Causal graph neural network prediction method for eliminating ad spurious correlations and media

Abstract

The application relates to a causal graph neural network prediction method and medium for eliminating AD false correlation. The method first constructs a graph structure of SNPs and brain images, extracts features using a time series graph network and maps the features to a disease quasi-time axis, and then divides stages and generates a graph sequence. The causal contribution degree between nodes is calculated through a time series prediction network to obtain a causal prior graph; direction correction is performed by using counterfactual flipping and graph convolution to obtain a refined causal graph. The features are subjected to causal enhancement, and double-mode information is fused through cross-attention. Finally, a classifier is input to realize AD staging prediction. The method overcomes the limitation that existing methods are difficult to distinguish between causality and false correlation in cross-sectional data, and through explicit modeling of disease time evolution and causal structure, the ability to capture early Alzheimer's disease and subtle pathological patterns is enhanced, thereby improving the accuracy, interpretability and cross-domain generalization ability of classification prediction.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence and neuroimaging analysis technology, specifically to a causal graph neural network prediction method and medium for eliminating spurious correlations in Alzheimer's disease (AD). Background Technology

[0002] Alzheimer's disease (AD), a progressive neurodegenerative disease with a clear temporal evolutionary pattern, requires precise early diagnosis and analysis of its pathological mechanisms. In recent years, with the development of neuroimaging and genomics, the use of multimodal data (such as structural magnetic resonance imaging (sMRI) and single nucleotide polymorphisms (SNPs)) to build deep learning models has become the mainstream paradigm for intelligent diagnosis of AD.

[0003] However, current AD diagnostic models based on multimodal data (such as brain imaging and genetic data), especially advanced methods relying on graph neural networks, still have significant limitations, mainly in the following three aspects: First, existing models are generally built based on cross-sectional data, mixing individuals at different disease stages. This approach makes the models prone to learning spurious correlations that change synchronously with disease stages but are not the true causes. For example, the model may capture certain patterns of brain volume changes that only stabilize in late-stage Alzheimer's disease and misclassify them as key diagnostic features, even though these features may not exist in the early stages of the disease or drive disease progression. This over-reliance on statistical co-occurrence leads to a lack of causal interpretability in the features learned by the model and a significant decline in generalization performance when applied to different data distributions (such as different scanning devices, collection centers, or population cohorts), i.e., weak "out-of-distribution generalization" ability.

[0004] Secondly, most existing graph neural network methods construct graph structures based on static correlations within modalities (such as using Pearson correlation coefficients to define connections between brain regions or gene loci). Such static graphs cannot characterize the dynamic evolution of AD pathophysiology over time. Disease progression is essentially a process of orderly changes in biomarkers and brain network topology over time, but static graph models cannot distinguish between early driving factors and later accompanying phenomena, thus making it difficult to capture the dynamic causal relationships of disease progression.

[0005] Furthermore, causal discovery from observational data to distinguish between causal correlations and spurious correlations is crucial for improving model interpretability and robustness. However, classical causal discovery algorithms (such as constraint-based methods) often suffer from the "Markov equivalence class" problem when processing cross-sectional data due to a lack of time dimension information. This means they cannot statistically distinguish whether variable A causes B or B causes A, resulting in causal graphs containing a large number of undirected edges and lacking directional information. Although some studies have attempted to introduce strong assumptions (such as additive noise models) to break directional symmetry, these assumptions are often difficult to uphold given the high heterogeneity and complexity of biomedical data, easily leading to erroneous causal inferences.

[0006] In summary, existing technologies face three core challenges in multimodal AD diagnosis: spurious correlations, lack of dynamic evolutionary information, and difficulty in determining causal orientation in cross-sectional data. Therefore, there is an urgent need for a novel method capable of reconstructing the disease evolution timeline from cross-sectional data, identifying and eliminating spurious correlations, and inferring robust causal orientations to achieve more accurate, interpretable, and generalizable AD diagnosis and pathological analysis. Summary of the Invention

[0007] This invention provides a causal graph neural network prediction method and medium for eliminating spurious correlations in Alzheimer's disease (AD). The purpose is to solve the problems of existing AD diagnosis methods based on cross-sectional multimodal data, which are difficult to distinguish and eliminate spurious correlations, cannot model the dynamic evolution of the disease, and have unreliable causal direction inferences, resulting in limited model classification performance, poor interpretability, and weak generalization ability.

[0008] To achieve the above objectives, the first aspect of the present invention provides a causal graph neural network prediction method for eliminating spurious correlations in Alzheimer's disease (AD), comprising the following steps: Based on the Pearson correlation coefficient, a bimodal undirected graph of single nucleotide polymorphism sites and brain imaging data was constructed. Each single nucleotide polymorphism site and each brain region were defined as a graph node, and brain imaging graph and gene graph were constructed respectively. The brain image map and gene map are processed by a temporally aware graph attention network to generate a graph embedding vector for each sample. The graph embedding vector is then mapped to a continuous pseudo-temporal value. The samples are then assigned to different stage buckets according to the pseudo-temporal value. The graphs in each stage bucket are aggregated to obtain a staged graph sequence containing node features and an adjacency matrix. Using the aforementioned phased graph sequence, a time-aware prediction network is used to calculate the causal contribution of different nodes under multiple time lags, and a hybrid causal prior graph is generated based on the causal contribution. The adjacency matrix elements corresponding to the directed edges in the hybrid causal prior graph are parameterized. The prediction error before and after flipping the direction of the directed edges is evaluated using a graph convolutional network. The flipping reward is calculated, and the direction of the hybrid causal prior graph is corrected based on the flipping reward to obtain a refined causal graph. The node features in the stage graph sequence are causally enhanced based on the refined causal graph, and a multi-head cross-attention mechanism is applied to perform bidirectional interactive fusion of the enhanced brain image modality features and gene modality features. The fused bimodal global feature vector is input into the classifier to complete the classification and prediction of different stages of AD.

[0009] Furthermore, in the step of constructing a bimodal undirected graph of single nucleotide polymorphism sites and brain imaging data based on Pearson correlation coefficient: The nodes of the brain image map are brain regions. The feature vector of each brain region node includes gray matter volume, white matter volume and cerebrospinal fluid volume. The edge features between nodes are obtained by calculating the Pearson correlation coefficient between brain region features. The nodes of the gene map are single nucleotide polymorphism sites. The features of each site node are represented by 0-1-2 encoding to indicate the genotype. The edge features between nodes are obtained by calculating the Pearson correlation coefficient between site codes. Thresholds are set for the brain imaging map and the gene map respectively, and only edges with Pearson correlation coefficient absolute values ​​greater than the corresponding thresholds are retained as initial topological connections.

[0010] Furthermore, the method for using a temporally aware graph attention network to process the brain image map and the gene map respectively, and generating a graph embedding vector for each sample, includes: Two stacked graph attention layers are used to encode the initial features of the nodes; The first graph attention layer adopts a multi-head attention mechanism. It calculates the attention coefficients between nodes and their neighbors through linear transformation and LeakyReLU activation function, and normalizes them within the set of neighbor nodes using the Softmax function. The normalized multi-head attention outputs are then concatenated to obtain the first layer of node features. The second graph attention layer adopts a single-head attention mechanism to fuse and reduce the dimensionality of the first-layer node features output by the first layer. The neighborhood features are weighted and summed through new attention weights to obtain the node embedding that fuses the structural information of the entire graph. Global average pooling is performed on the embedding features of all nodes after processing by the two graph attention layers to calculate the arithmetic mean of the feature vectors of all nodes and generate sample-level graph embedding vectors.

[0011] Furthermore, the graph embedding vector is mapped to continuous quasi-temporal values, calculated using the following formula:

[0012] in, Represents the pseudo-time series value of the sample. This represents the Sigmoid activation function. This is the weight vector of the fully connected regression layer. Let be the graph embedding vector of the sample. For bias terms; The pseudo-time series values ​​output by the regression layer are activated by the Sigmoid activation function. The constraint is within the range of 0 to 2, where 0 corresponds to the healthy control stage and 2 corresponds to the late stage of AD.

[0013] Furthermore, the method for allocating samples to different stage buckets based on the pseudo-time series values, and aggregating the graphs within each stage bucket to obtain a staged graph sequence containing node features and an adjacency matrix includes: The continuous range of the pseudo-time series value is evenly divided into K stage bucket intervals; Each sample is assigned to the corresponding stage bucket based on its pseudo-time series value. For each non-empty stage bucket, the node feature matrix and adjacency matrix of all samples in each non-empty stage bucket are averaged and aggregated to obtain the node feature matrix and adjacency matrix of that non-empty stage bucket. For empty stage buckets, a baseline blocking strategy is adopted: its node feature matrix is ​​obtained by randomly selecting sub-samples from the first non-empty stage bucket, calculating the mean of node features, and then multiplying it by the decay coefficient, and its adjacency matrix is ​​set to an all-zero matrix. The graph representations obtained by aggregating buckets at each stage are arranged in stage order to form a staged graph sequence containing a node feature matrix and an adjacency matrix.

[0014] Furthermore, methods for calculating the causal contribution of different nodes under various time lags include: Calculate the baseline prediction error of the target node when only its own historical state is used; Calculate the enhanced prediction error after incorporating the historical state of the source node; The initial causal contribution is calculated based on the difference between the baseline prediction error and the enhanced prediction error; A correction network constructed using a multilayer perceptron is used to correct the initial causal contribution to obtain the final causal contribution.

[0015] Furthermore, the method of evaluating the prediction error before and after flipping the directed edge direction using a graph convolutional network, calculating the flipping gain, and correcting the orientation of the mixed causal prior graph based on the flipping gain to obtain a refined causal graph includes: A learnable parameter matrix is ​​introduced and mixed with the adjacency matrix of the hybrid causal prior graph to generate a parameterized adjacency matrix. Based on the parameterized adjacency matrix and node features, the original prediction error of the graph-level prediction task is calculated using a graph convolutional network. For each directed edge in the hybrid causal prior graph, a counterfactual flip graph is constructed, wherein the direction of the directed edge is forcibly flipped, and the flip prediction error is calculated based on the counterfactual flip graph. Calculate the flipping reward for each directed edge based on the original prediction error and the flipping prediction error; and For edges whose flipping gains exceed a preset threshold, their direction is corrected to a reverse connection in the hybrid causal prior graph, thereby obtaining the refined causal graph.

[0016] Furthermore, methods for bidirectional interactive fusion of enhanced brain imaging modal features and genetic modal features using multi-head cross-attention mechanisms include: Using the global feature vector of the brain imaging modality as the query and the global feature vector of the gene modality as the key and value, information related to brain imaging in the gene modality is extracted through a multi-head attention mechanism to generate image-guided gene enhancement features. Using the global feature vector of the gene modality as the query and the global feature vector of the brain imaging modality as the key and value, the gene-related information in the brain imaging modality is extracted through the multi-head attention mechanism to generate gene-guided image enhancement features. For each attention head, the query vector, key vector, and value vector are calculated separately, and the attention weights are obtained by dot product and Softmax normalization. The value vector is weighted and aggregated using the attention weights to obtain the interactive output of each head; The outputs of all attention heads are concatenated, mapped through a linear projection layer, and then subjected to residual connections and layer normalization to obtain the final bidirectional interactive fusion feature.

[0017] Furthermore, prior to the step of constructing the bimodal undirected graph, a data preprocessing step is also included: Brain region features are extracted from the brain imaging data based on the AAL template; The single nucleotide polymorphism (SNP) site data were subjected to quality control and screening, and the genotypes were converted into numerical codes.

[0018] To achieve the above objectives, a second aspect of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, performs the steps of the causal graph neural network prediction method for eliminating spurious AD correlations.

[0019] The beneficial effects of this invention are: Compared with existing technologies, this invention provides a causal graph neural network prediction method and medium for eliminating spurious correlations in Alzheimer's disease (AD). It addresses the aforementioned problems by constructing an end-to-end causal graph neural network framework integrating quasi-temporal reconstruction, dynamic causal discovery, and reverse verification. First, a temporal-aware graph attention network is used to learn continuous quasi-temporal values ​​for each sample from cross-sectional multimodal data, and bucketing and aggregation are performed accordingly. This reconstructs a staged graph sequence simulating the dynamic evolution of the disease, providing temporal structural constraints for subsequent analysis, even in the absence of longitudinal data. Second, based on this sequence, a time-lag causal judgment module is designed to quantify causality by comparing the changes in prediction errors between nodes at different time lags. The model calculates contribution and generates a hybrid causal prior graph, transforming static associations into directional causal hypotheses, effectively distinguishing true causality from accompanying spurious correlations. Furthermore, a reverse causal confirmation module is introduced, using a parameterized adjacency matrix and a counterfactual flipping mechanism to verify and correct the directional robustness of directed edges, further eliminating erroneous causal directions caused by data noise or equivalence class confusion, thus improving the reliability of the causal structure. Finally, refined causal topology is used to enhance node features, and a multi-head cross-attention mechanism is used to achieve deep semantic alignment and complementary fusion between gene and image modalities, enabling the model to capture cross-modal synergistic pathogenic mechanisms. Finally, Focal Loss is used in the classification layer to alleviate class imbalance, thereby achieving the extraction and modeling of true causal features in the progression of Alzheimer's disease, significantly improving classification accuracy, model interpretability, and generalization ability on out-of-distribution data. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below.

[0021] Figure 1 This is a flowchart of a causal graph neural network prediction method for eliminating spurious AD correlations, as disclosed in an embodiment of the present invention.

[0022] Figure 2 This is a causal modeling framework diagram for eliminating spurious correlations disclosed in an embodiment of the present invention. Detailed Implementation

[0023] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0024] According to embodiments of the present invention, it should be noted that the steps shown in the flowcharts of the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the following methods, in some cases the steps shown or described may be executed in a different order than that shown here.

[0025] like Figure 1 As shown, this invention provides a causal graph neural network prediction method for eliminating spurious AD correlations, the method comprising the following steps: Step S100: Construct a bimodal undirected graph of single nucleotide polymorphism sites and brain imaging data based on Pearson correlation coefficient. Define each single nucleotide polymorphism site and each brain region as a graph node, and construct a brain imaging graph and a gene graph respectively. The pathogenesis of Alzheimer's disease (AD) involves complex interactions between genetic factors (such as single nucleotide polymorphisms, SNPs) and brain structures (quantified by structural magnetic resonance imaging, sMRI). However, traditional association analysis methods are prone to capturing spurious correlations that are covariant but not causal with disease progression stages. To provide a high-quality initial topological structure for the subsequent causal discovery module, this step first converts heterogeneous SNP and sMRI data into a unified graph-based representation, where nodes represent biological entities and edges represent strong statistical correlations between them. Understandably, this step abstracts complex biological features (brain region imaging features and genetic variations) into "nodes" in graph theory and quantifies the strength of their statistical correlations into "edges" in the graph, thereby transforming multimodal data into a topological representation suitable for graph neural network processing.

[0026] Specifically, two types of undirected graphs were constructed: sMRI-Graph (brain imaging graph) and gene-Graph (gene graph). The construction logic of the two is the same, but the biological entities they deal with are different.

[0027] For the construction of the sMRI-Graph: In this graph, each node represents a brain region. Using the AAL (Automatic Anatomical Labeling) atlas, a total of [number missing] sMRI regions were selected. One (in this embodiment) = 90) Representative brain regions are selected as nodes. For each brain region node (in A three-dimensional feature vector was defined for it. Here These represent the gray matter volume, white matter volume, and cerebrospinal fluid volume of the brain region, respectively. They are key anatomical indicators extracted from preprocessed sMRI images, reflecting the structural integrity of the brain region. The connectivity between nodes, i.e., the edges of the graph, is defined by calculating the Pearson correlation coefficient between feature vectors of two different brain regions.

[0028] Specifically, for any two nodes and Its edge features This is the correlation coefficient calculated from the feature vectors of their corresponding brain regions. The Pearson correlation coefficient measures the degree of linear correlation between two variables, with values ​​ranging from -1 to 1. To ensure that the constructed connections have biological significance rather than random noise, a threshold is set. Only when the absolute value of the correlation coefficient... Only at the node and Establish an edge between them. For sMRI-Graph, set... The aim is to preserve those strong functional or structural covariant connections in brain structural networks.

[0029] For the construction of the gene-Graph: In this graph, each node represents a selected SNP site. SNPs refer to DNA sequence polymorphisms caused by a single nucleotide variation at the genome level, and are important markers for studying the genetic basis of complex diseases. SNPs are selected from preprocessed gene data. (same,) = 90) SNP sites potentially associated with Alzheimer's disease are used as nodes in the graph. Features of each SNP node. A 0-1-2 encoding method is used, meaning that the individual's genotype (e.g., AA, Aa, aa) is encoded as 0, 1, or 2. Similar to sMRI-Graph, the construction of edges in gene-Graph is also based on the Pearson correlation coefficient. The correlation coefficient between the 0-1-2 encoded sequences of each pair of SNP nodes (i.e., each pair of SNP loci) is calculated to obtain the edge features. Similarly, a threshold is set to filter strongly correlated edges; for gene-Graph, this is used. Due to genetic characteristics such as linkage disequilibrium between gene loci, the distribution of correlation coefficients may differ from that of imaging features. Therefore, a relatively low threshold was used to capture a wider range of genetic covariance patterns.

[0030] Ultimately, two fundamental undirected graphs were obtained: and .in, Represents a set of nodes. This represents the set of edges that satisfy a threshold condition. For sMRI-Graph, a threshold is set. For gene-Graph, set a threshold. Only when Only at the node and Establish an edge between them.

[0031] Since sMRI-Graph and gene-Graph are highly consistent in data structure (node ​​feature matrix and adjacency matrix) and construction process, the generalized graph will be used uniformly in the subsequent method description to simplify notation and derivation. A graph representing either mode can be used, and all processing steps can be applied in parallel to both modes.

[0032] Step S200: Use a temporally aware graph attention network to process the brain image map and the gene map respectively, generate a graph embedding vector for each sample, map the graph embedding vector to a continuous pseudo-temporal value, then allocate the samples to different stage buckets according to the pseudo-temporal value, and aggregate the graphs in each stage bucket to obtain a stage graph sequence containing node features and adjacency matrix. This step aims to infer the potential developmental stage of each sample in the continuous course of Alzheimer's disease (AD) from static individual brain imaging and gene maps, i.e., pseudo-temporal sequence, and to construct a staged graph sequence that reflects the dynamic evolution of the disease based on this. After converting neuroimaging and genetic data into graph representations, this step mainly includes three stages: node feature encoding, graph-level representation aggregation, and pseudo-temporal regression, and finally obtains the temporal graph sequence through a bucketing strategy.

[0033] First, for any sample Its graph structure is represented as ,in It is the node feature matrix ( The total number of nodes. (for node feature dimensions) It is an adjacency matrix constructed based on Pearson correlation; These are node features.

[0034] To capture the complex high-order nonlinear dependencies between nodes in the graph, a two-layer stacked graph attention network (GAT) is used for encoding: The first layer of GAT (Multi-head Attention): Employs a multi-head attention mechanism to enrich feature representation capabilities. Let the number of attention heads be... For the first Size, node with his neighbors Attention coefficient between Calculated through a shared attention mechanism:

[0035] in, It is the learnable parameter vector of the attention head. This indicates the batch size during model training, i.e., the number of samples processed at a time. It is a linear transformation matrix ( , This represents the transformed output dimension of each attention head in a Graph Attention Network (GAT). and Representing nodes respectively and neighboring nodes eigenvectors, symbols The negative slope of the LeakyReLU activation function represents the vector concatenation operation. If there is no edge connecting the two nodes (i.e. , Representing the adjacency matrix The Middle Okay, number (elements of the column), then set Then, the Softmax function is used to optimize the nodes. The attention coefficients of all neighbors are normalized to obtain the final attention weights. :

[0036] in, Represents a node The set of neighbors. The output of each attention head is a weighted sum of the features of its neighboring nodes. Finally, The outputs of each layer are concatenated and then processed using Dropout (to prevent overfitting) and ReLU activation function to obtain the node features of the first layer output. :

[0037] in, This indicates the concatenation operator, specifically referring to... The output feature vectors of each attention head are concatenated; at this time, the node feature dimension is changed from the input dimension. Expand to (That is, 4*32=128 dimensions).

[0038] Second-layer GAT (single-head attention and dimensionality reduction): Employs a single-head attention mechanism ( This layer aims to fuse and reduce the dimensionality of the high-dimensional features generated in the first layer. It uses a novel transformation matrix. Project features onto the target hidden space dimension. .node The final embedding representation We obtain the following by weighted summation of neighborhood features:

[0039] in, These are the normalized attention weights calculated for the second layer; Represents a node The output feature vector after processing by the first layer of the Graph Attention Network (GAT). Output feature matrix. It incorporates multi-hop structure information from the entire graph, representing the high-level abstract features of each node.

[0040] To obtain a global representation at the sample level, global average pooling is performed on the final embeddings of all nodes, and the arithmetic mean is calculated to obtain the graph embedding vector. :

[0041] Next, a fully connected regression layer will be used to... It is mapped to a scalar and uses a scaled sigmoid function. By strictly constraining the scalar within the interval (0, 2), the pseudo-time series value of the sample is obtained. :

[0042] in, and The parameters are for the regression layer. This design gives the predicted values ​​a clear biological meaning: 0 corresponds to the healthy control (HC) stage, 2 corresponds to the late stage of Alzheimer's disease (AD), and the continuous values ​​in between quantitatively characterize the potential continuous evolution process from health to the late stage of disease.

[0043] To capture the dynamic graph structure at different stages of disease progression, a bucketing strategy based on quasi-temporal sequence was designed.

[0044] Sample binning: The pseudo-time series range [0, 2] is uniformly divided into... Each interval (bucket). Based on the sample's pseudo-time series values. Assign it to the one that meets the conditions The In the buckets, among which and It is the boundary of the bucket; This represents the final, quasi-time series value of the sample after calibration, which is strictly constrained within the [0,2] interval and is used to characterize the potential development stage of the sample in the continuous process of the disease.

[0045] Phase graph aggregation: For the first phase Each non-empty bucket contains a set of samples. To obtain a stable graph representation representing the stage of the disease, average aggregation was performed at both the node feature and topology levels:

[0046] in, Indicates the first The node feature matrix after aggregating multiple time buckets; Indicates the index number of the current time bucket (or stage) (e.g., bucket 1, bucket 2); Indicates belonging to the sample set The The original node feature matrix of each sample; Indicates belonging to a set The The original adjacency matrix of each sample; Indicates the first The adjacency matrix after aggregating time buckets.

[0047] This aggregation reduces individual noise, resulting in a representative map for this stage. Arranging all the buckets in the diagram in order yields the sequence of stage diagrams. ,in, This represents the total number of time buckets, which is the number of intervals into which the pseudo-timeline is evenly divided.

[0048] Empty bucket handling (baseline blocking strategy): In practice, some pseudo-time series intervals may lack samples. To ensure sequence integrity and prevent the transmission of false information, the following strategy is adopted: Feature decay padding: Select the first non-empty bucket as the reference set. For empty buckets ,from Random subsampling of a small number of samples (number) ), calculate its nodal characteristic mean, and multiply it by a decay coefficient. (In this embodiment, the value is set to 0.5), resulting in the filled feature matrix. :

[0049] in, A subset consisting of randomly sampled samples. Indicates the first The feature matrix of each individual provides low-intensity background noise while clearly identifying the stage as an "unobserved state".

[0050] Structure nullification blocking: This involves removing the adjacency matrix of empty buckets. Forced to be set to an all-zero matrix .

[0051]

[0052] In the message passing mechanism of graph neural networks, this is equivalent to severing the connection between all nodes in that "stage", fundamentally eliminating the risk of generating false causal edges across data gaps.

[0053] Through this step, a continuous disease evolution axis was successfully reconstructed from static cross-sectional data, and a series of graphs reflecting different disease development stages were obtained along this axis.

[0054] Step S300: Using the stage graph sequence, calculate the causal contribution of different nodes under multiple time lags through a time-aware prediction network, and generate a hybrid causal prior graph based on the causal contribution. This step utilizes the constraint of time lag—that is, the cause must occur before the effect—to distinguish between simple statistical associations and true causal relationships. Specifically, it is accomplished through four sub-steps: prediction error differencing, bidirectional nonlinear correction, multi-lag fusion, and graph construction based on net causal flow.

[0055] The input is a sequence of stage graphs. ,in , For the first The node feature matrix of the stage, This is the corresponding adjacency matrix.

[0056] First, a time-aware prediction network is constructed. The network consists of a multilayer perceptron (MLP, a type of feedforward neural network) and a ReLU activation function, used to fit the complex nonlinear dynamics of node states changing with "time" (stages).

[0057] For a given stage and time lag For node pairs Evaluation from source node To the target node Causal effects: Baseline prediction: using only the target node At the previous moment ( The historical state of ) To predict its current state Calculate the baseline forecast value. and its mean square error ;in It is a time-aware prediction network.

[0058] Enhanced prediction: Source node In the lag time ( The historical state of ) Prediction is introduced. To efficiently fuse features from different dimensions, an adaptive strategy is adopted, in which... For the stage and time lag Below is the fused feature vector used to enhance prediction: For multidimensional features (such as brain imaging), ), using channel splicing: .

[0059] For single-dimensional features (such as SNP, Signal superposition is used. .

[0060] Calculate the error of the enhanced prediction .

[0061] Calculate the initial causal contribution: at a specific lag Below, from arrive The causal contribution is defined as the average reduction in prediction error over all effective time steps:

[0062] in, To lag at a specific time Next, from the source node To the target node The initial causal contribution; In the stage The baseline prediction error (mean squared error, MSE); In the stage and lag Enhanced prediction error; ; A complete set of stage indexes; It is a lag An effective set of stages. The larger the value, the more likely it is to indicate the addition of... Historical information for prediction The more significant the improvement in the current state, the more likely it is that there is a problem. arrive The greater the likelihood of a causal relationship.

[0063] Directly by and The direction of correlation can be influenced by collinearity or unobserved confounding factors, leading to spurious symmetric correlations. Therefore, a correction network is introduced. (Also an MLP), biasing the initial causal contribution:

[0064] in, In lag Below, from arrive The corrected causal contribution.

[0065] Correction network Scoring based on reverse causality As input, by learning a complex nonlinear mapping, an estimated bias is output and then subtracted. This step effectively suppresses interference from redundant signals in bidirectional feedback, highlighting the causal driving force with a clear direction.

[0066] Biological regulatory mechanisms (such as gene expression or neural transmission) often have unknown and heterogeneous time delays. To automatically identify the most significant time windows of action, an energy-based weighted aggregation mechanism is used instead of a simple averaging.

[0067] First, calculate each lag. Relative importance (weight):

[0068] in, Lagging Adaptive weights, For a very small positive number (such as ), used to ensure numerical stability; For other lags The adjusted causal contribution. Weights This reflects the relative magnitude of the causal strength under this lag.

[0069] The node is obtained by weighting and summing the corrected causal contributions for all lags. right Overall causal contribution:

[0070] in, As the source node To the target node The overall causal contribution; To ultimately determine the type of connection between nodes (directed or undirected), compute node pairs. Net causal driving force between :

[0071] in, It was calculated using the same method. right The overall causal contribution. Net driving force. The size and sign of the causal effect reflect the net direction and intensity of the causal influence.

[0072] Subsequently, a preset adaptive threshold is used. To determine the type of edge: like Then it is determined that there exists from point to directed edges .

[0073] like Then it is determined that there exists from point to directed edges .

[0074] like If so, then the edge between the two is determined to be undirected. This indicates that there is synchronous change or bidirectional feedback / co-driving at the current resolution, and it is impossible or unnecessary to distinguish a clear direction.

[0075] Through the above process, step S300 finally outputs a hybrid causal prior graph. This graph not only includes explicit causal paths inferred from time lag (directed edges), but also retains those interactions that are strongly correlated but have unclear directionality (undirected edges), thus more realistically and comprehensively depicting the complex multimodal biological network topology behind Alzheimer's disease.

[0076] Step S400: Parameterize the adjacency matrix elements corresponding to the directed edges in the hybrid causal prior graph, evaluate the prediction error before and after flipping the direction of the directed edges using a graph convolutional network, calculate the flipping reward, and perform direction correction on the hybrid causal prior graph based on the flipping reward to obtain a refined causal graph. Although time lag analysis can infer most causal directions, it is still susceptible to "pseudo-synchronization" or misjudgment of causal directions due to limitations in data sampling rate or local statistical bias. This step introduces counterfactual reasoning and gradient-based optimization to examine the rationality of each directed edge from the perspective of the global prediction task, eliminate false causal directions, and output a refined causal graph.

[0077] To preserve the prior causal topology while allowing the model to automatically learn and adjust edge confidence via gradient descent, the elements corresponding to directed edges in the adjacency matrix are first parameterized. This is based on the adjacency matrix of the hybrid causal prior graph output from step S300. (Including a set of directed edges) (and undirected edges), introducing a learnable parameter matrix. Then, construct the parameterized adjacency matrix. :

[0078] in, (A value of 0.5) is the mixing coefficient, used to balance the fixed prior structure. With learnable parameterization The weights between them. It is the Sigmoid activation function, which will activate the parameter matrix. Mapping to the (0, 1) interval represents the continuous confidence level of the existence of the directed edge. This operation only "softens" directed edges, while undirected edges remain unchanged, thus respecting the initial causal discovery results during the optimization process while allowing room for correction.

[0079] To quantify the interpretability of the current parametric graph structure for node features, it is applied to a downstream graph-level prediction task (e.g., predicting the disease stage to which the graph belongs). The specific process is as follows: First, for the parameterized adjacency matrix Add self-loops (identity matrix) Then perform symmetric normalization to obtain the normalized adjacency matrix. ,in, It is the degree matrix. This operation helps stabilize the training of graph neural networks.

[0080] Next, the node feature matrix With normalized adjacency matrix Input a Graph Convolutional Network (GCN). The GCN generates new node representations by aggregating information from neighboring nodes. .

[0081] Then, the node representation Global average pooling is performed to obtain the global representation vector of the graph, which is then passed through a multilayer perceptron (MLP) prediction head to output graph-level predictions. , .

[0082] Finally, calculate the prediction loss. That is, the predicted value Compared with the actual target value Mean squared error (MSE) between the corresponding category labels (as shown in the figure): This loss measures the ability of the current causal graph structure to support accurate predictions.

[0083] To examine a directed edge (Assuming the original direction is) To determine if the direction is correct, a counterfactual reversal strategy is employed: the direction of that side is forcibly reversed. (i.e., order) , Meanwhile, keeping all other edges in the graph unchanged, construct a counterfactual graph adjacency matrix. Based on this flipped graph, the loss for the graph-level prediction task is recalculated, denoted as the counterfactual prediction loss. .

[0084] According to the basic principles of structural causal modeling (SCM), if the original direction If this is correct, then reversing the direction will disrupt the true generation mechanism behind the data, leading to a decrease in prediction performance (i.e., Conversely, if the prediction error decreases after flipping, it strongly suggests that the original direction may be incorrect.

[0085] To encode this logic in the optimization, a one-sided flip penalty term is defined for each directed edge. :

[0086] in, It is a tolerance threshold used to prevent minor random fluctuations from triggering corrections. The function ensures that the loss is significantly lower only when the flipped value is significantly lower. Only when this condition is met will a positive penalty value be generated, thereby encouraging the optimization process to avoid this "profitable" wrong direction.

[0087] The overall optimization objective of the model consists of three parts:

[0088] in, This represents the total loss function (or total optimization objective) of the model during the training phase of the reverse causality confirmation module. To ensure that the graph structure has good predictive capabilities; For reverse confirmation loss, weighting coefficient Controlling the intensity of its impact on the total loss, its role is to penalize those edges that exhibit erroneous orientation in counterfactual tests; For L2 regularization, the regularization coefficient is... Used for control parameter matrix To reduce complexity and prevent overfitting, Let f(x) denote the squared Frobenius norm of the matrix.

[0089] By minimizing The model simultaneously optimizes prediction accuracy and the reasonableness of causal direction, driving the parameter matrix. Adjust the structure to represent the correct causal direction.

[0090] After training converges, based on the optimal parameters Perform deterministic refining steps to transform continuous parameter confidence into discrete topology corrections.

[0091] For each prior directed edge Calculate its reversal profit :

[0092] Intuitively reflects the movement of the edge from Flip to The resulting change in predicted loss. If This indicates that the loss is reduced after flipping, meaning that the flipping direction is more reasonable.

[0093] To avoid misjudgments caused by numerical noise, a strict threshold for reversal payoff is set. (In this embodiment, it is set to 0.008). The final refined causal graph adjacency matrix. Generate according to the following rules:

[0094] in, This indicates that in the adjacency matrix of the hybrid causal prior graph generated in step S300, from the source node... To the target node The connection status; This indicates that in the adjacency matrix of the same mixed causal prior graph, from the target node... To the source node The connection status.

[0095] That is, only when the flipping benefit of an edge significantly exceeds the threshold is its original direction determined to be a false positive and a physical direction flipping correction is performed; otherwise, the original direction inferred in step S300 is retained.

[0096] Through this step, the model completes the robustness verification and calibration of the initial causal graph, and the final output is... It is a refined causal graph with a lower false positive rate and more reliable causal direction.

[0097] Step S500: Based on the refined causal graph, perform causal enhancement on the node features in the stage graph sequence, and apply a multi-head cross-attention mechanism to perform bidirectional interactive fusion of the enhanced brain image modality features and gene modality features; This step comprises three sub-processes: causal topology enhancement, intramodal importance filtering (gated attention pooling), and cross-modal bidirectional semantic alignment.

[0098] For a given batch of samples (batch size) Each sample is determined according to its corresponding quasi-time series stage. Extract the corresponding refined causal adjacency matrix Two feature representations are constructed for each sample: Individual representation: Preserving the original node feature matrix .in For the number of nodes, For node feature dimensions (such as in sMRI) In the gene This representation directly preserves the original heterogeneity information of each subject in terms of anatomical structure or genetic coding.

[0099] Causal context representation: generated by projecting individual characteristics onto the causal topological space of their corresponding stage, calculated as follows: ;in, For causal context representation; The original features are used. This operation utilizes a refined causal adjacency matrix. By weighted aggregation of node features, each node's features incorporate information from other nodes with which it has a causal relationship, thereby encoding group-level causal dynamics into individual representations.

[0100] Then, the two representations are fused. First, the original features are... By mapping to a unified hidden space through a linear projection layer, ReLU activation function, and layer normalization, the basic features are obtained. :

[0101] in, Layer normalization is used to eliminate dimensional differences between different biometric measurements; This represents a linear transformation layer, also known as a fully connected layer or an affine transformation layer.

[0102] Then, a residual connection mechanism is used to integrate causal context features. After linear projection, with weights Overlay The enhanced node features are obtained above. .

[0103]

[0104] Here It is an adjustable hyperparameter used to control the strength of the contribution of causal prior information to the final feature.

[0105] From Automatically focusing on the most disease-relevant features within each node (e.g., 90 brain regions or SNP sites), a gated attention pooling layer is introduced. This mechanism is applied to each node. Calculate an attention weight The calculation formula is as follows:

[0106] in, It is a matrix The corresponding number in the middle The feature vector of each node; and These are two learnable weight matrices, used for two parallel nonlinear transformation branches respectively; and (Sigmoid) is an activation function. This represents element-wise multiplication. The output of the Sigmoid function serves as a "gate," with a value between 0 and 1, which can suppress the contribution of node features irrelevant to the task. The function ensures that the sum of all weights is 1.

[0107] Finally, the sample-level graph embedding vector is obtained by weighted summation. This vector achieves semantic enhancement from local node features to macroscopic disease states. This operation is performed separately for the SNP and sMRI modalities to obtain their respective global vectors. and .

[0108] To capture the complex nonlinear synergistic effects between genotype and brain structural phenotype, a multi-head cross-attention mechanism is used for bidirectional interaction, rather than simple splicing.

[0109] SNP-guided image enhancement ( ): Using global vectors of brain imaging As a query (Q), using the SNP global vector This is analogous to asking, "In the currently observed brain structural patterns, which genetic variations provide the most critical explanation?" (Key, K; Value, V) Image-guided gene enhancement ( Conversely, with For query, For keys and values. This is equivalent to asking: "Given an individual's genetic risk background, which brain regions show the most significant structural changes that reflect this susceptibility?" Specifically, for each interaction direction (taking modality A querying modality B as an example): For each interaction direction (taking modality A querying modality B as an example), firstly, the input global representation vector... , The attention head is projected onto the attention space through a learnable linear mapping layer. (This embodiment is set) The calculation of the query, key, and value vectors is as follows:

[0110]

[0111]

[0112] in, For query vector; The key vector; It is a value vector; The primary query modality, To assist the reference mode, It is the first The weight matrix of the size, the subspace dimension The multi-head mechanism allows the model to capture diverse patterns of association between the brain and genes in parallel across different feature dimensions. The system calculates query vectors... With key vector The dot product is used to quantify the interaction strength between two modes. To prevent gradient vanishing, a scaling factor is introduced. And normalization is performed using the Softmax function:

[0113] in, For the first The attention weight matrix (or attention distribution) of each attention head; For the first Key vectors of attention heads The transpose of .

[0114] This scoring matrix represents the degree of attention modality A pays to specific information in modality B. In a biological sense, it identifies which genetic traits provide the strongest explanation for current brain structural changes or which areas of the brain show the most phenotypic expression of a particular genetic risk.

[0115] The calculated weights are used to define the value vector of the auxiliary modes. Perform weighted reconstruction to extract cross-modal context information:

[0116] in, For the first The output of each attention head.

[0117] The outputs of all heads are concatenated and then mapped through an output projection layer. Finally, residual connections and layer normalization are used to generate an enhanced interactive representation.

[0118] in, The enhanced interactive representation (final output); For layer normalization function; The output of the first attention head; For the first The output of each attention head.

[0119] Through the aforementioned two-way interaction, a bimodal augmentation feature with semantic alignment and complementary information was ultimately obtained. and .

[0120] Step S600: Input the fused dual-modal global feature vector into the classifier to complete the classification prediction of different stages of AD.

[0121] First, the two enhanced modal features obtained in step S500 are... and Perform horizontal concatenation to form a global fusion vector:

[0122] in, For global fusion vector, It is the total dimension of the features after splicing.

[0123] then, The system employs a linear projection layer and uses the GELU activation function for non-linear transformation and feature recombination. GELU is a commonly used smooth and non-monotonic activation function in deep learning, which facilitates richer feature representations.

[0124] Then, to enhance the model's generalization ability and prevent overfitting, a Dropout layer with a dropout rate of 0.5 is introduced to randomly "shut down" a portion of neurons during training.

[0125] Finally, the processed features are fed into the output layer (a linear layer) and mapped to the target class space. In the three binary classification tasks of this embodiment (AD vs. HC, AD vs. MCI, MCI vs. HC), the number of categories... The output of this layer is the unnormalized raw score, i.e., the logical value, which is used for subsequent loss function calculation.

[0126] In step S200, when training the quasi-temporal prediction model, a joint loss was used:

[0127] in, For joint losses, It is the mean squared error loss, which makes the predicted value close to the discrete clinical label (HC=0, MCI=1, AD=2). It is the marginal ranking loss, which forces the predicted values ​​of samples from different classes to maintain at least a gap. (AD vs. HC quest) ), tasks involving MCI This expands the predicted distribution. It is the balance coefficient.

[0128] To enable the model to learn the potential evolution sequence consistent with the actual disease stages, the predicted values ​​are first minimized using the MSE loss. With discrete clinical labels Differences between them:

[0129] Here, B represents the batch size, which ensures that the overall distribution trend of the pseudo-time series is positively correlated with disease severity. Specifically, each sample has its own clinical diagnostic category. To map discrete diagnostic categories to a continuously modelable disease progression axis, the following ordered numerical encoding method is used:

[0130] in, For healthy controls; Mild cognitive impairment; For Alzheimer's disease.

[0131] Relying solely on MSE can easily lead to predicted values ​​collapsing around the label, making it difficult to form a continuous distribution. Therefore, a margin-based ranking loss is introduced. In each training batch, all sample pairs are constructed... Only for sample pairs with different labels (i.e. Calculate sorting constraints:

[0132] in, For the set of valid sample pairs, Indicates the actual relative order. For a preset margin, samples from different disease stages are forced to maintain at least a certain distance along the pseudo-timeline of the disease. The distance effectively stretches the distribution of predicted scores, preventing confusion between samples of different classes at the boundaries. Notably, considering the varying degrees of significance of pathological differences across diagnostic tasks, a clinically-driven adaptive setting was employed: a strong interval constraint was applied to the AD vs. HC task, where pathological boundaries were significant. To maximize inter-class variance; however, for tasks involving the prodromal stage (AD vs. MCI and MCI vs. HC), the constraints were moderately relaxed considering the continuity of disease progression and the inherent ambiguity of boundaries. 0.7), to ensure separability while accommodating potential biological heterogeneity and preventing the model from overfitting to difficult samples.

[0133] To address the common class imbalance problem in medical data, the final classifier uses Focal Loss as the loss function. :

[0134] in, This is the class balance coefficient, used to adjust the weight ratio of positive and negative samples in the total loss. When the true label... hour, ;when hour, ( (This is the result of mapping the original logical value through the Sigmoid activation function). It directly reflects the model's confidence in classifying the current sample. is a non-negative, adjustable hyperparameter used to control the decay rate of simple sample weights. ).

[0135] Figure 2This is a framework diagram for causal modeling to eliminate spurious correlations. The model input consists of multimodal data from SNPs and sMRI. In the bimodal branch, the model first constructs a disease pseudo-temporal axis using a bucketing strategy to recover dynamic evolutionary information from static cross-sectional data. The core path includes a time-lag causal discovery module and a reverse causal confirmation module, which work together to infer robust causal directions and eliminate spurious correlations. Finally, the fused multi-view causal features are fed into a hierarchical classifier, which then feeds into a classification module to perform predictions for three independent binary classification tasks: AD vs. HC, AD vs. MCI, and MCI vs. HC. This effectively addresses the feature aliasing challenge between different pathological stages and achieves accurate diagnosis of different stages of AD.

[0136] This invention reconstructs the continuous disease evolution process from static cross-sectional data by constructing a bimodal undirected graph and generating a pseudo-temporal axis of disease using a time-aware graph attention network. Furthermore, it identifies directional causal contributions between nodes through a time-lag causal judgment module, generating a hybrid causal prior graph. Innovatively, it introduces a counterfactual flip-based reverse causal confirmation module to perform gradient verification and correction of causal directions, effectively eliminating spurious correlations caused by data bias or collinearity. Based on this, it employs causal topology enhancement and multi-head cross-attention mechanisms to achieve deep semantic alignment and complementary fusion between brain images and genetic features. Finally, by integrating the above causal constraints and fusion representations into a classifier, it significantly improves the ability to capture early subtle pathological patterns and diagnostic accuracy in classification tasks at different stages of Alzheimer's disease, while also enhancing the model's interpretability and cross-domain generalization ability. This provides an effective solution for reliable and interpretable disease stage prediction and mechanism analysis in the absence of longitudinal follow-up data.

[0137] To verify the effectiveness and superiority of the above method, the implementation process of the present invention will be further explained below with reference to specific datasets and experimental settings.

[0138] The data used in this embodiment comes from the ADNI database, a large public database of AD (http: / / adni.loni.usc.edu / ), and includes the ADNI-1, ADNI-GO, ADNI-2, and ADNI-3 queues. SNP data is based on Illumina 2.5M and Illumina Omni Quad arrays, and image data is based on T1W1-3D-MP RAGE. Detailed data is shown in Table 1.

[0139] Table 1 Demographic Information

[0140] SNP data underwent standard QC via PLINK, including missing rate filtering (removing SNPs and samples with ≥2% missing data), gender concordance testing, MAF ≥5% screening, and Hardy–Weinberg equilibrium test (p ≤ 1 × 10⁻⁶). -4 (Removal) and abnormal heterozygosity removal. Without considering covariates, a logistic linear regression model was then used to evaluate SNP-AD associations, and combined with AlzData highly correlated sites, resulting in 90 SNPs. Finally, the `recodeA` command in plink was used to convert these 90 SNPs into 0-1-2 encoded genotypes, which served as node features for constructing the gene-graph structure.

[0141] sMRI data were used to obtain brain region-level features based on the AAL template: For 3D images, DICOM was first converted to NII, and then statistical parametric mapping (SPM) and computational anatomy (CAT) were used to perform voxel-based texture analysis (VBM) preprocessing on the 3D brain images of all subjects. In T1-weighted image preprocessing, the images were first uniformly segmented and spatially normalized, and then registered to the MNI152 standard space. Based on this, 116 brain structural features were extracted from gray matter, white matter, and cerebrospinal fluid probability maps using the AAL brain atlas. Further structural validity screening was conducted, and finally 90 brain regions were determined as node features for constructing the sMRI-Graph structure.

[0142] This framework is implemented in PyTorch and runs on a workstation equipped with a single NVIDIA GeForce RTX 3090 GPU (24GB VRAM). To ensure robustness and reproducibility of the evaluation, a hierarchical 5-fold cross-validation strategy was employed, with the random seed fixed at 2024. Network parameters were updated using the Adam optimizer, and the batch size was set to 32. The initial learning rate was set to 1e-4, along with a weight decay of 1e-4 to prevent overfitting. The maximum number of training iterations was 150, and early stopping with a patience value of 20 was enabled.

[0143] Specific hyperparameters were configured for different functional stages to balance the optimization objectives.

[0144] In constructing the disease pseudo-temporal axis, to eliminate distribution clustering caused by Sigmoid soft saturation and ensure full coverage of the pseudo-temporal spectrum, a secondary calibration based on the training set extrema was introduced.

[0145] in, This is the final pseudo-timing value after secondary calibration; This is the raw, uncalibrated pseudo-time series prediction value of the sample, output by the time-aware GAT model in step S200. To prevent small constants from being divided by zero, The cutoff function ensures that the final pseudo-time series values ​​for all samples are accurate. The samples must fall strictly within the range [0,2]. This operation eliminates scale differences while maintaining the relative order of the samples, providing a unified metric for subsequent equally spaced binning.

[0146] In the Reverse Causal Confirmation module (RCCM), the regularization hyperparameters lambda and alpha are set to 0.5 and 0.1, respectively.

[0147] To fully evaluate the diagnostic performance of the ES-Causal framework, stratified 5-fold cross-validation was performed on three independent binary clinical tasks: Alzheimer's disease vs. healthy controls (AD vs. HC), AD vs. mild cognitive impairment (AD vs. MCI), and MCI vs. healthy controls (MCI vs. HC). Tables 2-4 below report in detail the key assessment measures for each task in the 5-fold validation: mean and standard deviation of accuracy (ACC), sensitivity (SEN), specificity (SPE), and area under the curve (AUC).

[0148] In the AD vs. HC task, the method of this invention achieved a mean accuracy of 90.65% and an AUC of 94.39%, demonstrating its strong ability to identify typical pathological features. In the more challenging early diagnosis tasks (AD vs. MCI and MCI vs. HC), the method maintained robust performance, with mean accuracies of 73.10% and 76.67%, respectively. Notably, the standard deviations of all indicators across different folds remained low (e.g., the standard deviation of accuracy for the AD vs. HC task was only ±1.98%). This low standard deviation reflects the excellent classification stability and generalization ability of ES-Causal, indicating that the method can effectively eliminate spurious correlations in specific data partitions, thereby capturing causal pathological features with cross-domain robustness.

[0149] Table 2. Classification performance of the proposed method evaluated by stratified 5-fold cross-validation (HC vs AD)

[0150] Table 3. Classification performance of the proposed method evaluated by hierarchical 5-fold cross-validation (MCI vs AD)

[0151] Table 4. Classification performance of the proposed method evaluated by stratified 5-fold cross-validation (HC vs MCI)

[0152] In the embodiments, the proposed ES-Causal framework was systematically compared with several representative state-of-the-art (SOTA) methods to validate its superiority in AD diagnosis. These baseline methods were carefully selected, covering a wide range from basic graphical models to specialized multimodal imaging genetic networks, aiming to comprehensively evaluate the contribution of our method.

[0153] To ensure fairness in the comparisons, all methods employed the same dataset partitioning and hierarchical 5-fold cross-validation protocol. GCN, BrainGNN, and CGB were trained using only sMRI features, while IG-GCN and the method of this invention utilized both SNP and sMRI data to evaluate the value of multimodal integration.

[0154] GCN is a fundamental graph convolutional neural network, serving as a baseline for static graph learning. It aggregates information based on a fixed adjacency matrix derived from static feature correlations (such as Pearson correlation coefficients), treating all highly correlated connections as valid edges without distinguishing underlying causal mechanisms. BrainGNN is an interpretable GNN specifically designed for brain image analysis, treating brain regions as nodes and the functional correlations between them as edges. It is included in the comparison to represent ROI-aware medical image analysis methods utilizing clustering-based pooling techniques, but it still heavily relies on association-based functional connections rather than causal structures. CGB is a recently proposed framework that introduces causal graph learning into brain network analysis. It utilizes transfer entropy to discover causal relationships and employs geometric curvature for graph refinement. A comparison with CGB allows us to evaluate the specific advantages of the proposed disease progression-based quasi-temporal causality over signal-based causality in classification tasks. IG-GCN is a multimodal image genetics framework that constructs hierarchical SNP graphs and image sparse graphs based on gene ontology. This comparison is crucial for demonstrating the effectiveness of the causal fusion strategy of this invention relative to existing SOTA multimodal integration methods.

[0155] As shown in Tables 5-7, the method of this invention achieved the best accuracy (ACC) and sensitivity (SEN) in all three diagnostic tasks, demonstrating the model's effectiveness in disease identification and its strong competitiveness in other metrics. For example, in the highly challenging MCI vs. AD task (Table 3), the method of this invention achieved an accuracy of 73.10%, significantly outperforming multimodal competitors IG-GCN (59.31%) and the causal baseline CGB (65.17%). Notably, the method of this invention consistently maintained the lowest standard deviation in accuracy and AUC across all tasks (e.g., the standard deviation of HC vs. AD accuracy was only 1.98%). These results strongly confirm that TDCJM effectively captures the dynamic evolution of disease pathology, while RCCM successfully eliminates spurious correlations and false positive edges. The synergistic effect of these two modules facilitates the direct extraction of discriminative causal features, thereby achieving superior and more stable classification performance compared to purely association-based methods or conventional causal methods.

[0156] Table 5. Classification results of different models (HC vs AD)

[0157] Table 6. Classification results of different models (MCI vs AD)

[0158] Table 7. Classification results of different models (HC vs MCI)

[0159] According to another aspect of the embodiments of this application, an electronic device is also provided, including a processor and a memory, wherein the processor is configured to implement the steps of the method when executing a computer program stored in the memory.

[0160] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0161] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For instance, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection may be through some interfaces; the indirect coupling or communication connection between units or modules may be electrical or other forms.

[0162] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0163] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0164] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A causal graph neural network prediction method for eliminating spurious correlations in Alzheimer's disease (AD), characterized in that, Includes the following steps: Based on the Pearson correlation coefficient, a bimodal undirected graph of single nucleotide polymorphism sites and brain imaging data was constructed. Each single nucleotide polymorphism site and each brain region were defined as a graph node, and brain imaging graph and gene graph were constructed respectively. The brain image map and gene map are processed by a temporally aware graph attention network to generate a graph embedding vector for each sample. The graph embedding vector is then mapped to a continuous pseudo-temporal value. The samples are then assigned to different stage buckets according to the pseudo-temporal value. The graphs in each stage bucket are aggregated to obtain a staged graph sequence containing node features and an adjacency matrix. Using the aforementioned phased graph sequence, a time-aware prediction network is used to calculate the causal contribution of different nodes under multiple time lags, and a hybrid causal prior graph is generated based on the causal contribution. The adjacency matrix elements corresponding to the directed edges in the hybrid causal prior graph are parameterized. The prediction error before and after flipping the direction of the directed edges is evaluated using a graph convolutional network. The flipping reward is calculated, and the direction of the hybrid causal prior graph is corrected based on the flipping reward to obtain a refined causal graph. The node features in the stage graph sequence are causally enhanced based on the refined causal graph, and a multi-head cross-attention mechanism is applied to perform bidirectional interactive fusion of the enhanced brain image modality features and gene modality features. The fused bimodal global feature vector is input into the classifier to complete the classification and prediction of different stages of AD.

2. The causal graph neural network prediction method for eliminating spurious AD correlations as described in claim 1, characterized in that, In the steps of constructing a bimodal undirected graph of single nucleotide polymorphism sites and brain imaging data based on Pearson correlation coefficient: The nodes of the brain image map are brain regions. The feature vector of each brain region node includes gray matter volume, white matter volume and cerebrospinal fluid volume. The edge features between nodes are obtained by calculating the Pearson correlation coefficient between brain region features. The nodes of the gene map are single nucleotide polymorphism sites. The features of each site node are represented by 0-1-2 encoding to indicate the genotype. The edge features between nodes are obtained by calculating the Pearson correlation coefficient between site codes. Thresholds are set for the brain imaging map and the gene map respectively, and only edges with Pearson correlation coefficient absolute values ​​greater than the corresponding thresholds are retained as initial topological connections.

3. The causal graph neural network prediction method for eliminating spurious AD correlations as described in claim 1, characterized in that, The method for generating a graph embedding vector for each sample by processing the brain image map and gene map using a time-aware graph attention network includes: Two stacked graph attention layers are used to encode the initial features of the nodes; The first graph attention layer adopts a multi-head attention mechanism. It calculates the attention coefficients between nodes and their neighbors through linear transformation and LeakyReLU activation function, and normalizes them within the set of neighbor nodes using the Softmax function. The normalized multi-head attention outputs are then concatenated to obtain the first layer of node features. The second graph attention layer adopts a single-head attention mechanism to fuse and reduce the dimensionality of the first-layer node features output by the first layer. The neighborhood features are weighted and summed through new attention weights to obtain the node embedding that fuses the structural information of the entire graph. Global average pooling is performed on the embedding features of all nodes after processing by the two graph attention layers to calculate the arithmetic mean of the feature vectors of all nodes and generate sample-level graph embedding vectors.

4. The causal graph neural network prediction method for eliminating spurious AD correlations as described in claim 1, characterized in that, The graph embedding vector is mapped to continuous quasi-temporal values, which are calculated using the following formula: in, Represents the pseudo-time series value of the sample. This represents the Sigmoid activation function. This is the weight vector of the fully connected regression layer. Let be the graph embedding vector of the sample. For bias terms; The pseudo-time series values ​​output by the regression layer are activated by the Sigmoid activation function. The constraint is within the range of 0 to 2, where 0 corresponds to the healthy control stage and 2 corresponds to the late stage of AD.

5. The causal graph neural network prediction method for eliminating spurious AD correlations as described in claim 1, characterized in that, The method for allocating samples to different stage buckets based on the pseudo-time series values, and aggregating the graphs within each stage bucket to obtain a stage graph sequence containing node features and an adjacency matrix includes: The continuous range of the pseudo-time series value is evenly divided into K stage bucket intervals; Each sample is assigned to the corresponding stage bucket based on its pseudo-time series value. For each non-empty stage bucket, the node feature matrix and adjacency matrix of all samples in each non-empty stage bucket are averaged and aggregated to obtain the node feature matrix and adjacency matrix of that non-empty stage bucket. For empty stage buckets, a baseline blocking strategy is adopted: its node feature matrix is ​​obtained by randomly selecting sub-samples from the first non-empty stage bucket, calculating the mean of node features, and then multiplying it by the decay coefficient, and its adjacency matrix is ​​set to an all-zero matrix. The graph representations obtained by aggregating buckets at each stage are arranged in stage order to form a staged graph sequence containing a node feature matrix and an adjacency matrix.

6. The causal graph neural network prediction method for eliminating spurious AD correlations as described in claim 1, characterized in that, Methods for calculating the causal contribution of different nodes under various time lags include: Calculate the baseline prediction error of the target node when only its own historical state is used; Calculate the enhanced prediction error after incorporating the historical state of the source node; The initial causal contribution is calculated based on the difference between the baseline prediction error and the enhanced prediction error; A correction network constructed using a multilayer perceptron is used to correct the initial causal contribution to obtain the final causal contribution.

7. The causal graph neural network prediction method for eliminating spurious AD correlations as described in claim 1, characterized in that, The method for evaluating the prediction error before and after flipping the direction of a directed edge using a graph convolutional network, calculating the flipping gain, and correcting the orientation of the mixed causal prior graph based on the flipping gain to obtain a refined causal graph includes: A learnable parameter matrix is ​​introduced and mixed with the adjacency matrix of the hybrid causal prior graph to generate a parameterized adjacency matrix. Based on the parameterized adjacency matrix and node features, the original prediction error of the graph-level prediction task is calculated using a graph convolutional network. For each directed edge in the hybrid causal prior graph, a counterfactual flip graph is constructed, wherein the direction of the directed edge is forcibly flipped, and the flip prediction error is calculated based on the counterfactual flip graph. Calculate the flipping reward for each directed edge based on the original prediction error and the flipping prediction error; and For edges whose flipping gains exceed a preset threshold, their direction is corrected to a reverse connection in the hybrid causal prior graph, thereby obtaining the refined causal graph.

8. The causal graph neural network prediction method for eliminating spurious AD correlations as described in claim 1, characterized in that, Methods for bidirectional interactive fusion of enhanced brain imaging modal features and genetic modal features using multi-head cross-attention mechanisms include: Using the global feature vector of the brain imaging modality as the query and the global feature vector of the gene modality as the key and value, information related to brain imaging in the gene modality is extracted through a multi-head attention mechanism to generate image-guided gene enhancement features. Using the global feature vector of the gene modality as the query and the global feature vector of the brain imaging modality as the key and value, the gene-related information in the brain imaging modality is extracted through the multi-head attention mechanism to generate gene-guided image enhancement features. For each attention head, the query vector, key vector, and value vector are calculated separately, and the attention weights are obtained by dot product and Softmax normalization. The value vector is weighted and aggregated using the attention weights to obtain the interactive output of each head; The outputs of all attention heads are concatenated, mapped through a linear projection layer, and then subjected to residual connections and layer normalization to obtain the final bidirectional interactive fusion feature.

9. The causal graph neural network prediction method for eliminating spurious AD correlations as described in claim 1, characterized in that, Prior to the step of constructing the bimodal undirected graph, a data preprocessing step is also included: Brain region features are extracted from the brain imaging data based on the AAL template; The single nucleotide polymorphism (SNP) site data were subjected to quality control and screening, and the genotypes were converted into numerical codes.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program is executed by the processor to perform the steps of the causal graph neural network prediction method for eliminating spurious AD correlations as described in any of claims 1-9.

Images