Insulin resistance dynamic intervention method and system based on metabolic flow graph network and reinforcement learning

Abstract

The application provides an insulin resistance dynamic intervention method and system based on a metabolic flow graph network and reinforcement learning, and belongs to the technical field of insulin resistance intervention. The method comprises the following steps: acquiring first time series data and performing preprocessing; based on a feature matrix obtained through the preprocessing, performing individualized parameter learning through a heterogeneous graph neural network, constructing an individualized metabolic flow graph network model to generate a metabolic state vector; inputting the metabolic state vector into a reinforcement learning intelligent agent to generate an individualized intervention strategy vector; outputting and executing the strategy, and simultaneously performing online co-evolution updating on the metabolic network model and the reinforcement learning intelligent agent based on new data and feedback after the execution. The application solves the technical problems of the static, universal and unable-to-adapt-to-individual-metabolic-dynamic-fluctuation and multi-objective-weighting of existing intervention schemes, and realizes precise, real-time, adaptive and interpretable dynamic individualized intervention.

Description

Technical Field

[0001] This application relates to the technical field of insulin resistance intervention, and in particular to a dynamic intervention method and system for insulin resistance based on metabolic flow graph networks and reinforcement learning. Background Technology

[0002] Insulin resistance is a core pathological basis for many major metabolic chronic diseases, including type 2 diabetes, non-alcoholic fatty liver disease, cardiovascular disease, and polycystic ovary syndrome. Essentially, it involves reduced sensitivity of insulin target tissues to the biological effects of insulin, directly leading to an imbalance in blood glucose regulation and further inducing systemic disorders of glucose, lipid, and protein metabolism, thus becoming a significant public health problem threatening human health. Currently, clinical intervention and management of insulin resistance rely primarily on standardized protocols recommended by industry guidelines. These mainly include fixed-ratio dietary control, standardized exercise recommendations, stepwise drug administration regimens, and regular blood glucose monitoring and follow-up. While these protocols have played a fundamental role in population-based prevention and control, the complexity of the human metabolic system and individual variability limit the effectiveness of existing intervention models in meeting the precise and personalized clinical needs, revealing numerous technical shortcomings that urgently need to be addressed.

[0003] The human metabolic system is a complex network system dynamically regulated by multiple factors such as genetic background, epigenetic regulation, dietary behavior, exercise status, circadian rhythm, and gut microbiota. Its core characteristics are significant differences in inherent metabolic phenotypes among individuals, and high-frequency dynamic fluctuations in the metabolic response of the same individual at different times and under different physiological states. Existing insulin resistance intervention technologies fail to fully adapt to this core characteristic of the metabolic system, exhibiting several key limitations: First, the static nature of intervention strategies. Traditional protocols are often fixed or adjusted periodically according to a 3-6 month follow-up cycle, failing to capture and respond to the intraday and diurnal dynamic changes in indicators such as postprandial blood glucose, insulin secretion, and energy metabolism, and struggling to match the metabolic state switching triggered by instantaneous physiological events such as eating, exercise, and sleep. Second, the lack of universality and personalization in protocol design. Existing guidelines are based on the average metabolic characteristics of the population, lacking quantitative analysis of individual unique metabolic pathway activity, enzyme kinetics, gene expression patterns, and lifestyle constraints, resulting in a "one-size-fits-all" intervention model. This can easily lead to poor intervention response and unsatisfactory efficacy in some patients, or even side effects such as hypoglycemia and exacerbation of metabolic disorders. Third, the systematic nature and feedback loop of the intervention system. The existing measures often focus on a single blood glucose indicator in isolation, neglecting the interrelationships and compensatory regulatory mechanisms between glucose metabolism, lipid metabolism, and protein metabolism, which can easily lead to secondary imbalances in the metabolic network. Furthermore, there is a lack of a rapid, automated effect evaluation system based on high-dimensional, continuous physiological and biochemical data after the implementation of the plan, as well as a lack of a real-time dynamic adjustment mechanism, making it impossible to form an intelligent intervention closed loop of "perception-decision-execution-evaluation-optimization". Fourth, the interpretability of emerging intervention models is insufficient. In recent years, machine learning and deep learning models based on pure data-driven approaches have been applied to blood glucose prediction and intervention decisions for insulin resistance. Although they can capture complex nonlinear patterns in the data, the models themselves are "black box" characteristics, lacking a clear connection between the decision-making process and the biological mechanisms of human metabolism. This makes it difficult to gain the trust of clinicians and patient compliance, and also limits the learning and model optimization at the metabolic mechanism level from failed cases.

[0004] With the development of precision medicine and digital health technologies, the maturity of technologies such as continuous glucose monitoring devices, wearable physiological monitoring terminals, and mobile health management applications has enabled the continuous collection of individual first-series and behavioral data at low cost, high frequency, and multiple dimensions, providing a data foundation for dynamic intervention in insulin resistance. Simultaneously, metabolic network models in systems biology are constantly being improved, providing a mechanistic framework for quantitatively simulating the regulatory patterns of the human metabolic system and analyzing the dynamic changes in metabolic pathways in computers. Deep reinforcement learning technology in artificial intelligence has demonstrated powerful capabilities in sequence decision-making and dynamic optimization problems, providing algorithmic support for personalized real-time intervention decisions. However, in the current technological system, data-driven AI algorithms and domain-knowledge-based metabolic mechanism models remain relatively separate. How to deeply integrate the two to construct an intelligent intervention system that can accurately understand the dynamic changes of complex metabolic systems, make personalized real-time intervention decisions, and possess biological interpretability has become a key technical challenge that urgently needs to be overcome in the field of insulin resistance intervention technology.

[0005] Therefore, there is an urgent need in this field to develop a novel dynamic intervention technology for insulin resistance that can effectively integrate real-time, multi-source, heterogeneous physiological and biochemical data of individuals, construct a dynamic and personalized metabolic system computational model based on prior knowledge of metabolic biology, and combine advanced decision-making intelligence algorithms to generate an intervention strategy that can adapt to dynamic fluctuations in metabolic state, taking into account efficacy, safety, user compliance, and other multi-objectives. At the same time, it can achieve biological interpretability of the decision-making process, fundamentally overcoming the technical defects of existing intervention programs that are static, universal, black-box, and lack closed-loop, and promoting the upgrade of insulin resistance intervention from group-based and standardized to individualized, dynamic, and precise. Summary of the Invention

[0006] This application aims to solve the above-mentioned technical problems and proposes an innovative dynamic intervention method and system for insulin resistance based on metabolic flow graph networks and multi-objective reinforcement learning.

[0007] A first aspect of the embodiment provides a dynamic intervention method for insulin resistance based on metabolic flow graph networks and multi-objective reinforcement learning, comprising the following steps: acquiring various first time-series data of the user; processing the first time-series data to obtain a standardized feature matrix; constructing a personalized metabolic flow graph network model based on the standardized feature matrix and a pre-built general metabolic network topology, and generating a metabolic state vector, wherein the metabolic state vector represents the state that adapts to the differences in the user's metabolic network and provides interpretability; constructing a pre-trained reinforcement learning agent based on a multi-objective deep deterministic policy gradient algorithm, inputting the metabolic state vector into the pre-trained reinforcement learning agent to generate an insulin resistance intervention policy vector for the user, thereby realizing personalized decision-making under multi-objective trade-offs; outputting the intervention policy vector, and based on the new data obtained after the policy execution and user feedback, performing online co-evolutionary updates on the personalized metabolic flow graph network model and the reinforcement learning agent to achieve real-time adaptive adjustment to match dynamic changes in metabolism.

[0008] Through the above technical means, the embodiments of this application construct a personalized metabolic flow graph network model, which maps abstract physiological and biochemical data into metabolic state vectors with biological significance, and combines multi-objective reinforcement learning agents to optimize decision-making, thereby achieving dynamic personalized intervention that takes into account efficacy, safety and user compliance, and solving the technical problem that intervention strategies in the prior art are static, universal and lack interpretability.

[0009] In some optional embodiments, the first time-series data is processed, including: imputing missing values ​​and correcting outliers in the first time-series data to obtain second time-series data; dividing the second time-series data into multiple second data segments based on an adaptive sliding window Bayesian change point detection algorithm; extracting time-domain features, frequency-domain features, and nonlinear dynamic features from each second data segment to form the feature vector of the second data segment; and using an attention-based segment feature fusion network to fuse the feature vectors of each second data segment to obtain a standardized feature matrix.

[0010] Through the above technical means, the embodiments of this application can transform the original messy, multi-source heterogeneous time-series data into a clean, regular, dynamic pattern-rich, and standardized feature matrix suitable for subsequent graph neural network and metabolic model input by using advanced missing value imputation, outlier correction, adaptive change point detection, multi-dimensional feature extraction, and attention-based fusion techniques. This lays a high-quality data foundation for subsequent accurate modeling.

[0011] In some optional embodiments, the Bayesian change point detection algorithm based on an adaptive sliding window segments the second time-series data, including: defining the running length. r t The time elapsed since the last variable point; at each new data point xt Upon arrival, calculate its non-parametric similarity score with the current segment. ,in K For Gaussian kernel function, h For bandwidth; calculate the probability of the change point. Among them, the variable point prior , l This is the decay rate parameter; when the probability of a change point exceeds the dynamic threshold. i t When determining that a change point occurs at time t, where , used for smooth detection.

[0012] Through the above technical means, the non-parametric similarity calculation and the change point prior that increases with the running length introduced in the embodiments of this application enable the segmentation algorithm to more accurately capture the switching of metabolic state due to events such as eating, exercise, and sleep, and avoid missegmentation or omission due to the complex distribution characteristics of physiological data, thereby obtaining state segments that are more in line with physiological reality.

[0013] In some optional embodiments, constructing a personalized metabolic flow graph network model includes: loading a general metabolic network topology, which is formalized as a heterogeneous directed hypergraph. H =( V,E,R ),in V It is a heterogeneous set of nodes containing metabolite nodes, enzyme nodes, hormone signaling molecule nodes, and abstract physiological compartment nodes. E A hyperboundary set, representing biochemical reactions, transport processes, or regulatory relationships. R The set of relationship types between nodes and hyperedges; mapping the normalized feature matrix to a heterogeneous directed hypergraph. H The corresponding type of node is used as the initial feature of the node; message passing and node feature updating are performed through a relation-aware heterogeneous graph neural network to learn personalized feature representations of nodes and hyperedges; based on the learned hyperedge features, the maximum reaction rate parameter of each reaction hyperedge is estimated in a personalized way; using the personalized parameters, a flux balance analysis with dynamic constraints is performed at time t to calculate the network flux distribution, and combined with the concentration of key metabolites, a metabolic state vector is generated.

[0014] Through the above technical means, this application's embodiments accurately represent complex biological regulatory networks by introducing heterogeneous directed hypergraphs, and utilize relation-aware heterogeneous graph neural networks to learn personalized adjustments of network parameters from individual data, thereby instantiating a general metabolic framework into an individual-specific dynamic model. Flux balance analysis with dynamic constraints further integrates real-time physiological data into the model calculation, ultimately outputting a vector that comprehensively, quantitatively, and interpretably characterizes the current state of an individual's systemic metabolic pathways, providing high-quality environmental awareness for intelligent decision-making.

[0015] In some optional embodiments, message passing and node feature updating are performed using a relation-aware heterogeneous graph neural network, including: for nodes v and relation types r Calculate from neighboring nodes Relationship-specific messages: ,in The normalized attention weights are calculated based on the relational attention mechanism. Learnable linear transformations specific to relations. Features of the neighboring nodes at the next higher level; merging messages from different relation types: ,in β r To compute relation type weights using a gated fusion network; to update node features using gated recurrent units: .

[0016] Through the above technical means, the RHGNN designed in this application embodiment can process information flow differently according to different biological relationship types (such as catalysis and inhibition). The relationship attention mechanism enables the model to focus on specific neighbors and relationships that are important to the current node, while GRU update allows nodes to retain a certain historical memory, thereby learning dynamic node and hyperedge feature representations that better reflect individual metabolic regulation characteristics.

[0017] In some optional embodiments, the maximum reaction rate parameter for each reaction hyperedge is individually estimated, specifically as follows: Where V_max_e_ref is the reference maximum value, scale e = s ( MLP θ ( h e )), MLP θ It is a multilayer perceptron. h e This represents the personalized features of the hyperedge. s This is the Sigmoid function.

[0018] Through the above technical means, the embodiments of this application link the abstract features learned by the graph neural network with specific enzyme kinetic parameters (V_max), realizing a direct mapping from data to mechanism parameters. This enables the key reaction rate of the metabolic network model to be adaptively adjusted according to individual characteristics, which is the core step in realizing truly personalized dynamic modeling.

[0019] In some alternative embodiments, the optimization problem for flux balance analysis with dynamic constraints is: , ,in, c o For the first o A coefficient vector of biological targets S p For personalized stoichiometry matrix, v It is the flux vector. v m in ( t )and v m axe ( t () represents the dynamic flux boundary. v m axe ( t The formula for calculating ) is: ,in,[ S ]( t )and[ I ]( t The substrate concentration and effector concentration at time t are respectively. n , m The Hill coefficient is used. K m It is the Michaelis constant. K i This is the suppression constant.

[0020] Through the above technical means, the embodiments of this application upgrade the traditional static flux balance analysis to a dynamic constraint version. The dynamic boundary formula integrates personalized enzyme maximization capacity. V m axe e ( t The real-time metabolite concentrations allow the calculated flux distribution to not only satisfy the conservation of matter (steady-state constraints) but also conform to the enzyme kinetic limitations under current physiological conditions, thus more realistically simulating the metabolic state of an individual at a specific moment. The multi-objective optimization framework reflects the biological reality that cells may simultaneously pursue multiple objectives (such as energy production and antioxidant capacity).

[0021] In some optional embodiments, a pre-trained reinforcement learning agent is constructed based on a multi-objective deep deterministic policy gradient algorithm, wherein: the state space is a metabolic state vector. S ( t Action space A ( t Let be an n-dimensional continuous vector, where each dimension corresponds to the normalized intensity of an intervention measure; the reward function is a multi-objective weighted sum: ,in, , sim c oh S is the cosine similarity calculation function. t arget is the target healthy metabolic state vector. β This is the scaling factor; , MLP c omp Based on user historical compliance characteristics H c omp A small, pre-trained network used to predict user actions. A t The probability of; , c This is the penalty coefficient.

[0022] Through the above technical means, the reinforcement learning framework designed in this application explicitly incorporates three often conflicting objectives: health improvement (encouraging states to move closer to the goal through cosine similarity index rewards), user feasibility (assessing compliance through predictive models), and system security (punishing drastic fluctuations in states). The multi-objective DDPG algorithm enables the agent to learn a series of Pareto optimal policies, providing the possibility of choosing different trade-offs in different scenarios (such as the rapid improvement period vs. the maintenance period), thereby greatly enhancing the practicality and flexibility of decision-making.

[0023] In some optional embodiments, online co-evolutionary updates of the reinforcement learning agent include: incorporating new interaction experiences generated after policy execution ( S ( t ), A ( t ), R r real ( t ), S ( t +1)) Stored into the priority experience replay pool based on non-dominated sorting and crowding distance, where the experience priority , For multi-target TD error vector d Scalar priority of computation r The Pareto front is determined based on the objective value of the experience; the critic network and conditional actor network of the agent are fine-tuned by periodically sampling from the priority experience replay pool; and the parameters of the conditional actor network are periodically perturbed and selected using a population-based multi-objective evolutionary algorithm to explore the Pareto front.

[0024] Through the above technical means, the online update mechanism of this application not only enables the agent to quickly adapt to the user's latest response pattern through priority experience replay and network fine-tuning, but more importantly, it introduces a multi-objective evolutionary algorithm. This algorithm, by performing population-level crossover, mutation, and selection on policy parameters, can actively explore the policy space, discover new and potentially better Pareto solutions, effectively preventing the agent from getting trapped in local optima and ensuring the continuous evolution and long-term effectiveness of decision-making capabilities.

[0025] In some optional embodiments, online co-evolutionary updates of the personalized metabolic flow graph network model include: online incremental learning of the parameters of the relationship-aware heterogeneous graph neural network using newly acquired user data, with the loss function being: ,in, L r reconstruct For data reconstruction loss, L s teady s father For metabolic homeostasis constraint loss, This is a flexible weighting consolidation term used to prevent catastrophic forgetting.

[0026] Through the above technical means, the embodiments of this application cleverly combine data-driven and mechanistic constraints when updating the metabolic network model. The reconstruction loss ensures that the model features can reflect the new data, the steady-state constraint loss provides regularization by utilizing the inherent mass conservation law of the metabolic network, and the elastic weight consolidation term protects the key parameters learned from historically important tasks (i.e., modeling past data), thereby avoiding the "forgetting" of old knowledge while adapting to new changes, and achieving stable and efficient online learning.

[0027] A second aspect embodiment provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the dynamic intervention method for insulin resistance based on metabolic flow graph network and multi-objective reinforcement learning as described in the above embodiment.

[0028] Additional aspects and advantages of this application will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of this application. Attached Figure Description

[0029] The above and / or additional aspects and advantages of this application will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart illustrating a dynamic intervention method for insulin resistance based on metabolic flow graph networks and multi-objective reinforcement learning, provided for some embodiments.

[0030] Figure 2 A flowchart illustrating step S2 provided for some embodiments.

[0031] Figure 3 A flowchart illustrating step S3 provided for some embodiments.

[0032] Figure 4 A flowchart illustrating step S4 provided for some embodiments.

[0033] Figure 5 A flowchart of step S5 provided for some embodiments.

[0034] Figure 6 A schematic diagram of the structure of an electronic device provided for some embodiments. Detailed Implementation

[0035] The embodiments of this application are described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application.

[0036] The following description, with reference to the accompanying drawings, describes a method and device for dynamic intervention of insulin resistance based on metabolic flow graph networks and multi-objective reinforcement learning, according to embodiments of this application. Addressing the technical problems mentioned in the background section regarding the static, universal nature of existing insulin resistance intervention programs and their lack of dynamic perception and closed-loop adaptive capabilities, this application provides an innovative solution. In this solution, firstly, advanced adaptive segmented fusion preprocessing technology is used to transform the user's original multi-source heterogeneous physiological data into a high-quality feature matrix; then, using heterogeneous graph neural networks and flux balance analysis with dynamic constraints, a personalized dynamic metabolic network model whose parameters can learn with individual data is constructed, outputting an interpretable metabolic state vector; next, based on this state vector, a multi-objective reinforcement learning agent generates a personalized intervention strategy that balances health, compliance, and stability; finally, through an online co-evolutionary mechanism, the feedback data after strategy execution is used to simultaneously update the mechanistic model and the decision model, forming a continuously self-optimizing intelligent closed loop. This solves the fundamental problem that existing technologies cannot achieve precise, real-time, adaptive, and interpretable dynamic personalized intervention.

[0037] Specifically, Figure 1 This is a flowchart illustrating a dynamic intervention method for insulin resistance based on metabolic flow graph networks and multi-objective reinforcement learning, provided as an embodiment of this application.

[0038] like Figure 1 As shown, the method includes the following steps: Step S1: Obtain various first-time series data of the user. The first-time series data represents the dynamic basic data of the user's metabolic physiology and biochemistry.

[0039] In this embodiment of the application, the first time-series data is multi-source heterogeneous information reflecting the user's metabolic state. Specifically, this may include: blood glucose data, typically a time series of interstitial glucose concentration (mmol / L) at points every 5-15 minutes; insulin level data: from regular (e.g., weekly) finger-prick or venous blood tests, including fasting insulin and insulin concentrations (μIU / mL) at specific postprandial time points (e.g., 30, 60, 120 minutes); lipid profile and liver and kidney function data: from monthly or quarterly blood tests, including values ​​and timestamps for indicators such as triglycerides (TG), total cholesterol (TC), high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), alanine aminotransferase (ALT), and creatinine (Cr); dietary diary data: including timestamps for each meal, food types, and estimated weights, which can be converted into a sequence of carbohydrate (CHO), fat (FAT), protein (PRO), and dietary fiber (FIBER) intake using a standard food composition database; and exercise and behavior data: including time series of activity types (e.g., walking, running), intensity (expressed in metabolic equivalents METs), duration, resting and exercise heart rates, and sleep stages. This data is collected through an electronic informed consent form, clearly informing users of the purpose, method, scope, and storage period of the data collection. Users must actively authorize the data collection process. The system adheres to the requirements of the "Ethical Review Regulations for Biomedical Research Involving Human Subjects," ensuring that data collection has a clear legal basis (i.e., explicit consent from the user), and that the scope of collection is strictly limited to the minimum necessary for insulin resistance intervention (such as blood glucose levels, dietary records, and exercise levels), refusing to collect irrelevant privacy data beyond the permitted scope.

[0040] During transmission of this raw data to electronic devices or cloud servers, SSL / TLS encryption protocols are used to encrypt the channel, preventing data theft or tampering during transmission. In the storage phase, all sensitive data (including but not limited to users' names, ID numbers, contact information, and specific physiological indicators) are encrypted using AES-256 or equivalent encryption algorithms. Database access is subject to strict hierarchical access control, with access to decrypted data only authorized to specific model training processes and certified medical personnel, ensuring data confidentiality and integrity, and complying with the provisions of the "Data Security Law of the People's Republic of China" regarding the protection of important data. Furthermore, directly identifying personal information (PIIs, such as names and mobile phone numbers) is removed or encrypted from the data, and the remaining physiological time-series data (such as blood glucose concentration and heart rate) is transformed into "de-identified data" that cannot identify specific individuals. In the process of constructing a personalized metabolic flow graph network model based on the de-identified data (step S3) and training the reinforcement learning agent (step S4), the model only learns the metabolic patterns of the group and the dynamic characteristics of individuals, and cannot reverse-engineer the user's true identity. This technical approach effectively implements the anonymization requirements of the Personal Information Protection Law of the People's Republic of China, safeguarding users' personal information rights.

[0041] Step S2: Process the first time series data to obtain the standardized feature matrix.

[0042] Those skilled in the art will understand that directly using raw data for training complex models can lead to performance degradation. This step aims to perform high-quality feature engineering. Figure 2 As shown, it specifically includes the following sub-steps:

[0043] Step S21 involves imputing missing values ​​and correcting outliers in the first time series data to obtain the second time series data. For continuous monitoring data (such as blood glucose), a combination of Time Series Decomposition (STL) and Kalman filtering is used for imputation. STL decomposes the sequence into trend, seasonality, and residuals, while Kalman filtering performs optimal estimation based on a state-space model. The combination of the two can effectively handle time series with noise and missing values. For sparse detection data (such as monthly blood lipids), a personalized Gaussian Process Regression (GPR) model is used for prediction imputation. For each user, a GPR model is trained using their historical detection data. For the time point t* to be imputed, its predicted distribution is as follows: ,in , .here, X Given the known detection time point, y For the corresponding indicator value, k It is a radial basis function (RBF) kernel. s nThis represents the noise variance. This method reflects the non-linear trends of physiological indicators better than linear interpolation. Outlier correction combines Isolation Forest with hard constraints on physiologically reasonable ranges. First, Isolation Forest is used to detect global outliers in the multivariate time-series space. Then, for each univariate sequence, if its value exceeds the clinically recognized physiological limits (e.g., blood glucose concentration below 2.2 mmol / L or above 33.3 mmol / L), it is marked as an outlier. For the detected outliers, the median of normal data within a reasonable time window (e.g., 1 hour) before and after the point is used to replace it to resist the influence of extreme values. The output is the second time-series dataset. D corrected .

[0044] Step S22: Based on the Bayesian change point detection algorithm using an adaptive sliding window, the second time-series data is divided into multiple second data segments. Since metabolic states change due to events such as eating and exercise, it is necessary to... D corrected The data is divided into segments with relatively uniform metabolic states. This application employs an improved Bayesian online change point detection (BOCPD) algorithm based on an adaptive sliding window. The runtime is defined as follows: r t This represents the time elapsed since the last variable point. For each newly arrived data point... x t (here x t It can be a vector of multivariate data or an aggregated scalar), calculate its relationship with the current segment (by...). r t−1 Nonparametric similarity score (defined): .in, K For standard Gaussian kernel function, h For bandwidth, adaptively determined according to the Silverman rule: h =0.9∗ minutes (standard deviation, interquartile range / 1.34)*n −1 / 5 .Score s ( t The higher the value, the better. x t The greater the difference from the current segment, the greater the probability of a change point. The variable-point prior design is a function that increases with the runtime: The parameter λ controls the tendency to "cut" long segments, and is usually set to keep the average segment length around 1-2 hours. When the probability of a change point... Exceeding the dynamic threshold i t At time t, a change point is determined. Threshold i tAdaptive adjustment based on recent change point frequency: ,in i base Use a base threshold (e.g., 0.5). N cp_last_hour This represents the number of change points detected in the past hour. Expected _ N cp The expected frequency of change points (e.g., one every 3 hours) is defined by α, which is an adjustment coefficient. This design allows for smooth detection, avoiding excessive meaningless segmentation during periods of frequent events (e.g., continuous eating). The final output is a series of change point positions {t}. cp1 ,t cp2 ,...,t cpM The timing sequence is divided into M+1 data segments. Seg 1, Seg 2,..., Seg M+1}

[0045] Step S23: Extract time-domain features, frequency-domain features, and nonlinear dynamic features from each second data segment to construct the feature vector of that second data segment. For each data segment... Seg k For each physiological indicator variable (such as blood glucose and heart rate) it contains, three types of features are extracted to form the original feature matrix of this segment. F raw_k ∈ R P×F Where P is the number of indicators and F is the total number of features extracted for each indicator.

[0046] (1) Time-domain characteristics: including mean, standard deviation, coefficient of variation, skewness, kurtosis, minimum, maximum, range, area under the curve (AUC), rising / falling slope (obtained by linear fitting), and coefficients based on autoregressive (AR) models (such as using parameters φ1 and φ2 of AR(2) models).

[0047] (2) Frequency domain features: Perform Fast Fourier Transform (FFT) on the sequence to extract the main frequency (the frequency corresponding to the largest amplitude value in the amplitude spectrum), amplitude spectrum entropy (measuring the degree of disorder in the spectrum), and the energy proportion of three typical physiological frequency bands: very low frequency (VLF, 0.003-0.04Hz, reflecting body temperature regulation, etc.), low frequency (LF, 0.04-0.15Hz, reflecting sympathetic nerve activity), and high frequency (HF, 0.15-0.4Hz, reflecting parasympathetic nerve activity).

[0048] (3) Nonlinear dynamic characteristics: Calculate the approximate entropy (ApEn) to measure the complexity and unpredictability of the time series; calculate the scaling index α of detrended volatility analysis (DFA) to distinguish different modes such as white noise (α≈0.5) and pink noise (α≈1.0, reflecting long-range correlation).

[0049] For a data segment containing P indicators, F features are extracted for each indicator, resulting in... F raw_k To reduce dimensionality and integrate cross-index information, sparse principal component analysis (SPCA) is performed on the matrix. The SPCA optimization problem is as follows: ,satisfy A T A = I k .in x i As a sample, B It is a sparse load matrix. A Let be the projection matrix, and λ control the sparsity. The scores of the first L sparse principal components (with a cumulative variance contribution rate > 80%) are taken as the fusion feature vector for that segment. f k ∈ R L .

[0050] Step S24: Using an attention-based segment feature fusion network, the feature vectors of each second data segment are fused to obtain a standardized feature matrix. Different time periods have different weights influencing the current state. This step designs a multi-head self-attention fusion network. The feature vectors of each segment are... f k Projecting the query onto the linear layer Q k ,key K k' ,value V k Calculate the attention weight of segment k relative to all segments: , where d is the vector dimension. The context-aware representation of segment k is: . All segments c k Mean pooling is performed to obtain C At the same time, from the whole D corrected Calculate global statistical features in the dataset g Examples include 24-hour average blood glucose, total daily calorie intake, and average daily activity level. A gated aggregation layer is used to determine the composition of the final representation: the gated vector. , where [;] indicates concatenation, and σ is the Sigmoid function. The final output is: , where ⊙ represents element-wise multiplication. Finally, for F fused Perform row standardization (z-score standardization) to obtain the final standardized feature matrix. X ∈ R D , where D is the feature dimension. This matrix serves as the input basis for all subsequent modeling steps.

[0051] Step S3: Using the standardized feature matrix and a pre-built general metabolic flow graph network, construct a user-based metabolic flow graph network model and generate a metabolic state vector. The metabolic state vector represents the state that adapts to the differences in the user's metabolic network and provides interpretability.

[0052] like Figure 3 As shown, step S3 is the core mechanism modeling step of this application, which aims to construct a dynamic, parameter-learnable computational model of an individual metabolic system.

[0053] Step S31: Load and represent the general metabolic network topology. The input is a predefined human core metabolic network based on a knowledge base (such as Reactome, KEGG, BiGGModels). This application formalizes it as a heterogeneous directed hypergraph. H =( V,E,R,φ,ψ ).

[0054] V It is a heterogeneous set of nodes, containing various types: metabolites (such as glucose, pyruvate, acetyl-CoA), enzymes / proteins (such as hexokinase, insulin receptor substrate 1IRS-1), hormone signaling molecules (such as insulin, glucagon), transcription factors, and abstract physiological compartment nodes (such as "muscle cytoplasm," "hepatocyte mitochondria," "blood circulation"). Node types are mapped through a mapping function. f : V -> T v definition.

[0055] E It is a set of superedges. A single superedge... e It can connect multiple input nodes (reactants, regulators) to multiple output nodes (products, regulated targets) to precisely represent biochemical reactions (e.g., "glucose + ATP → glucose-6-phosphate + ADP"), transport processes (e.g., "extracellular glucose → intracellular glucose (via GLUT4)"), or regulatory relationships (e.g., "active insulin receptor → phosphorylated IRS-1 (activation)"). Hyperedge types are... ψ : E -> Te definition.

[0056] R It is a set of relation types, such as "catalysis", "product", "substrate", "positive activation", "inhibition", "transportation", etc.

[0057] Each reaction superedge e Associate a stoichiometric vector S e The coefficients of reactants are negative, and the coefficients of products are positive. The stoichiometric vectors of all hyperedges constitute the stoichiometric matrix of the network. S In addition, each edge is associated with a set of kinetic reference parameters, such as the maximum reaction rate. V max_ref Michaelis constant K m Etc. This heterogeneous directed hypergraph H It is a static topological skeleton shared by all users and encapsulates domain knowledge.

[0058] This heterogeneous directed hypergraph is a general topological framework built on metabolic knowledge bases such as Reactome, KEGG, and BiGGModels. Its core feature is the accurate representation of metabolic systems through multiple types of nodes and hyperedges. The node type is defined by the mapping function φ:V→Tv, and the hyperedge type is defined by ψ:E→Te. Each reaction hyperedge is associated with a stoichiometric vector and kinetic reference parameters (Vmax_ref, Km, etc.). The stoichiometric vector is assigned according to the rule of "reactants are negative and products are positive". The stoichiometric vectors of all hyperedges constitute a general stoichiometric matrix S. This topological framework encapsulates the domain prior knowledge of human core metabolism, laying the mechanistic foundation for subsequent personalized learning.

[0059] Step S32: The standardized feature matrix is ​​mapped to the corresponding type nodes of the heterogeneous directed hypergraph H as the initial features of the nodes. The personalized feature representations of the nodes and hyperedges are learned by message passing and node feature updating through a relation-aware heterogeneous graph neural network.

[0060] Feature mapping follows the rule of "one-to-one correspondence between physiological indicators and node types". For example, blood glucose concentration features are mapped to the "blood circulation: glucose" node, and carbohydrate intake features are mapped to the "exogenous glucose input" node. The initial features of nodes without direct data are set as zero vectors or learnable type embedding vectors. The relation-aware heterogeneous graph neural network (RHGNN) adopts a three-level learning mechanism of "relationship-specific message passing - cross-relationship gated fusion - GRU node update". First, the normalized attention weights of neighbor nodes are calculated for each relation type and relation-specific messages are aggregated. Then, the normalized weights are assigned to different relation types and heterogeneous messages are fused through the gated fusion network. Finally, the node features are updated through the gated recurrent unit (GRU) to retain the historical memory of metabolic state. The node features are propagated through L layers (L≥3) to obtain the final personalized representation. The personalized features of the hyperedge are obtained by the aggregation of all node features connected to it through MLP. Finally, the personalized heterogeneous graph is output.

[0061] The normalized feature matrix output from step S2 X Mapping to a hypergraph H On the corresponding type of node, as the initial feature of the node. For example, blood glucose concentration features are mapped to the "blood circulation: glucose" node, and dietary carbohydrate intake features are mapped to the abstract "exogenous glucose input" node. For nodes without directly corresponding data (such as intracellular metabolites), their initial features are set to a zero vector or a learnable type embedding vector. Next, a relational heterogeneous graph neural network (RHGNN) is designed and trained. For each node in the graph... v In the l The layer update process is as follows:

[0062] (1) Relationship-specific message passing: For each relationship with a node v Connected relation types r ∈ R v From all relationships r and v Connected neighbor nodes u ∈ N r (v) Message aggregation. A relational attention mechanism is employed: first, neighbors are calculated. u right v Importance score: ,in a r It is a relationship r Learnable attention vectors It is a relationship r In the lThe learnable transformation matrix of the layer, This represents vector concatenation. Then, for those belonging to the same relation type... r Normalize the importance scores of all neighbors: Finally, relationships r Passed to the node v The message is: .

[0063] (2) Cross-relationship message fusion: Node v You will receive a collection of messages from different relationship types. To fuse these heterogeneous messages, a gated fusion network is used: a gate value is computed for each relation type. ,in U r and b r These are learnable parameters. s It's the Sigmoid function. Then the gate value is normalized: The merged message is: .

[0064] (3) Node feature update: A gated recurrent unit (GRU)-like update mechanism is adopted to combine the features of the node's previous layer with the fused message: . Specifically, , , , This allows node states to retain some historical memory, making it suitable for dynamic process modeling.

[0065] go through L After propagation through layers (e.g., L=3), the final personalized feature representations of all nodes are obtained. Each super edge e Features h e It can be obtained by aggregating the features of all the nodes it connects to, for example: The output is a heterogeneous graph that has undergone personalized learning. H personalized .

[0066] Step S33: Based on the learned hyperedge features, the maximum reaction rate parameter of each reaction hyperedge is estimated individually to obtain personalized kinetic parameters.

[0067] Personalized estimation of the maximum reaction rate achieves a direct mapping from abstract hyperedge features to specific kinetic parameters. Taking the personalized hyperedge feature *he* as input, a small multilayer perceptron (MLPθ, two-layer structure) is used to fit the relationship between the feature and the scaling factor. Then, the scaling factor *scale_e* is constrained to a reasonable range of [0.5, 2.0] by the Sigmoid function. Finally, the personalized maximum reaction rate is calculated according to the formula Vmax_e(t) = Vmax_e_ref × scale_e. This process adapts the core kinetic parameters of the metabolic network to individual metabolic characteristics and is the core step in instantiating a general network into an individual-specific network.

[0068] For each flux superedge representing a biochemical reaction e Utilizing the features it has learned h e Its maximum reaction rate V maxe Perform personalized estimation. Design a parameterized module: scale e = s ( MLP θ ( h e )),in MLP θ It is a small multilayer perceptron (e.g., two layers). s It is the Sigmoid function, which constrains the output within a reasonable range (e.g., [0.5, 2.0]). Therefore, the maximum reaction rate after personalization is: .in, V maxe_ref These are reference values ​​obtained from literature or cell-specific models. This step directly links the abstract features learned by the graph neural network with specific, physiologically meaningful kinetic parameters.

[0069] Step S34: Using the personalized parameters, perform a flux balance analysis with dynamic constraints at time t to calculate the network flux distribution and generate a metabolic state vector by combining the concentrations of key metabolites.

[0070] Dynamically constrained flux balance analysis (dcFBA) incorporates real-time physiological constraints into traditional static FBA. First, it obtains the key metabolite concentration vector c(t) and effector concentration based on the denormalized data of the mapping nodes. Then, it calculates the upper bound vmax(t) of the dynamic flux for each reaction hyperedge using the saturation kinetic formula, with the lower bound vmin(t) set according to the principle of "0 for irreversible reactions and negative minimum for reversible reactions." dcFBA constructs a multi-objective linear programming problem based on a personalized stoichiometric matrix Sp, satisfying the quasi-steady-state constraint Spv=0 and the dynamic capacity constraint vmin(t). If v ≤ vmax(t), the Pareto optimal flux solution set is obtained by combining the NSGA-II multi-objective evolutionary algorithm with a linear programming solver. The representative flux distribution v is selected according to the rule of "closest to the ideal point / maximum preset target weight / most similar to historical flux". The metabolic state vector is formed by concatenating the principal components of v*, the key metabolite concentration vector c(t) and the objective function value vector Z into a high-dimensional original vector. Then, the dimensionality is reduced by a pre-trained autoencoder to obtain a low-dimensional dense final metabolic state vector S(t). This vector is a quantitative and interpretable mathematical code for the current functional state of the individual's metabolic network.

[0071] At a given time t, metabolic flux is calculated using personalized network parameters. This application employs flux balance analysis with dynamic constraints (dcFBA). First, the concentration vectors of key metabolites are obtained based on real-time data (after denormalization) mapped to the nodes. c ( t ), and the concentration of potential effectors (such as insulin).

[0072] For each reaction hyperedge e Its flux v e ( t Dynamic upper bound of ) Calculated using the following adjustable saturation kinetic equation: in,[ S ]( t ) is the substrate concentration, [ I ]( t ) represents the inhibitor concentration (0 if none is present). n and m It is the Hill coefficient. K m It is the Michaelis constant. K i It is the suppression constant. The Hill coefficient can be slightly adjusted based on the learned features. Lower bound It is usually set to 0 (for irreversible reactions) or a negative minimum value (for reversible reactions).

[0073] dcFBA is typically expressed as a (multi-objective) linear programming problem: Subject to : S p * v =0 (quasi-steady-state constraint) v min ( t )≤ v ≤ v max ( t (Dynamic capacity constraints) in, S p It is a personalized stoichiometric matrix (S can be weighted if tissue-specific distribution is considered). v It is the vector of all reaction fluxes. c o It is the first o A coefficient vector of biological targets (e.g., c ATP The position for the ATP synthesis reaction is 1, and the position for the rest is 0; c ROS The reactive oxygen species generation reaction position is -1 to minimize its generation.

[0074] To efficiently solve for and obtain the Pareto front, a multi-objective evolutionary algorithm (NSGA-II) is combined with a local linear programming solver. NSGA-II performs a global search in the flux space to obtain a set of non-dominated solutions. Then, for each non-dominated solution, a linear programming solver such as the simplex method is used for local refinement to ensure that all constraints are satisfied and a local optimum is reached.

[0075] From the Pareto solution set, select the most representative flux distribution vector. v * Selection criteria can be: closest to the ideal point, or based on preset target weights (e.g., assigning weights based on current focus on blood sugar control). c glucose _ uptake The weighted sum of higher weights is the largest, or the distribution is most similar to the historical average flux distribution (in pursuit of stability).

[0076] Finally, a metabolic state vector is generated. A representative flux distribution is then applied. v * Key metabolite concentration vector c ( t ), and the objective function value vector Z The first few principal components are concatenated to form a high-dimensional original state vector.S raw ( t Then, dimensionality reduction is performed using a pre-trained autoencoder (its encoder part) to obtain a low-dimensional, dense final metabolic state vector. S ( t )∈ R K (For example, K=50). S ( t It is a comprehensive, quantitative, and interpretable mathematical code for the functional state of the user's whole-body metabolic network at the current moment.

[0077] Step S4: Construct a pre-trained reinforcement learning agent based on the multi-objective deep deterministic policy gradient algorithm, input the metabolic state vector into the pre-trained reinforcement learning agent, generate an insulin resistance intervention policy vector for the user, and realize personalized decision-making under multi-objective trade-off.

[0078] like Figure 4 As shown, this step is the core of the decision-making process, aiming to generate the optimal personalized intervention strategy based on the current metabolic state. The reinforcement learning agent is constructed based on the Multi-Objective Deep Deterministic Policy Gradient (MO-DDPG) algorithm.

[0079] Step S41, define the problem and agent architecture.

[0080] state space S That is, the metabolic state vector generated in step S3. S ( t ).

[0081] Action space A A continuous n-dimensional space A ( t )=[ a 1, a 2,..., a n ], each dimension a i ∈[0,1] corresponds to the normalized recommendation strength of an intervention. For example: a 1: The percentage of your baseline carbohydrate intake at the next meal; a 2: Daily aerobic exercise intensity (METs level); a 3: The percentage of resistance training duration (in minutes) relative to the maximum recommended value; a 4: Increase dietary fiber intake; a 5: Recommended duration of light activity after meals, etc. In actual output, these normalized values ​​are mapped to specific physical units through inverse transformation.

[0082] Reward function: designed as a weighted sum of three core objectives, but Pareto optimality is sought during training.

[0083] 1. Health Improvement Rewards R health Encourage metabolic state to move towards target healthy state. S target Proximity. Using an exponential form based on cosine similarity: .in, , β It is a scaling factor (e.g., β=2). S target It is a vector learned from metabolic data of healthy individuals or set by experts. The reward is sensitive to the direction of the state vector.

[0084] User compliance rewards R compliance Encourage the generation of solutions that are easy for users to implement. R compliance = MLP comp ( A t , H comp ).in, MLP comp It is a small neural network pre-trained on historical user behavior data to predict user actions. A t The probability of. H comp It is a user's historical compliance characteristic, including the average execution rate of various suggestions in the past week, the self-reported difficulty rating sequence, etc.

[0085] Metabolic stability reward R stability Punishing “aggressive” strategies that may cause drastic fluctuations in metabolic state in order to enhance safety. , c It is the penalty coefficient (e.g., γ=0.1).

[0086] Intelligent agent network architecture: Conditional Actor Network (Actorμ): Parameters i μ The input is the state. S Given a preference weight vector λ=[λ1,λ2,λ3] (λk≥0,Σλk=1), output deterministic actions. A = m ( S , l | i μNetwork structure: Input layer (K+3 dimensional) → Fully connected layer (256 neurons, ReLU) → Fully connected layer (128 neurons, ReLU) → Output layer (n dimensional, Tanh activation) → Sigmoid (constrains the output to [0,1]). Conditional input allows a network to express multiple strategies on the Pareto front.

[0087] Multiple critic networks (Critics{Q_k}): 3 in total, each corresponding to a reward objective. k A network of critics Q k The parameters are i Qk The input is the state. S and actions A Output scalar Q k ( S , A ), representing the target k The expected long-term return. Each Q k The network structure is independent: input layer (K+n dimensional) → fully connected layer (256 neurons, ReLU) → fully connected layer (128 neurons, ReLU) → output layer (1 dimensional).

[0088] Target Network: Actor Target Network m′ and critic target network Q′ k′ Parameters are updated via soft update ( t =0.005) Replicated from the main network: i μ′ ← tth μ +(1− t )θ μ′ , i Q′k ← tth Qk +(1− t ) i Q′k .

[0089] Priority Experience Replay Pool P : Store experience tuples ( S t , A t , R t , S t+1 ),in R t =[ R health , Rcompliance , R stability Its priority is dynamically adjusted based on multi-objective TD error and Pareto level.

[0090] The action space dimensions are set according to intervention types such as diet, exercise, and nutritional supplementation. Each dimension ai∈[0,1], and the actual output is mapped to physical units (grams, minutes, METs, etc.) through inverse transformation. The reward function is a weighted sum of three objectives: health improvement, user compliance, and metabolic stability. Rhealth measures the approximation of the metabolic state to the target health state Starget using cosine similarity and is calculated using the formula Rhealth=exp(βsim_cos(St+1,Starget))-1 (β is a scaling factor, β≥2). Rcompliance is obtained by predicting the action execution probability using MLPcomp pre-trained based on user historical compliance features Hcomp. Rstability penalizes the metabolic state through L2 norm. The state fluctuates drastically, calculated using the formula Rstability=-γ||St+1-St||2 (γ is the penalty coefficient, γ≤0.1); in the agent architecture, the conditional actor network takes the state S and the preference weight vector λ (Σλk=1) as input and outputs a deterministic intervention action. The network has a structure of "input layer → 256-neuron ReLU layer → 128-neuron ReLU layer → Tanh layer → Sigmoid layer"; an independent commentator network is configured for each reward objective, taking the state S and action A as input and outputting the long-term expected return of a single objective; the parameters of the objective network and the main network are synchronized through soft updates (τ=0.005), and the priority experience replay pool is used to store experience tuples (S(t), A(t), Rt, S(t+1)) to provide high-quality samples for subsequent training.

[0091] Step S42, Offline Training. The agent is trained offline in a simulated environment coupled with a personalized metabolic network model. It transforms intervention actions through an action effect model, calculates the next state and reward, and iteratively performs experience collection, network updates, and target network soft updates.

[0092] The agent is trained offline in a simulated environment coupled with the personalized metabolic network model from step S3. The simulated environment receives the state. S t and actions A t Through an "action effect model" A t This is converted into perturbations to metabolic network parameters (such as increasing exogenous glucose input or increasing muscle energy expenditure rate), and then dcFBA is run to calculate the next state. S t+1 And calculate the reward Rt The training process is executed cyclically: collecting experience, calculating priorities and storing them in the experience pool, sampling mini-batch experiences, updating the critic network and actor network, and softly updating the target network. The actor network is updated using a policy gradient based on Chebyshev scalarization. For multiple preference weights λ sampled from a batch of experiences, the loss is calculated separately for each. The gradient descent is then averaged. Through numerous iterations, the agent learns to find Pareto optimal policies under different preferences.

[0093] Step S43, Online Decision Making. During the application phase, for the current user's metabolic state vector... S ( t First, determine the preference weight λ. curr This can be automatically recommended by the system based on the user's recent condition (such as whether they frequently experience high blood sugar), or it can be selected by the doctor or user through the user interface (such as "prioritize improving blood sugar" or "prioritize improving feasibility"). S ( t )and l curr Inputting the actor network, forward propagation yields the intervention policy vector. A ( t Before the final output, it passes through a security filter for inspection. A ( t Is there a high risk involved (e.g., a predicted blood glucose level below 3.9 mmol / L at the next moment)? If a risk is detected, fine-tuning is performed according to preset rules. A ( t (e.g., proportionally reducing the intensity of all interventions), or triggering manual review. The final output... A ( t This refers to personalized intervention strategies.

[0094] Step S5: Output the intervention strategy vector, and based on the new data obtained after strategy execution and user feedback, perform online co-evolutionary updates on the personalized metabolic flow graph network model and the reinforcement learning agent to achieve real-time adaptive adjustment to match dynamic metabolic changes. Specifically, as... Figure 5 Step S5 includes:

[0095] Step S51, convert the action vector A ( t This is decoded into specific suggestions that users can understand. For example, combining a user's basal metabolic rate and daily habits, the carbohydrate intensity is... a The message 1 is translated into "It is recommended that you consume approximately XX grams of brown rice as your staple food in your next meal." The output is a structured daily plan, pushed to your mobile app.

[0096] Step S52: Collect feedback data and new experience data. After the user performs the intervention, the system collects new continuous blood glucose levels, potentially new test data, and user feedback on the degree of implementation and subjective feelings through the application in subsequent time windows (before and after the next meal). These new data are preprocessed using a real-time version of step S2 to obtain a new feature matrix. X new .Will X new Input the current version of the personalized metabolic network model (step S3) to calculate the new metabolic state. S ( t +1). According to S ( t +1) S ( t ), S target And user feedback on compliance ratings, to calculate the actual reward vector received. R real ( t This generates new empirical tuples. S ( t ), A ( t ), R real ( t ), S ( t +1)).

[0097] Step S53, using new data X new With new observations, the parameters of the relation-aware heterogeneous graph neural network are learned incrementally online. To avoid catastrophic forgetting, an elastic weight consolidation (EWC) strategy is employed. The loss function is designed as follows:

[0098] L reconstruct Reconstruction loss encourages the model to reconstruct the input data from the learned node features.

[0099] L steady_state : metabolic homeostasis constraint loss, for example This encourages model parameters to satisfy mass conservation.

[0100] The third item is the EWC item: F i It is a parameter i i Fisher's information diagonal matrix estimation on historical data measures the importance of this parameter to historical tasks. These are the parameter values ​​after training on historical tasks; this penalty significantly modifies important historical parameters. A small number of iterative updates are performed using mini-batch gradient descent to make the metabolic network model more closely reflect the user's latest metabolic response characteristics.

[0101] Step S54: Perform online evolution of the agent.

[0102] (1) Experience storage and priority calculation: new experience ( S ( t ), A ( t ), R real ( t ), S ( t +1) Store in the priority experience replay pool P Calculate the multi-objective TD error vector of this experience. d =[ d 1, d 2, d 3], of which d k = R real,k + c k* Q′ k ( S ( t +1), m′ ( S ( t +1),λ∣ i μ′ ))− Q k ( S ( t ), A ( t )). Calculate scalarization priority. p At the same time, based on this experience, the target value among all experiences ( R real + c * Q Perform a non-dominated sort to obtain its Pareto front level. r (Level 1 is the Pareto optimal frontier). The final adjusted sampling priority is... p adjusted = p / (1+ r This makes it more likely that experiences with higher Pareto levels will be sampled.

[0103] (2) Network fine-tuning: Periodically (e.g., daily) sample small batches of experience from the priority pool and perform several rounds of gradient descent updates on the critic network and conditional actor network according to the MO-DDPG update rules.

[0104] (3) Strategy Exploration and Evolution: To continuously explore the Pareto frontier, a small-scale population-based evolutionary strategy is adopted. A small population of actor network parameters (e.g., 5 individuals) is maintained. Crossover and mutation operations are performed on individuals in the population at regular intervals (e.g., weekly) (e.g., adding Gaussian noise to the parameters). The multi-objective performance (fitness) of each mutated individual is evaluated in an updated simulation environment (using a fine-tuned metabolic network model). The NSGA-II selection mechanism is used to select superior individuals to form a new generation population based on non-dominated ranking and crowding distance. The parameters of the best individuals in the new generation population are fused with the main actor network parameters (e.g., Polyak average). This allows the newly discovered strategies to be introduced into the main network.

[0105] Through the co-evolution of steps S53 and S54, both the metabolic network model (perception engine) and the reinforcement learning agent (decision engine) can improve themselves using new feedback data, enabling the entire system to adapt to the long-term drift of users' metabolic characteristics and changes in behavioral patterns, forming a truly self-learning, adaptive, and long-term effective personalized health management closed loop.

[0106] According to the method proposed in this application, a complete intelligent intervention system is constructed through the close coupling of four stages: innovative data preprocessing, personalized dynamic metabolic modeling, multi-objective reinforcement learning decision-making, and online co-evolutionary updating. Furthermore, through analysis of the interpretability reports generated by the system, doctors assessed that over 85% of the reports provided reasonable or highly reasonable biological justifications, significantly improving trust in AI decision-making.

[0107] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. The electronic device may include: a memory 601, a processor 602, and a computer program stored in the memory 601 and executable on the processor 602.

[0108] When the processor 602 executes the program, it implements the dynamic intervention method for insulin resistance based on metabolic flow graph network and multi-objective reinforcement learning provided in the above embodiments.

[0109] Furthermore, the electronic device also includes a communication interface 603 for communication between the memory 601 and the processor 602.

[0110] Memory 601 is used to store computer programs that can run on processor 602. Memory 601 may include high-speed random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0111] If the memory 601, processor 602, and communication interface 603 are implemented independently, then the communication interface 603, memory 601, and processor 602 can be interconnected via a bus to complete communication between them. The bus can be an Industry Standard Architecture (ISA) bus, a PCI bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into address bus, data bus, control bus, etc. For ease of representation, Figure 6 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0112] Optionally, in a specific implementation, if the memory 601, processor 602, and communication interface 603 are integrated on a single chip, then the memory 601, processor 602, and communication interface 603 can communicate with each other through an internal interface.

[0113] Processor 602 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present application.

[0114] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or more of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0115] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium. When executed, the program includes one or a combination of the steps of the method embodiments.

[0116] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0117] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. A method for generating a dynamic intervention strategy for insulin resistance, characterized in that, include: Acquire the user's first time-series data, which represents the dynamic basic data of the user's metabolic physiology and biochemistry. The first time series data is processed to obtain a standardized feature matrix; By utilizing a standardized feature matrix and a pre-built general metabolic flow graph network, a user-based metabolic flow graph network model is constructed and a metabolic state vector is generated. The metabolic state vector represents the state that adapts to the differences in user metabolic networks and provides interpretability. A pre-trained reinforcement learning agent is constructed based on a multi-objective deep deterministic policy gradient algorithm. The metabolic state vector is input into the pre-trained reinforcement learning agent to generate an insulin resistance intervention policy vector for the user, thereby realizing personalized decision-making under multi-objective trade-offs. The system outputs an intervention strategy vector and, based on new data and user feedback obtained after the strategy execution, performs online co-evolutionary updates on the personalized metabolic flow graph network model and the reinforcement learning agent to achieve real-time adaptive adjustment to match dynamic metabolic changes.

2. The method according to claim 1, characterized in that, The first time series data is processed to obtain a standardized feature matrix, including: Missing values ​​were imputed and outliers were corrected in the first time series data to obtain the second time series data; The Bayesian variable point detection algorithm based on adaptive sliding window divides the second time series data into multiple second data segments; Extract time-domain features, frequency-domain features, and nonlinear dynamic features from each second data segment to form the feature vector of that second data segment; By using an attention-based segment feature fusion network, the feature vectors of each second data segment are fused to obtain a standardized feature matrix.

3. The method according to claim 2, characterized in that, The Bayesian change point detection algorithm based on an adaptive sliding window divides the second time series data into multiple second data segments, including: Define runtime r t This represents the time elapsed since the current point was the last variable point. At each new data point x t Upon arrival, calculate its non-parametric similarity score with the current segment. ,in K For Gaussian kernel function, h For bandwidth; Calculate the probability of a change point Among them, the variable point prior λ is the attenuation rate parameter; When the probability of a change point exceeds the dynamic threshold θ t When determining that a change point occurs at time t, where , used for smooth detection.

4. The method according to claim 1, characterized in that, Construct a user-based metabolic flow graph network model and generate metabolic state vectors, including: Load the general metabolic flow graph network topology, which is formalized as a heterogeneous directed hypergraph H=(V,E,R), where V is a heterogeneous set of nodes containing metabolite nodes, enzyme nodes, hormone signaling molecule nodes and abstract nodes of physiological compartments, E is a set of hyperedges representing biochemical reactions, transport processes or regulatory relationships, and R is a set of relationship types between nodes and hyperedges. The standardized feature matrix is ​​mapped to the corresponding type node of the heterogeneous directed hypergraph H as the initial feature of the node. The message passing and node feature update are performed through the relation-aware heterogeneous graph neural network to learn the personalized feature representation of the node and the hyperedge. Based on the learned hyperedge features, the maximum reaction rate parameter of each reaction hyperedge is estimated individually. Using the personalized parameters, a flux balance analysis with dynamic constraints is performed at time t to calculate the network flux distribution and generate a metabolic state vector by combining the concentrations of key metabolites.

5. The method according to claim 4, characterized in that, Message passing and node feature updating are performed using a relation-aware heterogeneous graph neural network, including: For nodes v and relation types r Calculate from neighboring nodes Relationship-specific messages: ,in The normalized attention weights are calculated based on the relational attention mechanism. Learnable linear transformations specific to relations. Features of the neighboring nodes at the next higher level; Merging messages from different relationship types: ,in β r To calculate the relation type weights using a gated fusion network; Update node features using gated loop units: .

6. The method according to claim 4, characterized in that, The maximum reaction rate parameter for each reaction hyperedge is estimated individually, specifically as follows: , in, For reference maximum value, , MLP θ It is a multilayer perceptron. h e Let σ be the personalized feature representation of the reaction hyperedge, and let σ be the Sigmoid function.

7. The method according to claim 4, characterized in that, The optimization problem for flux balance analysis with dynamic constraints is as follows: in, c o For the first o A coefficient vector of biological targets S p For personalized stoichiometry matrix, v It is the flux vector. v m in (t) and v m ax(t) For dynamic flux boundary, dynamic flux boundary v m ax(t) The calculation formula is: ， in,[ S ]( t )and[ I ]( t The substrate concentration and effector concentration at time t are respectively. n , m The Hill coefficient is used. K m It is the Michaelis constant. K i This is the suppression constant.

8. The method according to claim 1, characterized in that, A pre-trained reinforcement learning agent is constructed based on a multi-objective deep deterministic policy gradient algorithm, wherein: The state space is a metabolic state vector. S ( t ); Action space A ( t ) is an n-dimensional continuous vector, where each dimension corresponds to the normalized intensity of an intervention measure; The reward function is a multi-objective weighted sum: in, , sim_cos This is the cosine similarity calculation function. Starget Let the target healthy metabolic state vector be... β This is the scaling factor; , MLP c omp Based on user historical compliance characteristics H c omp A small, pre-trained network used to predict user actions. A t The probability of; , γ is Penalty coefficient.

9. The method according to claim 8, characterized in that, Online co-evolutionary updates of reinforcement learning agents, including: The new interactive experiences generated after the strategy is executed S ( t ), A ( t ), R r eal ( t ), S ( t +1)), S ( t+ 1) Store the data in a priority experience replay pool based on non-dominated sorting and crowding distance, where the experience priority... , For multi-target TD error vector δ Scalar priority of computation r This experience is based on the Pareto frontier rank of the target value; Periodically sample from the priority experience replay pool to fine-tune the agent's critic network and conditional actor network; A population-based multi-objective evolutionary algorithm is used to periodically perturb and select the parameters of a conditional actor network in order to explore the Pareto front. This includes online co-evolutionary updates of personalized metabolic flow graph network models, including: The parameters of a relation-aware heterogeneous graph neural network are learned online incrementally using newly acquired user data, with the loss function being: in, L r econstruct For data reconstruction loss, L s teady s tate For metabolic homeostasis constraint loss, This is a flexible weighting consolidation term used to prevent catastrophic forgetting.

10. An electronic device, characterized in that, include: The memory, the processor, and the computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the dynamic intervention method for insulin resistance based on metabolic flow map network and multi-objective reinforcement learning as described in any one of claims 1-9.

Images