Dynamic health adaptive monitoring method and system using artificial intelligence

Abstract

The present invention provides a dynamic health adaptive monitoring method and system using artificial intelligence, which relates to the technical field of health monitoring, including: using multi-source heterogeneous health data, using a hierarchical diffusion probability model to reduce noise and combining with an improved attention autoencoder to extract preliminary features, then learning discriminative features through a contrast learning network, inputting the discriminative features into an adaptive wavelet neural network for multi-scale feature decomposition, and then inputting them into a dynamic graph neural network to obtain a health feature tensor, using a hierarchical memory enhancement network to obtain time series features, and combining a dynamic routing capsule network and a causal reasoning network to generate a health knowledge graph, inputting the time series features and the health knowledge graph into a deep Bayesian network for health risk assessment, using a hierarchical reinforcement learning network, combining a multi-agent collaborative optimization system and a deep deterministic policy gradient network, and obtaining an adaptive monitoring solution through a model predictive control network.

Description

Technical Field

[0001] The present invention relates to the technical field of health monitoring, and in particular to a dynamic health adaptive monitoring method and system using artificial intelligence. Background Art

[0002] Health monitoring is an important part of the healthcare field, which aims to track an individual's health status and identify potential health risks. With the advancement of technology, dynamic health monitoring methods have become possible, which can continuously collect and analyze health data, providing a more comprehensive and real-time understanding of an individual's health status;

[0003] Traditional health monitoring methods usually rely on manual data collection and regular physical examinations, which can be time-consuming and inefficient. In addition, traditional health monitoring methods may not be able to capture the dynamic changes in an individual's health status, leading to delayed detection and intervention of health problems. Existing AI-based health monitoring methods have problems such as being unable to handle multi-source heterogeneous health data, lacking consideration of time dependency and causality, and being unable to provide personalized monitoring solutions.

[0004] Therefore, a solution is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a dynamic health adaptive monitoring method and system using artificial intelligence, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a dynamic health adaptive monitoring method using artificial intelligence, comprising:

[0007] Obtain multi-source heterogeneous health data corresponding to the user, define the forward diffusion process and add noise through a hierarchical diffusion probability model based on continuous-time stochastic differential equations, solve the differential equations in combination with the backward diffusion process for noise reduction, generate noise-reduced data and add it to the improved attention autoencoder, calculate the correlation weights between features through a multi-head self-attention mechanism and encode them as potential representations, optimize network parameters in combination with reconstruction errors to generate preliminary features and add them to the contrastive learning network and perform random enhancement, generate positive sample pairs and learn discriminative features through a contrastive loss function, add the discriminative features to an adaptive wavelet neural network, perform multi-scale feature decomposition by dynamically selecting the optimal wavelet basis function and the number of decomposition layers, obtain multi-scale features and add them to a dynamic graph neural network, extract node representations in combination with an improved graph convolution operator and update edge weights in combination with a gated attention mechanism to obtain a health feature tensor;

[0008] The health feature tensor is added to the hierarchical memory enhancement network, and the feature information is stored by constructing a multi-layer memory unit with different time spans. The information flow is modified and the long-term and short-term dependencies are determined in combination with an adaptive gating mechanism to obtain the time series feature and add it to the dynamic routing capsule network. The time series feature is hierarchically expressed in combination with an improved iterative routing algorithm to obtain the hierarchical feature and add it to the causal reasoning network. The dependency between different health influencing factors is defined by a structural equation model and the causal hypothesis is verified in combination with a reverse intervention mechanism to obtain a verified causal relationship graph. The conditional random field in the probability graph attention network is combined to construct the conditional probability distribution between nodes and calculate the node importance weights to generate a health knowledge graph. The time series feature and the health knowledge graph are added to a deep Bayesian network, and the posterior distribution is approximated by variational inference and the uncertainty of the prediction result is estimated in combination with Monte Carlo sampling to obtain a health risk assessment result.

[0009] The health risk assessment results are added to the hierarchical reinforcement learning network. By constructing a hierarchical policy structure, the upper network generates abstract goals to guide the lower network to select specific actions. The selected candidate action plans are input into the multi-agent collaborative optimization system. The collaborative optimization of group decision-making is achieved by designing the communication protocol and reward distribution mechanism between agents. The optimized strategy is input into the deep deterministic policy gradient network. The deterministic policy function is combined with the value function and the policy parameters are optimized by the gradient ascent method to achieve online learning. The learned policy parameters are input into the model predictive control network. The future state sequence is predicted by the rolling time domain optimization method and the control strategy is dynamically adjusted to obtain an adaptive monitoring solution.

[0010] In an optional embodiment,

[0011] Obtain the multi-source heterogeneous health data corresponding to the user, define the forward diffusion process and add noise through the hierarchical diffusion probability model based on continuous-time stochastic differential equations, solve the differential equations in combination with the backward diffusion process for denoising, generate denoised data and add it to the improved attention autoencoder, calculate the correlation weights between features through the multi-head self-attention mechanism and encode them as potential representations, optimize the network parameters in combination with the reconstruction error to generate preliminary features and add them to the contrastive learning network and perform random enhancement, generate positive sample pairs and learn discriminative features through the contrastive loss function, add the discriminative features to the adaptive wavelet neural network, perform multi-scale feature decomposition by dynamically selecting the optimal wavelet basis function and the number of decomposition layers, obtain multi-scale features and add them to the dynamic graph neural network, extract node representations in combination with the improved graph convolution operator and update edge weights in combination with the gated attention mechanism, and obtain the health feature tensor including:

[0012] The continuous data of electrocardiogram signals, blood pressure values, and blood oxygen concentration and the discrete data of movement status and sleep quality are collected through the sensor array of the wearable device, the continuous data and the discrete data are time-aligned and format-unified, the data at different time intervals are resampled into a standardized multidimensional time series, the missing values ​​in the multidimensional time series are supplemented by a combination of forward filling and linear interpolation, and the abnormal values ​​in the multidimensional time series are processed by sliding window median filtering to obtain preprocessed health data;

[0013] Input the preprocessed health data into a hierarchical diffusion probability model, gradually inject Gaussian noise with increasing variance into the health data during the forward diffusion process, and construct a noise prediction network based on a U-shaped structure. The noise prediction network extracts multi-scale features and reconstructs detail information through a downsampling path and an upsampling path, respectively, adds a jump connection between the corresponding scales of the downsampling path and the upsampling path, predicts the distribution of the Gaussian noise through iterative optimization, and removes the noise during the reverse diffusion process to obtain the denoised health data;

[0014] Input the denoised health data into an improved attention autoencoder, which maps the input data to a high-dimensional feature space through a multi-layer perceptron and then divides it into multiple attention heads, each attention head calculates a query vector, a key vector and a value vector respectively, calculates the dot product of the query vector and the key vector to obtain an attention score, performs a weighted sum of the attention score and the value vector to obtain a context representation, connects the context representations of multiple attention heads and obtains a feature representation through a linear transformation, reconstructs the original data through a decoder and optimizes the network parameters to obtain the encoded health features;

[0015] Input the encoded health feature into a contrastive learning network, perform random time window cropping, amplitude modulation and random noise injection on the health feature to obtain an enhanced feature view, construct different feature views of the same health feature as a positive sample pair, construct feature views of different health features as a negative sample pair, constrain the distance of the positive sample pair in the feature space to be less than the distance of the negative sample pair through a contrast loss function, and obtain a discriminative health feature;

[0016] Input the discriminative health feature into an adaptive wavelet neural network, evaluate the expressive power of different wavelet basis functions based on the time-frequency characteristics of the signal and select the optimal basis function family, determine the number of decomposition layers according to the frequency composition of the signal, perform multi-scale decomposition on the health feature to obtain sub-band coefficients of different frequency bands, perform feature fusion on the sub-band coefficients at each scale to obtain a multi-scale health feature;

[0017] The multi-scale health features are constructed as a dynamic graph structure, and different health indicators are used as nodes in the graph. The neighborhood information of the nodes is aggregated through the graph convolution operator of the adaptive weight matrix, and the weight matrix is ​​dynamically updated based on the feature similarity between the nodes. The connection weights of the edges are adjusted according to the correlation strength between the nodes through the gated attention mechanism, and the node features extracted by the graph convolution are deeply fused with the edge features learned by the attention mechanism to output the health feature tensor.

[0018] In an optional embodiment,

[0019] The discriminative health features are input into the adaptive wavelet neural network, the expression capabilities of different wavelet basis functions are evaluated based on the time-frequency characteristics of the signal and the optimal basis function family is selected, the number of decomposition layers is determined according to the frequency composition of the signal, the health features are multi-scale decomposed to obtain sub-band coefficients of different frequency bands, and the sub-band coefficients are feature fused at each scale to obtain multi-scale health features including:

[0020] Inputting discriminative health features into an adaptive wavelet neural network, wherein the adaptive wavelet neural network comprises a wavelet basis function evaluation module, a decomposition layer number determination module and a feature fusion module;

[0021] The wavelet basis function evaluation module constructs a candidate wavelet basis function library, which includes a Haar wavelet basis function in a piecewise constant form, a Dobesi wavelet basis function family with different vanishing moment orders, a Schindler wavelet basis function family with different filter lengths, a Koivulet wavelet basis function family with symmetry, and a Mayer wavelet basis function with a smooth transition band in the frequency domain;

[0022] Performing time-frequency analysis on the health characteristics to obtain frequency domain characteristics, and calculating a reconstruction error evaluation value, a sparsity evaluation value, and a detail preservation evaluation value of a wavelet basis function based on the frequency domain characteristics;

[0023] Establishing a scoring matrix, normalizing the reconstruction error evaluation value, the sparsity evaluation value and the detail preservation evaluation value to obtain a normalized evaluation value, setting an adaptive weight coefficient based on the frequency domain feature, weighting the normalized evaluation value and the adaptive weight coefficient to obtain an expression ability score of each wavelet basis function, sorting the wavelet basis functions according to the expression ability score and selecting the wavelet basis function with the highest score as the optimal basis function;

[0024] The decomposition layer number determination module performs frequency analysis on the health feature to obtain frequency composition information, and determines the wavelet decomposition layer number according to the number of main frequency components and frequency band distribution in the frequency composition information;

[0025] Based on the optimal basis function and the wavelet decomposition layer number, the health feature is decomposed at multiple scales to obtain sub-band coefficients of different frequency bands, and the feature fusion module performs feature fusion on the sub-band coefficients of different frequency bands at various scales to obtain multi-scale health features;

[0026] Monitor the degree of change of the frequency domain characteristics, the rate of change of statistical characteristics and the fluctuation of reconstruction quality. When the degree of change exceeds a preset threshold, reselect the optimal basis function. When the reconstruction quality is lower than a preset reconstruction quality threshold, randomly initialize the adaptive weight coefficient.

[0027] In an optional embodiment,

[0028] The health feature tensor is added to the hierarchical memory enhancement network, and the feature information is stored by constructing a multi-layer memory unit with different time spans. The information flow is modified and the long-term and short-term dependencies are determined by combining the adaptive gating mechanism, and the time series features are obtained and added to the dynamic routing capsule network. The time series features are hierarchically expressed by combining the improved iterative routing algorithm, and the hierarchical features are obtained and added to the causal reasoning network. The dependency between different health influencing factors is defined by the structural equation model and the causal hypothesis is verified by combining the reverse intervention mechanism to obtain a verified causal relationship graph. The conditional probability distribution between nodes is constructed by combining the conditional random field in the probability graph attention network and the node importance weight is calculated to generate a health knowledge graph. The time series features and the health knowledge graph are added to the deep Bayesian network, and the posterior distribution is approximated by variational inference and the uncertainty of the prediction results is estimated by combining Monte Carlo sampling. The health risk assessment results include:

[0029] The healthy feature tensor is input into the normalization module for normalization processing to obtain a standardized feature, and the standardized feature is input into the hierarchical memory enhancement network. The current input feature reception ratio is regulated by the input gating value of the short-term memory unit, the historical state retention degree is adjusted by the forgetting gating value, and the feature output strength is regulated by the output gating value to generate a short-term feature. The short-term feature is input into the medium-term memory unit and a query vector is generated. The similarity score between the query vector and the short-term feature is calculated, and the similarity score is normalized to obtain the attention weight. The short-term feature is multiplied by the attention weight to generate a medium-term feature. The short-term feature and the medium-term feature are combined by the residual connection of the long-term memory unit and a gated update is performed to generate a long-term feature. The change rate of the short-term feature, the medium-term feature and the long-term feature is calculated, and the state update ratio of the corresponding memory unit is adjusted according to the change rate by using the adaptive gating mechanism and the long-term dependency is determined, and the time series feature is generated by fusion.

[0030] Input the time series features into multiple dedicated capsules of the dynamic routing capsule network, wherein each of the dedicated capsules corresponds to a specific health influencing factor, perform feature extraction on the time series features through a convolution transformation matrix and generate an activity vector, calculate the modulus length of the activity vector to represent the probability of feature existence, use the direction of the activity vector to represent the feature attribute to obtain a primary feature representation, initialize the routing coefficient matrix to be uniformly distributed, multiply the primary feature representation with the transformation matrix to generate a prediction vector, calculate the inner product of the prediction vector and the high-level capsule feature vector to obtain similarity, perform softmax normalization on the similarity to update the routing coefficient, perform weighted summation of the updated routing coefficient and the prediction vector to generate a feature vector of the high-level capsule, and repeat the update until the routing coefficient converges or reaches a maximum number of iterations to generate a hierarchical feature;

[0031] Input the hierarchical features into the causal reasoning network, calculate the correlation coefficients between different health influencing factors based on the hierarchical features to construct a correlation matrix, use the structural equation model to decompose the correlation matrix into a direct effect matrix and an indirect effect matrix to construct an initial causal graph, calculate the conditional mutual information and Fisher conditional independence statistics for each pair of nodes with a connection relationship, screen the direct causal relationship that meets the conditional independence hypothesis according to the significance level, uniformly sample within the value range of the target node to generate multiple groups of intervention values, substitute each group of intervention values ​​into the causal graph and calculate the response values ​​of other nodes except the target node based on the structural equation model, calculate the mean square error between the response value and the historical data, remove the corresponding causal edge when the mean square error exceeds a preset threshold, and repeatedly calculate and delete the causal edge until the causal graph structure is stable to generate a verified causal relationship graph;

[0032] Input the verified causal relationship graph into the probabilistic graph attention network, define a discrete state space for each node and count the state transition frequency between adjacent time steps, use maximum likelihood estimation to calculate the state transition probability to construct a transfer matrix, calculate the joint distribution probability between any two nodes based on the transfer matrix, calculate the corresponding conditional probability distribution for each node to represent the current node's dependence on other nodes, calculate the importance score of the node according to the conditional probability distribution and perform softmax normalization to obtain the attention weight, multiply the node feature with the corresponding attention weight and aggregate to generate a health knowledge graph that takes into account the importance of the node;

[0033] The time series features and the health knowledge graph are input into the deep Bayesian network, the mean vector and covariance matrix of the variational distribution are initialized, the variational objective function is constructed including the log-likelihood term and the KL divergence term, the gradient of the variational objective function with respect to the variational parameters is calculated, the variational parameters are iteratively optimized using stochastic gradient descent until convergence to obtain the approximate posterior distribution, multiple Monte Carlo samplings are performed from the approximate posterior distribution to obtain model parameter samples, the corresponding model prediction output is calculated for each group of parameter samples, the mean of all model prediction outputs is statistically calculated as the final prediction result, and the variance of the model prediction output is calculated to evaluate the uncertainty, and a health risk assessment result including a risk level and a confidence interval is generated.

[0034] In an optional embodiment,

[0035] Input the verified causal relationship graph into the probabilistic graph attention network, define a discrete state space for each node and count the state transition frequency between adjacent time steps, calculate the state transition probability using maximum likelihood estimation to construct a transfer matrix, and calculate the joint distribution probability between any two nodes based on the transfer matrix, including:

[0036] Inputting the verified causal relationship graph into the probabilistic graph attention network, the verified causal relationship graph includes multiple nodes representing different health influencing factors, determining the number of state categories of each node according to clinical medicine standards, dividing the indicator value of each node into multiple state intervals, establishing a mapping relationship from the indicator value to the state identifier, and generating a discrete state space of the node;

[0037] An observation time window is set based on the discrete state space, and state transition statistics are independently performed on each of the nodes within the observation time window. The state of the node at each time step is identified by traversing the time series data, and the state transition of the node between adjacent time steps is recorded to create a state transition frequency matrix, wherein the rows and columns of the state transition frequency matrix correspond to all possible states of the node respectively;

[0038] Sum all elements of each row in the state transition frequency matrix to obtain the total number of transition events of the corresponding state, divide each element in the state transition frequency matrix by the total number of transition events in its row to obtain a normalized transition probability, repeat the normalization operation for each row of the state transition frequency matrix to generate a transition probability matrix;

[0039] Extract the transition probability matrices of any two nodes, combine the two transition probability matrices to obtain a joint probability matrix representing the combination of the two node states, repeat the combination operation for all node pairs in the causal relationship graph, and generate the joint distribution probability between all node pairs in the causal relationship graph.

[0040] In an optional embodiment,

[0041] The health risk assessment results are added to the hierarchical reinforcement learning network. By constructing a hierarchical strategy structure, the upper network generates abstract goals to guide the lower network to select specific actions. The selected candidate action plans are input into the multi-agent collaborative optimization system. The collaborative optimization of group decision-making is achieved by designing the communication protocol and reward distribution mechanism between agents. The optimized strategy is input into the deep deterministic policy gradient network. The deterministic policy function is combined with the value function and the policy parameters are optimized by the gradient ascent method to achieve online learning. The learned policy parameters are input into the model predictive control network. The future state sequence is predicted by the rolling time domain optimization method and the control strategy is dynamically adjusted. The adaptive monitoring scheme includes:

[0042] Adding the health risk assessment results to a hierarchical reinforcement learning network, wherein the hierarchical reinforcement learning network generates an abstract goal in the upper network to guide the lower network to select specific actions by constructing a hierarchical strategy structure, wherein the hierarchical strategy structure includes a state encoding module, a goal generation module and a strategy mapping module, wherein the state encoding module encodes the risk level, abnormality degree and change trend in the health risk assessment results into a state vector, the goal generation module generates an abstract goal of monitoring frequency and warning threshold based on the state vector, and the strategy mapping module converts the abstract goal into a candidate action plan;

[0043] The candidate action plan is input into a multi-agent collaborative optimization system, the multi-agent collaborative optimization system sets a communication protocol and a reward allocation mechanism between agents, the communication protocol defines the transmission rules of status information messages, action suggestion messages and evaluation feedback messages, the reward allocation mechanism allocates a reward value to each agent according to data collection quality, prediction accuracy and strategy execution effect, and realizes collaborative optimization of group decision-making based on the communication protocol and the reward allocation mechanism;

[0044] The optimized strategy is input into a deep deterministic policy gradient network, which combines a deterministic policy function with a value function. The deterministic policy function outputs a specific monitoring action, and the value function evaluates the long-term value of the monitoring action. The policy parameters of the deterministic policy function are continuously optimized by a gradient ascent method, and the policy parameters are input into a model predictive control network. The model predictive control network predicts the future state sequence by a rolling horizon optimization method, searches for the optimal control sequence within the prediction range based on the future state sequence, executes the control action at the current moment and receives state feedback, dynamically adjusts the policy parameters, and generates an adaptive monitoring solution.

[0045] In an optional embodiment,

[0046] The model predictive control network predicts the future state sequence by a rolling horizon optimization method, searches for the optimal control sequence within the prediction range based on the future state sequence, executes the control action at the current moment and receives state feedback, dynamically adjusts the strategy parameters, and generates an adaptive monitoring scheme, including:

[0047] Constructing a basic architecture of a dynamic prediction model, the dynamic prediction model includes an input layer, a data preprocessing module, a feature extraction layer and a state encoding layer, the input layer receives a physiological indicator numerical sequence, a monitoring time series and an environmental parameter, the data preprocessing module performs standardization processing on the physiological indicator numerical sequence, the monitoring time series and the environmental parameter to obtain normalized data, the feature extraction layer extracts time series features from the normalized data through a bidirectional long short-term memory network, and the state encoding layer encodes the time series features into a state vector of fixed dimension;

[0048] Generate a future state sequence through a prediction controller based on the state vector, input the state vector into a prediction model to obtain a state prediction value at the next moment, and use the state prediction value as a new input to continue predicting to obtain a complete state sequence in a prediction time domain, wherein the prediction time domain is divided into a plurality of prediction steps, and each prediction step corresponds to a state prediction value;

[0049] The optimal control sequence is determined by adopting a hierarchical search strategy, wherein the first layer determines the reference value of the monitoring frequency according to the fluctuation degree of the complete state sequence, the second layer performs a local search near the reference value to obtain the optimized monitoring frequency, and the third layer dynamically adjusts the value range of the warning threshold based on the complete state sequence, and the optimal control sequence is determined by evaluating the comprehensive effects of different control schemes through a heuristic algorithm;

[0050] Sending the monitoring frequency and early warning threshold in the optimal control sequence to the monitoring device to start data collection, recording actual monitoring data, wherein the actual monitoring data includes a monitoring value sequence, a monitoring time sequence, and an early warning trigger record, calculating the error between the state prediction value and the actual monitoring data, constructing a loss function based on the error and using the stochastic gradient descent method to update the parameters of the dynamic prediction model, and triggering the adjustment of the dynamic prediction model structure when the error exceeds a set threshold;

[0051] The control strategy is re-optimized based on the updated dynamic prediction model, the monitoring frequency and the warning threshold are adaptively adjusted according to the weight relationship between the prediction accuracy and the monitoring burden, and the adjusted control strategy is used for the optimization process of the next prediction cycle.

[0052] A second aspect of an embodiment of the present invention provides a dynamic health adaptive monitoring system using artificial intelligence, including:

[0053] The first unit is used to obtain multi-source heterogeneous health data corresponding to the user, define the forward diffusion process and add noise through the hierarchical diffusion probability model based on continuous-time stochastic differential equations, solve the differential equations in combination with the backward diffusion process for noise reduction, generate noise-reduced data and add it to the improved attention autoencoder, calculate the correlation weights between features through the multi-head self-attention mechanism and encode them as potential representations, optimize the network parameters in combination with the reconstruction error to generate preliminary features and add them to the contrastive learning network and perform random enhancement, generate positive sample pairs and learn discriminative features through the contrastive loss function, add the discriminative features to the adaptive wavelet neural network, perform multi-scale feature decomposition by dynamically selecting the optimal wavelet basis function and the number of decomposition layers, obtain multi-scale features and add them to the dynamic graph neural network, extract node representations in combination with the improved graph convolution operator and update edge weights in combination with the gated attention mechanism to obtain a health feature tensor;

[0054] The second unit is used to add the health feature tensor to the hierarchical memory enhancement network, store feature information by constructing multi-layer memory units with different time spans, modify information flow and determine long-term and short-term dependencies in combination with an adaptive gating mechanism, obtain time series features and add them to the dynamic routing capsule network, express the time series features in a hierarchical manner in combination with an improved iterative routing algorithm, obtain hierarchical features and add them to the causal reasoning network, define the dependencies between different health influencing factors through a structural equation model and verify the causal hypothesis in combination with a reverse intervention mechanism, obtain a verified causal relationship graph, construct the conditional probability distribution between nodes in combination with the conditional random field in the probability graph attention network and calculate the node importance weights, generate a health knowledge graph, add the time series features and the health knowledge graph to a deep Bayesian network, approximate the posterior distribution through variational reasoning and estimate the uncertainty of the prediction results in combination with Monte Carlo sampling, and obtain a health risk assessment result;

[0055] The third unit is used to add the health risk assessment results to the hierarchical reinforcement learning network, construct a hierarchical policy structure so that the upper network generates abstract goals to guide the lower network to select specific actions, input the selected candidate action plans into the multi-agent collaborative optimization system, and realize the collaborative optimization of group decision-making by designing the communication protocol and reward distribution mechanism between agents. The optimized strategy is input into the deep deterministic policy gradient network, the deterministic policy function is combined with the value function and the policy parameters are optimized by the gradient ascent method to realize online learning, the learned policy parameters are input into the model predictive control network, and the future state sequence is predicted by the rolling time domain optimization method and the control strategy is dynamically adjusted to obtain an adaptive monitoring solution.

[0056] According to a third aspect of the embodiments of the present invention,

[0057] An electronic device is provided, comprising:

[0058] processor;

[0059] a memory for storing processor-executable instructions;

[0060] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0061] According to a fourth aspect of the embodiments of the present invention,

[0062] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0063] In the present invention, by combining multi-source heterogeneous health data, deep learning models and causal reasoning, it is possible to more comprehensively capture the user's health status and effectively identify potential risk factors, thereby improving the accuracy and reliability of health risk assessment and realizing personalized health management. By adopting hierarchical reinforcement learning and model predictive control, it is possible to adaptively adjust the monitoring plan according to the dynamic changes of the user's health status, and provide more targeted and timely health guidance, thereby better meeting the user's personalized needs. By using artificial intelligence technology, data processing, feature extraction, risk assessment and plan formulation are automatically performed, reducing the need for manual intervention, improving the efficiency and intelligence level of health monitoring, and reducing medical costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 A schematic diagram of a flow chart of a dynamic health adaptive monitoring method using artificial intelligence according to an embodiment of the present invention;

[0065] Figure 2 Schematic diagram of the structure of a dynamic health adaptive monitoring system using artificial intelligence according to an embodiment of the present invention. DETAILED DESCRIPTION

[0066] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0067] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0068] Figure 1FIG. 1 is a flow chart of a method for dynamic health adaptive monitoring using artificial intelligence according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0069] S1. Obtain the multi-source heterogeneous health data corresponding to the user, define the forward diffusion process and add noise through the hierarchical diffusion probability model based on the continuous-time stochastic differential equation, solve the differential equation for noise reduction in combination with the backward diffusion process, generate noise-reduced data and add it to the improved attention autoencoder, calculate the correlation weights between features through the multi-head self-attention mechanism and encode them as potential representations, optimize the network parameters in combination with the reconstruction error to generate preliminary features and add them to the contrastive learning network and perform random enhancement, generate positive sample pairs and learn discriminative features through the contrastive loss function, add the discriminative features to the adaptive wavelet neural network, perform multi-scale feature decomposition by dynamically selecting the optimal wavelet basis function and the number of decomposition layers, obtain multi-scale features and add them to the dynamic graph neural network, extract node representations in combination with the improved graph convolution operator and update edge weights in combination with the gated attention mechanism to obtain a health feature tensor;

[0070] The hierarchical diffusion probability model based on continuous-time stochastic differential equations is a model that describes and analyzes the diffusion process of data through continuous-time stochastic differential equations. It is usually used to process complex data with time series dependence and uncertainty, and captures diffusion behaviors at different levels through a hierarchical structure. The back diffusion is a back propagation diffusion process, which is used to optimize the parameters in the diffusion model. It is often used in deep learning and probabilistic graph models, and can effectively accelerate the training process and improve the accuracy of the model. The reconstruction error refers to the error between the reconstruction result of the input data and the original input data of the model in certain generative models (such as autoencoders, variational autoencoders, etc.). The contrast loss function is a loss function that optimizes the model by comparing the difference between positive samples and negative samples. The adaptive wavelet neural network is a hybrid model that combines wavelet transform and neural network. It uses wavelet transform to process the time-frequency characteristics of signals or images, and performs feature learning and nonlinear mapping through neural networks. The graph convolution operator is a convolution operation specifically used for graph data. The representation of each node is calculated by aggregating the features of nodes and their neighbors in the graph structure.

[0071] In an optional embodiment,

[0072] Obtain the multi-source heterogeneous health data corresponding to the user, define the forward diffusion process and add noise through the hierarchical diffusion probability model based on continuous-time stochastic differential equations, solve the differential equations in combination with the backward diffusion process for denoising, generate denoised data and add it to the improved attention autoencoder, calculate the correlation weights between features through the multi-head self-attention mechanism and encode them as potential representations, optimize the network parameters in combination with the reconstruction error to generate preliminary features and add them to the contrastive learning network and perform random enhancement, generate positive sample pairs and learn discriminative features through the contrastive loss function, add the discriminative features to the adaptive wavelet neural network, perform multi-scale feature decomposition by dynamically selecting the optimal wavelet basis function and the number of decomposition layers, obtain multi-scale features and add them to the dynamic graph neural network, extract node representations in combination with the improved graph convolution operator and update edge weights in combination with the gated attention mechanism, and obtain the health feature tensor including:

[0073] The continuous data of electrocardiogram signals, blood pressure values, and blood oxygen concentration and the discrete data of movement status and sleep quality are collected through the sensor array of the wearable device, the continuous data and the discrete data are time-aligned and format-unified, the data at different time intervals are resampled into a standardized multidimensional time series, the missing values ​​in the multidimensional time series are supplemented by a combination of forward filling and linear interpolation, and the abnormal values ​​in the multidimensional time series are processed by sliding window median filtering to obtain preprocessed health data;

[0074] Input the preprocessed health data into a hierarchical diffusion probability model, gradually inject Gaussian noise with increasing variance into the health data during the forward diffusion process, and construct a noise prediction network based on a U-shaped structure. The noise prediction network extracts multi-scale features and reconstructs detail information through a downsampling path and an upsampling path, respectively, adds a jump connection between the corresponding scales of the downsampling path and the upsampling path, predicts the distribution of the Gaussian noise through iterative optimization, and removes the noise during the reverse diffusion process to obtain the denoised health data;

[0075] Input the denoised health data into an improved attention autoencoder, which maps the input data to a high-dimensional feature space through a multi-layer perceptron and then divides it into multiple attention heads, each attention head calculates a query vector, a key vector and a value vector respectively, calculates the dot product of the query vector and the key vector to obtain an attention score, performs a weighted sum of the attention score and the value vector to obtain a context representation, connects the context representations of multiple attention heads and obtains a feature representation through a linear transformation, reconstructs the original data through a decoder and optimizes the network parameters to obtain the encoded health features;

[0076] Input the encoded health feature into a contrastive learning network, perform random time window cropping, amplitude modulation and random noise injection on the health feature to obtain an enhanced feature view, construct different feature views of the same health feature as a positive sample pair, construct feature views of different health features as a negative sample pair, constrain the distance of the positive sample pair in the feature space to be less than the distance of the negative sample pair through a contrast loss function, and obtain a discriminative health feature;

[0077] Input the discriminative health feature into an adaptive wavelet neural network, evaluate the expressive power of different wavelet basis functions based on the time-frequency characteristics of the signal and select the optimal basis function family, determine the number of decomposition layers according to the frequency composition of the signal, perform multi-scale decomposition on the health feature to obtain sub-band coefficients of different frequency bands, perform feature fusion on the sub-band coefficients at each scale to obtain a multi-scale health feature;

[0078] The multi-scale health features are constructed as a dynamic graph structure, and different health indicators are used as nodes in the graph. The neighborhood information of the nodes is aggregated through the graph convolution operator of the adaptive weight matrix, and the weight matrix is ​​dynamically updated based on the feature similarity between the nodes. The connection weights of the edges are adjusted according to the correlation strength between the nodes through the gated attention mechanism, and the node features extracted by the graph convolution are deeply fused with the edge features learned by the attention mechanism to output the health feature tensor.

[0079] The sliding window median filter is a filtering method based on a sliding window, which is used to remove noise in the signal. The median value is calculated in each window of the signal to replace the original data, thereby removing outliers or noise. The noise prediction network based on the U-shaped structure is a neural network model for noise prediction and denoising. The U-shaped structure is used to extract multi-scale features of input data and perform noise modeling. The random time window cropping is a technology for processing time series data, which is usually used for data cleaning or feature selection. A time window is randomly selected and the time series data is cropped within the window to reduce the impact of irrelevant or noise information and retain the part that is beneficial to model training.

[0080] Collect user health data from the sensor array of wearable devices, including continuous data such as ECG signals, blood pressure, and blood oxygen concentration, as well as discrete data such as exercise status and sleep quality. For example, the sampling frequency of ECG signals is 250Hz, blood pressure and blood oxygen concentration are measured once a minute, and exercise status and sleep quality are recorded once an hour. Time alignment and format unification of these data with different time intervals and formats are performed, the timestamps of all data are unified to the second level, and the discrete data are converted to numerical representation. Data with different time intervals are resampled, for example, all data are resampled to the same frequency, such as once a minute, using linear interpolation. For missing values ​​in multidimensional time series, a combination of forward filling and linear interpolation is used to complete them. For example, if the blood pressure value at a certain time point is missing, the blood pressure value at the previous time point is used to fill it; if the blood pressure values ​​at multiple consecutive time points are missing, linear interpolation is used to complete them. Finally, a sliding window median filter is used to process outliers in the multidimensional time series. For example, a median filter with a window size of 5 is used to remove spike noise in the ECG signal.

[0081] The preprocessed health data is input into the hierarchical diffusion probability model, and a U-shaped noise prediction network is used to gradually add Gaussian noise with increasing variance to the health data. The noise prediction network consists of a downsampling path and an upsampling path, which are used to extract multi-scale features and reconstruct detail information, respectively. A jump connection is added between the downsampling path and the upsampling path to transfer information between different scales. Through iterative optimization, the distribution of Gaussian noise is predicted, and the noise is gradually removed in the back diffusion process to obtain the denoised health data.

[0082] The denoised health data is input into the improved attention autoencoder, and the multi-head self-attention mechanism is used to calculate the correlation weights between features. Specifically, the input data is first mapped to a high-dimensional feature space through a multi-layer perceptron, and the features are divided into multiple attention heads. Each attention head calculates the query vector, key vector, and value vector respectively, and calculates the dot product between the query vector and the key vector to obtain the attention score. The attention score is weighted and summed with the value vector to obtain the context representation. The context representations of multiple attention heads are connected and linearly transformed to obtain the feature representation. Finally, the original data is reconstructed through the decoder, and the network parameters are optimized to obtain the encoded health features.

[0083] The encoded health features are input into the contrastive learning network. Data augmentation operations such as random time window cropping, amplitude modulation, and random noise injection are performed on the health features to generate different feature views. Different feature views of the same health feature are constructed as positive sample pairs, and feature views of different health features are constructed as negative sample pairs. Through the contrastive loss function, the distance between positive sample pairs in the feature space is constrained to be smaller than the distance between negative sample pairs, so as to learn discriminative health features.

[0084] Input the discriminative health features into the adaptive wavelet neural network. According to the time-frequency characteristics of the health data, select the optimal wavelet basis function and decomposition layer. For example, for the ECG signal, you can choose the Daubechies wavelet basis function and perform a 5-layer decomposition. Perform multi-scale decomposition on the health features to obtain sub-band coefficients of different frequency bands. Perform feature fusion on the sub-band coefficients at each scale to obtain multi-scale health features.

[0085] The multi-scale health features are constructed as a dynamic graph structure. Different health indicators are used as nodes in the graph, and the neighborhood information of the nodes is aggregated using the graph convolution operator with an adaptive weight matrix. The weight matrix is ​​dynamically updated according to the feature similarity between nodes. The gated attention mechanism is used to adjust the connection weight of the edge according to the correlation strength between nodes. The node features extracted by the graph convolution are deeply fused with the edge features learned by the attention mechanism, and finally the health feature tensor is output.

[0086] In this embodiment, by combining multiple deep learning models, key features in multi-source heterogeneous health data can be effectively extracted, the accuracy of health feature extraction can be improved, and model parameters can be adaptively adjusted according to individual differences of users to achieve personalized representation of health features. Through dynamic graph neural networks, health features can be represented as graph structures, and combined with a gated attention mechanism, the relationship between different health indicators can be explained, thereby enhancing the interpretability of health features.

[0087] In an optional embodiment,

[0088] The discriminative health features are input into the adaptive wavelet neural network, the expression capabilities of different wavelet basis functions are evaluated based on the time-frequency characteristics of the signal and the optimal basis function family is selected, the number of decomposition layers is determined according to the frequency composition of the signal, the health features are multi-scale decomposed to obtain sub-band coefficients of different frequency bands, and the sub-band coefficients are feature fused at each scale to obtain multi-scale health features including:

[0089] Inputting discriminative health features into an adaptive wavelet neural network, wherein the adaptive wavelet neural network comprises a wavelet basis function evaluation module, a decomposition layer number determination module and a feature fusion module;

[0090] The wavelet basis function evaluation module constructs a candidate wavelet basis function library, which includes a Haar wavelet basis function in a piecewise constant form, a Dobesi wavelet basis function family with different vanishing moment orders, a Schindler wavelet basis function family with different filter lengths, a Koivulet wavelet basis function family with symmetry, and a Mayer wavelet basis function with a smooth transition band in the frequency domain;

[0091] Performing time-frequency analysis on the health characteristics to obtain frequency domain characteristics, and calculating a reconstruction error evaluation value, a sparsity evaluation value, and a detail preservation evaluation value of a wavelet basis function based on the frequency domain characteristics;

[0092] Establishing a scoring matrix, normalizing the reconstruction error evaluation value, the sparsity evaluation value and the detail preservation evaluation value to obtain a normalized evaluation value, setting an adaptive weight coefficient based on the frequency domain feature, weighting the normalized evaluation value and the adaptive weight coefficient to obtain an expression ability score of each wavelet basis function, sorting the wavelet basis functions according to the expression ability score and selecting the wavelet basis function with the highest score as the optimal basis function;

[0093] The decomposition layer number determination module performs frequency analysis on the health feature to obtain frequency composition information, and determines the wavelet decomposition layer number according to the number of main frequency components and frequency band distribution in the frequency composition information;

[0094] Based on the optimal basis function and the wavelet decomposition layer number, the health feature is decomposed at multiple scales to obtain sub-band coefficients of different frequency bands, and the feature fusion module performs feature fusion on the sub-band coefficients of different frequency bands at various scales to obtain multi-scale health features;

[0095] Monitor the degree of change of the frequency domain characteristics, the rate of change of statistical characteristics and the fluctuation of reconstruction quality. When the degree of change exceeds a preset threshold, reselect the optimal basis function. When the reconstruction quality is lower than a preset reconstruction quality threshold, randomly initialize the adaptive weight coefficient.

[0096] The vanishing moment order refers to the characteristic that the high-order moments of certain wavelet basis functions tend to zero during the wavelet transform process, and is usually used to describe the smoothness of the wavelet basis function at certain scales. A higher vanishing moment order usually means that the wavelet basis function has stronger smoothness and locality. The Dobesi wavelet basis function family is a group of wavelet basis functions commonly used in multi-resolution analysis, which is characterized by the ability to perform effective signal decomposition at different scales. The frequency domain smooth transition band refers to a strip area applied to signal processing in the frequency domain, in which the signal spectrum transitions smoothly to reduce the edge effect of the frequency domain filter band.

[0097] Preprocess the discriminative health features, such as removing noise and normalizing, to ensure the quality of the input data. Then, a candidate wavelet basis function library is constructed, which contains many different types of wavelet basis functions, such as Haar wavelet, Dobecy wavelet, Sinlet wavelet, Koivreit wavelet and Mayer wavelet. These wavelet basis functions have different characteristics, such as compact support, orthogonality, symmetry, etc., which can adapt to different types of health feature signals.

[0098] Perform time-frequency analysis on the preprocessed healthy features, such as short-time Fourier transform or wavelet transform, to extract the frequency domain features of the signal. Based on the extracted frequency domain features, each candidate wavelet basis function is evaluated. The evaluation indicators include reconstruction error, sparsity, and detail preservation ability. Reconstruction error refers to the difference between the reconstructed signal and the original signal after the signal is decomposed and reconstructed using the wavelet basis function. Sparsity refers to the number of non-zero coefficients in the wavelet coefficients. The sparser the wavelet representation, the more conducive it is to feature extraction. Detail preservation ability refers to the ability of wavelet decomposition to retain the detail information of the original signal.

[0099] In order to comprehensively consider these three evaluation indicators, a scoring matrix is ​​constructed. The evaluation value of each indicator is normalized, and the normalized value is multiplied by the adaptive weight coefficient to obtain the expression ability score of each wavelet basis function. The setting of the adaptive weight coefficient can be adjusted according to the frequency domain characteristics. For example, if the signal has strong energy in the high-frequency part, the weight of the detail preservation ability can be increased. The wavelet basis functions are sorted according to the expression ability score, and the wavelet basis function with the highest score is selected as the optimal basis function.

[0100] For example, suppose a heartbeat signal is analyzed and its frequency domain features are extracted, and it is found that its energy is mainly concentrated in the low-frequency part. After evaluating the candidate wavelet basis functions, it is found that the Dobesi wavelet has the smallest reconstruction error and the highest sparsity, but the detail preservation ability is poor; while the Mayer wavelet has the best detail preservation ability, but the reconstruction error and sparsity are poor. Since the signal energy is mainly concentrated in the low frequency, the weight of the detail preservation ability is reduced, and the weight of the reconstruction error and sparsity is increased. In the end, the Dobesi wavelet has the highest comprehensive score and is selected as the optimal basis function.

[0101] After determining the optimal basis function, it is necessary to determine the number of wavelet decomposition layers. Perform frequency analysis on health characteristics to obtain frequency composition information, such as the number of main frequency components and frequency band distribution. Based on this information, determine the number of wavelet decomposition layers. The more decomposition layers there are, the more detailed frequency information can be obtained, but it also increases the computational complexity.

[0102] Assuming that the frequency range of the heartbeat signal is between 0.5Hz and 2Hz, in order to effectively separate the signals of different frequency bands, the number of decomposition layers can be set to 5. Based on the selected optimal basis function and the number of decomposition layers, the health features are decomposed at multiple scales to obtain the sub-band coefficients of different frequency bands. Then, the sub-band coefficients of different frequency bands are fused at each scale to obtain multi-scale health features. There are many ways to fuse features, such as simple splicing, weighted averaging, etc.

[0103] In order to ensure the adaptability of the algorithm, it is necessary to monitor the changes in frequency domain features. When the degree of change in frequency domain features exceeds the preset threshold, the optimal basis function and decomposition level are reselected. When the reconstruction quality is lower than the preset reconstruction quality threshold, the adaptive weight coefficient is reinitialized and the wavelet basis function is reselected.

[0104] In this embodiment, by selecting the optimal wavelet basis function and decomposition layer number, the time-frequency characteristics of the health characteristic signal can be better captured, thereby improving the accuracy of health assessment. By adaptively adjusting the wavelet basis function, decomposition layer number and weight coefficient, it can adapt to different types of health characteristic signals and different application scenarios. By selecting the appropriate decomposition layer number and wavelet basis function, the computational complexity can be reduced and the efficiency of the algorithm can be improved.

[0105] S2. Add the health feature tensor to the hierarchical memory-enhanced network, store feature information by constructing multi-layer memory units with different time spans, modify information flow and determine long-term and short-term dependencies in combination with an adaptive gating mechanism, obtain time series features and add them to the dynamic routing capsule network, express the time series features in a hierarchical manner in combination with an improved iterative routing algorithm, obtain hierarchical features and add them to the causal reasoning network, define the dependencies between different health influencing factors through a structural equation model and verify the causal hypothesis in combination with a reverse intervention mechanism, obtain a verified causal relationship graph, construct the conditional probability distribution between nodes in combination with the conditional random field in the probability graph attention network and calculate the node importance weights to generate a health knowledge graph, add the time series features and the health knowledge graph to a deep Bayesian network, approximate the posterior distribution through variational inference and estimate the uncertainty of the prediction results in combination with Monte Carlo sampling, and obtain the health risk assessment results;

[0106] The hierarchical memory enhancement network is a deep learning model that combines a memory network and a hierarchical structure, and is designed to enhance the model's memory ability for long-term dependencies and complex information. The information flow refers to the propagation process of data or signals in the system, which involves the conversion and processing of input signals at various levels to ultimately obtain outputs. The dynamic routing capsule network is a model based on a capsule network, designed to capture the spatial hierarchical relationship of data such as images and voices. The improved iterative routing algorithm is a key optimization method in the dynamic routing capsule network, which optimizes the information transmission path between capsule units through multiple iterations and layer-by-layer weight adjustment. The hierarchical expression refers to the step-by-step abstraction and representation of data through a hierarchical structure. In deep learning, hierarchical expression is usually reflected in different network levels, and each layer extracts different levels of data features. The causal reasoning network is a neural network model that uses causal relationships for reasoning and decision-making. By constructing a causal graph or a structural equation model, it can explain and predict the causal relationships between variables in the system. The structural equation model is a statistical model for analyzing the causal relationship between variables. Through the structural equation model, the relationship between multiple variables can be systematically studied, and the causal relationship between latent variables and observed variables can be evaluated. The reverse intervention machine is a mechanism for reversely inferring causal relationships. By applying reverse intervention in the causal network, the impact of intervention on the system is estimated. It is often used in causal reasoning and experimental design. The probabilistic graph attention network is a neural network that combines the probabilistic graph model and the attention mechanism. It uses the graph structure to model the dependency relationship between variables and dynamically adjusts the importance of different variables through the attention mechanism, thereby optimizing the learning efficiency and performance of the model. The conditional random field is a probabilistic graph model for labeling and segmentation tasks, which is particularly suitable for labeling sequence data and image data. The deep Bayesian network is a deep learning model based on the Bayesian network, which combines the advantages of probabilistic reasoning and deep learning, models complex probabilistic relationships through a hierarchical structure, and updates the probability distribution of nodes through Bayesian inference.

[0107] In an optional embodiment,

[0108] The health feature tensor is added to the hierarchical memory enhancement network, and the feature information is stored by constructing a multi-layer memory unit with different time spans. The information flow is modified and the long-term and short-term dependencies are determined by combining the adaptive gating mechanism, and the time series features are obtained and added to the dynamic routing capsule network. The time series features are hierarchically expressed by combining the improved iterative routing algorithm, and the hierarchical features are obtained and added to the causal reasoning network. The dependency between different health influencing factors is defined by the structural equation model and the causal hypothesis is verified by combining the reverse intervention mechanism to obtain a verified causal relationship graph. The conditional probability distribution between nodes is constructed by combining the conditional random field in the probability graph attention network and the node importance weight is calculated to generate a health knowledge graph. The time series features and the health knowledge graph are added to the deep Bayesian network, and the posterior distribution is approximated by variational inference and the uncertainty of the prediction results is estimated by combining Monte Carlo sampling. The health risk assessment results include:

[0109] The healthy feature tensor is input into the normalization module for normalization processing to obtain a standardized feature, and the standardized feature is input into the hierarchical memory enhancement network. The current input feature reception ratio is regulated by the input gating value of the short-term memory unit, the historical state retention degree is adjusted by the forgetting gating value, and the feature output strength is regulated by the output gating value to generate a short-term feature. The short-term feature is input into the medium-term memory unit and a query vector is generated. The similarity score between the query vector and the short-term feature is calculated, and the similarity score is normalized to obtain the attention weight. The short-term feature is multiplied by the attention weight to generate a medium-term feature. The short-term feature and the medium-term feature are combined by the residual connection of the long-term memory unit and a gated update is performed to generate a long-term feature. The change rate of the short-term feature, the medium-term feature and the long-term feature is calculated, and the state update ratio of the corresponding memory unit is adjusted according to the change rate by using the adaptive gating mechanism and the long-term dependency is determined, and the time series feature is generated by fusion.

[0110] Input the time series features into multiple dedicated capsules of the dynamic routing capsule network, wherein each of the dedicated capsules corresponds to a specific health influencing factor, perform feature extraction on the time series features through a convolution transformation matrix and generate an activity vector, calculate the modulus length of the activity vector to represent the probability of feature existence, use the direction of the activity vector to represent the feature attribute to obtain a primary feature representation, initialize the routing coefficient matrix to be uniformly distributed, multiply the primary feature representation with the transformation matrix to generate a prediction vector, calculate the inner product of the prediction vector and the high-level capsule feature vector to obtain similarity, perform softmax normalization on the similarity to update the routing coefficient, perform weighted summation of the updated routing coefficient and the prediction vector to generate a feature vector of the high-level capsule, and repeat the update until the routing coefficient converges or reaches a maximum number of iterations to generate a hierarchical feature;

[0111] Input the hierarchical features into the causal reasoning network, calculate the correlation coefficients between different health influencing factors based on the hierarchical features to construct a correlation matrix, use the structural equation model to decompose the correlation matrix into a direct effect matrix and an indirect effect matrix to construct an initial causal graph, calculate the conditional mutual information and Fisher conditional independence statistics for each pair of nodes with a connection relationship, screen the direct causal relationship that meets the conditional independence hypothesis according to the significance level, uniformly sample within the value range of the target node to generate multiple groups of intervention values, substitute each group of intervention values ​​into the causal graph and calculate the response values ​​of other nodes except the target node based on the structural equation model, calculate the mean square error between the response value and the historical data, remove the corresponding causal edge when the mean square error exceeds a preset threshold, and repeatedly calculate and delete the causal edge until the causal graph structure is stable to generate a verified causal relationship graph;

[0112] Input the verified causal relationship graph into the probabilistic graph attention network, define a discrete state space for each node and count the state transition frequency between adjacent time steps, use maximum likelihood estimation to calculate the state transition probability to construct a transfer matrix, calculate the joint distribution probability between any two nodes based on the transfer matrix, calculate the corresponding conditional probability distribution for each node to represent the current node's dependence on other nodes, calculate the importance score of the node according to the conditional probability distribution and perform softmax normalization to obtain the attention weight, multiply the node feature with the corresponding attention weight and aggregate to generate a health knowledge graph that takes into account the importance of the node;

[0113] The time series features and the health knowledge graph are input into the deep Bayesian network, the mean vector and covariance matrix of the variational distribution are initialized, the variational objective function is constructed including the log-likelihood term and the KL divergence term, the gradient of the variational objective function with respect to the variational parameters is calculated, the variational parameters are iteratively optimized using stochastic gradient descent until convergence to obtain the approximate posterior distribution, multiple Monte Carlo samplings are performed from the approximate posterior distribution to obtain model parameter samples, the corresponding model prediction output is calculated for each group of parameter samples, the mean of all model prediction outputs is statistically calculated as the final prediction result, and the variance of the model prediction output is calculated to evaluate the uncertainty, and a health risk assessment result including a risk level and a confidence interval is generated.

[0114] The convolution transformation matrix is ​​a matrix representation used in convolution operations. By converting convolution operations into matrix multiplications, computational efficiency can be improved, and it has important applications in large-scale image processing and deep neural network training. The activity vector refers to the output vector processed by the activation function in the neural network, which usually contains the nonlinear transformation results of the network on the input data and is a key component of model prediction and decision-making. The direct effect matrix is ​​a matrix that describes the direct effects in the system. It is usually used in causal reasoning and statistical analysis, and represents the direct causal relationship between the variables in the system in the absence of indirect effects. The indirect effect matrix describes the influence transmitted through other variables, which usually appears in structural equation models and is used to analyze the indirect effects transmitted through mediating paths between different variables. The conditional mutual information is a statistic used to measure the mutual dependence of two variables under the condition of given other variables. The Fisher conditional independence statistic is a method for testing the independence between variables, which is usually used in causal reasoning and statistical modeling. The significance level is the critical value used to decide whether to reject the null hypothesis in hypothesis testing.

[0115] Preprocess the collected health feature data. For example, the user's blood pressure, blood sugar, heart rate, weight, age, gender, family history and other health data are collected, and these data may have different dimensions and ranges. In order to eliminate the impact of dimensions and ranges, these health feature data are input into the normalization module for normalization processing, such as using the Min-Max normalization method to scale the value of each feature to between 0 and 1 to obtain standardized features.

[0116] Time series features are extracted using a hierarchical memory-augmented network. The standardized health features are input into the network. The network contains short-term, medium-term, and long-term memory units. Short-term memory units focus on the current health status, such as the most recent blood pressure value. Medium-term memory units focus on health trends over a period of time, such as changes in blood pressure over the past week. Long-term memory units focus on long-term health patterns, such as the average blood pressure over the past year. Each memory unit uses a gating mechanism to control the flow of information and the update of memory. For example, the input gate of the short-term memory unit controls the acceptance ratio of the current input feature, the forget gate controls the degree of retention of the historical state, and the output gate controls the intensity of the feature output. By combining memory information of different time spans, the network is able to capture the dynamic changes of health data and generate more representative time series features. Assuming that the user's blood pressure data has continued to rise over the past week, the short-term and medium-term memory units will capture this upward trend and reflect it in the generated time series features.

[0117] The extracted time series features are input into the dynamic routing capsule network for hierarchical expression. The dynamic routing capsule network consists of multiple dedicated capsules, each of which corresponds to a specific health influencing factor, such as blood pressure, blood sugar, etc. Each capsule extracts the time series features through the convolution transformation matrix and generates an activity vector. The modulus of the activity vector represents the probability of the feature's existence, and the direction represents the attribute of the feature. Through the iterative routing algorithm, the network can combine low-level features into high-level features, thereby realizing the hierarchical expression of time series features. For example, the network can combine the time series features of blood pressure and blood sugar into hierarchical features of cardiovascular risk.

[0118] The hierarchical features are input into the causal inference network to construct a causal relationship diagram between health influencing factors. First, the correlation coefficients between different health influencing factors are calculated based on the hierarchical features, and then the initial causal diagram is constructed using the structural equation model. In order to verify the causal hypothesis, the network uses a reverse intervention mechanism. For example, assuming that the network infers that high blood pressure causes heart disease, then the incidence of heart disease can be reduced by simulating interventions to lower blood pressure. If the simulation results are consistent with the actual situation, the causal relationship between high blood pressure and heart disease can be confirmed.

[0119] The verified causal graph is input into the probabilistic graph attention network to generate a health knowledge graph. The network constructs the conditional probability distribution between nodes based on the causal graph and calculates the importance weights of the nodes. For example, if high blood pressure is an important factor leading to heart disease, then the high blood pressure node will have a higher weight in the knowledge graph.

[0120] The time series features and health knowledge graph are input into the deep Bayesian network for health risk assessment. The network approximates the posterior distribution through variational inference and combines Monte Carlo sampling to estimate the uncertainty of the prediction results. For example, the network can predict that the probability of a user suffering from heart disease in the next year is 20%, with a confidence interval of 15% to 25%.

[0121] In this embodiment, by integrating technologies such as time series features, causal reasoning, and knowledge graphs, the dynamic changes of health data and the complex relationships between influencing factors can be more comprehensively considered, thereby improving the accuracy of the assessment. The constructed causal relationship diagram and health knowledge graph can clearly display the interactions and importance between different health influencing factors, thereby enhancing the interpretability of the assessment. Through Bayesian inference and Monte Carlo sampling, the uncertainty of the assessment results can be quantified, providing more reliable risk assessment results and confidence intervals.

[0122] In an optional embodiment,

[0123] Input the verified causal relationship graph into the probabilistic graph attention network, define a discrete state space for each node and count the state transition frequency between adjacent time steps, calculate the state transition probability using maximum likelihood estimation to construct a transfer matrix, and calculate the joint distribution probability between any two nodes based on the transfer matrix, including:

[0124] Inputting the verified causal relationship graph into the probabilistic graph attention network, the verified causal relationship graph includes multiple nodes representing different health influencing factors, determining the number of state categories of each node according to clinical medicine standards, dividing the indicator value of each node into multiple state intervals, establishing a mapping relationship from the indicator value to the state identifier, and generating a discrete state space of the node;

[0125] An observation time window is set based on the discrete state space, and state transition statistics are independently performed on each of the nodes within the observation time window. The state of the node at each time step is identified by traversing the time series data, and the state transition of the node between adjacent time steps is recorded to create a state transition frequency matrix, wherein the rows and columns of the state transition frequency matrix correspond to all possible states of the node respectively;

[0126] Sum all elements of each row in the state transition frequency matrix to obtain the total number of transition events of the corresponding state, divide each element in the state transition frequency matrix by the total number of transition events in its row to obtain a normalized transition probability, repeat the normalization operation for each row of the state transition frequency matrix to generate a transition probability matrix;

[0127] Extract the transition probability matrices of any two nodes, combine the two transition probability matrices to obtain a joint probability matrix representing the combination of the two node states, repeat the combination operation for all node pairs in the causal relationship graph, and generate the joint distribution probability between all node pairs in the causal relationship graph.

[0128] The state interval refers to the interval that describes the range of changes in the system state, which is usually used in reinforcement learning and control systems. Each state in the decision-making process will have a corresponding state interval, which is used to limit the behavior range of the system. The transition probability matrix is ​​used to describe the state transition probability in the Markov process, which represents the probability of transitioning from one state to another. It is a basic tool for analyzing and predicting changes in system states.

[0129] Get a verified causal relationship graph that describes the causal relationship between different health influencing factors, for example, smoking may cause lung cancer, lack of exercise may cause obesity, etc. The nodes in the graph represent different health influencing factors, such as blood pressure, blood sugar, BMI, exercise, smoking status, etc. The edges between the nodes represent the direction and strength of the causal relationship.

[0130] Determine the number of state categories and state intervals for each node. According to clinical medical standards, divide the indicator value of each node into multiple state intervals. For example, divide blood pressure into three states: normal, high, and high blood pressure; divide BMI into four states: thin, normal, overweight, and obese. And establish a mapping relationship from indicator value to state identifier. For example, blood pressure values ​​below 120 / 80mmHg correspond to the "normal" state, between 120 / 80mmHg and 140 / 90mmHg correspond to the "high" state, and above 140 / 90mmHg correspond to the "high blood pressure" state. In this way, a discrete state space for each node is generated.

[0131] For example, for the "blood pressure" node, it is assumed that it is divided into three states according to clinical medical standards: "normal", "high" and "hypertension". If the blood pressure measurement value at a certain time point is 130 / 85 mmHg, according to the predefined state interval, the blood pressure value falls into the "high" state interval, then the state identifier of the "blood pressure" node at that time point is "high".

[0132] Set an observation time window, such as one year, and perform state transition statistics for each node independently based on this time window. Traverse the time series data and identify the state of each node at each time step. For example, suppose a patient's blood pressure measurement data in one year is as follows: 115 / 75mmHg (normal) in the first month, 125 / 80mmHg (high) in the second month, 135 / 85mmHg (high) in the third month, 120 / 78mmHg (normal) in the fourth month, and so on. By traversing these 12 months of data, you can identify the state of blood pressure in each month.

[0133] Record the state transition of each node between adjacent time steps and create a state transition frequency matrix. The rows and columns of the matrix correspond to all possible states of the node. For example, the state transition frequency matrix of the "blood pressure" node is a 3x3 matrix, with rows and columns representing the three states of "normal", "high", and "hypertension". According to the sample data of the above patient, the state transition from the "normal" state in the first month to the "high" state in the second month, the element of the "normal" row and the "high" column in the frequency matrix is ​​increased by 1.

[0134] Normalize the state transition frequency matrix to generate a transition probability matrix. Sum all elements in each row of the frequency matrix to get the total number of transition events for the corresponding state. Divide each element in the frequency matrix by the total number of transition events in its row to get the normalized transition probability. For example, assuming that the number of times the "blood pressure" node transitions from the "normal" state to the "normal", "high", and "hypertension" states is 50, 30, and 20, respectively, then the total number of transition events in the "normal" state is 100, and the probabilities of transitioning to the "normal", "high", and "hypertension" states are 0.5, 0.3, and 0.2, respectively. Repeat the normalization operation for each row of the matrix to finally get the transition probability matrix.

[0135] Calculate the joint distribution probability between any two nodes. Extract the transition probability matrices of any two nodes, combine the two matrices, and obtain a joint probability matrix that represents the combination of the two node states. For example, combine the transition probability matrices of the "blood pressure" node and the "BMI" node to obtain a new matrix whose rows and columns represent all possible combinations of "blood pressure" and "BMI" states, respectively. Repeat this operation for all node pairs in the causal relationship graph to generate the joint distribution probability between all node pairs.

[0136] In this embodiment, by constructing a state transition probability matrix and a joint distribution probability, the correlation strength between different health influencing factors can be accurately quantified. By analyzing the health state transition of an individual at different time points, the evolution trend of the individual's future health state can be predicted. Based on the individual's health data and causal relationship diagram, a personalized health management plan can be provided for the individual, thereby improving the effectiveness of health management.

[0137] S3. Add the health risk assessment results to the hierarchical reinforcement learning network, and build a hierarchical policy structure to enable the upper network to generate abstract goals to guide the lower network to select specific actions. The selected candidate action plans are input into the multi-agent collaborative optimization system, and the collaborative optimization of group decision-making is achieved by designing communication protocols and reward distribution mechanisms between agents. The optimized strategy is input into the deep deterministic policy gradient network, and the deterministic policy function is combined with the value function and the policy parameters are optimized by the gradient ascent method to achieve online learning. The learned policy parameters are input into the model predictive control network, and the future state sequence is predicted by the rolling time domain optimization method and the control strategy is dynamically adjusted to obtain an adaptive monitoring solution.

[0138] The hierarchical strategy structure is a commonly used architecture in reinforcement learning and optimization algorithms. It decomposes the decision-making process into multiple levels, each level is responsible for handling decision-making problems at different levels to efficiently learn and execute strategies. The gradient ascent method is an optimization technique used to maximize the objective function by updating parameters. The rolling horizon optimization method is a strategy commonly used in dynamic optimization, especially when dealing with problems with uncertainty and dynamic changes.

[0139] In an optional embodiment,

[0140] The health risk assessment results are added to the hierarchical reinforcement learning network. By constructing a hierarchical strategy structure, the upper network generates abstract goals to guide the lower network to select specific actions. The selected candidate action plans are input into the multi-agent collaborative optimization system. The collaborative optimization of group decision-making is achieved by designing the communication protocol and reward distribution mechanism between agents. The optimized strategy is input into the deep deterministic policy gradient network. The deterministic policy function is combined with the value function and the policy parameters are optimized by the gradient ascent method to achieve online learning. The learned policy parameters are input into the model predictive control network. The future state sequence is predicted by the rolling time domain optimization method and the control strategy is dynamically adjusted. The adaptive monitoring scheme includes:

[0141] Adding the health risk assessment results to a hierarchical reinforcement learning network, wherein the hierarchical reinforcement learning network generates an abstract goal in the upper network to guide the lower network to select specific actions by constructing a hierarchical strategy structure, wherein the hierarchical strategy structure includes a state encoding module, a goal generation module and a strategy mapping module, wherein the state encoding module encodes the risk level, abnormality degree and change trend in the health risk assessment results into a state vector, the goal generation module generates an abstract goal of monitoring frequency and warning threshold based on the state vector, and the strategy mapping module converts the abstract goal into a candidate action plan;

[0142] The candidate action plan is input into a multi-agent collaborative optimization system, the multi-agent collaborative optimization system sets a communication protocol and a reward allocation mechanism between agents, the communication protocol defines the transmission rules of status information messages, action suggestion messages and evaluation feedback messages, the reward allocation mechanism allocates a reward value to each agent according to data collection quality, prediction accuracy and strategy execution effect, and realizes collaborative optimization of group decision-making based on the communication protocol and the reward allocation mechanism;

[0143] The optimized strategy is input into a deep deterministic policy gradient network, which combines a deterministic policy function with a value function. The deterministic policy function outputs a specific monitoring action, and the value function evaluates the long-term value of the monitoring action. The policy parameters of the deterministic policy function are continuously optimized by a gradient ascent method, and the policy parameters are input into a model predictive control network. The model predictive control network predicts the future state sequence by a rolling horizon optimization method, searches for the optimal control sequence within the prediction range based on the future state sequence, executes the control action at the current moment and receives state feedback, dynamically adjusts the policy parameters, and generates an adaptive monitoring solution.

[0144] The abstract goal refers to a relatively high-level goal set in a complex system to simplify problem handling, and the optimal control sequence refers to a control strategy sequence in a dynamic system that enables the system to achieve optimal performance under given constraints.

[0145] Conduct health risk assessment to obtain information such as individual risk level, abnormality, and change trend. For example, by collecting data on individual physiological indicators (such as blood pressure, blood sugar, heart rate, etc.), living habits (such as smoking, drinking, exercise, etc.), and environmental factors (such as air quality, noise pollution, etc.), and combining medical knowledge and statistical models, assess individual health risks. Suppose the health risk assessment result of an individual is: the risk level is medium, the abnormality is mild, and the change trend is rising.

[0146] The health risk assessment results are input into the state encoding module and converted into a state vector. For example, the risk level "intermediate" is encoded as the value "2" (low level is 1, high level is 3), the abnormality "mild" is encoded as the value "1" (moderate is 2, severe is 3), and the change trend "rising" is encoded as the value "1" (declining is -1, stable is 0). The final state vector is [2, 1, 1].

[0147] The target generation module generates abstract targets for monitoring frequency and warning threshold based on the state vector. For example, based on the state vector [2, 1, 1], the target generation module determines that the monitoring frequency is once a week and the warning threshold is when the blood pressure exceeds 140 mmHg.

[0148] The strategy mapping module converts abstract goals into candidate action plans. For example, the monitoring frequency "once a week" and the warning threshold "blood pressure exceeds 140 mmHg" are converted into specific candidate action plans: blood pressure data is collected once a week, and an alarm is issued if blood pressure exceeds 140 mmHg.

[0149] The candidate action plans are input into the multi-agent collaborative optimization system, which includes multiple agents, each of which represents a monitoring strategy. Agents exchange information through communication protocols, for example, sharing their respective status information, action suggestions, and evaluation feedback. The system rewards each agent based on indicators such as data collection quality, prediction accuracy, and strategy execution effect. Suppose there are two agents, one recommends increasing the monitoring frequency, and the other recommends lowering the warning threshold. Through multiple rounds of communication and reward distribution, the best monitoring plan is finally determined.

[0150] The optimized policy is input into a deep deterministic policy gradient network, which includes a deterministic policy function and a value function. The deterministic policy function outputs a specific monitoring action based on the current state, such as collecting blood pressure data. The value function evaluates the long-term value of the monitoring action, such as the degree of improvement in future health status. The network continuously optimizes the policy parameters through the gradient ascent method to maximize the long-term value.

[0151] The learned policy parameters are input into the model predictive control network, and the future state sequence is predicted through the rolling horizon optimization method, for example, the blood pressure change trend in the next week. Based on the prediction results, the optimal control sequence is found within the prediction range, for example, the monitoring frequency and warning threshold are adjusted. The control action at the current moment is executed, and state feedback is received, for example, the latest blood pressure data. Based on the feedback information, the policy parameters are dynamically adjusted to generate an adaptive monitoring plan. For example, if the blood pressure continues to rise, the monitoring frequency is increased or the warning threshold is lowered.

[0152] In this embodiment, a personalized monitoring plan is formulated based on the results of individual health risk assessment to avoid a one-size-fits-all approach and improve the pertinence and effectiveness of monitoring. The monitoring plan is dynamically adjusted according to changes in individual health status to achieve adaptive monitoring and improve monitoring flexibility. Through intelligent strategy optimization and predictive control, monitoring efficiency is improved, unnecessary monitoring is reduced, and monitoring costs are reduced.

[0153] In an optional embodiment,

[0154] The model predictive control network predicts the future state sequence by a rolling horizon optimization method, searches for the optimal control sequence within the prediction range based on the future state sequence, executes the control action at the current moment and receives state feedback, dynamically adjusts the strategy parameters, and generates an adaptive monitoring scheme, including:

[0155] Constructing a basic architecture of a dynamic prediction model, the dynamic prediction model includes an input layer, a data preprocessing module, a feature extraction layer and a state encoding layer, the input layer receives a physiological indicator numerical sequence, a monitoring time series and an environmental parameter, the data preprocessing module performs standardization processing on the physiological indicator numerical sequence, the monitoring time series and the environmental parameter to obtain normalized data, the feature extraction layer extracts time series features from the normalized data through a bidirectional long short-term memory network, and the state encoding layer encodes the time series features into a state vector of fixed dimension;

[0156] Generate a future state sequence through a prediction controller based on the state vector, input the state vector into a prediction model to obtain a state prediction value at the next moment, and use the state prediction value as a new input to continue predicting to obtain a complete state sequence in a prediction time domain, wherein the prediction time domain is divided into a plurality of prediction steps, and each prediction step corresponds to a state prediction value;

[0157] The optimal control sequence is determined by adopting a hierarchical search strategy, wherein the first layer determines the reference value of the monitoring frequency according to the fluctuation degree of the complete state sequence, the second layer performs a local search near the reference value to obtain the optimized monitoring frequency, and the third layer dynamically adjusts the value range of the warning threshold based on the complete state sequence, and the optimal control sequence is determined by evaluating the comprehensive effects of different control schemes through a heuristic algorithm;

[0158] Sending the monitoring frequency and early warning threshold in the optimal control sequence to the monitoring device to start data collection, recording actual monitoring data, wherein the actual monitoring data includes a monitoring value sequence, a monitoring time sequence, and an early warning trigger record, calculating the error between the state prediction value and the actual monitoring data, constructing a loss function based on the error and using the stochastic gradient descent method to update the parameters of the dynamic prediction model, and triggering the adjustment of the dynamic prediction model structure when the error exceeds a set threshold;

[0159] The control strategy is re-optimized based on the updated dynamic prediction model, the monitoring frequency and the warning threshold are adaptively adjusted according to the weight relationship between the prediction accuracy and the monitoring burden, and the adjusted control strategy is used for the optimization process of the next prediction cycle.

[0160] The prediction time domain refers to the time range used to predict the system behavior in the future in dynamic optimization and control problems. The prediction step size refers to the time interval or number of steps selected for future prediction in dynamic optimization. The monitoring burden refers to the resource consumption required in a system to ensure the normal operation of the system or respond to external interference.

[0161] Build the infrastructure of the dynamic prediction model. The architecture includes an input layer, a data preprocessing module, a feature extraction layer, and a state encoding layer. The input layer receives a sequence of physiological index values ​​(e.g., heart rate, blood pressure, blood oxygen saturation, etc.), a monitoring time series (e.g., monitoring timestamp), and environmental parameters (e.g., temperature, humidity, air pressure, etc.). Example data: heart rate 70 beats / minute, blood pressure 120 / 80 mmHg, blood oxygen 98%, temperature 25°C, humidity 50%, timestamp 2024-07-28 10:00:00. The data preprocessing module normalizes these input data, such as using the min-max normalization method to scale the data to between 0 and 1 to eliminate the influence of different indicator dimensions. The feature extraction layer uses a bidirectional long short-term memory network (BiLSTM) to extract time series features from the normalized data. The BiLSTM network can consider both past and future information at the same time, thereby better capturing the time dependency in the data. The state encoding layer encodes the extracted time series features into a state vector of fixed dimension for subsequent processing. For example, the time series features of indefinite length are encoded into a 128-dimensional vector.

[0162] Generate the future state sequence based on the state vector through the prediction controller. Input the current state vector into the prediction model (for example, another BiLSTM network) to obtain the state prediction value at the next moment. Input the prediction value into the prediction model again as a new input, and iterate the prediction to obtain the complete state sequence in the prediction time domain. The prediction time domain can be divided into multiple prediction steps, each of which corresponds to a state prediction value. For example, if the state of the next 24 hours is predicted, with a prediction step of 1 hour, 24 state prediction values ​​can be obtained.

[0163] A hierarchical search strategy is used to determine the optimal control sequence. The first layer determines the baseline value of the monitoring frequency according to the degree of fluctuation of the complete state sequence. For example, if the state fluctuation is large, the baseline value is set to a higher frequency, such as once an hour; if the state fluctuation is small, the baseline value is set to a lower frequency, such as once every 4 hours. The second layer performs a local search near the baseline value to obtain the optimized monitoring frequency. For example, based on the baseline value of once every 4 hours, try every 3 hours, every 5 hours, etc., and select the frequency with the best effect. The third layer dynamically adjusts the value range of the warning threshold based on the complete state sequence. For example, if the predicted state value is close to the abnormal value, the value range of the warning threshold is narrowed to improve the warning sensitivity. The comprehensive effect of different control schemes (combinations of different monitoring frequencies and warning thresholds) is evaluated through heuristic algorithms (such as genetic algorithms), and the optimal control sequence is finally determined.

[0164] Send the monitoring frequency and warning threshold in the optimal control sequence to the monitoring device to start data collection. Record the actual monitoring data, including the monitoring value sequence, monitoring time series, and warning trigger records. For example, record monitoring values ​​such as heart rate and blood pressure, as well as the time point when the warning is triggered. Calculate the error between the state prediction value and the actual monitoring data. Construct a loss function based on the error, such as mean square error, and use the stochastic gradient descent method to update the parameters of the dynamic prediction model. When the error exceeds the set threshold, trigger the adjustment of the dynamic prediction model structure, such as increasing the number of layers or hidden units of the BiLSTM network.

[0165] Re-optimize the control strategy based on the updated dynamic prediction model. Adaptively adjust the monitoring frequency and warning threshold according to the weight relationship between prediction accuracy and monitoring burden. For example, if the prediction accuracy is high, reduce the monitoring frequency to reduce the burden; if the monitoring burden is too heavy, appropriately reduce the prediction accuracy requirements and increase the monitoring frequency. Use the adjusted control strategy for the optimization process of the next prediction cycle.

[0166] In this embodiment, through dynamic prediction models and adaptive adjustment strategies, the monitoring plan can be adjusted in real time according to changes in individual status to achieve more accurate monitoring and timely detection of potential risks. By adaptively adjusting the monitoring frequency, unnecessary frequent monitoring can be avoided, the patient's physiological and psychological burden can be reduced, and the patient's compliance can be improved. Personalized monitoring plans can be formulated according to individual differences to better meet the needs of different individuals and improve the effectiveness of monitoring.

[0167] Figure 2 FIG. 1 is a schematic diagram of the structure of a dynamic health adaptive monitoring system using artificial intelligence according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0168] The first unit is used to obtain multi-source heterogeneous health data corresponding to the user, define the forward diffusion process and add noise through the hierarchical diffusion probability model based on continuous-time stochastic differential equations, solve the differential equations in combination with the backward diffusion process for noise reduction, generate noise-reduced data and add it to the improved attention autoencoder, calculate the correlation weights between features through the multi-head self-attention mechanism and encode them as potential representations, optimize the network parameters in combination with the reconstruction error to generate preliminary features and add them to the contrastive learning network and perform random enhancement, generate positive sample pairs and learn discriminative features through the contrastive loss function, add the discriminative features to the adaptive wavelet neural network, perform multi-scale feature decomposition by dynamically selecting the optimal wavelet basis function and the number of decomposition layers, obtain multi-scale features and add them to the dynamic graph neural network, extract node representations in combination with the improved graph convolution operator and update edge weights in combination with the gated attention mechanism to obtain a health feature tensor;

[0169] The second unit is used to add the health feature tensor to the hierarchical memory enhancement network, store feature information by constructing multi-layer memory units with different time spans, modify information flow and determine long-term and short-term dependencies in combination with an adaptive gating mechanism, obtain time series features and add them to the dynamic routing capsule network, express the time series features in a hierarchical manner in combination with an improved iterative routing algorithm, obtain hierarchical features and add them to the causal reasoning network, define the dependencies between different health influencing factors through a structural equation model and verify the causal hypothesis in combination with a reverse intervention mechanism, obtain a verified causal relationship graph, construct the conditional probability distribution between nodes in combination with the conditional random field in the probability graph attention network and calculate the node importance weights, generate a health knowledge graph, add the time series features and the health knowledge graph to a deep Bayesian network, approximate the posterior distribution through variational reasoning and estimate the uncertainty of the prediction results in combination with Monte Carlo sampling, and obtain a health risk assessment result;

[0170] The third unit is used to add the health risk assessment results to the hierarchical reinforcement learning network, construct a hierarchical policy structure so that the upper network generates abstract goals to guide the lower network to select specific actions, input the selected candidate action plans into the multi-agent collaborative optimization system, and realize the collaborative optimization of group decision-making by designing the communication protocol and reward distribution mechanism between agents. The optimized strategy is input into the deep deterministic policy gradient network, the deterministic policy function is combined with the value function and the policy parameters are optimized by the gradient ascent method to realize online learning, the learned policy parameters are input into the model predictive control network, and the future state sequence is predicted by the rolling time domain optimization method and the control strategy is dynamically adjusted to obtain an adaptive monitoring solution.

[0171] According to a third aspect of the embodiments of the present invention,

[0172] An electronic device is provided, comprising:

[0173] processor;

[0174] a memory for storing processor-executable instructions;

[0175] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0176] According to a fourth aspect of the embodiments of the present invention,

[0177] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0178] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0179] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic health adaptive monitoring method using artificial intelligence, characterized in that: include: Obtain multi-source heterogeneous health data corresponding to the user, define the forward diffusion process and add noise through a hierarchical diffusion probability model based on continuous-time stochastic differential equations, solve the differential equations in combination with the backward diffusion process for noise reduction, generate noise-reduced data and add it to the improved attention autoencoder, calculate the correlation weights between features through a multi-head self-attention mechanism and encode them as potential representations, optimize network parameters in combination with reconstruction errors to generate preliminary features and add them to the contrastive learning network and perform random enhancement, generate positive sample pairs and learn discriminative features through a contrastive loss function, add the discriminative features to an adaptive wavelet neural network, perform multi-scale feature decomposition by dynamically selecting the optimal wavelet basis function and the number of decomposition layers, obtain multi-scale features and add them to a dynamic graph neural network, extract node representations in combination with an improved graph convolution operator and update edge weights in combination with a gated attention mechanism to obtain a health feature tensor; The health feature tensor is added to the hierarchical memory enhancement network, and the feature information is stored by constructing a multi-layer memory unit with different time spans. The information flow is modified and the long-term and short-term dependencies are determined in combination with an adaptive gating mechanism to obtain the time series feature and add it to the dynamic routing capsule network. The time series feature is hierarchically expressed in combination with an improved iterative routing algorithm to obtain the hierarchical feature and add it to the causal reasoning network. The dependency between different health influencing factors is defined by a structural equation model and the causal hypothesis is verified in combination with a reverse intervention mechanism to obtain a verified causal relationship graph. The conditional random field in the probability graph attention network is combined to construct the conditional probability distribution between nodes and calculate the node importance weights to generate a health knowledge graph. The time series feature and the health knowledge graph are added to a deep Bayesian network, and the posterior distribution is approximated by variational inference and the uncertainty of the prediction result is estimated in combination with Monte Carlo sampling to obtain a health risk assessment result. The health risk assessment results are added to the hierarchical reinforcement learning network. By constructing a hierarchical policy structure, the upper network generates abstract goals to guide the lower network to select specific actions. The selected candidate action plans are input into the multi-agent collaborative optimization system. The collaborative optimization of group decision-making is achieved by designing the communication protocol and reward distribution mechanism between agents. The optimized strategy is input into the deep deterministic policy gradient network. The deterministic policy function is combined with the value function and the policy parameters are optimized by the gradient ascent method to achieve online learning. The learned policy parameters are input into the model predictive control network. The future state sequence is predicted by the rolling time domain optimization method and the control strategy is dynamically adjusted to obtain an adaptive monitoring solution.

2. The method according to claim 1, characterized in that: Obtain the multi-source heterogeneous health data corresponding to the user, define the forward diffusion process and add noise through the hierarchical diffusion probability model based on continuous-time stochastic differential equations, solve the differential equations in combination with the backward diffusion process for denoising, generate denoised data and add it to the improved attention autoencoder, calculate the correlation weights between features through the multi-head self-attention mechanism and encode them as potential representations, optimize the network parameters in combination with the reconstruction error to generate preliminary features and add them to the contrastive learning network and perform random enhancement, generate positive sample pairs and learn discriminative features through the contrastive loss function, add the discriminative features to the adaptive wavelet neural network, perform multi-scale feature decomposition by dynamically selecting the optimal wavelet basis function and the number of decomposition layers, obtain multi-scale features and add them to the dynamic graph neural network, extract node representations in combination with the improved graph convolution operator and update edge weights in combination with the gated attention mechanism, and obtain the health feature tensor including: The continuous data of electrocardiogram signals, blood pressure values, and blood oxygen concentration and the discrete data of movement status and sleep quality are collected through the sensor array of the wearable device, the continuous data and the discrete data are time-aligned and format-unified, the data at different time intervals are resampled into a standardized multidimensional time series, the missing values ​​in the multidimensional time series are supplemented by a combination of forward filling and linear interpolation, and the abnormal values ​​in the multidimensional time series are processed by sliding window median filtering to obtain preprocessed health data; Input the preprocessed health data into a hierarchical diffusion probability model, gradually inject Gaussian noise with increasing variance into the health data during the forward diffusion process, and construct a noise prediction network based on a U-shaped structure. The noise prediction network extracts multi-scale features and reconstructs detail information through a downsampling path and an upsampling path, respectively, adds a jump connection between the corresponding scales of the downsampling path and the upsampling path, predicts the distribution of the Gaussian noise through iterative optimization, and removes the noise during the reverse diffusion process to obtain the denoised health data; Input the denoised health data into an improved attention autoencoder, which maps the input data to a high-dimensional feature space through a multi-layer perceptron and then divides it into multiple attention heads, each attention head calculates a query vector, a key vector and a value vector respectively, calculates the dot product of the query vector and the key vector to obtain an attention score, performs a weighted sum of the attention score and the value vector to obtain a context representation, connects the context representations of multiple attention heads and obtains a feature representation through a linear transformation, reconstructs the original data through a decoder and optimizes the network parameters to obtain the encoded health features; Input the encoded health feature into a contrastive learning network, perform random time window cropping, amplitude modulation and random noise injection on the health feature to obtain an enhanced feature view, construct different feature views of the same health feature as a positive sample pair, construct feature views of different health features as a negative sample pair, constrain the distance of the positive sample pair in the feature space to be less than the distance of the negative sample pair through a contrast loss function, and obtain a discriminative health feature; Input the discriminative health feature into an adaptive wavelet neural network, evaluate the expressive power of different wavelet basis functions based on the time-frequency characteristics of the signal and select the optimal basis function family, determine the number of decomposition layers according to the frequency composition of the signal, perform multi-scale decomposition on the health feature to obtain sub-band coefficients of different frequency bands, perform feature fusion on the sub-band coefficients at each scale to obtain a multi-scale health feature; The multi-scale health features are constructed as a dynamic graph structure, and different health indicators are used as nodes in the graph. The neighborhood information of the nodes is aggregated through the graph convolution operator of the adaptive weight matrix, and the weight matrix is ​​dynamically updated based on the feature similarity between the nodes. The connection weights of the edges are adjusted according to the correlation strength between the nodes through the gated attention mechanism, and the node features extracted by the graph convolution are deeply fused with the edge features learned by the attention mechanism to output the health feature tensor.

3. The method according to claim 2, characterized in that The discriminative health features are input into the adaptive wavelet neural network, the expression capabilities of different wavelet basis functions are evaluated based on the time-frequency characteristics of the signal and the optimal basis function family is selected, the number of decomposition layers is determined according to the frequency composition of the signal, the health features are multi-scale decomposed to obtain sub-band coefficients of different frequency bands, and the sub-band coefficients are feature fused at each scale to obtain multi-scale health features including: Inputting discriminative health features into an adaptive wavelet neural network, wherein the adaptive wavelet neural network comprises a wavelet basis function evaluation module, a decomposition layer number determination module and a feature fusion module; The wavelet basis function evaluation module constructs a candidate wavelet basis function library, which includes a Haar wavelet basis function in a piecewise constant form, a Dobesi wavelet basis function family with different vanishing moment orders, a Schindler wavelet basis function family with different filter lengths, a Koivulet wavelet basis function family with symmetry, and a Mayer wavelet basis function with a smooth transition band in the frequency domain; Performing time-frequency analysis on the health characteristics to obtain frequency domain characteristics, and calculating a reconstruction error evaluation value, a sparsity evaluation value, and a detail preservation evaluation value of a wavelet basis function based on the frequency domain characteristics; Establishing a scoring matrix, normalizing the reconstruction error evaluation value, the sparsity evaluation value and the detail preservation evaluation value to obtain a normalized evaluation value, setting an adaptive weight coefficient based on the frequency domain feature, weighting the normalized evaluation value and the adaptive weight coefficient to obtain an expression ability score of each wavelet basis function, sorting the wavelet basis functions according to the expression ability score and selecting the wavelet basis function with the highest score as the optimal basis function; The decomposition layer number determination module performs frequency analysis on the health feature to obtain frequency composition information, and determines the wavelet decomposition layer number according to the number of main frequency components and frequency band distribution in the frequency composition information; Based on the optimal basis function and the wavelet decomposition layer number, the health feature is decomposed at multiple scales to obtain sub-band coefficients of different frequency bands, and the feature fusion module performs feature fusion on the sub-band coefficients of different frequency bands at various scales to obtain multi-scale health features; Monitor the degree of change of the frequency domain characteristics, the rate of change of statistical characteristics and the fluctuation of reconstruction quality. When the degree of change exceeds a preset threshold, reselect the optimal basis function. When the reconstruction quality is lower than a preset reconstruction quality threshold, randomly initialize the adaptive weight coefficient.

4. The method according to claim 1, characterized in that: The health feature tensor is added to the hierarchical memory enhancement network, and the feature information is stored by constructing a multi-layer memory unit with different time spans. The information flow is modified and the long-term and short-term dependencies are determined by combining the adaptive gating mechanism, and the time series features are obtained and added to the dynamic routing capsule network. The time series features are hierarchically expressed by combining the improved iterative routing algorithm, and the hierarchical features are obtained and added to the causal reasoning network. The dependency between different health influencing factors is defined by the structural equation model and the causal hypothesis is verified by combining the reverse intervention mechanism to obtain a verified causal relationship graph. The conditional probability distribution between nodes is constructed by combining the conditional random field in the probability graph attention network and the node importance weight is calculated to generate a health knowledge graph. The time series features and the health knowledge graph are added to the deep Bayesian network, and the posterior distribution is approximated by variational inference and the uncertainty of the prediction results is estimated by combining Monte Carlo sampling. The health risk assessment results include: The healthy feature tensor is input into the normalization module for normalization processing to obtain a standardized feature, and the standardized feature is input into the hierarchical memory enhancement network. The current input feature reception ratio is regulated by the input gating value of the short-term memory unit, the historical state retention degree is adjusted by the forgetting gating value, and the feature output strength is regulated by the output gating value to generate a short-term feature. The short-term feature is input into the medium-term memory unit and a query vector is generated. The similarity score between the query vector and the short-term feature is calculated, and the similarity score is normalized to obtain the attention weight. The short-term feature is multiplied by the attention weight to generate a medium-term feature. The short-term feature and the medium-term feature are combined by the residual connection of the long-term memory unit and a gated update is performed to generate a long-term feature. The change rate of the short-term feature, the medium-term feature and the long-term feature is calculated, and the state update ratio of the corresponding memory unit is adjusted according to the change rate by using the adaptive gating mechanism and the long-term dependency is determined, and the time series feature is generated by fusion. Input the time series features into multiple dedicated capsules of the dynamic routing capsule network, wherein each of the dedicated capsules corresponds to a specific health influencing factor, perform feature extraction on the time series features through a convolution transformation matrix and generate an activity vector, calculate the modulus length of the activity vector to represent the probability of feature existence, use the direction of the activity vector to represent the feature attribute to obtain a primary feature representation, initialize the routing coefficient matrix to be uniformly distributed, multiply the primary feature representation with the transformation matrix to generate a prediction vector, calculate the inner product of the prediction vector and the high-level capsule feature vector to obtain similarity, perform softmax normalization on the similarity to update the routing coefficient, perform weighted summation of the updated routing coefficient and the prediction vector to generate a feature vector of the high-level capsule, and repeat the update until the routing coefficient converges or reaches a maximum number of iterations to generate a hierarchical feature; Input the hierarchical features into the causal reasoning network, calculate the correlation coefficients between different health influencing factors based on the hierarchical features to construct a correlation matrix, use the structural equation model to decompose the correlation matrix into a direct effect matrix and an indirect effect matrix to construct an initial causal graph, calculate the conditional mutual information and Fisher conditional independence statistics for each pair of nodes with a connection relationship, screen the direct causal relationship that meets the conditional independence hypothesis according to the significance level, uniformly sample within the value range of the target node to generate multiple groups of intervention values, substitute each group of intervention values ​​into the causal graph and calculate the response values ​​of other nodes except the target node based on the structural equation model, calculate the mean square error between the response value and the historical data, remove the corresponding causal edge when the mean square error exceeds a preset threshold, and repeatedly calculate and delete the causal edge until the causal graph structure is stable to generate a verified causal relationship graph; Input the verified causal relationship graph into the probabilistic graph attention network, define a discrete state space for each node and count the state transition frequency between adjacent time steps, use maximum likelihood estimation to calculate the state transition probability to construct a transfer matrix, calculate the joint distribution probability between any two nodes based on the transfer matrix, calculate the corresponding conditional probability distribution for each node to represent the current node's dependence on other nodes, calculate the importance score of the node according to the conditional probability distribution and perform softmax normalization to obtain the attention weight, multiply the node feature with the corresponding attention weight and aggregate to generate a health knowledge graph that takes into account the importance of the node; The time series features and the health knowledge graph are input into the deep Bayesian network, the mean vector and covariance matrix of the variational distribution are initialized, the variational objective function is constructed including the log-likelihood term and the KL divergence term, the gradient of the variational objective function with respect to the variational parameters is calculated, the variational parameters are iteratively optimized using stochastic gradient descent until convergence to obtain the approximate posterior distribution, multiple Monte Carlo samplings are performed from the approximate posterior distribution to obtain model parameter samples, the corresponding model prediction output is calculated for each group of parameter samples, the mean of all model prediction outputs is statistically calculated as the final prediction result, and the variance of the model prediction output is calculated to evaluate the uncertainty, and a health risk assessment result including a risk level and a confidence interval is generated.

5. The method according to claim 4, characterized in that Input the verified causal relationship graph into the probabilistic graph attention network, define a discrete state space for each node and count the state transition frequency between adjacent time steps, calculate the state transition probability using maximum likelihood estimation to construct a transfer matrix, and calculate the joint distribution probability between any two nodes based on the transfer matrix, including: Inputting the verified causal relationship graph into the probabilistic graph attention network, the verified causal relationship graph includes multiple nodes representing different health influencing factors, determining the number of state categories of each node according to clinical medicine standards, dividing the indicator value of each node into multiple state intervals, establishing a mapping relationship from the indicator value to the state identifier, and generating a discrete state space of the node; An observation time window is set based on the discrete state space, and state transition statistics are independently performed on each of the nodes within the observation time window. The state of the node at each time step is identified by traversing the time series data, and the state transition of the node between adjacent time steps is recorded to create a state transition frequency matrix, wherein the rows and columns of the state transition frequency matrix correspond to all possible states of the node respectively; Sum all elements of each row in the state transition frequency matrix to obtain the total number of transition events of the corresponding state, divide each element in the state transition frequency matrix by the total number of transition events in its row to obtain a normalized transition probability, repeat the normalization operation for each row of the state transition frequency matrix to generate a transition probability matrix; Extract the transition probability matrices of any two nodes, combine the two transition probability matrices to obtain a joint probability matrix representing the combination of the two node states, repeat the combination operation for all node pairs in the causal relationship graph, and generate the joint distribution probability between all node pairs in the causal relationship graph.

6. The method according to claim 1, characterized in that The health risk assessment results are added to the hierarchical reinforcement learning network. By constructing a hierarchical strategy structure, the upper network generates abstract goals to guide the lower network to select specific actions. The selected candidate action plans are input into the multi-agent collaborative optimization system. The collaborative optimization of group decision-making is achieved by designing the communication protocol and reward distribution mechanism between agents. The optimized strategy is input into the deep deterministic policy gradient network. The deterministic policy function is combined with the value function and the policy parameters are optimized by the gradient ascent method to achieve online learning. The learned policy parameters are input into the model predictive control network. The future state sequence is predicted by the rolling time domain optimization method and the control strategy is dynamically adjusted. The adaptive monitoring scheme includes: Adding the health risk assessment results to a hierarchical reinforcement learning network, wherein the hierarchical reinforcement learning network generates an abstract goal in the upper network to guide the lower network to select specific actions by constructing a hierarchical strategy structure, wherein the hierarchical strategy structure includes a state encoding module, a goal generation module and a strategy mapping module, wherein the state encoding module encodes the risk level, abnormality degree and change trend in the health risk assessment results into a state vector, the goal generation module generates an abstract goal of monitoring frequency and warning threshold based on the state vector, and the strategy mapping module converts the abstract goal into a candidate action plan; The candidate action plan is input into a multi-agent collaborative optimization system, the multi-agent collaborative optimization system sets a communication protocol and a reward allocation mechanism between agents, the communication protocol defines the transmission rules of status information messages, action suggestion messages and evaluation feedback messages, the reward allocation mechanism allocates a reward value to each agent according to data collection quality, prediction accuracy and strategy execution effect, and realizes collaborative optimization of group decision-making based on the communication protocol and the reward allocation mechanism; The optimized strategy is input into a deep deterministic policy gradient network, which combines a deterministic policy function with a value function. The deterministic policy function outputs a specific monitoring action, and the value function evaluates the long-term value of the monitoring action. The policy parameters of the deterministic policy function are continuously optimized by a gradient ascent method, and the policy parameters are input into a model predictive control network. The model predictive control network predicts the future state sequence by a rolling horizon optimization method, searches for the optimal control sequence within the prediction range based on the future state sequence, executes the control action at the current moment and receives state feedback, dynamically adjusts the policy parameters, and generates an adaptive monitoring solution.

7. The method according to claim 6, characterized in that The model predictive control network predicts the future state sequence by a rolling horizon optimization method, searches for the optimal control sequence within the prediction range based on the future state sequence, executes the control action at the current moment and receives state feedback, dynamically adjusts the strategy parameters, and generates an adaptive monitoring scheme, including: Constructing a basic architecture of a dynamic prediction model, the dynamic prediction model includes an input layer, a data preprocessing module, a feature extraction layer and a state encoding layer, the input layer receives a physiological indicator numerical sequence, a monitoring time series and an environmental parameter, the data preprocessing module performs standardization processing on the physiological indicator numerical sequence, the monitoring time series and the environmental parameter to obtain normalized data, the feature extraction layer extracts time series features from the normalized data through a bidirectional long short-term memory network, and the state encoding layer encodes the time series features into a state vector of fixed dimension; Generate a future state sequence through a prediction controller based on the state vector, input the state vector into a prediction model to obtain a state prediction value at the next moment, and use the state prediction value as a new input to continue predicting to obtain a complete state sequence in a prediction time domain, wherein the prediction time domain is divided into a plurality of prediction steps, and each prediction step corresponds to a state prediction value; The optimal control sequence is determined by adopting a hierarchical search strategy, wherein the first layer determines the reference value of the monitoring frequency according to the fluctuation degree of the complete state sequence, the second layer performs a local search near the reference value to obtain the optimized monitoring frequency, and the third layer dynamically adjusts the value range of the warning threshold based on the complete state sequence, and the optimal control sequence is determined by evaluating the comprehensive effects of different control schemes through a heuristic algorithm; Sending the monitoring frequency and early warning threshold in the optimal control sequence to the monitoring device to start data collection, recording actual monitoring data, wherein the actual monitoring data includes a monitoring value sequence, a monitoring time sequence, and an early warning trigger record, calculating the error between the state prediction value and the actual monitoring data, constructing a loss function based on the error and using the stochastic gradient descent method to update the parameters of the dynamic prediction model, and triggering the adjustment of the dynamic prediction model structure when the error exceeds a set threshold; The control strategy is re-optimized based on the updated dynamic prediction model, the monitoring frequency and the warning threshold are adaptively adjusted according to the weight relationship between the prediction accuracy and the monitoring burden, and the adjusted control strategy is used for the optimization process of the next prediction cycle.

8. A dynamic health adaptive monitoring system using artificial intelligence, used to implement the method of any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain multi-source heterogeneous health data corresponding to the user, define the forward diffusion process and add noise through the hierarchical diffusion probability model based on continuous-time stochastic differential equations, solve the differential equations in combination with the backward diffusion process for noise reduction, generate noise-reduced data and add it to the improved attention autoencoder, calculate the correlation weights between features through the multi-head self-attention mechanism and encode them as potential representations, optimize the network parameters in combination with the reconstruction error to generate preliminary features and add them to the contrastive learning network and perform random enhancement, generate positive sample pairs and learn discriminative features through the contrastive loss function, add the discriminative features to the adaptive wavelet neural network, perform multi-scale feature decomposition by dynamically selecting the optimal wavelet basis function and the number of decomposition layers, obtain multi-scale features and add them to the dynamic graph neural network, extract node representations in combination with the improved graph convolution operator and update edge weights in combination with the gated attention mechanism to obtain a health feature tensor; The second unit is used to add the health feature tensor to the hierarchical memory enhancement network, store feature information by constructing multi-layer memory units with different time spans, modify information flow and determine long-term and short-term dependencies in combination with an adaptive gating mechanism, obtain time series features and add them to the dynamic routing capsule network, express the time series features in a hierarchical manner in combination with an improved iterative routing algorithm, obtain hierarchical features and add them to the causal reasoning network, define the dependencies between different health influencing factors through a structural equation model and verify the causal hypothesis in combination with a reverse intervention mechanism, obtain a verified causal relationship graph, construct the conditional probability distribution between nodes in combination with the conditional random field in the probability graph attention network and calculate the node importance weights, generate a health knowledge graph, add the time series features and the health knowledge graph to a deep Bayesian network, approximate the posterior distribution through variational reasoning and estimate the uncertainty of the prediction results in combination with Monte Carlo sampling, and obtain a health risk assessment result; The third unit is used to add the health risk assessment results to the hierarchical reinforcement learning network, construct a hierarchical policy structure so that the upper network generates abstract goals to guide the lower network to select specific actions, input the selected candidate action plans into the multi-agent collaborative optimization system, and realize the collaborative optimization of group decision-making by designing the communication protocol and reward distribution mechanism between agents. The optimized strategy is input into the deep deterministic policy gradient network, the deterministic policy function is combined with the value function and the policy parameters are optimized by the gradient ascent method to realize online learning, the learned policy parameters are input into the model predictive control network, and the future state sequence is predicted by the rolling time domain optimization method and the control strategy is dynamically adjusted to obtain an adaptive monitoring solution.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images