A data processing method for diabetes risk stratification aided assessment

Abstract

The application discloses a data processing method for diabetes risk stratification auxiliary evaluation, and belongs to the technical field of medical health data processing, machine learning, differentiable directed acyclic graph structure learning and graph neural network. The method solves the problems of label leakage, candidate directed structure instability and insufficient structure explanation in multi-source health data risk evaluation. The method obtains and pre-processes multi-source health sample data, constructs a diabetes phenotype label PHENO, and removes label source variables and label leakage variables. Under the conditions of a PHENO convergence node, a background prior variable and a self-loop prohibition constraint, differentiable directed acyclic graph structure learning is performed, and a stable candidate statistical correlation directed graph is formed by combining structure perception scoring and multiple stability screening. Then, a directed graph neural network is trained based on the graph, and risk scores, risk level probabilities and structure explanation data are outputted without inputting the real values of the PHENO. The method is mainly used for data processing of diabetes risk stratification auxiliary evaluation.

Description

Technical Field

[0001] This invention relates to the fields of medical and health data processing, machine learning, differentiable directed acyclic graph structure learning, and graph neural network technology, and particularly to a data processing method, computer device, and computer-readable storage medium for auxiliary assessment of diabetes risk stratification executed by a computer device. Background Technology

[0002] Risk stratification assessment for diabetes typically requires the integration of multi-source health data, including demographic information, questionnaire data, physical examination data, laboratory test data, dietary and nutritional information, and lifestyle information. Existing risk assessment methods often employ models such as logistic regression, random forests, support vector machines, gradient boosting trees, or multilayer perceptrons to directly predict and model candidate features. While these methods can output risk scores or classification probabilities, they generally suffer from the following shortcomings: First, directly using diagnostic, medication, or testing variables used to construct diabetes phenotype labels as predictive features can easily lead to label leakage, causing the model performance to be overestimated. Second, traditional feature selection methods are mostly based on correlation, mutual information, or model importance, making it difficult to express the directed structural relationships between variables oriented towards the target phenotype. Third, ordinary predictive models often cannot stably provide directly associated features and their upstream risk transmission paths, resulting in insufficient structural interpretability. Fourth, when learning directed graph structures from observed health data, single learning results are easily affected by random initialization, sample partitioning, missing data imputation, and noise perturbations, leading to insufficient edge direction and edge stability.

[0003] Differentiable directed acyclic graph (DAG) structure learning methods can transform directed graph structure learning into a continuous optimization problem, learning candidate directed structures from observed data through acyclicity constraints, sparsity constraints, and structural equation fitting. However, evaluating candidate graph structures solely based on data fitting errors may ignore the candidate graph's own acyclicity, sparsity, input / output edge distribution, and state information as the candidate structure changes during training, leading to problems such as pseudo-edges, unstable orientations, or local optima in the learned candidate structures. Therefore, a computer data processing scheme is needed that, while preventing label leakage, combines prior edge constraints, structure-aware scoring, stability screening, and directed graph neural network prediction to output risk-stratified auxiliary assessment data and corresponding structural interpretation data.

[0004] It should be noted that the directed edges learned from observed health data mainly represent candidate directed association structures obtained under the conditions of a preset set of variables, preset prior constraints, and statistical learning. They are not equivalent to medical causal mechanisms, nor are they used to directly make disease diagnoses, determine treatment plans, or make medication decisions. Summary of the Invention

[0005] I. Technical problems to be solved

[0006] This invention aims to solve the following technical problems: how to construct a diabetes phenotype label PHENO from multi-source healthy sample data, remove label source variables and label leakage variables, and form a candidate feature matrix without label leakage; how to learn candidate statistical association directed graphs under PHENO convergence node constraints, background prior variable constraints, and self-loop prohibition constraints; how to make the candidate graph structure learning process not only dependent on reconstruction errors through dynamic structure-aware mapping networks and structure-aware scoring items; how to extract stable candidate direct association features and candidate risk transmission paths through repeated learning and stability screening; and how to use stable candidate statistical association directed graphs for directed graph neural network training and risk stratification auxiliary assessment, while avoiding inputting the true value of PHENO in the prediction input stage.

[0007] II. Technical Solution

[0008] To address the aforementioned technical problems, this invention provides a data processing method executed by a computer device for auxiliary assessment of diabetes risk stratification. The method, consistent with the step numbers in the claims, includes the following steps:

[0009] S1: Obtain existing multi-source health sample data, preprocess the multi-source health sample data, and construct the diabetes phenotype label PHENO according to the preset phenotype construction rules;

[0010] S2: Remove the label source variables, label leakage variables, and process record variables that are not used as model prediction inputs related to the construction of PHENO to form a candidate feature matrix;

[0011] S3: Construct a graph learning variable set based on the candidate feature variables and corresponding PHENO variables in the candidate feature matrix, and generate an adjacency matrix mask according to preset edge constraints, wherein the preset edge constraints include at least PHENO node convergence constraints, self-loop prohibition constraints, and background prior variable constraints.

[0012] S4: Under the adjacency matrix mask constraint, differentiable directed acyclic graph structure learning is performed on the graph learning variable set to obtain candidate statistical association directed graphs, wherein the objective function of differentiable directed acyclic graph structure learning includes a data fitting loss term, a structure constraint term, and a structure-aware scoring term.

[0013] S5: Perform the differentiable directed acyclic graph structure learning multiple times, and determine the stability index of candidate directed edges based on the results of multiple learnings, screen stable candidate directed edges, and form a stable candidate statistical association directed graph;

[0014] S6: Using the stable candidate statistical association directed graph as the adjacency structure of the directed graph neural network, the directed graph neural network is trained, wherein the PHENO label is used as the supervised training target, and the PHENO node does not receive the actual PHENO value as the node signal input.

[0015] S7: Based on the trained directed graph neural network, perform directed message passing on the object to be evaluated, generate model output data, and generate structural interpretation data corresponding to the model output data based on the stable candidate statistical association directed graph;

[0016] The model output data is used to characterize the risk score, risk level representation value, or risk level corresponding probability of the subject to be evaluated in the preset risk stratification model; the data processing method only performs computer data processing on the acquired data, does not include the steps of conducting detection, sampling, diagnosis, treatment, or medication decisions on human or animal bodies, and does not directly make a diabetes diagnosis conclusion based on the model output data.

[0017] III. Beneficial Effects

[0018] Compared with the prior art, the present invention has at least the following beneficial effects:

[0019] (i) This invention removes the tag source variables, direct mapping variables, homologous derived variables and process record variables involved in the construction of PHENO after constructing the diabetes phenotype tag, which can reduce the problem of inflated risk prediction results caused by tag leakage.

[0020] (ii) This invention sets PHENO as a convergence node that only receives directed edges of other variables, and sets demographic variables as background prior variables, which can transform the prior edge constraints in the risk assessment task into an adjacency matrix mask, thereby improving the directional rationality of the candidate graph structure.

[0021] (III) This invention optimizes the candidate edge weight matrix by combining data fitting loss, acyclicity constraint term, sparse constraint term and structure-aware scoring term. It can introduce the acyclicity, sparsity and in-and-out edge states of the candidate graph structure itself into the scoring process, which helps to reduce the problems of pseudo edges, directional instability and local optima caused by relying solely on data fitting.

[0022] (iv) This invention obtains multiple candidate directed adjacency matrices based on repeated learning, and selects stable candidate directed edges according to edge stability, edge weight representation and direction consistency, which can reduce the impact of random seeds, sample perturbation or training subset changes on graph structure.

[0023] (v) This invention uses a stable candidate statistical association directed graph as the adjacency structure of a directed graph neural network, retains the PHENO target node but does not input the actual PHENO value, so that the model can aggregate upstream variable information along the candidate risk transmission path, while avoiding label leakage in the prediction stage.

[0024] (vi) The present invention outputs stable candidate direct correlation features and candidate risk transmission paths at the same time as the output model outputs data, so that the risk stratification auxiliary assessment results have structural interpretation data.

[0025] (vii) This invention only performs computer data processing on the acquired data, and does not include the steps of detection, sampling, diagnosis, treatment or medication decision-making on human or animal bodies, and does not make a diabetes diagnosis conclusion directly based on the model output data. Attached Figure Description

[0026] The accompanying drawings are used to further illustrate the technical solution of the present invention, wherein:

[0027] Figure 1 This is a schematic diagram of the overall flow of the data processing method provided in an embodiment of the present invention;

[0028] Figure 2 This invention provides a directed graph of candidate statistical associations and a schematic diagram of simulated candidate risk transmission paths for embodiments of the invention.

[0029] It should be noted that the above figures are used to illustrate the process and module relationship of the present invention and do not constitute a limitation on the scope of protection of the present invention. Detailed Implementation

[0030] The present invention will be further described below with reference to embodiments. It should be understood that the following embodiments are only used to explain the present invention and are not intended to limit the scope of protection of the present invention. Where there is no conflict, the technical features in the following embodiments can be combined with each other.

[0031] I. General Definition

[0032] In this specification, multi-source health sample data refers to acquired health-related data, which may include one or more of the following: demographic data, questionnaire data, physical examination data, laboratory test data, dietary and nutritional data, and lifestyle data. The candidate characteristic values ​​of the subject to be evaluated are also acquired data. This invention does not involve operations that involve testing, sampling, diagnosing, or treating human or animal bodies.

[0033] In this specification, PHENO represents the diabetes phenotype label constructed according to the preset phenotype construction rules; V represents the graph learning variable set; p represents the number of graph learning variables; X represents the matrix of variables to be learned; n represents the number of samples; M represents the adjacency matrix mask; and W represents the original learnable directed edge weight matrix. A represents the candidate edge weight matrix after masking constraints; A represents the directed candidate adjacency matrix; A ij =1 indicates that there exists a variable v i Pointer to variable v j Candidate directed edges; A ij =0 indicates that the variable v does not exist or is not retained. i Pointer to variable v j Candidate directed edges.

[0034] The graph learning variable set V is represented as: V = {v1, v2, ..., v...} p}

[0035] The matrix of variables to be learned, X, is represented as: X∈R n×p

[0036] Here, X includes the data columns corresponding to the candidate feature variables and the data columns corresponding to PHENO. It should be noted that the candidate feature matrix is ​​not separately denoted by other symbols in the specification; the candidate feature matrix, as the data matrix formed in S2, is used to form the learning variable matrix X in the graph structure learning stage after the graph learning variable set V is constructed together with the corresponding variables of PHENO in S3.

[0037] In this specification, the candidate statistical association directed graph represents the candidate directed association structure obtained under the conditions of a preset set of variables, preset edge constraints, a learning objective function for a differentiable directed acyclic graph structure, and statistical stability screening. This graph is used to express the candidate direct association characteristics and candidate risk transmission paths in risk stratification-assisted assessment, and is not equivalent to a medical causal mechanism, nor does it constitute a conclusion on the mechanism of disease occurrence.

[0038] II. Data Acquisition, Preprocessing, and PHENO Construction

[0039] In one embodiment, a computer device acquires existing multi-source health sample data. This multi-source health sample data may include one or more of the following: demographic data, questionnaire data, physical examination data, laboratory test data, dietary and nutritional data, and lifestyle data. The computer device performs variable name unification, variable unit conversion, variable value range verification, sample-level merging, and duplicate sample processing on the data from different sources to obtain an original merged data table.

[0040] In one embodiment, a computer device performs missing value processing, outlier handling, categorical variable encoding, continuous variable standardization, and sample screening on the original merged data table to obtain a preprocessed health characteristic data table. Missing value processing may include median imputation, mode imputation, model imputation, or missing indicator variable construction; outlier handling may include truncation, removal, or setting as missing values ​​based on a preset medically reasonable range; categorical variable encoding may include binary encoding, one-hot encoding, or ordered encoding; continuous variable standardization may include Z-score standardization or quantile normalization.

[0041] In one embodiment, the preset phenotype construction rules consist of one or more of the following: prior diabetes diagnosis information, information on hypoglycemic drug use, fasting blood glucose level, glycated hemoglobin level, oral glucose tolerance test level, or random blood glucose level. For the i-th sample, PHENO can be constructed as follows:

[0042]

[0043] Among them, PHENO i This represents the diabetes phenotype label for the i-th sample. Med represents the previous diabetes diagnosis information for the i-th sample. i Information on the use of hypoglycemic drugs, FPG i HbA1c represents the fasting blood glucose level. i OGTT indicates the glycated hemoglobin level. i RBG represents an indicator of oral glucose tolerance test. i Gamma represents a random blood glucose level. FPG γ HbA1c γ OGTT and γ RBG These represent the corresponding preset thresholds. All the above indicators are historical data or existing detection data acquired before implementing this method. The computer equipment generates labels based solely on this existing data and does not include any newly implemented detection, sampling, or diagnostic steps.

[0044] III. Removal of Tag Source Variables and Tag Leakage Variables

[0045] After constructing PHENO, the computer device determines the tag source variables participating in the PHENO construction according to preset phenotypic construction rules. The tag source variables include one or more of the following: prior diabetes diagnosis information variables, hypoglycemic drug use information variables, fasting blood glucose index variables, glycated hemoglobin index variables, oral glucose tolerance test index variables, or random blood glucose index variables used to construct PHENO.

[0046] The computer equipment also removes label leakage variables according to preset label leakage judgment rules. The label leakage variables include variables directly mapped from the label source variables, variables belonging to the same field as the label source variables, derived variables belonging to the same test item as the label source variables, derived variables belonging to the same questionnaire item as the label source variables, derived variables belonging to the same drug use record as the label source variables, and variables that can be transformed from the label source variables through deterministic rules.

[0047] The computer equipment further eliminates process record variables that are not used as inputs for model prediction. These process record variables include object unique identifier variables, serial number variables, sample number variables, visit number variables, data release date variables, sampling date variables, inspection date variables, or other process record variables that are not used as inputs for model prediction.

[0048] After the above elimination process, the computer device will form a candidate feature matrix without label leakage from the remaining candidate health features. The candidate feature matrix does not contain the actual values ​​of PHENO, the label source variables that participated in the construction of PHENO, or the label leakage variables that have a direct mapping, homology, or deterministic transformation relationship with the label source variables.

[0049] IV. Construction of Graph Learning Variable Sets and Generation of Adjacency Matrix Masks

[0050] In one embodiment, the computer device constructs a graph learning variable set V based on the candidate feature variables in the candidate feature matrix and the corresponding PHENO variables:

[0051] V = {v1, v2, ..., v} p}

[0052] Where p represents the number of graph learning variables, and the graph learning variable set V includes candidate feature variable nodes and PHENO nodes. PHENO nodes are set as pool nodes, so that PHENO nodes can only be the target nodes of candidate directed edges, and PHENO nodes are prohibited from being the source nodes of candidate directed edges pointing to any candidate feature variable node.

[0053] In one embodiment, the computer device sets one or more of the following demographic variables—age, gender, race, education level, income level, or others—that are determined before the risk stratification-assisted assessment and do not change with the current health data, as background prior variables, forming a background prior variable set B. These background prior variables represent variables that are determined before the risk stratification-assisted assessment and are not inversely determined by physical examination, laboratory tests, dietary nutrition, or lifestyle variables in the current health sample data. To avoid generating candidate directed edges such as "gender → age" or "income level → race" that do not conform to the meaning of the background prior variables, the background prior variables only serve as source nodes for candidate directed edges and not as target nodes for any candidate directed edge. Accordingly, any candidate directed edge pointing from any graph learning variable to a background prior variable is prohibited; the background prior variables are allowed to point to other graph learning variables not constrained by PHENO convergence node constraints, self-loop prohibition constraints, or other edge prohibition rules.

[0054] The computer device generates an adjacency matrix mask M based on the PHENO convergence node constraints, background prior variable constraints, and self-loop prohibition constraints.

[0055] M = [M ij ] p×p

[0056] Among them, M ij This indicates that variable v is learned from the i-th graph. i Learning variable v from the j-th graph j Whether candidate directed edges are allowed, M ij =0 indicates that v is prohibited i Point to v j Candidate directed edges, M ij =1 means that v is allowed i Point to v j Candidate directed edges.

[0057] The adjacency matrix mask M is determined according to the following rules:

[0058] When i = j, M ij =0, used to disable self-loops in graph learning variables;

[0059] When v i For PHENO node and v j When not a PHENO node, M ij =0, used to prevent the PHENO node from pointing to the candidate feature variable node;

[0060] When v j When ∈B, M ij =0, used to prevent any graph learning variable from pointing to a background prior variable;

[0061] For candidate directed edges not prohibited by the self-loop prohibition constraint, PHENO convergence node constraint, and background prior variable constraint, M ij =1.

[0062] The resulting adjacency matrix mask transforms the PHENO pooling node constraint, background prior variable constraint, and self-loop prohibition constraint into edge feasibility constraints during the differentiable graph structure learning process. Since the background prior variables serve only as source nodes for candidate directed edges and not as target nodes, the adjacency matrix mask prevents background prior variables from being pointed to in the opposite direction by other candidate feature variables, and also prevents multiple background prior variables from forming mutually pointing relationships that do not conform to the preset prior meaning. PHENO is only used during the training phase for constructing target phenotypic nodes in graph structure learning, and is not used as input for actual values ​​during the risk stratification-assisted assessment data processing phase of the object to be evaluated.

[0063] V. Differentiable Directed Acyclic Graph Structure Learning Based on Structure-Aware Scoring

[0064] In one embodiment, the computer device constructs a matrix of variables to be learned, X ∈ R, based on a graph learning variable set V. n×p Where n represents the number of samples, p represents the number of graph learning variables, and X includes the data columns corresponding to the candidate feature variables and the data columns corresponding to PHENO. The i-th sample vector in X is denoted as x. i Where i = 1, 2, ..., n. The computer device constructs the original learnable directed edge weight matrix W, W ∈ R. pxp Then, the adjacency matrix mask M is multiplied element-wise with W to obtain the candidate edge weight matrix constrained by the mask.

[0065]

[0066] Here, ⊙ denotes element-wise multiplication. For M ij Candidate directed edges with a value of 0, It is set to zero during optimization; for M ij Candidate directed edges with a value of 1, Participation in optimization is permitted. Thus, the PHENO convergence node constraint, background prior variable constraint, and self-loop prohibition constraint are transformed into edge feasibility constraints during the learning process of a differentiable directed acyclic graph structure.

[0067] The computer device is based on the candidate edge weight matrix after masking constraints. Construct a structural equation fitting model f θ And use structural equation modeling to fit model f θ The matrix of variables to be learned, X, is reconstructed to obtain the reconstructed variable matrix.

[0068]

[0069] Where θ represents the structural equation fitting model f θ The learnable parameters. f θ This can be a multilayer perceptron, a variable-level local connection network, a nonlinear structural equation fitting network, a graph neural network, or a combination thereof. Reconstructing the variable matrix. The i-th reconstructed sample vector in the dataset is denoted as . Where i = 1, 2, ..., n.

[0070] The computer device uses the learning variable matrix X and the reconstructed variable matrix as a basis. The difference between them determines the data fitting loss L. fit :

[0071]

[0072] Where Loss(·) represents the squared loss, negative log-likelihood loss, reconstruction error loss, or a combination thereof. In a specific embodiment, the squared reconstruction error can be expressed as follows:

[0073]

[0074] Among them, ||·|| F This represents the Frobenius norm.

[0075] The computer device uses the candidate edge weight matrix after masking constraints. Determine the acyclic constraint term

[0076]

[0077] Where tr(·) represents the matrix trace, exp(·) represents the matrix exponent, and p represents the number of graph learning variables. Used to characterize by The degree to which a given candidate directed graph violates the directed acyclic structure.

[0078] To ensure that candidate graph structure learning depends not only on the learnable variable matrix X and the reconstructed variable matrix The computer device further constructs a dynamic structure-aware mapping network to address the data fitting error between the two sets of data. and structure-aware descriptors in, Represents a dynamic structure-aware mapping network Learnable parameters, structure-aware descriptors Candidate edge weight matrix after masking The calculated structure-aware descriptor is used to characterize the candidate graph's structural information. It can include one or more of the following: acyclicity descriptor, edge sparsity descriptor, incoming edge strength vector, outgoing edge strength vector, absolute value statistics of candidate edge weights, or adjacency structure embedding vector.

[0079] In one embodiment, the incoming edge strength vector and the outgoing edge strength vector can be represented as:

[0080]

[0081] in, This represents the strength of the incoming edges of the j-th graph learning variable. This represents the outgoing edge strength of the i-th graph learning variable.

[0082] In one embodiment, a structure-aware descriptor It can be represented as:

[0083]

[0084] Wherein, Concat(·) represents the concatenation operation. Indicates the acyclicity descriptor. r represents the sparsity descriptor of the candidate edge weight matrix. in Let r represent the incoming edge strength vector. out Represent the edge strength vector. A statistic representing the absolute value of candidate edge weights, wherein the statistic may be one or more of the following: mean, variance, quantile, or maximum value.

[0085] Dynamic Structure-Aware Mapping Network This is used for joint mapping of sample vectors and structure-aware descriptors. Specifically, for the i-th sample, the computer device will map the i-th sample vector x from the variable matrix X to be learned. i With structure-aware descriptors After splicing or joint encoding, the data is input into the dynamic structure-aware mapping network. Obtain the first structure-aware embedding z i ; Reconstruct the variable matrix The i-th reconstructed sample vector in With structure-aware descriptors After splicing or joint encoding, the data is input into the dynamic structure-aware mapping network. Obtain the second structure-aware embedding

[0086]

[0087] Where i = 1, 2, ..., n, and [·] represents concatenation or joint encoding operations.

[0088] The first structure-aware embedding set is represented as: The second structure-aware embedding set is represented as: The computer device determines the structure-aware scoring item L based on the first structure-aware embedding set and the second structure-aware embedding set. s The structure-aware scoring item L s Used to characterize the matrix of variables to be learned, X, and the reconstructed matrix of variables. Differences in batch distribution in the structure-aware space.

[0089] In one embodiment, when the structure-aware scoring item L s When calculating the batch mean difference, it is determined according to the following formula:

[0090]

[0091] In another embodiment, when the structure-aware scoring item L s For the maximum mean difference, it is determined according to the following formula:

[0092]

[0093] Among them, MMD 2 This represents the maximum mean difference between the first and second structure-aware embedding sets. Through the above sample-by-sample mapping method... It does not act abstractly on the entire matrix, but rather on each sample vector x. i and corresponding reconstructed sample vector Perform structure-aware mapping separately; determine L by batch mean difference or maximum mean difference. s This clarifies the calculation object and calculation method of the structure perception scoring item.

[0094] In one specific embodiment, a dynamic structure-aware mapping network It can include an input splicing layer, two or three fully connected layers, a normalization layer, and a non-linear activation layer. For example, The input is a sample vector and a structure-aware descriptor. The network structure can be a combination of fully connected layers, normalization layers, activation layers, fully connected layers, activation layers, and output layers. Alternatively, variable-level local connection layers can be used, where each variable's corresponding mapping branch receives information from a preset variable block or a set of corresponding parent nodes.

[0095] Dynamic Structure-Aware Mapping Network In the process of learning graph structures, pre-training and alternating optimization are used for updates.

[0096] During the pre-training phase, the computer device fits the structural equation model f. θ and dynamic structure-aware mapping network Perform initial training to enable Able to initially distinguish between the learning variable matrix X and the reconstructed variable matrix Differences in the structure-perceived space.

[0097] The alternating optimization phase includes a first sub-step and a second sub-step.

[0098] In the first sub-step, the computer device is fixed to a dynamic structure-aware mapping network. And update the structural equation fitting model f θ And the original learnable directed edge weight matrix W, to minimize the data fitting loss L fit Structural perception scoring item L s sparse constraints and acyclic constraints

[0099] In the second sub-step, the computer equipment fixed structure equation fitting model f θ and the candidate edge weight matrix after masking To enhance the first structure-aware embedding set With the second structure-aware embedding set The distinguishability between them is used to update the parameters of the dynamic structure-aware mapping network for the target. The parameters of the dynamic structure-aware mapping network The update target can be expressed as:

[0100]

[0101] Therefore, dynamic structure-aware mapping network The update is no longer just about abstractly determining the measure of distributional difference, but rather by fixing... and θ, in the constrained set of parameters Internal enhancement and Distinguishability in the structure-aware space, thus enabling the structure-aware scoring item L s It can stably and repeatedly participate in the optimization of candidate edge weight matrix.

[0102] To prevent dynamic structure-aware mapping networks Degenerates into a structure-aware scoring item L s A failed mapping function, which a computer device can... Apply one or more of the following constraints: L2 regularization is applied to the weights; Clipping the gradient; Bounded clipping of parameters; Normalize the output; normalize the output of the product; The calculated structure-aware descriptor Using stopping gradient processing; for Employ spectral normalization or Lipschitz constraints; in updating The weight range is limited at that time.

[0103] In one embodiment, the parameter constraint set It can be represented as:

[0104]

[0105] in, This indicates the upper limit of the preset parameter norm. Represents a dynamic structure-aware mapping network The Lipschitz constant, This indicates the upper limit of the preset Lipschitz constant.

[0106] In another embodiment, to reduce computational complexity, the dynamic structure-aware mapping network can be frozen after pre-training. In the subsequent graph structure learning process, only the structure equation fitting model f is updated. θ And the original learnable directed edge weight matrix W. The freezing method is suitable for scenarios with a small sample size or requiring high training stability; the alternating optimization method is suitable for scenarios that need to dynamically adjust the scoring space according to changes in the candidate graph structure.

[0107] The computer device constructs an objective function for optimizing W and θ based on a data fitting loss term, a structure-aware scoring term, acyclic constraint term, and sparse constraint term:

[0108]

[0109] in, Represents the candidate edge weight matrix under mask constraints The sparse constraint term, where λ1 represents the sparse constraint weight, λ2 represents the structure-aware scoring term weight, and λ θ α represents the regularization weights of the structural equation fitting model, and ρ represents the augmented Lagrange coefficients.

[0110] During the optimization process, the computer equipment uses acyclic constraints. The changes adjust the augmented Lagrange coefficients α and ρ to promote the change from The determined candidate directed graphs satisfy the directed acyclic structure.

[0111] After optimization, the computer device processes the candidate edge weight matrix after masking constraints. Threshold pruning and sparsification are performed to obtain the directed candidate adjacency matrix A:

[0112]

[0113] Where, τ w This represents a preset edge weight threshold. The computer device forms a directed statistical association graph based on the directed candidate adjacency matrix A.

[0114] The structure-aware scoring item L s This is used to incorporate candidate graph structure information or acyclicity information into the graph structure learning scoring process, so that the learning of candidate statistical association directed graphs does not depend solely on the learnable variable matrix X and the reconstructed variable matrix. The data fitting error between them. By updating Time fixed and θ, and to enhance X with Distinguishability in structure-aware space is used for target updating. Simultaneously fix W and θ when updating It can make the structure perception scoring item L s The graph structure learning process has a clear training objective and a constrained parameter space, thereby reducing pseudo edges, unstable orientations, or local optimal candidate graph structures caused by simple data fitting.

[0115] VI. Extraction of Stable Candidate Structures

[0116] In one embodiment, the computer device acquires R repeated learning tasks, wherein each repeated learning task is formed based on one or more of different random seeds, different data resampling subsets, or different training subsets, where R is an integer greater than 1. The repeated learning tasks are used to avoid fixing the number of training subsets to the number of repetitions, allowing a single repeated learning task to correspond to one training subset, a random initialization process, a resampling subset, or a combination of a random seed, a resampling subset, and a training subset.

[0117] The computer device executes the R repeated learning tasks respectively, obtaining R candidate directed adjacency matrices and corresponding R masked candidate edge weight matrices:

[0118] A 1 A 2 A R

[0119]

[0120] Among them, A r This represents the candidate directed adjacency matrix obtained from the r-th repeated learning task. Let M represent the masked candidate edge weight matrix obtained from the r-th repeated learning task, where r = 1, 2, ..., R. Each repeated learning task is performed under the constraint of the adjacency matrix mask M, and satisfies the PHENO pool node constraint, background prior variable constraint, and self-loop prohibition constraint.

[0121] For any candidate directed edge v i →v j The computer device counts the number of times it appears in R candidate directed adjacency matrices, C. ij :

[0122]

[0123] The candidate directed edge v is determined according to the following formula. i →v j Edge stability S ij :

[0124]

[0125] Where 1(·) represents the indicator function, This indicates that in the r-th repeated learning task, the value of v is retained. i Point to v j Candidate directed edges; A r ij =0 indicates that the learning task in the r-th iteration did not retain the data generated by v. i Point to v j Candidate directed edges.

[0126] The computer device determines the edge weight representation B of the candidate directed edge based on the average absolute value of the edge weights in R masked candidate edge weight matrices. ij :

[0127]

[0128] in, This represents the masked candidate edge weight matrix obtained by learning the r-th differentiable directed acyclic graph structure. Zhongyou v i Point to v j The edge weights are represented by Norm(·), which indicates one of the following: min-max normalization, quantile normalization, or normalization according to the set of allowed edges.

[0129] The computer device compares candidate directed edges in opposite directions between pairs of the same variable, and determines the directional consistency D of the candidate directed edges based on the number of occurrences of the first candidate directed edge and the number of occurrences of the opposite candidate directed edge. ij :

[0130]

[0131] Where ε is a preset constant used to avoid the denominator being zero.

[0132] The computer device determines the edge score of candidate directed edges based on edge stability, edge weight representation, and direction consistency. ij :

[0133] Score ij =β1S ij +β2B ij +β3D ij

[0134] Wherein, β1, β2 and β3 are preset weight coefficients, β1, β2, β3>0 and β1+β2+β3=1.

[0135] Within a candidate edge set satisfying the PHENO convergence node constraint, background prior variable constraint, self-loop prohibition constraint, and acyclicity constraint, the computer equipment retains candidate directed edges whose edge scores are greater than a preset edge score threshold, or candidate directed edges whose edge scores rank within a preset proportion, as the candidate stable edge set. The candidate stable edge set can be represented as C:

[0136] C={v i →v j Score ij ≥τ s And M ij =1}

[0137] Where, τ s This represents a preset edge score threshold. Alternatively, it can be used to form a candidate stable edge set by ranking the candidate directed edges by their scores within a preset proportion and satisfying the adjacency matrix mask constraint.

[0138] To ensure that the stable candidate statistical association directed graph formed after multiple repeated learning iterations still satisfies the directed acyclic constraint, the computer device further performs acyclic filtering when selecting stable candidate directed edges. Specifically, the computer device initializes the stable candidate edge set E. stable :

[0139]

[0140] Computer equipment according to Score ij Sort the candidate directed edges in the candidate stable edge set C from high to low, and then iterate through the candidate directed edges v in turn. i →v j For the currently traversed candidate directed edge, the computer device determines whether to add the candidate directed edge to E. stable The question asks whether a directed cycle has formed. This determination can be achieved through topological sorting, depth-first search, or reachability matrix updates. If adding the current candidate directed edge does not form a directed cycle, then the candidate directed edge is retained in E. stable If adding the current candidate directed edge results in a directed cycle, then skip that candidate directed edge.

[0141] The above process can be expressed as: for according to Score ij Each candidate directed edge v in descending order i →v j If Graph(E) stable ∪v i →v j If E does not contain directed cycles, then E stable =E stable ∪v i →v j Otherwise, keep E stable The edge remains unchanged, and the process continues to iterate over the next candidate directed edge.

[0142] After the traversal is complete, the computer device is based on E stable Construct a stable candidate statistical association directed graph G stable Since a directed cycle check is performed before each candidate directed edge is added, the final stable candidate statistically associated directed graph G is obtained. stable It satisfies the PHENO pooling node constraint, background prior variable constraint, self-loop prohibition constraint, and directed acyclic constraint. This avoids the formation of new directed cycles due to opposite path combinations after edge aggregation of multiple DAGs learned in a single iteration, thus improving the reliability of stable candidate statistical association directed graphs as adjacency structures for subsequent directed graph neural networks.

[0143] After obtaining the stable candidate statistical association directed graph, the computer device extracts the stable candidate directed edges that directly point to the PHENO node, and identifies the corresponding source node as the stable candidate direct association feature.

[0144] The computer device also searches the stable candidate statistical association directed graph for upstream variable paths that can reach the PHENO node through one or more stable candidate directed edges, and identifies these upstream variable paths as candidate risk propagation paths. A candidate risk propagation path can be represented as:

[0145]

[0146] Among them, P q This represents a candidate risk transmission path. L represents the l-th graph learning variable node in the path. q This represents the path length of the q-th candidate risk transmission path.

[0147] The computer equipment determines the path score based on the mean, product, minimum, or weighted combination of the edge scores of each stable candidate directed edge in the candidate risk transmission path, and determines the path stability based on the mean, product, minimum, or weighted combination of the edge stability of each stable candidate directed edge in the candidate risk transmission path.

[0148] VII. Training of Directed Graph Neural Networks

[0149] In one embodiment, the computer device uses candidate feature variable nodes and PHENO nodes from a stable candidate statistical association directed graph as nodes of a directed graph neural network, and stable candidate directed edges from the stable candidate statistical association directed graph as directed edges of the directed graph neural network, constructing the adjacency structure of the directed graph neural network. PHENO nodes are retained in the adjacency structure but do not receive actual PHENO values ​​as input.

[0150] For training samples, the computer device maps candidate feature values ​​to node signals of candidate feature variable nodes, and sets the node signals of the corresponding PHENO nodes of the training samples as placeholder node signals with non-true values. The placeholder node signals with non-true values ​​can be preset placeholder vectors, zero vectors, mask vectors, or learnable target node vectors.

[0151] In one embodiment, the initial node representation of candidate feature variable node i can be represented as:

[0152]

[0153] The initial node representation of a PHENO node can be represented as follows:

[0154]

[0155] or:

[0156]

[0157] or:

[0158]

[0159] Among them, Enc i Denotes the node feature encoding function, x i e represents the input value of the corresponding candidate feature variable. mask Represents a preset placeholder vector, e learn This represents the learnable target node vector. The PHENO label is used only as a supervised training objective and is not used as the input to the actual node signal of the PHENO node.

[0160] In the l-th layer message passing of a directed graph neural network, the computer device determines the incoming edge adjacency, outgoing edge adjacency, and self-connection relationships based on a stable candidate statistical association directed graph, and updates the node representation. The directed message passing can be represented as:

[0161]

[0162]

[0163] in, This represents the incoming edge aggregation message of the i-th node in layer l. G represents the outgoing edge association message of the i-th node in layer l. in (·) and g out (·) represent the incoming edge message function and the outgoing edge message function, respectively. ji and e ij This represents the edge weight representation, edge score, or learnable edge embedding of the corresponding stable candidate directed edge. W represents the node representation of the i-th node in the (l+1)-th layer. s Let represent the self-connection parameter matrix, and σ(·) represent the nonlinear activation function.

[0164] After directed message passing through layer L, the computer device obtains the PHENO node representations and the full graph node representations corresponding to the training samples. The computer device reads out the PHENO node representations, or the concatenated representation of the PHENO node representations and the pooled representations of the full graph nodes, to obtain the graph-level risk representation:

[0165]

[0166] or:

[0167]

[0168] Where Readout(·) represents the readout function, Pool(·) represents the node pooling function, and z represents the graph-level risk representation.

[0169] The computer device inputs the graph-level risk representation z into the output layer, generates the predicted output corresponding to the training samples, and calculates the training loss based on the difference between the predicted output and the PHENO labels constructed during the training phase. If the output is a binary risk probability, the training loss can be a binary cross-entropy loss; if the output is a multi-risk level probability, the training loss can be a multi-class cross-entropy loss or a risk stratification loss.

[0170] In one embodiment, the binary classification risk probability output can be expressed as:

[0171]

[0172] The training loss can be expressed as:

[0173]

[0174] in, y represents the predicted output corresponding to the training sample.PHENO The PHENO supervision labels constructed during the training phase are represented by , BCE(·) represents the binary cross-entropy loss, and MLP(·) represents the output layer of the multilayer perceptron. λ represents the model parameters of a directed graph neural network. g This represents the regularization weights of the directed graph neural network model. The computer device updates the model parameters of the directed graph neural network based on the training loss, resulting in the trained directed graph neural network model.

[0175] VIII. Risk Stratification Auxiliary Assessment and Structural Interpretation Data Output

[0176] For the object to be evaluated, the computer device maps the candidate feature values ​​of the object to be evaluated to the node signals of the corresponding candidate feature variable nodes in the trained directed graph neural network model, and sets the node signals of the PHENO nodes corresponding to the object to be evaluated as placeholder node signals of non-true values, so that the PHENO nodes of the object to be evaluated do not receive the true PHENO values ​​as input in the risk stratification auxiliary assessment data processing stage.

[0177] The computer device, based on a trained directed graph neural network model, performs directed message passing on the node signals corresponding to the object to be evaluated according to the incoming edge adjacency, outgoing edge adjacency, and self-connection relationships of the stable candidate statistical association directed graph, to obtain the PHENO node representation corresponding to the object to be evaluated. The computer device reads out the PHENO node representation, or reads out the concatenated representation of the PHENO node representation and the pooled representation of the whole graph nodes, to obtain a graph-level risk representation, and the output layer generates the model output data.

[0178] In one embodiment, the risk probability can be expressed as:

[0179] p risk =sigmoid(MLP(z))

[0180] In another embodiment, the probabilities corresponding to multiple risk levels can be expressed as:

[0181] p level =softmax(MLP(z))

[0182] Where, p risk p represents the probability of risk corresponding to the object to be evaluated. level This represents the probability vector corresponding to different risk levels. The model output data can be used to characterize the risk score, risk level representation value, or risk level corresponding probability of the object under assessment in a preset risk stratification model.

[0183] The computer device outputs stable candidate direct association features and candidate risk propagation paths obtained from the stable candidate statistical association directed graph. It then associates these stable candidate direct association features and candidate risk propagation paths with the corresponding edge stability, direction consistency, edge weight representation, edge score, path length, path stability, path edge weight representation, or path score to generate structural interpretation data corresponding to the model output data.

[0184] It should be noted that the actual PHENO values ​​of the objects to be evaluated are not used as input to the trained directed graph neural network model. The model output data and structural interpretation data are auxiliary evaluation data processing results generated by computer equipment and are not used to directly make diabetes diagnoses, treatment plans, or medication decisions.

[0185] IX. Example 1: Specific Processing Flow Based on Public Health Sample Data

[0186] In one specific embodiment, the multi-source health sample data can be derived from publicly available or authorized health survey data, with a sample size of up to 12,000 cases. The original fields can include demographic fields, questionnaire fields, physical examination fields, laboratory test fields, dietary and nutritional fields, and lifestyle fields. Demographic fields can include age, sex, race, education level, and income level; physical examination fields can include BMI, waist circumference, systolic blood pressure, and diastolic blood pressure; dietary and nutritional fields can include frequency of staple food intake, frequency of vegetable intake, frequency of fruit intake, frequency of sugary beverage intake, and frequency of high-fat food intake; lifestyle fields can include exercise frequency, weekly duration of moderate-to-vigorous intensity exercise, smoking status, alcohol consumption frequency, sleep duration, and sedentary time; and family history fields can include a first-degree relative's history of diabetes.

[0187] The computer equipment first performs sample-level merging of the above data, standardizing units and field names. For example, BMI is standardized to kg / m². 2 Waist circumference is uniformly expressed in cm, and systolic and diastolic blood pressure are uniformly expressed in mmHg; continuous variables are imputed using median and Z-score standardization; categorical variables are encoded using one-hot or ordered coding; fields with missing values ​​exceeding a preset ratio can be removed or treated as missing indicator variables only.

[0188] The computer device constructs PHENO based on one or more of the following: previous diabetes diagnosis information, hypoglycemic drug usage information, fasting plasma glucose (FPG), glycated hemoglobin (HbA1c), oral glucose tolerance test (OGTT), or random blood glucose (RBG). For example, in a non-limiting embodiment, PHENO is set to 1 when any of the following conditions are met: a positive previous diagnosis, a history of hypoglycemic drug use, FPG reaching a preset threshold, HbA1c reaching a preset threshold, OGTT reaching a preset threshold, or RBG reaching a preset threshold; otherwise, PHENO is set to 0.

[0189] After constructing PHENO, the computer equipment removed the following fields: FPG, HbA1c, OGTT, RBG, diabetes diagnosis information, hypoglycemic drug use, diabetes status directly generated from the above fields, abnormality marker fields derived from the above fields, test date, sample number, visit number, and data release date. After removal, candidate feature variables such as age, gender, race, education level, income level, BMI, waist circumference, systolic blood pressure, diastolic blood pressure, exercise frequency, smoking status, alcohol consumption frequency, food consumption frequency, sleep duration, and family history were retained to form a candidate feature matrix without label leakage.

[0190] The computer device sets PHENO as the sink node and sets age, gender, ethnicity, education level, and income level as background prior variables. These background prior variables are designated only as source nodes for candidate directed edges, not as target nodes. Therefore, any candidate directed edge pointing to age, gender, ethnicity, education level, or income level from any graph learning variable is prohibited. Age, gender, ethnicity, education level, and income level can point to other graph learning variables not excluded by the edge prohibition rule. The computer device generates an adjacency matrix mask based on the PHENO sink node constraint, background prior variable constraints, and self-loop prohibition constraint. Subsequently, the computer device learns the candidate edge weight matrix under the constraints of the adjacency matrix mask and optimizes it based on data fitting loss, structure-aware scoring terms, sparsity constraints, and acyclicity constraints.

[0191] In an example with unrestricted parameters, the number of repeated graph structure learning iterations R can be set to 20 or 50; each iteration uses a different random seed and performs 80% no-replacement subsamples from the training samples; the sparsity constraint weight λ1 can be set to 0.002, and the structural equation fitting model regularization weight λ θ The value can be set to 0.002, and the weight λ2 of the structure-aware scoring item can be set to 0.01; the edge weight threshold τ w It can be set to 0.30; the edge score threshold τ sIt can be set to 0.60; if a sorting and filtering method is used, the top 10% to 30% of candidate directed edges by edge score can be retained. The above parameters are only examples of implementation, and can be adjusted according to the sample size, number of variables, and validation set performance in actual applications.

[0192] In an example of a non-restricted structure-aware mapping network, a dynamic structure-aware mapping network The input is the sample feature vector and the candidate edge weight matrix constrained by the mask. The obtained structure-aware descriptor The hidden layer dimensions can be 64 and 32 respectively, the output dimension can be 16, and the activation function can be ReLU. The pre-training phase can be trained for 10 epochs; in the alternating optimization phase, after every certain number of updates to the structure equation fitting model and candidate edge weight matrices, the structure equation fitting model and candidate edge weight matrices are fixed, and the dynamic structure-aware mapping network is updated. Once, and to The parameters are used for range clipping or gradient clipping.

[0193] In this embodiment, the structure-aware scoring item L s This can be determined based on a sample-by-sample structure-aware mapping. For the i-th sample, the computer device will assign the sample vector x... i With structure-aware descriptors Input dynamic structure-aware mapping network Obtain the first structure-aware embedding z i ; will reconstruct sample vectors With structure-aware descriptors Input dynamic structure-aware mapping network Obtain the second structure-aware embedding

[0194]

[0195] Where i = 1, 2, ..., n, and [·] represents concatenation or joint encoding operations.

[0196] The first structure-aware embedding set is represented as: The second structure-aware embedding set is represented as:

[0197] Structural perception scoring item L s It can be determined based on the difference in batch mean:

[0198]

[0199] Alternatively, it can be determined based on the maximum difference between the means:

[0200]

[0201] in, This represents the candidate edge weight matrix after being masked. Indicates by The calculated structure-aware descriptor, Represents a dynamic structure-aware mapping network. Structure-aware descriptor. It can include one or more of the following: acyclicity descriptor, edge sparsity descriptor, incoming edge strength vector, outgoing edge strength vector, absolute value statistics of candidate edge weights, or adjacency structure embedding vector.

[0202] Compared to graph structure learning methods that only use ordinary reconstruction loss, the aforementioned structure-aware scoring term enables the computer device to optimize the candidate edge weight matrix not only based on X and The system adjusts candidate edges based on reconstruction errors and also adjusts the scoring space based on the acyclicity, sparsity, edge strength distribution, and edge weight statistics of the current candidate graph. Therefore, when multiple candidate directed graphs have similar reconstruction errors, the computer can prioritize candidate structures with lower acyclicity violations, sparser edge distribution, and more stable orientation states. This reduces the search for invalid candidate edges, minimizes the impact of false edges and orientation oscillations on graph structure learning results, and reduces fluctuations in repeated learning results caused by random initialization, sample resampling, and changes in training subsets. These effects represent improvements in graph structure search space constraints, candidate edge selection stability, and model training input structure stability during computer data processing, without relying on manual diagnostic judgment or directly drawing disease diagnosis conclusions.

[0203] After completing R differentiable directed acyclic graph structure learning operations, the computer device obtains R candidate directed adjacency matrices and corresponding R masked candidate edge weight matrices. For any candidate directed edge v i →v j The computer device counts the number of times the candidate directed edge appears in R repeated learning iterations, C. ij And calculate the edge stability S ij Edge weight representation B ij and direction consistency D ij Then determine the edge score according to the following form. ij :

[0204] Score ij =β1S ij +β2B ij +β3D ij

[0205] Wherein, β1, β2, and β3 are preset weight coefficients, β1, β2, and β3 > 0 and β1 + β2 + β3 = 1.

[0206] To ensure that the stable candidate statistical association directed graph formed after multiple repeated learning iterations still satisfies the directed acyclic constraint, the computer device further performs acyclic filtering when selecting stable candidate directed edges. Specifically, the computer device first performs edge score threshold τ... s Alternatively, the candidate stable edge set C can be obtained by sorting the edges by their scores:

[0207] C={v i →v j Score ij ≥τ s And M ij =1}

[0208] Alternatively, sort the edge scores by a predetermined ratio that satisfies M. ij Candidate directed edges with a value of 1 form a candidate stable edge set C. Subsequently, the computer device initializes the stable candidate edge set E. stable :

[0209]

[0210] Computer equipment according to Score ij Sort the candidate directed edges in the candidate stable edge set C from high to low, and then iterate through the candidate directed edges v in turn. i →v j For the currently traversed candidate directed edge, the computer device determines whether to add the candidate directed edge to E. stable The question asks whether a directed cycle has formed. This determination can be achieved through topological sorting, depth-first search, or reachability matrix updates. If adding the current candidate directed edge does not form a directed cycle, then the candidate directed edge is retained in E. stable If adding the current candidate directed edge results in a directed cycle, then skip that candidate directed edge. The above process can be represented as:

[0211] According to Score ij Each candidate directed edge v in descending order i →v j :

[0212] If Graph(E) stable ∪v i →v j If E does not contain directed cycles, then: stable =E stable ∪v i →v j Otherwise, keep E stable The edge remains unchanged, and the process continues to iterate over the next candidate directed edge.

[0213] After the traversal is complete, the computer device is based on E stable Construct a stable candidate statistical association directed graph Gstable Since a directed cycle check is performed before each candidate directed edge is added, the final stable candidate statistically associated directed graph G is obtained. stable It satisfies the PHENO pooling node constraint, background prior variable constraint, self-loop prohibition constraint, and directed acyclic constraint. This avoids the formation of new directed cycles due to opposite path combinations after edge aggregation of multiple DAGs learned in a single iteration, thus improving the reliability of stable candidate statistical association directed graphs as adjacency structures for subsequent directed graph neural networks.

[0214] In an unrestricted example, if the edge score for candidate edge "Age→BMI" is 0.82, the edge score for candidate edge "BMI→PHENO" is 0.79, the edge score for candidate edge "Waist Circumference→BMI" is 0.73, and the edge score for candidate edge "BMI→Waist Circumference" is 0.61, the computer device will perform acyclic filtering in descending order of edge scores. If adding "Waist Circumference→BMI" first and then adding "BMI→Waist Circumference" would create a directed cycle, the computer device will skip "BMI→Waist Circumference" and retain the candidate directed edges with higher scores that do not violate acyclicity. This process can reduce structural conflicts caused by simultaneously retaining edges with opposite directions, achieving a balance between statistical stability and structural usability in the stable candidate statistical association directed graph.

[0215] After obtaining a stable candidate statistical association directed graph G stable Then, the computer device extracts stable candidate directed edges directly pointing to the PHENO node and identifies the corresponding source node as a stable candidate directly associated feature. For example, stable candidate directly associated features directly pointing to the PHENO node may include BMI, waist circumference, exercise frequency, and family history. The computer device also extracts G... stable The search can identify upstream variable paths leading to the PHENO node through one or more stable candidate directed edges, and determine these upstream variable paths as candidate risk transmission paths. For example, candidate risk transmission paths may include "age → BMI → PHENO", "gender → waist circumference → BMI → PHENO", and "eating frequency → BMI → PHENO". Each candidate risk transmission path can be associated with path length, path stability, path edge weight representation, and path score.

[0216] In an example of training an unrestricted directed graph neural network, the computer divides the samples into training, validation, and test sets in a ratio of 70%:15%:15%. The directed graph neural network can be configured with 2 or 3 layers, with a hidden dimension of 32 or 64. The optimizer can be Adam, the learning rate can be set to 0.001, the batch size can be set to 128, and early stopping is performed based on the validation set loss. During training, candidate feature variable nodes input the corresponding candidate feature values, and PHENO nodes input the zero vector, mask vector, or learnable placeholder vector. PHENO labels are only used to calculate the binary cross-entropy loss, risk stratification loss, or regression loss and are not used as the actual node signal input for PHENO nodes.

[0217] Due to the stable candidate statistical association directed graph G stable Having been validated through edge stability, edge weight representation, directional consistency, and acyclic filtering, directed graph neural networks (DNNs) can perform directed message passing based on a more stable adjacency structure during the training phase. This reduces the instability of message passing paths caused by frequent changes in candidate edge directions, conflicts between edges in opposite directions, or local directed cycles, thereby improving the consistency of node aggregation paths and the repeatability of model output during training. The effectiveness of this technique is reflected in the stabilization of the computer model's training input structure, the determinization of the graph neural network's message passing paths, and the reduction of fluctuations in model output data, rather than in making direct medical judgments about disease states.

[0218] In an unrestricted output example, the trained directed graph neural network outputs a risk probability for a given object to be evaluated:

[0219] p risk =0.83

[0220] The system outputs a high-risk level based on preset stratification rules. Structural interpretation data may include: stable candidate directly related features pointing directly to PHENO, such as BMI, waist circumference, exercise frequency, and family history; candidate risk transmission paths include "age → BMI → PHENO", "gender → waist circumference → BMI → PHENO", and "diet frequency → BMI → PHENO"; each directly related edge and candidate risk transmission path can be associated with edge stability, directional consistency, edge weight representation, edge score, path length, path stability, and path score. The above outputs represent only computer-generated auxiliary assessment data and structural interpretation data, and are not intended to directly lead to a diabetes diagnosis, treatment plan determination, or medication decision.

[0221] 10. Example 2: The role of the structure-aware scoring term relative to the simple reconstruction loss

[0222] In one embodiment, if only the data fitting loss is used as the scoring term, different candidate directed graphs may have similar reconstruction errors, but their acyclicity, edge distribution, and edge orientation stability may differ significantly. In this case, the optimization process may select candidate graphs with lower reconstruction errors but containing pseudo-edges or unstable orientations.

[0223] The structure-aware scoring item L of this invention s Will be The obtained graph structure information is introduced into a dynamic structure-aware mapping network. The scoring space is influenced by acyclicity constraints, sparsity, incoming edge strength, outgoing edge strength, and edge weight statistics. This becomes particularly relevant when candidate graphs exhibit strong acyclicity violations or abnormal edge distributions. X and in structure-sensing space The difference provides a more sensitive metric; when the candidate graph gradually satisfies the directed acyclic structure and the edge distribution tends to stabilize, The scoring process gradually becomes smoother due to bounded constraints, parameter pruning, or Lipschitz constraints.

[0224] Therefore, compared with simple reconstruction loss, the structure-aware scoring term can incorporate the information of the candidate graph structure itself into the optimization objective, so that the optimization direction not only pursues data fitting, but also takes into account the degree of acyclicity, sparsity and directional stability of the candidate graph, thereby reducing the probability of pseudo edges, directional oscillations and local optimal candidate graph structures.

[0225] XI. Other Instructions

[0226] Each step in the above embodiments can be executed by a single computer device or by multiple computer devices working together. Each module can be implemented through software, hardware, or a combination of both. The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Any equivalent substitutions or modifications made within the spirit and principles of the present invention regarding the step order, variable encoding method, missing data handling method, structural equation fitting function, dynamic structure-aware mapping network, number of layers in a directed graph neural network, readout method, risk level classification method, and output format should be included within the scope of protection of the present invention.

Claims

1. A data processing method for auxiliary assessment of diabetes risk stratification, characterized in that, Performed by computer devices, including: S1: Obtain existing multi-source health sample data, preprocess the multi-source health sample data, and construct the diabetes phenotype label PHENO according to the preset phenotype construction rules; S2: Remove the label source variables, label leakage variables, and process record variables that are not used as inputs for model prediction related to the construction of PHENO, and form a candidate feature matrix without label leakage; S3: Construct a graph learning variable set based on the candidate feature variables and corresponding PHENO variables in the candidate feature matrix, and generate an adjacency matrix mask according to preset edge constraints, wherein the preset edge constraints include at least PHENO node convergence constraints, self-loop prohibition constraints, and background prior variable constraints. S4: Under the adjacency matrix mask constraint, differentiable directed acyclic graph structure learning is performed on the graph learning variable set to obtain candidate statistical association directed graphs, wherein the objective function of differentiable directed acyclic graph structure learning includes a data fitting loss term, a structure constraint term, and a structure-aware scoring term. S5: Perform the differentiable directed acyclic graph structure learning multiple times, and determine the stability index of candidate directed edges based on the results of multiple learnings, screen stable candidate directed edges, and form a stable candidate statistical association directed graph; S6: Using the stable candidate statistical association directed graph as the adjacency structure of the directed graph neural network, the directed graph neural network is trained, wherein the PHENO label is used as the supervised training target, and the PHENO node does not receive the actual PHENO value as the node signal input. S7: Based on the trained directed graph neural network, perform directed message passing on the object to be evaluated, generate model output data, and generate structural interpretation data corresponding to the model output data based on the stable candidate statistical association directed graph; The model output data is used to characterize the risk score, risk level representation value, or risk level corresponding probability of the subject to be evaluated in the preset risk stratification model; the data processing method only performs computer data processing on the acquired data, does not include the steps of conducting detection, sampling, diagnosis, treatment, or medication decisions on human or animal bodies, and does not directly make a diabetes diagnosis conclusion based on the model output data.

2. The method according to claim 1, characterized in that, In step S1, existing multi-source healthy sample data is acquired, the multi-source healthy sample data is preprocessed, and a diabetes phenotype label PHENO is constructed according to preset phenotype construction rules, including: S11: Obtain existing multi-source health sample data, which includes one or more of the following: demographic data, questionnaire data, physical examination data, laboratory test data, dietary and nutritional data, and lifestyle data; S12: Perform variable name unification, variable unit conversion, variable value range verification, and sample-level merging on the multi-source health sample data to obtain the original merged data table; S13: Perform missing value processing, outlier processing, categorical variable encoding, continuous variable standardization, and sample screening on the original merged data table to obtain a preprocessed health feature data table; S14: Construct a diabetes phenotype label PHENO according to a preset phenotype construction rule, wherein the preset phenotype construction rule consists of one or more of the following: previous diabetes diagnosis information, hypoglycemic drug use information, fasting blood glucose index, glycated hemoglobin index, oral glucose tolerance test index, or random blood glucose index. The previously obtained diabetes diagnosis information, hypoglycemic drug usage information, fasting blood glucose index, glycated hemoglobin index, oral glucose tolerance test index, and random blood glucose index are all historical data or existing test data acquired before the data processing method is executed. The diabetes phenotype label PHENO is generated by computer equipment based on the historical data or existing test data according to the preset phenotype construction rules. S1 does not include the steps of conducting detection, sampling, diagnosis, treatment, or medication decisions on human or animal bodies.

3. The method according to claim 1, characterized in that, In step S2, label source variables, label leakage variables, and process record variables that are not used as model prediction inputs are removed from the PHENO construction process to form a candidate feature matrix without label leakage, including: S2.1: Based on the preset phenotype construction rules, determine the set of tag source variables participating in the construction of PHENO. The set of tag source variables includes one or more of the following variables used to construct PHENO: previous diabetes diagnosis information variables, hypoglycemic drug use information variables, fasting blood glucose index variables, glycated hemoglobin index variables, oral glucose tolerance test index variables, or random blood glucose index variables. S2.2: According to the preset label leakage judgment rules, remove the label source variables in the label source variable set, and remove the label leakage variables that have a direct mapping relationship, homo-derived relationship or deterministic transformation relationship with the label source variables. The label leakage variables include variables directly mapped from the label source variables, variables belonging to the same field as the label source variables, derived variables belonging to the same test item as the label source variables, derived variables belonging to the same questionnaire item as the label source variables, derived variables belonging to the same drug use record as the label source variables, and variables that can be transformed from the label source variables through deterministic rules. S2.3: Remove object unique identifier variables, sequence number variables, sample number variables, visit number variables, data release date variables, sampling date variables, inspection date variables, or other process record variables that are not used as inputs for model prediction; S2.4: The remaining candidate health features after removing the label source variables, label leakage variables and process record variables are used to form the candidate feature matrix without label leakage.

4. The method according to claim 1, characterized in that, In step S3, a graph learning variable set is constructed based on the candidate feature variables and PHENO corresponding variables in the candidate feature matrix, and an adjacency matrix mask is generated according to preset edge constraints, including: S3.1: Construct a graph learning variable set V = {v1, v2, ..., v...} based on the candidate feature variables in the candidate feature matrix and the corresponding PHENO variables. p }, where p is the number of graph learning variables, and the graph learning variable set V includes candidate feature variable nodes and PHENO nodes; S3.2: Set the PHENO node as the pool node, so that the PHENO node can only be the target node of the candidate directed edge, and prohibit the PHENO node from being the source node of the candidate directed edge pointing to any candidate feature variable node. S3.3: Set one or more of the following demographic variables as background prior variables: age, gender, race, education level, income level, or other demographic variables that have been determined before the risk stratification auxiliary assessment and do not change with the current health data, forming a set of background prior variables. The background prior variables are only used as source nodes of candidate directed edges and not as target nodes of any candidate directed edge. The background prior variables are allowed to point to other graph learning variables that have not been excluded by the edge prohibition rule. S3.4: Generate the adjacency matrix mask M based on the PHENO convergence node constraints, background prior variable constraints, and self-loop prohibition constraints. M∈{0，1} p×p Among them, M ij This indicates that variable v is learned from the i-th graph. i Learning variable v from the j-th graph j Whether candidate directed edges are allowed, M ij =0 indicates that v is prohibited i Point to v j Candidate directed edges, M ij =1 means that v is allowed i Point to v j Candidate directed edges; S3.5: The adjacency matrix mask M is determined according to the following rules: When i = j, M ij =0, used to disable self-loops in graph learning variables; When v i For PHENO node and v j When not a PHENO node, M ij =0, used to prevent the PHENO node from pointing to the candidate feature variable node; When v j When M belongs to the set of background prior variables, ij =0, used to prevent any graph learning variable from pointing to a background prior variable; For candidate directed edges not prohibited by the self-loop prohibition constraint, PHENO convergence node constraint, and background prior variable constraint, M ij =1; S3.6: Use the adjacency matrix mask M as an edge feasibility constraint for subsequent learning of differentiable directed acyclic graph structures; PHENO is used only during the training phase for constructing target phenotypic nodes in graph structure learning, and is not used as a true value input during the risk stratification auxiliary assessment data processing phase of the object to be evaluated.

5. The method according to claim 1, characterized in that, In step S4, under the adjacency matrix mask constraint, a differentiable directed acyclic graph structure is learned from the graph learning variable set to obtain a candidate statistical association directed graph, including: S4.1: Construct a matrix of variables to be learned, X, ∈ R, based on the graph learning variable set V. n×p Where n represents the number of samples, p represents the number of graph learning variables, and the matrix X of variables to be learned includes the data column corresponding to the candidate feature variables and the data column corresponding to PHENO; S4.2: Construct the original learnable directed edge weight matrix W, W∈R p×p The adjacency matrix mask M is then multiplied element-wise with the original learnable directed edge weight matrix W to obtain a candidate edge weight matrix constrained by the mask. Where ⊙ denotes element-wise multiplication, and W represents the original learnable directed edge weight matrix. Let M represent the candidate edge weight matrix after masking constraints. ij Candidate directed edges with = 0 in M is set to zero. ij Candidate directed edges with a value of 1 are in The optimization process is allowed; S4.3: Candidate edge weight matrix based on the mask constraint Construct a structural equation fitting model f θ And fit the model f using the structural equation. θ The matrix of variables to be learned, X, is reconstructed to obtain the reconstructed variable matrix. Where θ represents the structural equation fitting model f θ Learnable parameters; S4.4: Based on the learning variable matrix X and the reconstructed variable matrix The difference between them determines the data fitting loss L. fit : Where Loss(·) represents the expression used to measure X and The loss function is the difference between the two, wherein the loss function is squared loss, negative log-likelihood loss, reconstruction error loss or a combination thereof; S4.5: Based on the masked candidate edge weight matrix Determine the acyclic constraint term Where tr(·) represents the matrix trace, exp(·) represents the matrix exponent, p represents the number of graph learning variables, and the acyclicity constraint term... Used to characterize by The degree of violation of directed acyclic structures by the determined candidate directed graphs; S4.6: Fitting loss L based on the data fit Structural perception scoring item L s Acyclic constraint terms Construct an objective function with sparse constraints, and fit a model f to the original learnable directed edge weight matrix W and the structure equation. θ Optimize: in, Represents the candidate edge weight matrix under mask constraints The sparse constraint term, where λ1 represents the sparse constraint weight, λ2 represents the structure-aware scoring term weight, and λ θ represents the regularization weights of the structural equation fitting model, and α and ρ represent the augmented Lagrange coefficients; S4.7: During the optimization process, based on the aforementioned acyclic constraint term... The changes adjust the augmented Lagrange coefficients α and ρ to promote the change from The determined candidate directed graphs satisfy the directed acyclic structure; S4.8: After optimization, the mask-constrained candidate edge weight matrix is... Threshold pruning and sparsification are performed to obtain a directed candidate adjacency matrix A, where for any candidate directed edge v i →v j A is determined according to the following rules. ij : Where, τ w This indicates a preset edge weight threshold; S4.9: Form a candidate statistical association directed graph based on the directed candidate adjacency matrix A.

6. The method according to claim 5, characterized in that, In S4.6, the structure-aware scoring item L s Through dynamic structure-aware mapping network and structure-aware descriptors Determined, including: S4.6.1: Constructing a dynamic structure-aware mapping network and structure-aware descriptors in, Represents a dynamic structure-aware mapping network The learnable parameters of the structure-aware descriptor The candidate edge weight matrix constrained by the mask The calculated information is used to characterize candidate graph structure information, which includes one or more of the following: acyclicity descriptor, sparsity descriptor, incoming edge strength information, outgoing edge strength information, candidate edge weight statistics, or adjacency structure embedding information. S4.6.2: For the i-th sample, the i-th sample vector x in the matrix X to be learned is... i With the structure-aware descriptor After splicing or joint encoding, the data is input into the dynamic structure-aware mapping network. Obtain the first structure-aware embedding z i The reconstructed variable matrix The i-th reconstructed sample vector in With the structure-aware descriptor After splicing or joint encoding, the data is input into the dynamic structure-aware mapping network. Obtain the second structure-aware embedding Where i = 1, 2, ..., n, and [·] represents concatenation or joint encoding operation; S4.6.3: Based on the first structure-aware embedding set and the second structure-aware embedding set Determine the structural perception scoring item L s Wherein, the structure-aware scoring item L s This represents the difference between the batch mean and the maximum mean. When the structure-aware scoring item L s When calculating the batch mean difference, it is determined according to the following formula: When the structure-aware scoring item L s For the maximum mean difference, it is determined according to the following formula: Among them, the structure-aware scoring item L s Used to characterize the learning variable matrix X and the reconstructed variable matrix Differences in batch distribution in the structure-aware space.

7. The method according to claim 6, characterized in that, The dynamic structure-aware mapping network The graph structure learning process employs pre-training and alternating optimization methods for updates, including: S4.6.1.1: During the pre-training phase, the structural equation model f is fitted. θ and the dynamic structure-aware mapping network Perform initial training; S4.6.1.2: During the alternating optimization phase, the structural equation fitting model f is fixed. θ and the mask-constrained candidate edge weight matrix To enhance the first structure-aware embedding set With the second structure-aware embedding set Distinguishing between them is the basis for target update. The The update target is: in, This represents the set of constraints for the parameters of a dynamic structure-aware mapping network. This represents the regularized weights of the dynamic structure-aware mapping network parameters. The L2 regularization term represents the parameters of a dynamic structure-aware mapping network; S4.6.1.3: In updating the dynamic structure-aware mapping network At that time, the dynamic structure perception mapping network One or more of the following constraints are applied to the parameters, gradients, or output of the dynamic structure-aware mapping network: bounded constraints, norm constraints, gradient clipping, parameter clipping, output normalization, spectral normalization, or Lipschitz constraints, to limit the network. Degenerates into a structure-aware scoring item L s Invalid mapping function; S4.6.1.4: Optimizing the original learnable directed edge weight matrix W and the structure equation fitting model f θ At that time, the dynamic structure-aware mapping network is fixed. And make the structure-aware scoring item L s Participate in the optimization of the objective function described in claim 5; Among them, through updating When f is fixed θ and and with enhancement and The distinguishability between them is the objective, and it is fixed when updating W and θ. This enables the structure-aware scoring term L_s to stably participate in the optimization of the candidate edge weight matrix, thereby reducing pseudo edges, unstable orientations, or locally optimal candidate graph structures caused by simple data fitting.

8. The method according to claim 1, characterized in that, In step S5, the differentiable directed acyclic graph structure learning is performed multiple times, and the stability index of candidate directed edges is determined based on the results of multiple learnings. Stable candidate directed edges are then selected to form a stable candidate statistically correlated directed graph, including: S5.1: Obtain R repeated learning tasks, where each repeated learning task is formed based on one or more of different random seeds, different data resampling subsets, or different training subsets, and R is an integer greater than 1; S5.2: Perform the differentiable directed acyclic graph structure learning R times respectively to obtain R candidate directed adjacency matrices and corresponding R candidate edge weight matrices constrained by the mask. Each differentiable directed acyclic graph structure learning is performed under the constraint of the adjacency matrix mask M and satisfies the PHENO convergence node constraint, background prior variable constraint and self-loop prohibition constraint. S5.3: For any candidate directed edge v i →v j Count the number of times C it appears in the R candidate directed adjacency matrices. ij : The candidate directed edge v is determined according to the following formula. i →v j Edge stability S ij : Where 1(·) represents the indicator function, This indicates that in the r-th differentiable directed acyclic graph structure learning, the structure of v is preserved. i Point to v j Candidate directed edges, This indicates that in the r-th differentiable directed acyclic graph structure learning, the structure of v was not preserved. i Point to v j Candidate directed edges; S5.4: Based on the candidate directed edge v i →v j The candidate directed edge v is determined by the mean absolute value of the edge weights in the R masked candidate edge weight matrices. i →v j Edge weight representation of B ij : in, This represents the masked candidate edge weight matrix obtained by learning the r-th differentiable directed acyclic graph structure. Zhongyou v i Point to v j The edge weights, Norm(·) represent one of the following: min-max normalization, quantile normalization, or normalization according to the allowed edge set; S5.5: Compare the directed edges of candidates in opposite directions between pairs of the same variable, and determine the direction based on the first direction candidate directed edge v. i →v j The number of times C appears ij Candidate directed edge v in the opposite direction j →v i The number of times C appears ji Determine the candidate directed edge v i →v j Directional consistency D ij : Where ε is a preset constant used to avoid the denominator being zero; S5.6: According to the edge stability S ij Edge weight representation B ij and direction consistency D ij Determine the candidate directed edge v i →v i Edge score ij : Score ij =β1S ij +β2B ij +β3D ij Wherein, β1, β2 and β3 are preset weight coefficients, β1, β2, β3 > 0 and β1 + β2 + β3 = 1; S5.7: Within the set of candidate edges that satisfy the PHENO convergence node constraint, background prior variable constraint, self-loop prohibition constraint, and acyclicity constraint, retain the edge score. ij Greater than the preset edge score threshold τ s Candidate directed edges, or retain edge scores. ij Candidate directed edges that are ranked at the top of the preset proportion are added one by one in descending order of edge score. If a directed cycle is formed after adding them, they are skipped. S5.8: Construct a stable candidate statistical association directed graph based on the stable candidate directed edges; S5.9: Extract stable candidate directed edges that directly point to the PHENO node from the stable candidate statistical association directed graph, and determine the source node corresponding to the stable candidate directed edge as the stable candidate direct association feature; S5.10: Search for upstream variable paths from the stable candidate statistical association directed graph that can reach the PHENO node through one or more stable candidate directed edges, and determine the upstream variable paths as candidate risk transmission paths; S5.11: Determine the path score based on the mean, product, minimum, or weighted combination of the edge scores of each stable candidate directed edge in the candidate risk transmission path; determine the path stability based on the mean, product, minimum, or weighted combination of the edge stability of each stable candidate directed edge in the candidate risk transmission path; associate the stable candidate direct association features with the corresponding edge stability, direction consistency, edge weight representation, or edge score; and associate the candidate risk transmission path with the corresponding path length, path stability, path edge weight representation, or path score.

9. The method according to claim 1, characterized in that, In step S6, the directed graph neural network is trained using the stable candidate statistical association directed graph as the adjacency structure of the directed graph neural network, including: S6.1: Using the candidate feature variable nodes and PHENO nodes in the stable candidate statistical association directed graph as nodes of the directed graph neural network, and using the stable candidate directed edges in the stable candidate statistical association directed graph as directed edges of the directed graph neural network, the adjacency structure of the directed graph neural network is constructed. S6.2: Based on the candidate feature values ​​of each training sample in the training sample set, construct the node signal of the corresponding candidate feature variable node, and set the node signal of the PHENO node corresponding to the training sample as a placeholder node signal with non-true values, wherein the placeholder node signal with non-true values ​​is a preset placeholder vector, zero vector, mask vector or learnable target node vector. S6.3: The directed graph neural network is trained by using the PHENO labels constructed according to the preset phenotypic construction rules during the training phase as the supervised training target. The PHENO labels are used only as supervised labels for calculating classification loss, risk stratification loss or regression loss, and are not used as real node signals of PHENO nodes input into the directed graph neural network. S6.4: During the supervised training process, the incoming edge adjacency relationship, outgoing edge adjacency relationship and self-connection relationship are determined based on the stable candidate statistical association directed graph, and directed message passing is performed according to the incoming edge adjacency relationship, outgoing edge adjacency relationship and self-connection relationship to obtain the PHENO node representation or graph-level risk representation corresponding to the training sample. S6.5: Generate the predicted output corresponding to the training sample based on the PHENO node representation or graph-level risk representation corresponding to the training sample, calculate the training loss according to the difference between the predicted output and the PHENO label, update the model parameters of the directed graph neural network, and obtain the trained directed graph neural network model. The PHENO tags are used only as supervised learning targets during the training phase and are not used as actual node signal inputs for PHENO nodes during the construction of node signals and the transmission of directed messages in the directed graph neural network.

10. The method according to claim 1, characterized in that, In step S7, directed message passing is performed on the object to be evaluated based on the trained directed graph neural network to generate model output data, and structural interpretation data corresponding to the model output data is generated based on the stable candidate statistical association directed graph, including: S7.1: For the object to be evaluated, the candidate feature values ​​of the object to be evaluated are mapped to the node signals of the corresponding candidate feature variable nodes in the trained directed graph neural network model, and the node signals of the PHENO nodes corresponding to the object to be evaluated are set as placeholder node signals of non-true values, so that the PHENO nodes of the object to be evaluated do not receive the true PHENO values ​​as input in the risk stratification auxiliary assessment data processing stage. S7.2: Based on the trained directed graph neural network model, according to the incoming edge adjacency relationship, outgoing edge adjacency relationship and self-connection relationship of the stable candidate statistical association directed graph, the node signal corresponding to the object to be evaluated is transmitted in a directed message to obtain the PHENO node representation corresponding to the object to be evaluated. S7.3: Read out the PHENO node representation, or read out the concatenation representation of the PHENO node representation and the pooled representation of the whole graph nodes to obtain a graph-level risk representation; S7.4: Input the graph-level risk representation into the output layer to generate model output data. The model output data is used to represent the risk score, risk level representation value, or risk level corresponding probability of the object to be evaluated in the preset risk stratification model. S7.5: Output the stable candidate direct association features and candidate risk transmission paths obtained from the stable candidate statistical association directed graph, and associate the stable candidate direct association features and candidate risk transmission paths with the corresponding edge stability, direction consistency, edge weight representation, edge score, path length, path stability, path edge weight representation or path score to generate structural interpretation data corresponding to the model output data; The actual PHENO value of the object to be evaluated is not used as the input to the trained directed graph neural network model. The model output data and structural interpretation data are auxiliary evaluation data processing results generated by computer equipment and are not used to directly make diabetes diagnoses, treatment plans, or medication decisions.

Images