Artificial intelligence-based personalized anti-inflammatory diet recipe recommendation system and method

Abstract

The present application relates to the medical health information technology field, and discloses a personalized anti-inflammatory diet recipe recommendation system and method based on artificial intelligence; the present application constructs a user health portrait through collecting multi-source data such as wearable devices, electronic health records, gene sequencing, combines a dynamically updated anti-inflammatory knowledge graph, evaluates the user's quantitative inflammation load index and nutrient intervention target by using multi-modal deep learning based on the attention mechanism, adopts a hybrid intelligent optimization algorithm combining constraint satisfaction and evolutionary computing to generate the Pareto optimal anti-inflammatory recipe under the constraints of taste taboo, cost, cooking time and the like, and optimizes the model online through user physiological and subjective feedback; personalized, precise and dynamically adaptive anti-inflammatory diet recommendation is realized, and the scientificity and compliance are improved from group guidelines to individual intervention.

Description

Technical Field

[0001] This invention relates to the field of medical and health information technology, specifically to a personalized anti-inflammatory diet recipe recommendation system and method based on artificial intelligence. Background Technology

[0002] Chronic low-grade inflammation is widely recognized in the medical community as an important pathological basis for many chronic diseases, including cardiovascular disease, type 2 diabetes, autoimmune diseases, and certain types of cancer. Dietary structure is closely related to inflammation levels; specific food components such as omega-3 fatty acids and polyphenols have anti-inflammatory effects, while refined carbohydrates and trans fatty acids may promote inflammatory responses. Based on this understanding, the dietary inflammation index, a quantitative tool, has been proposed to assess the inflammatory potential of an individual's diet. However, translating this theory into effective daily dietary guidance still faces many challenges.

[0003] Currently, traditional anti-inflammatory dietary guidance mainly relies on general dietary guidelines or the personal experience of nutritionists. While general guidelines have some reference value, they fail to fully consider the significant differences among individuals in terms of genetic background, gut microbiota composition, underlying disease states, metabolic characteristics, and lifestyles. For the same food, inflammatory responses may vary significantly among individuals; for example, some populations exhibit genetic polymorphisms in their ability to metabolize specific food components. Although one-on-one consultations with nutritionists can provide a degree of personalization, limitations in the professional expertise of nutritionists, time commitment, and service accessibility make it difficult to meet the needs of large populations.

[0004] In existing technologies, some diet management applications or websites offer recipe recommendation functions. These systems typically generate recipes based on basic information manually entered by the user, such as age, gender, weight, and preset dietary goals, such as weight loss or muscle gain, using fixed rules or simple scoring mechanisms. However, these systems have significant limitations: they often focus only on the ratio of macronutrients, lacking in-depth consideration of the complex relationship between food micro-components and inflammatory pathways; they cannot dynamically integrate real-time physiological data from users, such as continuous glucose monitoring data or physiological indicator changes obtained from wearable devices; and they struggle to utilize actual user feedback to optimize subsequent recommendations, meaning they lack the ability to continuously learn from data. Furthermore, existing systems typically treat food selection as a static optimization problem, failing to construct a knowledge system encompassing multi-level entities and their complex interactions, including food components, inflammatory mediators, metabolic pathways, and disease phenotypes.

[0005] With the rapid development of artificial intelligence technology, especially the maturity of deep learning, knowledge graphs, and multimodal data fusion technologies, new technical paths have been provided to solve the aforementioned problems. How to utilize these advanced technologies to integrate multi-source heterogeneous data, construct a system capable of accurately assessing individual inflammatory states, dynamically understanding food-inflammation interaction mechanisms, and generating highly personalized, executable, and continuously evolving anti-inflammatory diet recommendations, has become a pressing technical problem in this field. To address this, a personalized anti-inflammatory diet recommendation system and method based on artificial intelligence are proposed. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a personalized anti-inflammatory diet recommendation system and method based on artificial intelligence, thereby resolving the problems mentioned in the background.

[0007] To achieve the above objectives, the present invention provides the following technical solution: Firstly, an AI-based personalized anti-inflammatory diet recommendation system includes: The multi-source data acquisition and preprocessing module is configured to collect users' time-series physiological data, static demographic data, clinical biochemical indicator data, metagenomic data, and dietary behavior log data from wearable devices, electronic health records, user terminals, and gene sequencing reports; and to clean, normalize, discretize, and perform feature engineering on the collected multi-source data to construct a structured user health profile vector. Preferably, the feature engineering processing of time-series physiological data in the multi-source data acquisition and preprocessing module includes extracting frequency and time domain features of blood glucose fluctuation amplitude, area under the curve, peak time, and nocturnal heart rate variability; the processing of metagenomic data includes calculating microbial α diversity, β diversity, and the relative abundance of specific anti-inflammatory probiotics.

[0008] The dynamic anti-inflammatory knowledge graph module is configured to store and manage multi-level entities and their relationships in the form of a graph database. The entities include food components, nutrients, bioactive compounds, metabolic pathways, cytokines, and disease phenotypes. The relationships include activation, inhibition, transformation, inclusion, and association. The entity and relationship weights in the knowledge graph are dynamically updated based on medical literature using natural language processing technology. Preferably, the relationships between entities in the dynamic anti-inflammatory knowledge graph module include the relationship between food and inflammation score quantified by the dietary inflammation index system, and the relationship between food components and cytokine regulation mined by retrieving medical literature; the knowledge graph is updated by periodically crawling medical literature databases and using a pre-trained biomedical relationship extraction model. A multimodal deep learning evaluation engine based on attention mechanism is configured to receive the user health profile vector and fuse it with the user-related subgraph embedding vector extracted from the dynamic anti-inflammatory knowledge graph module; using a deep neural network model including attention mechanism and Transformer architecture, it outputs the user's quantitative inflammatory load index, key pro-inflammatory pathway activity score and personalized nutrient intervention targets. Preferably, the multimodal deep learning evaluation engine further includes: Temporal encoders, based on long short-term memory networks or Transformer models, are used to encode users' temporal physiological data and extract time-dependent features; A graph neural network encoder, based on a graph attention network or GraphSAGE, is used to extract embedded representations of user-specific subgraphs from the dynamic anti-inflammatory knowledge graph module; The fusion layer employs tensor fusion or an attention-based cross-modal interaction network to deeply fuse the time-dependent features, user health profile vectors, and graph embedding representations. The output layer outputs multi-dimensional inflammatory status assessment results through a fully connected network. These multi-dimensional inflammatory status assessment results include the quantitative inflammatory load index, key pro-inflammatory pathway activity scores, and personalized nutrient intervention targets, as well as one or more selected from systemic inflammatory indices, intestinal inflammatory indices, and neuroinflammatory indices. Preferably, the multimodal deep learning evaluation engine is a pre-trained and fine-tuned model, whose pre-training tasks include mask attribute reconstruction and relation prediction, and whose fine-tuning tasks are based on supervised learning on a user dataset containing explicit inflammatory marker detection results; The multi-constraint, multi-objective recipe generation and optimization engine is configured to receive the quantitative inflammatory load index, the activity score of key pro-inflammatory pathways, and personalized nutrient intervention targets. Based on the dynamic anti-inflammatory knowledge graph module, it uses a multi-objective optimization algorithm that combines constraint satisfaction algorithm and evolutionary computation to generate a set of Pareto optimal candidate anti-inflammatory diet recipes under the constraints of macronutrient ratio, micronutrient upper limit, user taste restrictions, ingredient availability, and cooking time limit. Preferably, the multi-objective optimization algorithm used in the multi-constraint multi-objective recipe generation and optimization engine has the following optimization objective function: The first objective function is to maximize the overall anti-inflammatory score of the recommended recipes, which is calculated by weighting the anti-inflammatory index of each food in the recipe and its correlation with the user's specific pro-inflammatory pathways in the knowledge graph. The second objective function is to maximize the match between the recipe and the user's long-term dietary preferences. This match is calculated based on collaborative filtering of the user's historical diet logs and taste preference data. The third objective function minimizes the potential food-drug interaction risk in the recipe, which is assessed based on the user's current medication list and the food-drug interaction relationships in the knowledge graph.

[0009] The interactive feedback and model self-evolution module is configured to display the candidate anti-inflammatory diet recipes in a visual interface and collect real-time physiological feedback data and subjective experience feedback from users after executing the recipes. Using the feedback data, the model parameters of the multimodal deep learning evaluation engine and the multi-constraint multi-objective recipe generation and optimization engine are fine-tuned and optimized online through reinforcement learning or incremental learning techniques.

[0010] Secondly, a personalized anti-inflammatory diet recommendation method based on artificial intelligence, applied to the system described in the first aspect, includes the following steps: Step S1: Collect users' personal data from multiple heterogeneous data sources, perform data cleaning and feature engineering, and construct user health profiles. Step S2: Input the user's health profile into a pre-trained multimodal deep learning evaluation engine. This engine combines a dynamically updated anti-inflammatory knowledge graph to analyze and output a personalized inflammatory status evaluation report for the user. The report includes a quantitative inflammatory load index, a key pro-inflammatory pathway activity score, and nutrient intervention targets. Step S3: Construct a multi-objective optimization problem for recipe generation, where the decision variables are the foods selected from the ingredient database and their quantities, the constraints include the user's dietary restrictions and recommended nutrient intake range, and the optimization objectives are to maximize anti-inflammatory benefits, maximize taste preference satisfaction, and minimize the risk of food-drug interactions. Preferably, the constraints constructed in step S3 include hard constraints and soft constraints: hard constraints are conditions that must be met, including allergen exclusion and religious taboos; soft constraints include food costs not exceeding the budget and cooking time not exceeding the threshold. Violations of soft constraints are introduced into the objective function through penalty terms. Step S4: The multi-objective optimization problem is solved using a hybrid intelligent optimization algorithm that combines constraint satisfaction and evolutionary algorithm, generating a diverse set of Pareto optimal candidate recipes. Preferably, the hybrid intelligent optimization algorithm in step S4 specifically involves: firstly, using the constraint satisfaction algorithm to prune the search space to generate a high-quality initial recipe population; then, using a non-dominated sorting genetic algorithm with an elitist strategy for iterative evolution to approximate the Pareto front; Step S5: Based on preset rules or user historical selection preferences, select the final recommended recipe from the Pareto optimal candidate recipes and display it through the user interface; Step S6: Continuously track the user's implementation of the recommended recipes and the resulting physiological and psychological feedback, and store the feedback data in the feedback database; Step S7: Using the new data in the feedback database, the rating prediction model in the multimodal deep learning evaluation engine and the multi-objective optimization algorithm is updated through online learning or incremental training, thereby realizing the self-evolution of the recommendation system.

[0011] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention integrates multi-source heterogeneous data, including real-time physiological data from wearable devices, clinical biochemical indicators, and metagenomic data, to construct a comprehensive user health profile, achieving accurate quantitative assessment of individual inflammatory status and overcoming the limitations of traditional methods that rely on single-dimensional data for rough judgment.

[0012] 2. This invention constructs a dynamically updated anti-inflammatory knowledge graph, which structurally stores and associates multi-level entities such as food components, metabolic pathways, inflammatory mediators, and disease phenotypes, as well as their interactions. This enables recipe recommendations to be based on the ratio of macronutrients and delve into the anti-inflammatory mechanisms at the molecular level, thereby improving the scientific rigor and accuracy of the recommendations.

[0013] 3. This invention employs a multimodal deep learning evaluation engine based on an attention mechanism, which can effectively integrate temporal physiological features, static health profiles, and knowledge graph embedding representations to uncover key nutrient intervention targets that are highly correlated with specific pro-inflammatory pathways in individuals, thus achieving a technological leap from group dietary guidelines to personalized precision nutrition intervention.

[0014] 4. This invention utilizes a multi-constraint, multi-objective recipe generation and optimization engine to maximize the anti-inflammatory benefits of recipes and minimize the risks of food-drug interactions while satisfying realistic constraints such as user taste preferences, allergies, ingredient costs, and cooking time. The generated Pareto optimal candidate recipes are both scientifically sound and feasible, thus improving user compliance.

[0015] 5. This invention collects real-time physiological data and subjective experiences of users after executing a recipe through an interactive feedback and model self-evolution module, and uses reinforcement learning or incremental learning techniques to fine-tune and optimize the core model online, forming a data-driven, closed-loop iterative intelligent recommendation mechanism that enables the system to continuously learn and adapt to individual dynamic changes.

[0016] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0017] Figure 1This is a diagram illustrating the architecture of the AI-based personalized anti-inflammatory diet recommendation system of this invention. Figure 2 This is a structural diagram of the multimodal deep learning evaluation engine of the present invention; Figure 3 This is a flowchart of the multi-objective optimization algorithm of the present invention; Figure 4 This is a flowchart of the personalized anti-inflammatory diet recipe recommendation method based on artificial intelligence according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figures 1-4 The present invention relates to an AI-based personalized anti-inflammatory diet recommendation system and method. In this embodiment, the system generates a dynamically adaptive anti-inflammatory diet plan for users through multi-source data fusion, dynamic knowledge graph, deep learning evaluation, and multi-objective optimization.

[0020] Example 1: System Architecture and Module Implementation This embodiment describes the specific structure of the system and its internal working mechanism. The system includes a multi-source data acquisition and preprocessing module, a dynamic anti-inflammatory knowledge graph module, an attention-based multimodal deep learning evaluation engine, a multi-constraint, multi-objective recipe generation and optimization engine, and an interactive feedback and model self-evolution module.

[0021] 1. Multi-source data acquisition and preprocessing module This module is responsible for collecting individual user data from multiple data sources, and then cleaning, normalizing, discretizing, and performing feature engineering to ultimately construct a structured user health profile vector.

[0022] The data sources specifically include: Wearable devices, such as continuous glucose monitors and smartwatches, collect time-series physiological data, including blood glucose levels per minute, heart rate, heart rate variability, steps, sleep duration, etc.

[0023] Electronic health records: These records obtain users' clinical and biochemical data from the hospital information system, such as high-sensitivity C-reactive protein, interleukin-6, tumor necrosis factor-α, lipid profile, and fasting blood glucose; as well as demographic data such as age, gender, height, and weight.

[0024] User terminal: Collects static demographic data and dietary behavior log data filled in by users through mobile applications, including the type, quantity, and cooking method of each meal, as well as taste preferences, allergen information, religious taboos, etc.

[0025] Gene sequencing reports: Obtain users' genotype data, such as information on inflammation-related single nucleotide polymorphism sites; and metagenomic data, such as 16S rRNA sequencing results of gut microbiota.

[0026] The data preprocessing process is as follows: Cleaning: Outliers and missing values ​​are removed. For continuous blood glucose monitoring data, moving averages are used to smooth noise; for clinical indicators, missing values ​​are imputed or removed based on historical records.

[0027] Normalization: Standardizing features with different dimensions. Z-score normalization is used, with the formula:

[0028] in These are the original eigenvalues. This is the mean of the feature in the population. Standard deviation. Normalized features. The mean is 0 and the variance is 1.

[0029] Discretization: Binning certain continuous features. For example, age can be divided into youth, middle-aged, and elderly intervals. Using equal-frequency binning, each interval contains an approximate number of samples.

[0030] Feature engineering: Extracting high-order features from time-series physiological data. For example, blood glucose time series data. The extracted features include: Blood glucose fluctuation range:

[0031] Area under the curve: ,in The sampling interval is denoted as .

[0032] Peak time: The point at which blood glucose reaches its highest value.

[0033] Nocturnal heart rate variability: Frequency domain analysis was performed on the time series of heart rate intervals during the nighttime period to calculate the low-frequency power (LF) and high-frequency power (HF) and obtain the LF / HF ratio.

[0034] For metagenomic data, calculate microbial community α-diversity indices such as the Shannon index: Where S is the number of species, For the first The relative abundance of each species was calculated; β-diversity was measured using Bray-Curtis distance to assess community differences between samples. The relative abundance of specific anti-inflammatory probiotics, such as *Faecalibacterium prausnitzii* and *Roseburia*, was also extracted.

[0035] After the above processing, all features are concatenated into a high-dimensional vector, denoted as . That is, the user health profile vector, where For feature dimensions.

[0036] 2. Dynamic Anti-inflammatory Knowledge Graph Module This module uses the graph database Neo4j to store and manage multi-level entities and their relationships. Entity types include: food components such as curcumin and quercetin; nutrients such as vitamin D and omega-3 fatty acids; bioactive compounds such as short-chain fatty acids; metabolic pathways such as the NF-κB pathway and the COX-2 pathway; cytokines such as IL-6 and TNF-α; disease phenotypes such as rheumatoid arthritis and atherosclerosis; and inflammation scores such as the Dietary Inflammation Index (DII) score. Relationship types include: activation such as "TNF-α activates the NF-κB pathway"; inhibition such as "curcumin inhibits COX-2"; transformation such as "dietary fiber is converted into short-chain fatty acids"; inclusion such as "apples contain quercetin"; association such as "IL-6 is associated with cardiovascular disease risk"; and quantification such as "foods have a DII score," which is calculated using the Dietary Inflammation Index system. Furthermore, relationships between food components and cytokine regulation (such as "resveratrol inhibits IL-6 expression") mined from medical literature are also dynamically updated and incorporated into the graph.

[0037] Knowledge graph is defined as ,in For a collection of entities, Let be the set of relation edges. A collection of relation types. Each entity With initial eigenvectors Each relationship It has a learnable embedding representation. Furthermore, each edge... With a weight This indicates the strength of evidence for the relationship. The initial value is set based on the influence of the literature and can be dynamically updated later.

[0038] The knowledge graph update mechanism: The system periodically crawls medical literature databases such as PubMed and uses a pre-trained biomedical relation extraction model based on BERT for text mining to identify new entity relations or update the weights of existing relations from the latest published abstracts and full texts. For example, for a newly discovered relation "resveratrol inhibits IL-6 expression through the SIRT1 pathway," if the literature has a high impact factor, the weight of the corresponding edge will be increased. The update frequency is once a month.

[0039] 3. Attention-based multimodal deep learning evaluation engine The engine receives user health profile vectors. It extracts user-related subgraphs from the dynamic anti-inflammatory knowledge graph, generates subgraph embeddings, and then fuses multimodal information through an attention mechanism to output inflammation assessment results.

[0040] Subgraph extraction: Based on features in the user's health profile (such as disease phenotype, genotype, and gut microbiota characteristics), relevant entities are retrieved from the knowledge graph as seed nodes and expanded to 2-hop neighbors to form a user-specific subgraph. .

[0041] Graph Neural Network Encoding: A graph attention network is used to encode... Encode the nodes. Its initial characteristics are GAT aggregates neighbor information through a multi-head attention mechanism:

[0042] in For the number of attention heads, , For nodes The neighborhood group, For the first The linear transformation matrix of head attention, For non-linear activation functions such as ELU, Indicates splicing. Attention coefficient. The calculation is as follows:

[0043] in These are learnable parameter vectors. After passing through L layers of GAT, the final representation of each node is obtained. Then, subgraph-level embeddings are obtained through average pooling:

[0044] Time-series coding: For users' time-series physiological data, such as continuous blood glucose monitoring value sequences. A Transformer encoder is used to extract temporal-dependent features. First, positional encoding is performed, and then temporal features are obtained through a multi-head self-attention mechanism and a feedforward network. .

[0045] Multimodal fusion: Vectorizing user health profiles Subgraph embedding Time series characteristics The fusion process is performed using a cross-modal interaction network based on an attention mechanism: First, the importance weight of each modality to the final task is calculated.

[0046] in For a certain mode of feature vector, and These are learnable parameters. Then, a weighted fusion is performed:

[0047] Alternatively, tensor fusion can be used: ,in For the outer product, This is a vectorized operation, but it involves a large amount of computation. This embodiment uses the former.

[0048] Output layer: Feature fusion The system outputs multi-dimensional assessment results of inflammation status through a fully connected network.

[0049] Quantifying the inflammatory burden index Activated via sigmoid:

[0050] in and These are the learnable weight vector and the bias term, respectively. This index reflects the level of systemic inflammation; the higher the value, the more severe the inflammation.

[0051] Key pro-inflammatory pathway activity score ,for The activity probabilities of major pro-inflammatory pathways (such as NF-κB, JAK-STAT, and COX-2) were normalized using softmax:

[0052] in The total number of pro-inflammatory pathways, and The first The learnable weight vectors and bias terms of each pathway are normalized using softmax to obtain the activity probability of each pathway. Personalized nutrient intervention targets: Output a set of nutrients and their target adjustment amounts, such as increasing omega-3 intake by X grams / day and reducing saturated fat intake by Y grams / day. This can be achieved through a regression model:

[0053] in This is the weight matrix. For bias vectors, express The change in the recommended daily intake of a nutrient relative to baseline.

[0054] Furthermore, the output layer can output one or more of the following indices—systemic inflammation, gut inflammation, and neuroinflammation—through multiple independent fully connected branches to provide a more granular assessment of inflammation. For example, the systemic inflammation index can be obtained through sigmoid or linear regression, while the gut and neuroinflammation indices can be calculated based on weighted sums of specific gut-related or neurological-related features. The specific implementation is similar to the output of other indices in this embodiment and will not be elaborated further here.

[0055] The evaluation engine is pre-trained and fine-tuned. During the pre-training phase, mask attribute reconstruction and relation prediction tasks are performed on the graph. During the fine-tuning phase, supervised learning is performed using a user dataset containing the detection results of explicit inflammatory markers (such as hs-CRP and IL-6). The loss function is mean squared error or cross-entropy.

[0056] 4. Multi-constraint, multi-objective recipe generation and optimization engine The engine generates Pareto-optimal candidate recipes based on the inflammatory load index, pro-inflammatory pathway activity, and nutrient intervention targets output by the evaluation engine, combined with a dynamic anti-inflammatory knowledge graph.

[0057] Decision variables: Let there be a total of A variety of ingredients are available, each ingredient The portion size variable is The unit is grams. This represents the maximum recommended daily intake. The diet is then determined by a vector. express.

[0058] Constraints: Macronutrient ratio constraints: Carbohydrate energy ratio Protein energy ratio The proportion of energy supplied by fat Must be within the target range:

[0059] in , For ingredients Carbohydrate content per gram For the total energy, of which For ingredients Protein content per gram For ingredients Fat content per gram.

[0060] Micronutrient limits: For example, sodium intake should not exceed 2300mg. ,in For ingredients Sodium content per gram.

[0061] User taste restrictions: Excluding certain ingredients through hard constraints, i.e., prohibiting certain ingredients. ,make .

[0062] Ingredient availability: Based on the user's geographical location and the season, certain ingredients are limited to 0.

[0063] Cooking time limit: Total cooking time Not exceeding the user-defined threshold .

[0064] Objective function: Maximize overall anti-inflammatory score:

[0065] in For ingredients Dietary inflammation index score (positive value indicates pro-inflammatory, negative value indicates anti-inflammatory). The relevance weight of this food ingredient to a user-specific pro-inflammatory pathway in the knowledge graph is calculated as follows: , For ingredients With pathway The strength of the association (obtained from the knowledge graph).

[0066] Maximize the match between taste preferences:

[0067] in The vector representation of a user's historical dietary preferences is extracted from dietary logs using a collaborative filtering method. Cosine similarity is used for similarity.

[0068] Minimize the risk of food-drug interactions:

[0069] in For ingredients The potential interaction risk score with the user's current medication list is obtained from the "food ingredient-drug interaction" relationship in the knowledge graph.

[0070] Therefore, the multi-objective optimization problem can be expressed as:

[0071] Optimization Algorithm: A hybrid intelligent optimization algorithm combining constraint satisfaction and evolutionary computation is employed. First, a set of feasible solutions satisfying the hard constraints is quickly generated as the initial population using a constraint satisfaction algorithm (such as backtracking search). Then, the non-dominated sorting genetic algorithm NSGA-II with an elitist strategy is used for iterative evolution. The algorithm steps are as follows: Initialize population Size is .

[0072] For each individual Calculate its objective function value ( ).

[0073] Perform non-dominated ranking, assigning individuals to different frontiers. ,in It is a non-dominated solution set.

[0074] Calculate the crowding distance for each individual This is used to maintain population diversity. For the first For each target, individuals are sorted according to that target value. The distance between boundary individuals is set to infinity, and the distance between intermediate individuals is the difference between the target values ​​of their adjacent individuals.

[0075] Offspring populations are generated through tournament selection, simulated binary crossover, and polynomial mutation. Size is .

[0076] merge and get Size is ,right Perform non-dominated sorting and crowding calculation, and select the top Individuals form the next generation .

[0077] Repeat steps 2-6 until the maximum number of iterations is reached, and finally obtain the Pareto front. All solutions on the above, i.e., a diverse set of candidate recipes.

[0078] 5. Interactive Feedback and Model Self-Evolution Module This module displays candidate recipes through a visual interface and collects user feedback. Users can select a recipe from the Pareto front and execute it, with the system recording their selection preferences. During execution, the wearable device continuously collects real-time physiological data from the user, such as continuous blood glucose monitoring and heart rate variability; users can also submit subjective feedback through the application, such as feelings of satiety, discomfort, and compliance. This feedback data is stored in a feedback database.

[0079] Use feedback data to fine-tune and optimize the core model online: For the multimodal deep learning evaluation engine, the proximal strategy in reinforcement learning is used to optimize the PPO algorithm. User feedback (such as reduced blood sugar fluctuations and improved subjective feelings) is used as a reward signal to update network parameters to more accurately predict the inflammatory state.

[0080] For score prediction models (such as anti-inflammatory score weights) in multi-objective optimization algorithms, incremental learning techniques, such as online stochastic gradient descent, are used to update the edge weights in the knowledge graph based on new food-inflammation association data.

[0081] Specifically, for updating neural network parameters, the loss function is defined as the difference between the prediction and the feedback, and the mini-batch gradient descent method is used:

[0082] in For model parameters, For learning rate, This represents the mean squared error or cross-entropy loss.

[0083] Through this closed-loop feedback mechanism, the system can continuously adapt to individual dynamic changes and achieve self-evolution.

[0084] Example 2: Personalized Anti-inflammatory Diet Recommendation Method This embodiment describes a specific recommended method flow based on the above system, including the following steps S1 to S7.

[0085] Step S1: Data Acquisition and Preprocessing Personal data is collected from wearable devices, electronic health records, user terminals, and gene sequencing reports. The collected multi-source data is cleaned, normalized, discretized, and feature-engineered to construct a user health profile vector. The specific processing method is the same as that of the multi-source data acquisition and preprocessing module in Example 1.

[0086] Step S2: Multimodal Deep Learning Evaluation Vector user health profile The input is fed into a pre-trained multimodal deep learning evaluation engine. This engine first extracts a user-related subgraph from a dynamic anti-inflammatory knowledge graph. Subgraph embeddings are obtained through graph attention networks. Simultaneously, temporal features are obtained by encoding temporal physiological data using Transformer. Then, the three factors are fused through an attention mechanism to obtain a comprehensive feature. Finally, the inflammatory burden index is output through a fully connected layer. Key pro-inflammatory pathway activity score and nutrient intervention targets It can also optionally output one or more of the following: systemic inflammation index, intestinal inflammation index, and neuroinflammatory index. This step provides a personalized target for subsequent optimization.

[0087] Step S3: Construct a multi-objective optimization problem Based on the evaluation results, an optimization problem for recipe generation is defined. The decision variables are the foods selected from the ingredient database and their portion sizes. The constraints include hard and soft constraints: hard constraints, such as allergen exclusion and religious taboos, must be met; soft constraints, such as food costs not exceeding the budget and cooking time not exceeding a threshold, are introduced into the objective function through penalty terms. The optimization objective is:

[0088] in For anti-inflammatory score, For taste matching, This pertains to the risk of food-drug interactions. The specific calculation formulas for each objective function are given in Example 1. Violations of soft constraints are penalized through a penalty factor. Introducing an objective function, for example, the actual optimization objective becomes:

[0089] in, The total cost of ingredients for the recipe. Budget cap set for users As a penalty factor, violations of soft constraints are introduced into the objective function through penalty terms.

[0090] Step S4: Solve using a hybrid intelligent optimization algorithm A hybrid algorithm combining constraint satisfaction and evolutionary computation is employed to solve the aforementioned multi-objective optimization problem. First, a constraint satisfaction algorithm (such as backtracking search) is used to generate feasible solutions that satisfy the hard constraints as an initial population. Then, a non-dominated sorting genetic algorithm with an elitist strategy is used for iterative evolution. By performing a non-dominated sort on each individual in the population, multiple frontiers are obtained.

[0091] Calculate the distance of congestion.

[0092] Parents are selected through a tournament, and offspring are generated through simulated binary crossover and polynomial mutation.

[0093] Merge parent and child generations, and select the first generation. Individuals form a new generation.

[0094] Repeat until convergence, and you will get the Pareto optimal solution set, which is a diverse set of candidate recipes.

[0095] Step S5: Filtering and Recommendation Based on preset rules or user historical preferences, the final recommended recipe is selected from the Pareto optimal candidate recipes. For example, if the user's historical preferences show a greater emphasis on anti-inflammatory effects, then... The highest possible solution; if the user has repeatedly selected recipes with low cooking times, then a preference-weighted selection is made. The final recipe is presented visually through a user interface (such as a mobile app), including an ingredient list, cooking steps, nutritional analysis, etc.

[0096] Step S6: Feedback Tracking After a user follows the recommended recipe, the system continuously tracks their physiological and subjective feedback. Physiological feedback is collected in real time via wearable devices, such as continuous blood glucose monitoring data and heart rate variability data; subjective feedback is collected through in-app questionnaires, such as satiety scores, discomfort records, and adherence. All feedback data is stored in a feedback database.

[0097] Step S7: Model Update The evaluation engine and optimization model are regularly updated using new data from the feedback database. For the evaluation engine, incremental learning or online reinforcement learning algorithms are used to fine-tune parameters; for relation weights in the knowledge graph, Bayesian updates or weighted averages based on evidence strength are used. For example, for a food-inflammation relationship, new literature evidence causes its weight to change from... Updated to:

[0098] in As a cumulative amount of historical evidence, This reflects the impact of new literature. The updated model will take effect in the next recommendation, enabling the system to evolve itself.

[0099] Through the above steps, this invention achieves highly personalized, dynamically adaptive, and scientifically accurate anti-inflammatory diet recommendations, effectively solving the problem of traditional methods ignoring individual differences and dynamic changes.

Claims

1. A personalized anti-inflammatory diet recommendation system based on artificial intelligence, characterized in that, include: The multi-source data acquisition and preprocessing module is configured to collect users' time-series physiological data, static demographic data, clinical biochemical indicator data, metagenomic data, and dietary behavior log data from wearable devices, electronic health records, user terminals, and gene sequencing reports; and to clean, normalize, discretize, and perform feature engineering on the collected multi-source data to construct a structured user health profile vector. The dynamic anti-inflammatory knowledge graph module is configured to store and manage multi-level entities and their relationships in the form of a graph database. The entities include food components, nutrients, bioactive compounds, metabolic pathways, cytokines, and disease phenotypes. The relationships include activation, inhibition, transformation, inclusion, and association. The entity and relationship weights in the knowledge graph are dynamically updated based on medical literature using natural language processing technology. A multimodal deep learning evaluation engine based on attention mechanism is configured to receive the user health profile vector and fuse it with the user-related subgraph embedding vector extracted from the dynamic anti-inflammatory knowledge graph module; using a deep neural network model including attention mechanism and Transformer architecture, it outputs the user's quantitative inflammatory load index, key pro-inflammatory pathway activity score and personalized nutrient intervention targets. The multi-constraint, multi-objective recipe generation and optimization engine is configured to receive the quantitative inflammatory load index, the activity score of key pro-inflammatory pathways, and personalized nutrient intervention targets. Based on the dynamic anti-inflammatory knowledge graph module, it uses a multi-objective optimization algorithm that combines constraint satisfaction algorithm and evolutionary computation to generate a set of Pareto optimal candidate anti-inflammatory diet recipes under the constraints of macronutrient ratio, micronutrient upper limit, user taste restrictions, ingredient availability, and cooking time limit. The interactive feedback and model self-evolution module is configured to display the candidate anti-inflammatory diet recipes in a visual interface and collect real-time physiological feedback data and subjective experience feedback from users after executing the recipes. Using the feedback data, the model parameters of the multimodal deep learning evaluation engine and the multi-constraint multi-objective recipe generation and optimization engine are fine-tuned and optimized online through reinforcement learning or incremental learning techniques.

2. The personalized anti-inflammatory diet recommendation system based on artificial intelligence according to claim 1, characterized in that, The feature engineering processing of time-series physiological data in the multi-source data acquisition and preprocessing module includes extracting frequency and time domain features of blood glucose fluctuation amplitude, area under the curve, peak time, and nocturnal heart rate variability; the processing of metagenomic data includes calculating microbial α diversity, β diversity, and the relative abundance of specific anti-inflammatory probiotics.

3. The personalized anti-inflammatory diet recommendation system based on artificial intelligence according to claim 1, characterized in that, The relationships between entities in the dynamic anti-inflammatory knowledge graph module include the relationship between food and inflammation score quantified by the dietary inflammation index system, and the relationship between food components and cytokine regulation mined by retrieving medical literature; the knowledge graph is updated by periodically crawling medical literature databases and using a pre-trained biomedical relationship extraction model.

4. The personalized anti-inflammatory diet recommendation system based on artificial intelligence according to claim 1, characterized in that, The multimodal deep learning evaluation engine further includes: Temporal encoders, based on long short-term memory networks or Transformer models, are used to encode users' temporal physiological data and extract time-dependent features; A graph neural network encoder, based on a graph attention network or GraphSAGE, is used to extract embedded representations of user-specific subgraphs from the dynamic anti-inflammatory knowledge graph module; The fusion layer employs tensor fusion or an attention-based cross-modal interaction network to deeply fuse the time-dependent features, user health profile vectors, and graph embedding representations. The output layer outputs multi-dimensional inflammatory status assessment results through a fully connected network. These multi-dimensional inflammatory status assessment results include the quantitative inflammatory burden index, key pro-inflammatory pathway activity scores, and personalized nutrient intervention targets, as well as one or more selected from systemic inflammatory indices, intestinal inflammatory indices, and neuroinflammatory indices.

5. The personalized anti-inflammatory diet recommendation system based on artificial intelligence according to claim 4, characterized in that, The multimodal deep learning evaluation engine is a pre-trained and fine-tuned model. Its pre-training tasks include mask attribute reconstruction and relation prediction, and its fine-tuning tasks are based on supervised learning on a user dataset containing the detection results of explicit inflammatory markers.

6. The personalized anti-inflammatory diet recommendation system based on artificial intelligence according to claim 1, characterized in that, The multi-objective optimization algorithm used in the multi-constraint, multi-objective recipe generation and optimization engine includes the following optimization objective functions: The first objective function is to maximize the overall anti-inflammatory score of the recommended recipes, which is calculated by weighting the anti-inflammatory index of each food in the recipe and its correlation with the user's specific pro-inflammatory pathways in the knowledge graph. The second objective function is to maximize the match between the recipe and the user's long-term dietary preferences. This match is calculated based on collaborative filtering of the user's historical diet logs and taste preference data. The third objective function minimizes the potential food-drug interaction risk in the recipe, which is assessed based on the user's current medication list and the food-drug interaction relationships in the knowledge graph.

7. A personalized anti-inflammatory diet recommendation method based on artificial intelligence, applied to the system described in any one of claims 1 to 6, characterized in that, Includes the following steps: Step S1: Collect users' personal data from multiple heterogeneous data sources, perform data cleaning and feature engineering, and construct user health profiles. Step S2: Input the user's health profile into a pre-trained multimodal deep learning evaluation engine. This engine combines a dynamically updated anti-inflammatory knowledge graph to analyze and output a personalized inflammatory status evaluation report for the user. The report includes a quantitative inflammatory load index, a key pro-inflammatory pathway activity score, and nutrient intervention targets. Step S3: Construct a multi-objective optimization problem for recipe generation, where the decision variables are the foods selected from the ingredient database and their quantities, the constraints include the user's dietary restrictions and recommended nutrient intake range, and the optimization objectives are to maximize anti-inflammatory benefits, maximize taste preference satisfaction, and minimize the risk of food-drug interactions. Step S4: The multi-objective optimization problem is solved using a hybrid intelligent optimization algorithm that combines constraint satisfaction and evolutionary algorithm, generating a diverse set of Pareto optimal candidate recipes. Step S5: Based on preset rules or user historical selection preferences, select the final recommended recipe from the Pareto optimal candidate recipes and display it through the user interface; Step S6: Continuously track the user's implementation of the recommended recipes and the resulting physiological and psychological feedback, and store the feedback data in the feedback database; Step S7: Using the new data in the feedback database, the rating prediction model in the multimodal deep learning evaluation engine and the multi-objective optimization algorithm is updated through online learning or incremental training, thereby realizing the self-evolution of the recommendation system.

8. The method for recommending personalized anti-inflammatory diet recipes based on artificial intelligence according to claim 7, characterized in that, The constraints constructed in step S3 include hard constraints and soft constraints: hard constraints are conditions that must be met, including allergen exclusion and religious taboos; soft constraints include food costs not exceeding the budget and cooking time not exceeding the threshold. Violations of soft constraints are introduced into the objective function through penalty terms.

9. The method for recommending personalized anti-inflammatory diet recipes based on artificial intelligence according to claim 7, characterized in that, The hybrid intelligent optimization algorithm in step S4 is as follows: First, a high-quality initial recipe population is generated by using the constraint satisfaction algorithm pruning search space; then, a non-dominated sorting genetic algorithm with an elitist strategy is used for iterative evolution to approximate the Pareto front.

Images