagnostic aid; when and At that time, the reply to the user may be auxiliary information for the diagnosis and treatment of early-stage chronic obstructive pulmonary disease; when At that time, the user was initially identified as a patient with chronic obstructive pulmonary disease (COPD). The user was guided to complete the mMRC and CAT questionnaires to determine the group and stage. Subsequently, the knowledge question-answering model was called to generate a comprehensive response that included the classification, grouping, stage and treatment suggestions.

[0012] Preferably, the enhanced retrieval generation is invoked at different stages of the structured diagnosis and treatment process, specifically including: The authoritative medical guide text is segmented and converted into vectors using an embedding model before being stored in a vector database. When a response needs to be generated, the user's question or the current context is transformed into a query vector through an embedded model; Similarity matching is performed in the vector database to retrieve the text blocks most relevant to the query vector as reference information; The reference information is concatenated with the original question and then input into the currently invoked model to generate an enhanced response constrained by external knowledge.

[0013] Preferably, the embedding model is the BGE-M3 model, and the authoritative medical guideline is the GOLD guideline.

[0014] Preferably, the graphical user interface is used to receive user input in the form of text or uploaded report files, and to visually display the progress of the structured diagnosis and treatment process and the diagnosis and treatment assistance information generated by the model.

[0015] The beneficial effects of this invention are: (1) By constructing a specialized model with task separation and coordinating control under the constraints of a structured diagnosis and treatment process, this invention reduces the risk of knowledge confusion in a multi-task scenario by a single model and improves the accuracy and professionalism of the COPD diagnosis and treatment assistance process.

[0016] (2) By structurally modeling the professional consultation process, this invention enables clinical experience and diagnostic logic to be reflected in the diagnostic assistance process in a controllable manner, thereby enhancing the consistency and reliability of the diagnostic assistance results.

[0017] (3) By introducing search-enhanced generation technology, authoritative medical guidelines are used as external reference information to constrain the content generated by the model, thereby improving the evidence-based nature and credibility of diagnostic and treatment auxiliary information.

[0018] (4) The method of the present invention can assist in the standardized diagnosis and treatment of COPD in primary healthcare settings, and helps to improve the standardization level of the diagnosis and treatment process. Attached Figure Description

[0019] Figure 1 This is a flowchart of a structured diagnostic and treatment assistance method for COPD based on multi-LLM collaboration, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the structured consultation process according to an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the retrieval enhancement generation in an embodiment of the present invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments.

[0021] This application discloses a structured diagnostic and treatment assistance method for COPD based on multi-LLM collaboration, the process of which is as follows: Figure 1 As shown, it includes the following steps: S1. Acquire medical theoretical knowledge data such as medical literature and guidelines, as well as anonymized clinical case data. Through preprocessing, construct a task-separated professional dataset for multi-model fine-tuning.

[0022] In this embodiment, the theoretical knowledge data includes 167 Chinese review articles, 120 English review articles, 12 authoritative medical textbooks, and the 2024 edition of the GOLD guidelines, all selected by medical professionals. The clinical case data consists of 89,401 real-world outpatient cases obtained from the hospital information system.

[0023] Data preprocessing refers to organizing theoretical knowledge data and clinical case data into a JSON instruction dataset required for fine-tuning the large model. The JSON format required for fine-tuning is shown in Table 1.

[0024] Table 1. Components of the JSON instruction dataset required for fine-tuning

[0025] For theoretical knowledge data, original literature, textbooks, and guidelines were parsed into structured Markdown format. The theoretical knowledge dataset was constructed using the open-source large model dataset creation tool Easy Dataset. First, the data was automatically segmented into numerous text blocks based on different chapters within the Markdown format. When encountering excessively long paragraphs (exceeding the maximum segment length), a recursive segmentation algorithm was executed to ensure that each text block contained as much complete semantic information as possible. Next, for each text block, a question was automatically constructed based on the semantics of the text block using the locally deployed large model DeepSeek-R1 7B based on the Ollam tool. Then, DeepSeek-R1 7B was again used to automatically generate answers based on the text block context. Finally, the question-answer pairs obtained through Easy Dataset were compiled into a theoretical knowledge dataset of JSON instructions required for fine-tuning the large model. Specifically, in the theoretical knowledge dataset, "Instruction" represents the constructed questions, while "Output" represents the automatically generated answers. The final result was 25,153 fine-tuning instructions based on Chinese and English review literature, 2,151 fine-tuning instructions based on guidelines, and 5,631 fine-tuning instructions based on textbooks.

[0026] The construction of the clinical case dataset began with data cleaning to remove duplicates and null values. Next, data with weak relevance was filtered based on COPD-related keywords and a drug dictionary. Finally, the "chief complaint" and "present illness" fields were integrated as the "Input" of the JSON instruction clinical case dataset; the "diagnosis" and "medical order processing" fields were integrated as the "Output." The "Instruction" was uniformly set to "You are a respiratory physician; you will diagnose the patient and provide appropriate drug treatment based on the patient's chief complaint and present illness," allowing the model to clearly understand its task. All generated instruction data underwent manual review, ultimately resulting in 32,620 fine-tuning instructions for training.

[0027] S2. Select an open-source base model and fine-tune the instructions using theoretical knowledge datasets and clinical case datasets respectively to obtain a knowledge-based question-answering model and an auxiliary diagnostic model with separated responsibilities.

[0028] In this embodiment, the open-source foundation model uses Qwen2.5 7B, and instruction fine-tuning uses LoRa. The knowledge-based question-answering model is fine-tuned using a theoretical knowledge dataset to answer common COPD-related theoretical questions. The auxiliary diagnostic model is fine-tuned using a clinical case dataset, enabling it to deeply learn real-world clinical diagnostic logic, treatment decisions, and medication habits, thus obtaining more reasonable preliminary diagnostic and treatment assistance results based on the patient's chief complaint and medical history. This task-separated fine-tuning strategy ensures that each model achieves a higher level of expertise within its specific domain.

[0029] S3. Deploy the knowledge question answering model and the auxiliary diagnosis model, and set up a general model to perform semantic analysis on user input to output routing control parameters. Based on the control logic of the structured diagnosis and treatment process, use the routing control parameters to call and coordinate the knowledge question answering model and the auxiliary diagnosis model in an orderly manner.

[0030] This embodiment is based on, as follows Figure 2 The structured diagnosis and treatment process shown is used to describe the logical relationships between different stages of the consultation.

[0031] First, based on the user's question, the system determines whether the user is making a chief complaint by checking if it includes information about symptoms and duration. If the user's statement contains such information, the general model will determine that the user is making a chief complaint. Then, the general model will determine whether the user's symptoms are related to lung function. If not, under the control logic of the structured diagnosis and treatment process, a knowledge-based question-answering model trained on a theoretical knowledge dataset will be invoked to respond to the user. If the user's statement does not contain both symptom and duration information, it is determined that the user is engaging in routine consultation, and under the control logic of the structured diagnosis and treatment process, a knowledge-based question-answering model trained on a theoretical knowledge dataset will be invoked to respond to the user.

[0032] Next, if the user is presenting with symptoms related to lung function, the system will continue to collect the user's medical history and then ask if the user has undergone any lung function tests strongly correlated with a COPD diagnosis. If the user has not undergone lung function tests, the system will guide the user step-by-step through the COPD-SQ scale to assess COPD risk. During this process, global state variables will be used to store scores for different questions, and the scale score will be obtained by statistically analyzing the global state variables after the scale is completed. Based on the scale score, if there is no COPD risk, the consultation process will end directly, informing the user that there is no risk, and then the system will connect to a knowledge-based question-and-answer model to allow the user to ask other questions. If there is COPD risk, but the classification, grouping, and stage cannot be determined due to the lack of relevant lung function tests, an auxiliary diagnostic model will be invoked under the constraints of the structured treatment process. During its fine-tuning process, the auxiliary diagnostic model learns the correlation features between the chief complaint information, present medical history information, and treatment plan in previous case data. Based on the chief complaint information and present medical history information provided by the user in the preliminary consultation stage, it generates preliminary diagnosis and medication suggestions. Then, it connects to a knowledge question answering model to allow the user to ask other questions.

[0033] If the user has already undergone pulmonary function testing, they will be guided to input their measured FEV1%FVC and predicted / actual FEV1 values, which are relevant to the COPD diagnosis, or directly upload their pulmonary function test report. Next, the user will be guided to provide the result of the absolute eosinophil count (10*9 / L) from their complete blood count (CBC) report, which is also relevant to the COPD diagnosis. They can also choose to upload the report directly. However, since the CBC is an invasive procedure, users are not required to provide a result; they can also indicate their intention not to undergo the CBC, allowing the consultation to proceed to the next stage.

[0034] After obtaining the user's lung function information, a preliminary judgment can be made. If the measured FEV1%FVC is >= 70%, and the actual / predicted FEV1 is < 80%, the structured diagnostic process will directly provide the user with diagnostic support information suggesting possible impaired lung function with a retention ratio (PRISm). If the actual / predicted FEV1 is >= 80%, the structured diagnostic process will directly provide the user with diagnostic support information suggesting possible pre-COPD. Following this, a knowledge-based question-answering model will be used to answer the user's subsequent questions.

[0035] When the measured FEV1%FVC value is <70%, the user can be preliminarily identified as a COPD patient. Further questioning using the mMRC and CAT questionnaires will then confirm the user's possible classification, grouping, and stage. A knowledge-based question-and-answer model with extensive COPD-related knowledge will then provide a preliminary comprehensive response including the user's COPD classification, grouping, stage, medication recommendations, and non-pharmacological treatment suggestions. The user will be encouraged to continue asking questions about any remaining issues. Medication recommendations will also consider information provided during the conversation; if cough and phlegm-related symptoms are present, relevant medications will be suggested.

[0036] In this embodiment, the structured diagnosis and treatment process described above is divided into multiple consecutive consultation stages. For key judgment and classification tasks in different stages, a general model (driven by the Gemma-3 27B base language model) performs semantic analysis on the user input to generate a set of control parameters representing the current diagnosis and treatment stage and process status. The control parameter set includes at least: stage parameters indicating the current consultation stage, symptom parameters describing the user's symptom characteristics, and examination parameters identifying the examination information already obtained. These control parameters are stored after generation and passed as global state variables in subsequent stages of the structured diagnosis and treatment process.

[0037] The set of control parameters serves as the input to the process control logic, determining subsequent consultation stages and the corresponding professional model invocation methods. By recording and transmitting state variables at each consultation stage, the key information provided by the user is ensured to remain continuous and consistent throughout the process. This combines the semantic analysis capabilities of the large language model with the constraint mechanisms of the structured diagnosis and treatment process, achieving refined control over the overall consultation process.

[0038] S4. Use authoritative medical guidelines (GOLD guidelines) as an external knowledge base, and constrain and enhance the model output through Retrieval Enhancement Generation (RAG).

[0039] Specifically, such as Figure 3 As shown, Retrieval Enhancement Generation (RAG) first prepares the necessary external knowledge (this embodiment uses the 2024 GOLD Guidelines), processes it into chunks, and then uses the open-source embedding model BGE-M3 to convert each chunk into a machine-understandable vector form, storing it in a vector database as an external knowledge base. Next, when a user asks a question, the same embedding model BGE-M3 converts the user's question into a vector form, performs similarity matching with the vectors in the external knowledge base, and finds the most relevant information. Finally, the found relevant information, along with the user's original question, is sent to the knowledge question answering model, allowing the model to prioritize the use of this information when generating an answer, thereby enhancing the accuracy of the question answer.

[0040] S5. Receive user input through a graphical user interface and, under the constraints of a structured diagnosis and treatment process, call the corresponding model to output diagnostic and treatment auxiliary information.

[0041] Specifically, after a user submits input information through a graphical user interface, the structured diagnostic process processes the user input. First, the general model performs semantic analysis on the user input, generating process control parameters to represent the current consultation stage. When the process control parameters indicate that the user input is related to theoretical knowledge, a knowledge-based question-answering model is invoked to generate the corresponding theoretical answer. When the process control parameters indicate that the user input is related to symptom description, the structured diagnostic process guides the user to provide further diagnostic information, and, if necessary, an auxiliary diagnostic model is invoked to generate preliminary diagnostic assistance results. During the generation of diagnostic assistance information, when the stage of the structured diagnostic process requires reference to authoritative medical guidelines, the retrieval enhancement generation in S4 is invoked to provide reference constraints based on the external knowledge base for the model output. Finally, the invoked model generates diagnostic assistance information under the constraints of the structured diagnostic process and outputs it to the user through the graphical user interface.

[0042] In summary, this application provides a structured diagnostic and treatment assistance method for COPD based on multi-LLM collaboration. This method designs a structured diagnostic and treatment approach that follows a professional consultation process, incorporating the clinical experience and diagnostic and treatment logic of experts in a controllable manner into the diagnostic and treatment assistance process, thereby ensuring the rigor and safety of the diagnostic and treatment assistance process.

[0043] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A structured diagnostic and treatment assistance method for COPD based on multi-LLM collaboration, characterized in that, Includes the following steps: S1. Acquire medical theoretical knowledge data and clinical case data, and through preprocessing, construct a task-separated professional dataset for multi-model fine-tuning; S2. Select an open-source base model and fine-tune the instructions using theoretical knowledge datasets and clinical case datasets respectively to obtain a knowledge question-answering model and an auxiliary diagnosis model with separated responsibilities. S3. Deploy the knowledge question answering model and the auxiliary diagnosis model, and set up a general model to perform semantic analysis on user input to output routing control parameters. Based on the control logic of the structured diagnosis and treatment process, use the routing control parameters to call and coordinate the knowledge question answering model and the auxiliary diagnosis model in an orderly manner. S4. Use authoritative medical guidelines as an external knowledge base, and enhance the model output by retrieval and generation; S5. Receive user input through a graphical user interface and, under the constraints of a structured diagnosis and treatment process, call the corresponding model to output diagnostic and treatment auxiliary information.

2. The COPD structured diagnosis and treatment assistance method based on multi-LLM collaboration according to claim 1, characterized in that, The medical theoretical knowledge data includes Chinese and English medical review literature, medical textbooks, and authoritative medical guidelines; The clinical case data mentioned are real clinical case data after removing duplicates and null values.

3. The COPD structured diagnosis and treatment assistance method based on multi-LLM collaboration according to claim 2, characterized in that, The preprocessing includes: For medical theoretical knowledge data, the original theoretical knowledge data is converted into structured Markdown text and split into text blocks according to semantics. For each text block, question-answer pairs are automatically constructed through a large language model. The question-answer pairs are then organized into the first JSON instruction dataset, which serves as the theoretical knowledge dataset. For clinical case data, the desensitized clinical case data is cleaned and screened, and the chief complaint and present medical history information in the case are extracted as input, and the diagnosis and medical order information are extracted as output. The input and output are organized into a second JSON instruction dataset, which serves as the clinical case dataset.

4. The COPD structured diagnosis and treatment assistance method based on multi-LLM collaboration according to claim 3, characterized in that, The open-source base model is Qwen2.5 7B, and the instruction fine-tuning uses Lora.

5. The COPD structured diagnosis and treatment assistance method based on multi-LLM collaboration according to claim 4, characterized in that, The knowledge question-answering model is obtained by fine-tuning the theoretical knowledge dataset with instructions, and is used to answer theoretical questions related to COPD. The auxiliary diagnostic model is obtained by fine-tuning the clinical case dataset and is used to generate preliminary diagnostic and treatment assistance results based on the patient's chief complaint and present medical history.

6. The COPD structured diagnosis and treatment assistance method based on multi-LLM collaboration according to claim 5, characterized in that, The routing control parameters are a set of parameters generated by semantic analysis of user input through a general model, used to characterize the current diagnosis and treatment stage and status information; The parameter set includes at least: stage parameters, symptom parameters, and examination parameters.

7. The COPD structured diagnosis and treatment assistance method based on multi-LLM collaboration according to claim 6, characterized in that, The control logic of the structured diagnosis and treatment process includes: When the routing control parameters indicate that the user has not made a statement, the knowledge question-answering model is invoked. When routing control parameters instruct the user to make a chief complaint, ask the user whether they have undergone any pulmonary function tests that are strongly correlated with a COPD diagnosis: If the user has not had a lung function test, guide the user to answer the COPD-SQ scale and determine whether there is a risk of COPD. If there is a risk of COPD, call the auxiliary diagnostic model. If the user has undergone pulmonary function testing, obtain the measured values ​​of FEV1%FVC and the actual / predicted values ​​of FEV1, which are relevant to the diagnosis of COPD: when and At that time, the reply to the user may be to retain diagnostic and treatment auxiliary information about impaired lung function by retaining the ratio; when and At that time, the reply to the user may be auxiliary information for the diagnosis and treatment of early-stage chronic obstructive pulmonary disease; when At that time, the user was initially identified as a patient with chronic obstructive pulmonary disease (COPD). The user was guided to complete the mMRC and CAT questionnaires to determine the group and stage. Subsequently, the knowledge question-answering model was called to generate a comprehensive response that included the classification, grouping, stage and treatment suggestions.

8. The COPD structured diagnosis and treatment assistance method based on multi-LLM collaboration according to claim 7, characterized in that, The enhanced search generation is invoked at different stages of the structured diagnostic and treatment process, specifically including: The authoritative medical guide text is segmented and converted into vectors using an embedding model before being stored in a vector database. When a response needs to be generated, the user's question or the current context is transformed into a query vector through an embedded model; Similarity matching is performed in the vector database to retrieve the text blocks most relevant to the query vector as reference information; The reference information is concatenated with the original question and then input into the currently invoked model to generate an enhanced response constrained by external knowledge.

9. The COPD structured diagnosis and treatment assistance method based on multi-LLM collaboration according to claim 8, characterized in that, The embedded model is the BGE-M3 model, and the authoritative medical guideline is the GOLD guideline.

10. The COPD structured diagnosis and treatment assistance method based on multi-LLM collaboration according to claim 9, characterized in that, The graphical user interface is used to receive user input in the form of text or uploaded report files, and to visually display the progress of the structured diagnosis and treatment process and the diagnosis and treatment assistance information generated by the model.

Images