Digital diet management method and system for children with obesity

Abstract

The invention relates to the technical field of diet management, and discloses a digital diet management method and system for children obesity, and the method comprises the steps: data collection: collecting physiological data in real time through wearable equipment, including the blood glucose, heart rate variability and body fat rate of children, and collecting the diet data of children through a mobile phone application, the food type and component are automatically identified by adopting an image identification technology. The system comprises a data acquisition module, a modeling module, a reinforcement learning module, a causal intervention module, a federal learning module and a feedback module. The real-time performance and accuracy of personalized diet intervention are improved through federal learning and global optimization, and efficient learning is realized while data privacy is guaranteed; combining a dynamic graph neural network and causal reasoning to realize dynamic optimization of metabolism modeling and an intervention strategy; and through the virtual nutritionist and the gamification task, the interactivity and the participation degree of children are enhanced, the formation of healthy diet habits is promoted, and the personalized and intelligent health management level is remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of diet management, in particular to a digital diet management method and system for child obesity. BACKGROUND

[0002] Currently, the problem of child obesity is becoming increasingly serious, and many health management programs attempt to solve this problem through diet control. Most existing child obesity management systems rely on fixed diet recommendations, which are usually based on general nutritional standards and do not fully consider the individual differences of children. Therefore, the diet management solutions in the prior art have obvious deficiencies in personalization and real-time adjustment.

[0003] Traditional child obesity management methods often intervene through static diet plans or weight loss menus, but such methods often fail to respond to changes in children's physiological state and emotional fluctuations in a timely manner. For example, children's blood glucose concentration, body fat rate and other indicators may change over time, and existing technologies do not provide the ability to track and adjust strategies in real time. As a result, parents and children have difficulty obtaining more accurate diet guidance in a timely manner, resulting in ineffective interventions and even inability to adhere to them in the long term.

[0004] In addition, most existing systems lack effective data privacy protection measures. Since traditional diet management systems require centralized storage of large amounts of user data, this may lead to privacy leakage problems. With the continuous strengthening of data privacy laws, how to efficiently optimize diet strategies without exposing personal information has become a difficult problem to be solved. However, many current solutions lack effective technical means in this regard and cannot truly achieve personalized and compliant diet management.

[0005] In addition, the child diet management in the prior art is often dull and lacks appeal. Children have difficulty adhering to simple diet adjustments in the long term, and existing systems fail to effectively design interactivity and interest to encourage children to actively participate. Therefore, the existing methods have limited effectiveness in attracting children's sustained attention and behavior change. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a digital diet management method and system for child obesity, which solves the problems of insufficient personalization, data privacy protection and low user participation in child obesity management.

[0007] To achieve the above purpose, the present application realizes the following technical scheme: a digital diet management method for child obesity, comprising the following steps:

[0008] Data collection: physiological data is collected in real time through wearable devices, including the child's blood glucose, heart rate variability and body fat rate, and diet data of the child is collected through a mobile application, using image recognition technology to automatically identify food types and quantities;

[0009] Data preprocessing: clean and standardize the collected multi-modal data, generate a unified time series data matrix;

[0010] Metabolic kinetics modeling: based on the collected physiological data and dietary data, build a child metabolic kinetics model, use differential equations to describe the metabolism of blood glucose, insulin, fat, and dynamically calibrate the model parameters through variational autoencoder;

[0011] Hierarchical reinforcement learning decision generation: use hierarchical reinforcement learning algorithm to generate personalized short-term and long-term dietary adjustment strategies for children, including real-time dietary structure adjustment and long-term health goal optimization;

[0012] Causal reasoning and dynamic intervention: based on metabolic data, dietary habits and emotional changes of children, use causal reasoning to analyze the relationship between diet, metabolism and behavior, and generate dynamic intervention programs;

[0013] Federal learning and global optimization: based on federal learning, aggregate and update local training models;

[0014] Personalized feedback output: according to the generated personalized dietary intervention strategy, provide dietary report and health advice to parents and children in real time, and guide children to develop healthy dietary behavior through virtual nutritionist and gamified tasks.

[0015] Preferably, the dietary data collection step comprises:

[0016] Use the user's mobile phone application to take photos of the dining table, automatically identify the food categories and their quantities through image recognition technology, and generate a preliminary dietary log;

[0017] Compare the image recognition results with the manually input dietary data. If the recognition results are inconsistent with the manual input, the system will correct the food recognition results through hierarchical reinforcement learning model;

[0018] Real-time monitoring of children's physiological state through wearable devices, collecting blood glucose data, heart rate variability, body fat rate data;

[0019] Time synchronization of the collected physiological data, dietary data and environmental data to generate a unified time series data matrix.

[0020] Preferably, the metabolic kinetics modeling step comprises:

[0021] Based on the collected physiological data and dietary data, build a child metabolic kinetics model, use differential equations to describe the metabolism of blood glucose G, insulin I, fat L and fatty acid concentration F;

[0022] The parameters in the metabolic model are dynamically estimated by a variational autoencoder, and the parameters are updated using the variational inference formula.

[0023] The metabolic model is trained using the variational lower bound loss function, and high-precision metabolic parameter estimation is achieved by minimizing the ELBO loss function, and the parameters in the metabolic model are updated.

[0024] Preferably, the mathematical description of the metabolic kinetics model is:

[0025]

[0026] Where G is the blood glucose concentration, I is the insulin concentration, L is the body fat rate, F is the fatty acid concentration, D(t) is the dietary input, k1, k2, k3, k4, k5, k6, k7, k8, k9, k 10 are parameters of the metabolic model, represents the rate of change of blood glucose G with respect to time t, represents the rate of change of insulin I with respect to time t, represents the rate of change of body fat L with respect to time t, represents the rate of change of fatty acid F with respect to time t.

[0027] The variational inference formula is as follows:

[0028]

[0029] Where q φ (z|x) is the approximate posterior distribution of the latent variable z given the input data x, which is approximated by a parameterized distribution; represents that this distribution is a Gaussian distribution, where:

[0030] μ φ (x) is the mean of the latent variable z given the input x, which is obtained by processing the input data x through the neural network parameters φ; σ φ (x) is the standard deviation of the latent variable z given the input x, which is also obtained by processing the input data x through the neural network parameters φ; z represents the latent variable, which is usually used to represent the potential structure or abstract features of the input data; x represents the input data or observed data, such as images, text or other types of data;

[0031] p θ (k|z) represents the conditional probability of generating observed data k given the latent variable z, and the generation process is modeled by the decoder. This conditional probability is determined by the neural network parameters θ of the decoder; Decoder(z; θ) represents the neural network model of the decoder part, which generates data by inputting the latent variable z; k represents the generated data.

[0032] Preferably, the hierarchical reinforcement learning decision generation step comprises:

[0033] The hierarchical reinforcement learning model is used to generate a personalized diet adjustment strategy for children, which is divided into two strategy modules: a bottom layer and a top layer;

[0034] The bottom layer decision module receives the child's instant physiological state data, selects a short-term diet adjustment strategy, and uses the reward function R short to calculate the effect of instant diet adjustment;

[0035] The top layer decision module receives the child's long-term health goal data, selects a long-term diet pattern strategy, and uses the reward function R long to optimize long-term health goals;

[0036] Through the reinforcement learning model, information is exchanged between the bottom layer and top layer strategy modules, and the diet structure is dynamically adjusted to optimize the child's short-term and long-term health status.

[0037] Preferably, the reward function R short is calculated as follows:

[0038] R short = α1·Stability(G(t)) + α2·Satiety Index - α3·Craving Score ;

[0039] Wherein, α1, α2, and α3 are adjustment factors, Stability(G(t)) represents the stability of blood glucose concentration G at time t, Satiety Index represents the satiety index, and Craving Score represents the appetite score;

[0040] The reward function R long is calculated as follows

[0041] R long = β1·ΔBMI + β2·ΔInsulin Sensitivity - β3·Rebound Risk ;

[0042] Wherein, β1, β2, and β3 are adjustment factors, ΔBMI represents the change in body mass index, ΔInsulin Sensitivity represents the change in insulin sensitivity, and Rebound Risk represents the risk of weight rebound.

[0043] Preferably, the causal reasoning and dynamic intervention step comprises:

[0044] Constructing a multi-modal causal graph using a Bayesian causal discovery algorithm Identify the causal relationship between diet, metabolism and behavior;

[0045] The edge weight in the causal graph is dynamically updated by using a time-aware graph attention network, and the structure of the causal graph is adjusted according to the diet, metabolic state and emotional data of the child;

[0046] The influence chain between diet, metabolism and behavior is analyzed using causal reasoning, and the intervention effect of each node is calculated;

[0047] Based on the results of causal reasoning, a personalized diet adjustment plan and psychological intervention plan are generated through a dynamic optimization algorithm, and real-time intervention suggestions are pushed;

[0048] In each intervention period, the causal reasoning model is updated, and the intervention strategy is adjusted in real time.

[0049] Preferably, the federated learning and global optimization steps include:

[0050] Each user device locally trains a metabolic model and a reinforcement learning strategy;

[0051] The gradients of the local models are aggregated using a differential equation parameter federated aggregation algorithm to update the parameters of the global metabolic model;

[0052] Based on the federated learning algorithm, the training results of each terminal device are converged to optimize the global model and achieve efficient personalized learning.

[0053] The present application also provides a digital diet management system for children with obesity, which is applied to the above-mentioned digital diet management method for children with obesity, comprising:

[0054] The data acquisition module is used for collecting physiological data, diet data and environmental data of children, and preprocessing these data to generate a unified time series data matrix;

[0055] The metabolic kinetics modeling module is used for generating a metabolic kinetics model of the child according to the data, and dynamically calibrating the metabolic parameters in the model;

[0056] The reinforcement learning decision module is used for generating personalized short-term and long-term diet adjustment strategies according to the metabolic model output;

[0057] The causal reasoning and intervention module is used for analyzing the causal relationship between diet, metabolism and behavior, and generating a dynamic intervention plan according to the analysis results;

[0058] The federated learning module is used for aggregating the model updates of the local devices to the global model;

[0059] Feedback output module: for providing personalized health reports and dietary suggestions to parents and children in real time according to the dietary adjustment strategy, and guiding children to develop healthy dietary behaviors through virtual nutritionists and gamified tasks.

[0060] Preferably, the causal reasoning and intervention module comprises:

[0061] A causal graph construction module for identifying causal relationships between diet, metabolism and behavior;

[0062] A graph attention network module for dynamically updating edge weights of causal relationships and generating intervention plans.

[0063] The present application provides a kind of digital diet management method and system for child obesity. With the following beneficial effects:

[0064] 1, the present application is by adopting federal learning and global optimization technical scheme, by aggregating the local training result of different equipment, effectively improve the accuracy and real-time of individualized dietary intervention strategy, compared with traditional centralized data processing mode, the present application avoids the problem of data privacy disclosure, reduces the dependence on user data, while maintaining efficient individualized learning ability.

[0065] 2, the present application is by dynamically adjusting the edge weight of graph neural network, and combining causal reasoning to calculate intervention effect, realizes the precise optimization of metabolism modeling and dietary behavior strategy, compared with the intervention method based on fixed model in the prior art, the present application can automatically adjust intervention strategy according to the physical state and behavior feedback of user in real time, significantly improve the individualization and flexibility of health management.

[0066] 3, the present application is by virtual nutritionist and gamified task combination, provides a more interactive and attractive health diet guide mode, which not only increases the participation of children, but also promotes children to develop healthy eating habits through instant feedback and task motivation, compared with traditional simple dietary suggestions, the present application solves the problem that children are unwilling to accept healthy diet guidance through situational and personalized task design. BRIEF DESCRIPTION OF DRAWINGS

[0067] Figure 1 The flowchart of the method of the present application;

[0068] Figure 2 The framework diagram of the system of the present application. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be clearly and completely described with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present application.

[0070] Please refer to the drawings attached Figure 1 The embodiment of the present application provides a digital diet management method for child obesity, comprising the following steps:

[0071] S1, data acquisition: real-time acquisition of physiological data through wearable devices, including blood glucose, heart rate variability and body fat rate of children, and acquisition of children's diet data through mobile phone applications, automatic recognition of food types and quantities by image recognition technology;

[0072] Firstly, the physiological data of the child is collected through the wearable device, including blood glucose, heart rate variability and body fat rate. The wearable device includes a blood glucose sensor, a heart rate monitor and a body fat rate measuring device, which can obtain the physiological parameters of the child in real time and upload the data to the cloud or local storage through wireless transmission. By time synchronization processing of the data, the consistency of the data from different sources in time is ensured, and a unified time series data matrix is formed. This data matrix is used for subsequent metabolic dynamics modeling and personalized intervention generation.

[0073] Secondly, the diet data of the child is collected by the mobile phone application through image recognition technology. After the parents or children take photos of the dishes, the mobile phone application will automatically identify the food types and quantities in the photos. The image recognition technology is based on a deep learning model, which can automatically determine the food types and quantities in the dishes according to the appearance, color, shape and other characteristics of the food.

[0074] Taking photos of the dishes and generating a preliminary diet log:

[0075] After the user takes photos of the dishes through the mobile phone application, the system uses image recognition technology to identify the food types and quantities. The image recognition technology uses convolutional neural networks (CNN) and other deep learning algorithms to classify the various types of food in the dishes, automatically calculate the quantities of each type of food (such as the weight or volume of each type of food), and generate a preliminary diet log.

[0076] Specifically, the image recognition process can be achieved through the following steps:

[0077] Use image recognition algorithms to extract features of the food (such as color, texture, shape, etc.);

[0078] Classify the food based on a pre-trained deep neural network model;

[0079] The system calculates the portion of each food, usually by the proportion of pixels in the image or known food reference data.

[0080] Image recognition results are compared with manual input data:

[0081] To ensure the accuracy of the data, the system compares the image recognition results with the user's manually input diet data. The user can manually input the type and portion of food, and the system will correct the accuracy by calculating the difference between the image recognition results and the manually input data. If the recognition results are inconsistent with the manually input diet data, the system will use a hierarchical reinforcement learning model to correct the food recognition results. The role of the hierarchical reinforcement learning model is to adjust the accuracy of image recognition through real-time feedback. In each recognition error, the system will adjust the parameters of the recognition algorithm according to the historical data and current data, so as to improve the accuracy of recognition.

[0082] Specifically, the workflow of the hierarchical reinforcement learning model is as follows:

[0083] Bottom module: Compare the image recognition results with the user input, and adjust the image recognition algorithm parameters if there is a large error;

[0084] Top module: According to the correction result of the error, generate a new food recognition strategy, and feed back the corrected strategy to the bottom module for optimization of the recognition process.

[0085] While collecting diet data in real time, the system also monitors the physiological state of children in real time through wearable devices, including blood glucose level, heart rate variability and body fat rate. Real-time collection of these physiological data will be time-synchronized with diet data to ensure the time sequence consistency between diet behavior and physiological response. For this purpose, all data will be marked with time stamps and converted into standard time series data matrix.

[0086] After data synchronization, the system integrates physiological data, diet data and environmental data from different sources into a unified time series data matrix. This matrix contains the following key contents:

[0087] Timestamp: Mark the time point of data collection to ensure the correspondence of various data in time;

[0088] Blood glucose, heart rate variability and body fat rate data: collected by wearable devices, reflecting the physiological state of children in real time;

[0089] Diet data: generated by image recognition and manual input, including the type and portion of food on the children's plate;

[0090] Environmental data: including environmental temperature, humidity and other factors affecting children's metabolism.

[0091] The data matrix will be used as input for the metabolic kinetic model to generate subsequent health analysis and dietary adjustment strategies.

[0092] The image recognition model extracts food features in the image through a convolutional neural network (CNN) and predicts the food category and components. Let the input image be X, and the output be the food category C and the component Q. Their relationship can be represented as:

[0093] C = f CNN (X), Q = g CNN (X), where f CNN (X) is the prediction function of the food category in image recognition, and g CNN (X) is the prediction function of the component.

[0094] The hierarchical reinforcement learning model adjusts the results of food recognition based on the interaction between the bottom module and the top module. The specific optimization process updates the model parameters based on the feedback mechanism. Let the current food recognition result be R current , the target correction result be R target , and the correction strategy ΔR be optimized through the reinforcement learning algorithm:

[0095] where θ is the parameter of the model, is the expected reward obtained through the current strategy, and the optimization result is continuously adjusted through the feedback mechanism to achieve accurate correction of the recognition result.

[0096] Data synchronization and time series matrix generation: all collected data (blood glucose, heart rate variability, body fat rate, dietary data) are synchronized through timestamps to generate a unified time series data matrix, which is used for metabolic kinetic modeling in subsequent steps.

[0097] The present application can realize real-time collection of physiological data and dietary data of children, and ensure the accuracy and synchronization of data through various data processing technologies. The hierarchical reinforcement learning algorithm is used to correct the food recognition result, so that the system can dynamically adjust the dietary data collection strategy according to the specific situation of the children, improve the accuracy and individualization level of the system. In addition, the time series data matrix provides a solid data foundation for subsequent metabolic modeling and personalized health intervention, which can effectively support the digital management and intervention of childhood obesity.

[0098] S2, data preprocessing: cleaning and standardizing the collected multi-modal data to generate a unified time series data matrix;

[0099] In the embodiment, the data preprocessing step is used to clean and standardize the collected multi-modal data to generate a unified time series data matrix, providing a continuous and standardized data foundation for the construction of the metabolic kinetic model.

[0100] The data acquisition system comprises a physiological data acquisition unit and a dietary data acquisition unit, the physiological data acquisition unit is connected with a wearable sensor assembly, and is used for acquiring multi-dimensional physiological parameters of a user, such as blood glucose, body fat rate, heart rate variation rate and the like.

[0101] The dietary data acquisition unit comprises a camera acquisition part connected to a mobile terminal image recognition module, which can perform multi-category segmentation on a meal image, and identify food ingredients and intake quality.

[0102] The data cleaning module comprises an abnormality detection part and a missing value filling part, the abnormality detection part identifies and removes error data points based on statistical distribution and physiological reasonable threshold, and the missing value filling part reconstructs missing values in a time slice by a sliding window method or a local linear interpolation method.

[0103] The standardization module comprises a normalization processing part and a time synchronization part, the normalization processing part performs dimension normalization on different dimensional data, and adopts a z-score standardization method, so that each dimensional data has a uniform magnitude.

[0104] The standardization processing formula is as follows:

[0105] Wherein, X represents an original observation value, μ represents a historical sample mean, σ represents a sample standard deviation, and X norm represents normalized data.

[0106] The time synchronization part is used for multi-source data alignment, constructs a uniform time index set T={t1, t2, …, t n}, and interpolates and completes missing indexes in each data source timestamp set to generate a multi-dimensional homogeneous time sequence.

[0107] The output data matrix is in the form of:

[0108]

[0109] Wherein, n represents a time step number, m represents a category dimension of physiological or dietary data, x ij represents a data value of the i th time step and the j th feature.

[0110] The data preprocessing system further comprises a label generation module, the module is used for identifying a specific behavior event (such as eating time and exercise behavior), and embedding it into an additional dimension of a time sequence data matrix, which is used for model supervised training.

[0111] Through the above steps, multi-dimensional time series data with complete structure, continuous data, unified scale and consistent time sequence can be generated, providing high-quality input for subsequent metabolic modeling and parameter estimation, and effectively supporting stable convergence and improved generalization ability of the model.

[0112] The data preprocessing module and the model training module are connected through a communication bus. The generated data matrix can be directly input into the variational autoencoder model for inference and estimation of various parameters in the metabolic kinetics model, completing the closed-loop individual modeling process.

[0113] S3, metabolic kinetics modeling: based on the collected physiological data and dietary data, a child metabolic kinetics model is constructed, the metabolic processes of blood glucose, insulin and fat are described using differential equations, and the model parameters are dynamically calibrated by a variational autoencoder;

[0114] In this embodiment, the metabolic kinetics modeling step is the core processing link after data preprocessing and has relevance with the foregoing steps in terms of data flow and logical structure. The standardized multi-modal time series data matrix is input into the metabolic modeling module to describe the metabolic physiological response of children under specific dietary behavior intervention. This modeling process not only constructs a mathematical model reflecting the evolution of blood glucose, insulin, body fat rate and fatty acid concentration, but also further combines a deep generative model to dynamically estimate the model parameters to adapt to individual differences.

[0115] In this embodiment, first, a metabolic kinetics differential equation set based on physiological mechanisms is constructed, and ordinary differential form is used to simulate the evolution behavior of each variable over time.

[0116] Specifically, the metabolic model consists of the following differential equation system:

[0117]

[0118] where G is the blood glucose concentration, with units of mmol / L; I is the insulin concentration, with units of μU / mL; L is the body fat rate, representing the proportion of adipose tissue to body weight, which is a dimensionless number; F is the plasma fatty acid concentration, with units of mmol / L; D(t) is the dietary input function, representing the glucose intake rate at time t; k1, k2, k3, k4, k5, k6, k7, k8, k9, k 10 are rate constants of each metabolic reaction, which have positive values and reflect the intensity relationship in the metabolic pathway.

[0119] In general, equation (1) represents that blood glucose is jointly affected by degradation, interaction with fat, insulin regulation and external dietary input.

[0120] Equation (2) is used to describe the secretion mechanism of insulin under the stimulation of blood glucose and its inhibitory coupling relationship with fatty acids.

[0121] Equations (3) and (4) reflect the mutual conversion relationship and metabolic consumption between body fat and fatty acids, respectively.

[0122] As an option, the above equation set is solved by the Runge-Kutta fourth-order method to obtain the variable trajectory at discrete time points. To further enhance the individual expression ability of the model, a variational autoencoder is introduced to dynamically learn the rate parameters.

[0123] In a possible implementation, a variational autoencoder structure is constructed, and the encoder receives time series data input x and outputs the approximate posterior distribution of hidden variable z:

[0124]

[0125] where q φ (z|x) is the posterior approximate distribution of the hidden variable; μ φ (x) represents the mean of the latent variable generated by the input data x; σ φ (x) represents the standard deviation of the latent variable generated by the input data x; and φ is a set of learnable parameters of the encoder network.

[0126] The hidden variable z reflects the unobservable potential metabolic state or regulatory factor in the input data. Then, z is mapped to the model parameter space by the decoder:

[0127] k i = f i (z), i∈{1,…,10}; (6)

[0128] where f i (z) represents the neural mapping function of the i-th model parameter; the mapping function is composed of the decoder neural network, and the parameters are controlled by θ; and the output k i will replace the fixed constant in the original differential equation, forming a dynamic adjustment mechanism.

[0129] To optimize the encoder and decoder networks, the evidence lower bound (ELBO) is introduced as the objective function:

[0130]

[0131] In the above loss function:

[0132] p θ (x|z) is the conditional probability of reconstructing the observed data x given the hidden variable z; and p(z) is the prior distribution of the hidden variable, which is usually set to a standard normal distribution KL(·||·) is the Kullback-Leibler divergence, which measures the difference between the approximate posterior and the prior; the parameter θ is the set of trainable parameters of the decoder network.

[0133] Generally, the first term of the expected value is used to measure the model reconstruction ability, and the second term of the regular term limits the spatial distribution of the latent variable to prevent overfitting.

[0134] In this embodiment, the variational inference structure is unified with the differential modeling logic to form a unified modeling unit. The parameters output by the neural network are used to solve the differential equation, and in each training iteration, the parameters φ and θ are jointly optimized according to the gradient descent method.

[0135] In some embodiments, to avoid the problem of gradient disappearance, the hidden variable is represented as:

[0136]

[0137] where z is the hidden variable; μ φ (x) is the mean vector output by the encoder network, which is the center position of the input data x in the approximate posterior distribution; σ φ (x) is the standard deviation vector output by the encoder network, which represents the distribution uncertainty of the input data x in the latent space; ∈ is a noise item of the standard normal distribution, satisfying where I is the identity matrix.

[0138] This parameterization technique enables the backpropagation process to optimize the sampling results, which helps to improve the training stability.

[0139] As an implementation form, the metabolic model can also embed a control module to predict future metabolic trends using parameter sequences, generate dietary intervention suggestions, and realize a dynamic dietary adjustment mechanism.

[0140] In the above manner, a complete modeling technology chain from standardized data input, differential modeling, parameter estimation, training optimization to output control suggestions is constructed. The modeling structure has high time resolution, high variable coupling degree, adapts to individual physiological diversity, and meets the needs of clinical monitoring and intervention.

[0141] S4, hierarchical reinforcement learning decision generation: using hierarchical reinforcement learning algorithm to generate personalized short-term and long-term dietary adjustment strategies for children, including real-time dietary structure adjustment and long-term health goal optimization;

[0142] In this embodiment, the hierarchical reinforcement learning decision generation step is based on individual physiological state data output by the metabolic kinetics modeling module (S3) to establish a short-term and long-term dietary regulation strategy optimization system suitable for children using a hierarchical reinforcement learning algorithm. The system includes a real-time dietary structure adjustment mechanism and a long-term health goal adaptation mechanism, aiming to achieve strategy coordination and iterative optimization under multiple time scales through state-action interaction.

[0143] First, a double-layer strategy structure is established, including a top-layer strategy module and a bottom-layer strategy module. The bottom-layer strategy module receives immediate physiological state input variables, including blood glucose concentration G, insulin concentration I, body fat rate L, fatty acid concentration F, and their time change trends. The module generates immediate dietary behavior control signals and outputs adjustment actions such as carbohydrate intake adjustment, meal time optimization, or micronutrient ratio adjustment.

[0144] The immediate reward function of the bottom-layer strategy is defined as follows:

[0145] R short = α1·Stability(G(t)) + α2·Satiety Index - α3·Craving Score ;

[0146] Where α1, α2, and α3 are adjustment factors used to control the relative weight of each index on the immediate reward; Stability(G(t)) is a blood glucose stability measurement function, represented as: The smaller the rate of change of blood glucose concentration G at time t, the higher the stability.

[0147] Satiety Index is the satiety index, representing the subjective or predictive satiety intensity brought by the current dietary behavior, with a higher value indicating more effective satiety control; Craving Score represents the food craving score, derived from historical dietary preferences, neural model estimation, or subjective feeling quantification, reflecting the strength of current food desire.

[0148] The goal of the bottom-layer strategy is to maximize the following expected cumulative reward:

[0149] Where is the probability distribution of the bottom-layer strategy, θ is its parameter; γ is the time discount factor, 0 < γ < 1; T is the rolling execution time step number of the bottom-layer strategy.

[0150] A top-level policy module is then established to handle long-term health objective variables and formulate a macro-diet control template. The top-level policy input includes BMI trend, insulin sensitivity estimate, and body weight rebound risk prediction, generating a high-level diet control framework such as eating structure cycle, nutrition distribution pattern, and phased restriction path.

[0151] The top-level reward function is defined as follows:

[0152] R long = β1·ΔBMI + β2·ΔInsulin Sensitivity - β3·Rebound Risk ;

[0153] where β1, β2, and β3 are adjustment factors; ΔBMI represents the change in body mass index, defined as:

[0154] ΔBMI = BMI(t+T) - BMI(t);

[0155] ΔInsulin Sensitivity represents the change in insulin sensitivity, derived from the response estimate of variable I to the rate of change in blood glucose under unit insulin concentration, which can be characterized by the following function:

[0156]

[0157] The larger the function value, the higher the efficiency of unit insulin action;

[0158] Rebound Risk represents the body weight rebound risk estimate, calculated by a regression model constructed from historical body weight trajectory and diet behavior sequence.

[0159] The top-level policy optimization objective function is as follows:

[0160] where is the probability distribution of the top-level policy, φ is its policy parameter; K is the rolling optimization period of the top-level policy, generally much larger than the bottom-level period; the reward function R long can be calculated cumulatively within each top-level period and passed to the policy optimization network.

[0161] The hierarchical reinforcement learning system includes a policy generation part, a state evaluation part, a reward feedback part, and a policy update part, all of which are constructed by deep neural networks. The state evaluation part receives the prediction sequence G, I, L, F from the metabolic dynamics model and embeds it into the state vector. The policy generation part outputs the behavior distribution, and the policy update part performs backpropagation according to the actual environmental feedback.

[0162] In one implementation, the top-level policy outputs a nutritional structure constraint template to constrain the behavior selection space of the bottom-level policy. For example, it specifies upper and lower limits for protein intake, or a time window for sugar intake, and the bottom-level policy samples the optimal action within this range.

[0163] After the bottom-level strategy executes the action, the generated behavioral sequence and physiological response are fed back to the top-level module, forming a joint training loop. The information flow is:

[0164]

[0165]

[0166] Top-level behavior Decide on a diet template;

[0167] Top Reward R long Feedback is used for strategy optimization.

[0168] In some implementations, a soft actor-critic algorithm (SoftActor-Critic) is used for policy updates, enhancing exploration and avoiding local optima by adding a policy entropy regularization term. An experience replay mechanism is used during model training to maintain separate bottom-level and top-level trajectory pools, ensuring decoupling of the learning processes at each layer.

[0169] There is a strong coupling relationship between the hierarchical structure and metabolic modeling. The temporal dynamics of the physiological variables G, I, L, and F are derived from the differential model predictions described in S3. The prediction results are used as state inputs to ensure that the strategy generation has individualized and physiologically driven characteristics.

[0170] S5. Causal reasoning and dynamic intervention: Based on metabolic data, eating habits, and children's emotional changes, causal reasoning is used to analyze the relationship between diet, metabolism, and behavior to generate dynamic intervention plans;

[0171] In this embodiment, the causal reasoning and dynamic intervention module is used to construct a causal graph structure and analyze the causal links between variables based on metabolic modeling output, dietary behavior records and emotional state quantification results, and ultimately generate a periodically updated personalized intervention strategy.

[0172] First, a multimodal causal graph model based on Bayesian structural learning is established. As nodes, define the causal graph as The edge set ε represents the directed causal dependency relationship between variables.

[0173] The node set includes the following sub-variables:

[0174] Where G(t) is the blood glucose concentration at time t; I(t) is the insulin concentration at time t; L(t) is the body fat rate at time t; F(t) is the fatty acid concentration at time t; D(t) is the dietary behavior variable, including eating time, nutritional structure, etc.; M(t) is the emotional state variable, from the child psychological monitoring model or physiological signal modeling.

[0175] The Bayesian causal discovery method based on structure score is used for graph structure learning. The structure optimization objective is to maximize the posterior probability under the edge structure:

[0176] Where, is the observation data set; represents the likelihood of data under the current causal graph structure; is the structure prior, which controls the sparsity or structure interpretability of the graph.

[0177] Further, in order to enhance the time sensitivity of the graph structure, a time-aware graph attention network (Temporal Graph Attention Network, T-GAT) is used to dynamically adjust the edge weight.

[0178] The graph attention network includes an edge attention calculation part, a node state encoding part, and a time embedding module, and the input is the node state in the form of time series:

[0179] Where, is the graph embedding representation of node v at time t; is the neighbor node set of node v; τ(t) is the time embedding vector, which is used to describe the time context relationship between the interactions of nodes.

[0180] The network outputs the edge weight update value, which is used to correct the strength of the edge in the causal graph and dynamically adjust the graph structure, so as to capture the time-dependent causality of diet, metabolic state and emotion.

[0181] Then the causal reasoning algorithm is used to calculate the intervention effect of the key variable. For the variable to be intervened XXX (such as sugar intake), the expected effect of the target variable Y (such as G(t) or M(t)) is analyzed:

[0182] Where do(X=x) represents the case of forcibly intervening the variable X to the value x; is the expectation operator; IE(·) is the intervention effect value, which is used to measure the direct controllable influence between variables.

[0183] In order to realize the above intervention expectation calculation, the structural equation modeling (Structural Causal Model, SCM) is used to model the causal parent variable of each node:

[0184] V i = f i (Pa i ,∈ i ), where V i is the current variable node; Pa i is the set of causal parent nodes of V i ; ∈ i is the noise term; and f i represents the generating mapping function from the parent nodes to the current node.

[0185] The causal effect estimation is completed by applying do-operations, removing edges ∈ X→Y , making structural modifications and recalculating the generating expectation of node Y.

[0186] After obtaining the causal effect values of each node, the system enters the dynamic intervention generation phase. This phase includes an intervention decision generation part and a strategy optimization part. The former locates the adjustment variables according to the high intervention effect causal path, and the latter combines the target function to minimize metabolic load and psychological fluctuation amplitude, outputting the optimized intervention sequence.

[0187] In an exemplary embodiment, the intervention optimization target function is:

[0188]

[0189] where δD(t) is the adjustment amount of the dietary intervention variable; δM(t) is the psychological intervention measure variable; Var(·) represents the fluctuation intensity of the variable; λ1, λ2, and λ3 are weight control factors for metabolic stability, emotional stability, and intervention cost, respectively.

[0190] This optimization problem is solved by a dynamic programming solver or a reinforcement learning strategy executor, which outputs the next period intervention suggestion, such as adjusting the postprandial carbohydrate intake ratio, conducting game-like relaxation training, etc.

[0191] This module includes a graph modeling part, an attention network part, an intervention optimization part, and a control output part. The graph modeling part is connected to the S3 metabolic modeling output and the behavior recording system; the attention network part is connected to the emotion recognition module and the causal graph; the intervention optimization part is connected to the intervention execution device control interface; and the control output part is connected to the user interaction system.

[0192] After each intervention period ends, the system returns the actual observed metabolic and behavioral response data to the graph modeling part, triggering incremental updates of the graph structure and re-estimation of the causal model, achieving closed-loop dynamic adjustment. S6, Federated Learning and Global Optimization: Based on federated learning, the local training model is aggregated and updated;

[0193] In this embodiment, the federated learning and global optimization steps are used to update the metabolic model and reinforcement learning strategy globally based on the local data training results of user terminals while protecting the privacy of individual data.

[0194] First, a local metabolic kinetics model and a hierarchical reinforcement learning strategy network are deployed on each user device. The local model is used to predict the time variation of individual blood glucose concentration G, insulin concentration I, body fat rate L, and fatty acid concentration F, respectively, and output personalized dietary behavior adjustment strategies.

[0195] The local metabolic model is described in the form of system dynamics as follows:

[0196] where X i (t)=[G i (t),I i (t),L i (t),F i (t)] T is the metabolic state vector of the i-th terminal device user; D i (t) is the corresponding dietary input variable; θ i is the differential equation model parameter of the local user i; f i (·) is a family of nonlinear functions describing the metabolic variation law.

[0197] Each terminal device iteratively updates θ i locally to train the metabolic model and behavior strategy network. Then, the federated server initiates an aggregation request to collect the model gradient or parameter update uploaded by each device.

[0198] The federated aggregation process adopts a differential equation parameter aggregation mechanism (FedODE) to perform weighted averaging on the parameters uploaded by all devices. The process is defined as follows:

[0199] where N is the total number of user devices participating in federated update; n i is the number of local training samples of the i-th device; is the total number of samples; is the model parameter uploaded by the i-th device in the t-th round; is the global model parameter in the t+1-th round.

[0200] Further, the reinforcement learning strategy parameters also adopt a similar aggregation mechanism:

[0201] where, is the hierarchical strategy network parameter of the i-th device; is the aggregated global strategy parameter.

[0202] The federal server will distribute the aggregated parameters to all participating devices after each round. Each device takes the parameters as the initialization point, and continues to perform differential modeling and policy optimization combined with local data.

[0203] In an exemplary implementation, to enhance local adaptability, after receiving the global parameters, each terminal device performs a local fine-tuning, defined as follows:

[0204] Where η is the local fine-tuning learning rate; is the local loss function on the i-th device.

[0205] The local loss function includes two parts: model prediction error and behavior policy reward error:

[0206] Where, is the mean square error of the predicted state and the actual observation value; is the negative expectation based on the reinforcement learning reward; τ is the balance factor, which controls the importance weight of modeling error and policy reward.

[0207] The system includes a local modeling unit, a policy optimization unit, a federal aggregation unit, and a global broadcast unit. The local modeling unit is connected to the terminal perception system, which is used to collect physiological state X i (t) and dietary behavior D i (t); the policy optimization unit is connected to the hierarchical reinforcement learning module; the federal aggregation unit realizes parameter aggregation by a central server; the global broadcast unit is connected to each user device to perform synchronous update.

[0208] In an exemplary intervention cycle, the federal learning process is as follows:

[0209] First, each terminal device independently trains the local metabolic dynamics model and policy network;

[0210] Then, upload the current model parameters to the federal server;

[0211] The federal server performs parameter weighted aggregation to generate a global model;

[0212] The aggregated is issued to each device;

[0213] After each device performs local fine-tuning, it enters the next training cycle.

[0214] This mechanism forms a federal closed-loop learning path centered on "local modeling-federal aggregation-global optimization-individual retraining". During the federal optimization process, no raw data is transmitted, only model parameters are synchronized, ensuring the privacy of user data and the generalization ability of the system.

[0215] S7, personalized feedback output: according to the generated personalized dietary intervention strategy, real-time dietary report, health advice is provided to parents and children, and children are guided to healthy dietary behavior through virtual nutritionist and gamification tasks.

[0216] In this embodiment, the personalized feedback output module provides feedback information to parents and children in real time based on the generated dietary intervention strategy to guide children to healthy dietary behavior. This module includes four parts: health advice push, dietary report generation, virtual nutritionist guidance and gamification task design.

[0217] First, the system generates a dietary report for parents and children according to the global optimized personalized dietary intervention strategy D(t) combined with real-time data feedback from users. The report includes daily menu recommendations, nutritional component analysis, personalized dietary advice and health goal prompts. The specific report content is:

[0218] Where R i (t) is the dietary report of the i-th user (i.e. child); D i (t) is the personalized dietary intervention strategy, i.e. the recommended dietary content and intake time; is the nutritional component analysis of the diet, such as protein, carbohydrates, fat, etc.; G i (t) is the user's current health goal, such as blood glucose concentration, body fat rate, etc.; is the health advice, such as increasing exercise, reducing sugar intake, etc.

[0219] The report content is pushed through the parent and child interaction system, and the child can receive dietary guidance through the virtual nutritionist.

[0220] The virtual nutritionist guidance process uses conversational AI technology to provide dietary guidance to children through natural language generation based on user feedback and dietary habits. The decision-making process of the virtual nutritionist is driven by the following reinforcement learning model:

[0221] Where Q i (t,a) is the expected return of the i-th user at time t taking action a; R i (t,a) is the immediate reward of the current action a, such as whether a healthy diet has been chosen; γ is the discount factor; a ′ is the possible next action.

[0222] Through this reinforcement learning model, the virtual nutritionist can dynamically adjust the feedback strategy to guide children to develop healthy dietary habits.

[0223] In addition, the system further enhances the child's sense of participation by designing gamified tasks. Each task is related to a health goal, such as completing daily dietary recommendations, choosing low-sugar meals, etc. The task design follows the following objective function:

[0224] wherein, is the total score of the i-th user at time t; Score i (t) is a positive reward given according to whether the child has followed the dietary recommendations; Penalty i (t) is a penalty value for not following the dietary tasks; and α, β are the weights of the positive reward and the penalty, respectively.

[0225] The task completion is fed back to the user in real time, enhancing the interactive experience and the sense of goal achievement.

[0226] The system includes a dietary report generation unit, a virtual nutritionist guidance unit, a task design unit, and a feedback pushing unit. The dietary report generation unit generates reports based on user individualized strategies and health data; the virtual nutritionist guidance unit provides real-time guidance through an AI model; the task design unit designs daily tasks and calculates task scores; and the feedback pushing unit pushes all information to parent and child devices, supporting user interaction.

[0227] The operation process of the system is as follows:

[0228] First, individualized dietary reports and health recommendations are generated based on the globally optimized dietary intervention strategy; then, the virtual nutritionist guides the child's dietary behavior through a conversational AI and adjusts the feedback strategy according to the child's dietary choices; finally, the child's participation is enhanced through gamified task design, and positive or negative feedback is given according to the task completion.

[0229] The child obesity digital dietary management system described below can be mutually corresponding to the child obesity digital dietary management method described above.

[0230] Please refer to the attached Figure 2 A child obesity digital dietary management system, applied to the child obesity digital dietary management method described above, comprising:

[0231] Data acquisition module: for acquiring physiological data, dietary data and environmental data of children, and preprocessing these data to generate a unified time series data matrix;

[0232] Metabolic kinetics modeling module: for generating a metabolic kinetics model of the child according to the data, and dynamically calibrating metabolic parameters in the model;

[0233] Reinforcement learning decision module: for generating individualized short-term and long-term dietary adjustment strategies according to the metabolic model output;

[0234] Causal reasoning and intervention module: used to analyze the causal relationship between diet, metabolism and behavior, and generate dynamic intervention plans based on the analysis results;

[0235] Federated learning module: used to aggregate model updates of local devices into the global model;

[0236] Feedback output module: used to provide personalized health reports and dietary recommendations to parents and children in real time based on dietary adjustment strategies, and guide children to engage in healthy eating behaviors through virtual nutritionists and gamified tasks.

[0237] The Causal Reasoning and Intervention module includes:

[0238] A causal graph building module for identifying causal relationships between diet, metabolism, and behavior;

[0239] A graph attention network module is used to dynamically update the edge weights of causal relationships and generate intervention plans.

[0240] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, so they will not be repeated here.

[0241] While 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 these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A digital dietary management method for childhood obesity, characterized in that: The following steps are involved: Data Collection: Wearable devices collect real-time physiological data, including children's blood sugar, heart rate variability, and body fat percentage. Mobile apps also collect children's dietary data, using image recognition technology to automatically identify food types and portions. Data preprocessing: Clean and standardize the collected multimodal data to generate a unified time series data matrix; Metabolic kinetic modeling: Based on collected physiological and dietary data, a metabolic kinetic model for children is constructed. Differential equations are used to describe the metabolic processes of blood glucose, insulin, and fat. The model parameters are dynamically calibrated using a variational autoencoder. Hierarchical reinforcement learning decision generation: Hierarchical reinforcement learning algorithms are used to generate personalized short-term and long-term dietary adjustment strategies for children, including real-time dietary structure adjustment and long-term health goal optimization; Causal reasoning and dynamic intervention: Based on metabolic data, eating habits, and children's emotional changes, causal reasoning is used to analyze the relationship between diet, metabolism, and behavior, and to generate dynamic intervention plans; Federated learning and global optimization: Aggregate and update local training models based on federated learning; Personalized feedback output: Based on the generated personalized dietary intervention strategy, dietary reports and health recommendations are provided to parents and children in real time. Children are guided to adopt healthy eating behaviors through virtual nutritionists and gamified tasks.

2. A digital dietary management method for childhood obesity according to claim 1, characterized in that: The dietary data collection step includes: Use the user's mobile phone application to take a photo of the plate, automatically identify the food type and portion through image recognition technology, and generate a preliminary diet log; The image recognition results are compared with the dietary data manually entered by the user. If the recognition results are inconsistent with the manual input, the system will correct the food recognition results through the hierarchical reinforcement learning model; Use wearable devices to monitor children's physiological status in real time, collecting blood sugar data, heart rate variability, and body fat percentage data; The collected physiological data, dietary data, and environmental data are time-synchronized to generate a unified time series data matrix.

3. A digital dietary management method for childhood obesity according to claim 1, characterized in that: The metabolic kinetics modeling step includes: Based on the collected physiological and dietary data, a metabolic kinetic model of children was constructed, using differential equations to describe the metabolic processes of blood glucose G, insulin I, fat L, and fatty acid concentration F; Dynamically estimate the parameters in the metabolic model through a variational autoencoder and update the parameters using a variational inference formula; The metabolic model is trained using the variational lower bound loss function, and high-precision metabolic parameter estimation is achieved by minimizing the ELBO loss function, and the parameters in the metabolic model are updated.

4. A digital dietary management method for childhood obesity according to claim 1, characterized in that: The mathematical description of the metabolic kinetic model is: Among them, G is blood glucose concentration, I is insulin concentration, L is body fat percentage, FF is fatty acid concentration, D(t) is dietary input, k1, k2, k3, k4, k5, k6, k7, k8, k9, k 10 are all parameters of the metabolic model. represents the rate of change of blood glucose G with time t, represents the rate of change of insulin I with time t, represents the rate of change of body fat L over time t, represents the rate of change of fatty acid F with time t; The variational inference formula is as follows: Among them, q φ (z|x) is the approximate posterior distribution of the latent variable z given the input data x, approximated by a parameterized distribution; Indicates that this distribution is a Gaussian distribution, where: μ φ (x) is the mean value of the latent variable z after a given input x, which is obtained by processing the input data x through the neural network parameter φ; σ φ (x) is the standard deviation of the latent variable z given the input x, which is also obtained by processing the input data x through the neural network parameters φ; z represents the latent variable, which is usually used to represent the potential structure or abstract features of the input data; x represents the input data or observation data, such as images, text, or other types of data; p θ (k|z) represents the conditional probability of generating observation k given a latent variable z. The generation process is modeled by the decoder. This conditional probability is determined by the decoder's neural network parameters θ. Decoder(z;θ) represents the decoder's neural network model, which generates data by inputting the latent variable z. k represents the generated data.

5. A digital dietary management method for childhood obesity according to claim 1, characterized in that: The hierarchical reinforcement learning decision generation step includes: A hierarchical reinforcement learning model is used to generate personalized dietary adjustment strategies for children, which is divided into two strategy modules: bottom-level and top-level. The underlying decision module receives the child’s immediate physiological status data and selects a short-term diet adjustment strategy through the reward function R short Calculate the effects of immediate dietary modifications; The top-level decision module receives the long-term health goal data of the child and selects a long-term diet pattern strategy through the reward function R long Optimize long-term health goals; Through the reinforcement learning model, information interaction is carried out between the bottom and top strategy modules to dynamically adjust the dietary structure and optimize the short-term and long-term health status of children.

6. A digital dietary management method for childhood obesity according to claim 5, characterized in that: The reward function R short The calculation formula is as follows: R short =α1·Stability(G(t))+α2·Satiety Index -α3·Craving Score ; Among them, α1, α2, and α3 are regulatory factors, Stability (G(t)) represents the stability of blood glucose concentration G at time t, and Satiety Index Indicates satiety index, Craving Score represents appetite score; The reward function R long The calculation formula is as follows R long =β1·ΔBMI+β2·ΔInsulin Sensitivity -β3·Rebound Risk ; Among them, β1, β2 and β3 are regulatory factors, ΔBMI represents the change in body mass index, ΔInsulin Sensitivity Indicates the change in insulin sensitivity, Rebound Risk Indicates the risk of weight regain.

7. A digital dietary management method for childhood obesity according to claim 1, characterized in that: The causal reasoning and dynamic intervention steps include: Constructing a multimodal causal graph using the Bayesian causal discovery algorithm Identify causal relationships between diet, metabolism, and behavior; The time-aware graph attention network is used to dynamically update the edge weights in the causal graph, adjusting the structure of the causal graph based on the child's diet, metabolic status, and emotional data. Use causal reasoning to analyze the influence chain between diet, metabolism, and behavior, and calculate the intervention effect at each node; Based on the results of causal reasoning, a dynamic optimization algorithm is used to generate personalized dietary adjustment plans and psychological intervention plans, and real-time intervention suggestions are delivered; In each intervention cycle, the causal reasoning model is updated and the intervention strategy is adjusted in real time.

8. The digital dietary management method for childhood obesity according to claim 1, characterized in that: The federated learning and global optimization steps include: Each user device locally trains metabolic models and reinforcement learning strategies; The differential equation parameter federation aggregation algorithm is used to aggregate the gradients of the local model and update the parameters of the global metabolic model; Based on the federated learning algorithm, the training results of each terminal device are aggregated to optimize the global model and achieve efficient personalized learning.

9. A digital dietary management system for children with obesity, characterized in that: A digital dietary management method for childhood obesity as described in any one of claims 1 to 8, comprising: Data acquisition module: used to collect children's physiological data, dietary data and environmental data, and pre-process these data to generate a unified time series data matrix; Metabolic kinetics modeling module: used to generate a metabolic kinetics model for children based on data and dynamically calibrate the metabolic parameters in the model; Reinforcement learning decision module: used to generate personalized short-term and long-term dietary adjustment strategies based on metabolic model output; Causal reasoning and intervention module: used to analyze the causal relationship between diet, metabolism and behavior, and generate dynamic intervention plans based on the analysis results; Federated learning module: used to aggregate model updates of local devices into the global model; Feedback output module: used to provide personalized health reports and dietary recommendations to parents and children in real time based on dietary adjustment strategies, and guide children to engage in healthy eating behaviors through virtual nutritionists and gamified tasks.

10. A digital dietary management system for children with obesity according to claim 9, characterized in that: The causal reasoning and intervention module includes: A causal graph building module for identifying causal relationships between diet, metabolism, and behavior; A graph attention network module is used to dynamically update the edge weights of causal relationships and generate intervention plans.

Images