Type 1 diabetes mellitus special disease asking medical big model system

Abstract

The invention discloses a type 1 diabetes mellitus special disease asking medical big model system which supports multi-modal input such as voice, characters, pictures and videos, and after multi-modal recognition and classification, intention analysis is carried out by combining inherent information of a user and utilizing double knowledge bases. The intention of the user is judged through an innovative algorithm, and the intention is divided into a medicine asking question, a non-medicine asking question and an unrecognizable intention. And for different types, a defined artificial intelligence large model is adopted to generate professional replies, and the professional replies are output after compliance check. The system also has the functions of context perception retrieval, fuzzy intention reconstruction and the like, realizes accurate and personalized medical question service, effectively solves the defects of the existing type 1 diabetes medical service in the aspects of diagnosis, treatment scheme making, blood sugar monitoring and the like, and improves the medical service quality and efficiency.

Description

Technical Field

[0001] The present invention relates to a medical consultation technology, and in particular to a special disease consultation large model system for type 1 diabetes. Background Art

[0002] As an autoimmune disease, type 1 diabetes seriously affects the quality of life and health of patients. At the current medical level, patients with type 1 diabetes need to rely on insulin treatment for life, and at the same time closely monitor blood glucose changes to prevent the occurrence of serious complications such as cardiovascular diseases, kidney diseases, and retinopathy. According to authoritative medical statistics, the number of global type 1 diabetes patients is on the rise year by year, and the onset age is getting younger, so the demand for precise and personalized medical services is extremely urgent.

[0003] At present, there are many challenges in the current situation of medical services for type 1 diabetes. In terms of diagnosis and screening, the current diagnosis methods mostly rely on blood test indicators, which are cumbersome and may have a certain time lag. Patients are often diagnosed only after obvious symptoms appear, delaying the best intervention time. For example, some primary medical institutions are difficult to accurately identify type 1 diabetes patients in the early stage of the disease due to limitations in detection equipment and technology.

[0004] In the formulation of treatment plans, insulin treatment is the core means, but the precise adjustment of insulin dosage has always been a clinical problem. The insulin sensitivity of each patient is affected by many factors such as age, weight, exercise volume, and diet structure, with great differences. At present, doctors mainly adjust insulin dosage based on experience and limited data provided by patients during regular follow-up visits, making it difficult to achieve real-time and dynamic precise regulation, resulting in large fluctuations in patients' blood glucose levels and increasing the risk of complications. Blood glucose monitoring is an important part of the management of type 1 diabetes. Although traditional fingertip blood glucose meters are widely used, they can only provide single-point blood glucose values and cannot reflect the all-day blood glucose fluctuations. Although continuous glucose monitoring systems (CGMS) can continuously monitor blood glucose, they have high equipment costs, poor wearing comfort, and some patients have insufficient ability to interpret and apply the monitoring data, so they cannot fully play their roles.

[0005] In terms of exercise management and diet therapy, there is a lack of personalized and intelligent guidance programs. Existing exercise and diet recommendations are mostly general content and cannot be dynamically adjusted according to the patient's real-time physical condition and blood glucose data. During the implementation process, patients often have difficulty grasping key factors such as exercise intensity and dietary calorie intake, resulting in poor blood glucose control. Psychological support is also indispensable in the treatment of type 1 diabetes. Psychological stress caused by long-term illness, such as negative emotions like anxiety and depression, will have an adverse impact on the patient's treatment compliance and blood glucose control. However, in the current medical service system, psychological support services are relatively scarce, lacking effective evaluation and intervention mechanisms. For the management of chronic or acute complications of type 1 diabetes, the current medical services mostly provide passive treatment after the occurrence of complications, lacking effective early warning and active prevention measures. Once a patient develops severe complications, the treatment is difficult, costly, and the prognosis is poor. In terms of the convenience and accessibility of medical services, patients have limited access to professional medical advice. In most cases, patients need to go to the hospital to queue for treatment, consuming a lot of time and energy. Although online medical services have developed to some extent, there are problems such as fragmented information, lack of systematicness and professionalism, and cannot meet the patient's needs for one-stop and precise medical consultation services.

[0006] In summary, the current medical level and service status of type 1 diabetes are difficult to meet the growing needs of patients for precise, personalized, and convenient medical care, and an innovative technical solution is urgently needed to improve this dilemma. Summary of the Invention

[0007] The technical problem solved by the present invention is to provide a large model system for specialized medical consultation for type 1 diabetes, which effectively makes up for the deficiencies of existing type 1 diabetes medical services in aspects such as diagnosis, treatment plan formulation, and blood glucose monitoring, and improves the quality and efficiency of medical services.

[0008] The technical solution of the present invention is as follows: A large model system for specialized medical consultation for type 1 diabetes, which includes a multimodal input module, a multimodal recognition and classification module, a user information acquisition module, an intention analysis module, a model direction judgment module, an artificial intelligence model processing module, and a reply output and saving module; The multimodal input module inputs the user input information through an electronic terminal, including various modal input methods such as text input, voice input, picture input, and video input; The multimodal classification module is used to identify and classify the input multimodal information, including: Speech recognition, using an automatic speech recognition algorithm to convert the speech signal into text content; Picture recognition, using convolutional neural network technology to recognize the picture content; Text recognition directly performs preliminary parsing on the input text to determine the basic theme direction of the text content; Video recognition uses video frame extraction and analysis technologies, combines image recognition and action recognition algorithms to understand video content, and converts it into processable information; The user information acquisition module is used to directly obtain the user's inherent information of registration and record information after the user logs in to the system, and is used to combine and analyze the questions currently asked by the user, including Basic information, including at least height, weight and age, and calculating indicators related to diabetes management including body mass index; Diabetes history information, including the time of onset, previous treatment plans, and historical information on complications that have occurred; The intention analysis module combines the user input information and the user's inherent information to perform intention analysis in a dual database, and the dual database includes An unstructured database, which is a database that collects public information on type 1 diabetes on the Internet; A structured database, which integrates information on cases and clinical data on the treatment and maintenance of type 1 diabetes that are not publicly available within a specific organization or system to build a structured database; The intention analysis module uses a machine learning or deep learning model to match and classify among multiple preset intentions according to the user input content and user information; The model direction judgment module judges the direction of the user's intention to seek medical advice according to the intention analysis result of the intention analysis module. The direction of the intention to seek medical advice includes unrecognized intentions, non-medical intentions and medical intentions; The artificial intelligence processing module processes reply information for different medical intentions: For unrecognized intentions, a large model is used to generate prompt information to guide the user to clearly input the question again; For non-medical intentions, the knowledge of the large model in non-medical fields is used for general replies, and at the same time, the user is reminded that this system focuses on seeking medical advice on type 1 diabetes; For medical intentions, combined with the professional information in the dual database, the existing artificial intelligence large model is restricted in terms of format and reply scope to ensure the professionalism and accuracy of the reply; The reply output and saving module is used to save the generated reply result and output it to the user through the terminal. The output form corresponds to the input form, including text replies, voice replies, and picture annotation replies.

[0009] From the above description, it can be seen that this invention indeed has the following advantages: The type 1 diabetes-specific medical consultation large model system of the present invention provides a better, more convenient, more accurate and efficient solution for the current status of medical services related to medical consultation for type 1 diabetes: Improve the efficiency and accuracy of diagnosis and screening. Through multi-modal information collection, it can comprehensively capture the subtle symptoms and physical signs of patients. For example, the combination of local body pictures uploaded by patients and voice descriptions can assist doctors in detecting potential lesion signs earlier, shortening the diagnosis cycle, improving the accuracy of early diagnosis, and avoiding the deterioration of the condition due to delayed diagnosis. At the same time, based on a dual knowledge base and advanced algorithms, the system can deeply analyze various detection data, accurately identify the pre-diabetes and early symptoms of type 1 diabetes, and strive for the best treatment opportunity for patients.

[0010] Achieve precise and personalized adjustment of treatment plans such as insulin therapy. Combining the real-time input information and historical data of patients, the system can accurately analyze the individual insulin sensitivity. For example, based on the pictures of the patient's daily diet intake, exercise videos, and blood glucose monitoring data, the insulin dosage recommendation can be adjusted in real time. Compared with the traditional adjustment based on experience, the accuracy of dosage adjustment is greatly improved, effectively reducing blood glucose fluctuations and the risks of hypoglycemia or hyperglycemia caused by improper insulin dosage, significantly improving the blood glucose control effect, and delaying the progression of complications.

[0011] Optimize blood glucose monitoring and provide dynamic management. Using advanced video recognition and image recognition technologies, the system can automatically identify the data of the patient's home blood glucose monitoring equipment, and generate a comprehensive and accurate blood glucose fluctuation map in combination with the continuous glucose monitoring data. Without the need for patients to manually record, the system can provide personalized blood glucose management suggestions based on the map, such as adjusting diet, exercise, and medication time according to the blood glucose fluctuation pattern, realizing dynamic and precise blood glucose management, and enhancing the patient's self-management ability.

[0012] Provide comprehensive and personalized exercise and diet guidance. According to the patient's physical indicators (such as BMI, etc.), the analysis results of exercise videos, and the information recognized from diet pictures, the system formulates exclusive exercise and diet plans. For example, it designs a scientific weight loss exercise plan for overweight patients, adjusts the nutritional intake suggestions in real time according to the diet structure of each meal, and combines the real-time feedback of blood glucose changes to dynamically optimize the plan, helping patients effectively control blood glucose while improving their overall health.

[0013] Strengthen psychological support and management during special periods. By semantically analyzing the text input by patients, accurately understand the psychological state, and timely provide targeted psychological counseling resources. For example, when identifying the patient's anxiety, it pushes professional psychological popular science articles and psychological counseling channels. For special periods such as pregnancy, before and after surgery, the system provides customized management plans based on the knowledge base to ensure that patients can smoothly pass through the special stage and improve their quality of life and treatment compliance.

[0014] Effectively prevent and manage complications. Based on double-database case analysis and real-time data monitoring, the system can achieve early warning of chronic and acute complications. For example, by analyzing multi-index data such as blood glucose, blood pressure, and renal function, it can predict the risk of diabetic nephropathy several months in advance and provide preventive measures. When complications occur, accurate treatment suggestions are provided based on a large number of clinical cases to reduce the harm of complications and improve the quality of life and lifespan of patients.

[0015] Greatly improve the accessibility and convenience of medical services. Patients can consult a doctor through multi-modal input anytime and anywhere, and the system responds immediately, breaking the limitations of time and space. There is no need to frequently go to the hospital, reducing the patient's running around and waiting time in line. At the same time, the system integrates medical resources to provide one-stop medical consultation services. Patients can obtain full-process suggestions such as diagnosis, treatment, and rehabilitation on the same platform, significantly improving the efficiency of medical services and patient satisfaction. Description of the Drawings

[0016] Figure 1 It is a schematic diagram of the composition of the preferred embodiment of the type 1 diabetes dedicated medical consultation large model system of the present invention. Detailed Embodiments

[0017] For a clearer understanding of the technical features, objectives, and effects of the present invention, the detailed embodiments of the present invention are now described with reference to the accompanying drawings.

[0018] The type 1 diabetes dedicated medical consultation large model system proposed by the present invention is referred to Figure 1 As shown, in its preferred embodiment, the present invention includes a multi-modal input module, a multi-modal recognition and classification module, a user information acquisition module, an intention analysis module, a model direction judgment module, an artificial intelligence model processing module, and a reply output and saving module; The multi-modal input module inputs user information through an electronic terminal, including text input. The user directly inputs text questions related to type 1 diabetes through the built-in input method of the terminal, such as "Do I need to adjust my insulin dosage?"; voice input. The user uses the voice recording function of the terminal to speak out the question or describe the relevant situation, such as "I feel dizzy after exercise recently. Is there a problem with my blood sugar?"; picture input. The user uploads pictures related to type 1 diabetes through the photo-taking or attachment pasting function of the terminal, such as pictures of medicine packages, blood sugar test reports, food pictures, etc. And video input. The user can record and upload videos, such as showing their diet or exercise process to more comprehensively describe their own situation. It constitutes multiple modal input methods; The multimodal recognition and classification module is used to recognize and classify the input multimodal information, including: Speech recognition, which uses the automatic speech recognition (ASR) algorithm to convert speech signals into text content. For example, using end-to-end models based on deep learning (such as WaveNet, DeepSpeech, etc.), through feature extraction of speech signals (such as Mel-frequency cepstral coefficients MFCC) and sequence modeling, accurately recognize the semantic information in the speech. Image recognition, which uses convolutional neural network (CNN) technology, such as models like ResNet, VGG, etc., to recognize the content of images. For drug images, recognize the drug name, ingredients, etc.; for test report images, extract key data such as blood sugar values, insulin levels, etc.; for food images, recognize food types, approximate calories, etc.

[0019] For text recognition, directly perform a preliminary analysis on the input text to judge the basic theme direction of the text content, such as whether it involves symptom description, treatment plan inquiry, etc.

[0020] Video recognition, which uses video frame extraction and analysis technology, combines image recognition and action recognition algorithms to understand the video content; for example, recognize the types of user movement actions, eating behaviors, etc. in the video and convert them into processable information. Among them, image recognition algorithms include convolutional neural network (CNN), histogram of oriented gradients (HOG), and scale-invariant feature transform (SIFT) and other algorithms; action recognition algorithms include algorithms such as given smooth algorithms, 3D convolutional neural network (3D-CNN), and spatio-temporal interest points (STIP), etc., all of which can be used to achieve this.

[0021] The user information acquisition module is used to directly obtain the user's inherent information of registration and record information after the user logs in to the system, and is used to combine and analyze the questions currently asked by the user. It includes basic information, at least including height, weight, and age, which are used to calculate diabetes management-related indicators including body mass index (BMI), where BMI = weight (kg) / height 2 (m); diabetes history information, including historical information such as the time of onset, previous treatment plans, and complications that have occurred. These historical information are crucial for accurately analyzing the user's current questions.

[0022] The intent analysis module combines the user input information and the user's inherent information to perform intent analysis in a dual database. The dual database includes: The unstructured database is a database that collects public information on type 1 diabetes on the Internet; through natural language processing (NLP) technology, preprocessing such as keyword extraction and semantic understanding is performed. For example, use the TF-IDF algorithm to extract key terms in the text for quick retrieval of relevant information.

[0023] A structured database that integrates information on cases and clinical data related to the treatment and maintenance of type 1 diabetes, which are not publicly available within a specific organization or system, and constructs a structured database for efficient querying and analysis.

[0024] The intention classification algorithm of the intention analysis module adopts machine learning or deep learning models (such as support vector machine (SVM), long short-term memory network (LSTM), etc.), and based on the user input content and user information, it performs matching and classification among multiple preset intentions (such as diagnosis and screening related to type 1 diabetes, insulin therapy, blood glucose monitoring, exercise management, prevention, diet therapy, psychological support, management during special periods, management of chronic or acute complications, and basic characteristics of type 1 diabetes).

[0025] Specifically, the intentions to be recognized may include the following: Basic characteristics of T1DM (polydipsia, polyuria, thirst, C-peptide, weight loss, ketosis, onset characteristics, immune abnormalities, islet autoantibodies, insulin).

[0026] Management of acute complications of T1DM (hypoglycemia, diabetic ketoacidosis, asymptomatic, hyperosmolarity, coma, lactic acidosis, acute complications, infection, electrolyte disorders, first aid, palpitation, fruity breath, confusion, blood ketone).

[0027] Management of chronic complications of T1DM (blindness, numbness in feet, chest tightness, blurred vision, microalbuminuria, follow-up visit, early signs, gastroparesis, osteoporosis, complications, stress, skin itching, periodontitis, impotence, numbness, dialysis, glaucoma, cataract, ankle-brachial index, diarrhea, urinary tract infection).

[0028] Other treatment techniques or treatment progress of T1DM (immunotherapy, anti-CD3 monoclonal antibody, artificial pancreas, stem cells, genes, blood glucose watch, ns, MPC, closed-loop, dual hormone, hybrid closed-loop, AAPS, loop, Medtronic 670, Tandem Control-IQ, Omnipod 5, bionic pancreas, cure, new type, blood glucose trend, T cells, vaccine).

[0029] Screening for T1DM (such as high-risk groups, autoantibodies, genetics, family history, fasting blood glucose, polydipsia, polyuria, weight loss, islet function, glucose tolerance, positive antibodies, first-degree relatives, ketone body detection (blood ketone / urine ketone), interpretation, screening, neonatal / pregnancy screening, self-screening, diagnosis, typing).

[0030] Management of special periods of T1DM (school, pregnancy, lactation, elderly, honeymoon period, perioperative period, adolescence, insulin resistance, nephropathy period, infection period, jet lag, emergency surgery, fever period, diarrhea, vomiting, school age, long-distance travel).

[0031] Psychological support for T1DM (such as emotion management, family support, depression, anxiety, psychology, self-harm, irritability).

[0032] Blood glucose monitoring for T1DM (such as CGM, TIR / TBR / TAR, sensor, fasting blood glucose, postprandial blood glucose fluctuation, blood glucose meter, glycated hemoglobin, dynamic brand, wearing, calibration, AGP, blood glucose control target, urine glucose).

[0033] Insulin therapy for T1DM (such as insulin, individual differences in absorption, injection site rotation, psychology, insulin pump, continuous subcutaneous insulin infusion, CSII, intensive therapy, dual wave, square wave, adjustment techniques, allergy, insulin pen, needle-free injection technology, rebound hyperglycemia, lipohypertrophy, consumable replacement, pump failure, correction dose, dawn / dusk / Somogyi phenomenon, adolescence, large dose wizard function, temporary basal rate adjustment).

[0034] Diet therapy for T1DM (such as carbohydrates, GI, GL, meal distribution, sugar-free, protein, fat, dietary fiber, eating out, festivals, alcohol, ketogenic diet, lactation, Mediterranean diet, intermittent fasting, sugar substitutes, lactose intolerance, food exchange portion method, birthday, food, snacks).

[0035] Prevention of T1DM (Teplizumab, primary prevention, peptide immunotherapy, milk, environmental trigger factors, vitamin D, breastfeeding, gut microbiota, vaccines, regular physical examinations, new drugs, low carbon).

[0036] Exercise management for T1DM (such as running, heart rate, resistance, exercise contraindications, fast walking, swimming, delayed hypoglycemia, weight bearing).

[0037] Diagnosis of T1DM (polyuria, polydipsia, polyphagia, weight loss, C-peptide, glycated hemoglobin, diabetic ketoacidosis, differential diagnosis, MODY, LADA, OGTT, classification, diagnostic criteria, impaired fasting glucose, newly diagnosed, common misunderstandings, positive antibodies, honeymoon period, test reports).

[0038] Related to T2DM (such as insulin resistance, lifestyle intervention, oral hypoglycemic agents, obesity, weight loss, hyperlipidemia, hyperuricemia, hypertension).

[0039] Product consultation (such as insulin pump, dynamic monitor, CGM, insulin pen, blood collection needle, blood glucose meter, consumable purchase, tubing, baseplate).

[0040] AndroidAPS (micro bolus, COB, ISF, sensitivity factor, Autosens, DIY, closed loop, calculator, Carbs, Carbs On Board, Oref1, Android artificial pancreas system, profile, DIY risk, Bluetooth, pump communication, mobile phone).

[0041] Chatting or greeting phrases (such as hello, thank you, good morning, okay, got it).

[0042] Furthermore, integrating the unstructured knowledge base publicly available on the Internet and the internal non-public structured knowledge base provides richer and more accurate information support for intent analysis, improving the accuracy of intent judgment. For example, for some rare type 1 diabetes complication problems, the unstructured knowledge base may have the latest research reports, while the actual cases in the structured knowledge base can provide specific treatment experiences. The implementation method can include: Unstructured knowledge base processing: Use web information acquisition technology to collect publicly available Internet materials, and construct a knowledge graph through technologies such as named entity recognition (NER) and relation extraction in NLP. For example, identify entities such as "type 1 diabetes", "complications", "treatment methods" in the materials, and establish the relationships between them. When querying, use graph database technology (such as Neo4j) to quickly retrieve relevant knowledge.

[0043] Structured knowledge base processing: Use a relational database (such as MySQL) or a NoSQL database (such as MongoDB) to store cases and clinical data. Design a reasonable database table structure, such as a patient table (recording basic information), a treatment record table (recording medications, insulin doses, etc.), and a case table (recording the treatment processes and results of typical cases). When querying, quickly obtain relevant case data through SQL statements or database query interfaces.

[0044] The model direction judgment module judges the user's medical consultation intent direction according to the intent analysis result of the intent analysis module. The medical consultation intent direction includes unrecognizable intent, non-medical consultation intent, and medical consultation intent. Specifically, unrecognizable intent: When the user's input content is too vague, lacks key information, or exceeds the preset intent range, it is marked as this direction. Non-medical consultation questions: If the user asks general knowledge unrelated to type 1 diabetes (such as "Why is the sky blue?"), it belongs to this category. Medical consultation questions: The user's questions involve relevant medical content such as the diagnosis, treatment, and management of type 1 diabetes.

[0045] The artificial intelligence processing module processes the reply information for different medical consultation intents: For unrecognizable intent, call the large model to generate prompt information to guide the user to clearly input the question again. For example, "Your question is not very clear. Please describe your situation in detail so that we can better answer your question." For non-medical consultation intentions, use the knowledge of large language models in non-medical fields to give general responses, while reminding users that this system focuses on medical consultation for type 1 diabetes; for example, "Regarding this question, (general response). However, this system mainly provides you with medical consultations related to type 1 diabetes. If you have questions in this regard, you are welcome to ask at any time." For medical consultation intentions, combine the professional information in the dual databases to impose constraints on the format and response scope of existing large language models (such as DeepSeek) to ensure the professionalism and accuracy of the responses; for example, when answering questions related to insulin treatment, limit the large language model to provide specific dosage adjustment suggestions based on the latest clinical guidelines and valid cases in the knowledge base; such as "According to your situation, the insulin dosage can be increased or decreased by 10%-20% on the current basis, but please make sure to adjust it under the guidance of a doctor."

[0046] Response output and saving module, used to save the generated response results; stored in the system database for subsequent analysis and user query of historical records, and output to the user through the terminal. The output form corresponds to the input form, including text responses, voice responses (convert text to voice using text-to-speech TTS technology), and picture annotation responses; if it is a picture of a test report, key data will be marked on the picture and explained.

[0047] Taking a specific example, elaborate on the operation steps and process of the type 1 diabetes specialized disease medical consultation large language model system in the preferred embodiment of the present invention: User input link: Assume that the user inputs a query message through the mobile phone APP: The user first inputs the voice "I took this medicine today, but my blood sugar is still very high. What should I do?", and at the same time attaches a picture of the medicine, showing that the name of the medicine is "metformin".

[0048] Multimodal recognition and classification link: 1. Speech recognition: The mobile phone APP transmits the voice signal to the speech recognition module of the system, and uses the ASR algorithm to convert the voice into text "I took this medicine today, but my blood sugar is still very high. What should I do?".

[0049] 2. Picture recognition: Perform CNN recognition on the medicine picture uploaded by the user to determine that the medicine is "metformin", and quickly retrieve the basic information of this medicine from the knowledge base, such as it is mainly used for type 2 diabetes and is generally not used alone for type 1 diabetes, and is often used in combination with insulin, etc.

[0050] User information acquisition link: The system obtains the registration information of this user: height 170 cm, weight 70 kg, and calculates the BMI as 70 / 1.7 2≈24.22, which is on the high side of the normal range), 30 years old, has had type 1 diabetes for 5 years, and has been treated with insulin all the time.

[0051] Intention analysis stage: Text parsing: Perform NLP analysis on the text transcribed from speech "I took this medicine today, but my blood sugar is still very high. What should I do?", and extract keywords "taking medicine", "high blood sugar", and "what should I do".

[0052] Knowledge base retrieval: Combine the drug "metformin" identified by image recognition, and search for relevant information in the unstructured knowledge base, such as "Precautions for the use of metformin in patients with type 1 diabetes"; search for similar cases in the structured knowledge base, such as "Treatment cases for poor blood sugar control after combined use of insulin and metformin in patients with type 1 diabetes".

[0053] Intention classification: Determine through the intention classification model that the user's intention is "Drug treatment and blood sugar control issues in type 1 diabetes", which belongs to the comprehensive intention of "Insulin treatment (related to combined medication)" and "Blood sugar monitoring (treatment for poor control)" in the preset intentions.

[0054] Model direction judgment stage: It is clear that the user's question belongs to a "medical consultation question" because it involves the use of drugs and blood sugar control in type 1 diabetes, which is typical medical consultation content.

[0055] Artificial intelligence model processing stage: Limit conditions: According to the information in the knowledge base, limit the large model's reply to must include the following content: Inform the user of the general application principle of metformin in type 1 diabetes (not used alone, often combined with insulin), analyze the possible reasons for the user's high blood sugar, such as whether the drug combination is reasonable, whether the insulin dose needs to be adjusted, and whether there are changes in diet and exercise; provide preliminary suggestions, such as suggesting timely detection of blood sugar profiles, recording diet and exercise conditions, and contacting the attending doctor to adjust the treatment plan, etc.

[0056] Generate a reply: The large model generates a reply according to the above limit conditions: "The metformin you take is usually not used alone in the treatment of type 1 diabetes and is mostly combined with insulin. There may be various reasons for your high blood sugar. On the one hand, it is necessary to confirm whether the insulin dose has been reasonably adjusted according to recent physical conditions, such as activity level and diet changes; on the other hand, check whether the diet is strictly controlled and whether the exercise is regular. It is recommended that you promptly detect the all-day blood sugar profile (fasting and each postprandial period), record the diet and exercise conditions in detail, and contact your attending doctor as soon as possible so that the treatment plan can be adjusted based on comprehensive information." Reply output stage: Output reply: The system displays the text reply to the user through the mobile APP and provides a voice reply option at the same time. After the user clicks it, it can be converted into voice playback through TTS technology.

[0057] Save records: Save the user's input information (voice-to-text transcription, picture information), user information, intention analysis results, reply content, etc. to the system database, which is convenient for the user to query historical records later and also provides data support for the subsequent optimization of the system.

[0058] As described above, the type 1 diabetes-specific medical consultation large model system of the present invention integrates and recognizes various modal information such as voice, pictures, text, and videos, breaks the limitations of a single modality, and obtains user information more comprehensively and accurately. It integrates the unstructured knowledge base publicly available on the Internet and the internal non-public structured knowledge base to provide richer and more accurate information support for intention analysis and improve the accuracy of intention judgment. It limits the existing general artificial intelligence large model to make its replies in the field of type 1 diabetes medical consultation more professional, accurate, and in line with medical norms, avoiding misleading or inappropriate answers. It makes full use of information such as the user's height, weight, age, and diabetes history. It can provide accurate, professional, and personalized medical consultation services for type 1 diabetes patients.

[0059] In a preferred embodiment of the type 1 diabetes-specific medical consultation large model system of the present invention as described above, the multi-modal recognition and classification module integrates and recognizes various modal information of voice, pictures, text, and videos, and then obtains user information; among them, For voice and text, a joint coding method is adopted to map the text features of voice recognition and the directly input text features to the same feature space through an embedding layer, and a bidirectional LSTM or Transformer model is used for sequence modeling to capture semantic information.

[0060] For pictures, the image features extracted by using convolutional neural network technology (such as the feature vectors of drug pictures and the key data features of test reports) are concatenated with the text features or fused through an attention mechanism.

[0061] For videos, first obtain the key information of each frame through video frame extraction and image recognition technology, then process the dynamic information of the video in combination with time series analysis (such as LSTM), and finally fuse it with other modal features.

[0062] The integration of intent analysis can be to use the information retrieved from two knowledge bases as features and input them into an intent classification model. For example, for the question "Is blurred vision a complication of diabetes?" entered by the user, retrieve the knowledge graph information of "Type 1 diabetes common complication - retinopathy - blurred vision" from the unstructured knowledge base, and find case information such as "Type 1 diabetes patients with blurred vision cases - diagnosed with retinopathy - treatment measures" from the structured knowledge base. Combine this information to determine that the intent is "Diagnostic inquiry about Type 1 diabetes complications (retinopathy)".

[0063] In the preferred embodiment of the Type 1 diabetes specialized medical consultation large model system of the present invention as described above, the artificial intelligence model processing module is provided with a limited response mechanism to limit the general artificial intelligence large model, making its response in the field of Type 1 diabetes medical consultation more professional, accurate, and in line with medical norms, avoiding misleading or inappropriate answers. For example, when answering questions about insulin dose adjustment, limit the scope of the response content in the field of Type 1 diabetes medical consultation to answer according to the specific information of the patient and the dose adjustment range recommended by clinical guidelines; including Establish a medical matrix rule base, collect clinical guidelines, treatment specifications, etc. of Type 1 diabetes, and transform them into executable rules; for example, "The initial insulin dose for Type 1 diabetes patients is generally 0.5 - 1.0 U / (kg·d)", "When blood sugar control is poor, the adjustment range of each insulin dose should not exceed 20% of the original dose", etc.

[0064] Output constraint of the large model. After the large model generates a response, check it through rule matching and semantic analysis. If the response involves insulin dose adjustment, check whether it meets the dose range in the rule base; if it involves drug use, check whether it conforms to the treatment principles of Type 1 diabetes. If it does not meet the requirements, regenerate the response or prompt the model to correct it.

[0065] Dynamic limitation combined with user information: According to the specific information of the user, such as height, weight, age, and medical history, further refine the limitation conditions. For example, for users with a lighter weight, limit the lower limit of the dose range when answering questions about insulin dose adjustment; for users with a history of hypoglycemia, appropriately increase the lower limit value when answering questions about blood sugar control goals to avoid hypoglycemia.

[0066] In the preferred embodiment of the Type 1 diabetes specialized medical consultation large model system of the present invention as described above, it includes personalized response control based on user information, including: Personalized feature extraction: Digitally encode the user information, convert height and weight into BMI index, use age as a separate feature, and classify and encode the situation of disease duration and complications in the diabetes history; Personalized model training. In the intent analysis and response generation models, personalized features are added as inputs. For example, in the diet therapy advice model, a regression algorithm (such as linear regression) is used to establish a relationship model between BMI and daily calorie requirements: Calorie = a × BMI + b. The parameters a and b are obtained through training with the diet data of a large number of type 1 diabetes patients, and the daily diet calorie advice for the user is calculated based on the user's BMI.

[0067] Dynamic response adjustment. According to the new information input by the user each time, including weight changes and new symptoms, the personalized features are updated in real time, and the response content is recalculated and adjusted.

[0068] It should be noted that for "age as a separate feature", it means that the information of the user's age is processed as an independent and important parameter in the system. Age has an undeniable impact on the diagnosis, treatment, and management of type 1 diabetes. For example, adolescent patients are in the stage of growth and development, and their insulin requirements, diet plans, and blood glucose control goals are quite different from those of adults. Also, elderly patients may have more underlying diseases and different tolerances to drugs, which will all affect the treatment plan for type 1 diabetes. Therefore, by taking age as a separate feature in the present invention, the characteristics of different age groups can be fully considered in subsequent analysis and advice giving.

[0069] In a preferred embodiment of the type 1 diabetes specialized medical consultation model system of the present invention, there is also a medical consultation information security and privacy protection mechanism. Specifically, during the entire medical consultation process, the personal information and medical privacy of the user are strictly protected, meeting the requirements of relevant laws and regulations. For example, sensitive information such as the user's blood glucose test report and medical history is encrypted during transmission and storage and cannot be leaked without the user's authorization.

[0070] Data encryption of the present invention: The symmetric encryption algorithm (such as AES) is used to encrypt the sensitive information of the user for storage, and the asymmetric encryption algorithm (such as RSA) is used for data transmission encryption. For example, before the user's blood glucose test report is stored in the database, it is encrypted with the AES algorithm; when it is transmitted to the doctor's end (if the user authorizes), it is encrypted with the RSA algorithm.

[0071] Access control of the present invention: A strict user permission management system is established, and only modules or personnel authorized by the user (such as the user himself / herself, the attending doctor designated by the user) can access the relevant medical consultation information. The role-based access control (RBAC) model is adopted to assign different permissions to different roles (ordinary users, doctors, system administrators). For example, ordinary users can only view their own medical consultation records, doctors can view relevant information for diagnosis after the user authorizes, and system administrators can only perform system maintenance operations and cannot randomly view the specific medical consultation content of users.

[0072] The present invention adopts anonymization processing: when the system conducts data analysis and model training, user information is anonymized to remove information that can identify the user's identity (such as name, ID number). For example, users are identified with anonymous identifiers such as "User ID-1", "User ID-2", etc., to ensure privacy and security during data use.

[0073] The type 1 diabetes-specific medical consultation large model system of the present invention has a complete structure, clear process, and advanced technological innovation, and can provide accurate, professional, and personalized medical consultation services for type 1 diabetes patients. At the same time, it ensures information security and privacy, and has high practical value.

[0074] For the type 1 diabetes-specific medical consultation large model system of the present invention as described above, traditional dual knowledge bases are usually stored independently and are only associated through keyword matching. The present invention proposes a dynamic weight knowledge fusion network, which realizes deep collaboration between the two types of knowledge bases through cross-modal semantic mapping and adaptive weight allocation. In a preferred embodiment, the dual database is a collaborative operation dual database, including a cross-modal semantic mapping dual database architecture; the architectures of the unstructured database and structured database of the dual database are as follows: The unstructured database is set as knowledge graph K1, which stores publicly available type 1 diabetes knowledge data on the Internet using a dynamic semantic graph. The nodes include disease entities (such as "type 1 diabetes", "insulin resistance"), attributes (such as "pathogenesis", "clinical manifestations"), and relationships (such as "complications - retinopathy", "treatment - insulin injection"). Each node is attached with a semantically vector updated in real time, and the semantically vector is generated by the BERT-Med pre-training model, with a dimension d = 768; The structured database is set as a case database K2, and a spatio-temporal feature case table is set. Each case contains: Patient feature vector X = x 1, x 2,..., xn , such as BMI, disease duration, complication code; Diagnosis and treatment action sequence A = a 1, a 2,..., am , such as insulin dose adjustment, diet plan formulation; Efficacy feedback vector Y = y 1, y 2], such as blood glucose control compliance rate, complication incidence rate; Cases are stored through a time-series association index, supporting fast retrieval by combining similar patient characteristics and treatment stages.

[0075] In the preferred embodiment of the type 1 diabetes specialized medical consultation large model system of the present invention as described above, the dual database is a collaborative operation dual database, including an adaptive weight allocation method: Adopt the dynamic weight fusion algorithm (DWF-Net), design the cross-database data correlation degree calculation formula, and realize the dynamic weight allocation of the two databases: wij = α ⋅Sim KG ( ei , ej )+(1− α )⋅Sim Case ( Xi , Xj ); wij : The association weight between the entity ei in the knowledge graph K1 and the case Xj in the case database K2.

[0076] α ∈[0,1] is the modality balance factor; (automatically adjusted when the user inputs in multiple modalities, with voice / text input emphasizing α = 0.7 and image / video input emphasizing α = 0.3).

[0077] Sim KG ( ei , ej ) is the knowledge graph semantic similarity, calculated based on cosine similarity: Sim KG =( v i v j ) / (‖ v i ‖‖v j ‖); where vi , vj are the semantic vectors of the entities ei , ej .

[0078] Sim Case ( Xi , Xj ) is the case feature similarity, using the dynamic DTW algorithm with time decay: ; where ωt is the time weight of the diagnosis and treatment stage (if it is a recent case, the weight is higher, ω t =0.9 T−t ), d (⋅) is the Euclidean distance.

[0079] The present invention passes wijQuantify the association strength between knowledge graph entities and case data. When the user inputs a specific diagnosis and treatment problem (such as "insulin dose adjustment"), the algorithm preferentially activates the case database K2 with high weights, and combines the mechanistic explanations in the knowledge graph K1 to generate composite response information of "case experience combined with theoretical basis".

[0080] In the preferred embodiment of the type 1 diabetes specialized medical question and answer large model system of the present invention as described above, the steps of the coordinated operation of the dual databases are as follows: Input parsing: Decompose the user's question (such as "how to handle large blood glucose fluctuations after insulin injection") into a knowledge graph query entity set E = {insulin injection, blood glucose fluctuation} and a case feature set X q = {user BMI, disease duration}; Cross-database retrieval: For K1: Query the associated entity path of E through the graph database; (such as "insulin injection → pharmacokinetics → blood glucose fluctuation mechanism") For K2: Through X q Retrieve similar cases and calculate Sim Case And screen the top 5 cases; Weight assignment: Calculate according to the input modality (such as when the user uploads a blood glucose fluctuation curve picture at the same time, α = 0.3) wij to generate a fused knowledge vector Z = wij ⋅( vi + Xj ); Knowledge output: Input Z into the large model, and attach knowledge source labels at the same time (such as "K1 - guideline recommendation", "K2 - case verification") to ensure the traceability of the response.

[0081] In the preferred embodiment of the type 1 diabetes specialized medical question and answer large model system of the present invention as described above, in the intention analysis module, it includes a context-aware dual-modal knowledge retrieval method, including: Retrieval model architecture: For the multi-modal data input by the user, including text T, voice feature S, and picture feature I, generate a context vector through a cross-modal fusion layer: C : C =MultiHeadAttention( T ; S ; I );where T ; S ; I represents multi-modal feature splicing, and the dimension d = 1024; Construct an entity and relationship inverted index for K1, and each entity stores the associated diagnosis and treatment stage; (such as "insulin adjustment" is associated with "initial treatment stage", "complication stage").

[0082] Build a feature and efficacy index for K2 to support quickly locating cases according to patient characteristics combined with efficacy goals; (such as the case set of "BMI > 25 and blood sugar not up to standard").

[0083] Dynamic retrieval score calculation: Composite retrieval score including semantic relevance and case utility: Score( q , k )= β ⋅SemanticSim( q , k )+(1− β )⋅UtilitySim( q , k ); q is the user query vector, i.e., the context vector C; k is the knowledge base entry, which is an entity in K1 or a case in K2; β ∈[0,1] is the goal-oriented factor. When the user asks a clear question β =0.8, and when it is a fuzzy consultation β =0.5; Semantic relevance SemanticSim. For K1 entities, calculate the cosine similarity between C and the entity semantic vector vk ; for K2 cases, calculate the weighted similarity between C and the case feature vector Xk , and the weights are obtained by training the user's historical interaction data; Case utility UtilitySim measures the practical value of a case to the current user. The formula is: UtilitySim = (efficacy feedback k ⋅ time decay factor) / (1 + feature difference degree); where efficacy feedback k is the blood sugar control compliance rate in the case (normalized from 0 to 1). The time decay factor = 0.9 Δt , Δt is the time from the case to the present, in months; the feature difference degree = ‖ Xq − Xk ‖ 2 , the Euclidean distance between the user's current features and the case features. Dynamically balance semantic matching and practical value through β to avoid the problem of "high semantic relevance but low clinical utility" in traditional retrieval, such as retrieving a treatment plan that is theoretically correct but not suitable for the user's BMI.

[0084] The following is an example of the execution steps: Query vector generation: The user inputs the text "How can adolescent patients with type 1 diabetes adjust their exercise programs" and uploads a picture of the physical fitness test report (including height, weight, and maximum oxygen uptake data). Through a multimodal encoder, C is generated, which embeds semantic features such as "adolescent", "exercise program", and "physical fitness test data".

[0085] Parallel retrieval of two knowledge bases: For K1: Retrieve the knowledge path of "type 1 diabetes - exercise therapy - adolescent" to obtain the exercise intensity calculation formula, such as "target heart rate = (220 - age) × 60% - 70%", and precautions, such as "avoid exercising on an empty stomach".

[0086] For K 2: Retrieve similar cases based on the user's physical fitness test data (height 160 cm, weight 50 kg, BMI = 19.5), screen cases of "adolescent + normal BMI + incidence of post-exercise hypoglycemia", and calculate the top 3 cases with the highest UtilitySim.

[0087] Score fusion and ranking: Combined with β = 0.7 (the user clearly asks for a program, emphasizing semantic association), calculate the Score of each retrieval result, and preferentially return knowledge with "high semantic relevance and high case utility" (such as including both the exercise formula recommended by the guidelines and successful adjustment cases of patients with similar body types).

[0088] Knowledge data filtering and output: Conduct medical compliance verification on the retrieval results (such as excluding suggestions beyond the guideline dosage range), and finally output a response that integrates the theoretical basis of the knowledge graph and real case experience.

[0089] The preferred embodiment of the present invention as described above has the ability: Dynamic fusion of two knowledge bases: Through cross-modal semantic mapping (Sim KG ) and case timeliness weight ( ωt ), solve the problem of "data islands" in traditional two knowledge bases, and achieve collaborative output of "theoretical knowledge explanation + real case verification".

[0090] Context-aware retrieval: Introduce the goal-oriented factor (β) and case utility (UtilitySim) to make the retrieval results more in line with the user's individual characteristics (such as BMI, disease duration), and avoid the problem of "general suggestions not being applicable to special cases".

[0091] The present invention breaks through the simple architecture of "storage - retrieval" of traditional knowledge bases, and realizes intelligent fusion and dynamic adaptation of knowledge through algorithm innovation, providing a new knowledge management paradigm for precision medicine artificial intelligence systems.

[0092] In the preferred embodiment of the type 1 diabetes disease-specific medical consultation large model system of the present invention as described above, in the intention analysis module, a fuzzy intention reconstruction method is used to reconstruct fuzzy (such as "I feel uncomfortable") and out-of-preset intentions (such as asking about diabetes-related insurance policies). The method includes: Execute the problem decomposition network: decompose the original problem into symptom keywords and potential demand vectors; for example, "I feel uncomfortable" is decomposed into ("uncomfortable", potential demand vector v = 0.4 for association with medical history + 0.6 for association with life advice). Adversarial training is adopted, including a generator simulating users' fuzzy questions and a discriminator distinguishing real questions from fuzzy questions to improve the decomposition accuracy.

[0093] Cross-library guiding question generation: If the intention cannot be recognized, extract high-frequency question templates from the dynamic semantic graph library and generate guiding questions in combination with the user's historical data: q guide =Transformer( v history ⊕Top5Template( K 1)); v history is the average value of the semantic vectors of the user's previous 3 medical consultations; Top5Template is the 5 most frequently asked question templates of similar patients in the graph; Specifically, v history is the average value of the semantic vectors of the user's previous 3 medical consultations, representing the characteristics of the user's historical medical consultation information; Top 5 Template ( K 1) is the feature representation corresponding to the high-frequency question template extracted from the knowledge graph K 1 (dynamic semantic graph repository) source. Through the "⊕" operation, the features of these two parts are connected in a certain order to form a new feature vector. This new feature vector contains the user's historical medical consultation information and the high-frequency question template information from the knowledge source, and then it is input into the Transformer model for processing to generate guiding questions q guide to help the system better understand the user's intention and guide the user to clearly express their needs.

[0094] For non-medical consultation questions, filter through the non-medical association degree of the diagnosis and treatment decision case pool: If there is no relevant decision record in the diagnosis and treatment decision case pool and the semantic graph association degree < 0.3, then trigger the general knowledge module (such as encyclopedia Q&A) to answer, but with a clear hint of non-medical scope: such as "This question is not in the medical category, and the following is general information" Fuzzy intention confidence calculation: Conf(i )=(SemanticMatch( i , K 1)+CaseMatch( i , K 2)) / 2⋅(1−OutlierScore( i )); OutlierScore( i ) is the outlier score based on the Isolation Forest algorithm. The higher the value, the more likely it is a non-medical consultation question; When Conf( i ) < 0.5, trigger the intention reconstruction process.

[0095] In a preferred embodiment of the present invention, the intention classification model of the present invention makes judgments using the following method: Model selection and training: Use a deep learning model (such as a Long Short-Term Memory network LSTM) or a machine learning model (such as a Support Vector Machine SVM). Train using a large amount of labeled data (type 1 diabetes medical consultation cases labeled with specific intentions), for example: In the LSTM model, the integrated feature sequence (text vector, image feature, user information feature) is input by time step, and the long-term and short-term dependencies are captured through the hidden layer memory unit, and finally the probability distribution of the intention category is output.

[0096] In the SVM model, the feature vector is used as the input, mapped to a high-dimensional space through a kernel function (such as a radial basis kernel function), and the optimal classification hyperplane is found to judge the intention category.

[0097] Intention category judgment: The preset intentions include "Type 1 diabetes insulin treatment", "Blood glucose monitoring and control", "Diet treatment consultation", etc. The model calculates the matching scores of each intention according to the feature and knowledge base matching results. For example, in the above user input, the matching scores of the features "insulin injection" and "high fasting blood glucose" with the intention of "Type 1 diabetes insulin treatment" are 0.8, and the matching score with the intention of "Blood glucose monitoring and control" is 0.7. After comprehensive judgment, the main intention is "Type 1 diabetes insulin treatment (dose adjustment related)", and the secondary intention is "Blood glucose monitoring and control (analysis of abnormal fasting blood glucose)". When making a decision output, the requirements are: Determination of intention priority: When there are multiple matching intentions, determine the main intention and secondary intention according to the preset priority rules (such as treatment-related intentions take precedence over monitoring-related intentions) or the probability values output by the model.

[0098] Output result: Output the finally determined intention category (such as "Insulin therapy for type 1 diabetes - Dose adjustment consultation") to the model direction judgment module, providing a basis for subsequent judgment of whether it belongs to a medical consultation question, a non-medical consultation question, or an intention that cannot be recognized.

[0099] In a preferred embodiment of the present invention, the present invention can construct a dedicated medical rule base according to the data content of the unstructured database and the structured database, and transform the clinical guidelines, diagnosis and treatment specifications, etc. of type 1 diabetes into executable rules.

[0100] For example: Insulin therapy rules: Clearly define rules such as "The initial insulin dose for type 1 diabetes patients is generally ()" and "The dose adjustment amplitude each time should not exceed 20% of the original dose". When the large model generates a reply involving insulin dose, it must follow these rules. If content that violates the rules, such as "It is recommended to double the insulin dose", appears, the system will intercept and correct it.

[0101] Complication management rules: Specify "The standard process for the fluid replacement rate and insulin infusion rate when diabetic ketoacidosis occurs", ensuring that the large model complies with the specifications when replying to the treatment of acute complications. Through this rule constraint, the reply logic of the large model strictly follows medical professional standards, avoiding misleading suggestions.

[0102] In the preferred embodiment of the type 1 diabetes dedicated medical consultation large model system of the present invention as described above, when the user information obtained by the user information acquisition module includes artificial pancreas control exclusive qualification information, an artificial pancreas control database is added to the intention analysis module. The artificial pancreas control database includes historical data related to type 1 diabetes and other artificial pancreas controls recorded in the artificial pancreas control system used by the user.

[0103] Preferably, the structured database proposed by the present invention can include data in the fully closed-loop artificial pancreas control system, that is, it includes relevant data information such as the collected data and control data related to type 1 diabetes of patients wearing an artificial pancreas device for real-time monitoring and timed insulin injection, which can be better used in the present invention as more effective database support for a specified user.

[0104] The present invention is a type 1 diabetes dedicated medical consultation large model system. Through more targeted and precise multiple algorithms, using multiple artificial intelligence large models, it constitutes a large model system that can be used for type 1 diabetes dedicated medical consultation, providing more accurate and personalized blood glucose detection and control suggestions for a large number of type 1 diabetes users, and in daily life, providing the most effective suggestions for various activities in the control of type 1 diabetes. It plays a more comprehensive and secure role in escorting for a large number of type 1 diabetes patients all day long.

[0105] The above is only a schematic specific embodiment of the present invention and is not intended to limit the scope of the present invention. Any equivalent changes and modifications made by those skilled in the art without departing from the concept and principle of the present invention shall fall within the scope of protection of the present invention.

Claims

1. 1 Type 1 diabetes specialist consultation model system, characterized by: It includes multimodal input module, multimodal recognition and classification module, user information acquisition module, intention analysis module, model direction judgment module, artificial intelligence model processing module, and reply output and storage module; The multi-modal input module inputs user input information through the electronic terminal, including multiple modal input methods such as text input, voice input, picture input and video input; The multimodal classification module is used to identify and classify the input multimodal information, including: Speech recognition, using automatic speech recognition algorithms to convert speech signals into text content; Image recognition, using convolutional neural network technology to identify the content of the image; Text recognition: directly analyze the input text and determine the basic theme direction of the text content; Video recognition uses video frame extraction and analysis technology, combined with image recognition and action recognition algorithms, to understand the video content and convert it into processable information; The user information acquisition module is used for the system to directly acquire the user's inherent information of the user's registration and record information after the user logs into the system, and is used for analyzing the question currently asked by the user, including: Basic information, including at least height, weight and age, and calculation of parameters relevant to diabetes management including body mass index; Diabetes history information, including duration of illness, previous treatment plans, and historical information on complications that have occurred; The intent analysis module combines user input information and user inherent information to perform intent analysis in a dual database. The dual database includes: The unstructured database is a database that collects public information related to type 1 diabetes from the Internet; A structured database is constructed by integrating non-public information on cases and clinical data on the treatment and maintenance of type 1 diabetes within a specific organization or system; The intent analysis module uses machine learning or deep learning models to match and classify multiple preset intents based on user input and user information; A model direction judgment module, which judges the user's intention direction of asking for medical advice according to the intention analysis result of the intention analysis module, and the intention direction of asking for medical advice includes unrecognizable intention, non-inquiry intention and inquiry intention; Artificial intelligence processing module, which processes information based on different medical consultation intentions: For unrecognizable intents, the model generates prompts to guide users to re-enter the question clearly; For non-medical consultation intentions, the large model is used to provide general responses in the non-medical field, while reminding users that this system focuses on medical consultations for type 1 diabetes; For medical consultation intentions, the existing artificial intelligence model is constrained in terms of format and scope of responses by combining professional information in the dual databases to ensure the professionalism and accuracy of responses; The reply output and saving module is used to save the generated reply results and output them to the user through the terminal. The output form corresponds to the input form, including text replies, voice replies, and picture annotation replies.

2. The type 1 diabetes medical consultation model system according to claim 1, characterized in that: The multimodal recognition and classification module integrates and recognizes multiple modal information of voice, picture, text and video, and then obtains user information; wherein, For speech and text, a joint encoding method is used to map the text features of speech recognition and the text features of direct input to the same feature space through an embedding layer, and a bidirectional LSTM or Transformer model is used for sequence modeling to capture semantic information; For pictures, the image features extracted by convolutional neural network technology are spliced ​​with text features or fused through attention mechanism; For videos, we first use video frame extraction and image recognition technology to obtain the key information of each frame, then combine it with time series analysis to process the dynamic information of the video, and finally fuse it with other modal features.

3. The type 1 diabetes medical consultation model system according to claim 1, characterized in that: The artificial intelligence model processing module is provided with a limited reply mechanism to limit the general artificial intelligence large model, and limit its reply content scope in the field of type 1 diabetes medical consultation to reply according to the patient's specific information and the dosage adjustment range recommended by the clinical guidelines; including: Establish a medical matrix rule library to collect clinical guidelines and treatment specifications for type 1 diabetes and convert them into executable rules; The output constraints of the big model are checked through rule matching and semantic analysis after the big model generates a response. If the response involves insulin dosage adjustment, check whether it complies with the dosage range in the rule base. If it involves drug use, check whether it complies with the treatment principles of type 1 diabetes. If not, regenerate the response or prompt the model to be modified. Dynamic limitation based on user information: further refine the limitation conditions based on the user's specific information.

4. The type 1 diabetes medical consultation model system according to claim 3, characterized in that: Includes personalized response controls based on user information, including: Personalized feature extraction: digitally encode user information, convert height and weight into BMI index, use age as a separate feature, and classify and encode diabetes history including duration of illness and complications; Personalized model training: adding personalized features as input to the intent analysis and response generation models; Dynamically adjust responses, updating personalized features in real time, recalculating and adjusting responses based on new information the user enters, including weight changes and new symptoms.

5. The type 1 diabetes medical consultation model system according to claim 1, characterized in that: The dual database is a collaborative dual database, including a cross-modal semantic mapping dual database architecture; the architecture of the unstructured database and the structured database of the dual database is: The unstructured database is set as the knowledge graph K1, which uses a dynamic semantic graph to store the type 1 diabetes knowledge data publicly available on the Internet. The nodes include disease entities, attributes, and relationships. Each node is attached with a real-time updated semantic vector, which is generated by the BERT-Med pre-training model with a dimension of d=768. The structured database is set as the case database K2, and the spatiotemporal feature case table is set. Each case contains: Patient feature vector X =[ x 1, x 2,..., xn ], including BMI, duration of illness, or complication codes; Diagnosis and treatment action sequence A =[ a 1, a 2,..., 2. ], including insulin dose adjustment or diet plan formulation; Efficacy feedback vector Y =[ y 1, y 2], including the rate of achievement of target blood sugar control or the incidence of complications; Cases are stored through time-series correlation indexes, supporting rapid retrieval based on similar patient characteristics and diagnosis and treatment stages.

6. The type 1 diabetes medical consultation model system according to claim 5, characterized in that: The dual database is a collaborative dual database, including an adaptive weight allocation method: Adopt the dynamic weight fusion algorithm, design the cross-database data association calculation formula, and realize the dynamic weight allocation of the two databases: wij = α ⋅Sim KG ( ei , ej )+(1− α )⋅Sim Case ( Xi , Xj ); wij : Entities in the knowledge graph K1 ei Cases in case database K2 Xj The association weight of α ∈[0,1] is the modal balance factor; Sim KG ( ei , ej ) is the semantic similarity of the knowledge graph, calculated based on cosine similarity: Sim KG =( v i v j ) / (‖ v i ‖‖v j ‖);in vi , vj For Entity ei , ej The semantic vector of Sim Case ( Xi , Xj ) is the case feature similarity, using the dynamic DTW algorithm with time decay: ,in ωt is the time weight of the diagnosis and treatment stage, d (⋅) is the Euclidean distance; pass wij Quantify the strength of the association between knowledge graph entities and case data. When the user input involves specific diagnosis and treatment issues, the algorithm preferentially activates the high-weight case database K2, and combines it with the mechanistic explanation in the knowledge graph K1 to generate composite response information that combines case experience with theoretical basis.

7. The type 1 diabetes medical consultation model system according to claim 6, characterized in that: The steps for the dual database collaborative operation are: Input parsing: Decompose the user question into the knowledge graph query entity set E = {insulin injection, blood sugar fluctuation} and the case feature set X q ={user BMI, duration of illness}; Cross-database retrieval: For K1: query the associated entity path of E through the graph database; for K2: query the associated entity path of E through X q Retrieve similar cases and calculate Sim Case And filter the top 5 cases; Weight distribution: calculated based on input modality wij , generate fusion knowledge vector Z = wij ⋅( vi + Xj ); Knowledge output: Input Z into the big model and attach knowledge source labels to ensure that the answer is traceable.

8. The type 1 diabetes medical consultation model system according to claim 7, characterized in that: The intention analysis module includes a context-aware bimodal knowledge retrieval method, including: Retrieval model architecture: For the multimodal data input by the user, including text T, voice features S, and image features I, a context vector is generated through the cross-modal fusion layer: C : C =MultiHeadAttention([ T ; S ; I ]);in[ T ; S ; I ] represents multimodal feature concatenation, dimension d=1024; Construct an entity and relationship inverted index for K1, with each entity storing the associated diagnosis and treatment stage; Construct feature and efficacy indexes for K2 to support rapid case location based on patient characteristics combined with efficacy targets; Dynamic retrieval score calculation: A composite search score that includes semantic relevance and case usefulness: Score( q , k )= β ⋅SemanticSim( q , k )+(1− β )⋅UtilitySim( q , k ); q is the user query vector, i.e., the context vector C; k is the knowledge base entry, which is an entity in K1 or a case in K2; β ∈[0,1] is the goal-oriented factor. β =0.8, fuzzy consultation β =0.5; SemanticSim: for K1 entity, calculate C and entity semantic vector vk cosine similarity; for the K2 case, calculate C and the case feature vector Xk The weighted similarity of , where the weight is trained by the user's historical interaction data; The case utility UtilitySim measures the practical value of the case to the current user. The formula is: UtilitySim = (Effectiveness Feedback k ⋅Time decay factor) / (1+characteristic difference); where efficacy feedback k , is the blood sugar control target rate in the case; time decay factor = 0.9 Δt , Δt is the time since the case, in months; feature difference = ‖ Xq − Xk ‖2, the Euclidean distance between the user’s current features and the case features.

9. The type 1 diabetes medical consultation model system according to claim 8, characterized in that: The intention analysis module reconstructs the fuzzy and beyond-preset intentions using a fuzzy intention reconstruction method, the method comprising: Execute question decomposition network: decompose the original question into symptom keywords and potential demand vectors; use adversarial training, including the generator simulating user fuzzy questions and the discriminator distinguishing between real questions and fuzzy questions, to improve decomposition accuracy; Cross-database guided question generation: If the intent cannot be identified, high-frequency question templates are extracted from the dynamic semantic graph library and guided questions are generated based on the user's historical data: q guide =Transformer( v history ⊕Top5Template( K 1)); v history It is the average semantic vector of the user's past three medical consultations; Top5Template is the five most frequently asked question templates by similar patients in the graph; For non-medical questions, the non-medical relevance of the diagnosis and treatment decision case pool is used for filtering: if there is no relevant decision record in the diagnosis and treatment decision case pool, and the semantic graph relevance is <0.3, the general knowledge module is triggered to answer, but with additional reminders that clearly define the non-medical scope; Fuzzy intent confidence calculation: Conf( i )=(SemanticMatch( i , K 1)+CaseMatch( i , K 2)) / 2⋅(1−OutlierScore( i )); OutlierScore( i ) is the outlier score based on the isolation forest algorithm. The higher the value, the more likely it is a non-medical question; When Conf( i )<0.5, the intention reconstruction process is triggered.

10. The type 1 diabetes medical consultation model system according to claim 1, characterized in that: When the user information acquired by the user information acquisition module includes exclusive qualification information for artificial pancreas control, an artificial pancreas control database is added to the intention analysis module, and the artificial pancreas control database includes historical data on type 1 diabetes and other artificial pancreas controls recorded in the artificial pancreas control system used by the user.

Images