Diabetes personalized diet recommendation system based on multi-model fusion

Abstract

The invention, which relates to the field of diabetes nursing, discloses a multi-model fusion personalized diet recommendation system for diabetes mellitus, comprising a recipe recommendation decision module based on a system architecture, a blood sugar prediction module, and an intelligent question and answer interaction module. The nutritional requirements and the blood sugar control target of the patient are met at the same time; the blood glucose prediction module predicts the postprandial blood glucose level according to the current blood glucose level of the user and the recommended recipe; the intelligent question and answer interaction module provides convenient natural language interaction, collects user feedback and dynamically adjusts recommended content. Compared with the prior art, the method has the advantages that scientific diet recommendation is provided, the recommendation result is optimized through a user feedback mechanism, and the recommendation accuracy and practicability are ensured; a convenient health management tool is provided, and the compliance of the patient on diet management is enhanced; the health condition of the diabetic patient can be improved, the burden of a medical system is relieved, and the public health level is improved.

Description

Technical Field

[0001] The present invention relates to the field of diabetes care, and specifically to a multi-model fusion personalized diabetes diet recommendation system. Background Art

[0002] Dietary management, as a crucial component of diabetes prevention and treatment, directly impacts patients' blood sugar control and the development of complications. However, diabetic patients generally lack nutritional knowledge, and existing dietary management systems often lack personalized and dynamic adjustments, failing to effectively address patients' blood sugar fluctuations and dietary preferences. Currently, some diabetes and dietary management apps still lack personalized recommendations, dynamic adjustments, and user experience. First, the physical conditions and dietary habits of different diabetic patients vary, yet some apps offer generic recommendations that fail to fully consider individual needs and preferences, resulting in diet plans that lack specificity and practicality. For example, the "Sugar Nurse" diabetes management app only recommends dietary recommendations by weight for staple foods, meat, and vegetables, without specific recommendations for types or combinations. Second, these apps also have limitations in dynamic adjustments. Patients' blood sugar levels and physical indicators change over time, but some apps fail to automatically adjust dietary recommendations based on these changes, making it difficult to meet the management needs of patients at different stages of their disease. Furthermore, user experience needs to be improved. Some apps have complex interface designs, poorly designed layouts, and cumbersome workflows. Alternatively, they contain a large amount of dietary health-related information, such as short videos and product promotions, which obscure key information and hinder the efficient implementation of blood sugar management functions.

[0003] Therefore, it is particularly important to develop a system that can provide scientific and personalized dietary recommendations based on patients' glycemic responses and dynamic preferences. Summary of the Invention

[0004] To solve the above technical problems, the present invention provides a technical solution: a multi-model fusion personalized diabetes diet recommendation system, including a recipe recommendation decision module based on the system architecture, a blood glucose prediction module, and an intelligent question-and-answer interaction module. The recipe recommendation decision module is used to analyze the relationship between recipes, ingredients, and nutrients, and generate scientific and personalized recipe recommendations to ensure that both the patient's nutritional needs and blood glucose control goals are met. The recipe recommendation decision module is connected to the blood glucose prediction module via a channel based on a secure transmission protocol.

[0005] The blood sugar prediction module predicts the post-meal blood sugar level based on the user's current blood sugar level and the recommended recipes, ensuring that the recommended recipes will keep the patient's blood sugar level within the normal range;

[0006] The intelligent question-and-answer interaction module provides convenient natural language interaction, collects user feedback and dynamically adjusts recommended content, ensuring that the system continuously learns and adapts to users' health needs.

[0007] Preferably, the recipe recommendation decision module uses a knowledge graph model to generate food information. The knowledge graph model intuitively represents the relationship between recipes and ingredients in the form of nodes and edges, so that the system can quickly query and analyze the nutritional components of recipes, and generate personalized diet recommendations based on the user's health data and preferences. In the process of recommending recipes to users, the system filters out recipe nodes that meet the meal type from the knowledge graph based on the user's health data and meal requirements, and then calculates the health score of each recipe. The health score includes blood sugar score and nutrition score. The health score of each recipe is generated by combining the blood sugar score and nutrition score, and the final recommendation score is calculated in combination with the recipe's preference score. Recipes with higher recommendation scores will be recommended to users first. Health data includes height, weight, age, gender, diabetes type, and meal requirements include breakfast, lunch, and dinner.

[0008] Preferably, the knowledge graph model captures the user's preferences for recipes and ingredients, making recommendations more flexible and accurate. By combining the user's health data and blood sugar prediction model, the knowledge graph can dynamically adjust the recommendation algorithm to provide more accurate and scientific dietary advice. Nodes include recipes and ingredients, edges represent the inclusion relationship between recipes and ingredients, and edge weights represent the weight of ingredients. Recipe nodes include three attributes: name, meal number, and preparation steps. Meals include breakfast, lunch, and dinner. Ingredient nodes include six attributes: name, type, carbohydrate content, protein content, fat content, and dietary fiber content. Types include vegetables, meat, staple food, fruit, and others.

[0009] Preferably, the blood glucose score is based on a blood glucose prediction model, which predicts the blood glucose values 60 minutes, 120 minutes and 180 minutes after the meal through the user's pre-meal blood glucose value, the nutritional components of the recipe and the amount of insulin injection, and scores according to the blood glucose control standard; the nutritional components include carbohydrates, fat and fiber.

[0010] Preferably, the nutritional score is calculated by comparing the nutritional content of the recipe with the user's nutritional needs, which are based on their energy consumption and meal proportions, where breakfast accounts for 30%, lunch accounts for 40%, and dinner accounts for 30%, combined with the recommended intake of carbohydrates, protein, fat and dietary fiber.

[0011] Preferably, the recipe recommendation decision module prompts the user to provide feedback on the recipe rating and post-meal blood glucose level after the user has eaten, and dynamically updates the preference ratings for recipes and ingredients based on the user's feedback.

[0012] Preferably, the blood glucose prediction module uses a feedforward neural network to predict the changes in the user's blood glucose after a meal. The input of the model includes the blood glucose level before the meal, the amount of insulin injection before the meal, and the diet record. The output is the predicted blood glucose level 60, 120, and 180 minutes after the meal. The training data used by the initial model is the blood glucose level recorded by adult patients with type 2 diabetes using continuous blood glucose monitoring. The blood glucose level recorded by the user after the meal will be stored in the back-end database. The model uses this data for regular retraining to improve the personalization of blood glucose prediction; the diet record includes carbohydrate, fat, and dietary fiber intake.

[0013] Preferably, the intelligent question-answering interaction module adopts the Llama language model, which is fine-tuned by LoRA to adapt to the dietary management needs in the field of diabetes. The Llama model learns and adapts to diabetes-related tasks through diabetes literature data, including text generation, question-answering and accurate completion of dialogue tasks, and provides effective health advice according to user needs.

[0014] Preferably, the intelligent question-and-answer interaction module provides two-way feedback to the user through an event-driven architecture, and the system executes the corresponding business logic by triggering events, including receiving recipe recommendation results and pushing them to the user, pushing a recipe rating sheet to the user after recommending a recipe, and regularly reminding patients to take medication and record blood sugar.

[0015] The advantages of the present invention compared with the existing technology are: making full use of the health data of diabetic patients to provide scientific diet recommendations, the knowledge graph organizes recipes, ingredients and their nutritional components in a structured manner to ensure the accuracy and scalability of the data; the blood glucose prediction model uses a feedforward neural network to accurately predict the blood glucose values 60, 120 and 180 minutes after the meal through the user's pre-meal blood glucose, recipe nutritional components and insulin injection volume, and evaluates it in combination with the blood glucose control standards of diabetic patients. The recommendation algorithm comprehensively considers the health score and user preference score to dynamically generate personalized recipes; through the user feedback mechanism, the system can continuously optimize the recommendation results to ensure the accuracy and practicality of the recommendations; the application of the knowledge graph enables the relationship between recipes and ingredients to be efficiently queried and analyzed, and the blood glucose prediction model, combined with deep learning technology, can accurately predict post-meal blood glucose values in real time, providing a scientific basis for diet recommendations; the integration of large language models enables the system to interact with users in the form of natural language, generate detailed recipe descriptions, and analyze user feedback. By dynamically updating preference scores and regularly retraining the model, the system can continuously optimize the recommendation algorithm to improve user experience and recommendation effects; the overall architecture integrates a variety of AI technologies and has high technical Advancedness and practical value; the system can be promoted to hospitals, clinics and primary medical institutions to assist nutritionists and doctors in diagnosis and treatment, and improve patient care efficiency; for diabetic patients, the system reduces the complexity of self-management, provides convenient health management tools, and enhances patients' compliance with dietary management; the core technology of the system has wide scalability and can be applied to the dietary management of other chronic diseases such as hypertension and hyperlipidemia, providing comprehensive personalized solutions for chronic disease patients; it helps to improve the health status of diabetic patients, reduce the burden on the medical system, and promote the digital and popular transformation of chronic disease management through intelligent health management tools, improve the level of public health, and provide convenience for patients with different levels of diabetes management awareness. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram of the knowledge graph of the present invention.

[0017] Figure 2 It is a framework diagram of the present invention.

[0018] Figure 3 This is a diagram of the usage of the intelligent question-answering interactive module.

[0019] Figure 4 This is a standard blood sugar score table. DETAILED DESCRIPTION

[0020] The present invention will be described in further detail below with reference to the accompanying drawings.

[0021] Example 1

[0022] like Figure 1 、 Figure 2 and Figure 3 As shown, this embodiment provides a multi-model fusion personalized diet recommendation system for diabetes, which integrates multi-model technology to provide personalized diet recommendation services for diabetic patients. The system first analyzes the patient's health data and dietary preferences through the recipe recommendation decision module to generate scientific and personalized recipe recommendations. Next, the blood glucose prediction module predicts post-meal blood glucose changes based on the recommended recipes and the patient's current blood glucose level to ensure that the recommended recipes meet the blood glucose control goals. Finally, the intelligent question-answering interaction module provides a natural language interactive interface to collect user feedback and dynamically adjust the recommended content to achieve continuous learning and optimization of the system.

[0023] In terms of module architecture and interaction, the recipe recommendation decision module uses a knowledge graph model to intuitively represent the relationship between recipes and ingredients in the form of nodes and edges. Nodes include recipes and ingredients, while edges represent inclusion relationships and ingredient weights. Based on the user's health data (such as height, weight, age, etc.) and meal requirements (such as breakfast, lunch, and dinner), the module filters out recipe nodes that meet the requirements from the knowledge graph and calculates a health score for each recipe. The health score comprehensively considers the blood sugar score and nutritional score, combined with the user's preference score, to generate a final recommendation score, giving priority to recipes with high scores.

[0024] The blood glucose prediction module uses a feedforward neural network model. Its inputs include pre-meal blood glucose levels, pre-meal insulin injections, and dietary records (such as carbohydrate, fat, and dietary fiber intake). It outputs predicted blood glucose levels 60, 120, and 180 minutes after a meal. The module uses continuous glucose monitoring records from adults with type 2 diabetes as initial training data and periodically retrains using post-meal blood glucose data to further personalize predictions.

[0025] The intelligent question-and-answer interaction module is based on the Llama language model and fine-tuned through LoRA to adapt to the needs of diabetic diet management. The module can learn and adapt to diabetes-related tasks, including text generation, question-and-answering, and conversation, providing effective health advice tailored to user needs. Furthermore, the module employs an event-driven architecture for two-way feedback with users, including receiving and pushing recipe recommendations, pushing recipe ratings, and providing scheduled medication reminders and blood sugar tracking.

[0026] In terms of functionality, the system generates personalized recipe recommendations based on the user's health data, dietary preferences, and meal frequency requirements. The recommendation process comprehensively considers blood sugar control and nutritional needs to ensure that the recommended recipes are both scientific and tailored to individual tastes. The system can predict the user's post-meal blood sugar levels, helping them plan their diet appropriately and control blood sugar fluctuations. At the same time, the system collects and stores the user's post-meal blood sugar data for model retraining and personalized optimization. The system provides an intelligent question-and-answer interactive interface that supports natural language input and output. Users can interact with the system via voice or text to obtain health advice, provide recipe ratings, and more. The system dynamically adjusts recommendations based on user feedback, enabling continuous learning and improvement.

[0027] In summary, this system provides personalized and scientific diet recommendation services for diabetic patients by integrating multi-model technology, helping users to better control blood sugar levels and meet nutritional needs.

[0028] Example 2

[0029] like Figure 1 As shown, this embodiment provides a diet recommendation system based on a knowledge graph (KG). The diet recommendation system needs to comprehensively consider whether the recipe meets the patient's nutritional needs, the recipe's impact on the patient's blood sugar level, and the user's own dietary preferences. To this end, the knowledge graph is an effective form of representing recipes.

[0030] A knowledge graph is a structured semantic knowledge base that represents real-world knowledge through a graph structure of entities and relationships, supporting applications such as information retrieval, data integration, and intelligent reasoning. A knowledge graph consists of nodes and edges, where nodes represent entities and edges represent relationships between them. The advantage of using a knowledge graph to represent recipes is that it can efficiently organize and associate complex recipes, ingredients, and their nutritional information in a structured manner. The knowledge graph intuitively represents the relationships between recipes and ingredients through nodes and edges, enabling the system to quickly query and analyze the nutritional content of recipes and generate personalized dietary recommendations based on the user's health data and preferences. Most importantly, the knowledge graph can capture user preferences for recipes (at the macro level) and ingredients (at the micro level), making recommendations more flexible and accurate. By combining user health data and blood sugar prediction models, the knowledge graph can dynamically adjust its recommendation algorithm to provide more accurate and scientific dietary advice.

[0031] In the graph, nodes include recipes (such as scrambled eggs with tomatoes) and ingredients (such as tomatoes, eggs, and rapeseed oil). Edges represent the inclusion relationship between recipes and ingredients, with edge weights representing the weight of the ingredients (g). Recipe nodes have three attributes: name, meal number (breakfast, lunch, or dinner), and preparation steps. Ingredient nodes have six attributes: name, type (vegetables, meat, staple food, fruit, or other), carbohydrate content, protein content, fat content, and dietary fiber content.

[0032] When recommending recipes to users, the system first filters recipe nodes from the knowledge graph based on the user's health data (such as height, weight, age, gender, diabetes type, etc.) and meal frequency requirements (breakfast, lunch, dinner). The system then calculates a health score for each recipe, which consists of two components: a blood sugar score and a nutritional score.

[0033]

[0034] The formula for calculating the total blood glucose score is:

[0035]

[0036] Among them, scorei represents the blood glucose score at each time point (60, 120, and 180 minutes after a meal).

[0037] The blood glucose score is based on a blood glucose prediction model. It predicts the blood glucose levels 60 minutes, 120 minutes, and 180 minutes after a meal using the user's pre-meal blood glucose level, the nutritional composition of the diet (carbohydrates, fat, fiber), and the amount of insulin injected. It is then scored based on the blood glucose control standards for patients with type 2 diabetes proposed in the "Guidelines for the Prevention and Treatment of Diabetes in China (2024 Edition)."

[0038] The Nutrition Score assesses how a recipe meets the user's nutritional needs, which are based on their total daily energy expenditure (TDEE) and specific meal requirements (breakfast, lunch, or dinner). The Nutrition Score is influenced by the recipe's carbohydrate, protein, fat, and fiber content and measures how well these contents match the user's meal requirements. First, the meal energy requirement is calculated based on the user's TDEE and meal type, with the ratios for different meals being: breakfast (0.3), lunch (0.4), and dinner (0.3). The user's daily nutritional requirements are then adjusted by the corresponding ratios to determine the meal's nutritional requirements. The total energy content of a recipe is calculated based on its macronutrient content, with carbohydrates and protein providing 4 kcal per gram, and fat providing 9 kcal per gram. The Energy Ratio is the ratio of the recipe's energy content to the user's meal energy requirement. Each nutrient in the recipe (carbohydrates, protein, fat, and fiber) is then scaled by the Energy Ratio. The Nutrition Score is calculated by comparing the recipe's scaled nutritional content to the user's meal nutritional requirements. If a recipe's nutrient ratio deviates from 1 (i.e., the optimal match to the user's needs), the score is penalized. The formula for calculating the nutritional score is as follows:

[0039]

[0040] Among them, nutrient i Refers to the scaled nutrient content of the recipe, meal_nutrient_needs i Refers to the user's nutritional needs for that meal.

[0041] It can be seen that the nutritional score is calculated by comparing the nutritional content of the recipe with the user's nutritional needs, which are based on their energy consumption and meal ratio (30% for breakfast, 40% for lunch, and 30% for dinner), and are evaluated in combination with the recommended intake of carbohydrates, protein, fat and dietary fiber.

[0042] The final health score is the product of the glucose score and the nutrient score, which are combined to reflect the user's blood sugar management and the nutritional quality of the meal. In order to prevent the nutrient score from taking too much weight in the health score and ensure that blood sugar level is the main consideration of the nutrient score, the nutrient score is normalized using a logarithmic function. The calculation method of the comprehensive health score is as follows:

[0043] Health Score=log 10 (Nutrient Score)×Glucose Score

[0044] The system will combine the blood sugar score and nutritional score to generate a health score for each recipe, and combine the recipe preference score to calculate the final recommendation score. Recipes with higher recommendation scores will be recommended to users first. After the user has finished the meal, the system will prompt the user to provide feedback on the recipe score (0-10 points) and the post-meal blood sugar value. Based on user feedback, the system will dynamically update the preference scores of recipes and ingredients. In this way, the system can continuously optimize the recommendation algorithm and provide more personalized and accurate dietary recommendations. In addition, the system will regularly collect users' post-meal blood sugar data and retrain the blood sugar prediction model to further improve the accuracy and scientific nature of the recommendations. The entire process not only takes into account the nutritional information of the recipe and the health needs of the user, but also realizes the dynamic optimization of the recommendation mechanism through user feedback, ultimately providing users with a scientific, practical and personalized diet management plan.

[0045] In addition, this solution also uses a preference score calculation. The update of the recipe preference score adopts a weighted average method, combining the existing score with the new score. The weighting factor is determined by the magnitude of the change in the ingredient score. The specific process is as follows: Get the current score: Retrieve the current recipe preference score and ingredient score from the database. Calculate the ingredient score difference: Calculate the absolute difference between the current recipe score and the score of each ingredient. A large difference indicates that the ingredient may not be the main reason for the change in the recipe score, but rather an improper combination of the ingredients that make up the recipe. Weight calculation: The weight is calculated based on the average difference in the ingredient scores. If the difference is small, the weight of the new score is larger, making its impact more obvious; if the difference is large, the weight is reduced, making the overall recipe score account for a larger proportion. The formula is as follows:

[0046]

[0047] Updated Ratings: The new recipe preference rating is calculated as follows:

[0048] Preference Score = existing score × w + new score × (1-w)

[0049] This formula combines existing and new ratings and adjusts them based on weights, allowing the system to appropriately update recipe preference scores. Rating storage and priority adjustment: Updated recipe ratings are stored in the database, and the priority queue is adjusted to reflect the latest ratings. When recommending recipes, the system selects the highest health score from a group of recipes with higher preference scores, ensuring recipe diversity, healthiness, and user preference.

[0050] Example 3

[0051] like Figure 2As shown, this embodiment provides a blood glucose prediction module based on a feedforward neural network (FNN). The feedforward neural network (FNN) is a basic artificial neural network structure with the characteristics of a simple and clear structure and a strong ability to adaptively change weights. It can adapt well to the local regularity of the data and performs well in short-term prediction tasks. In this project, FNN is used to predict changes in the user's blood glucose after a meal. The input of the model includes the blood glucose level before the meal, the amount of insulin injected before the meal, and a detailed diet record (carbohydrate, fat, dietary fiber intake), and the output is the predicted blood glucose level 60, 120, and 180 minutes after the meal. The training data used for the initial model is the blood glucose level recorded by adult patients with type 2 diabetes using continuous glucose monitoring (CGM). As the user uses it, the blood glucose level recorded by the user after the meal will be stored in the back-end database. The model uses this data for retraining regularly to improve the personalization of blood glucose prediction.

[0052] Example 4

[0053] like Figure 3 As shown, this embodiment provides an intelligent interactive assistant based on Llama, a high-performance open source large language model with powerful natural language processing capabilities. In this project, the Llama3 model is fine-tuned (LoRA technology) to adapt to the dietary management needs in the field of diabetes. LoRA (Low-Rank Adaptation of Large Language Models) is a lightweight model fine-tuning technology that freezes the pre-trained weights of the model and learns the adaptation parameters only on the inserted low-rank matrix, significantly reducing the number of parameters that need to be trained and stored. During the fine-tuning training phase, the Llama model uses diabetes literature data to learn and adapt to diabetes-related tasks, including text generation, question answering, and accurate completion of dialogue tasks, and can provide effective health advice based on user needs. The intelligent dialogue assistant can not only provide patients with basic diabetes pathology knowledge and daily life advice, but also provide two-way feedback with users through an event-driven architecture (EDA). The system executes corresponding business logic through event triggering, such as receiving recipe recommendation results and pushing them to users, pushing recipe rating tables to users after recommending recipes, and regularly reminding patients to take medication and record blood sugar. Thanks to Llama3's advanced natural language processing technology, the system can achieve efficient and friendly human-computer interaction. Users can ask questions or express their needs through dialog boxes and receive instant feedback and support, greatly reducing the difficulty of information retrieval, screening and other operations.

[0054] Example 5

[0055] like Figures 1 to 3As shown, this embodiment provides a multi-model fusion personalized diet recommendation system for diabetes, integrating multi-model collaborative recommendation: the project uses knowledge graphs for recipe recommendations, FNN for blood sugar prediction, and combines the Llama assistant with personalized dynamic feedback to achieve multi-model collaborative work.

[0056] Intelligent real-time adjustment: Through real-time dynamic updates based on user health data (such as blood sugar levels, dietary preferences, etc.), the system can not only accurately predict post-meal blood sugar responses, but also continuously optimize dietary recommendations, thereby improving the self-management ability and quality of life of diabetic patients.

[0057] Simple Interaction: By introducing the Llama conversational assistant, the system can interact with users in a concise and natural conversational manner. This interactive approach enhances user engagement and system adaptability, while lowering technical barriers to entry, ensuring ease of use for patients and further improving system accessibility and user experience.

[0058] This system accurately predicts blood sugar levels using an FNN model architecture. Its recommendation algorithm design takes into account both user preferences and dietary diversity: prioritizing recipes with high user preference significantly reduces recipe diversity, while random recipe recommendations fail to account for user taste preferences. Feedback optimization and dynamic updates of user blood sugar data are also available: users' post-meal blood sugar data is collected to enable regular model adjustments.

[0059] In terms of key technical indicators, blood glucose prediction error: RMSE ≤ 2 mmol / L. User satisfaction: target ≥ 85% (based on user feedback scores). Computational efficiency: calculation time per recommendation ≤ 1 second. Llama3 optimization effect: domain question answering accuracy increased by over 20%.

[0060] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and descriptions are merely illustrative of the principles of the present invention. Various changes and improvements may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and improvements are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-model fusion personalized diabetes diet recommendation system, including a recipe recommendation decision module, a blood glucose prediction module, and an intelligent question-answering interaction module based on the system architecture, characterized by: The recipe recommendation decision module is used to analyze the relationship between recipes, ingredients, and nutrients to generate scientific and personalized recipe recommendations to ensure that the patient's nutritional needs and blood sugar control goals are met at the same time; the recipe recommendation decision module is connected to the blood sugar prediction module via a channel based on a secure transmission protocol; The blood glucose prediction module predicts the post-meal blood glucose level based on the user's current blood glucose level and the recommended recipes, ensuring that the recommended recipes will keep the patient's blood glucose level within a normal range; The intelligent question-and-answer interaction module provides convenient natural language interaction, collects user feedback and dynamically adjusts recommended content, ensuring that the system continuously learns and adapts to the user's health needs.

2. The multi-model fusion personalized diabetes diet recommendation system according to claim 1, characterized in that: The recipe recommendation decision module uses a knowledge graph model to generate food information. The knowledge graph model intuitively represents the relationship between recipes and ingredients in the form of nodes and edges, enabling the system to quickly query and analyze the nutritional content of recipes and generate personalized dietary recommendations based on the user's health data and preferences. In the process of recommending recipes to users, the system filters out recipe nodes that meet the meal type from the knowledge graph based on the user's health data and meal needs, and then calculates the health score of each recipe. The health score includes blood sugar score and nutrition score. The health score of each recipe is generated by combining the blood sugar score and nutrition score, and the final recommendation score is calculated in combination with the recipe preference score. Recipes with higher recommendation scores will be recommended to users first; the health data includes height, weight, age, gender, and diabetes type, and meal needs include breakfast, lunch, and dinner.

3. The multi-model fusion personalized diabetes diet recommendation system according to claim 2, characterized in that: The knowledge graph model captures users' preferences for recipes and ingredients, making recommendations more flexible and accurate. By combining users' health data and blood sugar prediction models, the knowledge graph can dynamically adjust the recommendation algorithm to provide more accurate and scientific dietary advice. Nodes include recipes and ingredients, edges represent the inclusion relationship between recipes and ingredients, and edge weights represent the weight of ingredients. Recipe nodes include three attributes: name, meal number, and preparation steps. Meals include breakfast, lunch, and dinner. Ingredient nodes include six attributes: name, type, carbohydrate content, protein content, fat content, and dietary fiber content. Types include vegetables, meat, staple food, fruit, and others.

4. The multi-model fusion diabetes personalized diet recommendation system according to claim 2, characterized in that: The blood sugar score is based on a blood sugar prediction model. It predicts the blood sugar levels 60 minutes, 120 minutes, and 180 minutes after a meal using the user's pre-meal blood sugar level, the nutritional components of the meal, and the amount of insulin injected, and scores the user based on blood sugar control standards. The nutritional components include carbohydrates, fat, and fiber.

5. The multi-model fusion personalized diabetes diet recommendation system according to claim 2, characterized in that: The nutritional score is calculated by comparing the nutritional content of the recipe with the user's nutritional needs, which are based on their energy consumption and meal ratio, with breakfast accounting for 30%, lunch accounting for 40%, and dinner accounting for 30%, combined with the recommended intake of carbohydrates, protein, fat and dietary fiber.

6. The multi-model fusion diabetes personalized diet recommendation system according to claim 1, characterized in that: The recipe recommendation decision module prompts the user to provide feedback on the recipe score and post-meal blood glucose level after the meal, and dynamically updates the recipe and ingredient preference scores based on the user's feedback.

7. The multi-model fusion diabetes personalized diet recommendation system according to claim 1, characterized in that: The blood glucose prediction module uses a feedforward neural network to predict changes in the user's blood glucose after a meal. The model's input includes pre-meal blood glucose levels, pre-meal insulin injection amounts, and dietary records. The output is the predicted blood glucose levels 60, 120, and 180 minutes after the meal. The training data used in the initial model is the blood glucose levels recorded by adult patients with type 2 diabetes using continuous blood glucose monitoring. The user's blood glucose levels recorded after meals will be stored in the back-end database. The model regularly uses this data for retraining to improve the personalization of blood glucose predictions. The dietary records include carbohydrate, fat, and dietary fiber intake.

8. The multi-model fusion personalized diabetes diet recommendation system according to claim 1, characterized in that: The intelligent question-answering interaction module adopts the Llama language model and is fine-tuned by LoRA to adapt to the dietary management needs in the field of diabetes. The Llama model learns and adapts to diabetes-related tasks through diabetes literature data, including the precise completion of text generation, question-answering, and dialogue tasks, and provides effective health advice based on user needs.

9. The multi-model fusion diabetes personalized diet recommendation system according to claim 1, characterized in that: The intelligent question-and-answer interaction module provides two-way feedback to users through an event-driven architecture. The system executes corresponding business logic through event triggering, including receiving recipe recommendation results and pushing them to users, pushing recipe rating sheets to users after recommending recipes, and regularly reminding patients to take medication and record blood sugar.

Images