Knowledge graph-based health status assessment and product matching method and system

Abstract

The invention provides a health state evaluation and product matching method and system based on a knowledge graph, and relates to the technical field of knowledge graphs, and the method comprises the steps: obtaining user health data, building a health knowledge graph, mapping the user data into a feature vector, and employing a multi-scale time convolution network to extract a time sequence change feature; and constructing a dynamic health feature association network, generating a health state evaluation result by using a graph convolutional neural network and a causal discovery algorithm, calculating a collaborative enhancement coefficient based on a product feature vector, and predicting an intervention effect to determine an optimal product combination. According to the method, personalized health assessment and precise product matching are realized, and the health management efficiency is improved.

Description

Technical Field

[0001] This invention relates to knowledge graph technology, and more particularly to a method and system for health status assessment and product matching based on knowledge graphs. Background Technology

[0002] As people's health awareness continues to rise, health status assessment and product recommendation based on personal health data are gradually becoming important aspects of health management. Knowledge graphs, as a structured knowledge representation method, can effectively integrate the complex relationships between medical knowledge, health rules, and product characteristics, providing knowledge support for health status assessment and product matching.

[0003] However, existing health status assessment and product matching technologies have the following shortcomings: First, existing technologies typically use static thresholds to process the correlations between user health characteristics, making it difficult to adapt to individual differences among users and the dynamic changes in their health indicators. This results in a lack of flexibility and personalization in the construction of feature association networks. Second, in the health status assessment process, existing technologies often ignore the causal relationships between health characteristics, relying solely on correlation analysis for judgment. This makes it difficult to reveal the essential causes and development trends of health status, thus affecting the accuracy and interpretability of the assessment results. Finally, existing technologies typically use simple rule matching or similarity calculations in the product matching stage, failing to fully consider the synergistic effects between different products and their comprehensive intervention effects on user health status, leading to poor performance in recommending product combinations. Summary of the Invention

[0004] This invention provides a knowledge graph-based method and system for health status assessment and product matching, which can solve the problems in the prior art.

[0005] A first aspect of this invention provides a knowledge graph-based method for health status assessment and product matching, comprising:

[0006] Acquire user health data; establish a health knowledge graph that includes medical knowledge nodes, health assessment rule nodes, and product feature nodes;

[0007] The user health data is mapped to a user feature vector; a multi-scale temporal convolutional network is used to extract the temporal variation features of the user feature vector in daily, weekly and monthly cycles, the fluctuation standard deviation of the temporal variation features is calculated, a dynamic association threshold is generated, feature associations are filtered based on the dynamic association threshold, and a dynamic health feature association network is constructed.

[0008] The dynamic health feature association network is input into a graph convolutional neural network to obtain feature representations. The feature representations are then processed by a structured causal discovery algorithm to establish a feature causal relationship network. A state transition probability matrix is ​​constructed based on the feature causal relationship network. Causal intervention analysis is then performed based on the state transition probability matrix and the health assessment rule nodes in the health knowledge graph to generate a health status assessment result.

[0009] Construct a product feature vector that includes efficacy and mechanism of action dimensions; calculate the synergistic enhancement coefficient between candidate products based on the product feature vector; predict the intervention effect on the health status assessment result for each candidate product combination; and determine the optimal product combination based on the intervention effect and the synergistic enhancement coefficient.

[0010] In one alternative implementation,

[0011] A multi-scale temporal convolutional network is used to extract the temporal variation features of the user feature vector in daily, weekly, and monthly cycles. The standard deviation of the fluctuation of the temporal variation features is calculated to generate a dynamic association threshold. Based on the dynamic association threshold, feature associations are filtered, and a dynamic health feature association network is constructed, including:

[0012] Within a preset time window, the first-order difference value of the user feature vector is calculated to generate a fluctuation feature matrix, and the fluctuation density index is calculated based on the fluctuation feature matrix.

[0013] A multi-scale temporal convolutional network is constructed, which includes daily-scale convolutional branches, weekly-scale convolutional branches, and monthly-scale convolutional branches. The dynamic expansion rate of each branch is calculated using the hyperbolic tangent function based on the fluctuation density index, and the dynamic expansion rate is smoothed by performing an exponential moving average on the dynamic expansion rate.

[0014] The user feature vector is input into the multi-scale temporal convolutional network, and convolution is performed using the smooth dilation rate to extract temporal variation features.

[0015] The temporal consistency index and prediction error index of the temporal variation features are calculated. The temporal consistency index, prediction error index and signal-to-noise ratio score are combined to form a reliability vector. The reliability vector is input into a nonlinear mapping function to obtain feature weights. The temporal variation features are weighted and fused according to the feature weights to obtain multi-scale fusion features. At the same time, temporal smoothing constraints and entropy regularization constraints are introduced.

[0016] The feature fluctuation standard deviation is calculated based on the multi-scale fusion features. A dynamic correlation threshold is calculated based on the feature fluctuation standard deviation and the health risk score. The dynamic correlation threshold increases with the increase of the feature fluctuation standard deviation and decreases with the increase of the health risk score.

[0017] Calculate the association strength of feature pairs in the multi-scale fusion features, retain the association strengths that are greater than the dynamic association threshold, and construct a dynamic health feature association network.

[0018] In one alternative implementation,

[0019] The dynamic health feature association network is input into a graph convolutional neural network to obtain feature representations. These feature representations are then processed using a structured causal discovery algorithm to establish a feature causal relationship network, including:

[0020] A multi-head attention map convolutional layer is constructed. Attention coefficients between nodes in the dynamic health feature association network are calculated based on attention vectors. Feature transformation and multi-head feature aggregation are performed on node features according to these attention coefficients to obtain an initial feature representation. This initial feature representation is then input into a temporal-aware residual module, where features at different levels are adaptively fused through gated skip connections to obtain temporal-enhanced features. A multi-scale feature pyramid is constructed, and convolutional processing and feature pooling are performed on the temporal-enhanced features with different receptive fields to generate a multi-scale feature sequence. This multi-scale feature sequence is then upsampled and fused based on learnable weights to obtain hierarchical features.

[0021] Based on the maximization of neural mutual information, the conditional mutual information between feature pairs in the hierarchical features is calculated, and an initial feature causal relationship network is constructed based on the conditional mutual information.

[0022] Calculate the direct and indirect causal effects between feature pairs in the initial feature causal relationship network, and calculate the time-varying causal strength based on the direct and indirect causal effects;

[0023] An adaptive threshold is generated based on the time-varying causal strength and health risk score, and causal associations greater than the adaptive threshold are retained to construct a feature causal relationship network.

[0024] In one alternative implementation,

[0025] Calculating the conditional mutual information between feature pairs in the hierarchical features based on maximizing neural mutual information, and constructing an initial feature causal relationship network based on the conditional mutual information includes:

[0026] The hierarchical features are subjected to time dimension displacement transformation and feature dimension perturbation transformation to generate comparative feature samples of the hierarchical features.

[0027] Calculate the initial estimate of the conditional mutual information between the hierarchical features and the contrastive feature samples;

[0028] A conditional variable attention network is constructed, with the initial estimate as input, and conditional variable weights are generated based on attention calculation. The most relevant conditional variable is selected according to the conditional variable weights.

[0029] The hierarchical features are projected onto multiple orthogonal feature spaces to obtain multi-view feature representations, and multi-view conditional mutual information is calculated based on the multi-view feature representations and the most relevant conditional variables.

[0030] Calculate the temporal fluctuation intensity of the hierarchical features, construct an adaptive kernel function based on the temporal fluctuation intensity, and apply the adaptive kernel function to the multi-view conditional mutual information calculation to obtain the optimized multi-view conditional mutual information;

[0031] Product feature association information is extracted from the health knowledge graph, and the product feature association strength is calculated based on graph reasoning.

[0032] The optimized multi-view conditional mutual information and the product feature association strength are combined using a Bayesian weighted fusion method to construct an initial feature causal relationship network. The weight of each edge in the initial feature causal relationship network is calculated using the Bayesian weighted fusion method.

[0033] In one alternative implementation,

[0034] Based on the aforementioned feature causal relationship network, a state transition probability matrix is ​​constructed. Causal intervention analysis is then performed based on the state transition probability matrix and the health assessment rule nodes in the health knowledge graph to generate health status assessment results, including:

[0035] A multidimensional health state vector is constructed, which includes physiological state components, psychological state components, and behavioral state components. A state transition space is constructed based on the node features in the feature causal relationship network, and a state distribution sequence of the multidimensional health state vector is generated in the state transition space.

[0036] Based on the aforementioned feature causal relationship network, all possible state transition paths are extracted, and each state transition path and its corresponding state distribution sequence are input into a multi-layer neural network to obtain a path importance score.

[0037] Attention weights are constructed based on the path importance scores, and the transition probabilities of different state transition paths are weighted and combined according to the attention weights to obtain a state transition probability matrix.

[0038] Based on the medical knowledge nodes in the health knowledge graph, evaluation rule nodes for different health states are extracted, and the evaluation rule nodes are represented as rule vectors. A rule attention computation network is constructed based on the rule vectors.

[0039] The state transition probability matrix is ​​input into the rule-based attention computation network to obtain a rule-enhanced state representation. Based on the rule-enhanced state representation, the direct intervention effect and the indirect intervention effect are calculated. A hierarchical intervention evaluation structure is constructed based on the direct intervention effect and the indirect intervention effect. The hierarchical intervention evaluation structure includes a state-level intervention unit, a transition-level intervention unit, and a rule-level intervention unit.

[0040] The multidimensional health status vector is input into the hierarchical intervention assessment structure, and sub-assessment results are generated based on each intervention unit. The health status assessment result is obtained by adaptively weighting and fusing the sub-assessment results.

[0041] In one alternative implementation,

[0042] Constructing product feature vectors that include efficacy and mechanism of action dimensions; calculating synergistic enhancement coefficients among candidate products based on the product feature vectors; predicting the intervention effect on the health status assessment results for each candidate product combination; and determining the optimal product combination based on the intervention effect and the synergistic enhancement coefficients, including:

[0043] Product efficacy node information and active ingredient mechanism information are extracted from the health knowledge graph and weighted and fused to obtain the product feature vector.

[0044] The product feature vector and health status assessment results are encoded into a composite state vector; a deep inverse reward network and a two-layer attention mechanism are constructed based on the composite state vector, and a candidate product combination is generated and evaluated using an attention-based policy network, and the synergistic enhancement coefficient is obtained by fusion.

[0045] Based on the product feature vector, predict the probability of the health status impact of the product portfolio, construct a health status transition sequence, substitute the probability of the health status impact into the health status transition sequence, and obtain the predicted health status values ​​at multiple time points; calculate the state fluctuation amplitude at adjacent time points based on the predicted health status values, and calculate the cumulative intervention effect and intervention stability of the product portfolio based on the state fluctuation amplitude.

[0046] The optimal product combination is determined based on the synergistic enhancement coefficient, cumulative intervention effect, and intervention stability.

[0047] In one alternative implementation,

[0048] The product feature vector and health status assessment results are encoded into a composite state vector. A deep inverse reward network and a two-layer attention mechanism are constructed based on this composite state vector. An attention-based policy network is used to generate and evaluate candidate product combinations, and the resulting synergistic enhancement coefficients are obtained by fusing them.

[0049] The product feature vector and health status assessment results are encoded into a composite state vector, which includes a product efficacy intensity component, a mechanism component, and a health impact component; valid expert recommendation records are selected based on the degree of improvement of the health impact component in the target composite state vector.

[0050] A deep inverse reward network is constructed. Based on the effective expert recommendation records, the implicit reward function of expert decision-making is learned through a multi-layer neural network. The loss function of the deep inverse reward network includes a state prediction error term, a reward consistency constraint term, and a temporal smoothing constraint term.

[0051] A two-layer attention mechanism is constructed based on the composite state vector: the first layer is a product-level attention module, which calculates the product combination importance weight based on the similarity matrix of product features in the composite state vector; the second layer is a feature-level attention module, which calculates the feature association weight based on the interaction matrix of the efficacy intensity component and the mechanism component; the product combination importance weight and the feature association weight are fused to obtain the combination attention weight;

[0052] Construct an attention-based policy network, input the composite state vector, and generate a probability distribution of candidate product combinations based on the combined attention weights; use Thompson sampling to sample from the probability distribution to generate combination schemes, and evaluate the combination schemes based on the implicit reward function;

[0053] The reward value of the deep inverse reward network, the combined attention weights, and the evaluation value of the combined scheme are weighted and fused to obtain the collaborative enhancement coefficient.

[0054] A second aspect of this invention provides a knowledge graph-based health status assessment and product matching system, comprising:

[0055] The first unit is used to acquire user health data and establish a health knowledge graph that includes medical knowledge nodes, health assessment rule nodes, and product feature nodes.

[0056] The second unit is used to map the user health data into user feature vectors; it uses a multi-scale temporal convolutional network to extract the temporal variation features of the user feature vectors in daily, weekly and monthly cycles, calculates the fluctuation standard deviation of the temporal variation features, generates a dynamic association threshold, filters feature associations based on the dynamic association threshold, and constructs a dynamic health feature association network.

[0057] The third unit is used to input the dynamic health feature association network into a graph convolutional neural network to obtain feature representations, process the feature representations through a structured causal discovery algorithm, establish a feature causal relationship network, construct a state transition probability matrix based on the feature causal relationship network, and perform causal intervention analysis according to the state transition probability matrix and the health assessment rule nodes in the health knowledge graph to generate health status assessment results.

[0058] The fourth unit is used to construct a product feature vector that includes efficacy and mechanism of action dimensions; calculate the synergistic enhancement coefficient between candidate products based on the product feature vector; predict the intervention effect on the health status assessment result for each candidate product combination; and determine the optimal product combination based on the intervention effect and the synergistic enhancement coefficient.

[0059] A third aspect of the present invention provides an electronic device, comprising:

[0060] processor;

[0061] Memory used to store processor-executable instructions;

[0062] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0063] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0064] The present invention has the following beneficial effects:

[0065] By constructing a comprehensive health knowledge graph that includes medical knowledge nodes, health assessment rule nodes, and product feature nodes, and combining it with a multi-scale temporal convolutional network to extract the temporal features of user health data, a precise and dynamic assessment of user health status is achieved, overcoming the limitations of traditional static assessment methods.

[0066] We employ dynamic correlation thresholding and graph convolutional neural network techniques to process the correlation of health features, and establish a feature causal relationship network through a structured causal discovery algorithm. This makes health status assessment interpretable, enables us to trace the root causes of health problems, and provides a scientific basis for subsequent interventions.

[0067] In terms of product matching, by constructing multi-dimensional product feature vectors and calculating the synergistic enhancement coefficient between candidate products, the intervention effect of product combinations on health status can be predicted, thereby achieving personalized and accurate product recommendations, which significantly improves the effectiveness of product matching and user satisfaction. Attached Figure Description

[0068] Figure 1This is a flowchart illustrating the knowledge graph-based health status assessment and product matching method according to an embodiment of the present invention.

[0069] Figure 2 A comparison chart of the accuracy of conditional mutual information estimation methods;

[0070] Figure 3 A time-dynamic analysis chart of the effects of health interventions;

[0071] Figure 4 Flowchart for generating product collaborative enhancement coefficients based on deep inverse reward networks and two-layer attention mechanisms. Detailed Implementation

[0072] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0073] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0074] Figure 1 This is a flowchart illustrating the knowledge graph-based health status assessment and product matching method of this invention, as shown in the embodiment. Figure 1 As shown, the method includes:

[0075] Acquire user health data; establish a health knowledge graph that includes medical knowledge nodes, health assessment rule nodes, and product feature nodes;

[0076] The user health data is mapped to a user feature vector; a multi-scale temporal convolutional network is used to extract the temporal variation features of the user feature vector in daily, weekly and monthly cycles, the fluctuation standard deviation of the temporal variation features is calculated, a dynamic association threshold is generated, feature associations are filtered based on the dynamic association threshold, and a dynamic health feature association network is constructed.

[0077] The dynamic health feature association network is input into a graph convolutional neural network to obtain feature representations. The feature representations are then processed by a structured causal discovery algorithm to establish a feature causal relationship network. A state transition probability matrix is ​​constructed based on the feature causal relationship network. Causal intervention analysis is then performed based on the state transition probability matrix and the health assessment rule nodes in the health knowledge graph to generate a health status assessment result.

[0078] Construct a product feature vector that includes efficacy and mechanism of action dimensions; calculate the synergistic enhancement coefficient between candidate products based on the product feature vector; predict the intervention effect on the health status assessment result for each candidate product combination; and determine the optimal product combination based on the intervention effect and the synergistic enhancement coefficient.

[0079] A health knowledge graph is constructed, integrating three key nodes: medical knowledge nodes (containing professional medical information such as disease symptoms, etiology, and diagnostic criteria), health assessment rule nodes (containing judgment rules such as health risk assessment standards and clinical guidelines), and product characteristic nodes (containing characteristics such as product efficacy, mechanism of action, and applicable population). A structured health knowledge network is built through a three-stage process: First, knowledge extraction is performed, using existing natural language processing technologies to extract structured information from medical literature, clinical guidelines, and product instructions. Key entities are identified using a named entity recognition method based on Conditional Random Fields (CRF), and relationships between entities are extracted using relation extraction techniques based on dependency parsing. Second, entity linking is implemented, mapping the same concept expressed differently (e.g., "hypertension" and "elevated blood pressure") to a unified medical ontology through medical terminology normalization. A dictionary-based matching method combined with TF-IDF similarity calculation is used to solve the problem of ambiguous terms. Finally, relational reasoning is performed, combining rule-based reasoning methods to fill in implicit relationships in the knowledge graph, such as inferring "Product A may affect mechanism C" from "Product A contains ingredient B" and "Ingredient B inhibits mechanism C". The system also uses traditional association rule mining algorithms to discover potential relationships between entities and applies path analysis-based methods to capture indirect relationships. These technologies together construct a multidimensional health knowledge network that includes medical knowledge, health assessment rules, and product characteristics.

[0080] In one optional implementation, a multi-scale temporal convolutional network is used to extract the temporal variation features of the user feature vector over daily, weekly, and monthly periods. The standard deviation of the fluctuation of the temporal variation features is calculated to generate a dynamic association threshold. Based on the dynamic association threshold, feature associations are filtered, and a dynamic health feature association network is constructed, including:

[0081] Within a preset time window, the first-order difference value of the user feature vector is calculated to generate a fluctuation feature matrix, and the fluctuation density index is calculated based on the fluctuation feature matrix.

[0082] A multi-scale temporal convolutional network is constructed, which includes daily-scale convolutional branches, weekly-scale convolutional branches, and monthly-scale convolutional branches. The dynamic expansion rate of each branch is calculated using the hyperbolic tangent function based on the fluctuation density index, and the dynamic expansion rate is smoothed by performing an exponential moving average on the dynamic expansion rate.

[0083] The user feature vector is input into the multi-scale temporal convolutional network, and convolution is performed using the smooth dilation rate to extract temporal variation features.

[0084] The temporal consistency index and prediction error index of the temporal variation features are calculated. The temporal consistency index, prediction error index and signal-to-noise ratio score are combined to form a reliability vector. The reliability vector is input into a nonlinear mapping function to obtain feature weights. The temporal variation features are weighted and fused according to the feature weights to obtain multi-scale fusion features. At the same time, temporal smoothing constraints and entropy regularization constraints are introduced.

[0085] The feature fluctuation standard deviation is calculated based on the multi-scale fusion features. A dynamic correlation threshold is calculated based on the feature fluctuation standard deviation and the health risk score. The dynamic correlation threshold increases with the increase of the feature fluctuation standard deviation and decreases with the increase of the health risk score.

[0086] Calculate the association strength of feature pairs in the multi-scale fusion features, retain the association strengths that are greater than the dynamic association threshold, and construct a dynamic health feature association network.

[0087] For example, the user feature vector is processed within a preset time window. Assume the user feature vector is X, containing multi-dimensional health indicator data, such as blood pressure, blood sugar, and heart rate, each forming a time series. Within the preset 30-day time window, the first-order difference value is calculated for each feature dimension, i.e., the difference between adjacent time points, generating a fluctuation feature matrix D. Specifically, if the value of the original feature vector at time t is X(t), then the first-order difference value D(t) = X(t) - X(t-1). Taking blood pressure as an example, if the systolic blood pressure for two consecutive days is 120 mmHg and 125 mmHg respectively, then the difference value is 5 mmHg. Based on the fluctuation feature matrix D, the fluctuation density index P is calculated, which reflects the frequency of feature changes within the time window. The fluctuation density index P can be obtained by calculating the proportion of differences exceeding a preset threshold; for example, if the diurnal difference in blood pressure exceeds 5 mmHg for 15 days within 30 days, then the fluctuation density is 0.5.

[0088] Next, a multi-scale temporal convolutional network with three convolutional branches is constructed: a daily-scale convolutional branch, a weekly-scale convolutional branch, and a monthly-scale convolutional branch. The daily-scale convolutional branch consists of three one-dimensional convolutional layers, each containing 32 3×1 convolutional kernels, primarily capturing changes in health indicators over 1-2 days. The weekly-scale convolutional branch consists of three one-dimensional convolutional layers, each containing 24 5×1 convolutional kernels, focusing on capturing cyclical patterns over 3-7 days. The monthly-scale convolutional branch consists of three one-dimensional convolutional layers, each containing 16 7×1 convolutional kernels, used to identify long-term health trends over 20-30 days. Each branch's convolutional layers are followed by batch normalization layers and ReLU activation functions to improve training stability and expressive power. Finally, each branch uses a global average pooling layer to convert feature sequences of different lengths into fixed-dimensional feature vectors. For each branch, the dynamic expansion rate r is calculated using the hyperbolic tangent function based on the volatility density index P. The dilation rate determines the receptive field size of the convolution kernel. The more frequent the fluctuations, the smaller the dilation rate, in order to capture finer-grained changes. For example, when the blood pressure fluctuation density P is 0.5, the initial dilation rate obtained by mapping through the hyperbolic tangent function may be 2; while when the fluctuation density P is 0.8, the initial dilation rate may be reduced to 1 to better capture frequent fluctuations. An exponential moving average is performed on the initial dilation rate r to obtain a smoothed dilation rate r', avoiding model instability caused by drastic changes in the dilation rate. If the historical smoothed dilation rate is r'(t-1) and the currently calculated dilation rate is r(t), then the new smoothed dilation rate r'(t) = 0.9 × r'(t-1) + 0.1 × r(t).

[0089] The user feature vector X is input into the constructed multi-scale temporal convolutional network, and convolution is performed using a smooth dilation rate r' to extract temporal variation features F. The daily-scale branch uses a small convolutional kernel (e.g., 3×1) and a small dilation rate (e.g., 1-2) to capture short-term changes; the weekly-scale branch uses a medium-sized convolutional kernel (e.g., 5×1) and a medium dilation rate (e.g., 3-7) to capture medium-term changes; and the monthly-scale branch uses a large convolutional kernel (e.g., 7×1) and a large dilation rate (e.g., 7-15) to capture long-term changes.

[0090] For the extracted temporal variation features F, a temporal consistency index C, a prediction error index E, and a signal-to-noise ratio (SNR) score S are calculated to form a reliability vector [C, E, S]. The temporal consistency index C is obtained by calculating the similarity of features within a continuous time window; a higher value indicates a more stable feature. The prediction error index E is obtained by comparing the difference between predicted and actual values; a lower value indicates a more predictable feature. The SNR score S is obtained by calculating the ratio of feature signal strength to noise level; a higher value indicates better feature quality. The reliability vector [C, E, S] is input into a nonlinear mapping function, which can be a sigmoid function, to control the weights of each index. For example, for the blood pressure feature, if its temporal consistency index is 0.8, its prediction error index is 0.1, and its SNR is 15, its weight might be calculated as 0.9; while for the weight feature, which fluctuates significantly, its weight might only be 0.6. The temporal variation features F are then weighted and fused according to their weights w to obtain a multi-scale fused feature G. Simultaneously, temporal smoothing constraints are introduced to ensure that the fused features change smoothly over time; entropy regularization constraints are also introduced to ensure that the features contain sufficient information.

[0091] The feature fluctuation standard deviation σ is calculated based on the multi-scale fusion feature G, that is, the standard deviation of G at each time point within the time window. A dynamic association threshold T is calculated based on the feature fluctuation standard deviation σ and the health risk score R: T = α × σ - β × R + γ, where α, β, and γ are positive hyperparameters. When the feature fluctuation standard deviation σ increases, the dynamic association threshold T increases, filtering out more associations that may be caused by noise; when the health risk score R increases, the dynamic association threshold T decreases, retaining more associations that may indicate health problems. For example, when the fluctuation standard deviation is 0.2 and the health risk score is 0.7, if α = 1, β = 0.5, and γ = 0.3, then the dynamic association threshold T = 0.2 - 0.35 + 0.3 = 0.15.

[0092] The correlation strength L between feature pairs in the multi-scale fusion feature G can be calculated using methods such as Pearson correlation coefficient or mutual information. Correlation strengths greater than the dynamic correlation threshold T are retained, constructing a dynamic health feature correlation network N. In network N, nodes represent health features, edges represent correlations between features, and edge weights represent correlation strengths. For example, if the correlation strength between blood pressure and heart rate is 0.25 and the dynamic correlation threshold is 0.15, this correlation is retained in the network; however, if the correlation strength between blood pressure and body temperature is 0.12, which is below the threshold, this correlation is not established in the network.

[0093] The multi-scale temporal convolutional network implementation scheme of this invention solves the problem of capturing the multi-period variation characteristics of health data. By dynamically adjusting the receptive field size through dynamic expansion rate, it effectively handles health fluctuations at different time scales; by employing reliability vector weighted fusion, it improves the stability of feature extraction; and the dynamic association threshold mechanism can intelligently adjust according to the standard deviation of fluctuation and health risk, preserving key associations while filtering noise, significantly improving the accuracy and reliability of subsequent causal analysis.

[0094] In one optional implementation, the dynamic health feature association network is input into a graph convolutional neural network to obtain feature representations. The feature representations are then processed using a structured causal discovery algorithm to establish a feature causal relationship network, including:

[0095] A multi-head attention map convolutional layer is constructed. Attention coefficients between nodes in the dynamic health feature association network are calculated based on attention vectors. Feature transformation and multi-head feature aggregation are performed on node features according to these attention coefficients to obtain an initial feature representation. This initial feature representation is then input into a temporal-aware residual module, where features at different levels are adaptively fused through gated skip connections to obtain temporal-enhanced features. A multi-scale feature pyramid is constructed, and convolutional processing and feature pooling are performed on the temporal-enhanced features with different receptive fields to generate a multi-scale feature sequence. This multi-scale feature sequence is then upsampled and fused based on learnable weights to obtain hierarchical features.

[0096] Based on the maximization of neural mutual information, the conditional mutual information between feature pairs in the hierarchical features is calculated, and an initial feature causal relationship network is constructed based on the conditional mutual information.

[0097] Calculate the direct and indirect causal effects between feature pairs in the initial feature causal relationship network, and calculate the time-varying causal strength based on the direct and indirect causal effects;

[0098] An adaptive threshold is generated based on the time-varying causal strength and health risk score, and causal associations greater than the adaptive threshold are retained to construct a feature causal relationship network.

[0099] For example, when constructing a multi-head attention map convolutional layer, the system assigns an initial feature vector of dimension 64 to each healthy feature node. For any two connected nodes i and j, the attention coefficient between them is calculated. The feature vectors of nodes i and j are transformed using the parameter matrix W respectively to obtain transformed feature vectors. These two transformed feature vectors are concatenated, and then an inner product operation is performed using the attention vector a. Finally, the LeakyReLU activation function is applied to obtain the unnormalized attention coefficients. The softmax function is used to normalize the attention coefficients of all neighboring nodes of node i to obtain the final attention coefficients.

[0100] Eight attention heads are set up, each with an output dimension of 8. Attention coefficients and feature transformations are calculated independently for each attention head. Taking blood pressure and blood glucose as two health feature nodes as an example, the initial attention coefficient is 0.23, which becomes 0.18 after normalization, representing the influence of the blood pressure feature on the blood glucose feature. The output features of the eight attention heads are concatenated and then processed through the ELU activation function to obtain an initial feature representation with a dimension of 64.

[0101] The initial feature representation is input into the temporal-aware residual module, which contains a gated skip connection structure. For each time step, the feature vectors of the current time step and the feature vectors of the historical time steps are extracted and processed through two independent fully connected layers to obtain the current information representation and the historical information representation. The temporal gating weights are calculated using the sigmoid function, with values ​​ranging from 0 to 1, to adaptively adjust the importance of the current and historical information. The weighted current and historical information are added together and then residually connected to the input features to form the temporal augmentation feature.

[0102] When constructing the multi-scale feature pyramid, three different convolutional kernel sizes—3×3, 5×5, and 7×7—were applied to the temporal enhancement features to capture feature dependencies at the local, intermediate, and global scales, respectively. Taking heart rate variability features as an example, the 3×3 convolutional kernel captures short-term fluctuation patterns, the 5×5 kernel captures medium-term trends, and the 7×7 kernel captures long-term health status changes. Average pooling and max pooling operations were performed on the three convolutional results to obtain six sets of multi-scale features. Learnable weights were used to upsample and fuse these multi-scale features. The weights were automatically optimized through backpropagation, with an initial value of 0.167 for each. The final hierarchical feature structure has a dimension of 128, encompassing multi-scale temporal health information.

[0103] When calculating conditional mutual information based on maximizing neural mutual information, two parameterized functions, T(X,Y) and T(X,Y,Z), are designed to estimate the mutual information between features X and Y, and the conditional mutual information between X and Y under feature Z, respectively. These two functions are implemented using a multilayer perceptron, and the conditional mutual information is estimated by maximizing the difference between the expectation of T(X,Y) and the expectation of T(X,Y,Z). An initial feature causal relationship network is then constructed based on the calculated conditional mutual information.

[0104] When calculating direct causal effects, a counterfactual inference framework is used to estimate the magnitude of the causal effect by comparing changes in characteristic states before and after the intervention. For a characteristic pair (X, Y), keeping other characteristics constant, X is increased by one standard deviation, and the change in Y is observed as the direct causal effect. Indirect causal effects are calculated through path analysis, multiplying and summing all direct effects along the indirect path from X to Y. Taking the causal relationship between sleep quality and blood pressure as an example, the direct effect is 0.31, the indirect effect through the intermediate node of metabolic rate is 0.15, and the total causal effect is 0.46.

[0105] For time-varying causal strength, an exponentially weighted moving average method is used, comprehensively considering both historical and current causal effects. The smoothing factor α is set to 0.8, the weight of the current time window is 0.8, the weight of the previous time window is 0.16, and the weight of earlier time windows is 0.04. This processing can capture the dynamic changes in causal relationships over time; for example, the causal strength of exercise intensity on heart rate rises to 0.72 after strenuous exercise and drops to 0.25 at rest.

[0106] The average μ and standard deviation σ of the time-varying causal strength of all causal relationships are calculated as a baseline reference. The health risk score ranges from 0 to 100, calculated comprehensively based on the degree of abnormality of health characteristics. When the risk score is higher than 70, the base threshold is set to μ-0.5σ; when the risk score is between 30 and 70, the base threshold is set to μ; when the risk score is lower than 30, the base threshold is set to μ+0.5σ. Then, the threshold is dynamically adjusted based on the time-varying causal strength of each node: if the time-varying causal strength of a certain characteristic fluctuates by more than 2σ, its threshold is further reduced by 0.1σ to capture potentially significant changes. For a user with a risk score of 85 and large fluctuations in the time-varying causal strength of blood pressure, the system's final threshold is adjusted to μ-0.6σ, retaining 23 causal relationships; while for a healthy user with a risk score of 25 and stable indicators, the threshold is maintained at μ+0.5σ, retaining only 8 strong causal relationships.

[0107] This invention efficiently extracts spatiotemporal correlation information of dynamic health features through multi-head attention map convolution and temporal-aware residual module; multi-scale feature pyramid and neural mutual information maximization technology accurately identify complex causal relationships between features; time-varying causal strength and adaptive threshold mechanism effectively filter weak causal associations, and the constructed feature causal relationship network is more robust and interpretable, providing a solid causal reasoning foundation for health status assessment and overcoming the limitations of traditional correlation analysis.

[0108] In one optional implementation, the conditional mutual information between feature pairs in the hierarchical features is calculated based on maximizing neural mutual information, and the initial feature causal relationship network is constructed based on the conditional mutual information, including:

[0109] The hierarchical features are subjected to time dimension displacement transformation and feature dimension perturbation transformation to generate comparative feature samples of the hierarchical features.

[0110] Calculate the initial estimate of the conditional mutual information between the hierarchical features and the contrastive feature samples;

[0111] A conditional variable attention network is constructed, with the initial estimate as input, and conditional variable weights are generated based on attention calculation. The most relevant conditional variable is selected according to the conditional variable weights.

[0112] The hierarchical features are projected onto multiple orthogonal feature spaces to obtain multi-view feature representations, and multi-view conditional mutual information is calculated based on the multi-view feature representations and the most relevant conditional variables.

[0113] Calculate the temporal fluctuation intensity of the hierarchical features, construct an adaptive kernel function based on the temporal fluctuation intensity, and apply the adaptive kernel function to the multi-view conditional mutual information calculation to obtain the optimized multi-view conditional mutual information;

[0114] Product feature association information is extracted from the health knowledge graph, and the product feature association strength is calculated based on graph reasoning.

[0115] The optimized multi-view conditional mutual information and the product feature association strength are combined using a Bayesian weighted fusion method to construct an initial feature causal relationship network. The weight of each edge in the initial feature causal relationship network is calculated using the Bayesian weighted fusion method.

[0116] For example, for received hierarchical features, the time-dimensional shift transformation refers to translating the original feature sequence along the time axis, such as moving the feature value at time t to time t+δ, where δ can be a time step of 1, 2, or larger. The feature-dimensional perturbation transformation adds random noise to the original feature values, such as adding a random perturbation following a normal distribution N(0, 0.01) to the feature value x. Combining these two transformations can generate diverse comparative samples, enhancing the model's robust understanding of feature relationships.

[0117] A neural network architecture is employed to estimate the conditional mutual information between hierarchical features and contrastive feature samples. This neural network comprises three input branches: the first branch processes feature X, the second branch processes feature Y, and the third branch processes the conditional variable Z. Each branch consists of three fully connected layers with 64 neurons each, using the ReLU activation function. The outputs of the three branches are fused through a connection layer and then passed through two fully connected layers to output the estimated conditional mutual information. During training, the system is optimized using the InfoNCE loss function, with a batch size of 128 and a learning rate of 0.001, and trained for 500 epochs until convergence.

[0118] The conditional variable attention network employs a multi-head attention mechanism with four heads, each with an attention dimension of 16. The network receives all possible conditional variables as input and calculates an attention score for each. These attention scores are normalized using a softmax function to obtain the conditional variable weights. Conditional variables with weights greater than a preset threshold of 0.15 are selected as the most relevant. For example, for user motion state characteristics, the system might identify sleep quality, heart rate variability, and daily activity level as the most relevant conditional variables, with attention weights of 0.35, 0.28, and 0.22, respectively.

[0119] The system projects hierarchical features into multiple orthogonal feature spaces. Projection is achieved using orthogonal matrices, generating three mutually orthogonal projection matrices with dimensions equal to the original feature dimensions. Each projection matrix is ​​generated through a Gram-Schmidt orthogonalization process to ensure the independence of feature representations from different perspectives. The original features are multiplied by each projection matrix to obtain feature representations from three different perspectives. Based on these multi-perspective feature representations and the most relevant conditional variables selected in the previous step, the system calculates multi-perspective conditional mutual information. The purpose of multi-perspective computation is to capture different aspects of the relationships between features and improve the reliability of the estimation.

[0120] The system calculates the temporal volatility intensity of hierarchical features, specifically by calculating the standard deviation of each feature within a continuous time window. For example, for the heart rate feature, the system calculates the standard deviation within a 10-minute window, which might yield a volatility intensity value of 5.8. Based on the calculated temporal volatility intensity, the system constructs an adaptive kernel function. Features with higher volatility intensity correspond to a wider kernel bandwidth, while features with lower volatility intensity correspond to a narrower kernel bandwidth. For example, the kernel bandwidth for the heart rate feature is set to 5.8 × 0.5 = 2.9, while the more stable body temperature feature might have a kernel bandwidth of 1.2. The system applies the adaptive kernel function to the calculation of multi-view conditional mutual information, obtaining optimized multi-view conditional mutual information. The adaptive kernel function can adjust the sensitivity of mutual information estimation according to the inherent volatility properties of the features, improving the estimation accuracy.

[0121] The system extracts product feature association information from a health knowledge graph. It uses a graph attention network to process the knowledge graph, extracting subgraph structures related to the currently analyzed features. Based on graph inference, the system calculates the product feature association strength, with values ​​ranging from 0 to 1; a higher value indicates a stronger association.

[0122] The system combines the optimized multi-view conditional mutual information with the product feature association strength using a Bayesian weighted fusion method. This method considers the uncertainty of both information sources and performs a weighted average. The weights are adaptively determined based on their respective confidence levels; for example, the weight of the multi-view conditional mutual information might be set to 0.65, and the weight of the product feature association strength might be set to 0.35. Through fusion, the system constructs an initial feature causal relationship network. Each edge in the network represents a causal relationship between features, and the edge weight is calculated using the Bayesian weighted fusion method. For example, for the causal relationship strength between exercise volume and heart rate, the fused weight might be 0.65 × 0.82 + 0.35 × 0.75 = 0.796.

[0123] Figure 2 This figure compares the accuracy of three different conditional mutual information estimation methods across different feature dimensions. The horizontal axis represents the feature dimension (0-140), and the vertical axis represents the accuracy (R²) of the conditional mutual information estimation. 2 Values, ranging from 0.5 to 1.1). The method of this invention (circular marker, solid line): As the feature dimension increases, the accuracy steadily improves from 0.90 to 0.985, demonstrating excellent performance and stability, especially maintaining high accuracy in high-dimensional feature spaces (80-140 dimensions). Traditional mutual information estimation (triangle marker, dashed line): The accuracy decreases significantly with the increase of feature dimension, indicating that the method has a serious performance degradation problem in high-dimensional spaces. Basic neural mutual information estimation (square marker, dotted line): The performance is between the former two, gradually decreasing to 0.68 with the increase of dimension. Although it is better than the traditional method, it still has obvious limitations. The significant advantages of the method of this invention compared with the other two methods are: (1) It always maintains the highest accuracy in the entire dimension range; (2) The performance does not decrease but increases with the increase of dimension; (3) In high-dimensional feature spaces (100-140 dimensions), its accuracy is higher than both the traditional method and the basic neural mutual information estimation method. This fully demonstrates the excellent performance and technological innovation of this invention in handling high-dimensional conditional mutual information estimation tasks.

[0124] This invention generates high-quality comparative samples through time-dimensional displacement transformation and feature-dimensional perturbation transformation; conditional variable attention network and multi-view feature representation technology significantly improve the accuracy of conditional mutual information estimation; adaptive kernel function dynamically adjusts according to the intensity of temporal fluctuations, enhancing the robustness of mutual information calculation; and Bayesian weighted fusion is performed by combining product feature association information from knowledge graphs, so that the constructed causal relationship network not only conforms to data characteristics but also satisfies professional knowledge constraints, greatly improving the accuracy of causal discovery.

[0125] In one optional implementation, a state transition probability matrix is ​​constructed based on the feature causal relationship network. Causal intervention analysis is then performed based on the state transition probability matrix and the health assessment rule nodes in the health knowledge graph to generate a health status assessment result, including:

[0126] A multidimensional health state vector is constructed, which includes physiological state components, psychological state components, and behavioral state components. A state transition space is constructed based on the node features in the feature causal relationship network, and a state distribution sequence of the multidimensional health state vector is generated in the state transition space.

[0127] Based on the aforementioned feature causal relationship network, all possible state transition paths are extracted, and each state transition path and its corresponding state distribution sequence are input into a multi-layer neural network to obtain a path importance score.

[0128] Attention weights are constructed based on the path importance scores, and the transition probabilities of different state transition paths are weighted and combined according to the attention weights to obtain a state transition probability matrix.

[0129] Based on the medical knowledge nodes in the health knowledge graph, evaluation rule nodes for different health states are extracted, and the evaluation rule nodes are represented as rule vectors. A rule attention computation network is constructed based on the rule vectors.

[0130] The state transition probability matrix is ​​input into the rule-based attention computation network to obtain a rule-enhanced state representation. Based on the rule-enhanced state representation, the direct intervention effect and the indirect intervention effect are calculated. A hierarchical intervention evaluation structure is constructed based on the direct intervention effect and the indirect intervention effect. The hierarchical intervention evaluation structure includes a state-level intervention unit, a transition-level intervention unit, and a rule-level intervention unit.

[0131] The multidimensional health status vector is input into the hierarchical intervention assessment structure, and sub-assessment results are generated based on each intervention unit. The health status assessment result is obtained by adaptively weighting and fusing the sub-assessment results.

[0132] For example, in constructing a multidimensional health state vector, the system creates a vector representation that includes physiological, psychological, and behavioral components. The physiological components include physiological indicators such as blood pressure, blood sugar levels, and body temperature, represented by normalized actual measurements. The psychological components include psychological indicators such as mood swings, stress levels, and sleep quality, quantified through questionnaire scores and wearable device monitoring data. The behavioral components include behavioral indicators such as exercise frequency, dietary habits, and social activities, quantified based on user logs and smart device records. For a user, their physiological component blood pressure is 135 mmHg, normalized to 0.72; their psychological component sleep quality score is 6.5 out of 10, normalized to 0.65; and their behavioral component average daily exercise time is 35 minutes, normalized to 0.58.

[0133] Based on the node features in the feature causal relationship network, the system constructs a state transition space. Taking the user mentioned above as an example, blood pressure is divided into three states: normal (90-120 mmHg), slightly high (121-139 mmHg), and high blood pressure (≥140 mmHg); sleep quality is divided into three states: good (8-10 points), average (5-7 points), and poor (0-4 points); and exercise frequency is divided into three states: high (≥60 minutes / day), moderate (30-59 minutes / day), and low (<30 minutes / day). Through the combination of these categorized states, a state transition space is constructed, with a total of 27 possible state combinations. The system generates the user's state distribution sequence in this space, tracking the transition process from "slightly high blood pressure, average sleep, moderate exercise" to "normal blood pressure, good sleep, high exercise frequency" over 30 days, and recording the daily state distribution probability.

[0134] The system extracts all possible state transition paths from a feature causal relationship network. One of the user's main transition paths is "increased exercise frequency → decreased blood pressure → improved sleep quality → improved mood". The system constructs a multi-layer neural network to evaluate the importance of the path. The network contains three hidden layers: the first layer has 64 neurons, the second layer has 32 neurons, and the third layer has 16 neurons. ReLU activation and a dropout rate of 0.2 are used to prevent overfitting. The network inputs are the path representation (encoded as a feature transition sequence) and the corresponding state distribution sequence (a 30-day state probability distribution), and the output is a single importance score. For the user's transition path from medium to high exercise frequency, the system calculates a high importance score of 0.85, indicating that this path has a significant impact on the user's health improvement.

[0135] The system normalizes the importance scores of all main paths to generate attention weights. For the five main paths extracted by the user, their importance scores are 0.85, 0.72, 0.65, 0.43, and 0.25, respectively, and their attention weights after normalization using the softmax function are 0.29, 0.25, 0.22, 0.15, and 0.09, respectively. Based on these attention weights, the system assigns weighted probabilities to state transitions on each path. For example, the transition "increased exercise frequency → decreased blood pressure" is assigned a weight of 0.29; the transition "dietary adjustment → decreased blood sugar" is assigned a weight of 0.25. Through this weighted combination, the system constructs a 27×27 state transition probability matrix, where each element P... ij This represents the probability of transitioning from state i to state j.

[0136] The system extracts assessment rule nodes for different health states from a health knowledge graph. The health knowledge graph contains various medical knowledge nodes, such as disease nodes, symptom nodes, and clinical guideline nodes. For this user's high blood pressure, the system first locates the "hypertension" disease node and its associated "prehypertension" medical knowledge node. Then, it extracts assessment rules from the associated clinical guideline nodes, such as "increase monitoring frequency if systolic blood pressure exceeds 130 mmHg for three consecutive days" and "seek medical attention if accompanied by dizziness." The system also extracts supplementary rules from related cardiovascular risk medical knowledge nodes to ensure the comprehensiveness of the assessment. Each rule is encoded into a rule vector, and the BERT model is used to process the rule text, extracting 128-dimensional semantic features. These features are then adjusted to a uniform-dimensional rule vector through a fully connected layer.

[0137] Based on rule vectors, the system constructs a rule-based attention computation network, which includes a rule embedding layer, a multi-head self-attention layer, and a rule-state interaction layer. The rule embedding layer adds positional encoding to each rule; the multi-head self-attention layer uses eight attention heads to compute the associations between rules; and the rule-state interaction layer computes the matching degree between the state and the rule. The system flattens the previously obtained state transition probability matrix into a vector and converts it into a state representation vector through a two-layer fully connected network (256 neurons in the first layer and 128 neurons in the second layer).

[0138] For the user's state, the system calculates the attention weights of all rules, finding that rules related to "high blood pressure" receive a high attention weight of 0.27, while the average attention weight of other rules is only 0.08. The system adds the state representation to the attention-weighted rule representation to obtain a rule-enhanced state representation. Based on the rule-enhanced state representation, the system calculates the direct and indirect intervention effects of various interventions. The direct intervention effect is calculated using a forward prediction network: the system first predicts the expected values ​​of health indicators in the user's current state, then predicts the expected values ​​of the indicators after implementing a specific intervention; the difference between the two is the direct effect. The indirect intervention effect is calculated through path analysis: the system identifies all intermediate paths from the intervention variable to the target variable and calculates the effect transmitted along each path. For the intervention "increase aerobic exercise," the system identifies three main indirect paths: reducing resting heart rate by improving cardiopulmonary function, thereby reducing blood pressure (effect of 5.2 mmHg); improving insulin sensitivity by reducing body fat percentage, thereby reducing blood glucose (effect of 0.5 mmol / L); and improving sleep quality by increasing endorphin levels, thereby reducing stress hormone levels (effect of reducing stress score by 2.1 points). The system assesses the magnitude, duration, and stability of the effect for each path, forming a complete map of intervention effects.

[0139] Based on the calculated direct and indirect intervention effects, the system constructs a hierarchical intervention assessment structure, comprising three intervention units: a state-level intervention unit, a transition-level intervention unit, and a rule-level intervention unit. The state-level intervention unit focuses on assessing direct intervention effects, providing immediately feasible intervention suggestions for the user's current state, such as "reducing sodium intake can directly lower blood pressure by 8.5 mmHg" or "increasing protein intake can increase basal metabolic rate by 12%." This unit quantifies the immediate effects of interventions and is suitable for health problems requiring rapid improvement. The transition-level intervention unit utilizes a state transition probability matrix and indirect intervention effects to assess the long-term impact of different interventions on state transitions. This unit calculates the time sensitivity and cumulative effect of intervention pathways, such as "adjusting dietary structure can reduce the risk of type 2 diabetes by 25% within 6 months." The rule-level intervention unit combines medical rules and intervention effects to provide personalized recommendations that conform to professional guidelines. This unit assesses the applicability, contraindications, and best practices for interventions, such as "meeting the criteria for prehypertension, it is recommended to monitor blood pressure daily, control sodium intake to no more than 5g, and increase potassium intake," and "meeting the early characteristics of metabolic syndrome, it is recommended to adopt the Mediterranean diet pattern combined with intermittent fasting." This unit ensures that intervention recommendations comply with medical standards and the user's specific health condition.

[0140] The user's multidimensional health status vector is input into the hierarchical intervention assessment structure, and each unit generates a sub-assessment result. The status layer unit outputs "Current blood pressure is high (135 mmHg), it is recommended to monitor twice daily"; the transition layer unit outputs "Maintaining the current intervention plan, the probability of blood pressure returning to normal within 8 weeks is 76%"; the rule layer unit outputs "Meets the criteria for prehypertension, it is recommended to increase exercise intensity and duration, and reduce sodium intake". The system adaptively sets the weights of the three units to 0.3, 0.4, and 0.3 based on the user's risk level and adherence history, and the weighted fusion yields the final health status assessment result.

[0141] Figure 3 This graph presents a dynamic analysis of the effects of health interventions over time, illustrating the changes in the intervention effects of three different health status assessment methods. The horizontal axis represents the intervention time (in weeks), from 0 to 20 weeks; the vertical axis represents the improvement rate of health indicators (percentage), from 0% to 100%. The graph compares three methods: the method of this invention (solid line), the traditional statistical correlation method (dashed line), and the random forest method (dotted line). The results show that the method of this invention is significantly superior to the other two methods in both the speed of improvement and the final effect. The method of this invention reaches a critical turning point in week 3 (specially marked in the graph), with the improvement rate of health indicators reaching 50%, while the traditional statistical correlation method requires 16 weeks and the random forest method requires 6 weeks to reach the same level. In terms of long-term effects, the method of this invention also has considerable advantages. The graph marks the "critical period of intervention" (approximately weeks 3-8), during which the differences in the effects of the three methods begin to widen significantly, and the method of this invention shows a clear trend of accelerated improvement during this stage. This result demonstrates that the hierarchical intervention assessment structure of the present invention can promote health status improvement more quickly and effectively, and exhibits better stability in long-term maintenance, which is particularly important for chronic health problems that require long-term management.

[0142] This invention comprehensively represents a user's health status through a multidimensional health state vector, combines causal network analysis with the complex relationships between health factors, and incorporates professional medical rules for assessment. The hierarchical intervention assessment structure achieves a comprehensive evaluation, from direct state adjustment and transfer path optimization to medical rule guidance, significantly improving the accuracy and interpretability of health assessments. The system can provide users with personalized and targeted health intervention suggestions, effectively supporting health management decisions and possessing significant value for chronic disease prevention and health promotion.

[0143] In one optional implementation, a product feature vector is constructed, comprising efficacy and mechanism of action dimensions; a synergistic enhancement coefficient between candidate products is calculated based on the product feature vector; the intervention effect on the health status assessment result is predicted for each candidate product combination; and the optimal product combination is determined based on the intervention effect and the synergistic enhancement coefficient, including:

[0144] Product efficacy node information and active ingredient mechanism information are extracted from the health knowledge graph and weighted and fused to obtain the product feature vector.

[0145] The product feature vector and health status assessment results are encoded into a composite state vector; a deep inverse reward network and a two-layer attention mechanism are constructed based on the composite state vector, and a candidate product combination is generated and evaluated using an attention-based policy network, and the synergistic enhancement coefficient is obtained by fusion.

[0146] Based on the product feature vector, predict the probability of the health status impact of the product portfolio, construct a health status transition sequence, substitute the probability of the health status impact into the health status transition sequence, and obtain the predicted health status values ​​at multiple time points; calculate the state fluctuation amplitude at adjacent time points based on the predicted health status values, and calculate the cumulative intervention effect and intervention stability of the product portfolio based on the state fluctuation amplitude.

[0147] The optimal product combination is determined based on the synergistic enhancement coefficient, cumulative intervention effect, and intervention stability.

[0148] For example, product efficacy node information is extracted from the health knowledge graph, including its regulatory effects on health indicators such as liver function, blood lipids, and blood sugar. For instance, for product A, its efficacy nodes include "lowering blood lipids" and "regulating immunity." Simultaneously, its active ingredient mechanism information is extracted, such as "inhibiting cholesterol synthesis" and "promoting fat metabolism." The system assigns weights of 0.6 and 0.4 to the efficacy and mechanism information, respectively, and obtains a unified product feature vector through weighted combination. Taking product A as an example, its feature vector contains a total of 24 features across the efficacy and mechanism dimensions, with each feature ranging from 0 to 1, representing the strength of the corresponding efficacy or mechanism.

[0149] The system transforms the user's health status assessment results into a target health status vector, which includes target improvement values ​​for health indicators such as blood glucose, blood pressure, and blood lipids. For example, if the user's current blood glucose level is 7.2 mmol / L and the target value is 5.6 mmol / L, then the target improvement value for blood glucose is -1.6 mmol / L. The system concatenates the product feature vector and the health status assessment results to form a composite state vector. A deep inverse reward network is constructed based on this composite state vector, using the ReLU activation function. A two-layer attention mechanism is employed. Candidate product combinations are generated through a policy network, with the composite state vector as input and the product combination probability as output.

[0150] Based on product feature vectors, the system constructs a health status impact prediction model. This model adopts a gradient boosting tree structure, containing 100 decision trees with a maximum depth of 5 and a learning rate of 0.05. The model input is the set of feature vectors of the product combination, and the output is the probability of the combination's impact on each health indicator. The system constructs a health status transition sequence, including the initial state (T0), short-term state (T1, one week later), medium-term state (T2, one month later), and long-term state (T3, three months later). Substituting the health status impact probability into the health status transition sequence, the predicted health status values ​​at time points T1, T2, and T3 are obtained. Taking blood glucose as an example, the user's initial value is 7.2 mmol / L. After using product combination AB, the predicted values ​​are 6.8 mmol / L at T1, 6.3 mmol / L at T2, and 5.8 mmol / L at T3. The system calculates the state fluctuation amplitude between adjacent time points, such as |T1-T0|=0.4, |T2-T1|=0.5, and |T3-T2|=0.5. The cumulative intervention effect (the difference between the initial and final states, which is 1.4 mmol / L in this example) and intervention stability (the variance of the fluctuation range, which is 0.003 in this example) are calculated based on the fluctuation range.

[0151] The system determines the optimal product combination based on the synergistic enhancement coefficient, cumulative intervention effect, and intervention stability. The system constructs a comprehensive scoring function, assigning weights of 0.3, 0.5, and 0.2 to the synergistic enhancement coefficient, cumulative intervention effect, and intervention stability, respectively. For example, product combination AB has a synergistic enhancement coefficient of 1.4, a cumulative intervention effect of 1.4 mmol / L (corresponding to a standardized value of 0.87), and an intervention stability of 0.003 (corresponding to a standardized value of 0.92), then its comprehensive score is 0.3 × 1.4 + 0.5 × 0.87 + 0.2 × 0.92 = 1.055. The system scores all candidate product combinations and selects the combination with the highest score as the optimal recommendation.

[0152] In practical applications, the system also considers the cost factor of product combinations. By setting a budget threshold, it filters out product combinations that match the user's budget. Simultaneously, the system further adjusts the product combinations based on the user's personal preferences (such as taste preferences and preferred methods of consumption), improving the personalization of the recommendations and user acceptance.

[0153] The product feature vector of this invention combines efficacy and mechanism of action dimensions to comprehensively characterize product features; the synergistic enhancement coefficient accurately quantifies the interaction between products; the probability of health status impact and the transition sequence analysis technology scientifically predict the long-term intervention effect of the product portfolio; the cumulative intervention effect and intervention stability index assess the sustained impact of the intervention. The multi-objective optimization mechanism that integrates synergistic enhancement, intervention effect and stability significantly improves the accuracy and personalization of product portfolio recommendations.

[0154] In one optional implementation, the product feature vector and health status assessment result are encoded into a composite state vector; a deep inverse reward network and a two-layer attention mechanism are constructed based on the composite state vector, and a candidate product combination is generated and evaluated using an attention-based policy network. The resulting synergistic enhancement coefficients include:

[0155] The product feature vector and health status assessment results are encoded into a composite state vector, which includes a product efficacy intensity component, a mechanism component, and a health impact component; valid expert recommendation records are selected based on the degree of improvement of the health impact component in the target composite state vector.

[0156] A deep inverse reward network is constructed. Based on the effective expert recommendation records, the implicit reward function of expert decision-making is learned through a multi-layer neural network. The loss function of the deep inverse reward network includes a state prediction error term, a reward consistency constraint term, and a temporal smoothing constraint term.

[0157] A two-layer attention mechanism is constructed based on the composite state vector: the first layer is a product-level attention module, which calculates the product combination importance weight based on the similarity matrix of product features in the composite state vector; the second layer is a feature-level attention module, which calculates the feature association weight based on the interaction matrix of the efficacy intensity component and the mechanism component; the product combination importance weight and the feature association weight are fused to obtain the combination attention weight;

[0158] Construct an attention-based policy network, input the composite state vector, and generate a probability distribution of candidate product combinations based on the combined attention weights; use Thompson sampling to sample from the probability distribution to generate combination schemes, and evaluate the combination schemes based on the implicit reward function;

[0159] The reward value of the deep inverse reward network, the combined attention weights, and the evaluation value of the combined scheme are weighted and fused to obtain the collaborative enhancement coefficient.

[0160] For example, the result Figure 4The flowchart illustrating the product synergistic enhancement coefficient generation based on a deep inverse reward network and a two-layer attention mechanism is as follows: The product feature vector and health status assessment results are encoded to form a composite state vector containing product efficacy intensity, mechanism of action, and health impact components. In practical applications, the product efficacy intensity component uses values ​​from 0 to 1 to represent the strength of each efficacy indicator. For example, for a certain nutritional supplement, its antioxidant efficacy intensity is 0.8, and its immune regulation efficacy intensity is 0.6. The mechanism of action component uses a multi-dimensional vector to represent the degree of the product's effect on different physiological mechanisms, such as a metabolic regulation index of 0.7 and a cell repair index of 0.5. The health impact component is quantified as the degree of improvement in target health indicators, such as a blood glucose level improvement index of 0.4 and a fatigue recovery index of 0.6.

[0161] For screening valid expert recommendation records, the system sets a health impact improvement threshold of 0.3. This means that a recommendation is considered valid if the recorded health indicators show an improvement of more than 30%. For example, if a user's blood pressure improved from an abnormal value of 140 / 90 mmHg to 130 / 85 mmHg after using a product combination recommended by an expert, an improvement of approximately 7% is below the threshold and therefore not included in the valid record. However, another recommended combination reduced blood glucose levels from 9.0 mmol / L to 6.5 mmol / L, an improvement of approximately 28%, close to the threshold, and was included in the valid record database after comprehensive evaluation of multiple indicators.

[0162] In constructing the deep inverse reward network, the system employs a five-layer neural network structure to learn the implicit reward function of expert decision-making. The implicit reward function refers to the decision-making criteria and value judgment system that experts inherently follow but do not explicitly express when recommending product combinations. Machine learning techniques are used to extract these criteria from expert behavior and quantify them into numerical reward values. For example, experts may prefer to recommend combinations with complementary antioxidant and immune-regulating effects rather than combinations with overlapping antioxidant effects, or prioritize "multi-target synergistic effects" rather than "high-dose effects of a single target." These preference patterns constitute the core content of the implicit reward function. The neural network constructed by the system includes one input layer, three hidden layers, and one output layer. The input layer receives a composite state vector with a dimension equal to the sum of the number of product features and the number of health state indicators, typically 120 dimensions in typical applications. The three hidden layers have 128, 256, and 128 nodes respectively, and the ReLU activation function is used to enhance non-linear expressive power. The output layer is a single node, and the Sigmoid activation function is used to limit the output range to between 0 and 1, representing a normalized reward value.

[0163] The system collects 5,000 valid expert recommendation records from a health management database as training data. Each record includes product combination information, the user's initial health status, the degree of improvement in health status after use, and an expert rating. For example, the record "ER-1042" includes product combination {A,C,E}, initial status "high blood pressure + poor sleep quality", health improvement "blood pressure decreased by 12% + sleep quality improved by 35%", and an expert rating of 0.82. The system converts these records into training samples, with the input being a composite state vector and the label being the expert rating. It then uses a backpropagation algorithm to learn the decision rules implicit in the expert ratings.

[0164] The network training utilizes collected effective expert recommendation records. The loss function comprises three components: a state prediction error term calculated using mean squared error with a weight of 0.5; a reward consistency constraint term ensuring similar states receive similar rewards with a weight of 0.3; and a temporal smoothing constraint term ensuring the continuity of state changes with a weight of 0.2. During training, the initial learning rate is set to 0.001, the Adam optimizer is used, the batch size is 64, and the number of iterations is 200.

[0165] For example, for a new combination {A,C,G}, the system constructs its composite state vector and inputs it into the network, obtaining a reward value of 0.82, indicating that the combination meets the expert decision-making criteria and may have a good synergistic effect. The system can also analyze the impact of different factors on the reward value by perturbing the input vector, providing a more transparent explanation of the decision, such as "replacing product C with product H will reduce the reward value by 0.15, mainly because the key complementary mechanism is lost."

[0166] The two-layer attention mechanism is constructed based on composite state vectors. During the product-level attention module calculation, the feature vector of each product is first extracted. For each pair of products, cosine similarity is calculated to form a similarity matrix. For example, for three products A, B, and C, the resulting similarity matrix might be: A vs. A = 1.0, A vs. B = 0.75, A vs. C = 0.4, B vs. B = 1.0, B vs. C = 0.65, and C vs. C = 1.0. These similarity values ​​are then converted into importance weights using the Softmax function, resulting in a weight of 0.45 for product A, 0.35 for product B, and 0.2 for product C. The feature-level attention module then constructs an interaction matrix based on efficacy intensity and mechanism components to reflect the degree of association between different features. Taking a certain combination as an example, the interaction value between antioxidant efficacy and cell repair mechanism is 0.6, and the interaction value between immune regulation efficacy and metabolic regulation mechanism is 0.8. After applying the Softmax function, feature association weights are obtained, with the efficacy-mechanism pair "immune regulation-metabolic regulation" receiving the highest weight of 0.25. Finally, the product portfolio importance weight and feature association weight are combined using a weighted average method to obtain the portfolio attention weight.

[0167] The attention-based policy network receives a composite state vector as input and processes it through a three-layer neural network: the number of nodes in the input layer equals the dimension of the composite state vector, the first hidden layer contains 64 nodes, the second hidden layer contains 32 nodes, and the number of nodes in the output layer equals the total number of candidate products. Combined attention weights are applied between each layer to adjust the connection strength, making the network focus more on important products and key feature combinations. The output layer generates a probability distribution for each product being selected using a Softmax function. In a healthy state, the system may generate the following probability distribution: product A has a probability of 0.4, product B 0.3, product C 0.2, and product D 0.1. Using Thompson sampling to sample five times from this distribution, possible combinations are {A,B}, {A,C}, {A,B,C}, {A,D}, and {B,C}. These combinations are evaluated using the aforementioned deep inverse reward network, yielding reward values ​​of 0.85, 0.7, 0.95, 0.6, and 0.65, respectively.

[0168] The synergy enhancement coefficient is calculated by comprehensively considering the reward value of the deep inverse reward network, the combined attention weights, and the evaluation value of the combined scheme. For the highest-scoring combination scheme {A,B,C} in the example above, its deep reward network reward value is 0.95, its combined attention weight is 0.85, and its combined evaluation value is 0.9. A weighted average is used (weights are 0.4, 0.3, and 0.3 respectively) to obtain a final synergy enhancement coefficient of 0.91, indicating that this combination scheme has a very high synergistic effect and is suitable for recommendation to users. The system sorts the candidate combinations according to the synergy enhancement coefficient and selects the scheme with the highest coefficient as the final recommendation result.

[0169] The deep inverse reward network of this invention learns the implicit reward function from effective expert recommendation records, effectively simulating the expert decision-making process; a two-layer attention mechanism captures product-level and feature-level association patterns respectively; an attention-based policy network and Thompson sampling technique achieve a balance between product portfolio generation and evaluation; and a weighted fusion mechanism of reward value, attention weight, and evaluation value ensures the comprehensiveness of the synergy enhancement coefficient calculation. This method effectively solves the industry problem of the difficulty in quantifying product synergy effects and improves the synergistic effect of product portfolios.

[0170] A second aspect of this invention provides a knowledge graph-based health status assessment and product matching system, comprising:

[0171] The first unit is used to acquire user health data and establish a health knowledge graph that includes medical knowledge nodes, health assessment rule nodes, and product feature nodes.

[0172] The second unit is used to map the user health data into user feature vectors; it uses a multi-scale temporal convolutional network to extract the temporal variation features of the user feature vectors in daily, weekly and monthly cycles, calculates the fluctuation standard deviation of the temporal variation features, generates a dynamic association threshold, filters feature associations based on the dynamic association threshold, and constructs a dynamic health feature association network.

[0173] The third unit is used to input the dynamic health feature association network into a graph convolutional neural network to obtain feature representations, process the feature representations through a structured causal discovery algorithm, establish a feature causal relationship network, construct a state transition probability matrix based on the feature causal relationship network, and perform causal intervention analysis according to the state transition probability matrix and the health assessment rule nodes in the health knowledge graph to generate health status assessment results.

[0174] The fourth unit is used to construct a product feature vector that includes efficacy and mechanism of action dimensions; calculate the synergistic enhancement coefficient between candidate products based on the product feature vector; predict the intervention effect on the health status assessment result for each candidate product combination; and determine the optimal product combination based on the intervention effect and the synergistic enhancement coefficient.

[0175] A third aspect of the present invention provides an electronic device, comprising:

[0176] processor;

[0177] Memory used to store processor-executable instructions;

[0178] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0179] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0180] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0181] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions 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 health status assessment and product matching method based on knowledge graphs, characterized in that, include: Acquire user health data; establish a health knowledge graph that includes medical knowledge nodes, health assessment rule nodes, and product feature nodes; The user health data is mapped to a user feature vector; a multi-scale temporal convolutional network is used to extract the temporal variation features of the user feature vector in daily, weekly and monthly cycles, the fluctuation standard deviation of the temporal variation features is calculated, a dynamic association threshold is generated, feature associations are filtered based on the dynamic association threshold, and a dynamic health feature association network is constructed. The dynamic health feature association network is input into a graph convolutional neural network to obtain feature representations. The feature representations are then processed by a structured causal discovery algorithm to establish a feature causal relationship network. A state transition probability matrix is ​​constructed based on the feature causal relationship network. Causal intervention analysis is then performed based on the state transition probability matrix and the health assessment rule nodes in the health knowledge graph to generate a health status assessment result. Construct a product feature vector that includes efficacy and mechanism of action dimensions; calculate the synergistic enhancement coefficient between candidate products based on the product feature vector; predict the intervention effect on the health status assessment result for each candidate product combination; and determine the optimal product combination based on the intervention effect and the synergistic enhancement coefficient.

2. The method according to claim 1, characterized in that, A multi-scale temporal convolutional network is used to extract the temporal variation features of the user feature vector in daily, weekly, and monthly cycles. The standard deviation of the fluctuation of the temporal variation features is calculated to generate a dynamic association threshold. Based on the dynamic association threshold, feature associations are filtered, and a dynamic health feature association network is constructed, including: Within a preset time window, the first-order difference value of the user feature vector is calculated to generate a fluctuation feature matrix, and the fluctuation density index is calculated based on the fluctuation feature matrix. A multi-scale temporal convolutional network is constructed, which includes daily-scale convolutional branches, weekly-scale convolutional branches, and monthly-scale convolutional branches. The dynamic expansion rate of each branch is calculated using the hyperbolic tangent function based on the fluctuation density index, and the dynamic expansion rate is smoothed by performing an exponential moving average on the dynamic expansion rate. The user feature vector is input into the multi-scale temporal convolutional network, and convolution is performed using the smooth dilation rate to extract temporal variation features. The temporal consistency index and prediction error index of the temporal variation features are calculated. The temporal consistency index, prediction error index and signal-to-noise ratio score are combined to form a reliability vector. The reliability vector is input into a nonlinear mapping function to obtain feature weights. The temporal variation features are weighted and fused according to the feature weights to obtain multi-scale fusion features. At the same time, temporal smoothing constraints and entropy regularization constraints are introduced. The feature fluctuation standard deviation is calculated based on the multi-scale fusion features. A dynamic correlation threshold is calculated based on the feature fluctuation standard deviation and the health risk score. The dynamic correlation threshold increases with the increase of the feature fluctuation standard deviation and decreases with the increase of the health risk score. Calculate the association strength of feature pairs in the multi-scale fusion features, retain the association strengths that are greater than the dynamic association threshold, and construct a dynamic health feature association network.

3. The method according to claim 1, characterized in that, The dynamic health feature association network is input into a graph convolutional neural network to obtain feature representations. These feature representations are then processed using a structured causal discovery algorithm to establish a feature causal relationship network, including: A multi-head attention map convolutional layer is constructed. Attention coefficients between nodes in the dynamic health feature association network are calculated based on attention vectors. Feature transformation and multi-head feature aggregation are performed on node features according to these attention coefficients to obtain an initial feature representation. This initial feature representation is then input into a temporal-aware residual module, where features at different levels are adaptively fused through gated skip connections to obtain temporal-enhanced features. A multi-scale feature pyramid is constructed, and convolutional processing and feature pooling are performed on the temporal-enhanced features with different receptive fields to generate a multi-scale feature sequence. This multi-scale feature sequence is then upsampled and fused based on learnable weights to obtain hierarchical features. Based on the maximization of neural mutual information, the conditional mutual information between feature pairs in the hierarchical features is calculated, and an initial feature causal relationship network is constructed based on the conditional mutual information. Calculate the direct and indirect causal effects between feature pairs in the initial feature causal relationship network, and calculate the time-varying causal strength based on the direct and indirect causal effects; An adaptive threshold is generated based on the time-varying causal strength and health risk score, and causal associations greater than the adaptive threshold are retained to construct a feature causal relationship network.

4. The method according to claim 3, characterized in that, Calculating the conditional mutual information between feature pairs in the hierarchical features based on maximizing neural mutual information, and constructing an initial feature causal relationship network based on the conditional mutual information includes: The hierarchical features are subjected to time dimension displacement transformation and feature dimension perturbation transformation to generate comparative feature samples of the hierarchical features. Calculate the initial estimate of the conditional mutual information between the hierarchical features and the contrastive feature samples; A conditional variable attention network is constructed, with the initial estimate as input, and conditional variable weights are generated based on attention calculation. The most relevant conditional variable is selected according to the conditional variable weights. The hierarchical features are projected onto multiple orthogonal feature spaces to obtain multi-view feature representations, and multi-view conditional mutual information is calculated based on the multi-view feature representations and the most relevant conditional variables. Calculate the temporal fluctuation intensity of the hierarchical features, construct an adaptive kernel function based on the temporal fluctuation intensity, and apply the adaptive kernel function to the multi-view conditional mutual information calculation to obtain the optimized multi-view conditional mutual information; Product feature association information is extracted from the health knowledge graph, and the product feature association strength is calculated based on graph reasoning. The optimized multi-view conditional mutual information and the product feature association strength are combined using a Bayesian weighted fusion method to construct an initial feature causal relationship network. The weight of each edge in the initial feature causal relationship network is calculated using the Bayesian weighted fusion method.

5. The method according to claim 3, characterized in that, Based on the aforementioned feature causal relationship network, a state transition probability matrix is ​​constructed. Causal intervention analysis is then performed based on the state transition probability matrix and the health assessment rule nodes in the health knowledge graph to generate health status assessment results, including: A multidimensional health state vector is constructed, which includes physiological state components, psychological state components, and behavioral state components. A state transition space is constructed based on the node features in the feature causal relationship network, and a state distribution sequence of the multidimensional health state vector is generated in the state transition space. Based on the aforementioned feature causal relationship network, all possible state transition paths are extracted, and each state transition path and its corresponding state distribution sequence are input into a multi-layer neural network to obtain a path importance score. Attention weights are constructed based on the path importance scores, and the transition probabilities of different state transition paths are weighted and combined according to the attention weights to obtain a state transition probability matrix. Based on the medical knowledge nodes in the health knowledge graph, evaluation rule nodes for different health states are extracted, and the evaluation rule nodes are represented as rule vectors. A rule attention computation network is constructed based on the rule vectors. The state transition probability matrix is ​​input into the rule-based attention computation network to obtain a rule-enhanced state representation. Based on the rule-enhanced state representation, the direct intervention effect and the indirect intervention effect are calculated. A hierarchical intervention evaluation structure is constructed based on the direct intervention effect and the indirect intervention effect. The hierarchical intervention evaluation structure includes a state-level intervention unit, a transition-level intervention unit, and a rule-level intervention unit. The multidimensional health status vector is input into the hierarchical intervention assessment structure, and sub-assessment results are generated based on each intervention unit. The health status assessment result is obtained by adaptively weighting and fusing the sub-assessment results.

6. The method according to claim 1, characterized in that, Construct product feature vectors that include efficacy and mechanism of action dimensions; calculate synergistic enhancement coefficients among candidate products based on the product feature vectors; predict the intervention effect on the health status assessment results for each candidate product combination; Determining the optimal product mix based on the intervention effect and the synergistic enhancement coefficient includes: Product efficacy node information and active ingredient mechanism information are extracted from the health knowledge graph and weighted and fused to obtain the product feature vector. The product feature vector and health status assessment results are encoded into a composite state vector; a deep inverse reward network and a two-layer attention mechanism are constructed based on the composite state vector, and a candidate product combination is generated and evaluated using an attention-based policy network, and the synergistic enhancement coefficient is obtained by fusion. Based on the product feature vector, predict the probability of the health status impact of the product portfolio, construct a health status transition sequence, substitute the probability of the health status impact into the health status transition sequence, and obtain the predicted health status values ​​at multiple time points; calculate the state fluctuation amplitude at adjacent time points based on the predicted health status values, and calculate the cumulative intervention effect and intervention stability of the product portfolio based on the state fluctuation amplitude. The optimal product combination is determined based on the synergistic enhancement coefficient, cumulative intervention effect, and intervention stability.

7. The method according to claim 6, characterized in that, The product feature vector and health status assessment results are encoded into a composite state vector. A deep inverse reward network and a two-layer attention mechanism are constructed based on this composite state vector. An attention-based policy network is used to generate and evaluate candidate product combinations, and the resulting synergistic enhancement coefficients are obtained by fusing them. The product feature vector and health status assessment results are encoded into a composite state vector, which includes a product efficacy intensity component, a mechanism component, and a health impact component; valid expert recommendation records are selected based on the degree of improvement of the health impact component in the target composite state vector. A deep inverse reward network is constructed. Based on the effective expert recommendation records, the implicit reward function of expert decision-making is learned through a multi-layer neural network. The loss function of the deep inverse reward network includes a state prediction error term, a reward consistency constraint term, and a temporal smoothing constraint term. A two-layer attention mechanism is constructed based on the composite state vector: the first layer is a product-level attention module, which calculates the product combination importance weight based on the similarity matrix of product features in the composite state vector; the second layer is a feature-level attention module, which calculates the feature association weight based on the interaction matrix of the efficacy intensity component and the mechanism component; the product combination importance weight and the feature association weight are fused to obtain the combination attention weight; Construct an attention-based policy network, input the composite state vector, and generate a probability distribution of candidate product combinations based on the combined attention weights; use Thompson sampling to sample from the probability distribution to generate combination schemes, and evaluate the combination schemes based on the implicit reward function; The reward value of the deep inverse reward network, the combined attention weights, and the evaluation value of the combined scheme are weighted and fused to obtain the collaborative enhancement coefficient.

8. A health status assessment and product matching system based on knowledge graphs, used to implement the method of any one of claims 1-7, characterized in that, include: The first unit is used to acquire user health data and establish a health knowledge graph that includes medical knowledge nodes, health assessment rule nodes, and product feature nodes. The second unit is used to map the user health data into user feature vectors; it uses a multi-scale temporal convolutional network to extract the temporal variation features of the user feature vectors in daily, weekly and monthly cycles, calculates the fluctuation standard deviation of the temporal variation features, generates a dynamic association threshold, filters feature associations based on the dynamic association threshold, and constructs a dynamic health feature association network. The third unit is used to input the dynamic health feature association network into a graph convolutional neural network to obtain feature representations, process the feature representations through a structured causal discovery algorithm, establish a feature causal relationship network, construct a state transition probability matrix based on the feature causal relationship network, and perform causal intervention analysis according to the state transition probability matrix and the health assessment rule nodes in the health knowledge graph to generate health status assessment results. The fourth unit is used to construct a product feature vector that includes efficacy and mechanism of action dimensions; calculate the synergistic enhancement coefficient between candidate products based on the product feature vector; predict the intervention effect on the health status assessment result for each candidate product combination; and determine the optimal product combination based on the intervention effect and the synergistic enhancement coefficient.

9. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to 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 the processor, they implement the method described in any one of claims 1 to 7.

Images