Intelligent nutrition correction tracking early warning method and system

Abstract

The invention relates to the technical field of artificial intelligence and nutrition health management, and discloses an intelligent nutrition correction tracking early warning method and system.The intelligent nutrition correction tracking early warning method comprises the steps that a hierarchical heterogeneity causal effect model is constructed, and individual features are mapped to a potential type space; representing a causal relationship between nutrition intervention and result variables by using a structural causal model, and evaluating an intervention effect by using a dual robust estimator; personalized grouping is realized by identifying individual subgroups; a reinforcement learning algorithm based on context awareness is combined with an anti-fact reward function, and a personalized nutrition intervention strategy is optimized; key feature combinations are extracted through a differentiable decision tree and SHAP value analysis, and multi-level causal interpretation is generated. According to the method, anti-factual reasoning, reinforcement learning and causal relationship extraction technologies are fused, accurate prediction, evaluation and early warning of the individual nutrition intervention effect are realized, and the personalized precision of nutrition intervention and the user compliance are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence and nutritional health management technology, and more specifically, to an intelligent nutritional correction tracking and early warning method and system. Background Art

[0002] With the rapid development of precision medicine and personalized health management, nutritional intervention, as a key health management tool, is playing an increasingly important role in the prevention and treatment of chronic diseases. Current nutritional intervention approaches primarily adopt a static, one-size-fits-all model, viewing the nutritional intervention system as a network structure composed of independent nodes, such as individual characteristics, intervention measures, and response outcomes. This heterogeneous data is mostly optimized and decided based on existing nutritional management technologies, making it difficult to handle the high-dimensional, complex, and dynamic nature of personalized nutritional interventions.

[0003] Therefore, how to deal with the high-dimensional, high-complexity and dynamically changing individualized nutritional intervention work, so as to accurately capture the differences in individual response patterns, strictly distinguish between cause and effect and correlation, dynamically optimize intervention strategies and provide clearly explained intelligent nutritional correction tracking and early warning methods to improve the personalized precision, accuracy and user compliance of nutritional intervention has become a technical problem that needs to be solved urgently. Summary of the Invention

[0004] The present invention provides an intelligent nutrition correction tracking and early warning method and system to solve the technical problem of how to handle high-dimensional, high-complexity and dynamically changing individualized nutrition intervention work in the prior art.

[0005] The present invention provides an intelligent nutrition correction tracking and early warning method, comprising the following steps: Construct a hierarchical heterogeneous causal effect model to map individual characteristics into latent type space and model the conditional average treatment effect; Based on the hierarchical heterogeneous causal effect model, a structural causal model was used to represent the causal relationship between nutritional intervention and outcome variables, and a doubly robust estimator was applied to evaluate the intervention effect. Based on the results of the causal relationship assessment, a recursive heterogeneity test algorithm is used to identify subgroups of individuals with similar response patterns to achieve personalized clustering. Targeting the identified subpopulations, a context-aware reinforcement learning algorithm combined with a counterfactual reward function is used to optimize personalized nutritional intervention strategies. Key feature combinations are extracted through differentiable decision trees and SHAP value analysis to generate multi-level causal explanations, enabling patients and doctors to understand the effects of intervention strategies and their causes.

[0006] Furthermore, the step of constructing a hierarchical heterogeneous causal effect model includes: Collect individual historical nutritional intervention data, including individual characteristic information, intervention measures and corresponding response results, to form a basic data set; Construct a feature mapping network to map individual feature vectors to the latent type space; Train a treatment effect prediction function that receives a latent type vector and an intervention as input and outputs the predicted treatment effect.

[0007] Furthermore, the step of using a structural causal model to represent the causal relationship between nutritional intervention and outcome variables includes: Construct a structural causal model to represent exogenous variables, endogenous variables, function sets, and probability distributions of exogenous variables; The doubly robust estimator was applied to calculate the average treatment effect; Representation learning is introduced to optimize the estimation process and feature representation is learned through neural networks.

[0008] Furthermore, the step of using the recursive heterogeneity test algorithm to identify subgroups of individuals with similar response patterns includes: Construct a response vector to represent the individual's response characteristics to different intervention measures; Define a similarity measurement function to calculate the distance between response vectors; Iteratively optimize subgroup division to maximize intra-group similarity and inter-group differences.

[0009] Furthermore, the steps of combining the context-aware reinforcement learning algorithm with the counterfactual reward function include: The nutritional intervention problem is modeled as a Markov decision process, which includes state space, action space, state transition probability, reward function and discount factor; Construct a counterfactual reward function based on the estimated conditional mean treatment effect; The SoftActor-Critic algorithm is used to learn the optimal intervention strategy.

[0010] Furthermore, the step of extracting key feature combinations through differentiable decision trees and SHAP value analysis includes: Train a differentiable decision tree model to predict the response pattern subgroup to which an individual belongs; Calculate SHAP value to analyze feature importance; Build a multi-level explanation generator to generate different levels of explanation content according to user types.

[0011] Furthermore, the feature mapping network includes three fully connected layers, namely an input layer, a hidden layer and an output layer, and the hidden layer uses batch normalization and Dropout technology to prevent overfitting.

[0012] Furthermore, the similarity measurement function is implemented by weighted combination of Euclidean distance, cosine similarity and dynamic time warping distance, and is used to measure the overall reaction amplitude difference, reaction pattern direction difference and temporal reaction pattern difference.

[0013] Furthermore, the counterfactual reward function trains the base model through a model-independent meta-learning algorithm and quickly adapts to each individual. The multi-objective reward function balances nutritional status improvement, compliance, and side effect control.

[0014] An intelligent nutrition correction tracking and early warning system is used to run an intelligent nutrition correction tracking and early warning method, comprising: Heterogeneous causal model building module, used to build hierarchical heterogeneous causal effect models, map individual characteristics into latent type space, and model conditional average treatment effects; The causal relationship representation and evaluation module is used to represent the causal relationship between nutritional interventions and outcome variables using a hierarchical heterogeneous causal effect model and to apply a doubly robust estimator to evaluate the intervention effect; The response pattern detection and clustering module is used to obtain causal relationship assessment results and use a recursive heterogeneity test algorithm to identify subgroups of individuals with similar response patterns to achieve personalized clustering; A reinforcement learning intervention optimization module is used to optimize personalized nutritional intervention strategies for different identified subpopulations based on a context-aware reinforcement learning algorithm combined with a counterfactual reward function; A multi-level self-explanatory generation module is used to extract key feature combinations through differentiable decision trees and SHAP value analysis to generate multi-level causal explanations, enabling patients and doctors to understand the effects of intervention strategies and their causes.

[0015] The beneficial effects of the present invention lie in: through a data acquisition module, a multi-source data fusion module, a host-microbe interaction network construction module, a spatiotemporal feature extraction module, a cascade effect propagation module, and a differential mutual verification module; and by constructing a multi-source data fusion model, a host-microbe interaction network, a spatiotemporal feature extraction algorithm, a cascade effect propagation algorithm, and a differential mutual verification mechanism, the present invention addresses the core shortcomings of traditional methods in identifying the synergistic effects of microbial functional pathways and host genes, effectively identifying the dynamic changes in the microbiome during the incubation period, and interactions at the functional pathway level. Specifically, it significantly improves prediction accuracy. In the prediction of autoimmune diseases such as inflammatory bowel disease, rheumatoid arthritis, and multiple sclerosis, the accuracy rate increased by 38%, the sensitivity increased by 45%, and the false positive rate decreased by 27%. Furthermore, by accurately identifying the changes in the microbiome during the incubation period of the disease, ultra-early intervention is achieved, advancing the prevention window by 2-5 years and significantly improving the success rate of intervention. By leveraging the host-microbe interaction network and the cascade effect propagation algorithm, this method analyzes the interaction mechanism between the functional pathways of the intestinal microbiome and host genes, providing a new perspective for understanding the pathogenesis of diseases. Through a differential mutual verification mechanism, the model demonstrates strong generalization across datasets from diverse regions, populations, and disease types, overcoming the generalization issues of existing models with heterogeneous data. Combined with counterfactual inference methods, it can predict the effects of different interventions at different stages of the disease incubation period, providing a quantitative basis for personalized risk management and optimizing the timing and approach of interventions. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of an intelligent nutrition correction tracking and early warning method provided in an embodiment of the present invention; Figure 2 This is a module diagram of an intelligent nutrition correction tracking and early warning system provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The subject matter described herein will now be discussed with reference to example embodiments. It should be understood that these embodiments are discussed solely to enable those skilled in the art to better understand and implement the subject matter described herein, and that the functions and arrangements of the elements discussed may be varied without departing from the scope of this specification. Various examples may omit, substitute, or add various processes or components as needed. Furthermore, features described in some examples may be combined in other examples.

[0018] At least one embodiment of the present invention discloses an intelligent nutrition correction tracking and early warning method, such as Figure 1 Shown, including: Step 1: Construct a hierarchical heterogeneous causal effect model to map individual characteristics into the latent type space and model the conditional average treatment effect.

[0019] The model models the conditional mean treatment effect via a metafunction: ; in, represents the individual feature vector, Nutritional interventions, represents the function that maps individual features to the latent class space, represents the mapping function from the type space to the treatment effect, represents the conditional average treatment effect. The specific steps include: 1.1 Collect individual historical nutritional intervention data, including individual characteristic information, intervention measures and corresponding response results, to form a basic data set ,in, represents a feature set, represents a set of intervention measures, Represents a collection of response results.

[0020] 1.2 Building a feature mapping network , the network adopts a multi-layer perceptron structure, inputting individual feature vectors , output potential type vector , realizing the mapping from the original feature space to the potential type space. Network parameters The optimization is performed by minimizing the following loss function: ; in, Represents an individual The actual reaction results, represents the treatment effect prediction function, represents the regularization coefficient, represents the number of samples, represents the parameters of the treatment effect prediction function, Representation parameters The L2 norm squared, represents the loss function of the feature mapping network, Represents an individual The eigenvector of Represents an individual Interventions received.

[0021] It should be noted that the feature mapping network consists of three fully connected layers: the input layer, the hidden layer, and the output layer. The input layer dimension is equal to the feature vector dimension. The hidden layer uses 256 neurons with the activation function ReLU (Rectified Linear Unit). The output layer dimension is equal to the latent space dimension (default is 32). To prevent overfitting, batch normalization and dropout (with a probability of 0.3) are added after the hidden layer. The feature vector contains continuous features (such as age and body mass index) and categorical features (such as gender and activity level). The categorical features are converted to low-dimensional dense vectors through the embedding layer and then concatenated with the continuous features before being input into the network.

[0022] In some embodiments, the feature mapping network can adopt a variational auto-encoder (VAE) structure, which introduces randomness to make the latent space smoother and help capture the uncertainty of individual features. In this case, the loss function needs to add a KL divergence term to constrain the latent space distribution: ; in, is the balance parameter, is the posterior distribution, is the prior distribution (usually the standard normal distribution), represents the Kullback-Leibler divergence, represents the loss function of the feature mapping network, represents the total loss function of the variational autoencoder, represents the potential type vector, represents the individual feature vector, Indicates network parameters.

[0023] 1.3 Simultaneous training of treatment effect prediction functions , which receives a vector of potential types and interventions As input, it outputs the predicted treatment effect.

[0024] function Using an interactive network structure, the parameters are optimized through the following loss function : ; in, Represents an individual The potential type vector of represents the regularization coefficient, Representation parameters The L2 norm squared, Represents an individual The actual reaction results, represents the number of samples, represents the loss function of the treatment effect prediction function, Represents an individual The eigenvector of Represents an individual interventions received, represents the treatment effect prediction function, Represents the feature mapping network.

[0025] The specific implementation of the interactive network structure is as follows: first, the potential type vector and intervention vectors The model passes through two independent fully connected layers, then calculates their interaction features (including element-wise products and absolute differences). Finally, the original and interaction features are concatenated, and a multi-layer perceptron (MLP) is used to output the predicted treatment effect. This structure effectively captures the interactive relationship between individual characteristics and intervention measures, improving prediction accuracy. In practical application scenarios, such as nutritional intervention for celiac disease patients, this model can accurately predict patients' personalized responses to gluten-free diet interventions, identify rapid and slow responders, and formulate differentiated intervention strategies in advance.

[0026] In some embodiments, the interaction network can be replaced with an interaction module based on a self-attention mechanism, which is particularly suitable for handling the combined effects of multiple interventions.

[0027] At this point, feature interaction Calculation formula ; in, 、 、 are the query, key, and value matrices, respectively, which are generated by linear transformation of the potential type vector and the intervention measure vector. represents the feature dimension, represents the softmax function, represents the matrix transpose, represents the transpose multiplication of the Q matrix and the K matrix, represents the square root of the feature dimension d. This attention mechanism can dynamically adjust the importance of interactions between different interventions and individual characteristics. For example, in trace element supplementation interventions, the self-attention mechanism can identify the complex interactions between elements such as iron, zinc, and vitamin C, enabling more precise and personalized supplementation plans.

[0028] Step 2: Based on the hierarchical heterogeneous causal effect model, a structural causal model was used to represent the causal relationship between the nutritional intervention and the outcome variables, and a doubly robust estimator was applied to evaluate the intervention effect.

[0029] The specific steps include: 2.1 Constructing a structural causal model: ; in, represents the set of exogenous variables (including potential confounders), represents the set of endogenous variables (including intervention variables and the outcome variable ), represents the set of functions that determine the values ​​of endogenous variables, represents the probability distribution of the exogenous variable, represents a structural causal model, represents the set of exogenous variables, represents the set of endogenous variables, Represents a set of functions, represents the probability distribution of the exogenous variable, represents the intervention variable, Represents the outcome variable.

[0030] 2.2 Applying the doubly robust estimator to calculate the average treatment effect : ; in, Indicates intervention The potential results of In the control intervention The potential results of Express expectations.

[0031] The specific calculation uses a dual machine learning method: in, Represents an individual The intervention indicator variable, Represents an individual The observation results, Represents an individual The eigenvector of represents the number of samples, represents the estimate of the mean treatment effect, represents the potential outcome model for the intervention group, represents the potential outcome model for the control group, represents the propensity score estimate.

[0032] To enhance the accuracy of causal inference, potential outcome models and propensity score models Both were implemented using an ensemble learning approach. Specifically, for the potential outcome model, a gradient boosting tree (maximum depth 5, number of trees 100) and a deep neural network (three-layer structure, hidden layer dimension [128, 64]) were combined to obtain the final prediction results through weighted averaging (weight ratio of 0.7:0.3). For the propensity score model, a binary ensemble of logistic regression and random forest was used (weight ratio of 0.4:0.6). This ensemble learning strategy can balance the advantages and disadvantages of different models and improve the robustness of the estimate. In the nutritional intervention scenario of metabolic syndrome, this method can accurately distinguish the differences in the effects of the two intervention programs of low sodium diet and DASH diet in patients with hypertension, excluding the influence of confounding factors such as age and weight.

[0033] In some embodiments, Bayesian nonparametric methods can be used instead of traditional machine learning models, such as Gaussian process regression (GPR) for potential outcome models and Bayesian additive regression trees (BART) for propensity score models. The advantage of such methods is that they can naturally quantify the uncertainty of the prediction, thereby providing more reliable estimates of the treatment effect and its confidence interval. For example, in a study of vitamin D supplementation in the elderly, the Bayesian method not only provides a point estimate of the supplementation effect, but also provides a credible interval for the effect of different individuals, making clinical decisions more cautious and personalized.

[0034] 2.3 Introducing the representation learning optimization estimation process, learning feature representation through neural networks: ; Apply counterfactual estimation in representation space: ; parameter Optimize by minimizing the following objective function: ; in, represents the estimate of the mean treatment effect, represents the result prediction loss, represents the balance loss, represents the regularization coefficient, Representation parameters The L2 norm squared, represents the estimate of the mean treatment effect, represents the number of samples, represents the potential outcome model for the intervention group, represents the potential outcome model for the control group, Representation characteristics The expression, Indicates that the parameter is neural network, Represents an individual The eigenvector of Representation parameters The loss function of .

[0035] The representation learning network adopts an encoder-decoder structure. The encoder maps high-dimensional features to a low-dimensional representation space through a three-layer fully connected network, and the decoder verifies the degree of information retention of the representation through reverse mapping. In particular, The Maximum Mean Discrepancy (MMD) metric is introduced into the loss to ensure that the distributions of the treatment and control groups in the representation space are similar, reducing the impact of selection bias. In the context of nutritional interventions for the elderly, representation learning methods can effectively handle the comparison between multivitamin supplementation and single nutrient supplementation, accurately estimating the causal effects of different interventions even when there are significant deviations in the observed data.

[0036] In some implementations, an adversarial learning strategy can be used to optimize the balance of the representation space and introduce a discriminator network. Try to distinguish the representations of the treatment group and the control group. At this point, the representation learning network The goal becomes to generate representations that the discriminator cannot distinguish; forming the following adversarial goal: in, and represent the characteristics of the treatment group and the control group, respectively. represents the feature representation function, represents the parameters of the learning network, represents the parameters of the discriminator network, and represent minimization and maximization operations respectively, and denote the expectations about the characteristics of the treatment and control groups, respectively, represents the natural logarithm, Represents resistance to loss, represents the discriminator network.

[0037] This adversarial balancing strategy is particularly suitable for situations where the characteristic distributions of the treatment group and the control group are quite different. For example, in a comparative study of dietary interventions between obese and normal-weight people, adversarial learning significantly reduced the estimation error caused by selection bias.

[0038] Step 3: Based on the results of the causal relationship assessment, a recursive heterogeneity test algorithm is used to identify subgroups of individuals with similar response patterns to achieve personalized clustering.

[0039] The specific steps include: 3.1 Constructing the reaction vector Represents an individual Response characteristics to different interventions, calculated based on historical intervention data and predictive models : ; in, Indicates 1, 2, A set of predefined interventions, represents the number of intervention measures, Represents an individual The reaction vector, represents the treatment effect prediction function, represents the feature mapping network, Represents an individual The eigenvector of represents the parameters of the treatment effect prediction function, Represents the parameters of the feature mapping network.

[0040] When constructing the response vector, in order to solve the sparsity problem in the intervention space, this application uses low-rank matrix decomposition technology. Specifically, the intervention space is mapped to a low-dimensional feature space (the dimension is usually 10% to 20% of the total number of interventions), and the matrix decomposition is used to solve the sparsity problem in the intervention space. ,in, is the individual latent factor, To intervene in potential factors, represents the individual-intervention response matrix, It is approximately equal to, represents the transpose of the matrix, thus achieving reasonable predictions of unobserved responses.

[0041] In some implementations, collaborative filtering models based on deep neural networks can be used instead of traditional matrix decomposition methods to capture complex patterns of individual-intervention interactions through nonlinear transformations. For example, Neural Collaborative Filtering (NCF) combines the advantages of multilayer perceptrons and matrix decomposition to learn more complex interaction functions: ; in, and are the embedding vectors of users and items respectively, represents element-wise product, represents vector concatenation, Indicates the predicted user About the project reaction, represents the activation function, represents the transposed weight vector, Represents a multilayer perceptron. In nutritional intervention data, this approach can more accurately predict individuals’ potential responses to untried interventions, especially for individuals with complex dietary habits and multiple nutritional needs.

[0042] 3.2 Apply the recursive heterogeneity test algorithm to identify subgroups. First, define the similarity measurement function Calculate the distance between the response vectors. Then initialize the subgroup centers through hierarchical clustering: ; in, They are 1, 2, adaptively determined subpopulations, The number of subpopulations determined for adaptation.

[0043] Similarity measurement function This is achieved by weighted combination of three distance metrics: Euclidean distance (measures the overall response amplitude difference), cosine similarity (measures the response pattern direction difference), and dynamic time warping distance (measures the temporal response pattern difference). The weight ratio is determined by cross-validation, with a typical value of 0.3:0.4:0.3. This application innovatively introduces the conditional heterogeneity test statistic: ; in, Represents an individual The treatment effect, represents the treatment effect of individual j, is the preset threshold, is the indicator function, Representing a collection The cardinality, Representing a collection The heterogeneity test statistic, represents a subgroup of individuals, and Represents individuals and individuals The response vector of . This statistic is used to assess the homogeneity within the subgroup and guide the subgroup division process.

[0044] In some embodiments, a similarity function based on deep metric learning can be used to learn the optimal distance metric through a twin network or a triplet network. The core idea of ​​this method is to learn a mapping function , so that the distance between similar samples in the embedding space is small, while the distance between different samples is large: ; in, 、 、 are anchor samples, positive samples, and negative samples respectively. represents the boundary parameter, Indicates taking the maximum value, represents the distance function, represents the triplet loss, represents the activity factor function, represents the protein factor function, represents a function of nutritional status factors. In nutritional response pattern analysis, this learned distance metric can adaptively discover complex similarity rules. For example, in a weight management program, it can identify groups of individuals with different weight change curves but similar long-term stability.

[0045] 3.3 Iterative optimization of subgroup division. The current subgroup division results are: ; in, Respectively represent 1, 2, adaptively determined subpopulations, Number of subpopulations determined for adaptation: ; in, Indicates subgroup A collection of individuals, Represents an individual Response vectors and subpopulations The distance from the center, Represents an individual The reaction vector, Indicates subgroup The center vector of Indicates subgroup The center vector of Indicates except Other subgroup indexes except Indicates subgroup The radius threshold is determined by maximizing the intra-group similarity and inter-group difference in the iterative process.

[0046] Update Subpopulation Center: ; in, Indicates subgroup The number of individuals in the process is calculated until convergence or the maximum number of iterations is reached.

[0047] The core of the recursive heterogeneity test algorithm lies in the adaptive splitting and merging strategy. In each round of iteration, the heterogeneity score of each subpopulation is first calculated, and the subpopulations that exceed the threshold are tried to be divided into two parts; then the heterogeneity increment after merging any two subpopulations is evaluated, and the subpopulation pair with the smallest increment is merged. This strategy can efficiently discover the natural clustering structure of the data without pre-specifying the number of subpopulations. In the dietary habit intervention study, the algorithm successfully identified three typical reaction pattern subpopulations: high-fat diet sensitivity, carbohydrate sensitivity, and mixed reaction type, providing a scientific basis for personalized dietary recommendations.

[0048] In some embodiments, spectral clustering or density peak clustering can be used instead of traditional distance-based clustering methods to better handle non-convex data distributions. In particular, density peak clustering automatically determines the number of clusters and is insensitive to noise points by identifying points with high local density and far distance from high-density points as cluster centers: First, the local density of each point is calculated. : ; in, For all points Sum, , Indicates the number of points, exist 1 if yes, otherwise 0. is the cutoff distance, Indicates a point and point The method calculates the distance between each point and the nearest point with a higher density; then calculates the distance between each point and the nearest point with a higher density; finally, the point with the larger value is selected as the cluster center. This method is very effective in identifying groups with special nutritional needs. For example, among diabetic patients, it can identify subgroups with different pathological mechanisms such as insulin resistance and insulin secretion deficiency, thereby formulating differentiated nutritional intervention plans.

[0049] Step 4: Optimize personalized nutritional intervention strategies for the identified subpopulations using a context-aware reinforcement learning algorithm combined with a counterfactual reward function.

[0050] The specific steps include: 4.1 Modeling the nutritional intervention problem as a Markov decision process: ; in, It is a mathematical framework for reinforcement learning. represents the state space (containing individual nutritional states and contextual features), represents the action space (feasible nutritional interventions), represents the state transition probability, represents the reward function, Represents the discount factor.

[0051] In the state representation design, this application uses the state vector Divided into three parts: basic features (including demographics and basic physiological indicators), current nutritional status (including current levels of various nutritional indicators) and situational characteristics (Including psychological state, social environment, seasonal factors, etc.). The introduction of contextual features is a significant improvement over traditional reinforcement learning state design, capable of capturing key external factors that influence intervention effectiveness. In the context of blood sugar management for diabetic patients, incorporating contextual factors such as work stress and life events increased the adaptability of intervention strategies by 35%, avoiding a one-size-fits-all approach.

[0052] In some implementations, a graph neural network (GNN) can be used to represent the state more efficiently, especially when there are complex dependencies between state variables. Specifically, the state variables are constructed as a graph structure. , where the node represents the characteristic variable, edge Represents the dependencies between variables (which can be determined through domain knowledge or correlation analysis). Then, a comprehensive representation is extracted through a graph convolutional network (GCN) or a graph attention network (GAT): ; in, Representation node In the The feature representation of the layer, Representation node The neighbor set of represents the attention weight, For the The weight matrix of the layer, Representation node In the The feature representation of the layer, Representation node In the The feature representation of the layer, represents the activation function, represents the network layer index, represents the neighbor node index, Represents the central node index. This graph-based state representation can better capture the interactions between nutritional indicators. For example, in mineral balance management, it explicitly models the complex dependencies between indicators such as calcium, phosphorus, and vitamin D, resulting in more coordinated intervention strategies.

[0053] 4.2 Constructing Counterfactual Reward Function : ; in, Indicates that the status Take action The estimated conditional mean treatment effect of To evaluate the status Take action expected effect.

[0054] Constructing multi-objective reward functions simultaneously : ; in, Indicates the improvement of nutritional status, Indicates compliance, Indicates the severity of the side effects, 、 、 are the first, second and third weight coefficients respectively, Represents the time step status, Represents the time step action, Represents the time step status, represents the counterfactual reward function, represents the multi-objective reward function, represents the estimated conditional mean treatment effect.

[0055] The counterfactual reward function is implemented through a meta-learning framework, that is, a basic model is trained through the Model-Agnostic Meta-Learning (MAML) algorithm, and then quickly adapted for each individual. Specifically, the meta-model parameters The following objective function is optimized: ; Specifically, ; in, Represents the minimize operation, represents the meta-learning model parameters, For all samples Sum, , represents the number of samples, Represents a sample The loss function is Indicates that the parameter is Model, Represents a sample Task-specific parameters for represents the meta-learning rate, Indicates about The gradient operator.

[0056] This meta-learning architecture enables the system to quickly adapt to new individuals and produce accurate counterfactual reward estimates even when data is scarce. Represents an individual The loss function is Indicates that the parameter is model.

[0057] In some implementations, a causal-enhanced counterfactual reward function can be used to introduce a structural causal model (SCM) to explicitly model the causal relationship between state variables, interventions, and rewards. The advantage of this approach is that it can calculate more accurate intervention effects through do-calculus: ; in, represents the counterfactual reward function, Indicates the expected value, represents the outcome variable, Represents a variable Intervene and set to value , represents the state variable, Represents the value of a state variable. In practice, causal graph structures are discovered from data using structured learning algorithms (such as the PC algorithm or the FCI algorithm), and then verified and corrected using expert knowledge. In long-term dietary pattern interventions, causally enhanced counterfactual reward functions can distinguish between direct and indirect intervention effects. For example, identifying the direct benefits of reducing saturated fat intake on cardiovascular health and the indirect benefits through weight loss, thereby generating more targeted intervention strategies.

[0058] 4.3 Using the Soft Actor-Critic (SAC) algorithm to learn the optimal strategy The algorithm contains a policy network Sum value function network , optimize the parameters through the following objective function and : ; in, represents the experience replay buffer, represents the entropy regularization coefficient, represents the loss function of the value function network, represents the loss function of the policy network, represents the discount factor, Indicates status, Indicates action, Indicates reward, Indicates the next state, Indicates the next action. Express expectations, represents the natural logarithm, Indicates that the parameter is The strategic network, Indicates that the parameter is The value function network.

[0059] In the implementation of the SAC algorithm, this application introduces the Prioritized Experience Replay technology, which assigns sampling weights based on the temporal difference (TD) error of the samples to enhance the learning of key state transitions. At the same time, the entropy regularization coefficient Through automatic adjustment mechanism optimization, cumulative rewards are maximized while maintaining an exploration-exploitation balance. In the scenario of formulating long-term weight loss strategies, the improved SAC algorithm demonstrates excellent stability and convergence. Compared with traditional reinforcement learning methods, it can produce smoother and more physiologically consistent intervention strategies, avoiding the common phenomenon of weight loss rebound.

[0060] In some implementations, a hierarchical reinforcement learning architecture can be used to decompose complex nutritional intervention decisions into multiple levels. For example, the upper-level strategy is responsible for setting long-term nutritional goals and overall strategic direction, and the lower-level strategy is responsible for selecting specific daily intervention measures. Generate abstract actions or subgoals , lower-level strategy Based on the current state and sub-goals Select a specific action: ; in, Represents high-level strategy, Represents the low-level strategy, represents the sub-goal, Indicates the current state, Indicates specific actions. Represents a sampling operation.

[0061] This hierarchical architecture can effectively handle long-term decision-making problems, improving the interpretability and robustness of the strategy. In long-term nutritional status adjustment, hierarchical reinforcement learning can simultaneously optimize short-term interventions and long-term nutritional goals. For example, in prenatal nutrition management, the upper-level strategy sets different nutritional priorities based on the stage of pregnancy (such as focusing on folic acid supplementation in the early stages, high-quality protein intake in the middle stages, and iron and calcium supply in the late stages), while the lower-level strategy is responsible for formulating daily dietary recommendations that meet these priorities.

[0062] Step 5: Extract key feature combinations through differentiable decision trees and SHAP value analysis to generate multi-level causal explanations, enabling patients and doctors to understand the effects of intervention strategies and their causes.

[0063] The specific steps include: 5.1 Training a Differentiable Decision Tree Model Predicting individuals The response pattern subgroup to which it belongs. The structure of the tree is determined by the parameters The decision, including splitting features and thresholds, is optimized by minimizing the following loss function: ; in, Represents an individual The subpopulation label, represents the cross entropy loss, represents the structural complexity penalty, represents the regularization coefficient, represents the loss function of the decision tree, represents the number of samples, represents the decision tree model, Represents an individual The eigenvector of Represents the set of parameters for a decision tree.

[0064] The differentiable decision tree of this application is implemented by using a soft decision boundary, i.e. (sigmoid) function instead of the hard threshold split in the traditional decision tree: ; in, represents the classification parameter, Controls the steepness of the boundary, is the splitting threshold.

[0065] This soft decision boundary makes the training process of the decision tree end-to-end differentiable and can be optimized by gradient descent. At the same time, to enhance the interpretability of the decision tree, a tree structure regularization term is introduced: ; in, represents the regularization function, represents the decision tree parameter set, represents the L1 regularization coefficient, represents the structural sparsity regularization coefficient, Indicates the decision tree parameters, represents the indicator function, Indicates that for all parameters Sum, , represents the number of parameters. The first term is L1 regularization, which promotes parameter sparsity, and the second term is structural sparsity regularization, which controls the complexity of the tree. In the context of trace element supplementation intervention, differentiable decision trees can accurately identify the key determinants of individual responses to trace element supplementation such as zinc and selenium, providing concise and intuitive classification rules.

[0066] In some implementations, an integrated interpretable model, such as the Scalable Bayesian Rule List (SBRL) or the Bayesian Optimal Classification Tree (BART), can be used to further improve classification accuracy while maintaining interpretability. In particular, SBRL generates a series of concise if-then rules through posterior inference, each with a clear probabilistic interpretation: ; in, Represents rules, Represents a given feature Time Tags The conditional probability of Representation Rules Next label The conditional probability of Represents a sample Whether the rules are met The indicator function, For all rules This rule list format is particularly friendly to non-technical users.

[0067] 5.2 Calculate SHAP value to analyze feature importance. For each prediction , calculate the features SHAP value: ; in, represents a feature set, represents a feature subset, represents the total number of features, Representation subset The number of features in Indicates that only a subset of features is used The prediction results, Representation characteristics The SHAP value of A subset of features prediction results.

[0068] Feature importance is sorted by the absolute value of SHAP value: ; in, Representation characteristics The importance score of Represents the total number of features.

[0069] In SHAP value calculation, this application adopts a strategy that combines the kernel-based estimation method (KernelSHAP) with the tree ensemble method (TreeSHAP) to balance computational efficiency and accuracy. In particular, the conditional SHAP value calculation is introduced: ; in, represents the conditional feature set, represents the conditional SHAP value, Representation characteristics SHAP value, Represents a sample The eigenvector of Represents the expected value, allowing exploration of the importance of a feature in a specific context.

[0070] In addition, by calculating the feature interaction SHAP value: ; in, Representation characteristics and features The interactive SHAP value of represents the set of all features, represents a feature subset, represents the total number of features, Indicates that only a subset of features is used The prediction function, Representing factorial operations, it identifies important interactions between features. In the analysis of nutritional supplement effects, this method successfully discovered important interactions between vitamin D and calcium absorption, and iron and vitamin C, providing an important basis for targeted nutritional intervention.

[0071] In some implementations, causal SHAP can be used instead of traditional SHAP values ​​to more accurately reflect the causal relationship between features and predicted results rather than just statistical correlation by combining causal graph structures for feature attribution.

[0072] Causal SHAP by intervening in distribution Instead of the traditional conditional distribution Calculate feature contribution: ; in, Representation characteristics The causal SHAP value of represents the sample feature vector, Representation characteristics The value of Indicates the removal of features All features except represents the prediction model, Indicates the expected value, Indicates the budget for the intervention. represents a probability distribution, and expectations are calculated based on the intervention distribution. This approach can avoid bias in feature importance estimates caused by confounding factors. In nutritional metabolism research, causal SHAP can distinguish between direct and indirect effects. For example, accurately attributing changes in blood lipid levels to whether they are due to the direct effect of dietary fat or through indirect pathways such as weight change, providing more precise guidance for intervention decisions.

[0073] 5.3 Build a multi-level explanation generator to generate different levels of explanation content based on user type (professional or general user). For professionals, we generate in-depth technical explanations with statistical evidence and detailed causal pathways; for general users, we generate easy-to-understand explanations that focus on key influencing factors and practical suggestions. The explanation content includes four aspects: description of reaction type, key influencing factors, interactions, and practical suggestions.

[0074] The multi-level explanation generator is implemented based on a templated NLG (Natural Language Generation) system and includes three core modules: document planning, micro-planning, and language implementation. In the document planning stage, the explanation content and structure are selected based on the user role; in the micro-planning stage, the specific expression method and level of terminology usage are determined; in the language implementation stage, the final natural language explanation is generated. The system also integrates a visual explanation component that can generate a variety of intuitive charts such as feature importance bar charts, feature interaction heat maps, and decision tree visualizations to enhance the comprehension of explanations. In the doctor-patient communication scenario, this multi-level explanation system significantly improves patients' understanding and acceptance of nutritional intervention plans, and doctors can also obtain more detailed decision-making basis, which promotes more accurate clinical decision-making.

[0075] In some embodiments, a comparative explanation generation strategy can be used to make explanations more intuitive by comparing them with reference cases or counterfactual scenarios. For example, a comparative explanation can be generated such as, compared to others with similar conditions but different responses to interventions, you are unique in that if your vitamin D level increases by 10 ng / ml, your bone density improvement is expected to increase by 15%. Studies have shown that such comparative explanations can significantly improve users' understanding and acceptance of intervention recommendations. In clinical nutrition management, comparative explanations help patients more clearly understand the relationship between personal characteristics and intervention effects, thereby improving compliance with and effectiveness of treatment plans.

[0076] This implementation method was applied to a nutritional intervention program for patients with pre-type 2 diabetes. The program aims to help patients control their blood sugar levels and delay or prevent the onset of diabetes through personalized nutritional intervention plans. The program enrolled 500 patients with pre-diabetes (fasting blood sugar 6.1-7.0 mmol / L or 2-hour glucose tolerance test blood sugar 7.8-11.1 mmol / L), aged 35-65 years, without severe complications. The program lasted six months, and each patient was equipped with a smart blood sugar monitoring device and a nutrition tracking app to collect real-time blood sugar data and dietary information.

[0077] Implementation process example: Construct a hierarchical heterogeneous causal effect model.

[0078] First, baseline patient data was collected, including demographic characteristics (age, gender, BMI, etc.), physiological indicators (fasting blood glucose, glycated hemoglobin, insulin resistance index, etc.), and lifestyle habits (dietary structure, activity level, etc.), totaling 42 characteristic dimensions. Furthermore, 10 different nutritional interventions implemented in the initial phase (such as a low-carb diet, a Mediterranean diet, and intermittent fasting) and their corresponding blood sugar fluctuation responses were recorded.

[0079] The feature mapping network is implemented using a three-layer fully connected neural network with an input layer dimension of 42, a hidden layer with 256 neurons, a batch normalization layer followed by a dropout layer (with a dropout rate of 0.3), and an output layer dimension of 32. Continuous features (such as blood sugar levels and BMI) are directly input, while categorical features (such as gender and dietary preferences) are converted to 8-dimensional vectors through an embedding layer and then concatenated with the continuous features. The network is trained by minimizing the following loss function: in, represents the loss function, Indicates patient The blood sugar response change value, Indicates patient The eigenvector of Indicates the nutritional interventions applied, represents the feature mapping network, represents the network parameters, represents the treatment effect prediction function, Denotes the prediction function parameter. After 2000 rounds of iterative training, the model's prediction mean square error on the test set was reduced from 1.26mmol / L using the traditional method to 0.82mmol / L.

[0080] Representing and evaluating causal relationships of nutritional interventions.

[0081] A structural causal model was constructed, with patient characteristics (such as age, BMI, baseline blood glucose, etc.) as exogenous variables and intervention measures and blood glucose response as endogenous variables. The average treatment effect of different intervention measures was calculated using a doubly robust estimator: in, represents the estimated mean treatment effect, Indicates patient The eigenvector of Indicates patient Whether to accept the intervention (1 means accepted, 0 means not accepted), Indicates patient The observation results, represents the potential outcome model (predicting Given features ), Represents a propensity score model (predicting a given characteristic probability of receiving intervention under the given

[0082] For potential outcome models , an ensemble model combining gradient boosting trees (100 trees, maximum depth 5) and a three-layer neural network (hidden layer dimensions 128 and 64) with a weight ratio of 0.7:0.3. , an ensemble of logistic regression and random forest (100 trees) was used with a weight ratio of 0.4:0.6.

[0083] The model analysis found that the average treatment effect of a low-carb diet on patients with high insulin resistance was -0.85mmol / L (reduced blood sugar), while it was only -0.32mmol / L for patients with low insulin resistance. In contrast, traditional analysis methods were unable to distinguish this heterogeneous effect and only produced an average effect of -0.58mmol / L.

[0084] Detect and stratify individual nutritional intervention response patterns.

[0085] Construct a response vector based on the patient's predicted response to different interventions , indicating that the patient Characteristics of responses to all 10 interventions: in, Indicates patient The reaction vector, Indicates patient The eigenvector of Indicates the intervention measures, represents the feature mapping network, represents the treatment effect prediction function.

[0086] The similarity of response patterns between patients was calculated using a weighted similarity metric function that combines Euclidean distance (weight 0.3), cosine similarity (weight 0.4), and dynamic time warping distance (weight 0.3). The recursive heterogeneity test algorithm started with an initial set of five subgroups and, after eight rounds of iteration, automatically identified four optimal subgroups: Subgroup 1 (32% of patients): highly sensitive to carbohydrate restriction and relatively insensitive to fat intake; Subgroup 2 (28% of patients): highly sensitive to increased dietary fiber and moderately sensitive to protein adjustments; Subgroup 3 (24% of patients): highly sensitive to meal time adjustments (such as intermittent fasting); Subgroup 4 (16% of patients): mixed response type, requiring multifactorial comprehensive intervention.

[0087] Compared with traditional demographic-based clustering, this clustering method increased the accuracy of predicting patients' blood sugar response from 72% to 89%.

[0088] Optimize nutritional intervention strategies based on causal effects.

[0089] The nutritional intervention problem is modeled as a Markov decision process. The state space includes the patient's basic characteristics (age, gender, etc.), current nutritional status (various blood sugar and metabolic indicators), and contextual characteristics (work stress, seasonal factors, etc.). The action space includes 17 fine-grained nutritional intervention measures.

[0090] Constructing the counterfactual reward function: ; in, represents the counterfactual reward function, Indicates status, Indicates action, Denotes the estimated conditional mean treatment effect. A general model is trained using a model-agnostic meta-learning algorithm, which is then rapidly adapted for each subpopulation.

[0091] Using multiple objective reward functions simultaneously: in, represents the total reward function, Indicates time status, Indicates time action, Indicates time status, Indicates blood sugar change value, Indicates patient compliance, Indicates the degree of side effects.

[0092] The intervention strategy was trained using a Soft Actor-Critic algorithm, with prioritized experience replay introduced to enhance learning efficiency for key state transitions. Through 5,000 training iterations, the optimal intervention strategy was learned for each subgroup. Compared to traditional fixed intervention plans based on clinical guidelines, this dynamic strategy, which can be adjusted based on real-time patient feedback, increased the blood glucose target achievement rate from 62% to 83%.

[0093] Generating multilevel self-explanatory explanations of nutritional intervention effects.

[0094] To predict the subgroup to which the patient belongs, a differentiable decision tree model is trained, and end-to-end differentiable training is achieved through soft decision boundaries: in, represents the loss function of the decision tree, Indicates patient The subpopulation label, Indicates patient The eigenvector of represents a differentiable decision tree model, represents the cross entropy loss, represents the structural complexity penalty, Represents the parameters of the decision tree.

[0095] SHAP value analysis was used to identify key influencing factors. For example, for patients in subgroup 1, baseline insulin resistance index, fasting blood glucose level, and carbohydrate intake ratio were the three most important features for predicting their response to intervention, with SHAP values ​​of 0.28, 0.23, and 0.19, respectively.

[0096] Build a multi-level explanation generator to generate different levels of explanations for patients and doctors: The physician version of the explanation includes detailed causal path analysis, statistical evidence, and precise predictive indicators; The patient version uses plain language to emphasize key influencing factors and practical suggestions. For example, if you are particularly sensitive to carbohydrates, reducing your carbohydrate intake by 30g per day can lower your blood sugar by an average of 0.7mmol / L.

[0097] Technical effect verification: Technical effect 1: Improved accuracy of personalized nutritional intervention.

[0098] This method was applied to a 6-month intervention in 500 patients with prediabetes and significantly improved the precision of personalized intervention compared to traditional standardized protocols based on clinical guidelines. With the traditional approach, patients were randomly assigned to several standardized diets, resulting in an average blood sugar reduction of 0.42 mmol / L, with a target compliance rate of 62%. However, with this method, by identifying individual response patterns and providing targeted intervention, patients' blood sugars decreased by an average of 0.68 mmol / L, increasing the target compliance rate to 83% and improving personalized precision by 38%.

[0099] By analyzing the degree of match between the actual and predicted effects of interventions, this method achieved an 86% prediction accuracy rate, a 32% improvement over the 65% achieved by traditional methods. Specifically, in terms of identifying ineffective interventions, this method reduced the number of ineffective interventions by 35%, avoiding unnecessary costs and time.

[0100] The top 40% of patients with high insulin resistance achieved a 2.3-fold increase in weight loss after using this method compared to traditional methods, with an average additional weight loss of 3.2 kg, which further verifies the importance of accurately identifying the causal relationship of individual nutritional interventions.

[0101] Technical effect 2: Improved user understanding and compliance.

[0102] A comparative experiment found that the control group, which received traditional text explanations, had a 65% accuracy rate in understanding the intervention plan and a 58% six-month compliance rate. However, the experimental group, which received the multi-level self-explanatory method, had a 94% accuracy rate in understanding the intervention plan, a 45% improvement, and a 41% improvement in six-month compliance rate, reaching 82%.

[0103] Using a standardized user satisfaction questionnaire, the experimental group's satisfaction score (out of a maximum of 100) reached 87, significantly higher than the control group's score of 62. Specifically, on the question "Why is this intervention right for me?", the experimental group's understanding score was 92, while the control group's was only 59.

[0104] The incidence of adverse reactions reported by patients in the experimental group was 12%, significantly lower than the 28% in the control group, a 57% reduction. A detailed analysis revealed that this was primarily due to the system's ability to quickly adjust intervention strategies based on real-time patient feedback, avoiding potential risks.

[0105] Based on the above results, this implementation method has demonstrated its significant technical effects in improving the accuracy of personalized nutritional intervention and enhancing user understanding and compliance in actual applications, providing patients with prediabetes with a more accurate, safe and efficient nutritional intervention solution.

[0106] The above describes an embodiment of the present invention, but this embodiment is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Ordinary technicians in this field can also make more forms of equivalent embodiments based on the inspiration of this embodiment, all of which are protected by this embodiment.

Claims

1. An intelligent nutrition correction tracking and early warning method, characterized in that: The following steps are involved: Construct a hierarchical heterogeneous causal effect model to map individual characteristics into latent type space and model the conditional average treatment effect; Based on the hierarchical heterogeneous causal effect model, a structural causal model was used to represent the causal relationship between nutritional intervention and outcome variables, and a doubly robust estimator was applied to evaluate the intervention effect. Based on the results of the causal relationship assessment, a recursive heterogeneity test algorithm is used to identify subgroups of individuals with similar response patterns to achieve personalized clustering. Targeting the identified subpopulations, a context-aware reinforcement learning algorithm combined with a counterfactual reward function is used to optimize personalized nutritional intervention strategies. Key feature combinations are extracted through differentiable decision trees and SHAP value analysis to generate multi-level causal explanations, enabling patients and doctors to understand the effects of intervention strategies and their causes.

2. The intelligent nutrition correction tracking and early warning method according to claim 1, characterized in that: The steps of constructing a hierarchical heterogeneous causal effect model include: Collect individual historical nutritional intervention data, including individual characteristic information, intervention measures and corresponding response results, to form a basic data set; Construct a feature mapping network to map individual feature vectors to the latent type space; Train a treatment effect prediction function that receives a latent type vector and an intervention as input and outputs the predicted treatment effect.

3. The intelligent nutrition correction tracking and early warning method according to claim 1, characterized in that: The steps of using the structural causal model to represent the causal relationship between the nutritional intervention and the outcome variable include: Construct a structural causal model to represent exogenous variables, endogenous variables, function sets, and probability distributions of exogenous variables; The doubly robust estimator was applied to calculate the average treatment effect; Representation learning is introduced to optimize the estimation process and feature representation is learned through neural networks.

4. The intelligent nutrition correction tracking and early warning method according to claim 1, characterized in that: The step of using a recursive heterogeneity testing algorithm to identify subgroups of individuals with similar response patterns comprises: Construct a response vector to represent the individual's response characteristics to different intervention measures; Define a similarity measurement function to calculate the distance between response vectors; Iteratively optimize subgroup division to maximize intra-group similarity and inter-group differences.

5. The intelligent nutrition correction tracking and early warning method according to claim 1, characterized in that: The steps of combining the context-aware reinforcement learning algorithm with the counterfactual reward function include: The nutritional intervention problem is modeled as a Markov decision process, which includes state space, action space, state transition probability, reward function and discount factor; Construct a counterfactual reward function based on the estimated conditional mean treatment effect; The SoftActor-Critic algorithm is used to learn the optimal intervention strategy.

6. The intelligent nutrition correction tracking and early warning method according to claim 1, characterized in that: The steps of extracting key feature combinations through differentiable decision trees and SHAP value analysis include: Train a differentiable decision tree model to predict the response pattern subgroup to which an individual belongs; Calculate SHAP value to analyze feature importance; Build a multi-level explanation generator to generate different levels of explanation content according to user types.

7. The intelligent nutrition correction tracking and early warning method according to claim 2, characterized in that: The feature mapping network includes three fully connected layers, namely an input layer, a hidden layer and an output layer. The hidden layer uses batch normalization and Dropout technology to prevent overfitting.

8. The intelligent nutrition correction tracking and early warning method according to claim 4, characterized in that: The similarity measurement function is implemented by weighted combination of Euclidean distance, cosine similarity and dynamic time warping distance, and is used to measure the overall response amplitude difference, response pattern direction difference and temporal response pattern difference.

9. The intelligent nutrition correction tracking and early warning method according to claim 5, characterized in that: The counterfactual reward function trains the base model through a model-independent meta-learning algorithm and quickly adapts to each individual. The multi-objective reward function balances nutritional status improvement, compliance, and side effect control.

10. An intelligent nutrition correction tracking and early warning system, used to run an intelligent nutrition correction tracking and early warning method according to any one of claims 1 to 9, characterized in that: include: Heterogeneous causal model building module, used to build hierarchical heterogeneous causal effect models, map individual characteristics into latent type space, and model conditional average treatment effects; The causal relationship representation and evaluation module is used to represent the causal relationship between nutritional interventions and outcome variables using a hierarchical heterogeneous causal effect model and to apply a doubly robust estimator to evaluate the intervention effect; The response pattern detection and clustering module is used to obtain causal relationship assessment results and use a recursive heterogeneity test algorithm to identify subgroups of individuals with similar response patterns to achieve personalized clustering; A reinforcement learning intervention optimization module is used to optimize personalized nutritional intervention strategies for different identified subpopulations based on a context-aware reinforcement learning algorithm combined with a counterfactual reward function; A multi-level self-explanatory generation module is used to extract key feature combinations through differentiable decision trees and SHAP value analysis to generate multi-level causal explanations, enabling patients and doctors to understand the effects of intervention strategies and their causes.

Images