Disease portrait generation method and device, electronic equipment and storage medium

Abstract

The invention provides a disease portrait generation method and device, electronic equipment and a storage medium, and belongs to the technical field of artificial intelligence, and the method comprises the steps: constructing a feature association map according to multi-source data; inputting the atlas into a hierarchical graph attention network, learning a direct association and combination relationship of feature nodes by using a hierarchical structure, and obtaining an attention weight; candidate key feature nodes are screened accordingly; performing causal effect estimation on the candidate key feature nodes to obtain a causal effect value representing real causal intensity; and finally, based on the attention weight and the causal effect value, determining an association level and generating a disease portrait. According to the method, the depth feature mining capability of the hierarchical graph attention network and the counterfeit elimination and true storage capability of causal inference are fused, so that automatic discovery of the high-order combination relationship of the disease features and quantitative verification of the true causal are realized, and a dynamic disease portrait with depth correlation information and high-credibility causal interpretation can be generated.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method, apparatus, electronic device, and storage medium for generating disease profiles. Background Technology

[0002] With the rapid development of medical informatization, multi-source heterogeneous medical data, such as electronic medical records, genomics, and wearable devices, are experiencing explosive growth. Building comprehensive and accurate disease profiles based on this massive amount of data can help doctors gain a deeper understanding of the mechanisms of disease development, identify high-risk groups, and formulate personalized intervention plans, which has significant clinical value for improving the level of precision medicine.

[0003] Currently, existing technologies for constructing disease profiles typically employ traditional statistical methods (such as logistic regression and proportional hazards models) or conventional machine learning and deep learning models (such as random forests and convolutional neural networks). These methods usually involve cleaning and standardizing the collected patient clinical data, and then training a model to fit the mapping relationship between features and disease outcomes. Finally, based on the feature importance score or statistical significance index output by the model, the top-ranked features are selected as key influencing factors of the disease, thereby generating a static profile describing the distribution of disease features.

[0004] The existing technical solutions described above still have shortcomings in practical applications. First, when constructing disease profiles, existing methods struggle to effectively capture and characterize the complex high-order interactions between features from massive multi-source data. This often results in the generated profiles missing key pathological information at the feature combination level, leading to insufficient information mining depth. Second, disease profiles generated by existing methods typically only reflect the statistical correlation between features and diseases, failing to effectively identify whether this correlation is a genuine causal relationship or a spurious correlation caused by confounding factors. This makes the final profiles prone to including non-causal interference features, lacking a quantitative assessment of the true causal strength between features and diseases, thereby reducing the accuracy and reference value of disease profiles in clinical decision support. Summary of the Invention

[0005] This invention provides a method, apparatus, electronic device, and storage medium for generating disease profiles, which addresses the technical problem that existing technologies struggle to automatically uncover high-order combination relationships between features when processing complex, multi-source medical data, and are unable to effectively distinguish between statistical correlation and true causality. This results in disease profiles lacking deep correlation information and easily containing spurious correlation features, making it difficult to accurately guide clinical decision-making.

[0006] This invention provides a method for generating disease profiles, comprising: A feature association map is constructed based on multi-source data related to the target disease, and the feature association map includes feature nodes that characterize each feature in the multi-source data; The feature association graph is input into a preset hierarchical graph attention network to obtain the attention weight of each feature node determined by the hierarchical graph attention network. At least one candidate key feature node is selected from all the feature nodes based on the attention weight; Causal effect estimation is performed on the candidate key feature nodes to obtain causal effect values ​​that characterize the causal strength between the candidate key feature nodes and the target disease; Based on the attention weight and the causal effect value, the association level of the candidate key feature nodes is determined, so as to generate a disease profile according to the association level.

[0007] According to a disease profile generation method provided by the present invention, the hierarchical graph attention network includes a first network layer and a second network layer; The step of obtaining the attention weights of each feature node determined by the hierarchical graph attention network includes: The first network layer aggregates information about neighboring nodes determined by the edges of the feature association graph to generate a first-order feature representation that characterizes the direct association information between feature nodes. The first-order feature representations are aggregated using the second network layer to generate second-order feature representations that characterize the combination relationships between feature nodes; The attention coefficients generated during the generation of the first-order feature representation and / or the second-order feature representation are extracted to obtain the attention weight of each feature node.

[0008] According to a disease profile generation method provided by the present invention, the step of estimating the causal effect of the candidate key feature nodes to obtain a causal effect value characterizing the causal strength between the candidate key feature nodes and the target disease includes: Target sample data for causal verification is extracted from the multi-source data. The target sample data includes the feature values ​​of multiple sample objects on the candidate key feature nodes and the true label values ​​on the target disease. Obtain pre-trained propensity score model and outcome regression model; The target sample data is input into the propensity score model, and the propensity score of each sample object on the candidate key feature node is output. The target sample data is input into the result regression model, and the predicted probability of each sample object having the target disease is output. The average processing effect of the candidate key feature nodes is calculated based on the propensity score, the predicted probability of disease, the feature value, and the true label value. The average treatment effect is determined as the causal effect value between the candidate key feature node and the target disease.

[0009] According to a disease profile generation method provided by the present invention, the step of calculating the average processing effect of the candidate key feature nodes based on the propensity score, the predicted probability of disease, the feature value, and the true label value includes: Sample weights are constructed using the propensity score and the feature values; Calculate the prediction residual between the true label value and the predicted disease probability value; The predicted residuals are weighted based on the sample weights, and the weighted residuals are fused with the predicted disease probability to obtain the causal effect estimate. The average treatment effect is obtained by averaging the estimated causal effects of all sample objects.

[0010] According to a disease profile generation method provided by the present invention, the step of aggregating information of neighboring nodes determined by edges based on the feature association graph using the first network layer to generate a first-order feature representation characterizing the direct association information between feature nodes includes: For any target feature node in the feature association graph, obtain the initial feature vectors of the target feature node and its neighboring nodes; The initial eigenvector is linearly transformed using a linear transformation matrix to obtain the transformed eigenvector; Based on the transformed feature vector, the correlation coefficient between the target feature node and each neighbor node is calculated; The correlation coefficients are normalized using an activation function to obtain normalized attention coefficients. The transformed feature vectors of the neighboring nodes are weighted and aggregated using the normalized attention coefficients to obtain the first-order feature representation of the target feature node.

[0011] According to a disease profile generation method provided by the present invention, the step of aggregating the first-order feature representations using a second network layer to generate second-order feature representations characterizing the combination relationships between feature nodes includes: For the connected feature node pairs, perform element-wise product operation or concatenation operation on the first-order feature representations of the feature node pairs to generate feature interaction vectors; The importance score of the feature interaction vector in representing the feature combination effect is calculated using a multi-head attention mechanism; The feature interaction vectors are aggregated based on the importance scores to generate the second-order feature representation that characterizes the meta-features formed by the combined interactions between the feature nodes.

[0012] According to a disease profile generation method provided by the present invention, the step of extracting attention coefficients generated during the generation of the first-order feature representation and / or the second-order feature representation to obtain the attention weight of each feature node includes: Obtain the attention coefficient matrix output by the multi-head attention mechanism in the first network layer and / or the second network layer in the hierarchical graph attention network; For any feature node, sum or average all attention coefficients related to that feature node in the attention coefficient matrix to obtain a comprehensive importance score; The comprehensive importance score is determined as the attention weight of any feature node.

[0013] According to a disease profile generation method provided by the present invention, the step of selecting at least one candidate key feature node from all the feature nodes based on the attention weight includes: Sort all the feature nodes from high to low according to the attention weight; The top K feature nodes in the sorting results are selected as the candidate key feature nodes, where K is a preset positive integer; Alternatively, feature nodes whose attention weight is greater than a preset weight threshold can be selected as candidate key feature nodes.

[0014] According to a disease profile generation method provided by the present invention, determining the association level of the candidate key feature nodes based on the attention weight and the causal effect value includes: If the attention weight is greater than a first preset threshold and the causal effect value is greater than a second preset threshold, the association level of the candidate key feature node is determined to be strongly correlated. If the attention weight is greater than the first preset threshold but the causal effect value is less than or equal to the second preset threshold, the association level of the candidate key feature node is determined to be pseudo-correlation or weak correlation. If the attention weight is less than or equal to the first preset threshold but the causal effect value is greater than the second preset threshold, the association level of the candidate key feature node is determined to be potentially relevant.

[0015] According to a disease profile generation method provided by the present invention, the step of constructing a feature association map based on multi-source data related to the target disease includes: The statistical correlation between any two features in the multi-source data is calculated using statistical methods. If the statistical correlation meets the preset conditions, an edge is established between the feature nodes corresponding to the two features, and the value of the statistical correlation is used as the initial weight of the edge.

[0016] A method for generating a disease profile according to the present invention further includes: In response to the acquisition of new multi-source data, the feature association map is updated using the new multi-source data; Based on the updated feature association graph, the hierarchical graph attention network, the propensity scoring model, and the result regression model are incrementally trained. The attention weights are re-determined using the trained hierarchical graph attention network, and the causal effect is re-estimated using the trained propensity score model and outcome regression model to generate the updated disease profile based on the newly added multi-source data.

[0017] According to a method for generating a disease profile provided by the present invention, the step of generating a disease profile based on the association level includes: Generate a visual knowledge graph or multi-dimensional radar chart containing the candidate key feature nodes; In the visualized knowledge graph, the size or color of the nodes is used to represent the magnitude of the attention weights. The strength of the causal effect value is characterized by the thickness or type of the connection between nodes; Alternatively, the disease profile may explicitly output a textual description of the combination relationship between the feature nodes corresponding to the second-order feature representation.

[0018] The present invention also provides a disease profile generation device, comprising the following modules: The graph construction module is used to construct a feature association graph based on multi-source data related to the target disease. The feature association graph includes feature nodes that characterize each feature in the multi-source data. The weight acquisition module is used to input the feature association map into a preset hierarchical graph attention network and acquire the attention weight of each feature node determined by the hierarchical graph attention network. A feature filtering module is used to filter at least one candidate key feature node from all the feature nodes according to the attention weight; The causal estimation module is used to estimate the causal effect of the candidate key feature nodes to obtain a causal effect value that characterizes the causal strength between the candidate key feature nodes and the target disease. The profile generation module is used to determine the association level of the candidate key feature nodes based on the attention weight and the causal effect value, so as to generate a disease profile according to the association level.

[0019] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the disease profile generation method as described above.

[0020] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the disease profile generation method as described above.

[0021] The disease profile generation method, apparatus, electronic device, and storage medium provided by this invention, by integrating the deep feature mining capability of hierarchical graph attention network with the ability to distinguish true from false causal inference, realize the automatic discovery of high-order combination relationships of disease features and the quantitative verification of true causality, and can generate dynamic disease profiles that have both deep correlation information and highly credible causal explanatory power. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart illustrating the disease profile generation method provided by the present invention.

[0024] Figure 2 This is one of the schematic diagrams illustrating the calculation process of attention weights for each feature node in the feature association graph provided by this invention.

[0025] Figure 3 This is a schematic diagram of the calculation process for the causal effect value provided by the present invention.

[0026] Figure 4 This is a schematic diagram of the calculation process for the average processing effect of candidate key feature nodes provided by the present invention.

[0027] Figure 5 This is a schematic diagram of the calculation process for the first-order feature representation of the direct association information between feature nodes provided by the present invention.

[0028] Figure 6 This is a schematic diagram of the calculation process for the second-order feature representation of the combination relationship between feature nodes provided by the present invention.

[0029] Figure 7 This is the second schematic diagram of the calculation process of attention weights for each feature node in the feature association graph provided by this invention.

[0030] Figure 8 This is a schematic diagram of the candidate key feature node screening process provided by the present invention.

[0031] Figure 9 This is a schematic diagram of the real-time updating process of disease profiles provided by the present invention.

[0032] Figure 10 This is a schematic diagram of the disease profile generation device provided by the present invention.

[0033] Figure 11 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0035] It should be noted that, in the description of this invention, the terms "comprising," "including," or any other variations thereof are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0036] The terms "first," "second," etc., used in this invention are used to distinguish similar objects, not to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that embodiments of the invention can be implemented in orders other than those illustrated or described herein, and the objects distinguished by "first," "second," etc., are generally of the same class, without limiting the number of objects; for example, a first object can be one or more. Furthermore, "and / or" indicates at least one of the connected objects, and the character " / " generally indicates that the preceding and following objects are in an "or" relationship.

[0037] It should be noted that the acquisition of multi-source data related to the target disease in this application was carried out in compliance with the relevant data protection laws and policies of the country where the application is located, and with the authorization of the respective owners.

[0038] A disease profile is a systematic description of a disease's typical characteristics, high-risk populations, clinical manifestations, and related factors. Existing methods for constructing disease profiles largely rely on the following two approaches: One approach is an inductive method based on expert experience, where medical experts manually define profile labels and correlation strengths based on clinical guidelines, literature reviews, and personal experience. This method is highly subjective, difficult to scale, and cannot handle massive amounts of new medical data from multiple sources (such as genomics and wearable device data).

[0039] Another approach is based on traditional statistical models, such as logistic regression and decision trees, to extract risk factors from data. This method also has significant limitations, primarily in three aspects: first, traditional models struggle to capture complex interactions and nonlinear relationships between features; second, they cannot distinguish between genuine causal drivers and confounding associations, potentially leading to misleading disease profiles; and third, the constructed disease profiles are one-off and static, unable to be dynamically updated with new data, and treat each feature as an independent entity, ignoring their networked connections in high-dimensional space.

[0040] In view of this, this invention provides a novel method that, through an innovative hierarchical graph attention network, not only considers single features but also captures the complex, networked interactions between features, generating more in-depth meta-feature labels. Furthermore, it creatively integrates deep learning and causal inference, adding causal insight to its powerful predictive capabilities, making the generated profiles more effective in guiding clinical interventions rather than merely providing statistical warnings.

[0041] Furthermore, this invention completely changes the model that relies on experts to manually define labels, realizing end-to-end automated generation from data to profiles, greatly improving efficiency and coverage. Even better, this invention introduces an online learning model; as new data continuously flows in, the feature association graph and correlation classification can be automatically updated periodically, enabling disease profiles to evolve dynamically and reflect the latest medical understanding.

[0042] The following is combined with Figures 1-11 This invention describes specific embodiments of the disease profile generation method, apparatus, electronic device, and storage medium provided by the present invention.

[0043] This invention provides a method for generating disease profiles. The executing entity can be a program or module running in an electronic device such as a server, personal computer, or mobile terminal, or it can be a dedicated hardware circuit or a virtual computing instance on a cloud computing platform. In this embodiment, a disease profile generation system deployed on a medical information center server (whose core is the processor within the server) is used as the executing entity for illustration.

[0044] Figure 1 This is a flowchart illustrating the disease profile generation method provided by the present invention, as shown below. Figure 1 As shown, this embodiment provides a method for generating disease profiles. By integrating deep learning models with causal inference techniques, it can automatically mine deep correlations related to diseases from massive amounts of data and verify their causality, thereby generating more in-depth disease profiles with clinical guidance value. Specifically, this includes, but is not limited to, the following steps: Step 11: Construct a feature association map based on multi-source data related to the target disease.

[0045] In this embodiment, the target disease refers to a specific disease that requires profiling analysis, such as type 2 diabetes, hypertension, or Alzheimer's disease.

[0046] Multi-source data refers to a collection of data related to a target disease that originates from different sources and has various formats. For example, when analyzing type 2 diabetes, multi-source data may include: structured data from electronic medical records, such as patient age, gender, height, weight, fasting blood glucose, glycated hemoglobin, and other test indicators; unstructured text, such as doctors' written medical records and discharge summaries; genomic data, such as gene mutation sites related to insulin resistance; lifestyle data collected by wearable devices, such as daily steps and sleep duration; and even environmental data related to the target disease, such as the regional air quality index.

[0047] The feature association graph includes feature nodes that characterize each feature in the multi-source data. In this embodiment, the feature association graph can be constructed by abstracting the complex multi-source data into a computer-processable feature association graph.

[0048] Specifically, a feature association graph is a mathematical model that includes feature nodes and edges connecting them. A feature node is a graph unit representing a specific characteristic in multi-source data. For example, numerical features such as "age," "Body Mass Index (BMI)," and "fasting blood glucose level," as well as categorical features such as "smoking status" and "the presence of a mutation in a specific gene (e.g., TCF7L2)," can all be abstracted as a feature node in the feature association graph. Edges represent an initial association between feature nodes. This initial association can be based on pre-existing medical knowledge or on preliminary statistical analysis of the multi-source data. For example, the correlation coefficient between the features "BMI" and "fasting blood glucose" in the data can be calculated. If the correlation coefficient meets a pre-defined condition (e.g., greater than a certain pre-defined threshold), an edge is established between their corresponding feature nodes. The resulting feature association graph resembles a vast social network, intuitively showcasing the intricate potential connections between hundreds or thousands of risk factors associated with type 2 diabetes.

[0049] Step 12: Input the feature association graph into a preset hierarchical graph attention network (HGAT) to obtain the attention weight of each feature node determined by the hierarchical graph attention network.

[0050] HGAT can be regarded as a deep learning model specifically designed for processing non-Euclidean spatial data such as graphs. Its hierarchical nature enables it not only to learn the direct relationships between feature nodes, but also to discover higher-order combination relationships generated by the combined action of multiple feature nodes (referred to as meta-features in subsequent embodiments of the present invention).

[0051] Using the feature association graph constructed in the previous step as input to HGAT, the hierarchical graph attention network simulates the propagation and aggregation of information within the graph. During this process, the model dynamically evaluates the importance of each neighboring node to the current node through a defined attention mechanism.

[0052] The process of obtaining the attention weights determined by the hierarchical graph attention network is essentially extracting the importance score assigned to each feature node in the hierarchical graph attention network by the model during the learning process. An attention weight is a numerical value that reflects the degree to which the hierarchical graph attention network considers a particular feature node to contribute to the prediction of the target disease. Feature nodes with higher weights are generally considered by the model to be more critical predictors.

[0053] Step 13: Select at least one candidate key feature node from all the feature nodes according to the attention weight.

[0054] After obtaining the attention weights of all feature nodes, the filtering step in the feature association graph can be performed. The purpose of this step is to focus on a small subset of the most noteworthy features from a massive amount of data, thereby improving the efficiency and accuracy of subsequent analysis.

[0055] There are several ways to use the screening method. For example, a number K can be set, and the top K feature nodes with the highest attention weights can be selected as candidate key feature nodes; or a weight threshold can be set, and all feature nodes with attention weights greater than the threshold can be selected as candidate key feature nodes. These screened feature nodes are the suspects that the model considers to be most closely related to the disease, but whether they are the real culprits still needs further verification.

[0056] Step 14: Estimate the causal effect of the candidate key feature nodes to obtain the causal effect value that characterizes the causal strength between the candidate key feature nodes and the target disease.

[0057] Causal effect estimation is an important step in this embodiment that distinguishes it from traditional association analysis. It adopts causal inference methods in statistics to eliminate the interference of confounding factors and explore the true impact of a change in a feature on disease outcome.

[0058] For example, moderate alcohol consumption might be included among the selected candidate key feature nodes. Data suggests that moderate drinkers may have a lower risk of type 2 diabetes, thus giving it a high attentional weight. However, this association is likely spurious because moderate drinkers may also have unobserved confounding factors such as healthier eating habits and higher income levels. The purpose of causal effect estimation is to isolate these confounding factors and calculate the net effect of "moderate drinking" itself on blood glucose levels.

[0059] This step yields a causal effect value. The causal effect value (e.g., Average Treatment Effect, ATE) is a quantitative indicator used to characterize the true causal strength between the corresponding feature node and the target disease. A significantly positive causal effect value indicates that the feature node is a risk factor for the disease, a significantly negative value indicates that it is a protective factor, and a causal effect value close to zero indicates that there may be no direct causal relationship between the feature node and the disease.

[0060] Step 15: Based on the attention weight and the causal effect value, determine the association level of the candidate key feature nodes, so as to generate a disease profile according to the association level.

[0061] After obtaining the predicted importance (i.e., attention weight) and causal truth (i.e., causal effect value) of candidate key feature nodes, the step of determining the association level can be performed. This is a comprehensive judgment under a dual standard. For example, a judgment strategy can be set: When the attention weight and causal effect value of a feature node are both high, its association level can be determined as "strong causal association".

[0062] When the attention weight is high but the causal effect value is low, it can be identified as a "strong predictive association".

[0063] When attention weight is low but causal effect value is high, it can be identified as a "potential causal association".

[0064] Finally, a disease profile can be generated based on the determined association level. The disease profile is the final structured or visual report output by this invention. The process of generating a disease profile involves filling the aforementioned graded feature information into a preset profile template. For example, for the disease profile corresponding to type 2 diabetes, features with strong causal associations (such as "high BMI" and "insulin resistance-related genes") can be displayed as core pathological features, and features with strong predictive associations (such as region and occupation) can be used as high-risk population identification labels. These features can be presented intuitively in various forms such as text descriptions, network diagrams, and radar charts, providing in-depth and reliable decision support for clinicians and researchers.

[0065] The disease profile generation method provided by this invention integrates the deep feature mining capability of hierarchical graph attention network with the ability to distinguish true from false causal inference, thereby realizing the automatic discovery of high-order combination relationships of disease features and the quantitative verification of true causality. It can generate dynamic disease profiles that have both deep correlation information and highly credible causal explanation.

[0066] Based on the above embodiments, a further detailed explanation will be given regarding the step of obtaining the attention weight of each feature node determined by the hierarchical graph attention network, using an example. The hierarchical graph attention network used in this embodiment includes at least a first network layer and a second network layer. This hierarchical structure enables a progressively deeper exploration from direct feature associations to feature combination relationships.

[0067] Figure 2 This is one of the schematic diagrams illustrating the calculation process of attention weights for each feature node in the feature association graph provided by this invention, such as... Figure 2 As shown, obtaining the attention weights of each feature node determined by the hierarchical graph attention network mainly includes, but is not limited to: Step 121: Aggregate the information of neighboring nodes determined by the edges of the feature association graph using the first network layer to generate a first-order feature representation that represents the direct association information between feature nodes.

[0068] Specifically, after the hierarchical graph attention network receives the feature association graph, the first network layer begins its work. Its main task is to learn the relationship between feature nodes and their directly connected neighbor nodes. For each feature node in the feature association graph (e.g., the "fasting blood glucose" feature node), the first network layer pays attention to all its neighbor nodes (e.g., "insulin level," "glycated hemoglobin," "BMI," etc.). Through the attention mechanism, the first network layer calculates an attention coefficient for each neighbor node. This attention coefficient represents the importance of that neighbor node to the state update of the "fasting blood glucose" feature node in the current directly associated context.

[0069] Subsequently, the first network layer weights and aggregates the feature information of all neighboring nodes according to their attention coefficients, thereby generating a first-order feature representation of the feature node "fasting blood glucose" after incorporating the feature information of all its first-order neighbors. The first-order feature representation can be understood as a richer, more context-aware vector representation of each feature node after absorbing the directly related feature information from its surrounding neighbors. It captures relatively simple and direct dependencies between feature information, such as direct, point-to-point associations like "the higher the insulin level, the higher the fasting blood glucose may be."

[0070] Step 122: Aggregate the first-order feature representations using the second network layer to generate second-order feature representations that characterize the combination relationships between feature nodes.

[0071] After the first network layer completes its computation, its output first-order feature representation will serve as the input to the second network layer. The design goal of the second network layer is to achieve a higher level of abstraction, namely, to automatically discover the combinatorial relationships between feature nodes (referred to as meta-features).

[0072] It is important to note that the second network layer is no longer limited to analyzing the relationships between individual feature nodes, but focuses on identifying feature combinations with specific clinical significance formed by the combined effects of multiple feature nodes.

[0073] The aggregation process using the second network layer can be viewed as a concept formation process. For example, the second network layer might discover that the three feature nodes "high BMI," "hypertension," and "hyperlipidemia" (which are now their respective first-order feature representations) frequently appear together in the data and collectively point to a high risk of diabetes. At this point, the second network layer, through its internal aggregation and transformation mechanisms, will fuse these three first-order feature representations into a new, higher-order representation. This representation semantically corresponds to the clinical concept of "metabolic syndrome," and this newly generated higher-order representation is the second-order feature representation.

[0074] Therefore, second-order feature representation is a vector expression used to characterize more complex and non-linear combination relationships between feature nodes, reflecting the synergistic effect of multiple risk factors, and realizing the automatic identification of clinically meaningful feature combinations or meta-features from isolated data points.

[0075] Step 123: Extract the attention coefficients generated during the generation of the first-order feature representation and / or the second-order feature representation to obtain the attention weight of each feature node.

[0076] In the computation process of each layer of the hierarchical graph attention network, whether it is the direct dependence between computing nodes in the first network layer or the recognition of feature combinations in the second network layer, the attention mechanism will generate a large number of attention coefficients. These attention coefficients are the internal basis for the model to make decisions.

[0077] This embodiment extracts and synthesizes these attention coefficients to obtain the final attention weights used to evaluate the importance of each original feature node. For example, the final attention weight of a feature node can be obtained by summing (e.g., summing, averaging, or more complex mapping) the attention coefficients involved in all network layers and all attention heads. If a feature node (such as the "BMI" feature node) is important in the direct association of the first network layer and frequently participates in constituting important meta-features (such as "metabolic syndrome") in the second network layer, then the final attention weight obtained through this extraction and summarization method will be high.

[0078] This invention, through the introduction of a hierarchical graph attention network, particularly the re-aggregation of first-order feature representations via a second network layer, can automatically discover and quantify complex combinations of features from disease-related multi-source data, rather than simply analyzing the isolated effects of individual features. This layer-by-layer mining capability, moving from point-to-point associations to higher-order combinations, provides a deeper understanding of the complex pathological mechanisms of diseases, offering a solid technical foundation for generating more comprehensive and in-depth disease profiles.

[0079] Based on the above embodiments, the following embodiment will describe in detail the implementation of the step of estimating the causal effect of the at least one candidate key feature node to obtain a causal effect value that characterizes the causal strength between the candidate key feature node and the target disease.

[0080] Figure 3 This is a schematic diagram of the calculation process for the causal effect value provided by the present invention, as shown below. Figure 3 As shown, the specific steps include, but are not limited to, the following: Step 141: Extract target sample data for causal verification from the multi-source data. The target sample data includes feature values ​​of multiple sample objects at the candidate key feature nodes and the true label values ​​for the target disease.

[0081] The target sample data is a subset of data extracted from multi-source data, used to perform causal inference. The sample object typically refers to an independent individual, such as a patient. The feature value refers to the specific state of a sample object at a candidate key feature node, usually treated as a binary variable, such as 1 representing the presence of the candidate key feature and 0 representing its absence. The true label value refers to the actual record of whether the sample object suffers from the target disease, such as 1 representing the disease and 0 representing the absence. In addition, the target sample data also contains a series of covariates that may be related to both the candidate key feature node and the target disease. These covariates are confounding factors, which need to be controlled and excluded for causal inference.

[0082] Optionally, for the i-th sample object (e.g., a specific patient) in the dataset consisting of target sample data extracted from multi-source data for causal verification, its data can be formally represented as a tuple (Xi, Ti, Yi), where: Xi represents a covariate vector that contains all other relevant features of the sample object besides the candidate key features, such as age, gender, BMI, etc. These are potential confounding factors that need to be controlled.

[0083] Ti represents the feature value, indicating whether the sample object has the candidate key feature currently being analyzed. For example, Ti=1 means it has the feature, and Ti=0 means it does not.

[0084] Yi is the true label value, which represents the true record of whether the sample object has the target disease. For example, Yi=1 indicates that the object has the disease, and Yi=0 indicates that the object does not have the disease.

[0085] Therefore, the target sample data is a set {(Xi, Ti, Yi)} containing N sample objects, where {i=1 to N}.

[0086] Step 142: Obtain the pre-trained Propensity Score Model and Outcome Regression Model.

[0087] The propensity score model and the outcome regression model are two core components for performing dual robust estimation. They can be pre-trained, meaning that the two models have already completed parameter learning on a large-scale, representative dataset and can be directly used for subsequent prediction tasks without retraining each time a disease profile is generated.

[0088] Optionally, the goal of the propensity score model is to simulate the feature allocation mechanism, which outputs the conditional probability that the sample object is assigned to a treatment group for the input covariate vector X of a sample object. This probability value is the propensity score.

[0089] Propensity score models can be implemented using algorithms such as logistic regression, gradient boosting machines (e.g., XGBoost), or neural networks.

[0090] The goal of the outcome regression model is to simulate the conditional expectation of the outcome under given covariates and treatment conditions; that is, for an input sample object with treatment state T and covariate X, output a predicted value of its probability of having the target disease. The outcome regression model can be implemented by algorithms such as generalized linear models or deep neural networks.

[0091] Step 143: Input the target sample data into the propensity score model and output the propensity score of each sample object on the candidate key feature node.

[0092] Furthermore, this embodiment utilizes the two pre-trained models described above for inference and prediction. For each sample object, the propensity score model provides a propensity score, which quantifies the degree to which the sample object naturally tends to possess the candidate key feature currently being analyzed due to its other background features (covariates).

[0093] Step 144: Input the target sample data into the result regression model and output the predicted probability value of each sample object on the target disease.

[0094] Accordingly, the resulting regression model may provide two probability predictions of having the target disease: one is the prediction assuming the sample has the candidate key feature, and the other is the prediction assuming it does not have the candidate key feature.

[0095] Step 145: Calculate the average processing effect of the candidate key feature nodes based on the propensity score, the predicted probability of disease, the feature value, and the true label value.

[0096] Step 146: The average treatment effect is determined as the causal effect value between the candidate key feature node and the target disease.

[0097] Considering that simply comparing the average results of the treatment and control groups can be biased due to confounding factors, this embodiment uses a comprehensive calculation strategy to systematically utilize all the information obtained in the previous step to correct for this bias. Specifically: the disease probability prediction value provides a preliminary estimate of the causal effect based on regression adjustment; the propensity score can be used to construct inverse probability weights to balance the differences in the distribution of covariates between the two groups; and the feature values ​​and the true label values ​​are used to calculate the prediction residuals of the regression model (i.e., the difference between the predicted value and the true value), and thus correct the preliminary estimate.

[0098] By applying this comprehensive calculation strategy to all sample objects in the dataset, an unbiased average treatment effect (ATE) estimate is finally obtained, which is then determined as the final causal effect value of the candidate key feature node.

[0099] Below, we will continue to focus on "type 2 diabetes" as the target disease and take "long-term use of metformin" as a candidate key feature node to illustrate the specific process of this embodiment.

[0100] First, a target sample dataset containing 10,000 patients is extracted from the electronic medical record database. For the i-th patient, the covariate vector Xi (including age, gender, BMI, family history, etc.), the feature value Ti of whether or not they have been "taking metformin long-term" (1 indicates yes, 0 indicates no), and the true label value Yi of whether or not they were ultimately diagnosed with "type 2 diabetes" (1 indicates yes, 0 indicates no) are recorded.

[0101] Load a pre-trained XGBoost-based propensity score model and a deep neural network-based outcome regression model. For the i-th patient, input their covariate Xi into the propensity score model to obtain their propensity score e(Xi). For example, an older and overweight patient might have a score of 0.6, indicating that they have a 60% probability of being advised to take medication due to background factors.

[0102] Furthermore, by inputting its covariate Xi into the outcome regression model, two disease probability prediction values ​​are obtained: μ1(Xi) (let's say it is the disease prediction probability after taking the medication) and μ0(Xi) (let's say it is the disease prediction probability after not taking the medication).

[0103] The actual observed values ​​(Ti, Yi) and model predicted values ​​(e(Xi), μ1(Xi), μ0(Xi)) for each patient are substituted into a comprehensive calculation formula, such as the dual robust estimation formula. The dual robust estimation formula calculates the corrected causal effect for each patient through weighting and residual correction.

[0104] Finally, the adjusted effects of all 10,000 patients were averaged to obtain the final mean treatment effect (ATE). If the calculated ATE is -0.15, this constitutes the causal effect value of the characteristic information "long-term use of metformin", indicating that after excluding confounding factors such as age and weight, this behavior itself can reduce the risk of disease by an average of 15%.

[0105] This invention introduces a processing flow that includes data extraction, pre-trained model application, and comprehensive calculation, which enables effective causal effect estimation of candidate key features. This allows the solution to not only discover correlations but also verify causality, thereby greatly improving the scientific validity and credibility of the final disease profile.

[0106] Figure 4 This is a schematic diagram of the calculation process for the average processing effect of candidate key feature nodes provided by the present invention. As an optional embodiment, such as... Figure 4 As shown, this embodiment provides a specific implementation method for calculating the average processing effect of the candidate key feature nodes based on the propensity score, the predicted disease probability value, the feature value, and the true label value, including but not limited to the following steps: Step 1451: Construct sample weights using the propensity score and the feature value.

[0107] Step 1452: Calculate the prediction residual between the true label value and the predicted disease probability value.

[0108] Step 1453: The predicted residual is weighted based on the sample weights, and the weighted residual is fused with the predicted disease probability value to obtain the causal effect estimate.

[0109] Step 1454: The causal effect estimates of all sample objects are averaged to obtain the average treatment effect.

[0110] Specifically, this embodiment provides an implementation method for calculating causal effect estimates based on dual robust estimation. The advantage of dual robust estimation is that as long as either the bias score model or the outcome regression model is correctly specified, the final causal effect estimate will be unbiased, thereby greatly enhancing the robustness of the estimation results.

[0111] In step 1451, an environment of a randomized controlled trial is simulated using inverse probability weighting (IPW) to eliminate confounding effects caused by imbalanced covariate distributions. For the i-th sample in the target sample data, its sample weight Wi is constructed as follows: (1) If the sample object belongs to the treatment group, that is, its feature value Ti=1, then its sample weight is Wi =1 / e(Xi), where e(Xi) is its propensity score.

[0112] (2) If the sample belongs to the control group, that is, its feature value Ti=0, then its sample weight is Wi=1 / (1-e(Xi)).

[0113] By assigning such weights to each sample, the distributions of the treatment and control groups across all covariates can be made more consistent in the weighted sample population, thereby correcting for selection bias.

[0114] Next, step 1452 is performed to quantify the predictive accuracy of the regression model. The prediction residual refers to the difference between the model's predicted probability of disease for each sample and its actual disease status. For the i-th sample, its prediction residual Resid... i The calculation method is as follows: (1) If the sample belongs to the treatment group (Ti=1), its prediction residual is Resid. i =Yi-μ1(Xi), where Yi is its true label value and μ1(Xi) is the probability of disease under treatment predicted by the regression model.

[0115] (2) If the sample belongs to the control group (Ti=0), its prediction residual is Resid. i =Yi-μ0(Xi), where μ0(Xi) is the probability of disease predicted by the outcome regression model in the control state. The above prediction residuals capture information that the outcome regression model could not fully explain.

[0116] Furthermore, step 1453 is executed to organically integrate the ideas based on regression adjustment with those based on inverse probability weighting.

[0117] Specifically, for the i-th sample object, first calculate its causal effect estimates τ1i and τ0i under the two potential outcomes: τ1i=μ1(Xi) +(Ti / e(Xi))*(Yi-μ1(Xi)); τ0i=μ0(Xi)+((1-Ti) / (1-e(Xi)))*(Yi-μ0(Xi)); As can be seen, the estimated value of each potential outcome consists of two parts: one part is the direct prediction of the outcome regression model, such as μ1(Xi); the other part is the correction term obtained by weighting the predicted residuals based on the sample weights, such as (Ti / e(Xi))*(Yi-μ1(Xi)). The above calculation method ensures the dual robustness: if the outcome regression model is perfect, i.e., the predicted residual is 0, then the correction term is 0, and the estimated value equals the model prediction; if the propensity score model is perfect, then the expected value of the weighted residual is 0, and the estimated value also tends to be unbiased.

[0118] Finally, after performing step 1454, which involves calculating the estimated values ​​τ1i and τ0i for each sample object under the two potential outcomes, the expected values ​​of these two potential outcomes at the population level can be obtained by averaging all target sample data. E[Y(1)]=(1 / N)*Σ(τ1i); E[Y(0)]=(1 / N)*Σ(τ0i); Where N is the total number of samples; Σ represents the summation of all sample objects i=1 to N; E[Y(1)] represents the expected value of the average potential outcome of the target disease under the assumption that all sample objects in the target sample data are treated; E[Y(0)] represents the expected value of the average potential outcome of the target disease under the assumption that all sample objects in the target sample data are not treated.

[0119] Ultimately, the average treatment effect obtained, which is the causal effect value ATE output by this embodiment after double robust correction, is the difference between these two expected values: ATE=E[Y(1)]-E[Y(0)].

[0120] The technical solution of this embodiment provides a method for calculating the average treatment effect based on dual robust estimation. By constructing sample weights, calculating prediction residuals, fusing weighted residuals with predicted values, and finally obtaining the average difference, it provides a specific technical implementation of how to perform causal inference with high quality, which greatly enhances the feasibility of the solution.

[0121] Figure 5 This is a schematic diagram of the calculation process for the first-order feature representation of the direct association information between feature nodes provided by the present invention, as shown below. Figure 5 As shown, this embodiment provides another specific implementation method for generating a first-order feature representation characterizing the direct association information between feature nodes by aggregating information of neighboring nodes determined by the edges of the feature association graph using the first network layer. This mainly includes, but is not limited to: Step 1211: For any target feature node in the feature association graph, obtain the initial feature vector of the target feature node and its neighboring nodes.

[0122] The initial feature vector is the raw numerical representation of each feature node before it is input into the hierarchical graph attention network. For numerical features, the initial feature vector can be its normalized value; for categorical features, it can be its one-hot encoding vector; in more complex implementations, all features can be mapped to a dense vector of uniform dimension through a shared embedding layer, which serves as their initial feature vector.

[0123] Step 1212: Perform a linear transformation on the initial eigenvector using a linear transformation matrix to obtain the transformed eigenvector.

[0124] To enhance the model's expressive power, this embodiment will perform dimensionality increase or feature space transformation on the initial feature vector. Specifically, this can be achieved using a trainable linear transformation matrix (also known as the weight matrix W). For any node's initial feature vector h, its transformed feature vector h' can be calculated using the following formula: h' = W * h.

[0125] After this operation, all nodes will be mapped to a new, higher-dimensional feature space, so that subsequent attention mechanisms can better learn and distinguish features.

[0126] Step 1213: Based on the transformed feature vector, calculate the correlation coefficient between the target feature node and each neighbor node.

[0127] For any target feature node i and one of its neighbor nodes j, the correlation coefficient (or attention score) e between them can be calculated using an attention function. ij A common implementation is to transform the feature vector h' of the two nodes. i and h' j The data is concatenated and then input into a single-layer fully connected neural network. The parameters of the fully connected neural network are learnable, and the resulting correlation coefficient e... ij The magnitude of reflects the importance of the information of neighbor node j to the target feature node i under the current task.

[0128] Step 1214: Normalize the correlation coefficient using an activation function to obtain the normalized attention coefficient.

[0129] To make the importance scores of different neighboring nodes comparable, this embodiment will normalize the correlation coefficients obtained in the previous step.

[0130] Softmax can be used as the activation function. Specifically, for the target feature node i, its correlation coefficient {e_i} with all its neighboring nodes j is... ij All of these will be input into the Softmax function, resulting in a set of normalized attention coefficients {α} that sum to 1. ij}. Each α ij This represents the weight proportion of neighbor node j when it exerts influence on target feature node i.

[0131] Step 1215: Use the normalized attention coefficient to perform weighted aggregation on the transformed feature vectors of the neighboring nodes to obtain the first-order feature representation of the target feature node.

[0132] Specifically, for the target feature node i, its final first-order feature representation h'' i This is achieved by transforming the feature vector {h' of all its neighboring nodes. j According to the corresponding normalized attention coefficient {α} ij The result is obtained by weighted summation.

[0133] Optionally, a nonlinear activation function (such as ReLU or ELU) can be applied to the weighted sum. This newly generated first-order feature representation h'' i A vector is a first-order feature representation that integrates information from all neighboring nodes and is dynamically weighted by an attention mechanism, representing directly related information.

[0134] This invention provides an aggregation process based on an attention mechanism within the first network layer, which enables the model to automatically and dynamically learn the direct correlation strength between different features, rather than relying on fixed, predefined weights. This allows the generated first-order feature representation to more accurately capture the complex nonlinear dependencies between features.

[0135] Figure 6 This is a schematic diagram illustrating the calculation process of the second-order feature representation of the combination relationship between feature nodes provided by the present invention, as shown below. Figure 6 As shown, this embodiment provides another specific implementation method for generating a second-order feature representation that characterizes the combination relationship between feature nodes by aggregating the first-order feature representation using the second network layer, which mainly includes, but is not limited to: Step 1221: For the connected feature node pairs, perform element-wise product operation or concatenation operation on the first-order feature representations of the feature node pairs to generate feature interaction vectors.

[0136] Step 1222: Calculate the importance score of the feature interaction vector in representing the feature combination effect using a multi-head attention mechanism.

[0137] Step 1223: Aggregate the feature interaction vectors based on the importance scores to generate the second-order feature representation that characterizes the meta-features formed by the combined interactions between the feature nodes.

[0138] After the first network layer outputs the first-order feature representations of all feature nodes, the task of the second network layer is to explore the combinatorial effects among these first-order feature representations. In order to explicitly construct combinatorial relationships, this embodiment first operates on feature node pairs that are connected in the feature association graph, or, in a more complex implementation, on all possible feature node pairs.

[0139] For any pair of feature nodes, such as node i and node j, their respective first-order feature representations h'' are obtained. i and h'' j Then, an interaction vector representing the combined effect of the two is generated through interactive operations. Optional interaction operation methods include: Method 1 involves multiplying the corresponding elements of the two first-order feature representation vectors. This operation can capture the co-occurrence or co-enhancement / weakening patterns of the two first-order feature representations across various dimensions.

[0140] Method 2 directly concatenates two first-order feature representation vectors into a longer vector. This method preserves all the original information and delegates the task of learning the combination relationship to subsequent network layers.

[0141] The generated feature interaction vector is no longer a representation of a single feature semantically, but a combination of two first-order feature representations.

[0142] After generating a large number of feature interaction vectors, considering that not all feature combinations have equal clinical significance, this embodiment introduces a multi-head attention mechanism to filter out important combinations. This mechanism filters these feature interaction vectors from different perspectives and calculates an importance score for each interaction vector. For example, one attention head may focus more on metabolism-related feature combinations, while another attention head may focus more on inflammation-related feature combinations. In this way, the multi-head attention mechanism can more comprehensively capture the diversity of feature combinations.

[0143] After calculating the importance score for each feature interaction vector, this implementation proceeds to perform an aggregation operation. Specifically, all feature interaction vectors are weighted and summed according to their corresponding importance scores, and the resulting vector is the second-order feature representation.

[0144] It should be noted that this second-order feature representation no longer directly corresponds to any single original feature node, but rather represents the meta-feature formed by the combination and interaction between the most important feature nodes in the entire feature association graph. It is a highly condensed numerical expression that semantically represents a certain high-order clinical concept (such as "metabolic syndrome", "cytokine storm").

[0145] Next, taking the generation of a disease profile for "type 2 diabetes" as an example, we assume that the second network layer is analyzing the first-order feature representation output by the first network layer.

[0146] First, the two connected feature nodes "high BMI" and "hypertension" were selected. After obtaining their respective 256-dimensional first-order feature representations, element-wise product operations were used to generate a 256-dimensional feature interaction vector representing the interaction between "high BMI and hypertension". Similarly, feature interaction vectors were generated for other feature node pairs (such as "hyperlipidemia and insulin resistance").

[0147] These feature interaction vectors are input into a pre-built multi-head attention mechanism module. Assume this module has eight attention heads, each independently calculating a set of importance scores for all feature interaction vectors. One attention head might, after learning, discover that the combination of "high BMI and hypertension interaction" is crucial for predicting diabetes, and therefore assign it a very high importance score.

[0148] Finally, all feature interaction vectors are weighted and summed according to the comprehensive importance score given by the multi-head attention mechanism. Since combinations such as "high BMI and hypertension interaction" and "high blood lipids and insulin resistance interaction" have higher importance scores, they contribute more to the final weighted sum. The resulting aggregated vector is the second-order feature representation obtained in this round of calculation. This second-order feature representation conceptually points to the meta-feature of "metabolic syndrome," even though the model is unaware of this medical term, it has captured the essence of this combined concept from the data level.

[0149] The embodiments of the present invention clearly define the process of discovering and aggregating feature combinations within the second network layer, realizing the ability to mine meta-features with higher-level semantics from data. This makes the generated disease profile no longer just a list of isolated risk factors, but can reveal complex pathological patterns under the synergistic effect of multiple factors, greatly improving the depth and insight of the profile.

[0150] Figure 7 This is the second schematic diagram illustrating the calculation process of the attention weights of each feature node in the feature association graph provided by this invention, as shown below. Figure 7As shown, this embodiment further provides a specific implementation method for extracting the attention coefficients generated during the generation of the first-order feature representation and / or the second-order feature representation to obtain the attention weight of each feature node, including but not limited to: Step 1231: Obtain the attention coefficient matrix output by the multi-head attention mechanism in the first network layer and / or the second network layer in the hierarchical graph attention network.

[0151] Step 1232: For any feature node, sum or average all attention coefficients related to the feature node in the attention coefficient matrix to obtain a comprehensive importance score.

[0152] Step 1233: Determine the comprehensive importance score as the attention weight of any feature node.

[0153] In the computation of hierarchical graph attention networks, especially when using a multi-head attention mechanism, each layer and each attention head generates an attention coefficient matrix. This attention coefficient matrix records the degree of mutual attention between feature nodes or feature interaction vectors under that layer and that attention head.

[0154] For example, in the first network layer, there will be one or more attention coefficient matrices, whose elements α ij This represents the attention coefficient of node j to node i. In the second network layer, there is also an attention coefficient matrix representing the attention between feature interaction vectors. The purpose of this operation is to obtain all these intermediate results related to attention calculation generated during the model's forward propagation.

[0155] After acquiring a large number of attention coefficient matrices, this embodiment integrates and attributes this information, which is scattered across different layers and attention heads, back to each original feature node. For any feature node (e.g., node k), all attention coefficients associated with that feature node are identified, which may include: (1) In the first network layer, when node k is the target node, the attention coefficients it receives from all its neighboring nodes.

[0156] (2) In the first network layer, when node k is a neighbor node, the attention coefficient it applies to all other nodes.

[0157] (3) In the second network layer, the attention coefficients obtained by the feature interaction vector formed by node k.

[0158] After identifying all these attention coefficients, an aggregation operation, such as summation or averaging, can be used to obtain a comprehensive importance score. This score is a quantitative indicator that fully reflects the role of a feature node in the complex operations of the entire hierarchical graph attention network. A higher comprehensive importance score indicates that the feature node plays a more prominent role, whether as a direct influencing factor or as a component of key meta-features. Therefore, this embodiment directly determines this comprehensive importance score as the final, global attention weight for that feature node.

[0159] The present invention provides a method for extracting and integrating information from a multi-layered, multi-headed attention coefficient matrix to obtain the global attention weight of each original feature node. This method solves the technical problem of how to transform the complex, distributed attention information within a deep learning model into a simple, intuitive, and globally important indicator that can be used for ranking and filtering.

[0160] Based on the above embodiments, as an optional embodiment, a specific implementation method is provided for filtering at least one candidate key feature node from all the feature nodes according to the attention weight, which mainly includes, but is not limited to: First, all the feature nodes are sorted from high to low according to the attention weight.

[0161] Specifically, a list can be constructed of all the acquired feature nodes and their corresponding attention weights, and then the list can be sorted in descending order, with the feature node with the highest attention weight at the beginning and the one with the lowest weight at the end. This sorted list intuitively reflects the feature sequence from most important to least important as perceived by the model.

[0162] After sorting, this embodiment provides two parallel filtering strategies: Strategy 1: Select the top K feature nodes in the sorting results as the candidate key feature nodes, where K is a preset positive integer.

[0163] The strategy employed here is a fixed-number Top-K strategy. In this embodiment, K is a preset positive integer that can be adjusted based on the specific application scenario and computing resources. For example, in preliminary exploratory analysis, K can be set to a larger value (e.g., 100) to include more possible features. When building a streamlined predictive model, K can be set to a smaller value (e.g., 20). During the filtering process, the top K feature nodes are selected directly from the top of the sorted list and identified as candidate key feature nodes for subsequent analysis.

[0164] The advantage of implementing this strategy is that it is simple and intuitive, and it can ensure that the number of features selected is constant, which facilitates subsequent processing.

[0165] Strategy 2: Select feature nodes whose attention weight is greater than a preset weight threshold as candidate key feature nodes.

[0166] This employs a selection strategy based on an importance threshold. In this implementation, a preset weight threshold can be set, representing the minimum importance standard that the model considers a feature node worthy of attention. Next, the sorted feature list is traversed, selecting all feature nodes whose attention weight is greater than the preset weight threshold and identifying them as candidate key feature nodes.

[0167] The advantage of this strategy is that its selection criteria are based on the absolute importance of the features themselves rather than their relative ranking. If there are many important features in a given analysis, the selected candidate set will be larger; conversely, it will be smaller. This approach is more flexible and better suited to situations where the distribution of feature importance is uneven across different diseases and datasets.

[0168] Based on the above embodiments, as an optional embodiment, a specific implementation method is provided for determining the association level of the candidate key feature nodes based on the attention weight and the causal effect value.

[0169] Specifically, in this embodiment, the association level of the candidate key feature nodes can be determined by constructing a two-dimensional judgment matrix, which combines the predictive ability of the feature nodes represented by attention weights with the causal strength represented by causal effect values, to classify the candidate key feature nodes in a more refined and clinically significant manner.

[0170] In this embodiment, at least two thresholds can be preset: a first preset threshold for evaluating the importance of attention weights, and a second preset threshold for evaluating the significance of causal effect values. Based on these two thresholds, each candidate key feature node can be classified and determined as follows: If the attention weight is greater than a first preset threshold and the causal effect value is greater than a second preset threshold, the correlation level of the candidate key feature node is determined to be strong correlation.

[0171] If the attention weight is greater than the first preset threshold but the causal effect value is less than or equal to the second preset threshold, the association level of the candidate key feature node is determined to be pseudo-correlation or weak correlation.

[0172] If the attention weight is less than or equal to the first preset threshold but the causal effect value is greater than the second preset threshold, the association level of the candidate key feature node is determined to be potentially relevant.

[0173] When the attention weight of a candidate key feature node exceeds a first preset threshold, indicating that it is a strong predictor in the disease prediction model, and the absolute value of its causal effect value also exceeds a second preset threshold, indicating that it has a statistically significant true causal relationship with the disease, its association level is determined to be strong correlation, or strong causal association. These feature nodes with strong causal associations are the most valuable features, not only effectively predicting disease risk, but also likely being key links in the disease development pathway, and potential intervention targets.

[0174] When a candidate key feature node has a high attention weight, exceeding a first preset threshold, but its absolute causal effect value is small and fails to exceed a second preset threshold, it indicates that although the feature node is statistically highly correlated with the disease outcome, this correlation is likely not caused by direct causation but rather by other unobserved confounding factors. Therefore, this embodiment will determine its association level as a spurious correlation or weak correlation, or a strong predictive non-causal association. Although such features can be used as markers of risk prediction, they should not be considered as direct intervention targets.

[0175] When the attention weight of a candidate key feature node is not prominent and fails to exceed the first preset threshold, but the absolute value of its causal effect value significantly exceeds the second preset threshold, it means that although the feature node is not assigned a high weight in a complex prediction model due to collinearity with other features, it does have an independent causal effect on the disease. This embodiment will determine its association level as potential correlation, or potential causal association. These features are easily overlooked by traditional association analysis methods, but they may represent new pathological mechanisms or intervention targets worthy of in-depth research.

[0176] Of course, for features whose attention weight and causal effect value are both below their respective thresholds, their association level can be determined as irrelevant or they can be directly excluded from the disease profile.

[0177] Figure 8 This is a schematic diagram of the candidate key feature node screening process provided by the present invention, as shown below. Figure 8 As shown, this embodiment provides a specific implementation method for constructing a feature association map based on multi-source data related to a target disease, which mainly includes, but is not limited to: Step 111: Calculate the statistical correlation between any two features in the multi-source data based on statistical methods.

[0178] Step 112: If the statistical correlation meets the preset conditions, establish an edge between the feature nodes corresponding to the two features, and use the value of the statistical correlation as the initial weight of the edge.

[0179] Specifically, after abstracting all features into feature nodes, in order to determine whether there is an initial correlation between them, it is necessary to further quantify the strength of their relationship. In this embodiment, for any two features (e.g., feature A and feature B), their corresponding numerical columns are extracted from multi-source data, and then the statistical correlation between them is calculated.

[0180] There are various statistical methods that can be used, and the specific choice depends on the data type of the feature. For example, for two continuous numerical features (such as "BMI" and "fasting blood glucose"), their Pearson correlation coefficient or Spearman rank correlation coefficient can be calculated; for two categorical features (such as "smoker" and "gender"), their mutual information can be calculated or the association strength can be obtained through the chi-square test; and for one continuous feature and one categorical feature, the continuous feature can be binned and calculated according to the categorization, or methods such as point bicollinear correlation coefficient can be used. The specific implementation details of the above methods will not be elaborated here.

[0181] After calculating the statistical correlation between all feature pairs, we can use these values ​​to decide whether to create edges in the feature association graph.

[0182] Optionally, this embodiment introduces a preset condition as the basis for edge construction. This preset condition can be a simple threshold, such as "the absolute value of statistical correlation is greater than 0.3" or "the mutual information value is greater than 0.1", etc.

[0183] This embodiment iterates through all feature pairs, and if the statistical correlation meets a preset condition, an undirected edge is established between the feature nodes corresponding to the two features. This approach is equivalent to performing a sparsification process on the fully connected graph, retaining only the statistically significant correlations, which helps reduce the complexity and noise of subsequent model calculations.

[0184] Furthermore, to enrich the information contained in the feature association graph, this embodiment uses the statistical correlation values ​​as the initial weights of the edges. This means that each edge in the feature association graph not only represents a relationship but also indicates the strength of that relationship through its weight. For example, the weight of the edge between "BMI" and "fasting blood glucose" might be 0.6, while the weight of the edge between "age" and "sleep duration" might be 0.2. These initial weights can serve as useful prior information for the learning process of the hierarchical graph attention network.

[0185] The method proposed in this embodiment is to automatically construct edges and their initial weights in a feature association graph based on statistical correlation. This data-driven graph construction method, compared with relying on manual experience to define or simple fully connected graphs, can more objectively and efficiently capture the inherent association structure in the data, thereby improving the accuracy and efficiency of the entire disease profile generation method.

[0186] Figure 9 This is a schematic diagram of the real-time updating process of disease profiles provided by the present invention, such as... Figure 9 As shown, this embodiment introduces an online model update mechanism, designed to address the real-world needs of continuously growing medical data and evolving disease understanding. Through this mechanism, the disease profile generation method provided by this invention is no longer a static, one-off analysis tool, but a dynamic system capable of self-learning and self-improvement. Specifically, after completing one profile generation, this embodiment can also periodically or under specific conditions execute the following steps: Step 21: In response to the acquisition of new multi-source data, update the feature association map using the new multi-source data.

[0187] Step 22: Based on the updated feature association graph, perform incremental training on the hierarchical graph attention network, the propensity score model, and the result regression model.

[0188] Step 23: Re-determine the attention weights using the trained hierarchical graph attention network.

[0189] Step 24: Re-estimate the causal effect using the trained propensity score model and outcome regression model to generate the updated disease profile based on the newly added multi-source data.

[0190] The triggering condition for model updates set in this embodiment can be the acquisition of newly added multi-source data (hereinafter referred to as new data). This can be a time-driven event, for example, set to perform an update once every month; or it can be a data volume-driven event, for example, triggering an update when the newly collected patient data reaches 1,000 cases.

[0191] After an update is triggered, the feature association graph will be updated using the newly added data. The update methods may include, but are not limited to, updating feature nodes, updating edges and weights.

[0192] If new features, never before seen, appear in the new data, such as a new laboratory test indicator or gene locus, a new feature node will be created and added to the feature association graph. Additionally, the statistical correlation between all feature pairs will be recalculated by combining the existing and new data. Based on the new correlation values, new edges may be added to the existing feature association graph, or old edges may be deleted, and the initial weights of all edges will be updated.

[0193] When the feature association map and data foundation change, the model needs to be updated to adapt to these changes, for example, by using incremental training. Compared to a complete retraining from scratch on the full dataset, incremental training uses existing model parameters as a starting point and only performs several rounds of retraining on new data or updated maps. This approach has lower computational costs and faster training speed. Specifically: For hierarchical graph attention networks, incremental training enables them to learn representations of new features and adjust their judgments on the strength of associations between existing features and between new and old features.

[0194] For propensity score models and outcome regression models, incremental training enables them to more accurately predict propensity scores and disease incidence probabilities under new population distributions and data characteristics.

[0195] After completing the incremental training of the model, all model components have been updated to the latest state. At this point, these trained models can be used to execute the complete process of the disease profile generation method provided in the previous examples, namely, redetermining attention weights, re-estimating causal effects, and finally generating an updated disease profile. This new disease profile not only includes insights from the original data but also incorporates new knowledge and trends brought by the newly added data, thus more accurately reflecting the latest understanding of the target disease at the current point in time.

[0196] Based on the above embodiments, as an optional embodiment, generating a disease profile according to the association level specifically includes: Generate a visual knowledge graph or multi-dimensional radar chart containing the candidate key feature nodes; In the visualized knowledge graph, the size or color of the nodes represents the magnitude of the attention weight, and the thickness or type of the lines connecting the nodes represents the strength of the causal effect value; or, in the disease profile, the text description related to the combination relationship between the feature nodes corresponding to the second-order feature representation is explicitly output.

[0197] This embodiment provides a specific and visual implementation of the step of generating disease profiles based on association levels. The aim is to transform calculated abstract data, such as attention weights, causal effect values, and association levels, into an intuitive form that is easy for users to understand and interpret, thereby maximizing the application value of disease profiles. Specifically, after determining the association levels of candidate key feature nodes, this information is organized and presented in a meaningful way.

[0198] This embodiment provides two optional visualization formats: Method 1 presents information in a graph structure by treating each candidate key feature node as a visual node in a visualized knowledge graph and drawing connecting lines based on known associations (such as edges existing in the feature association graph) or newly discovered associations (such as jointly constituting a meta-feature). This method is very suitable for displaying network relationships and complex dependencies between features.

[0199] Method two presents the data in the form of a multi-dimensional radar chart, which is essentially a way to display numerical values ​​on multiple coordinate axes. Multiple evaluation dimensions can be set, such as "predictive importance," "causal strength," and "data availability." Then, the scores of each candidate key feature node on these dimensions are plotted as a radar chart. This method is very suitable for comprehensive evaluation and comparison of individual features from multiple perspectives.

[0200] Secondly, in order to contain richer information in the visualization charts, this embodiment further defines the information encoding method. This encoding method encodes numerical information through the attributes of visual elements, so that users can intuitively grasp the core points.

[0201] For example, the radius of a node can be proportional to its attention weight, or a color gradient from cool (low weight) to warm (high weight) can be used to fill the node. This way, users can immediately see which features the model considers the most important predictors.

[0202] For example, the strength of the causal effect calculated between two feature nodes can be mapped to the thickness of the line connecting them. Alternatively, different types of lines, such as solid lines, dashed lines, and arrows, can be used to represent the direction and type of causal relationship.

[0203] Finally, this embodiment also provides a way to directly output meta-feature information. The second network layer generates a second-order feature representation of the meta-feature. Although this is an internal mathematical vector, it is traceable to which original feature nodes contributed the most to the generation of this representation. This embodiment translates this abstract meta-feature into a human-readable text description. For example, if a second-order feature representation is mainly composed of the aggregation of three first-order feature representations, "high BMI," "hypertension," and "hyperlipidemia," then a text can be explicitly output in the final disease profile report, such as: "A strongly correlated combination of clinical features was detected: metabolic syndrome, caused by the synergistic effect of high BMI, hypertension, and hyperlipidemia."

[0204] Figure 10 This is a schematic diagram of the disease profile generation device provided by the present invention, as shown below. Figure 10 As shown, it mainly includes, but is not limited to: The graph construction module 1 is used to construct a feature association graph based on multi-source data related to the target disease. The feature association graph includes feature nodes that characterize each feature in the multi-source data.

[0205] The weight acquisition module 2 is used to input the feature association map into a preset hierarchical graph attention network and obtain the attention weight of each feature node determined by the hierarchical graph attention network.

[0206] Feature filtering module 3 is used to filter at least one candidate key feature node from all the feature nodes according to the attention weight.

[0207] Causal estimation module 4 is used to estimate the causal effect of the candidate key feature nodes to obtain a causal effect value that characterizes the causal strength between the candidate key feature nodes and the target disease.

[0208] The profile generation module 5 is used to determine the association level of the candidate key feature nodes based on the attention weight and the causal effect value, so as to generate a disease profile according to the association level.

[0209] It should be noted that the disease profile generation device provided by the present invention can execute the disease profile generation method described in any of the above embodiments during specific operation, which will not be elaborated in this embodiment.

[0210] The disease profile generation method provided by this invention integrates the deep feature mining capability of hierarchical graph attention network with the ability to distinguish true from false causal inference, thereby realizing the automatic discovery of high-order combination relationships of disease features and the quantitative verification of true causality. It can generate dynamic disease profiles that have both deep correlation information and highly credible causal explanation.

[0211] Figure 11This is a schematic diagram of the structure of the electronic device provided by the present invention, such as... Figure 11 As shown, the electronic device may include: a processor 1110, a communications interface 1120, a memory 1130, and a communications bus 1140, wherein the processor 1110, the communications interface 1120, and the memory 1130 communicate with each other through the communications bus 1140. Processor 1110 can call logical instructions in memory 1130 to execute a disease profile generation method, the method comprising: constructing a feature association map based on multi-source data related to a target disease, the feature association map including feature nodes characterizing each feature in the multi-source data; inputting the feature association map into a preset hierarchical graph attention network to obtain an attention weight for each feature node determined by the hierarchical graph attention network; selecting at least one candidate key feature node from all the feature nodes based on the attention weight; performing causal effect estimation on the candidate key feature node to obtain a causal effect value characterizing the causal strength between the candidate key feature node and the target disease; and determining the association level of the candidate key feature node based on the attention weight and the causal effect value, so as to generate a disease profile based on the association level.

[0212] Furthermore, the logical instructions in the aforementioned memory 1130 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0213] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions, wherein when the program instructions are executed by a computer, the computer is able to execute the disease profile generation method provided in the above embodiments, the method comprising: constructing a feature association map based on multi-source data related to a target disease, the feature association map comprising feature nodes characterizing each feature in the multi-source data; inputting the feature association map into a preset hierarchical graph attention network to obtain an attention weight for each feature node determined by the hierarchical graph attention network; selecting at least one candidate key feature node from all the feature nodes according to the attention weight; performing causal effect estimation on the candidate key feature node to obtain a causal effect value characterizing the causal strength between the candidate key feature node and the target disease; and determining the association level of the candidate key feature node based on the attention weight and the causal effect value, so as to generate a disease profile according to the association level.

[0214] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the disease profile generation method provided in the above embodiments. The method includes: constructing a feature association map based on multi-source data related to a target disease, the feature association map including feature nodes characterizing each feature in the multi-source data; inputting the feature association map into a preset hierarchical graph attention network to obtain an attention weight for each feature node determined by the hierarchical graph attention network; selecting at least one candidate key feature node from all the feature nodes based on the attention weight; performing causal effect estimation on the candidate key feature node to obtain a causal effect value characterizing the causal strength between the candidate key feature node and the target disease; and determining the association level of the candidate key feature node based on the attention weight and the causal effect value, so as to generate a disease profile based on the association level.

[0215] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0216] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0217] 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 of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for generating disease profiles, characterized in that, include: A feature association map is constructed based on multi-source data related to the target disease, and the feature association map includes feature nodes that characterize each feature in the multi-source data; The feature association graph is input into a preset hierarchical graph attention network to obtain the attention weight of each feature node determined by the hierarchical graph attention network. At least one candidate key feature node is selected from all the feature nodes based on the attention weight; Causal effect estimation is performed on the candidate key feature nodes to obtain causal effect values ​​that characterize the causal strength between the candidate key feature nodes and the target disease; Based on the attention weight and the causal effect value, the association level of the candidate key feature nodes is determined, so as to generate a disease profile according to the association level.

2. The disease profile generation method according to claim 1, characterized in that, The hierarchical graph attention network includes a first network layer and a second network layer; The step of obtaining the attention weights of each feature node determined by the hierarchical graph attention network includes: The first network layer aggregates information about neighboring nodes determined by the edges of the feature association graph to generate a first-order feature representation that characterizes the direct association information between feature nodes. The first-order feature representations are aggregated using the second network layer to generate second-order feature representations that characterize the combination relationships between feature nodes; The attention coefficients generated during the generation of the first-order feature representation and / or the second-order feature representation are extracted to obtain the attention weight of each feature node.

3. The disease profile generation method according to claim 1, characterized in that, The step of estimating the causal effect of the candidate key feature nodes to obtain a causal effect value characterizing the causal strength between the candidate key feature nodes and the target disease includes: Target sample data for causal verification is extracted from the multi-source data. The target sample data includes the feature values ​​of multiple sample objects on the candidate key feature nodes and the true label values ​​on the target disease. Obtain pre-trained propensity score model and outcome regression model; The target sample data is input into the propensity score model, and the propensity score of each sample object on the candidate key feature node is output. The target sample data is input into the result regression model, and the predicted probability of each sample object having the target disease is output. The average processing effect of the candidate key feature nodes is calculated based on the propensity score, the predicted probability of disease, the feature value, and the true label value. The average treatment effect is determined as the causal effect value between the candidate key feature node and the target disease.

4. The disease profile generation method according to claim 3, characterized in that, The step of calculating the average processing effect of the candidate key feature nodes based on the propensity score, the predicted disease probability value, the feature value, and the true label value includes: Sample weights are constructed using the propensity score and the feature values; Calculate the prediction residual between the true label value and the predicted disease probability value; The predicted residuals are weighted based on the sample weights, and the weighted residuals are fused with the predicted disease probability to obtain the causal effect estimate. The average treatment effect is obtained by averaging the estimated causal effects of all sample objects.

5. The disease profile generation method according to claim 2, characterized in that, The step of using the first network layer to aggregate information of neighboring nodes determined by edges based on the feature association graph to generate a first-order feature representation characterizing the direct association information between feature nodes includes: For any target feature node in the feature association graph, obtain the initial feature vectors of the target feature node and its neighboring nodes; The initial eigenvector is linearly transformed using a linear transformation matrix to obtain the transformed eigenvector; Based on the transformed feature vector, the correlation coefficient between the target feature node and each neighbor node is calculated; The correlation coefficients are normalized using an activation function to obtain normalized attention coefficients. The transformed feature vectors of the neighboring nodes are weighted and aggregated using the normalized attention coefficients to obtain the first-order feature representation of the target feature node.

6. The disease profile generation method according to claim 2, characterized in that, The step of aggregating the first-order feature representations using the second network layer to generate second-order feature representations characterizing the combination relationships between feature nodes includes: For the connected feature node pairs, perform element-wise product operation or concatenation operation on the first-order feature representations of the feature node pairs to generate feature interaction vectors; The importance score of the feature interaction vector in representing the feature combination effect is calculated using a multi-head attention mechanism; The feature interaction vectors are aggregated based on the importance scores to generate the second-order feature representation that characterizes the meta-features formed by the combined interactions between the feature nodes.

7. The disease profile generation method according to claim 2, characterized in that, The extraction of attention coefficients generated during the generation of the first-order feature representation and / or the second-order feature representation, to obtain the attention weight of each feature node, includes: Obtain the attention coefficient matrix output by the multi-head attention mechanism in the first network layer and / or the second network layer in the hierarchical graph attention network; For any feature node, sum or average all attention coefficients related to that feature node in the attention coefficient matrix to obtain a comprehensive importance score; The comprehensive importance score is determined as the attention weight of any feature node.

8. The disease profile generation method according to claim 1, characterized in that, The step of selecting at least one candidate key feature node from all the feature nodes based on the attention weight includes: Sort all the feature nodes from high to low according to the attention weight; The top K feature nodes in the sorting results are selected as the candidate key feature nodes, where K is a preset positive integer; Alternatively, feature nodes whose attention weight is greater than a preset weight threshold can be selected as candidate key feature nodes.

9. The disease profile generation method according to claim 1, characterized in that, Determining the association level of the candidate key feature nodes based on the attention weight and the causal effect value includes: If the attention weight is greater than a first preset threshold and the causal effect value is greater than a second preset threshold, the association level of the candidate key feature node is determined to be strongly correlated. If the attention weight is greater than the first preset threshold but the causal effect value is less than or equal to the second preset threshold, the association level of the candidate key feature node is determined to be pseudo-correlation or weak correlation. If the attention weight is less than or equal to the first preset threshold but the causal effect value is greater than the second preset threshold, the association level of the candidate key feature node is determined to be potentially relevant.

10. The disease profile generation method according to claim 1, characterized in that, The construction of a feature association map based on multi-source data related to the target disease includes: The statistical correlation between any two features in the multi-source data is calculated using statistical methods. If the statistical correlation meets the preset conditions, an edge is established between the feature nodes corresponding to the two features, and the value of the statistical correlation is used as the initial weight of the edge.

11. The disease profile generation method according to claim 3, characterized in that, Also includes: In response to the acquisition of new multi-source data, the feature association map is updated using the new multi-source data; Based on the updated feature association graph, the hierarchical graph attention network, the propensity scoring model, and the result regression model are incrementally trained. The attention weights are re-determined using the trained hierarchical graph attention network, and the causal effect is re-estimated using the trained propensity score model and outcome regression model to generate the updated disease profile based on the newly added multi-source data.

12. The disease profile generation method according to claim 2, characterized in that, The step of generating a disease profile based on the association level includes: Generate a visual knowledge graph or multi-dimensional radar chart containing the candidate key feature nodes; In the visualized knowledge graph, the size or color of the nodes represents the magnitude of the attention weight, and the thickness or type of the lines connecting the nodes represents the strength of the causal effect value; or, in the disease profile, the text description related to the combination relationship between the feature nodes corresponding to the second-order feature representation is explicitly output.

13. A disease profile generation device, characterized in that, include: The graph construction module is used to construct a feature association graph based on multi-source data related to the target disease. The feature association graph includes feature nodes that characterize each feature in the multi-source data. The weight acquisition module is used to input the feature association map into a preset hierarchical graph attention network and acquire the attention weight of each feature node determined by the hierarchical graph attention network. A feature filtering module is used to filter at least one candidate key feature node from all the feature nodes according to the attention weight; The causal estimation module is used to estimate the causal effect of the candidate key feature nodes to obtain a causal effect value that characterizes the causal strength between the candidate key feature nodes and the target disease. The profile generation module is used to determine the association level of the candidate key feature nodes based on the attention weight and the causal effect value, so as to generate a disease profile according to the association level.

14. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the disease profile generation method as described in any one of claims 1 to 12.

15. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the disease profile generation method as described in any one of claims 1 to 12.

Images