Personalized old people diet recommendation method based on large language model

Abstract

The invention discloses a personalized old people diet recommendation method based on a large language model, and the method comprises the following steps: collecting the individual data, including food preference, diet history, electronic health records and disease history, of old people through an individual model; loading nutrition data of the group through the group model, wherein the nutrition data comprises a public dietary guide and a nutrition standard; performing information retrieval and conversion through a scheduler; according to the method, personalized diet recommendation is generated through a large language model, a recommendation result is fed back to a user, personalized diet recommendation is realized by combining individual health data and causal reasoning on the basis of specific requirements of old people, the health management effect and recommendation scientificity are effectively improved, and causal reasoning is used for establishing a causal graph and carrying out anti-factual inference, so that the user experience is improved. And the influence of the nutrients on the health result is quantified, so that the subsequent recommendation is more accurate.

Description

Technical Field

[0001] The present invention relates to the field of large models and personalized diet recommendation, and in particular to a personalized diet recommendation method for the elderly based on a large language model. Background Art

[0002] As it is difficult for the elderly to consume a variety of food groups, it is often difficult to achieve ideal dietary diversity. In addition, the decline in cognitive ability also makes it difficult for the elderly to effectively avoid unhealthy food choices. The lack of dietary diversity and quality will have a negative impact on the prognosis of chronic diseases such as diabetes and hypertension, and further deteriorate the health of the elderly, thus falling into a vicious cycle of health deterioration and unhealthy diet.

[0003] Existing personalized food recommendation systems can be roughly divided into four categories: content-based filtering, hybrid filtering, collaborative filtering, and deep learning-based methods.

[0004] In food recommendation systems, content-based filtering methods are widely used for personalized recipe recommendations. These methods mainly rely on matrix decomposition techniques and are good at mining the potential interactive features between users and recipes, including ingredients, cooking methods, and nutritional information, rather than relying solely on user ratings.

[0005] Hybrid filtering (HF) strategies, especially combined with content-based filtering methods, are at the core of many innovations in this field. Research shows that the N-neighbor algorithm commonly used in hybrid filtering predicts user preferences by identifying N users with similar tastes to the user, and uses metrics such as Pearson correlation or cosine similarity to measure similarity. In addition, research explores clustering techniques such as SPARK and enhanced clustering methods to help achieve more accurate recommendations by grouping users into groups with similar interests.

[0006] The core goal of collaborative filtering methods is to identify recipes or other users that are similar to user preferences. To improve the recommendation effect, technologies such as TF-IDF, Tidal Trust Algorithm, and K-Nearest Neighbor (KNN) are often used. TF-IDF is an information retrieval technology that can quantify the importance of terms in a corpus, thereby helping to evaluate the value of certain ingredients in a recipe. The KNN algorithm infers the possibility of recommending a recipe by finding the "neighbors" closest to the user, thereby achieving personalized recommendations.

[0007] Deep learning methods have also received widespread attention in food recommendation systems in recent years. For example, a study called food-RecSys combined BMI, BMR, k-NN, and BPNN techniques and showed effectiveness in improving public awareness of healthy weight

[17] . Ju

[18] proposed another transformer-based deep learning model that specifically recommends restaurant dishes based on users' nutritional needs, thereby developing an accurate restaurant food recommendation system.

[0008] Generally speaking, the above methods have obvious shortcomings in responding to the needs of the elderly population, and fail to effectively combine the particularities of the elderly population in chronic disease management, dietary preferences, and daily operations. In particular, in terms of personalization of nutritional needs, scientific nature of dietary recommendations, and ease of use of system interfaces, these methods do not address the complex health conditions and individual needs of the elderly population, resulting in their recommendation effects and user experience being unable to fully meet the actual needs of the elderly population.

[0009] After searching, the invention patent with Chinese application number CN202310975201.2 discloses a healthy diet recipe recommendation method and system. The division module is used to obtain edible food for the consumer, which is divided into several food databases according to different food types; the calculation module is used to calculate the necessary nutrient content of each food in the food database; the initial recipe acquisition module is used to obtain the consumer's physical condition data, select the corresponding food from the food database according to the physical condition data and the necessary nutrient content, and obtain the initial recipe; the recommended recipe acquisition module is used to obtain the consumer's taste change requirements, filter food from the initial recipe according to the taste change requirements, and obtain recommended recipes. The recommended recipes obtained contain the necessary nutrient content calculated according to the consumer's physical condition data.

[0010] Compared with the existing technology, the invention patent with Chinese application number CN202310975201.2 can provide consumers with necessary nutrition. At the same time, the food in the recommended recipe is obtained according to the consumer's changing taste needs, which can meet the consumer's actual dietary needs in the near future.

[0011] However, in the actual use of the above method, a large amount of garbage and abnormal data will exist in the early data collection and processing process. These data have not been effectively processed, which makes the trained neural network not accurate enough. At the same time, a large amount of non-text data is directly ignored when acquiring data, and the causal relationship between the data has not been effectively verified, which makes the subsequent dietary recommendations deviate from reality. Therefore, a personalized dietary recommendation method for the elderly based on a large language model is needed. Summary of the invention

[0012] The purpose of the present invention is to solve the shortcomings of the prior art that junk and abnormal data are not effectively processed, a large amount of non-text data is directly ignored when acquiring data, and the causal relationship between data is not effectively verified, and a personalized diet recommendation method for the elderly based on a large language model is proposed.

[0013] In order to achieve the above object, the present invention adopts the following technical solutions: A personalized diet recommendation method for the elderly based on a large model, comprising the following methods: Collect individual data of elderly users through personal models, including food preferences, dietary history, electronic health records and disease history, including various data formats, extracted representations, embeddings and text descriptions; Loading nutritional data of the group through the group model, including public dietary guidelines and nutritional standards; Information retrieval and conversion are performed through the scheduler. The scheduler plays a core role in improving the effectiveness of food recommendations and is responsible for three tasks: First, it retrieves the most relevant information from the individual model and the group model based on the user query. Second, it converts non-text information into a text format for the large language model to process. Finally, the large language model is guided by prompts to generate personalized recommendations to avoid irrelevant or spurious outputs. The specific steps include the following: a. Retrieve relevant information from individual models and group models based on user queries; b. Convert non-text information into text format; c. Provide accurate personalized recommendations based on the user’s past data; Generate personalized diet recommendations through a large language model and provide feedback to users; The large language model is used in the method to generate responses. It is responsible for processing the comprehensive information from the scheduler, which combines personal background and group knowledge, and finally generates personalized, nutrition-oriented food recommendations to the user.

[0014] The above technical solution further includes: The data collected by the personal model is processed using causal discovery and causal reasoning to determine the causal impact of nutrition on health outcomes, thereby providing personalized health support for the elderly population; The steps of causal discovery and causal reasoning are as follows: Step 1: Use the SAM algorithm to perform causal discovery, build a causal network based on the individual nutritional intake data and health outcome data of elderly users, and generate a causal graph; Step 2: Use the DoWhy library to perform causal reasoning on the causal graph generated in the first step to derive the specific impact of nutrition on health outcomes; Step 3: Use regression analysis to calculate the direct effects of nutrients on health outcomes and the indirect effects through mediating variables.

[0015] To achieve a balance between model complexity and data fitting, SAM uses the sparsity regularization parameter λ to control the complexity of the causal graph. λ is set to 0.3 to ensure that the generated graph strikes a balance between simplicity and retention of important edges. The learning rate of the generator is set to 0.001, the learning rate of the discriminator is set to 0.0005, and the adversarial balance coefficient α is set to 1.

[0016] During the recommendation process, the scheduler retrieves relevant information from the individual model and the group model, converts non-text information into text format, and guides the large language model through prompt optimization to generate accurate personalized diet recommendation results. The scheduler extracts various non-text data present in the system, including time series data, numerical information in spreadsheets, and charts.

[0017] The large language model generates nutrition-oriented dietary recommendations based on the information provided by the scheduler, combined with personal background and group knowledge.

[0018] The personal model also includes a feedback module and a health tracking module. The feedback module collects user preferences and satisfaction with recommendations and continuously optimizes recommendation results based on user feedback. The user health tracking module displays the historical changes in the user's health indicators and adjusts personalized dietary recommendations based on the data.

[0019] The retrieval function of the scheduler is implemented through a two-stage retrieval process. In the first stage, the BM25 retrieval algorithm is used to retrieve nutritional effects closely related to the elderly user's query from the personal model and the population model. In the second stage, specific foods containing these key nutrients are retrieved from the food database.

[0020] The present invention has the following beneficial effects: 1. In the present invention, by combining individual health data and causal reasoning, personalized dietary recommendations are implemented based on the specific needs of the elderly, effectively improving the health management effect and the scientific nature of the recommendations. Causal reasoning quantifies the impact of nutrients on health outcomes by constructing a causal graph and performing counterfactual inference, thereby making subsequent recommendations more accurate; 2. In the present invention, population models are used to efficiently integrate group nutrition data, simplify the data processing process, and combine group knowledge and individual characteristics to provide scientific and comprehensive personalized dietary recommendations. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Figure 1 This is a flow chart of a personalized diet recommendation method for the elderly based on a large language model proposed by the present invention. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention. Example 1

[0023] A personalized dietary recommendation method for the elderly based on a large model, see Figure 1 , the method comprises the following steps: 101: The personal model incorporates individual data of elderly users (such as food preferences, dietary history, and disease history) into the recommendation process, providing personalized health support for the system and ensuring that dietary recommendations match the user's unique needs.

[0024] 102: Population models are responsible for loading nutrition data from groups, such as public dietary guidelines and nutrition standards, to ensure that dietary recommendations are science-based and consistent with broader health trends and norms.

[0025] 103: The scheduler plays a core role in food recommendation by retrieving personal data provided by the individual model and group data provided by the population model, converting non-text information into text format, and using prompt optimization to guide the large language model to generate accurate personalized diet recommendations.

[0026] 104: The large language model generates personalized, nutrition-oriented dietary recommendations based on the information summarized by the scheduler, combined with personal background and group knowledge, and feeds back to the user to ensure that the recommendation results are accurate and explainable.

[0027] In summary, the embodiments of the present invention achieve personalized dietary recommendations for the elderly by combining individual data, group nutritional knowledge and the capabilities of a large language model. This not only improves the scientificity and accuracy of the recommendations, but also fully considers the health needs of the elderly group and the usability of the system, providing a more comprehensive and effective healthy diet solution for the elderly. Example 2

[0028] The scheme in Example 1 is further introduced below with reference to specific examples, as described below for details: 201: Implementing personal models using causal discovery and causal inference methods; The above step 201 mainly includes: 1) Using the SAM algorithm for causal discovery Through the structure-agnostic modeling algorithm, a causal network is constructed based on the individual nutritional intake data and health outcome data of elderly users. SAM is a causal discovery algorithm based on neural networks, which combines the principles of adversarial learning to identify causal relationships between variables through the interaction of generators and discriminators.

[0029] The generator is responsible for generating candidate causal graphs, and the discriminator verifies the rationality of causal relationships by evaluating the conditional independence and distribution asymmetry of the candidate graphs. In order to achieve a balance between model complexity and data fitting, SAM uses the sparsity regularization parameter λ to control the complexity of the causal graph. The recommended setting is 0.3 to ensure that the generated graph strikes a balance between simplicity and the retention of important edges. The setting of the learning rate is also crucial. The learning rate of the generator is set to 0.001, and the learning rate of the discriminator is set to 0.0005. This slightly higher generator learning rate can speed up the generation of the graph, while the lower discriminator learning rate can prevent the model from overfitting the data.

[0030] In addition, in order to enhance the stability of adversarial learning, the adversarial balance coefficient α is set to 1, which ensures that the generator and discriminator maintain an appropriate weight relationship during training. In terms of data processing, the data noise parameter σ can be set to 0.2. This setting can adapt to data with different noise levels. A lower noise parameter is suitable for situations with higher data quality, while a higher noise parameter helps to enhance the robustness of causal discovery in a high-noise environment. The number of training iterations (epochs) is set to 2000. Experimental verification shows that a longer training time helps to generate a more stable causal graph. At the same time, the final number of iterations needs to be dynamically adjusted based on the performance of the validation set. These parameters of SAM interact with each other during training. For example, a higher learning rate needs to be combined with a slightly larger sparsity regularization parameter to control the model from generating overly complex causal graphs, and the adversarial balance coefficient also needs to be adjusted synchronously with the learning rate and regularization parameter to maintain the effect of adversarial learning.

[0031] The causal discovery tool library is used to process the nutritional intake and health outcome data. The causal graph generated by SAM is a directed acyclic graph (DAG), which intuitively shows the possible causal paths between the intake of each nutrient and the health outcome. This causal graph not only includes direct effects, but also reveals indirect effects through mediating variables, forming a complete causal network, laying a solid foundation for subsequent causal reasoning. These recommended parameter settings and experimental results prove that SAM can achieve a good balance between sparsity, accuracy, and stability.

[0032] 2) Using the DoWhy library for causal reasoning The DoWhy library is used to perform causal reasoning on the causal graph generated in the first step to quantify the impact of nutrition on health outcomes. First, the causal graph is imported into the DoWhy framework as a structural input to define the causal relationship model. In the process of model definition, the causal relationship is set based on the following aspects: First, the generation of the causal graph is based on the SAM algorithm and the Causal Discovery Tool Library (CDT). The causal graph has preliminarily captured the causal relationship between variables through conditional independence tests and distribution asymmetry analysis, which provides a reliable structural foundation for the model. Secondly, combined with domain knowledge, such as causal hypothesis in nutrition and medical literature, the rationality of the causal path between variables is further verified. For example, the effect of certain nutrients (such as vitamin D) on bone density improvement has been confirmed by a large number of studies, and this knowledge provides support for model definition. In addition, the time series characteristics of the data itself also provide a basis for causal inference. For example, the potential causal relationship can be captured by associating changes in daily nutrient intake with lagged changes in health indicators.

[0033] After defining the model, the average treatment effect (ATE) is calculated using methods in the DoWhy library to quantify the average causal impact of a specific intervention (such as increasing the intake of a certain nutrient) on health outcomes. The calculation of ATE relies on counterfactual inference logic, which infers the net effect of the intervention by simulating changes in health indicators under different intervention scenarios. Finally, the causal inference results are verified through sensitivity analysis and robustness testing to ensure that the inferred causal effects are credible and robust. For example, by adjusting the potential impact of unobserved variables on the model, it is assessed whether the causal conclusions are significantly offset by hidden variables or confounding factors.

[0034] This process effectively quantifies the overall impact of nutritional intake on health outcomes and verifies the accuracy of causal effects. By combining causal graph generation, domain knowledge support, and counterfactual inference, the DoWhy framework can provide scientific, reasonable, and explainable causal reasoning results.

[0035] 3) Conduct mediation analysis Further analysis of the causal paths in the causal diagram was performed to find mediating variables to gain a deeper understanding of the causal mechanism. The mediating variables found in the causal diagram were identified, which may be the intermediate processes by which nutrients affect health outcomes through indirect pathways. Subsequently, statistical methods such as regression analysis or structural equation modeling (SEM) were used to calculate the direct effects of nutrients on health outcomes and the indirect effects through mediating variables.

[0036] 202: Using a comprehensive food database to implement a population model; The above step 202 mainly includes: 1) Load food information Food information is loaded through the Cronometer food recorder database, including the ingredients, nutritional content, texture, smell and taste of each food. In order to increase the richness of the data, relevant nutritional information is also selected and loaded from other open API databases such as Nutritionix. In this way, the population model obtains sufficiently comprehensive data, covering the detailed nutritional composition of different foods, to support the subsequent steps.

[0037] 2) Transmit information After successfully loading the nutrient data, the population model passes this information directly to the system’s scheduler for further integration and utilization of the data.

[0038] 203: Implement the information retrieval, data conversion and prompt optimization functions of the scheduler; The above step 203 mainly includes: 1) Implement the information retrieval function of the scheduler in stages The retrieval function of the scheduler is implemented through a two-stage retrieval process. In the first stage, the BM25 retrieval algorithm is used to retrieve nutritional effects that are closely related to the elderly user's query from personal models and population models. For example, when a user requests recommendations related to "enhancing deep sleep", the BM25 algorithm can quickly find nutrients that are known to have a significant impact on deep sleep. Next, the second stage focuses on retrieving specific foods that contain these key nutrients from the food database. To do this, the system ranks foods according to the nutrient content per calorie and ultimately selects the top ten foods to ensure that the recommended foods can effectively meet the health needs of elderly users.

[0039] 2) Implement the data conversion function of the scheduler through predefined templates The transcription function of the scheduler is used to convert non-text information into text descriptions suitable for processing by large language models. First, the scheduler extracts various non-text data present in the system, such as time series data, numerical information in spreadsheets, charts, etc., which are usually not directly understood by large language models. In order to perform data conversion, the scheduler applies a set of predefined templates to convert these different types of data into natural language. For example, for time series data, the scheduler extracts the changing trends and key values ​​and describes them in simple sentences as "the user's vitamin intake has been on an upward trend in the past week, with an average daily intake of X mg." For food nutrition data in spreadsheets, the scheduler describes the specific ingredients of each food (such as calories, vitamins, and mineral content) in order, such as "each 100 grams of spinach contains 23 kcal of energy, 2.9 grams of protein, and 558 mg of potassium."

[0040] 3) Implement the scheduler's prompt optimization function through instruction templates When implementing prompt optimization, the scheduler provides detailed prompts to the large language model through instruction templates, such as "Please make recommendations based only on the provided food ingredient list and nutritional effects." These prompts guide the large language model to strictly follow specific information without generating irrelevant or false content. In addition, we adopted the thought chain method of zero-shot learning to further enhance users' trust in recommendations. To this end, the scheduler attached the instruction "step by step Explain your recommendation", requiring the large language model to explain the process and basis of the recommendation step by step to ensure the transparency and consistency of the recommendation results, helping users better understand the recommended content.

[0041] 204: Generating Personalized Dietary Recommendations for the Elderly Using Large-Scale Training Language Models; The above step 204 mainly includes: 1) Receive data input from the scheduler The large language model first receives integrated data from the scheduler, including individual health data provided by the personal model and group nutrition data provided by the population model. These data have been processed by the scheduler and converted into a text format suitable for model understanding.

[0042] 2) Generate recommendations based on user background and group knowledge The large language model combines input data with its internally trained knowledge to generate personalized, nutrition-oriented dietary recommendations taking into account the user's individual background (such as dietary preferences and health status) as well as group health trends. To ensure that the recommendations are interpretable and relevant, when generating recommendations, the large language model optimizes instructions based on prompts to ensure that the recommended content strictly follows the prompts provided by the scheduler to avoid generating irrelevant or illusory content. At the same time, by gradually explaining the basis for the recommendations (such as the impact of nutrients on health), the recommendations are well interpretable.

[0043] 3) Feedback personalized recommendations to users Finally, the generated recommendation results are fed back to users in a form that is easy for users to understand. The presentation of the recommendation results includes a combination of pictures and texts, video tutorials, and detailed text descriptions to help users more intuitively understand the daily diet arrangements and the health benefits behind them. The recommended content not only provides daily meal combinations, nutritional analysis, and ingredient selection suggestions, but also includes explanations of the health reasons behind each recommendation, such as how certain ingredients can help improve specific health problems or enhance nutritional intake, so that users can have a deeper understanding and trust in the recommendation results. Example 3

[0044] Based on the carrier used in the method of Example 1-2, see the following description for details. This example mainly introduces a user interface design of a personalized elderly diet recommendation system based on a large model, which aims to enable the elderly and their families to conveniently view and manage diet recommendations and optimize the recommendation effect based on user feedback. The user interface design is intuitive, convenient and interactive, ensuring that elderly users can smoothly obtain information and perform personalized operations during use: 1) User interface presentation The interface shows each recommended meal in detail, including a list of ingredients, specific cooking methods, and nutritional analysis. It is also equipped with high-definition pictures to help users intuitively understand the required ingredients and the finished product effect. For the recommended dishes, the interface will display the searched related cooking video links. These videos can demonstrate how to prepare food step by step, helping users to understand how to handle ingredients, cooking steps, and time control by watching videos. The links are presented as simple buttons, and users can watch related videos by simply clicking, helping elderly users to implement the recommended plan more conveniently. In addition, the interface also includes a daily progress panel that shows the comparison of the user's current nutritional intake with the recommended values. These data are presented in simple charts, which are easy for the elderly to understand and allow them to grasp their dietary health status for the day in real time.

[0045] 2) Interactive management tools Users can rate each recommended dish or fill out feedback, such as taste preferences, availability of ingredients, etc. The system will record these feedbacks and use them to optimize the next recommendation results. The interface allows users or family members to personalize dietary recommendations based on the physical condition of the elderly, such as replacing certain ingredients or adjusting the portion size of dishes. At the same time, the system will recalculate the nutrients based on these modifications and update the recommendations in real time. The interface also integrates health tracking functions for elderly users, showing historical changes in health indicators such as weight, blood sugar, and blood pressure, and further adjusting dietary recommendations based on these changes to make recommendations more in line with the user's actual health status.

[0046] 3) Dynamic optimization function Based on the long-term feedback data of elderly users, the system will automatically analyze the changes in user preferences and health data trends. In this way, the system will gradually optimize the recommendation model to continuously improve its adaptability to user needs. Users can view their historical recommendations and adjustment records to understand how the system has been optimized based on their feedback. This not only increases the transparency of the system, but also allows the elderly and their families to have more trust in the reliability of the recommendations. In addition, the system is designed with a "Family Interaction" section, where family members can view dietary recommendations and health changes synchronously with the user's interface through the mobile terminal, and provide additional support for the elderly. This function takes into account the role of family members in particular to ensure that the elderly get more help in diet management.

[0047] It should be pointed out here that the device description in the above embodiment corresponds to the method description in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0048] The execution subjects of the above-mentioned processor and memory can be devices with computing functions such as computers, single-chip microcomputers, and microcontrollers. In specific implementation, the embodiments of the present invention do not limit the execution subjects and are selected according to the needs of actual applications.

[0049] The data signal is transmitted between the memory and the processor via a bus, which is not described in detail in the embodiment of the present invention.

[0050] Based on the same inventive concept, an embodiment of the present invention further provides a computer-readable storage medium, the storage medium includes a stored program, and when the program is running, the device where the storage medium is located is controlled to execute the method steps in the above embodiment.

[0051] The computer-readable storage medium includes but is not limited to a flash memory, a hard disk, a solid-state drive, and the like.

[0052] It should be pointed out here that the description of the readable storage medium in the above embodiment corresponds to the description of the method in the embodiment, and the embodiment of the present invention will not be described in detail here.

[0053] In the above embodiments, all or part of the embodiments may be implemented by software, hardware, firmware or any combination thereof. When implemented by software, all or part of the embodiments may be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions according to the embodiments of the present invention are generated.

[0054] The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. Computer instructions may be stored in a computer-readable storage medium or transmitted via a computer-readable storage medium. The computer-readable storage medium may be any available medium that can be accessed by the computer or a data storage device such as a server or a data center that includes one or more available media. The available medium may be a magnetic medium or a semiconductor medium, etc.

[0055] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A personalized diet recommendation method for the elderly based on a large language model, characterized in that: The following methods are included: Collect individual data of elderly users through personal models, including food preferences, dietary history, electronic health records and disease history; Loading nutritional data of the group through the group model, including public dietary guidelines and nutritional standards; Information retrieval and conversion is performed through the scheduler, which includes the following steps: a. Retrieve relevant information from individual models and group models based on user queries; b. Converting non-text information in the retrieved relevant information into text format; c. Provide accurate personalized recommendations based on the user’s past data; Generate personalized diet recommendations through a large language model and provide feedback to users.

2. The personalized diet recommendation method for the elderly based on a large language model according to claim 1, characterized in that: The data collected by the individual models are processed using causal discovery and causal inference to determine the causal effects of nutrition on health outcomes; The steps of causal discovery and causal reasoning are as follows: Step 1: Use the SAM algorithm to perform causal discovery, build a causal network based on the individual nutritional intake data and health outcome data of elderly users, and generate a causal graph; Step 2: Use the DoWhy library to perform causal reasoning on the causal graph generated in the first step to derive the specific impact of nutrition on health outcomes; Step 3: Use regression analysis to calculate the direct effects of nutrients on health outcomes and the indirect effects through mediating variables.

3. The personalized diet recommendation method for the elderly based on a large language model according to claim 2, characterized in that: To achieve a balance between model complexity and data fitting, SAM uses the sparsity regularization parameter λ to control the complexity of the causal graph. λ is set to 0.3 to ensure that the generated graph strikes a balance between simplicity and retention of important edges. The learning rate of the generator is set to 0.001, the learning rate of the discriminator is set to 0.0005, and the adversarial balance coefficient α is set to 1.

4. The personalized diet recommendation method for the elderly based on a large language model according to claim 1, characterized in that: The scheduler retrieves relevant information from the individual model and the group model during the recommendation process, converts non-text information into text format, and guides the large language model through prompt optimization to generate accurate personalized diet recommendation results.

5. The personalized diet recommendation method for the elderly based on a large language model according to claim 4, characterized in that: The large language model generates nutrition-oriented dietary recommendations based on the information provided by the scheduler, combined with personal background and group knowledge.

6. The personalized diet recommendation method for the elderly based on a large language model according to claim 1, characterized in that: The personal model also includes a feedback module and a health tracking module. The feedback module collects user preferences and satisfaction with recommendations and continuously optimizes recommendation results based on user feedback. The user health tracking module displays the historical changes in the user's health indicators and adjusts personalized dietary recommendations based on the data.

7. The personalized diet recommendation method for the elderly based on a large language model according to claim 5, characterized in that: The retrieval function of the scheduler is implemented through a two-stage retrieval process. In the first stage, the BM25 retrieval algorithm is used to retrieve nutritional effects closely related to the elderly user's query from the personal model and the population model. In the second stage, specific foods containing these key nutrients are retrieved from the food database.

8. The personalized diet recommendation method for the elderly based on a large language model according to claim 4, characterized in that: The scheduler extracts various non-text data present in the system, including time series data, numerical information in spreadsheets, and charts.

Images