Standardized Parkinson's disease medical inquiry system based on large language model

Abstract

The invention discloses a standardized Parkinson's disease medical inquiry system based on a large language model. The system comprises an inquiry knowledge base module and a vector retrieval knowledge base constructed based on Parkinson's disease inquiry clinical guidelines and scales customized by experts in the industry. The chief complaint inquiry module is used for collecting unbiased symptoms of three main symptoms of Parkinson's disease, namely motion retardation, muscular stiffness and tremor, according to natural narration of the patient; the scale type inquiry module is used for intelligent question asking, intelligent understanding replying and symptom detail asking based on the RAG technology; the digital doctor interaction module realizes real-time interaction with the patient; and an information induction and summarization module. Through RAG, a large language model technology and a digital human technology, real-time interaction between a digital doctor image and a patient is highly simulated, intelligent questioning and accurate understanding of the description of the patient can be realized in combination with the current context, and extremely high inquiry accuracy is achieved. The whole interrogation process is fully automatically realized, and the interrogation efficiency is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of medical artificial intelligence, and particularly to a standardized Parkinson's disease medical interview system based on a large language model, a large language model system based on medical authoritative guidelines and scales, which realizes automated interviews for Parkinson's disease. Background Art

[0002] Parkinson's Disease (PD) is the second most common neurodegenerative disease after Alzheimer's disease. The clinical symptoms of the disease are complex, and its typical symptoms are Parkinsonism, including three main symptoms: bradykinesia, muscle rigidity, and tremor. In addition to motor symptoms, patients often have non-motor symptoms such as cognitive, sleep, emotional, and autonomic dysfunction. As the disease progresses, patients gradually lose their ability to work and their ability to move independently. In the late stage of the disease, patients may become completely rigid, have difficulty moving, and even be bedridden, and ultimately often die unfortunately due to complications such as pneumonia and pressure sores.

[0003] In the diagnosis and treatment of Parkinson's disease, medical history taking is a crucial initial step. Parkinsonism, as a clinical symptom, is not only seen in Parkinson's disease but also in various neurological disorders, collectively known as Parkinsonian Syndrome. Given the significant differences in etiology, treatment response, and prognosis between Parkinsonian Syndrome and Parkinson's disease, clinical differential diagnosis is particularly important. Doctors need to obtain key information such as the patient's chief complaint, medical history, symptom characteristics, and lifestyle habits through detailed medical history taking, and pay particular attention to non-motor symptoms to provide a reliable basis for subsequent diagnosis and treatment plan formulation. This process often requires a large amount of time and effort. Traditional medical history taking mainly relies on the doctor's experience and the patient's subjective description. Clinically, for standardized diagnosis and treatment and evaluation, various authoritative guidelines and clinical scales are widely used, such as the Movement Disorder Society (MDS) diagnostic criteria for Parkinson's disease, the Unified Parkinson's Disease Rating Scale (UPDRS), the Non-Motor Symptoms scale (NMSS), and the 39-item Parkinson's Disease Questionnaire (PDQ-39). These guidelines and scales are summarized from a large number of clinical studies and industry experts, and have been verified for reliability and validity, with high reliability and practicality. In clinical practice, a standardized pre-medical history taking mode based on guidelines and scales has gradually taken shape. Doctors can make a preliminary judgment quickly based on the results of pre-medical history taking, and conduct targeted further supplementary medical history taking, physical examination, and auxiliary examinations to improve the accuracy of diagnosis. However, the pre-medical history taking method still requires a large amount of time and labor costs. Due to the professionalism of the questions recommended by each guideline and the information collected by the scale, patients cannot fill in the information by themselves. It is necessary to arrange junior doctors to have repeated conversations with patients for each question and confirm their understanding of the corresponding questions, and summarize and record the questions, so that pre-medical history taking can only be carried out in hospitals with sufficient medical resources.

[0004] Combining large language models with a standardized pre-consultation model offers new possibilities for breaking through the current medical consultation dilemma. The artificial intelligence of large language models has demonstrated excellent capabilities in natural language conversations in various fields, and this ability also highly aligns with the needs of medical consultations. However, applying large language models to medical consultations still faces several key challenges: First, when applied in medical consultations, the risk of generating incorrect information cannot be ignored. This phenomenon is usually referred to as "hallucination," which may lead to serious medical misguidance. Second, the complexity of medical consultations requires orderly and comprehensive collection of multi-faceted information. Existing models still lack in reasonably decomposing the consultation process into multiple sub-tasks and ensuring the completeness of information collection. Finally, the language generated by the model may lack medical terminology and standardized expressions, which may affect patients' accurate understanding of questions and thus affect the accuracy of diagnosis. The above defects all make the current consultation system based on large language models lack medical norms and are difficult to be practically applied. Summary of the Invention

[0005] Aiming at the current problem of automated Parkinson's disease consultation, the present invention provides a standardized Parkinson's disease medical consultation system based on a large language model, which is used to automatically conduct pre-consultations for Parkinson's disease patients to assist in improving the accuracy and efficiency of diagnosis, reducing the workload of doctors, and saving medical resources.

[0006] The object of the present invention is achieved through the following technical solutions: A standardized Parkinson's disease medical consultation system based on a large language model, comprising:

[0007] A consultation knowledge base module, which is used to store a vector retrieval knowledge base constructed based on the clinical authoritative guidelines and scales customized by industry experts;

[0008] A chief complaint consultation module, which is used to determine whether the patient has the three main symptoms of bradykinesia, muscle rigidity, and tremor;

[0009] A scale-based consultation module, which includes an intelligent question-asking sub-module, a reply understanding sub-module, and a detailed follow-up sub-module;

[0010] The intelligent question-asking sub-module is used to connect the historical chat records of patient interaction into a single text to form a query, use the vector representation of the query to retrieve relevant questions in the consultation knowledge base, generate prompt words based on the large language model, screen out questions that conform to the current conversation environment, and rewrite the questions based on the prompt word rewriting template;

[0011] The reply understanding sub-module is used to use the large language model to determine whether the patient understands the question; if the patient answers irrelevantly, it will perform intelligent interruption and adjust the way of asking questions; judge whether the patient has the symptoms described in the question through the patient's reply;

[0012] The detailed questioning sub-module is used to determine whether the intelligent questioning sub-module has asked about the details of the onset. If not, it will ask about the details of the onset.

[0013] The digital doctor interaction module is used to construct a virtual doctor image and realize real-time human-computer interaction with patients.

[0014] The information induction and summary module is used to organize the results of the medical interview.

[0015] Furthermore, the medical interview knowledge base module includes:

[0016] (1) Organize the medical interview clinical guidelines and scales into questions line by line and store them as documents.

[0017] (2) Use an open-source vector model to calculate the vector representation of each question to obtain a binary tuple of question to vector.

[0018] (3) Import the binary tuple into a vector retrieval engine to construct a medical interview knowledge base.

[0019] Furthermore, the chief complaint interview module includes:

[0020] (1) Use a large language model to understand the chief complaint.

[0021] (2) Adopt the few-shot technique and list the specific manifestations of the three main symptoms of bradykinesia, muscle rigidity, and tremor in the prompt words.

[0022] (3) Judge whether the patient has the three main symptoms.

[0023] Furthermore, the chief complaint interview module includes:

[0024] (1) Collect Parkinson's public corpus and hospital interview data, and manually annotate the three main symptom datasets.

[0025] (2) Fine-tune and train the 7B large language model using Lora and P-Tuning techniques.

[0026] (3) Judge whether the patient has the three main symptoms.

[0027] Furthermore, the intelligent questioning sub-module of the scale-based medical interview module includes:

[0028] (1) Select the most recent N chat records of the patient to construct a query.

[0029] (2) Use an open-source vector model to calculate the vector representation of the query.

[0030] (3) Retrieve the top-k semantically relevant results from the medical interview knowledge base and sort them according to the vector similarity score.

[0031] (4) Filter out the questions that have already been asked, use the large language model to generate prompt words, and screen the most context - appropriate questions from the top - k results;

[0032] (5) Use the large language model to rewrite the questions, use easy - to - understand expressions, avoid using professional medical terms, and illustrate with examples in combination with the actual situation.

[0033] Furthermore, the response understanding sub - module of the scale - type interrogation module is used for:

[0034] (1) Judge whether the patient correctly understands the question;

[0035] (2) Judge whether the patient has the symptoms described in the question.

[0036] Furthermore, the detailed questioning sub - module of the scale - type interrogation module is used for:

[0037] (1) Ask about the onset time of the symptoms;

[0038] (2) Ask about the laterality of the onset of the symptoms;

[0039] (3) Ask about details such as whether the symptoms are getting worse.

[0040] Furthermore, the intelligent interruption of the scale - type interrogation module is used to splice the context, as part of the prompt words, revise the way of asking questions, and ask questions again.

[0041] Furthermore, the digital doctor interaction module includes:

[0042] (1) A voice acquisition module, which is used to acquire the patient's voice and convert it into text;

[0043] (2) A question generation module, which is used to process the patient's reply and generate new questions;

[0044] (3) An information output module, which is used to synthesize the questions into voice and generate the corresponding digital doctor image.

[0045] Furthermore, the information induction and summary module is used for:

[0046] (1) Organize the interrogation results in the format of <question, whether there are symptoms, time, left - right side, frequency>;

[0047] (2) Generate an interrogation result table for doctors' reference.

[0048] The beneficial effects of the present invention are as follows:

[0049] (1) The present invention proposes a medical consultation method based on large language models and knowledge bases. Compared with the consultation systems implemented by traditional intelligent question-and-answer technologies, the present invention uses more advanced RAG technology and large language model technology, combines authoritative guidelines and scales to standardize the consultation process, and retains the flexibility of large language models in language understanding, resulting in a better patient question-and-answer experience and higher consultation accuracy.

[0050] (2) The present invention proposes a method for intelligent question asking based on RAG technology, which can screen out the questions that best match the current context from the consultation knowledge base. At the same time, a large language model is used to rewrite the questions. After rewriting, actual examples are combined to avoid using professional medical terms, making them easy to understand. Reasonable questions and correct expressions enable patients to accurately understand the questions asked, thereby improving the accuracy of the consultation and ensuring the coherence and smoothness of the consultation.

[0051] (3) The present invention proposes an intelligent interruption method based on large language model technology. During the consultation process, reasonable prompt words are designed, and large language model technology is used to determine whether the questions asked are accurately understood by the patient. In the case where the patient's answer is off-topic, the way of asking questions is further rewritten in combination with the context to ensure that the patient accurately understands the questions, thereby improving the accuracy of the consultation and excluding non-standard information.

[0052] (4) The present invention proposes an interaction method integrating digital human technology, in which the consultation system interacts with patients in the image of a real digital doctor. The image of the digital doctor highly replicates the image of a real doctor and is indistinguishable from a real doctor in terms of interactive experience, enhancing the consultation experience.

[0053] (5) The present invention significantly improves the efficiency of the consultation: all links of the consultation are automated in this system, and the consultation process does not require the participation of doctors, which can greatly reduce the burden on doctors and save medical resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a structural block diagram of a consultation system based on a large language model of RAG according to an embodiment of the present invention;

[0055] Figure 2 is an implementation flowchart of the chief complaint consultation module in the present invention;

[0056] Figure 3 is the implementation process of the scale-based consultation module in the present invention;

[0057] Figure 4 is an implementation flowchart of intelligent question asking based on a RAG consultation knowledge base;

[0058] Figure 5 is a digital doctor interaction interface integrating digital human technology;

[0059] Figure 6 This is an example of a summary table for the results of medical history inquiries. Detailed implementation

[0060] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0061] As Figure 1 shown, it is a structural block diagram of a medical history inquiry system based on the RAG (Retrieval-Augmented Generation) large language model according to an embodiment of the present invention. By embedding clinical guidelines and scales into the medical history inquiry process, the system can conduct medical history inquiries according to established steps and logic, ensuring the comprehensiveness and standardization of information collection, avoiding omission of key symptoms or indicators, constructing a standardized and professional automatic medical history inquiry system, and completing the preliminary medical history inquiry work for Parkinson's disease. The system can reduce the workload of doctors, save medical resources, enable doctors to have more time and energy to focus on the individual differences and special needs of patients, and can flexibly adjust the medical history inquiry content and process according to the needs of different medical institutions, and is applicable to hospitals at all levels and community health service centers. The system specifically includes the following modules:

[0062] 1. Medical history inquiry knowledge base module. First, organize the authoritative guidelines and scales customized by industry experts in a line-by-line format (each question occupies one line) and store them as doc documents. The guidelines and scales include the MDS Parkinson's disease diagnostic criteria, MDS-UPDRS, NMSS, and PDQ39 to construct a medical history inquiry knowledge base. Subsequently, parse the stored documents and use an open-source vector model to calculate the vector representations (embeddings) of each line, thereby obtaining the binary group <question, embedding> of the question and its corresponding vector. The open-source vector model uses the BCEmbedding model, which has very strong bilingual and cross-lingual capabilities and is a commonly used vector representation model in RAG. Finally, input these binary groups into a vector retrieval engine to realize the construction of the medical history inquiry knowledge base. The vector retrieval engine uses the open-source FAISS framework. FAISS is an efficient vector similarity search library, known for its high performance, rich indexing methods, and ease of use. It supports a variety of indexing techniques, is suitable for large-scale data sets, and can be easily integrated into various applications.

[0063] 2. Chief complaint inquiry module. Tongyi Qianwen 2.5 (110B) is used to understand the chief complaint and evaluate whether the patient presents the typical symptoms of Parkinson's disease, namely the three main symptoms of bradykinesia, muscle rigidity, and tremor. Through open-ended questions such as "What's wrong with you?", an unbiased description of the disease symptoms from the patient is obtained, and symptom descriptions are acquired from the patient's natural narrative. When constructing the prompts for the large model, utilize its few-shot learning ability, and detail the specific manifestation forms of the above three symptoms in the patient in the prompts to enhance the recognition accuracy of the model. The few-shot ability of the large model refers to the ability of the model to effectively predict or generate for new tasks or unseen data after receiving a very small number (usually several) of labeled samples. In addition to this, a discriminant model for the three main symptoms can also be trained to determine the three main symptoms. The first step is to construct a training dataset. The training data has two sources: publicly available corpora in the field of Parkinson's and hospital inquiry record data. Then, the collected raw data is manually labeled: if the three main symptoms exist, the label is 1; if the three main symptoms do not exist, the label is 0. The second step is to train the main symptom discriminant model. The base model uses the pre-trained Tongyi Qianwen 7B model and is fine-tuned using methods such as Lora and P-Tuning. The determination accuracy of the fine-tuned model reaches 99%. If the patient does not show any of the symptoms of bradykinesia, muscle rigidity, and tremor, the inquiry process is terminated. If it is diagnosed that the patient has the symptoms of bradykinesia, muscle rigidity, or tremor, proceed with the inquiry according to Figure 3 the inquiry until all questions in the inquiry knowledge base are answered.

[0064] 3. Scale-based inquiry module, including an intelligent question-asking sub-module, a response understanding sub-module, and a detail follow-up sub-module; the main process of the inquiry is: intelligent question-asking -> response understanding -> follow-up of details, and this process is polled until all questions in the inquiry knowledge base are answered.

[0065] The intelligent question-asking sub-module asks intelligent questions based on the RAG inquiry knowledge base, Figure 4This is its implementation process. Based on the patient's answer, the most suitable question in the medical interview knowledge base is selected for questioning to ensure smooth and coherent communication with the patient. For example, if the patient replies "I haven't been feeling very well recently", questions such as "How's your sleep at night? Are there any difficulties, such as trouble falling asleep or tossing and turning for more than half an hour without being able to fall asleep" are selected from the medical interview knowledge base. If the patient mentions symptoms related to mental state, mental health questions are prioritized. The specific steps for the intelligent questioning function implemented based on the RAG technology in this invention are as follows: First, construct a query. Select the most recent N chat records of interactions with the patient, where N is set to 6, and connect them into a single text using the line break character (\n) to form the query. Second, perform vectorization processing on the query. Use the BCEmbedding model to calculate the embedding representation of the query. Then, enter the first-stage retrieval. Utilize the embedding representation of the query to retrieve the top-k questions in the medical interview knowledge base that are semantically most relevant. This step is directly implemented using the vector recall ability of the FAISS index, and the retrieved top-k results can be sorted according to the vector similarity scores. Subsequently, in the second-stage retrieval, filter out the questions that have already been asked, and then generate prompts using a large language model (Tongyi Qianwen 2.5 110B) to further select the question that best fits the current conversation context from the remaining results. Finally, use the large language model again to rewrite the selected question, and the rewritten result is the rewrite-question. When rewriting, the question is rephrased in an easy-to-understand language, avoiding the use of professional medical terms, and illustrated with examples based on the actual situation so that the patient can understand accurately. An example of the prompt rewrite template is as follows:

[0066] rewrite_template = 'Assume you are a doctor asking a patient about their condition. Please rewrite the question according to the context: {question}\nRequirements: Without changing the original meaning of the question, it should be concise and clear, avoid using professional medical terms, list practical examples, use an easy-to-understand expression, and imitate the way of a doctor's inquiry, showing sufficient patience to the user.\nContext: {context}'

[0067] Here, question is the question selected from the medical interview knowledge base, and context is the above-mentioned query, that is, the most recent N chat records of interactions with the patient.

[0068] The reply understanding sub-module is used for the patient to reply to the rewrite-question generated by the intelligent questioning sub-module. Through the large language model (Tongyi Qianwen 2.5 110B), it is determined whether the patient understands and answers the question rewrite-question.

[0069] If the patient fails to understand the question and gives an irrelevant answer, the intelligent interruption function is activated, such as Figure 3 shown. In the actual medical interview process, it often happens that the patient fails to correctly understand the question or the expression is unclear. Therefore, after receiving the patient's reply, it is necessary to use a large model to evaluate whether the patient accurately understands the question. If it is found that the patient's answer deviates from the topic, the chat context is spliced and used as part of the prompt to input Tongyi Qianwen 2.5 110B. Combining with the patient's reply, using the understanding ability of the large language model, revise the way of asking questions and ask the patient again to guide the patient to understand the questions asked. Jump to the reply understanding sub-module. If the patient still cannot understand the question after multiple attempts, skip this question and then jump to the intelligent question sub-module.

[0070] The large language model (Tongyi Qianwen 2.5 110B) judges whether the patient has the symptoms described in the question through the patient's reply, such as Figure 3 . If not, use the intelligent question function to ask the patient other questions and jump to the intelligent question sub-module; if so, ask for the details of the onset through the detailed questioning sub-module.

[0071] In the said detailed questioning sub-module, if the patient has already mentioned the details of the onset in the reply, including the onset time, the left or right side of the onset symptoms, the severity of the onset, whether the onset symptoms have worsened, etc., then jump to the intelligent question sub-module to ask other questions; otherwise, ask the above details of the onset until all questions in the medical interview knowledge base have been asked.

[0072] 4. The digital doctor interaction module, as Figure 5 shown, is the digital doctor interaction interface of the medical interview system. The above process realizes the function of interacting with the patient in text form. After integrating digital human technology, it can realize real-time interaction with the patient in the form of a digital image, which can not only ask questions to the patient but also receive the patient's voice reply. The digital doctor interaction module consists of three main components: a voice collection module, a question generation module, and an information output module. Voice collection module: Use hardware devices (which can be a microphone on the computer or other external professional microphones) to capture the patient's voice response and convert it into text form. The open-source funasr model can be used to convert voice to text. Question generation module: Process the patient's text reply and generate new questions (in text form) based on the large language model of RAG technology, that is, the rewrite-question generated by the above scale-based medical interview module. Information output module: Convert the generated rewrite-question into voice and display it to the patient through the image of the digital doctor. The dialogue experience between the patient and the digital doctor is close to real human communication, significantly improving the patient's medical interview experience.

[0073] 5. Information induction and summary module. When all the questions in the consultation knowledge base have been completed, the consultation can be ended. Organize them into a table in the format of <question, whether there are symptoms, time, left or right side, frequency> (as shown in Figure 6 ), and upload it for the doctor's reference.

[0074] The present invention is not limited to the above best implementation modes. Anyone can derive other various forms of systems under the inspiration of this patent. All equivalent changes and modifications made in accordance with the scope of the patent application of the present invention shall fall within the scope covered by this patent.

Claims

1. A standardized Parkinson's disease medical consultation system based on a large language model, characterized in that: include: The medical consultation knowledge base module is used to store the vector retrieval knowledge base built based on the authoritative clinical consultation guidelines and scales customized by industry experts; The main complaint questioning module is used to determine whether the patient has the three main symptoms of bradykinesia, muscle rigidity and tremor; The scale-based questioning module includes the intelligent questioning submodule, the response understanding submodule, and the detail inquiry submodule; The intelligent question submodule is used to connect the historical chat records of patient interactions into a single text to form a query, use the vector representation of the query to retrieve relevant questions in the medical consultation knowledge base, generate prompt words based on the large language model, filter out questions that meet the current dialogue environment, and rewrite the questions based on the prompt word rewriting template; The response understanding submodule is used to determine whether the patient understands the question using a large language model; If the patient's answer is not relevant to the question, intelligent interruption will be carried out to adjust the questioning method; the patient's response will be used to determine whether the patient has the symptoms described in the question; The detail inquiry submodule is used to determine whether the intelligent questioning submodule has already inquired about the details of the onset of the disease, and if not, to inquire about the details of the onset of the disease; Digital doctor interaction module, used to build a virtual doctor image and realize real-time human-computer interaction with patients; The information summary module is used to organize the consultation results.

2. The standardized Parkinson's disease medical consultation system based on a large language model according to claim 1 is characterized in that: The medical consultation knowledge base module includes: (1) Organize clinical guidelines and scales into questions by row and store them as documents; (2) Use the open source vector model to calculate the vector representation of each question and obtain a binary pair from question to vector; (3) Import the bigrams into the vector search engine to build a medical consultation knowledge base.

3. The standardized Parkinson's disease medical consultation system based on a large language model according to claim 1 is characterized in that: The chief complaint consultation module includes: (1) Use a large language model to understand the main complaint; (2) Using the few-shot technique, the specific manifestations of the three main symptoms of bradykinesia, muscle rigidity, and tremor are listed in the prompt words; (3) Determine whether the patient has the three main symptoms.

4. The standardized Parkinson's disease medical consultation system based on a large language model according to claim 1, characterized in that: The chief complaint consultation module includes: (1) Collect Parkinson's disease public corpus and hospital consultation data, and manually annotate the three main disease data sets; (2) Use Lora and P-Tuning technology to fine-tune and train the 7B large language model; (3) Determine whether the patient has the three main symptoms.

5. The standardized Parkinson's disease medical consultation system based on a large language model according to claim 1 is characterized in that: The intelligent questioning submodule of the scale-based questioning module includes: (1) Select the N most recent chat records with the patient to construct a query; (2) Use open source vector models to calculate the vector representation of the query; (3) Retrieve the top-k semantically relevant results from the medical inquiry knowledge base and sort them according to their vector similarity scores; (4) Filter out questions that have already been asked, use a large language model to generate prompt words, and select the questions that best match the current context from the top-k results; (5) Use a large language model to rephrase the question, use easy-to-understand expressions, avoid using professional medical terms, and provide examples based on actual situations.

6. The standardized Parkinson's disease medical consultation system based on a large language model according to claim 1, characterized in that: The response understanding submodule of the scale-based questioning module is used to: (1) Determine whether the patient understands the question correctly; (2) Determine whether the patient has the symptoms described in the question.

7. The standardized Parkinson's disease medical consultation system based on a large language model according to claim 1, characterized in that: The detail inquiry submodule of the scale-based inquiry module is used to: (1) Asking about the onset of symptoms; (2) Ask about the laterality of symptom onset; (3) Ask for details such as whether the symptoms have worsened.

8. The standardized Parkinson's disease medical consultation system based on a large language model according to claim 1, characterized in that: The intelligent interruption of the scale-based questioning module is used to splice the context, as part of the prompt words, revise the way of asking questions, and ask questions again.

9. The standardized Parkinson's disease medical consultation system based on a large language model according to claim 1, characterized in that: The digital doctor interaction module includes: (1) Voice collection module, used to collect patient voice and convert it into text; (2) Question generation module, which is used to process patient responses and generate new questions; (3) An information output module, which is used to synthesize the question into speech and generate a corresponding digital doctor image.

10. The standardized Parkinson's disease medical consultation system based on a large language model according to claim 1, characterized in that: The information summarization module is used for: (1) Organize the results of the medical consultation in the format of <question, whether there are symptoms, time, left or right side, frequency>; (2) Generate a consultation result table for the doctor’s reference.

Images